diff --git "a/openai_whisper-large-v2_949MB/TextDecoder.mlmodelc/model.mil" "b/openai_whisper-large-v2_949MB/TextDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/openai_whisper-large-v2_949MB/TextDecoder.mlmodelc/model.mil" @@ -0,0 +1,9627 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}})] +{ + func main(tensor cache_length, tensor decoder_key_padding_mask, tensor encoder_output_embeds, tensor input_ids, tensor key_cache, tensor kv_cache_update_mask, tensor value_cache) { + tensor var_80_axis_0 = const()[name = tensor("op_80_axis_0"), val = tensor(0)]; + tensor var_80_batch_dims_0 = const()[name = tensor("op_80_batch_dims_0"), val = tensor(0)]; + tensor var_80_validate_indices_0 = const()[name = tensor("op_80_validate_indices_0"), val = tensor(false)]; + tensor embed_tokens_weight_to_fp16 = const()[name = tensor("embed_tokens_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor var_80_cast_fp16 = gather(axis = var_80_axis_0, batch_dims = var_80_batch_dims_0, indices = input_ids, validate_indices = var_80_validate_indices_0, x = embed_tokens_weight_to_fp16)[name = tensor("op_80_cast_fp16")]; + tensor var_84_axis_0 = const()[name = tensor("op_84_axis_0"), val = tensor(0)]; + tensor var_84_batch_dims_0 = const()[name = tensor("op_84_batch_dims_0"), val = tensor(0)]; + tensor var_84_validate_indices_0 = const()[name = tensor("op_84_validate_indices_0"), val = tensor(false)]; + tensor embed_positions_weight_to_fp16 = const()[name = tensor("embed_positions_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132774528)))]; + tensor cache_length_to_int16_dtype_0 = const()[name = tensor("cache_length_to_int16_dtype_0"), val = tensor("int16")]; + tensor cast_0 = cast(dtype = cache_length_to_int16_dtype_0, x = cache_length)[name = tensor("cast_0")]; + tensor var_84_cast_fp16_cast_int16 = gather(axis = var_84_axis_0, batch_dims = var_84_batch_dims_0, indices = cast_0, validate_indices = var_84_validate_indices_0, x = embed_positions_weight_to_fp16)[name = tensor("op_84_cast_fp16_cast_int16")]; + tensor hidden_states_1_cast_fp16 = add(x = var_80_cast_fp16, y = var_84_cast_fp16_cast_int16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_98_axes_0 = const()[name = tensor("op_98_axes_0"), val = tensor([2])]; + tensor var_98_cast_fp16 = expand_dims(axes = var_98_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_98_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([3])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_98_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280])]; + tensor var_103_axis_0 = const()[name = tensor("op_103_axis_0"), val = tensor(1)]; + tensor var_103_cast_fp16_0, tensor var_103_cast_fp16_1, tensor var_103_cast_fp16_2, tensor var_103_cast_fp16_3, tensor var_103_cast_fp16_4, tensor var_103_cast_fp16_5, tensor var_103_cast_fp16_6, tensor var_103_cast_fp16_7, tensor var_103_cast_fp16_8, tensor var_103_cast_fp16_9, tensor var_103_cast_fp16_10, tensor var_103_cast_fp16_11, tensor var_103_cast_fp16_12, tensor var_103_cast_fp16_13, tensor var_103_cast_fp16_14, tensor var_103_cast_fp16_15, tensor var_103_cast_fp16_16, tensor var_103_cast_fp16_17, tensor var_103_cast_fp16_18, tensor var_103_cast_fp16_19, tensor var_103_cast_fp16_20, tensor var_103_cast_fp16_21, tensor var_103_cast_fp16_22, tensor var_103_cast_fp16_23, tensor var_103_cast_fp16_24, tensor var_103_cast_fp16_25, tensor var_103_cast_fp16_26, tensor var_103_cast_fp16_27, tensor var_103_cast_fp16_28, tensor var_103_cast_fp16_29, tensor var_103_cast_fp16_30, tensor var_103_cast_fp16_31 = split(axis = var_103_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_103_cast_fp16")]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280, 1280])]; + tensor var_138_axis_0 = const()[name = tensor("op_138_axis_0"), val = tensor(1)]; + tensor var_138_cast_fp16_0, tensor var_138_cast_fp16_1, tensor var_138_cast_fp16_2, tensor var_138_cast_fp16_3, tensor var_138_cast_fp16_4, tensor var_138_cast_fp16_5, tensor var_138_cast_fp16_6, tensor var_138_cast_fp16_7, tensor var_138_cast_fp16_8, tensor var_138_cast_fp16_9, tensor var_138_cast_fp16_10, tensor var_138_cast_fp16_11, tensor var_138_cast_fp16_12, tensor var_138_cast_fp16_13, tensor var_138_cast_fp16_14, tensor var_138_cast_fp16_15, tensor var_138_cast_fp16_16, tensor var_138_cast_fp16_17, tensor var_138_cast_fp16_18, tensor var_138_cast_fp16_19, tensor var_138_cast_fp16_20, tensor var_138_cast_fp16_21, tensor var_138_cast_fp16_22, tensor var_138_cast_fp16_23, tensor var_138_cast_fp16_24, tensor var_138_cast_fp16_25, tensor var_138_cast_fp16_26, tensor var_138_cast_fp16_27, tensor var_138_cast_fp16_28, tensor var_138_cast_fp16_29, tensor var_138_cast_fp16_30, tensor var_138_cast_fp16_31 = split(axis = var_138_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_138_cast_fp16")]; + tensor var_176 = const()[name = tensor("op_176"), val = tensor(3)]; + tensor var_183 = const()[name = tensor("op_183"), val = tensor(1)]; + tensor var_184 = const()[name = tensor("op_184"), val = tensor(true)]; + tensor var_196 = const()[name = tensor("op_196"), val = tensor([1])]; + tensor channels_mean_1_cast_fp16 = reduce_mean(axes = var_196, keep_dims = var_184, x = inputs_1_cast_fp16)[name = tensor("channels_mean_1_cast_fp16")]; + tensor zero_mean_1_cast_fp16 = sub(x = inputs_1_cast_fp16, y = channels_mean_1_cast_fp16)[name = tensor("zero_mean_1_cast_fp16")]; + tensor zero_mean_sq_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = zero_mean_1_cast_fp16)[name = tensor("zero_mean_sq_1_cast_fp16")]; + tensor var_200 = const()[name = tensor("op_200"), val = tensor([1])]; + tensor var_201_cast_fp16 = reduce_mean(axes = var_200, keep_dims = var_184, x = zero_mean_sq_1_cast_fp16)[name = tensor("op_201_cast_fp16")]; + tensor var_202_to_fp16 = const()[name = tensor("op_202_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_203_cast_fp16 = add(x = var_201_cast_fp16, y = var_202_to_fp16)[name = tensor("op_203_cast_fp16")]; + tensor denom_1_epsilon_0 = const()[name = tensor("denom_1_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_1_cast_fp16 = rsqrt(epsilon = denom_1_epsilon_0, x = var_203_cast_fp16)[name = tensor("denom_1_cast_fp16")]; + tensor out_1_cast_fp16 = mul(x = zero_mean_1_cast_fp16, y = denom_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; + tensor obj_1_mean_0_to_fp16 = const()[name = tensor("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133921472)))]; + tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133924096)))]; + tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133926720)))]; + tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133929344)))]; + tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("obj_1_cast_fp16")]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; + tensor var_223 = const()[name = tensor("op_223"), val = tensor([1, 1])]; + tensor pretrained_out_1_pad_type_0 = const()[name = tensor("pretrained_out_1_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_1_pad_0 = const()[name = tensor("pretrained_out_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133931968))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134751232))), name = tensor("layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134751360)))]; + tensor pretrained_out_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_223, groups = var_183, pad = pretrained_out_1_pad_0, pad_type = pretrained_out_1_pad_type_0, strides = var_221, weight = layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("pretrained_out_1_cast_fp16")]; + tensor var_227 = const()[name = tensor("op_227"), val = tensor([1, 1])]; + tensor var_229 = const()[name = tensor("op_229"), val = tensor([1, 1])]; + tensor input_1_pad_type_0 = const()[name = tensor("input_1_pad_type_0"), val = tensor("custom")]; + tensor input_1_pad_0 = const()[name = tensor("input_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134753984)))]; + tensor input_1_cast_fp16 = conv(dilations = var_229, groups = var_183, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = var_227, weight = layers_0_self_attn_q_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_233 = const()[name = tensor("op_233"), val = tensor([1, 1])]; + tensor var_235 = const()[name = tensor("op_235"), val = tensor([1, 1])]; + tensor lora_out_1_pad_type_0 = const()[name = tensor("lora_out_1_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1_pad_0 = const()[name = tensor("lora_out_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_3_weight_0_to_fp16 = const()[name = tensor("lora_out_3_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134795008)))]; + tensor lora_out_3_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_235, groups = var_183, pad = lora_out_1_pad_0, pad_type = lora_out_1_pad_type_0, strides = var_233, weight = lora_out_3_weight_0_to_fp16, x = input_1_cast_fp16)[name = tensor("lora_out_3_cast_fp16")]; + tensor query_1_cast_fp16 = add(x = pretrained_out_1_cast_fp16, y = lora_out_3_cast_fp16)[name = tensor("query_1_cast_fp16")]; + tensor var_245 = const()[name = tensor("op_245"), val = tensor([1, 1])]; + tensor var_247 = const()[name = tensor("op_247"), val = tensor([1, 1])]; + tensor pretrained_out_3_pad_type_0 = const()[name = tensor("pretrained_out_3_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_3_pad_0 = const()[name = tensor("pretrained_out_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134836032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135655296))), name = tensor("layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_3_cast_fp16 = conv(dilations = var_247, groups = var_183, pad = pretrained_out_3_pad_0, pad_type = pretrained_out_3_pad_type_0, strides = var_245, weight = layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("pretrained_out_3_cast_fp16")]; + tensor var_251 = const()[name = tensor("op_251"), val = tensor([1, 1])]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 1])]; + tensor input_3_pad_type_0 = const()[name = tensor("input_3_pad_type_0"), val = tensor("custom")]; + tensor input_3_pad_0 = const()[name = tensor("input_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135655424)))]; + tensor input_3_cast_fp16 = conv(dilations = var_253, groups = var_183, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = var_251, weight = layers_0_self_attn_k_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor var_257 = const()[name = tensor("op_257"), val = tensor([1, 1])]; + tensor var_259 = const()[name = tensor("op_259"), val = tensor([1, 1])]; + tensor lora_out_5_pad_type_0 = const()[name = tensor("lora_out_5_pad_type_0"), val = tensor("custom")]; + tensor lora_out_5_pad_0 = const()[name = tensor("lora_out_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_7_weight_0_to_fp16 = const()[name = tensor("lora_out_7_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135696448)))]; + tensor lora_out_7_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_259, groups = var_183, pad = lora_out_5_pad_0, pad_type = lora_out_5_pad_type_0, strides = var_257, weight = lora_out_7_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("lora_out_7_cast_fp16")]; + tensor current_key_1_cast_fp16 = add(x = pretrained_out_3_cast_fp16, y = lora_out_7_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; + tensor var_270 = const()[name = tensor("op_270"), val = tensor([1, 1])]; + tensor var_272 = const()[name = tensor("op_272"), val = tensor([1, 1])]; + tensor pretrained_out_5_pad_type_0 = const()[name = tensor("pretrained_out_5_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_5_pad_0 = const()[name = tensor("pretrained_out_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135737472))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136556736))), name = tensor("layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136556864)))]; + tensor pretrained_out_5_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_272, groups = var_183, pad = pretrained_out_5_pad_0, pad_type = pretrained_out_5_pad_type_0, strides = var_270, weight = layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("pretrained_out_5_cast_fp16")]; + tensor var_276 = const()[name = tensor("op_276"), val = tensor([1, 1])]; + tensor var_278 = const()[name = tensor("op_278"), val = tensor([1, 1])]; + tensor input_5_pad_type_0 = const()[name = tensor("input_5_pad_type_0"), val = tensor("custom")]; + tensor input_5_pad_0 = const()[name = tensor("input_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136559488)))]; + tensor input_5_cast_fp16 = conv(dilations = var_278, groups = var_183, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = var_276, weight = layers_0_self_attn_v_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor var_282 = const()[name = tensor("op_282"), val = tensor([1, 1])]; + tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 1])]; + tensor lora_out_9_pad_type_0 = const()[name = tensor("lora_out_9_pad_type_0"), val = tensor("custom")]; + tensor lora_out_9_pad_0 = const()[name = tensor("lora_out_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_11_weight_0_to_fp16 = const()[name = tensor("lora_out_11_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136600512)))]; + tensor lora_out_11_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_284, groups = var_183, pad = lora_out_9_pad_0, pad_type = lora_out_9_pad_type_0, strides = var_282, weight = lora_out_11_weight_0_to_fp16, x = input_5_cast_fp16)[name = tensor("lora_out_11_cast_fp16")]; + tensor current_value_1_cast_fp16 = add(x = pretrained_out_5_cast_fp16, y = lora_out_11_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; + tensor var_291_axes_0 = const()[name = tensor("op_291_axes_0"), val = tensor([1])]; + tensor var_291_cast_fp16 = expand_dims(axes = var_291_axes_0, x = kv_cache_update_mask)[name = tensor("op_291_cast_fp16")]; + tensor var_292_axes_0 = const()[name = tensor("op_292_axes_0"), val = tensor([2])]; + tensor var_292_cast_fp16 = expand_dims(axes = var_292_axes_0, x = var_291_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor var_294_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_294_cast_fp16")]; + tensor var_177_to_fp16 = const()[name = tensor("op_177_to_fp16"), val = tensor(0x1p+0)]; + tensor var_295_cast_fp16 = sub(x = var_177_to_fp16, y = var_292_cast_fp16)[name = tensor("op_295_cast_fp16")]; + tensor var_296_cast_fp16 = mul(x = var_103_cast_fp16_0, y = var_295_cast_fp16)[name = tensor("op_296_cast_fp16")]; + tensor key_1_cast_fp16 = add(x = var_294_cast_fp16, y = var_296_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_298_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor var_300_cast_fp16 = mul(x = var_138_cast_fp16_0, y = var_295_cast_fp16)[name = tensor("op_300_cast_fp16")]; + tensor value_1_cast_fp16 = add(x = var_298_cast_fp16, y = var_300_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_303 = const()[name = tensor("op_303"), val = tensor([1, 20, 64, -1])]; + tensor var_304_cast_fp16 = reshape(shape = var_303, x = query_1_cast_fp16)[name = tensor("op_304_cast_fp16")]; + tensor var_305_to_fp16 = const()[name = tensor("op_305_to_fp16"), val = tensor(0x1p-3)]; + tensor var_306_cast_fp16 = mul(x = var_304_cast_fp16, y = var_305_to_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_307 = const()[name = tensor("op_307"), val = tensor([1, 20, 64, -1])]; + tensor var_308_cast_fp16 = reshape(shape = var_307, x = key_1_cast_fp16)[name = tensor("op_308_cast_fp16")]; + tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; + tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; + tensor mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_306_cast_fp16, y = var_308_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor var_312_axes_0 = const()[name = tensor("op_312_axes_0"), val = tensor([1])]; + tensor var_312_cast_fp16 = expand_dims(axes = var_312_axes_0, x = decoder_key_padding_mask)[name = tensor("op_312_cast_fp16")]; + tensor var_313_axes_0 = const()[name = tensor("op_313_axes_0"), val = tensor([2])]; + tensor var_313_cast_fp16 = expand_dims(axes = var_313_axes_0, x = var_312_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_316_cast_fp16 = softmax(axis = var_176, x = mh_w_3_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor var_317 = const()[name = tensor("op_317"), val = tensor([1, 20, 64, -1])]; + tensor var_318_cast_fp16 = reshape(shape = var_317, x = value_1_cast_fp16)[name = tensor("op_318_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(true)]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_318_cast_fp16, y = var_316_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_321 = const()[name = tensor("op_321"), val = tensor([1, 1280, 1, -1])]; + tensor input_7_cast_fp16 = reshape(shape = var_321, x = attn_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor var_328 = const()[name = tensor("op_328"), val = tensor([1, 1])]; + tensor var_330 = const()[name = tensor("op_330"), val = tensor([1, 1])]; + tensor pretrained_out_7_pad_type_0 = const()[name = tensor("pretrained_out_7_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_7_pad_0 = const()[name = tensor("pretrained_out_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136641536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137460800))), name = tensor("layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137460928)))]; + tensor pretrained_out_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_330, groups = var_183, pad = pretrained_out_7_pad_0, pad_type = pretrained_out_7_pad_type_0, strides = var_328, weight = layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = tensor("pretrained_out_7_cast_fp16")]; + tensor var_334 = const()[name = tensor("op_334"), val = tensor([1, 1])]; + tensor var_336 = const()[name = tensor("op_336"), val = tensor([1, 1])]; + tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("custom")]; + tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137463552)))]; + tensor input_9_cast_fp16 = conv(dilations = var_336, groups = var_183, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = var_334, weight = layers_0_self_attn_o_proj_loraA_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor([1, 1])]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 1])]; + tensor lora_out_13_pad_type_0 = const()[name = tensor("lora_out_13_pad_type_0"), val = tensor("custom")]; + tensor lora_out_13_pad_0 = const()[name = tensor("lora_out_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_15_weight_0_to_fp16 = const()[name = tensor("lora_out_15_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137504576)))]; + tensor lora_out_15_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_342, groups = var_183, pad = lora_out_13_pad_0, pad_type = lora_out_13_pad_type_0, strides = var_340, weight = lora_out_15_weight_0_to_fp16, x = input_9_cast_fp16)[name = tensor("lora_out_15_cast_fp16")]; + tensor obj_7_cast_fp16 = add(x = pretrained_out_7_cast_fp16, y = lora_out_15_cast_fp16)[name = tensor("obj_7_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1])]; + tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_355, keep_dims = var_184, x = inputs_3_cast_fp16)[name = tensor("channels_mean_3_cast_fp16")]; + tensor zero_mean_3_cast_fp16 = sub(x = inputs_3_cast_fp16, y = channels_mean_3_cast_fp16)[name = tensor("zero_mean_3_cast_fp16")]; + tensor zero_mean_sq_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = zero_mean_3_cast_fp16)[name = tensor("zero_mean_sq_3_cast_fp16")]; + tensor var_359 = const()[name = tensor("op_359"), val = tensor([1])]; + tensor var_360_cast_fp16 = reduce_mean(axes = var_359, keep_dims = var_184, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor var_361_to_fp16 = const()[name = tensor("op_361_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_362_cast_fp16 = add(x = var_360_cast_fp16, y = var_361_to_fp16)[name = tensor("op_362_cast_fp16")]; + tensor denom_3_epsilon_0 = const()[name = tensor("denom_3_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_3_cast_fp16 = rsqrt(epsilon = denom_3_epsilon_0, x = var_362_cast_fp16)[name = tensor("denom_3_cast_fp16")]; + tensor out_3_cast_fp16 = mul(x = zero_mean_3_cast_fp16, y = denom_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; + tensor obj_9_gamma_0_to_fp16 = const()[name = tensor("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137545600)))]; + tensor obj_9_beta_0_to_fp16 = const()[name = tensor("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137548224)))]; + tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor("obj_9_cast_fp16")]; + tensor var_380 = const()[name = tensor("op_380"), val = tensor([1, 1])]; + tensor var_382 = const()[name = tensor("op_382"), val = tensor([1, 1])]; + tensor pretrained_out_9_pad_type_0 = const()[name = tensor("pretrained_out_9_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_9_pad_0 = const()[name = tensor("pretrained_out_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137550848))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138370112))), name = tensor("layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138370240)))]; + tensor pretrained_out_9_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_382, groups = var_183, pad = pretrained_out_9_pad_0, pad_type = pretrained_out_9_pad_type_0, strides = var_380, weight = layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor("pretrained_out_9_cast_fp16")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 1])]; + tensor var_388 = const()[name = tensor("op_388"), val = tensor([1, 1])]; + tensor input_11_pad_type_0 = const()[name = tensor("input_11_pad_type_0"), val = tensor("custom")]; + tensor input_11_pad_0 = const()[name = tensor("input_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138372864)))]; + tensor input_11_cast_fp16 = conv(dilations = var_388, groups = var_183, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = var_386, weight = layers_0_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor var_392 = const()[name = tensor("op_392"), val = tensor([1, 1])]; + tensor var_394 = const()[name = tensor("op_394"), val = tensor([1, 1])]; + tensor lora_out_17_pad_type_0 = const()[name = tensor("lora_out_17_pad_type_0"), val = tensor("custom")]; + tensor lora_out_17_pad_0 = const()[name = tensor("lora_out_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_19_weight_0_to_fp16 = const()[name = tensor("lora_out_19_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138413888)))]; + tensor lora_out_19_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_394, groups = var_183, pad = lora_out_17_pad_0, pad_type = lora_out_17_pad_type_0, strides = var_392, weight = lora_out_19_weight_0_to_fp16, x = input_11_cast_fp16)[name = tensor("lora_out_19_cast_fp16")]; + tensor query_3_cast_fp16 = add(x = pretrained_out_9_cast_fp16, y = lora_out_19_cast_fp16)[name = tensor("query_3_cast_fp16")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([1, 1])]; + tensor var_406 = const()[name = tensor("op_406"), val = tensor([1, 1])]; + tensor pretrained_out_11_pad_type_0 = const()[name = tensor("pretrained_out_11_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_11_pad_0 = const()[name = tensor("pretrained_out_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138454912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139274176))), name = tensor("layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_11_cast_fp16 = conv(dilations = var_406, groups = var_183, pad = pretrained_out_11_pad_0, pad_type = pretrained_out_11_pad_type_0, strides = var_404, weight = layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_11_cast_fp16")]; + tensor var_410 = const()[name = tensor("op_410"), val = tensor([1, 1])]; + tensor var_412 = const()[name = tensor("op_412"), val = tensor([1, 1])]; + tensor input_13_pad_type_0 = const()[name = tensor("input_13_pad_type_0"), val = tensor("custom")]; + tensor input_13_pad_0 = const()[name = tensor("input_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139274304)))]; + tensor input_13_cast_fp16 = conv(dilations = var_412, groups = var_183, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = var_410, weight = layers_0_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_13_cast_fp16")]; + tensor var_416 = const()[name = tensor("op_416"), val = tensor([1, 1])]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1])]; + tensor lora_out_21_pad_type_0 = const()[name = tensor("lora_out_21_pad_type_0"), val = tensor("custom")]; + tensor lora_out_21_pad_0 = const()[name = tensor("lora_out_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_23_weight_0_to_fp16 = const()[name = tensor("lora_out_23_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139315328)))]; + tensor lora_out_23_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_418, groups = var_183, pad = lora_out_21_pad_0, pad_type = lora_out_21_pad_type_0, strides = var_416, weight = lora_out_23_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("lora_out_23_cast_fp16")]; + tensor key_3_cast_fp16 = add(x = pretrained_out_11_cast_fp16, y = lora_out_23_cast_fp16)[name = tensor("key_3_cast_fp16")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 1])]; + tensor var_431 = const()[name = tensor("op_431"), val = tensor([1, 1])]; + tensor pretrained_out_13_pad_type_0 = const()[name = tensor("pretrained_out_13_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_13_pad_0 = const()[name = tensor("pretrained_out_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139356352))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140175616))), name = tensor("layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140175744)))]; + tensor pretrained_out_13_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_431, groups = var_183, pad = pretrained_out_13_pad_0, pad_type = pretrained_out_13_pad_type_0, strides = var_429, weight = layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_13_cast_fp16")]; + tensor var_435 = const()[name = tensor("op_435"), val = tensor([1, 1])]; + tensor var_437 = const()[name = tensor("op_437"), val = tensor([1, 1])]; + tensor input_15_pad_type_0 = const()[name = tensor("input_15_pad_type_0"), val = tensor("custom")]; + tensor input_15_pad_0 = const()[name = tensor("input_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140178368)))]; + tensor input_15_cast_fp16 = conv(dilations = var_437, groups = var_183, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = var_435, weight = layers_0_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_15_cast_fp16")]; + tensor var_441 = const()[name = tensor("op_441"), val = tensor([1, 1])]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1])]; + tensor lora_out_25_pad_type_0 = const()[name = tensor("lora_out_25_pad_type_0"), val = tensor("custom")]; + tensor lora_out_25_pad_0 = const()[name = tensor("lora_out_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_27_weight_0_to_fp16 = const()[name = tensor("lora_out_27_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140219392)))]; + tensor lora_out_27_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_443, groups = var_183, pad = lora_out_25_pad_0, pad_type = lora_out_25_pad_type_0, strides = var_441, weight = lora_out_27_weight_0_to_fp16, x = input_15_cast_fp16)[name = tensor("lora_out_27_cast_fp16")]; + tensor value_3_cast_fp16 = add(x = pretrained_out_13_cast_fp16, y = lora_out_27_cast_fp16)[name = tensor("value_3_cast_fp16")]; + tensor var_450 = const()[name = tensor("op_450"), val = tensor([1, 20, 64, -1])]; + tensor var_451_cast_fp16 = reshape(shape = var_450, x = query_3_cast_fp16)[name = tensor("op_451_cast_fp16")]; + tensor var_452_to_fp16 = const()[name = tensor("op_452_to_fp16"), val = tensor(0x1p-3)]; + tensor var_453_cast_fp16 = mul(x = var_451_cast_fp16, y = var_452_to_fp16)[name = tensor("op_453_cast_fp16")]; + tensor var_454 = const()[name = tensor("op_454"), val = tensor([1, 20, 64, -1])]; + tensor var_455_cast_fp16 = reshape(shape = var_454, x = key_3_cast_fp16)[name = tensor("op_455_cast_fp16")]; + tensor mh_w_5_transpose_x_0 = const()[name = tensor("mh_w_5_transpose_x_0"), val = tensor(true)]; + tensor mh_w_5_transpose_y_0 = const()[name = tensor("mh_w_5_transpose_y_0"), val = tensor(false)]; + tensor mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_453_cast_fp16, y = var_455_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor obj_13_cast_fp16 = softmax(axis = var_176, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor var_459 = const()[name = tensor("op_459"), val = tensor([1, 20, 64, -1])]; + tensor var_460_cast_fp16 = reshape(shape = var_459, x = value_3_cast_fp16)[name = tensor("op_460_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(true)]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_460_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_463 = const()[name = tensor("op_463"), val = tensor([1, 1280, 1, -1])]; + tensor input_17_cast_fp16 = reshape(shape = var_463, x = attn_3_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor var_470 = const()[name = tensor("op_470"), val = tensor([1, 1])]; + tensor var_472 = const()[name = tensor("op_472"), val = tensor([1, 1])]; + tensor pretrained_out_15_pad_type_0 = const()[name = tensor("pretrained_out_15_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_15_pad_0 = const()[name = tensor("pretrained_out_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140260416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141079680))), name = tensor("layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141079808)))]; + tensor pretrained_out_15_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_472, groups = var_183, pad = pretrained_out_15_pad_0, pad_type = pretrained_out_15_pad_type_0, strides = var_470, weight = layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = tensor("pretrained_out_15_cast_fp16")]; + tensor var_476 = const()[name = tensor("op_476"), val = tensor([1, 1])]; + tensor var_478 = const()[name = tensor("op_478"), val = tensor([1, 1])]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("custom")]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141082432)))]; + tensor input_19_cast_fp16 = conv(dilations = var_478, groups = var_183, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = var_476, weight = layers_0_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor var_482 = const()[name = tensor("op_482"), val = tensor([1, 1])]; + tensor var_484 = const()[name = tensor("op_484"), val = tensor([1, 1])]; + tensor lora_out_29_pad_type_0 = const()[name = tensor("lora_out_29_pad_type_0"), val = tensor("custom")]; + tensor lora_out_29_pad_0 = const()[name = tensor("lora_out_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_31_weight_0_to_fp16 = const()[name = tensor("lora_out_31_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141123456)))]; + tensor lora_out_31_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_484, groups = var_183, pad = lora_out_29_pad_0, pad_type = lora_out_29_pad_type_0, strides = var_482, weight = lora_out_31_weight_0_to_fp16, x = input_19_cast_fp16)[name = tensor("lora_out_31_cast_fp16")]; + tensor obj_11_cast_fp16 = add(x = pretrained_out_15_cast_fp16, y = lora_out_31_cast_fp16)[name = tensor("obj_11_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_493 = const()[name = tensor("op_493"), val = tensor([1])]; + tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_493, keep_dims = var_184, x = inputs_5_cast_fp16)[name = tensor("channels_mean_5_cast_fp16")]; + tensor zero_mean_5_cast_fp16 = sub(x = inputs_5_cast_fp16, y = channels_mean_5_cast_fp16)[name = tensor("zero_mean_5_cast_fp16")]; + tensor zero_mean_sq_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = zero_mean_5_cast_fp16)[name = tensor("zero_mean_sq_5_cast_fp16")]; + tensor var_497 = const()[name = tensor("op_497"), val = tensor([1])]; + tensor var_498_cast_fp16 = reduce_mean(axes = var_497, keep_dims = var_184, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_498_cast_fp16")]; + tensor var_499_to_fp16 = const()[name = tensor("op_499_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_500_cast_fp16 = add(x = var_498_cast_fp16, y = var_499_to_fp16)[name = tensor("op_500_cast_fp16")]; + tensor denom_5_epsilon_0 = const()[name = tensor("denom_5_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_5_cast_fp16 = rsqrt(epsilon = denom_5_epsilon_0, x = var_500_cast_fp16)[name = tensor("denom_5_cast_fp16")]; + tensor out_5_cast_fp16 = mul(x = zero_mean_5_cast_fp16, y = denom_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; + tensor input_21_gamma_0_to_fp16 = const()[name = tensor("input_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141164480)))]; + tensor input_21_beta_0_to_fp16 = const()[name = tensor("input_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141167104)))]; + tensor input_21_epsilon_0_to_fp16 = const()[name = tensor("input_21_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = batch_norm(beta = input_21_beta_0_to_fp16, epsilon = input_21_epsilon_0_to_fp16, gamma = input_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_514 = const()[name = tensor("op_514"), val = tensor([1, 1])]; + tensor var_516 = const()[name = tensor("op_516"), val = tensor([1, 1])]; + tensor pretrained_out_17_pad_type_0 = const()[name = tensor("pretrained_out_17_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_17_pad_0 = const()[name = tensor("pretrained_out_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141169728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144446592))), name = tensor("layers_0_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_0_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144446720)))]; + tensor pretrained_out_17_cast_fp16 = conv(bias = layers_0_fc1_pretrained_bias_to_fp16, dilations = var_516, groups = var_183, pad = pretrained_out_17_pad_0, pad_type = pretrained_out_17_pad_type_0, strides = var_514, weight = layers_0_fc1_pretrained_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = tensor("pretrained_out_17_cast_fp16")]; + tensor var_520 = const()[name = tensor("op_520"), val = tensor([1, 1])]; + tensor var_522 = const()[name = tensor("op_522"), val = tensor([1, 1])]; + tensor input_23_pad_type_0 = const()[name = tensor("input_23_pad_type_0"), val = tensor("custom")]; + tensor input_23_pad_0 = const()[name = tensor("input_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_0_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144457024)))]; + tensor input_23_cast_fp16 = conv(dilations = var_522, groups = var_183, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = var_520, weight = layers_0_fc1_loraA_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor var_526 = const()[name = tensor("op_526"), val = tensor([1, 1])]; + tensor var_528 = const()[name = tensor("op_528"), val = tensor([1, 1])]; + tensor lora_out_33_pad_type_0 = const()[name = tensor("lora_out_33_pad_type_0"), val = tensor("custom")]; + tensor lora_out_33_pad_0 = const()[name = tensor("lora_out_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_35_weight_0_to_fp16 = const()[name = tensor("lora_out_35_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144498048)))]; + tensor lora_out_35_bias_0_to_fp16 = const()[name = tensor("lora_out_35_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144661952)))]; + tensor lora_out_35_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_528, groups = var_183, pad = lora_out_33_pad_0, pad_type = lora_out_33_pad_type_0, strides = var_526, weight = lora_out_35_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("lora_out_35_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = pretrained_out_17_cast_fp16, y = lora_out_35_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_mode_0 = const()[name = tensor("input_27_mode_0"), val = tensor("EXACT")]; + tensor input_27_cast_fp16 = gelu(mode = input_27_mode_0, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor var_540 = const()[name = tensor("op_540"), val = tensor([1, 1])]; + tensor var_542 = const()[name = tensor("op_542"), val = tensor([1, 1])]; + tensor pretrained_out_19_pad_type_0 = const()[name = tensor("pretrained_out_19_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_19_pad_0 = const()[name = tensor("pretrained_out_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144672256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147949120))), name = tensor("layers_0_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_0_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147949248)))]; + tensor pretrained_out_19_cast_fp16 = conv(bias = layers_0_fc2_pretrained_bias_to_fp16, dilations = var_542, groups = var_183, pad = pretrained_out_19_pad_0, pad_type = pretrained_out_19_pad_type_0, strides = var_540, weight = layers_0_fc2_pretrained_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = tensor("pretrained_out_19_cast_fp16")]; + tensor var_546 = const()[name = tensor("op_546"), val = tensor([1, 1])]; + tensor var_548 = const()[name = tensor("op_548"), val = tensor([1, 1])]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("custom")]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_0_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_0_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147951872)))]; + tensor input_29_cast_fp16 = conv(dilations = var_548, groups = var_183, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = var_546, weight = layers_0_fc2_loraA_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor var_552 = const()[name = tensor("op_552"), val = tensor([1, 1])]; + tensor var_554 = const()[name = tensor("op_554"), val = tensor([1, 1])]; + tensor lora_out_37_pad_type_0 = const()[name = tensor("lora_out_37_pad_type_0"), val = tensor("custom")]; + tensor lora_out_37_pad_0 = const()[name = tensor("lora_out_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_39_weight_0_to_fp16 = const()[name = tensor("lora_out_39_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148115776)))]; + tensor lora_out_39_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_554, groups = var_183, pad = lora_out_37_pad_0, pad_type = lora_out_37_pad_type_0, strides = var_552, weight = lora_out_39_weight_0_to_fp16, x = input_29_cast_fp16)[name = tensor("lora_out_39_cast_fp16")]; + tensor hidden_states_3_cast_fp16 = add(x = pretrained_out_19_cast_fp16, y = lora_out_39_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor var_570 = const()[name = tensor("op_570"), val = tensor(3)]; + tensor var_577 = const()[name = tensor("op_577"), val = tensor(1)]; + tensor var_578 = const()[name = tensor("op_578"), val = tensor(true)]; + tensor var_590 = const()[name = tensor("op_590"), val = tensor([1])]; + tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_590, keep_dims = var_578, x = inputs_7_cast_fp16)[name = tensor("channels_mean_7_cast_fp16")]; + tensor zero_mean_7_cast_fp16 = sub(x = inputs_7_cast_fp16, y = channels_mean_7_cast_fp16)[name = tensor("zero_mean_7_cast_fp16")]; + tensor zero_mean_sq_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = zero_mean_7_cast_fp16)[name = tensor("zero_mean_sq_7_cast_fp16")]; + tensor var_594 = const()[name = tensor("op_594"), val = tensor([1])]; + tensor var_595_cast_fp16 = reduce_mean(axes = var_594, keep_dims = var_578, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_595_cast_fp16")]; + tensor var_596_to_fp16 = const()[name = tensor("op_596_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_597_cast_fp16 = add(x = var_595_cast_fp16, y = var_596_to_fp16)[name = tensor("op_597_cast_fp16")]; + tensor denom_7_epsilon_0 = const()[name = tensor("denom_7_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_7_cast_fp16 = rsqrt(epsilon = denom_7_epsilon_0, x = var_597_cast_fp16)[name = tensor("denom_7_cast_fp16")]; + tensor out_7_cast_fp16 = mul(x = zero_mean_7_cast_fp16, y = denom_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor obj_15_gamma_0_to_fp16 = const()[name = tensor("obj_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148156800)))]; + tensor obj_15_beta_0_to_fp16 = const()[name = tensor("obj_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148159424)))]; + tensor obj_15_epsilon_0_to_fp16 = const()[name = tensor("obj_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("obj_15_cast_fp16")]; + tensor var_615 = const()[name = tensor("op_615"), val = tensor([1, 1])]; + tensor var_617 = const()[name = tensor("op_617"), val = tensor([1, 1])]; + tensor pretrained_out_21_pad_type_0 = const()[name = tensor("pretrained_out_21_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_21_pad_0 = const()[name = tensor("pretrained_out_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148162048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148981312))), name = tensor("layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148981440)))]; + tensor pretrained_out_21_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_617, groups = var_577, pad = pretrained_out_21_pad_0, pad_type = pretrained_out_21_pad_type_0, strides = var_615, weight = layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor("pretrained_out_21_cast_fp16")]; + tensor var_621 = const()[name = tensor("op_621"), val = tensor([1, 1])]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1])]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("custom")]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148984064)))]; + tensor input_31_cast_fp16 = conv(dilations = var_623, groups = var_577, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = var_621, weight = layers_1_self_attn_q_proj_loraA_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1])]; + tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 1])]; + tensor lora_out_41_pad_type_0 = const()[name = tensor("lora_out_41_pad_type_0"), val = tensor("custom")]; + tensor lora_out_41_pad_0 = const()[name = tensor("lora_out_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_43_weight_0_to_fp16 = const()[name = tensor("lora_out_43_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149025088)))]; + tensor lora_out_43_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_629, groups = var_577, pad = lora_out_41_pad_0, pad_type = lora_out_41_pad_type_0, strides = var_627, weight = lora_out_43_weight_0_to_fp16, x = input_31_cast_fp16)[name = tensor("lora_out_43_cast_fp16")]; + tensor query_5_cast_fp16 = add(x = pretrained_out_21_cast_fp16, y = lora_out_43_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor var_639 = const()[name = tensor("op_639"), val = tensor([1, 1])]; + tensor var_641 = const()[name = tensor("op_641"), val = tensor([1, 1])]; + tensor pretrained_out_23_pad_type_0 = const()[name = tensor("pretrained_out_23_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_23_pad_0 = const()[name = tensor("pretrained_out_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149066112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149885376))), name = tensor("layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_23_cast_fp16 = conv(dilations = var_641, groups = var_577, pad = pretrained_out_23_pad_0, pad_type = pretrained_out_23_pad_type_0, strides = var_639, weight = layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor("pretrained_out_23_cast_fp16")]; + tensor var_645 = const()[name = tensor("op_645"), val = tensor([1, 1])]; + tensor var_647 = const()[name = tensor("op_647"), val = tensor([1, 1])]; + tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("custom")]; + tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149885504)))]; + tensor input_33_cast_fp16 = conv(dilations = var_647, groups = var_577, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = var_645, weight = layers_1_self_attn_k_proj_loraA_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor var_651 = const()[name = tensor("op_651"), val = tensor([1, 1])]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1])]; + tensor lora_out_45_pad_type_0 = const()[name = tensor("lora_out_45_pad_type_0"), val = tensor("custom")]; + tensor lora_out_45_pad_0 = const()[name = tensor("lora_out_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_47_weight_0_to_fp16 = const()[name = tensor("lora_out_47_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149926528)))]; + tensor lora_out_47_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_653, groups = var_577, pad = lora_out_45_pad_0, pad_type = lora_out_45_pad_type_0, strides = var_651, weight = lora_out_47_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("lora_out_47_cast_fp16")]; + tensor current_key_3_cast_fp16 = add(x = pretrained_out_23_cast_fp16, y = lora_out_47_cast_fp16)[name = tensor("current_key_3_cast_fp16")]; + tensor var_664 = const()[name = tensor("op_664"), val = tensor([1, 1])]; + tensor var_666 = const()[name = tensor("op_666"), val = tensor([1, 1])]; + tensor pretrained_out_25_pad_type_0 = const()[name = tensor("pretrained_out_25_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_25_pad_0 = const()[name = tensor("pretrained_out_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149967552))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150786816))), name = tensor("layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150786944)))]; + tensor pretrained_out_25_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_666, groups = var_577, pad = pretrained_out_25_pad_0, pad_type = pretrained_out_25_pad_type_0, strides = var_664, weight = layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor("pretrained_out_25_cast_fp16")]; + tensor var_670 = const()[name = tensor("op_670"), val = tensor([1, 1])]; + tensor var_672 = const()[name = tensor("op_672"), val = tensor([1, 1])]; + tensor input_35_pad_type_0 = const()[name = tensor("input_35_pad_type_0"), val = tensor("custom")]; + tensor input_35_pad_0 = const()[name = tensor("input_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150789568)))]; + tensor input_35_cast_fp16 = conv(dilations = var_672, groups = var_577, pad = input_35_pad_0, pad_type = input_35_pad_type_0, strides = var_670, weight = layers_1_self_attn_v_proj_loraA_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor([1, 1])]; + tensor var_678 = const()[name = tensor("op_678"), val = tensor([1, 1])]; + tensor lora_out_49_pad_type_0 = const()[name = tensor("lora_out_49_pad_type_0"), val = tensor("custom")]; + tensor lora_out_49_pad_0 = const()[name = tensor("lora_out_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_51_weight_0_to_fp16 = const()[name = tensor("lora_out_51_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150830592)))]; + tensor lora_out_51_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_678, groups = var_577, pad = lora_out_49_pad_0, pad_type = lora_out_49_pad_type_0, strides = var_676, weight = lora_out_51_weight_0_to_fp16, x = input_35_cast_fp16)[name = tensor("lora_out_51_cast_fp16")]; + tensor current_value_3_cast_fp16 = add(x = pretrained_out_25_cast_fp16, y = lora_out_51_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; + tensor var_688_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_688_cast_fp16")]; + tensor var_690_cast_fp16 = mul(x = var_103_cast_fp16_1, y = var_295_cast_fp16)[name = tensor("op_690_cast_fp16")]; + tensor key_5_cast_fp16 = add(x = var_688_cast_fp16, y = var_690_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_692_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_692_cast_fp16")]; + tensor var_694_cast_fp16 = mul(x = var_138_cast_fp16_1, y = var_295_cast_fp16)[name = tensor("op_694_cast_fp16")]; + tensor value_5_cast_fp16 = add(x = var_692_cast_fp16, y = var_694_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_697 = const()[name = tensor("op_697"), val = tensor([1, 20, 64, -1])]; + tensor var_698_cast_fp16 = reshape(shape = var_697, x = query_5_cast_fp16)[name = tensor("op_698_cast_fp16")]; + tensor var_699_to_fp16 = const()[name = tensor("op_699_to_fp16"), val = tensor(0x1p-3)]; + tensor var_700_cast_fp16 = mul(x = var_698_cast_fp16, y = var_699_to_fp16)[name = tensor("op_700_cast_fp16")]; + tensor var_701 = const()[name = tensor("op_701"), val = tensor([1, 20, 64, -1])]; + tensor var_702_cast_fp16 = reshape(shape = var_701, x = key_5_cast_fp16)[name = tensor("op_702_cast_fp16")]; + tensor mh_w_7_transpose_x_0 = const()[name = tensor("mh_w_7_transpose_x_0"), val = tensor(true)]; + tensor mh_w_7_transpose_y_0 = const()[name = tensor("mh_w_7_transpose_y_0"), val = tensor(false)]; + tensor mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_700_cast_fp16, y = var_702_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_710_cast_fp16 = softmax(axis = var_570, x = mh_w_9_cast_fp16)[name = tensor("op_710_cast_fp16")]; + tensor var_711 = const()[name = tensor("op_711"), val = tensor([1, 20, 64, -1])]; + tensor var_712_cast_fp16 = reshape(shape = var_711, x = value_5_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(true)]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_712_cast_fp16, y = var_710_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_715 = const()[name = tensor("op_715"), val = tensor([1, 1280, 1, -1])]; + tensor input_37_cast_fp16 = reshape(shape = var_715, x = attn_5_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor var_722 = const()[name = tensor("op_722"), val = tensor([1, 1])]; + tensor var_724 = const()[name = tensor("op_724"), val = tensor([1, 1])]; + tensor pretrained_out_27_pad_type_0 = const()[name = tensor("pretrained_out_27_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_27_pad_0 = const()[name = tensor("pretrained_out_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150871616))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151690880))), name = tensor("layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151691008)))]; + tensor pretrained_out_27_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_724, groups = var_577, pad = pretrained_out_27_pad_0, pad_type = pretrained_out_27_pad_type_0, strides = var_722, weight = layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = tensor("pretrained_out_27_cast_fp16")]; + tensor var_728 = const()[name = tensor("op_728"), val = tensor([1, 1])]; + tensor var_730 = const()[name = tensor("op_730"), val = tensor([1, 1])]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("custom")]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151693632)))]; + tensor input_39_cast_fp16 = conv(dilations = var_730, groups = var_577, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = var_728, weight = layers_1_self_attn_o_proj_loraA_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor var_734 = const()[name = tensor("op_734"), val = tensor([1, 1])]; + tensor var_736 = const()[name = tensor("op_736"), val = tensor([1, 1])]; + tensor lora_out_53_pad_type_0 = const()[name = tensor("lora_out_53_pad_type_0"), val = tensor("custom")]; + tensor lora_out_53_pad_0 = const()[name = tensor("lora_out_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_55_weight_0_to_fp16 = const()[name = tensor("lora_out_55_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151734656)))]; + tensor lora_out_55_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_736, groups = var_577, pad = lora_out_53_pad_0, pad_type = lora_out_53_pad_type_0, strides = var_734, weight = lora_out_55_weight_0_to_fp16, x = input_39_cast_fp16)[name = tensor("lora_out_55_cast_fp16")]; + tensor obj_21_cast_fp16 = add(x = pretrained_out_27_cast_fp16, y = lora_out_55_cast_fp16)[name = tensor("obj_21_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_749 = const()[name = tensor("op_749"), val = tensor([1])]; + tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_749, keep_dims = var_578, x = inputs_9_cast_fp16)[name = tensor("channels_mean_9_cast_fp16")]; + tensor zero_mean_9_cast_fp16 = sub(x = inputs_9_cast_fp16, y = channels_mean_9_cast_fp16)[name = tensor("zero_mean_9_cast_fp16")]; + tensor zero_mean_sq_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = zero_mean_9_cast_fp16)[name = tensor("zero_mean_sq_9_cast_fp16")]; + tensor var_753 = const()[name = tensor("op_753"), val = tensor([1])]; + tensor var_754_cast_fp16 = reduce_mean(axes = var_753, keep_dims = var_578, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_754_cast_fp16")]; + tensor var_755_to_fp16 = const()[name = tensor("op_755_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_756_cast_fp16 = add(x = var_754_cast_fp16, y = var_755_to_fp16)[name = tensor("op_756_cast_fp16")]; + tensor denom_9_epsilon_0 = const()[name = tensor("denom_9_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_9_cast_fp16 = rsqrt(epsilon = denom_9_epsilon_0, x = var_756_cast_fp16)[name = tensor("denom_9_cast_fp16")]; + tensor out_9_cast_fp16 = mul(x = zero_mean_9_cast_fp16, y = denom_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; + tensor obj_23_gamma_0_to_fp16 = const()[name = tensor("obj_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151775680)))]; + tensor obj_23_beta_0_to_fp16 = const()[name = tensor("obj_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151778304)))]; + tensor obj_23_epsilon_0_to_fp16 = const()[name = tensor("obj_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_23_cast_fp16")]; + tensor var_774 = const()[name = tensor("op_774"), val = tensor([1, 1])]; + tensor var_776 = const()[name = tensor("op_776"), val = tensor([1, 1])]; + tensor pretrained_out_29_pad_type_0 = const()[name = tensor("pretrained_out_29_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_29_pad_0 = const()[name = tensor("pretrained_out_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151780928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152600192))), name = tensor("layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152600320)))]; + tensor pretrained_out_29_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_776, groups = var_577, pad = pretrained_out_29_pad_0, pad_type = pretrained_out_29_pad_type_0, strides = var_774, weight = layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_23_cast_fp16)[name = tensor("pretrained_out_29_cast_fp16")]; + tensor var_780 = const()[name = tensor("op_780"), val = tensor([1, 1])]; + tensor var_782 = const()[name = tensor("op_782"), val = tensor([1, 1])]; + tensor input_41_pad_type_0 = const()[name = tensor("input_41_pad_type_0"), val = tensor("custom")]; + tensor input_41_pad_0 = const()[name = tensor("input_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152602944)))]; + tensor input_41_cast_fp16 = conv(dilations = var_782, groups = var_577, pad = input_41_pad_0, pad_type = input_41_pad_type_0, strides = var_780, weight = layers_1_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_786 = const()[name = tensor("op_786"), val = tensor([1, 1])]; + tensor var_788 = const()[name = tensor("op_788"), val = tensor([1, 1])]; + tensor lora_out_57_pad_type_0 = const()[name = tensor("lora_out_57_pad_type_0"), val = tensor("custom")]; + tensor lora_out_57_pad_0 = const()[name = tensor("lora_out_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_59_weight_0_to_fp16 = const()[name = tensor("lora_out_59_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152643968)))]; + tensor lora_out_59_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_788, groups = var_577, pad = lora_out_57_pad_0, pad_type = lora_out_57_pad_type_0, strides = var_786, weight = lora_out_59_weight_0_to_fp16, x = input_41_cast_fp16)[name = tensor("lora_out_59_cast_fp16")]; + tensor query_7_cast_fp16 = add(x = pretrained_out_29_cast_fp16, y = lora_out_59_cast_fp16)[name = tensor("query_7_cast_fp16")]; + tensor var_798 = const()[name = tensor("op_798"), val = tensor([1, 1])]; + tensor var_800 = const()[name = tensor("op_800"), val = tensor([1, 1])]; + tensor pretrained_out_31_pad_type_0 = const()[name = tensor("pretrained_out_31_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_31_pad_0 = const()[name = tensor("pretrained_out_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152684992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153504256))), name = tensor("layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_31_cast_fp16 = conv(dilations = var_800, groups = var_577, pad = pretrained_out_31_pad_0, pad_type = pretrained_out_31_pad_type_0, strides = var_798, weight = layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_31_cast_fp16")]; + tensor var_804 = const()[name = tensor("op_804"), val = tensor([1, 1])]; + tensor var_806 = const()[name = tensor("op_806"), val = tensor([1, 1])]; + tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("custom")]; + tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153504384)))]; + tensor input_43_cast_fp16 = conv(dilations = var_806, groups = var_577, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = var_804, weight = layers_1_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_43_cast_fp16")]; + tensor var_810 = const()[name = tensor("op_810"), val = tensor([1, 1])]; + tensor var_812 = const()[name = tensor("op_812"), val = tensor([1, 1])]; + tensor lora_out_61_pad_type_0 = const()[name = tensor("lora_out_61_pad_type_0"), val = tensor("custom")]; + tensor lora_out_61_pad_0 = const()[name = tensor("lora_out_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_63_weight_0_to_fp16 = const()[name = tensor("lora_out_63_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153545408)))]; + tensor lora_out_63_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_812, groups = var_577, pad = lora_out_61_pad_0, pad_type = lora_out_61_pad_type_0, strides = var_810, weight = lora_out_63_weight_0_to_fp16, x = input_43_cast_fp16)[name = tensor("lora_out_63_cast_fp16")]; + tensor key_7_cast_fp16 = add(x = pretrained_out_31_cast_fp16, y = lora_out_63_cast_fp16)[name = tensor("key_7_cast_fp16")]; + tensor var_823 = const()[name = tensor("op_823"), val = tensor([1, 1])]; + tensor var_825 = const()[name = tensor("op_825"), val = tensor([1, 1])]; + tensor pretrained_out_33_pad_type_0 = const()[name = tensor("pretrained_out_33_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_33_pad_0 = const()[name = tensor("pretrained_out_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153586432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154405696))), name = tensor("layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154405824)))]; + tensor pretrained_out_33_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_825, groups = var_577, pad = pretrained_out_33_pad_0, pad_type = pretrained_out_33_pad_type_0, strides = var_823, weight = layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_33_cast_fp16")]; + tensor var_829 = const()[name = tensor("op_829"), val = tensor([1, 1])]; + tensor var_831 = const()[name = tensor("op_831"), val = tensor([1, 1])]; + tensor input_45_pad_type_0 = const()[name = tensor("input_45_pad_type_0"), val = tensor("custom")]; + tensor input_45_pad_0 = const()[name = tensor("input_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154408448)))]; + tensor input_45_cast_fp16 = conv(dilations = var_831, groups = var_577, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = var_829, weight = layers_1_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_45_cast_fp16")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([1, 1])]; + tensor var_837 = const()[name = tensor("op_837"), val = tensor([1, 1])]; + tensor lora_out_65_pad_type_0 = const()[name = tensor("lora_out_65_pad_type_0"), val = tensor("custom")]; + tensor lora_out_65_pad_0 = const()[name = tensor("lora_out_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_67_weight_0_to_fp16 = const()[name = tensor("lora_out_67_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154449472)))]; + tensor lora_out_67_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_837, groups = var_577, pad = lora_out_65_pad_0, pad_type = lora_out_65_pad_type_0, strides = var_835, weight = lora_out_67_weight_0_to_fp16, x = input_45_cast_fp16)[name = tensor("lora_out_67_cast_fp16")]; + tensor value_7_cast_fp16 = add(x = pretrained_out_33_cast_fp16, y = lora_out_67_cast_fp16)[name = tensor("value_7_cast_fp16")]; + tensor var_844 = const()[name = tensor("op_844"), val = tensor([1, 20, 64, -1])]; + tensor var_845_cast_fp16 = reshape(shape = var_844, x = query_7_cast_fp16)[name = tensor("op_845_cast_fp16")]; + tensor var_846_to_fp16 = const()[name = tensor("op_846_to_fp16"), val = tensor(0x1p-3)]; + tensor var_847_cast_fp16 = mul(x = var_845_cast_fp16, y = var_846_to_fp16)[name = tensor("op_847_cast_fp16")]; + tensor var_848 = const()[name = tensor("op_848"), val = tensor([1, 20, 64, -1])]; + tensor var_849_cast_fp16 = reshape(shape = var_848, x = key_7_cast_fp16)[name = tensor("op_849_cast_fp16")]; + tensor mh_w_11_transpose_x_0 = const()[name = tensor("mh_w_11_transpose_x_0"), val = tensor(true)]; + tensor mh_w_11_transpose_y_0 = const()[name = tensor("mh_w_11_transpose_y_0"), val = tensor(false)]; + tensor mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_847_cast_fp16, y = var_849_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor obj_27_cast_fp16 = softmax(axis = var_570, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 20, 64, -1])]; + tensor var_854_cast_fp16 = reshape(shape = var_853, x = value_7_cast_fp16)[name = tensor("op_854_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(true)]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_854_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_857 = const()[name = tensor("op_857"), val = tensor([1, 1280, 1, -1])]; + tensor input_47_cast_fp16 = reshape(shape = var_857, x = attn_7_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor var_864 = const()[name = tensor("op_864"), val = tensor([1, 1])]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([1, 1])]; + tensor pretrained_out_35_pad_type_0 = const()[name = tensor("pretrained_out_35_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_35_pad_0 = const()[name = tensor("pretrained_out_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154490496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155309760))), name = tensor("layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155309888)))]; + tensor pretrained_out_35_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_866, groups = var_577, pad = pretrained_out_35_pad_0, pad_type = pretrained_out_35_pad_type_0, strides = var_864, weight = layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_47_cast_fp16)[name = tensor("pretrained_out_35_cast_fp16")]; + tensor var_870 = const()[name = tensor("op_870"), val = tensor([1, 1])]; + tensor var_872 = const()[name = tensor("op_872"), val = tensor([1, 1])]; + tensor input_49_pad_type_0 = const()[name = tensor("input_49_pad_type_0"), val = tensor("custom")]; + tensor input_49_pad_0 = const()[name = tensor("input_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155312512)))]; + tensor input_49_cast_fp16 = conv(dilations = var_872, groups = var_577, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = var_870, weight = layers_1_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor var_876 = const()[name = tensor("op_876"), val = tensor([1, 1])]; + tensor var_878 = const()[name = tensor("op_878"), val = tensor([1, 1])]; + tensor lora_out_69_pad_type_0 = const()[name = tensor("lora_out_69_pad_type_0"), val = tensor("custom")]; + tensor lora_out_69_pad_0 = const()[name = tensor("lora_out_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_71_weight_0_to_fp16 = const()[name = tensor("lora_out_71_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155353536)))]; + tensor lora_out_71_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_878, groups = var_577, pad = lora_out_69_pad_0, pad_type = lora_out_69_pad_type_0, strides = var_876, weight = lora_out_71_weight_0_to_fp16, x = input_49_cast_fp16)[name = tensor("lora_out_71_cast_fp16")]; + tensor obj_25_cast_fp16 = add(x = pretrained_out_35_cast_fp16, y = lora_out_71_cast_fp16)[name = tensor("obj_25_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor var_887 = const()[name = tensor("op_887"), val = tensor([1])]; + tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_887, keep_dims = var_578, x = inputs_11_cast_fp16)[name = tensor("channels_mean_11_cast_fp16")]; + tensor zero_mean_11_cast_fp16 = sub(x = inputs_11_cast_fp16, y = channels_mean_11_cast_fp16)[name = tensor("zero_mean_11_cast_fp16")]; + tensor zero_mean_sq_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = zero_mean_11_cast_fp16)[name = tensor("zero_mean_sq_11_cast_fp16")]; + tensor var_891 = const()[name = tensor("op_891"), val = tensor([1])]; + tensor var_892_cast_fp16 = reduce_mean(axes = var_891, keep_dims = var_578, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_893_to_fp16 = const()[name = tensor("op_893_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_894_cast_fp16 = add(x = var_892_cast_fp16, y = var_893_to_fp16)[name = tensor("op_894_cast_fp16")]; + tensor denom_11_epsilon_0 = const()[name = tensor("denom_11_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_11_cast_fp16 = rsqrt(epsilon = denom_11_epsilon_0, x = var_894_cast_fp16)[name = tensor("denom_11_cast_fp16")]; + tensor out_11_cast_fp16 = mul(x = zero_mean_11_cast_fp16, y = denom_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; + tensor input_51_gamma_0_to_fp16 = const()[name = tensor("input_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155394560)))]; + tensor input_51_beta_0_to_fp16 = const()[name = tensor("input_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155397184)))]; + tensor input_51_epsilon_0_to_fp16 = const()[name = tensor("input_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = batch_norm(beta = input_51_beta_0_to_fp16, epsilon = input_51_epsilon_0_to_fp16, gamma = input_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor var_908 = const()[name = tensor("op_908"), val = tensor([1, 1])]; + tensor var_910 = const()[name = tensor("op_910"), val = tensor([1, 1])]; + tensor pretrained_out_37_pad_type_0 = const()[name = tensor("pretrained_out_37_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_37_pad_0 = const()[name = tensor("pretrained_out_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155399808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158676672))), name = tensor("layers_1_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_1_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158676800)))]; + tensor pretrained_out_37_cast_fp16 = conv(bias = layers_1_fc1_pretrained_bias_to_fp16, dilations = var_910, groups = var_577, pad = pretrained_out_37_pad_0, pad_type = pretrained_out_37_pad_type_0, strides = var_908, weight = layers_1_fc1_pretrained_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = tensor("pretrained_out_37_cast_fp16")]; + tensor var_914 = const()[name = tensor("op_914"), val = tensor([1, 1])]; + tensor var_916 = const()[name = tensor("op_916"), val = tensor([1, 1])]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("custom")]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_1_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158687104)))]; + tensor input_53_cast_fp16 = conv(dilations = var_916, groups = var_577, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = var_914, weight = layers_1_fc1_loraA_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor var_920 = const()[name = tensor("op_920"), val = tensor([1, 1])]; + tensor var_922 = const()[name = tensor("op_922"), val = tensor([1, 1])]; + tensor lora_out_73_pad_type_0 = const()[name = tensor("lora_out_73_pad_type_0"), val = tensor("custom")]; + tensor lora_out_73_pad_0 = const()[name = tensor("lora_out_73_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_75_weight_0_to_fp16 = const()[name = tensor("lora_out_75_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158728128)))]; + tensor lora_out_75_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_922, groups = var_577, pad = lora_out_73_pad_0, pad_type = lora_out_73_pad_type_0, strides = var_920, weight = lora_out_75_weight_0_to_fp16, x = input_53_cast_fp16)[name = tensor("lora_out_75_cast_fp16")]; + tensor input_55_cast_fp16 = add(x = pretrained_out_37_cast_fp16, y = lora_out_75_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_mode_0 = const()[name = tensor("input_57_mode_0"), val = tensor("EXACT")]; + tensor input_57_cast_fp16 = gelu(mode = input_57_mode_0, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor var_934 = const()[name = tensor("op_934"), val = tensor([1, 1])]; + tensor var_936 = const()[name = tensor("op_936"), val = tensor([1, 1])]; + tensor pretrained_out_39_pad_type_0 = const()[name = tensor("pretrained_out_39_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_39_pad_0 = const()[name = tensor("pretrained_out_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158892032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162168896))), name = tensor("layers_1_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_1_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162169024)))]; + tensor pretrained_out_39_cast_fp16 = conv(bias = layers_1_fc2_pretrained_bias_to_fp16, dilations = var_936, groups = var_577, pad = pretrained_out_39_pad_0, pad_type = pretrained_out_39_pad_type_0, strides = var_934, weight = layers_1_fc2_pretrained_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = tensor("pretrained_out_39_cast_fp16")]; + tensor var_940 = const()[name = tensor("op_940"), val = tensor([1, 1])]; + tensor var_942 = const()[name = tensor("op_942"), val = tensor([1, 1])]; + tensor input_59_pad_type_0 = const()[name = tensor("input_59_pad_type_0"), val = tensor("custom")]; + tensor input_59_pad_0 = const()[name = tensor("input_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_1_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_1_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162171648)))]; + tensor input_59_cast_fp16 = conv(dilations = var_942, groups = var_577, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = var_940, weight = layers_1_fc2_loraA_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor var_946 = const()[name = tensor("op_946"), val = tensor([1, 1])]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor([1, 1])]; + tensor lora_out_77_pad_type_0 = const()[name = tensor("lora_out_77_pad_type_0"), val = tensor("custom")]; + tensor lora_out_77_pad_0 = const()[name = tensor("lora_out_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_79_weight_0_to_fp16 = const()[name = tensor("lora_out_79_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162335552)))]; + tensor lora_out_79_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_948, groups = var_577, pad = lora_out_77_pad_0, pad_type = lora_out_77_pad_type_0, strides = var_946, weight = lora_out_79_weight_0_to_fp16, x = input_59_cast_fp16)[name = tensor("lora_out_79_cast_fp16")]; + tensor hidden_states_5_cast_fp16 = add(x = pretrained_out_39_cast_fp16, y = lora_out_79_cast_fp16)[name = tensor("hidden_states_5_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_964 = const()[name = tensor("op_964"), val = tensor(3)]; + tensor var_971 = const()[name = tensor("op_971"), val = tensor(1)]; + tensor var_972 = const()[name = tensor("op_972"), val = tensor(true)]; + tensor var_984 = const()[name = tensor("op_984"), val = tensor([1])]; + tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_984, keep_dims = var_972, x = inputs_13_cast_fp16)[name = tensor("channels_mean_13_cast_fp16")]; + tensor zero_mean_13_cast_fp16 = sub(x = inputs_13_cast_fp16, y = channels_mean_13_cast_fp16)[name = tensor("zero_mean_13_cast_fp16")]; + tensor zero_mean_sq_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = zero_mean_13_cast_fp16)[name = tensor("zero_mean_sq_13_cast_fp16")]; + tensor var_988 = const()[name = tensor("op_988"), val = tensor([1])]; + tensor var_989_cast_fp16 = reduce_mean(axes = var_988, keep_dims = var_972, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_989_cast_fp16")]; + tensor var_990_to_fp16 = const()[name = tensor("op_990_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_991_cast_fp16 = add(x = var_989_cast_fp16, y = var_990_to_fp16)[name = tensor("op_991_cast_fp16")]; + tensor denom_13_epsilon_0 = const()[name = tensor("denom_13_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_13_cast_fp16 = rsqrt(epsilon = denom_13_epsilon_0, x = var_991_cast_fp16)[name = tensor("denom_13_cast_fp16")]; + tensor out_13_cast_fp16 = mul(x = zero_mean_13_cast_fp16, y = denom_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; + tensor obj_29_gamma_0_to_fp16 = const()[name = tensor("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162376576)))]; + tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162379200)))]; + tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_29_cast_fp16")]; + tensor var_1009 = const()[name = tensor("op_1009"), val = tensor([1, 1])]; + tensor var_1011 = const()[name = tensor("op_1011"), val = tensor([1, 1])]; + tensor pretrained_out_41_pad_type_0 = const()[name = tensor("pretrained_out_41_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_41_pad_0 = const()[name = tensor("pretrained_out_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162381824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163201088))), name = tensor("layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163201216)))]; + tensor pretrained_out_41_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1011, groups = var_971, pad = pretrained_out_41_pad_0, pad_type = pretrained_out_41_pad_type_0, strides = var_1009, weight = layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor("pretrained_out_41_cast_fp16")]; + tensor var_1015 = const()[name = tensor("op_1015"), val = tensor([1, 1])]; + tensor var_1017 = const()[name = tensor("op_1017"), val = tensor([1, 1])]; + tensor input_61_pad_type_0 = const()[name = tensor("input_61_pad_type_0"), val = tensor("custom")]; + tensor input_61_pad_0 = const()[name = tensor("input_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163203840)))]; + tensor input_61_cast_fp16 = conv(dilations = var_1017, groups = var_971, pad = input_61_pad_0, pad_type = input_61_pad_type_0, strides = var_1015, weight = layers_2_self_attn_q_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor var_1021 = const()[name = tensor("op_1021"), val = tensor([1, 1])]; + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([1, 1])]; + tensor lora_out_81_pad_type_0 = const()[name = tensor("lora_out_81_pad_type_0"), val = tensor("custom")]; + tensor lora_out_81_pad_0 = const()[name = tensor("lora_out_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_83_weight_0_to_fp16 = const()[name = tensor("lora_out_83_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163244864)))]; + tensor lora_out_83_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1023, groups = var_971, pad = lora_out_81_pad_0, pad_type = lora_out_81_pad_type_0, strides = var_1021, weight = lora_out_83_weight_0_to_fp16, x = input_61_cast_fp16)[name = tensor("lora_out_83_cast_fp16")]; + tensor query_9_cast_fp16 = add(x = pretrained_out_41_cast_fp16, y = lora_out_83_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor var_1033 = const()[name = tensor("op_1033"), val = tensor([1, 1])]; + tensor var_1035 = const()[name = tensor("op_1035"), val = tensor([1, 1])]; + tensor pretrained_out_43_pad_type_0 = const()[name = tensor("pretrained_out_43_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_43_pad_0 = const()[name = tensor("pretrained_out_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163285888))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164105152))), name = tensor("layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_43_cast_fp16 = conv(dilations = var_1035, groups = var_971, pad = pretrained_out_43_pad_0, pad_type = pretrained_out_43_pad_type_0, strides = var_1033, weight = layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor("pretrained_out_43_cast_fp16")]; + tensor var_1039 = const()[name = tensor("op_1039"), val = tensor([1, 1])]; + tensor var_1041 = const()[name = tensor("op_1041"), val = tensor([1, 1])]; + tensor input_63_pad_type_0 = const()[name = tensor("input_63_pad_type_0"), val = tensor("custom")]; + tensor input_63_pad_0 = const()[name = tensor("input_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164105280)))]; + tensor input_63_cast_fp16 = conv(dilations = var_1041, groups = var_971, pad = input_63_pad_0, pad_type = input_63_pad_type_0, strides = var_1039, weight = layers_2_self_attn_k_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor var_1045 = const()[name = tensor("op_1045"), val = tensor([1, 1])]; + tensor var_1047 = const()[name = tensor("op_1047"), val = tensor([1, 1])]; + tensor lora_out_85_pad_type_0 = const()[name = tensor("lora_out_85_pad_type_0"), val = tensor("custom")]; + tensor lora_out_85_pad_0 = const()[name = tensor("lora_out_85_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_87_weight_0_to_fp16 = const()[name = tensor("lora_out_87_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164146304)))]; + tensor lora_out_87_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1047, groups = var_971, pad = lora_out_85_pad_0, pad_type = lora_out_85_pad_type_0, strides = var_1045, weight = lora_out_87_weight_0_to_fp16, x = input_63_cast_fp16)[name = tensor("lora_out_87_cast_fp16")]; + tensor current_key_5_cast_fp16 = add(x = pretrained_out_43_cast_fp16, y = lora_out_87_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; + tensor var_1058 = const()[name = tensor("op_1058"), val = tensor([1, 1])]; + tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1])]; + tensor pretrained_out_45_pad_type_0 = const()[name = tensor("pretrained_out_45_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_45_pad_0 = const()[name = tensor("pretrained_out_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164187328))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165006592))), name = tensor("layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165006720)))]; + tensor pretrained_out_45_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1060, groups = var_971, pad = pretrained_out_45_pad_0, pad_type = pretrained_out_45_pad_type_0, strides = var_1058, weight = layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor("pretrained_out_45_cast_fp16")]; + tensor var_1064 = const()[name = tensor("op_1064"), val = tensor([1, 1])]; + tensor var_1066 = const()[name = tensor("op_1066"), val = tensor([1, 1])]; + tensor input_65_pad_type_0 = const()[name = tensor("input_65_pad_type_0"), val = tensor("custom")]; + tensor input_65_pad_0 = const()[name = tensor("input_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165009344)))]; + tensor input_65_cast_fp16 = conv(dilations = var_1066, groups = var_971, pad = input_65_pad_0, pad_type = input_65_pad_type_0, strides = var_1064, weight = layers_2_self_attn_v_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor var_1070 = const()[name = tensor("op_1070"), val = tensor([1, 1])]; + tensor var_1072 = const()[name = tensor("op_1072"), val = tensor([1, 1])]; + tensor lora_out_89_pad_type_0 = const()[name = tensor("lora_out_89_pad_type_0"), val = tensor("custom")]; + tensor lora_out_89_pad_0 = const()[name = tensor("lora_out_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_91_weight_0_to_fp16 = const()[name = tensor("lora_out_91_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165050368)))]; + tensor lora_out_91_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1072, groups = var_971, pad = lora_out_89_pad_0, pad_type = lora_out_89_pad_type_0, strides = var_1070, weight = lora_out_91_weight_0_to_fp16, x = input_65_cast_fp16)[name = tensor("lora_out_91_cast_fp16")]; + tensor current_value_5_cast_fp16 = add(x = pretrained_out_45_cast_fp16, y = lora_out_91_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; + tensor var_1082_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_1082_cast_fp16")]; + tensor var_1084_cast_fp16 = mul(x = var_103_cast_fp16_2, y = var_295_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = var_1082_cast_fp16, y = var_1084_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_1086_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_1086_cast_fp16")]; + tensor var_1088_cast_fp16 = mul(x = var_138_cast_fp16_2, y = var_295_cast_fp16)[name = tensor("op_1088_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = var_1086_cast_fp16, y = var_1088_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([1, 20, 64, -1])]; + tensor var_1092_cast_fp16 = reshape(shape = var_1091, x = query_9_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1093_to_fp16 = const()[name = tensor("op_1093_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1094_cast_fp16 = mul(x = var_1092_cast_fp16, y = var_1093_to_fp16)[name = tensor("op_1094_cast_fp16")]; + tensor var_1095 = const()[name = tensor("op_1095"), val = tensor([1, 20, 64, -1])]; + tensor var_1096_cast_fp16 = reshape(shape = var_1095, x = key_9_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor mh_w_13_transpose_x_0 = const()[name = tensor("mh_w_13_transpose_x_0"), val = tensor(true)]; + tensor mh_w_13_transpose_y_0 = const()[name = tensor("mh_w_13_transpose_y_0"), val = tensor(false)]; + tensor mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_1094_cast_fp16, y = var_1096_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_1104_cast_fp16 = softmax(axis = var_964, x = mh_w_15_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor var_1105 = const()[name = tensor("op_1105"), val = tensor([1, 20, 64, -1])]; + tensor var_1106_cast_fp16 = reshape(shape = var_1105, x = value_9_cast_fp16)[name = tensor("op_1106_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(true)]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_1106_cast_fp16, y = var_1104_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1109 = const()[name = tensor("op_1109"), val = tensor([1, 1280, 1, -1])]; + tensor input_67_cast_fp16 = reshape(shape = var_1109, x = attn_9_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor var_1116 = const()[name = tensor("op_1116"), val = tensor([1, 1])]; + tensor var_1118 = const()[name = tensor("op_1118"), val = tensor([1, 1])]; + tensor pretrained_out_47_pad_type_0 = const()[name = tensor("pretrained_out_47_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_47_pad_0 = const()[name = tensor("pretrained_out_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165091392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165910656))), name = tensor("layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165910784)))]; + tensor pretrained_out_47_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1118, groups = var_971, pad = pretrained_out_47_pad_0, pad_type = pretrained_out_47_pad_type_0, strides = var_1116, weight = layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = tensor("pretrained_out_47_cast_fp16")]; + tensor var_1122 = const()[name = tensor("op_1122"), val = tensor([1, 1])]; + tensor var_1124 = const()[name = tensor("op_1124"), val = tensor([1, 1])]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("custom")]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165913408)))]; + tensor input_69_cast_fp16 = conv(dilations = var_1124, groups = var_971, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = var_1122, weight = layers_2_self_attn_o_proj_loraA_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 1])]; + tensor var_1130 = const()[name = tensor("op_1130"), val = tensor([1, 1])]; + tensor lora_out_93_pad_type_0 = const()[name = tensor("lora_out_93_pad_type_0"), val = tensor("custom")]; + tensor lora_out_93_pad_0 = const()[name = tensor("lora_out_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_95_weight_0_to_fp16 = const()[name = tensor("lora_out_95_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165954432)))]; + tensor lora_out_95_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1130, groups = var_971, pad = lora_out_93_pad_0, pad_type = lora_out_93_pad_type_0, strides = var_1128, weight = lora_out_95_weight_0_to_fp16, x = input_69_cast_fp16)[name = tensor("lora_out_95_cast_fp16")]; + tensor obj_35_cast_fp16 = add(x = pretrained_out_47_cast_fp16, y = lora_out_95_cast_fp16)[name = tensor("obj_35_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([1])]; + tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_1143, keep_dims = var_972, x = inputs_15_cast_fp16)[name = tensor("channels_mean_15_cast_fp16")]; + tensor zero_mean_15_cast_fp16 = sub(x = inputs_15_cast_fp16, y = channels_mean_15_cast_fp16)[name = tensor("zero_mean_15_cast_fp16")]; + tensor zero_mean_sq_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = zero_mean_15_cast_fp16)[name = tensor("zero_mean_sq_15_cast_fp16")]; + tensor var_1147 = const()[name = tensor("op_1147"), val = tensor([1])]; + tensor var_1148_cast_fp16 = reduce_mean(axes = var_1147, keep_dims = var_972, x = zero_mean_sq_15_cast_fp16)[name = tensor("op_1148_cast_fp16")]; + tensor var_1149_to_fp16 = const()[name = tensor("op_1149_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1150_cast_fp16 = add(x = var_1148_cast_fp16, y = var_1149_to_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor denom_15_epsilon_0 = const()[name = tensor("denom_15_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_15_cast_fp16 = rsqrt(epsilon = denom_15_epsilon_0, x = var_1150_cast_fp16)[name = tensor("denom_15_cast_fp16")]; + tensor out_15_cast_fp16 = mul(x = zero_mean_15_cast_fp16, y = denom_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; + tensor obj_37_gamma_0_to_fp16 = const()[name = tensor("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165995456)))]; + tensor obj_37_beta_0_to_fp16 = const()[name = tensor("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165998080)))]; + tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("obj_37_cast_fp16")]; + tensor var_1168 = const()[name = tensor("op_1168"), val = tensor([1, 1])]; + tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([1, 1])]; + tensor pretrained_out_49_pad_type_0 = const()[name = tensor("pretrained_out_49_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_49_pad_0 = const()[name = tensor("pretrained_out_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166000704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166819968))), name = tensor("layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166820096)))]; + tensor pretrained_out_49_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1170, groups = var_971, pad = pretrained_out_49_pad_0, pad_type = pretrained_out_49_pad_type_0, strides = var_1168, weight = layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor("pretrained_out_49_cast_fp16")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([1, 1])]; + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor([1, 1])]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("custom")]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166822720)))]; + tensor input_71_cast_fp16 = conv(dilations = var_1176, groups = var_971, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = var_1174, weight = layers_2_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor var_1180 = const()[name = tensor("op_1180"), val = tensor([1, 1])]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([1, 1])]; + tensor lora_out_97_pad_type_0 = const()[name = tensor("lora_out_97_pad_type_0"), val = tensor("custom")]; + tensor lora_out_97_pad_0 = const()[name = tensor("lora_out_97_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_99_weight_0_to_fp16 = const()[name = tensor("lora_out_99_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166863744)))]; + tensor lora_out_99_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1182, groups = var_971, pad = lora_out_97_pad_0, pad_type = lora_out_97_pad_type_0, strides = var_1180, weight = lora_out_99_weight_0_to_fp16, x = input_71_cast_fp16)[name = tensor("lora_out_99_cast_fp16")]; + tensor query_11_cast_fp16 = add(x = pretrained_out_49_cast_fp16, y = lora_out_99_cast_fp16)[name = tensor("query_11_cast_fp16")]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 1])]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([1, 1])]; + tensor pretrained_out_51_pad_type_0 = const()[name = tensor("pretrained_out_51_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_51_pad_0 = const()[name = tensor("pretrained_out_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166904768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167724032))), name = tensor("layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_51_cast_fp16 = conv(dilations = var_1194, groups = var_971, pad = pretrained_out_51_pad_0, pad_type = pretrained_out_51_pad_type_0, strides = var_1192, weight = layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_51_cast_fp16")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([1, 1])]; + tensor var_1200 = const()[name = tensor("op_1200"), val = tensor([1, 1])]; + tensor input_73_pad_type_0 = const()[name = tensor("input_73_pad_type_0"), val = tensor("custom")]; + tensor input_73_pad_0 = const()[name = tensor("input_73_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167724160)))]; + tensor input_73_cast_fp16 = conv(dilations = var_1200, groups = var_971, pad = input_73_pad_0, pad_type = input_73_pad_type_0, strides = var_1198, weight = layers_2_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_73_cast_fp16")]; + tensor var_1204 = const()[name = tensor("op_1204"), val = tensor([1, 1])]; + tensor var_1206 = const()[name = tensor("op_1206"), val = tensor([1, 1])]; + tensor lora_out_101_pad_type_0 = const()[name = tensor("lora_out_101_pad_type_0"), val = tensor("custom")]; + tensor lora_out_101_pad_0 = const()[name = tensor("lora_out_101_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_103_weight_0_to_fp16 = const()[name = tensor("lora_out_103_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167765184)))]; + tensor lora_out_103_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1206, groups = var_971, pad = lora_out_101_pad_0, pad_type = lora_out_101_pad_type_0, strides = var_1204, weight = lora_out_103_weight_0_to_fp16, x = input_73_cast_fp16)[name = tensor("lora_out_103_cast_fp16")]; + tensor key_11_cast_fp16 = add(x = pretrained_out_51_cast_fp16, y = lora_out_103_cast_fp16)[name = tensor("key_11_cast_fp16")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([1, 1])]; + tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 1])]; + tensor pretrained_out_53_pad_type_0 = const()[name = tensor("pretrained_out_53_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_53_pad_0 = const()[name = tensor("pretrained_out_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167806208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168625472))), name = tensor("layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168625600)))]; + tensor pretrained_out_53_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1219, groups = var_971, pad = pretrained_out_53_pad_0, pad_type = pretrained_out_53_pad_type_0, strides = var_1217, weight = layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_53_cast_fp16")]; + tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([1, 1])]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([1, 1])]; + tensor input_75_pad_type_0 = const()[name = tensor("input_75_pad_type_0"), val = tensor("custom")]; + tensor input_75_pad_0 = const()[name = tensor("input_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168628224)))]; + tensor input_75_cast_fp16 = conv(dilations = var_1225, groups = var_971, pad = input_75_pad_0, pad_type = input_75_pad_type_0, strides = var_1223, weight = layers_2_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_75_cast_fp16")]; + tensor var_1229 = const()[name = tensor("op_1229"), val = tensor([1, 1])]; + tensor var_1231 = const()[name = tensor("op_1231"), val = tensor([1, 1])]; + tensor lora_out_105_pad_type_0 = const()[name = tensor("lora_out_105_pad_type_0"), val = tensor("custom")]; + tensor lora_out_105_pad_0 = const()[name = tensor("lora_out_105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_107_weight_0_to_fp16 = const()[name = tensor("lora_out_107_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168669248)))]; + tensor lora_out_107_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1231, groups = var_971, pad = lora_out_105_pad_0, pad_type = lora_out_105_pad_type_0, strides = var_1229, weight = lora_out_107_weight_0_to_fp16, x = input_75_cast_fp16)[name = tensor("lora_out_107_cast_fp16")]; + tensor value_11_cast_fp16 = add(x = pretrained_out_53_cast_fp16, y = lora_out_107_cast_fp16)[name = tensor("value_11_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([1, 20, 64, -1])]; + tensor var_1239_cast_fp16 = reshape(shape = var_1238, x = query_11_cast_fp16)[name = tensor("op_1239_cast_fp16")]; + tensor var_1240_to_fp16 = const()[name = tensor("op_1240_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1241_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_1240_to_fp16)[name = tensor("op_1241_cast_fp16")]; + tensor var_1242 = const()[name = tensor("op_1242"), val = tensor([1, 20, 64, -1])]; + tensor var_1243_cast_fp16 = reshape(shape = var_1242, x = key_11_cast_fp16)[name = tensor("op_1243_cast_fp16")]; + tensor mh_w_17_transpose_x_0 = const()[name = tensor("mh_w_17_transpose_x_0"), val = tensor(true)]; + tensor mh_w_17_transpose_y_0 = const()[name = tensor("mh_w_17_transpose_y_0"), val = tensor(false)]; + tensor mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_1241_cast_fp16, y = var_1243_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor obj_41_cast_fp16 = softmax(axis = var_964, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor var_1247 = const()[name = tensor("op_1247"), val = tensor([1, 20, 64, -1])]; + tensor var_1248_cast_fp16 = reshape(shape = var_1247, x = value_11_cast_fp16)[name = tensor("op_1248_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(true)]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_1248_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 1280, 1, -1])]; + tensor input_77_cast_fp16 = reshape(shape = var_1251, x = attn_11_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor var_1258 = const()[name = tensor("op_1258"), val = tensor([1, 1])]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([1, 1])]; + tensor pretrained_out_55_pad_type_0 = const()[name = tensor("pretrained_out_55_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_55_pad_0 = const()[name = tensor("pretrained_out_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168710272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169529536))), name = tensor("layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169529664)))]; + tensor pretrained_out_55_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1260, groups = var_971, pad = pretrained_out_55_pad_0, pad_type = pretrained_out_55_pad_type_0, strides = var_1258, weight = layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_77_cast_fp16)[name = tensor("pretrained_out_55_cast_fp16")]; + tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([1, 1])]; + tensor var_1266 = const()[name = tensor("op_1266"), val = tensor([1, 1])]; + tensor input_79_pad_type_0 = const()[name = tensor("input_79_pad_type_0"), val = tensor("custom")]; + tensor input_79_pad_0 = const()[name = tensor("input_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169532288)))]; + tensor input_79_cast_fp16 = conv(dilations = var_1266, groups = var_971, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = var_1264, weight = layers_2_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor([1, 1])]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 1])]; + tensor lora_out_109_pad_type_0 = const()[name = tensor("lora_out_109_pad_type_0"), val = tensor("custom")]; + tensor lora_out_109_pad_0 = const()[name = tensor("lora_out_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_111_weight_0_to_fp16 = const()[name = tensor("lora_out_111_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169573312)))]; + tensor lora_out_111_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1272, groups = var_971, pad = lora_out_109_pad_0, pad_type = lora_out_109_pad_type_0, strides = var_1270, weight = lora_out_111_weight_0_to_fp16, x = input_79_cast_fp16)[name = tensor("lora_out_111_cast_fp16")]; + tensor obj_39_cast_fp16 = add(x = pretrained_out_55_cast_fp16, y = lora_out_111_cast_fp16)[name = tensor("obj_39_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_1281 = const()[name = tensor("op_1281"), val = tensor([1])]; + tensor channels_mean_17_cast_fp16 = reduce_mean(axes = var_1281, keep_dims = var_972, x = inputs_17_cast_fp16)[name = tensor("channels_mean_17_cast_fp16")]; + tensor zero_mean_17_cast_fp16 = sub(x = inputs_17_cast_fp16, y = channels_mean_17_cast_fp16)[name = tensor("zero_mean_17_cast_fp16")]; + tensor zero_mean_sq_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = zero_mean_17_cast_fp16)[name = tensor("zero_mean_sq_17_cast_fp16")]; + tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([1])]; + tensor var_1286_cast_fp16 = reduce_mean(axes = var_1285, keep_dims = var_972, x = zero_mean_sq_17_cast_fp16)[name = tensor("op_1286_cast_fp16")]; + tensor var_1287_to_fp16 = const()[name = tensor("op_1287_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1288_cast_fp16 = add(x = var_1286_cast_fp16, y = var_1287_to_fp16)[name = tensor("op_1288_cast_fp16")]; + tensor denom_17_epsilon_0 = const()[name = tensor("denom_17_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_17_cast_fp16 = rsqrt(epsilon = denom_17_epsilon_0, x = var_1288_cast_fp16)[name = tensor("denom_17_cast_fp16")]; + tensor out_17_cast_fp16 = mul(x = zero_mean_17_cast_fp16, y = denom_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; + tensor input_81_gamma_0_to_fp16 = const()[name = tensor("input_81_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169614336)))]; + tensor input_81_beta_0_to_fp16 = const()[name = tensor("input_81_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169616960)))]; + tensor input_81_epsilon_0_to_fp16 = const()[name = tensor("input_81_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_81_cast_fp16 = batch_norm(beta = input_81_beta_0_to_fp16, epsilon = input_81_epsilon_0_to_fp16, gamma = input_81_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor var_1302 = const()[name = tensor("op_1302"), val = tensor([1, 1])]; + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([1, 1])]; + tensor pretrained_out_57_pad_type_0 = const()[name = tensor("pretrained_out_57_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_57_pad_0 = const()[name = tensor("pretrained_out_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169619584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172896448))), name = tensor("layers_2_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_2_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172896576)))]; + tensor pretrained_out_57_cast_fp16 = conv(bias = layers_2_fc1_pretrained_bias_to_fp16, dilations = var_1304, groups = var_971, pad = pretrained_out_57_pad_0, pad_type = pretrained_out_57_pad_type_0, strides = var_1302, weight = layers_2_fc1_pretrained_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = tensor("pretrained_out_57_cast_fp16")]; + tensor var_1308 = const()[name = tensor("op_1308"), val = tensor([1, 1])]; + tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([1, 1])]; + tensor input_83_pad_type_0 = const()[name = tensor("input_83_pad_type_0"), val = tensor("custom")]; + tensor input_83_pad_0 = const()[name = tensor("input_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_2_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172906880)))]; + tensor input_83_cast_fp16 = conv(dilations = var_1310, groups = var_971, pad = input_83_pad_0, pad_type = input_83_pad_type_0, strides = var_1308, weight = layers_2_fc1_loraA_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 1])]; + tensor var_1316 = const()[name = tensor("op_1316"), val = tensor([1, 1])]; + tensor lora_out_113_pad_type_0 = const()[name = tensor("lora_out_113_pad_type_0"), val = tensor("custom")]; + tensor lora_out_113_pad_0 = const()[name = tensor("lora_out_113_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_115_weight_0_to_fp16 = const()[name = tensor("lora_out_115_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172947904)))]; + tensor lora_out_115_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_1316, groups = var_971, pad = lora_out_113_pad_0, pad_type = lora_out_113_pad_type_0, strides = var_1314, weight = lora_out_115_weight_0_to_fp16, x = input_83_cast_fp16)[name = tensor("lora_out_115_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = pretrained_out_57_cast_fp16, y = lora_out_115_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_mode_0 = const()[name = tensor("input_87_mode_0"), val = tensor("EXACT")]; + tensor input_87_cast_fp16 = gelu(mode = input_87_mode_0, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([1, 1])]; + tensor var_1330 = const()[name = tensor("op_1330"), val = tensor([1, 1])]; + tensor pretrained_out_59_pad_type_0 = const()[name = tensor("pretrained_out_59_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_59_pad_0 = const()[name = tensor("pretrained_out_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173111808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176388672))), name = tensor("layers_2_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_2_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176388800)))]; + tensor pretrained_out_59_cast_fp16 = conv(bias = layers_2_fc2_pretrained_bias_to_fp16, dilations = var_1330, groups = var_971, pad = pretrained_out_59_pad_0, pad_type = pretrained_out_59_pad_type_0, strides = var_1328, weight = layers_2_fc2_pretrained_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = tensor("pretrained_out_59_cast_fp16")]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 1])]; + tensor var_1336 = const()[name = tensor("op_1336"), val = tensor([1, 1])]; + tensor input_89_pad_type_0 = const()[name = tensor("input_89_pad_type_0"), val = tensor("custom")]; + tensor input_89_pad_0 = const()[name = tensor("input_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_2_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_2_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176391424)))]; + tensor input_89_cast_fp16 = conv(dilations = var_1336, groups = var_971, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = var_1334, weight = layers_2_fc2_loraA_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([1, 1])]; + tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([1, 1])]; + tensor lora_out_117_pad_type_0 = const()[name = tensor("lora_out_117_pad_type_0"), val = tensor("custom")]; + tensor lora_out_117_pad_0 = const()[name = tensor("lora_out_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_119_weight_0_to_fp16 = const()[name = tensor("lora_out_119_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176555328)))]; + tensor lora_out_119_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1342, groups = var_971, pad = lora_out_117_pad_0, pad_type = lora_out_117_pad_type_0, strides = var_1340, weight = lora_out_119_weight_0_to_fp16, x = input_89_cast_fp16)[name = tensor("lora_out_119_cast_fp16")]; + tensor hidden_states_7_cast_fp16 = add(x = pretrained_out_59_cast_fp16, y = lora_out_119_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor(3)]; + tensor var_1365 = const()[name = tensor("op_1365"), val = tensor(1)]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor(true)]; + tensor var_1378 = const()[name = tensor("op_1378"), val = tensor([1])]; + tensor channels_mean_19_cast_fp16 = reduce_mean(axes = var_1378, keep_dims = var_1366, x = inputs_19_cast_fp16)[name = tensor("channels_mean_19_cast_fp16")]; + tensor zero_mean_19_cast_fp16 = sub(x = inputs_19_cast_fp16, y = channels_mean_19_cast_fp16)[name = tensor("zero_mean_19_cast_fp16")]; + tensor zero_mean_sq_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = zero_mean_19_cast_fp16)[name = tensor("zero_mean_sq_19_cast_fp16")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor([1])]; + tensor var_1383_cast_fp16 = reduce_mean(axes = var_1382, keep_dims = var_1366, x = zero_mean_sq_19_cast_fp16)[name = tensor("op_1383_cast_fp16")]; + tensor var_1384_to_fp16 = const()[name = tensor("op_1384_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1385_cast_fp16 = add(x = var_1383_cast_fp16, y = var_1384_to_fp16)[name = tensor("op_1385_cast_fp16")]; + tensor denom_19_epsilon_0 = const()[name = tensor("denom_19_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_19_cast_fp16 = rsqrt(epsilon = denom_19_epsilon_0, x = var_1385_cast_fp16)[name = tensor("denom_19_cast_fp16")]; + tensor out_19_cast_fp16 = mul(x = zero_mean_19_cast_fp16, y = denom_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; + tensor obj_43_gamma_0_to_fp16 = const()[name = tensor("obj_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176596352)))]; + tensor obj_43_beta_0_to_fp16 = const()[name = tensor("obj_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176598976)))]; + tensor obj_43_epsilon_0_to_fp16 = const()[name = tensor("obj_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("obj_43_cast_fp16")]; + tensor var_1403 = const()[name = tensor("op_1403"), val = tensor([1, 1])]; + tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([1, 1])]; + tensor pretrained_out_61_pad_type_0 = const()[name = tensor("pretrained_out_61_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_61_pad_0 = const()[name = tensor("pretrained_out_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176601600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177420864))), name = tensor("layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177420992)))]; + tensor pretrained_out_61_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1405, groups = var_1365, pad = pretrained_out_61_pad_0, pad_type = pretrained_out_61_pad_type_0, strides = var_1403, weight = layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor("pretrained_out_61_cast_fp16")]; + tensor var_1409 = const()[name = tensor("op_1409"), val = tensor([1, 1])]; + tensor var_1411 = const()[name = tensor("op_1411"), val = tensor([1, 1])]; + tensor input_91_pad_type_0 = const()[name = tensor("input_91_pad_type_0"), val = tensor("custom")]; + tensor input_91_pad_0 = const()[name = tensor("input_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177423616)))]; + tensor input_91_cast_fp16 = conv(dilations = var_1411, groups = var_1365, pad = input_91_pad_0, pad_type = input_91_pad_type_0, strides = var_1409, weight = layers_3_self_attn_q_proj_loraA_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor var_1415 = const()[name = tensor("op_1415"), val = tensor([1, 1])]; + tensor var_1417 = const()[name = tensor("op_1417"), val = tensor([1, 1])]; + tensor lora_out_121_pad_type_0 = const()[name = tensor("lora_out_121_pad_type_0"), val = tensor("custom")]; + tensor lora_out_121_pad_0 = const()[name = tensor("lora_out_121_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_123_weight_0_to_fp16 = const()[name = tensor("lora_out_123_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177464640)))]; + tensor lora_out_123_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1417, groups = var_1365, pad = lora_out_121_pad_0, pad_type = lora_out_121_pad_type_0, strides = var_1415, weight = lora_out_123_weight_0_to_fp16, x = input_91_cast_fp16)[name = tensor("lora_out_123_cast_fp16")]; + tensor query_13_cast_fp16 = add(x = pretrained_out_61_cast_fp16, y = lora_out_123_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor var_1427 = const()[name = tensor("op_1427"), val = tensor([1, 1])]; + tensor var_1429 = const()[name = tensor("op_1429"), val = tensor([1, 1])]; + tensor pretrained_out_63_pad_type_0 = const()[name = tensor("pretrained_out_63_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_63_pad_0 = const()[name = tensor("pretrained_out_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177505664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178324928))), name = tensor("layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_63_cast_fp16 = conv(dilations = var_1429, groups = var_1365, pad = pretrained_out_63_pad_0, pad_type = pretrained_out_63_pad_type_0, strides = var_1427, weight = layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor("pretrained_out_63_cast_fp16")]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([1, 1])]; + tensor var_1435 = const()[name = tensor("op_1435"), val = tensor([1, 1])]; + tensor input_93_pad_type_0 = const()[name = tensor("input_93_pad_type_0"), val = tensor("custom")]; + tensor input_93_pad_0 = const()[name = tensor("input_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178325056)))]; + tensor input_93_cast_fp16 = conv(dilations = var_1435, groups = var_1365, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = var_1433, weight = layers_3_self_attn_k_proj_loraA_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor var_1439 = const()[name = tensor("op_1439"), val = tensor([1, 1])]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([1, 1])]; + tensor lora_out_125_pad_type_0 = const()[name = tensor("lora_out_125_pad_type_0"), val = tensor("custom")]; + tensor lora_out_125_pad_0 = const()[name = tensor("lora_out_125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_127_weight_0_to_fp16 = const()[name = tensor("lora_out_127_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178366080)))]; + tensor lora_out_127_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1441, groups = var_1365, pad = lora_out_125_pad_0, pad_type = lora_out_125_pad_type_0, strides = var_1439, weight = lora_out_127_weight_0_to_fp16, x = input_93_cast_fp16)[name = tensor("lora_out_127_cast_fp16")]; + tensor current_key_7_cast_fp16 = add(x = pretrained_out_63_cast_fp16, y = lora_out_127_cast_fp16)[name = tensor("current_key_7_cast_fp16")]; + tensor var_1452 = const()[name = tensor("op_1452"), val = tensor([1, 1])]; + tensor var_1454 = const()[name = tensor("op_1454"), val = tensor([1, 1])]; + tensor pretrained_out_65_pad_type_0 = const()[name = tensor("pretrained_out_65_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_65_pad_0 = const()[name = tensor("pretrained_out_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178407104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179226368))), name = tensor("layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179226496)))]; + tensor pretrained_out_65_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1454, groups = var_1365, pad = pretrained_out_65_pad_0, pad_type = pretrained_out_65_pad_type_0, strides = var_1452, weight = layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor("pretrained_out_65_cast_fp16")]; + tensor var_1458 = const()[name = tensor("op_1458"), val = tensor([1, 1])]; + tensor var_1460 = const()[name = tensor("op_1460"), val = tensor([1, 1])]; + tensor input_95_pad_type_0 = const()[name = tensor("input_95_pad_type_0"), val = tensor("custom")]; + tensor input_95_pad_0 = const()[name = tensor("input_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179229120)))]; + tensor input_95_cast_fp16 = conv(dilations = var_1460, groups = var_1365, pad = input_95_pad_0, pad_type = input_95_pad_type_0, strides = var_1458, weight = layers_3_self_attn_v_proj_loraA_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor var_1464 = const()[name = tensor("op_1464"), val = tensor([1, 1])]; + tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([1, 1])]; + tensor lora_out_129_pad_type_0 = const()[name = tensor("lora_out_129_pad_type_0"), val = tensor("custom")]; + tensor lora_out_129_pad_0 = const()[name = tensor("lora_out_129_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_131_weight_0_to_fp16 = const()[name = tensor("lora_out_131_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179270144)))]; + tensor lora_out_131_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1466, groups = var_1365, pad = lora_out_129_pad_0, pad_type = lora_out_129_pad_type_0, strides = var_1464, weight = lora_out_131_weight_0_to_fp16, x = input_95_cast_fp16)[name = tensor("lora_out_131_cast_fp16")]; + tensor current_value_7_cast_fp16 = add(x = pretrained_out_65_cast_fp16, y = lora_out_131_cast_fp16)[name = tensor("current_value_7_cast_fp16")]; + tensor var_1476_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_1476_cast_fp16")]; + tensor var_1478_cast_fp16 = mul(x = var_103_cast_fp16_3, y = var_295_cast_fp16)[name = tensor("op_1478_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = var_1476_cast_fp16, y = var_1478_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_1480_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_1480_cast_fp16")]; + tensor var_1482_cast_fp16 = mul(x = var_138_cast_fp16_3, y = var_295_cast_fp16)[name = tensor("op_1482_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = var_1480_cast_fp16, y = var_1482_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_1485 = const()[name = tensor("op_1485"), val = tensor([1, 20, 64, -1])]; + tensor var_1486_cast_fp16 = reshape(shape = var_1485, x = query_13_cast_fp16)[name = tensor("op_1486_cast_fp16")]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1488_cast_fp16 = mul(x = var_1486_cast_fp16, y = var_1487_to_fp16)[name = tensor("op_1488_cast_fp16")]; + tensor var_1489 = const()[name = tensor("op_1489"), val = tensor([1, 20, 64, -1])]; + tensor var_1490_cast_fp16 = reshape(shape = var_1489, x = key_13_cast_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor mh_w_19_transpose_x_0 = const()[name = tensor("mh_w_19_transpose_x_0"), val = tensor(true)]; + tensor mh_w_19_transpose_y_0 = const()[name = tensor("mh_w_19_transpose_y_0"), val = tensor(false)]; + tensor mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_1488_cast_fp16, y = var_1490_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_1498_cast_fp16 = softmax(axis = var_1358, x = mh_w_21_cast_fp16)[name = tensor("op_1498_cast_fp16")]; + tensor var_1499 = const()[name = tensor("op_1499"), val = tensor([1, 20, 64, -1])]; + tensor var_1500_cast_fp16 = reshape(shape = var_1499, x = value_13_cast_fp16)[name = tensor("op_1500_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(true)]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_1500_cast_fp16, y = var_1498_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1503 = const()[name = tensor("op_1503"), val = tensor([1, 1280, 1, -1])]; + tensor input_97_cast_fp16 = reshape(shape = var_1503, x = attn_13_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor var_1510 = const()[name = tensor("op_1510"), val = tensor([1, 1])]; + tensor var_1512 = const()[name = tensor("op_1512"), val = tensor([1, 1])]; + tensor pretrained_out_67_pad_type_0 = const()[name = tensor("pretrained_out_67_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_67_pad_0 = const()[name = tensor("pretrained_out_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179311168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180130432))), name = tensor("layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180130560)))]; + tensor pretrained_out_67_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1512, groups = var_1365, pad = pretrained_out_67_pad_0, pad_type = pretrained_out_67_pad_type_0, strides = var_1510, weight = layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_97_cast_fp16)[name = tensor("pretrained_out_67_cast_fp16")]; + tensor var_1516 = const()[name = tensor("op_1516"), val = tensor([1, 1])]; + tensor var_1518 = const()[name = tensor("op_1518"), val = tensor([1, 1])]; + tensor input_99_pad_type_0 = const()[name = tensor("input_99_pad_type_0"), val = tensor("custom")]; + tensor input_99_pad_0 = const()[name = tensor("input_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180133184)))]; + tensor input_99_cast_fp16 = conv(dilations = var_1518, groups = var_1365, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = var_1516, weight = layers_3_self_attn_o_proj_loraA_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor var_1522 = const()[name = tensor("op_1522"), val = tensor([1, 1])]; + tensor var_1524 = const()[name = tensor("op_1524"), val = tensor([1, 1])]; + tensor lora_out_133_pad_type_0 = const()[name = tensor("lora_out_133_pad_type_0"), val = tensor("custom")]; + tensor lora_out_133_pad_0 = const()[name = tensor("lora_out_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_135_weight_0_to_fp16 = const()[name = tensor("lora_out_135_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180174208)))]; + tensor lora_out_135_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1524, groups = var_1365, pad = lora_out_133_pad_0, pad_type = lora_out_133_pad_type_0, strides = var_1522, weight = lora_out_135_weight_0_to_fp16, x = input_99_cast_fp16)[name = tensor("lora_out_135_cast_fp16")]; + tensor obj_49_cast_fp16 = add(x = pretrained_out_67_cast_fp16, y = lora_out_135_cast_fp16)[name = tensor("obj_49_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_1537 = const()[name = tensor("op_1537"), val = tensor([1])]; + tensor channels_mean_21_cast_fp16 = reduce_mean(axes = var_1537, keep_dims = var_1366, x = inputs_21_cast_fp16)[name = tensor("channels_mean_21_cast_fp16")]; + tensor zero_mean_21_cast_fp16 = sub(x = inputs_21_cast_fp16, y = channels_mean_21_cast_fp16)[name = tensor("zero_mean_21_cast_fp16")]; + tensor zero_mean_sq_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = zero_mean_21_cast_fp16)[name = tensor("zero_mean_sq_21_cast_fp16")]; + tensor var_1541 = const()[name = tensor("op_1541"), val = tensor([1])]; + tensor var_1542_cast_fp16 = reduce_mean(axes = var_1541, keep_dims = var_1366, x = zero_mean_sq_21_cast_fp16)[name = tensor("op_1542_cast_fp16")]; + tensor var_1543_to_fp16 = const()[name = tensor("op_1543_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1544_cast_fp16 = add(x = var_1542_cast_fp16, y = var_1543_to_fp16)[name = tensor("op_1544_cast_fp16")]; + tensor denom_21_epsilon_0 = const()[name = tensor("denom_21_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_21_cast_fp16 = rsqrt(epsilon = denom_21_epsilon_0, x = var_1544_cast_fp16)[name = tensor("denom_21_cast_fp16")]; + tensor out_21_cast_fp16 = mul(x = zero_mean_21_cast_fp16, y = denom_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor obj_51_gamma_0_to_fp16 = const()[name = tensor("obj_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180215232)))]; + tensor obj_51_beta_0_to_fp16 = const()[name = tensor("obj_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180217856)))]; + tensor obj_51_epsilon_0_to_fp16 = const()[name = tensor("obj_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_51_cast_fp16")]; + tensor var_1562 = const()[name = tensor("op_1562"), val = tensor([1, 1])]; + tensor var_1564 = const()[name = tensor("op_1564"), val = tensor([1, 1])]; + tensor pretrained_out_69_pad_type_0 = const()[name = tensor("pretrained_out_69_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_69_pad_0 = const()[name = tensor("pretrained_out_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180220480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181039744))), name = tensor("layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181039872)))]; + tensor pretrained_out_69_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1564, groups = var_1365, pad = pretrained_out_69_pad_0, pad_type = pretrained_out_69_pad_type_0, strides = var_1562, weight = layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_51_cast_fp16)[name = tensor("pretrained_out_69_cast_fp16")]; + tensor var_1568 = const()[name = tensor("op_1568"), val = tensor([1, 1])]; + tensor var_1570 = const()[name = tensor("op_1570"), val = tensor([1, 1])]; + tensor input_101_pad_type_0 = const()[name = tensor("input_101_pad_type_0"), val = tensor("custom")]; + tensor input_101_pad_0 = const()[name = tensor("input_101_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181042496)))]; + tensor input_101_cast_fp16 = conv(dilations = var_1570, groups = var_1365, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = var_1568, weight = layers_3_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor var_1574 = const()[name = tensor("op_1574"), val = tensor([1, 1])]; + tensor var_1576 = const()[name = tensor("op_1576"), val = tensor([1, 1])]; + tensor lora_out_137_pad_type_0 = const()[name = tensor("lora_out_137_pad_type_0"), val = tensor("custom")]; + tensor lora_out_137_pad_0 = const()[name = tensor("lora_out_137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_139_weight_0_to_fp16 = const()[name = tensor("lora_out_139_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181083520)))]; + tensor lora_out_139_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1576, groups = var_1365, pad = lora_out_137_pad_0, pad_type = lora_out_137_pad_type_0, strides = var_1574, weight = lora_out_139_weight_0_to_fp16, x = input_101_cast_fp16)[name = tensor("lora_out_139_cast_fp16")]; + tensor query_15_cast_fp16 = add(x = pretrained_out_69_cast_fp16, y = lora_out_139_cast_fp16)[name = tensor("query_15_cast_fp16")]; + tensor var_1586 = const()[name = tensor("op_1586"), val = tensor([1, 1])]; + tensor var_1588 = const()[name = tensor("op_1588"), val = tensor([1, 1])]; + tensor pretrained_out_71_pad_type_0 = const()[name = tensor("pretrained_out_71_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_71_pad_0 = const()[name = tensor("pretrained_out_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181124544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181943808))), name = tensor("layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_71_cast_fp16 = conv(dilations = var_1588, groups = var_1365, pad = pretrained_out_71_pad_0, pad_type = pretrained_out_71_pad_type_0, strides = var_1586, weight = layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_71_cast_fp16")]; + tensor var_1592 = const()[name = tensor("op_1592"), val = tensor([1, 1])]; + tensor var_1594 = const()[name = tensor("op_1594"), val = tensor([1, 1])]; + tensor input_103_pad_type_0 = const()[name = tensor("input_103_pad_type_0"), val = tensor("custom")]; + tensor input_103_pad_0 = const()[name = tensor("input_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181943936)))]; + tensor input_103_cast_fp16 = conv(dilations = var_1594, groups = var_1365, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = var_1592, weight = layers_3_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_103_cast_fp16")]; + tensor var_1598 = const()[name = tensor("op_1598"), val = tensor([1, 1])]; + tensor var_1600 = const()[name = tensor("op_1600"), val = tensor([1, 1])]; + tensor lora_out_141_pad_type_0 = const()[name = tensor("lora_out_141_pad_type_0"), val = tensor("custom")]; + tensor lora_out_141_pad_0 = const()[name = tensor("lora_out_141_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_143_weight_0_to_fp16 = const()[name = tensor("lora_out_143_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181984960)))]; + tensor lora_out_143_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1600, groups = var_1365, pad = lora_out_141_pad_0, pad_type = lora_out_141_pad_type_0, strides = var_1598, weight = lora_out_143_weight_0_to_fp16, x = input_103_cast_fp16)[name = tensor("lora_out_143_cast_fp16")]; + tensor key_15_cast_fp16 = add(x = pretrained_out_71_cast_fp16, y = lora_out_143_cast_fp16)[name = tensor("key_15_cast_fp16")]; + tensor var_1611 = const()[name = tensor("op_1611"), val = tensor([1, 1])]; + tensor var_1613 = const()[name = tensor("op_1613"), val = tensor([1, 1])]; + tensor pretrained_out_73_pad_type_0 = const()[name = tensor("pretrained_out_73_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_73_pad_0 = const()[name = tensor("pretrained_out_73_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182025984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182845248))), name = tensor("layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182845376)))]; + tensor pretrained_out_73_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1613, groups = var_1365, pad = pretrained_out_73_pad_0, pad_type = pretrained_out_73_pad_type_0, strides = var_1611, weight = layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_73_cast_fp16")]; + tensor var_1617 = const()[name = tensor("op_1617"), val = tensor([1, 1])]; + tensor var_1619 = const()[name = tensor("op_1619"), val = tensor([1, 1])]; + tensor input_105_pad_type_0 = const()[name = tensor("input_105_pad_type_0"), val = tensor("custom")]; + tensor input_105_pad_0 = const()[name = tensor("input_105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182848000)))]; + tensor input_105_cast_fp16 = conv(dilations = var_1619, groups = var_1365, pad = input_105_pad_0, pad_type = input_105_pad_type_0, strides = var_1617, weight = layers_3_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_105_cast_fp16")]; + tensor var_1623 = const()[name = tensor("op_1623"), val = tensor([1, 1])]; + tensor var_1625 = const()[name = tensor("op_1625"), val = tensor([1, 1])]; + tensor lora_out_145_pad_type_0 = const()[name = tensor("lora_out_145_pad_type_0"), val = tensor("custom")]; + tensor lora_out_145_pad_0 = const()[name = tensor("lora_out_145_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_147_weight_0_to_fp16 = const()[name = tensor("lora_out_147_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182889024)))]; + tensor lora_out_147_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1625, groups = var_1365, pad = lora_out_145_pad_0, pad_type = lora_out_145_pad_type_0, strides = var_1623, weight = lora_out_147_weight_0_to_fp16, x = input_105_cast_fp16)[name = tensor("lora_out_147_cast_fp16")]; + tensor value_15_cast_fp16 = add(x = pretrained_out_73_cast_fp16, y = lora_out_147_cast_fp16)[name = tensor("value_15_cast_fp16")]; + tensor var_1632 = const()[name = tensor("op_1632"), val = tensor([1, 20, 64, -1])]; + tensor var_1633_cast_fp16 = reshape(shape = var_1632, x = query_15_cast_fp16)[name = tensor("op_1633_cast_fp16")]; + tensor var_1634_to_fp16 = const()[name = tensor("op_1634_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1635_cast_fp16 = mul(x = var_1633_cast_fp16, y = var_1634_to_fp16)[name = tensor("op_1635_cast_fp16")]; + tensor var_1636 = const()[name = tensor("op_1636"), val = tensor([1, 20, 64, -1])]; + tensor var_1637_cast_fp16 = reshape(shape = var_1636, x = key_15_cast_fp16)[name = tensor("op_1637_cast_fp16")]; + tensor mh_w_23_transpose_x_0 = const()[name = tensor("mh_w_23_transpose_x_0"), val = tensor(true)]; + tensor mh_w_23_transpose_y_0 = const()[name = tensor("mh_w_23_transpose_y_0"), val = tensor(false)]; + tensor mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_1635_cast_fp16, y = var_1637_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor obj_55_cast_fp16 = softmax(axis = var_1358, x = mh_w_23_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor var_1641 = const()[name = tensor("op_1641"), val = tensor([1, 20, 64, -1])]; + tensor var_1642_cast_fp16 = reshape(shape = var_1641, x = value_15_cast_fp16)[name = tensor("op_1642_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(true)]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_1642_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_1645 = const()[name = tensor("op_1645"), val = tensor([1, 1280, 1, -1])]; + tensor input_107_cast_fp16 = reshape(shape = var_1645, x = attn_15_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor var_1652 = const()[name = tensor("op_1652"), val = tensor([1, 1])]; + tensor var_1654 = const()[name = tensor("op_1654"), val = tensor([1, 1])]; + tensor pretrained_out_75_pad_type_0 = const()[name = tensor("pretrained_out_75_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_75_pad_0 = const()[name = tensor("pretrained_out_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182930048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183749312))), name = tensor("layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183749440)))]; + tensor pretrained_out_75_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1654, groups = var_1365, pad = pretrained_out_75_pad_0, pad_type = pretrained_out_75_pad_type_0, strides = var_1652, weight = layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_107_cast_fp16)[name = tensor("pretrained_out_75_cast_fp16")]; + tensor var_1658 = const()[name = tensor("op_1658"), val = tensor([1, 1])]; + tensor var_1660 = const()[name = tensor("op_1660"), val = tensor([1, 1])]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("custom")]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183752064)))]; + tensor input_109_cast_fp16 = conv(dilations = var_1660, groups = var_1365, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = var_1658, weight = layers_3_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor var_1664 = const()[name = tensor("op_1664"), val = tensor([1, 1])]; + tensor var_1666 = const()[name = tensor("op_1666"), val = tensor([1, 1])]; + tensor lora_out_149_pad_type_0 = const()[name = tensor("lora_out_149_pad_type_0"), val = tensor("custom")]; + tensor lora_out_149_pad_0 = const()[name = tensor("lora_out_149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_151_weight_0_to_fp16 = const()[name = tensor("lora_out_151_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183793088)))]; + tensor lora_out_151_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1666, groups = var_1365, pad = lora_out_149_pad_0, pad_type = lora_out_149_pad_type_0, strides = var_1664, weight = lora_out_151_weight_0_to_fp16, x = input_109_cast_fp16)[name = tensor("lora_out_151_cast_fp16")]; + tensor obj_53_cast_fp16 = add(x = pretrained_out_75_cast_fp16, y = lora_out_151_cast_fp16)[name = tensor("obj_53_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor var_1675 = const()[name = tensor("op_1675"), val = tensor([1])]; + tensor channels_mean_23_cast_fp16 = reduce_mean(axes = var_1675, keep_dims = var_1366, x = inputs_23_cast_fp16)[name = tensor("channels_mean_23_cast_fp16")]; + tensor zero_mean_23_cast_fp16 = sub(x = inputs_23_cast_fp16, y = channels_mean_23_cast_fp16)[name = tensor("zero_mean_23_cast_fp16")]; + tensor zero_mean_sq_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = zero_mean_23_cast_fp16)[name = tensor("zero_mean_sq_23_cast_fp16")]; + tensor var_1679 = const()[name = tensor("op_1679"), val = tensor([1])]; + tensor var_1680_cast_fp16 = reduce_mean(axes = var_1679, keep_dims = var_1366, x = zero_mean_sq_23_cast_fp16)[name = tensor("op_1680_cast_fp16")]; + tensor var_1681_to_fp16 = const()[name = tensor("op_1681_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1682_cast_fp16 = add(x = var_1680_cast_fp16, y = var_1681_to_fp16)[name = tensor("op_1682_cast_fp16")]; + tensor denom_23_epsilon_0 = const()[name = tensor("denom_23_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_23_cast_fp16 = rsqrt(epsilon = denom_23_epsilon_0, x = var_1682_cast_fp16)[name = tensor("denom_23_cast_fp16")]; + tensor out_23_cast_fp16 = mul(x = zero_mean_23_cast_fp16, y = denom_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; + tensor input_111_gamma_0_to_fp16 = const()[name = tensor("input_111_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183834112)))]; + tensor input_111_beta_0_to_fp16 = const()[name = tensor("input_111_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183836736)))]; + tensor input_111_epsilon_0_to_fp16 = const()[name = tensor("input_111_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_111_cast_fp16 = batch_norm(beta = input_111_beta_0_to_fp16, epsilon = input_111_epsilon_0_to_fp16, gamma = input_111_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor var_1696 = const()[name = tensor("op_1696"), val = tensor([1, 1])]; + tensor var_1698 = const()[name = tensor("op_1698"), val = tensor([1, 1])]; + tensor pretrained_out_77_pad_type_0 = const()[name = tensor("pretrained_out_77_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_77_pad_0 = const()[name = tensor("pretrained_out_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183839360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187116224))), name = tensor("layers_3_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_3_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187116352)))]; + tensor pretrained_out_77_cast_fp16 = conv(bias = layers_3_fc1_pretrained_bias_to_fp16, dilations = var_1698, groups = var_1365, pad = pretrained_out_77_pad_0, pad_type = pretrained_out_77_pad_type_0, strides = var_1696, weight = layers_3_fc1_pretrained_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = tensor("pretrained_out_77_cast_fp16")]; + tensor var_1702 = const()[name = tensor("op_1702"), val = tensor([1, 1])]; + tensor var_1704 = const()[name = tensor("op_1704"), val = tensor([1, 1])]; + tensor input_113_pad_type_0 = const()[name = tensor("input_113_pad_type_0"), val = tensor("custom")]; + tensor input_113_pad_0 = const()[name = tensor("input_113_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_3_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187126656)))]; + tensor input_113_cast_fp16 = conv(dilations = var_1704, groups = var_1365, pad = input_113_pad_0, pad_type = input_113_pad_type_0, strides = var_1702, weight = layers_3_fc1_loraA_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor var_1708 = const()[name = tensor("op_1708"), val = tensor([1, 1])]; + tensor var_1710 = const()[name = tensor("op_1710"), val = tensor([1, 1])]; + tensor lora_out_153_pad_type_0 = const()[name = tensor("lora_out_153_pad_type_0"), val = tensor("custom")]; + tensor lora_out_153_pad_0 = const()[name = tensor("lora_out_153_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_155_weight_0_to_fp16 = const()[name = tensor("lora_out_155_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187167680)))]; + tensor lora_out_155_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_1710, groups = var_1365, pad = lora_out_153_pad_0, pad_type = lora_out_153_pad_type_0, strides = var_1708, weight = lora_out_155_weight_0_to_fp16, x = input_113_cast_fp16)[name = tensor("lora_out_155_cast_fp16")]; + tensor input_115_cast_fp16 = add(x = pretrained_out_77_cast_fp16, y = lora_out_155_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_mode_0 = const()[name = tensor("input_117_mode_0"), val = tensor("EXACT")]; + tensor input_117_cast_fp16 = gelu(mode = input_117_mode_0, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor var_1722 = const()[name = tensor("op_1722"), val = tensor([1, 1])]; + tensor var_1724 = const()[name = tensor("op_1724"), val = tensor([1, 1])]; + tensor pretrained_out_79_pad_type_0 = const()[name = tensor("pretrained_out_79_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_79_pad_0 = const()[name = tensor("pretrained_out_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187331584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190608448))), name = tensor("layers_3_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_3_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190608576)))]; + tensor pretrained_out_79_cast_fp16 = conv(bias = layers_3_fc2_pretrained_bias_to_fp16, dilations = var_1724, groups = var_1365, pad = pretrained_out_79_pad_0, pad_type = pretrained_out_79_pad_type_0, strides = var_1722, weight = layers_3_fc2_pretrained_weight_to_fp16_palettized, x = input_117_cast_fp16)[name = tensor("pretrained_out_79_cast_fp16")]; + tensor var_1728 = const()[name = tensor("op_1728"), val = tensor([1, 1])]; + tensor var_1730 = const()[name = tensor("op_1730"), val = tensor([1, 1])]; + tensor input_119_pad_type_0 = const()[name = tensor("input_119_pad_type_0"), val = tensor("custom")]; + tensor input_119_pad_0 = const()[name = tensor("input_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_3_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_3_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190611200)))]; + tensor input_119_cast_fp16 = conv(dilations = var_1730, groups = var_1365, pad = input_119_pad_0, pad_type = input_119_pad_type_0, strides = var_1728, weight = layers_3_fc2_loraA_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor var_1734 = const()[name = tensor("op_1734"), val = tensor([1, 1])]; + tensor var_1736 = const()[name = tensor("op_1736"), val = tensor([1, 1])]; + tensor lora_out_157_pad_type_0 = const()[name = tensor("lora_out_157_pad_type_0"), val = tensor("custom")]; + tensor lora_out_157_pad_0 = const()[name = tensor("lora_out_157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_159_weight_0_to_fp16 = const()[name = tensor("lora_out_159_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190775104)))]; + tensor lora_out_159_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1736, groups = var_1365, pad = lora_out_157_pad_0, pad_type = lora_out_157_pad_type_0, strides = var_1734, weight = lora_out_159_weight_0_to_fp16, x = input_119_cast_fp16)[name = tensor("lora_out_159_cast_fp16")]; + tensor hidden_states_9_cast_fp16 = add(x = pretrained_out_79_cast_fp16, y = lora_out_159_cast_fp16)[name = tensor("hidden_states_9_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_1752 = const()[name = tensor("op_1752"), val = tensor(3)]; + tensor var_1759 = const()[name = tensor("op_1759"), val = tensor(1)]; + tensor var_1760 = const()[name = tensor("op_1760"), val = tensor(true)]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1])]; + tensor channels_mean_25_cast_fp16 = reduce_mean(axes = var_1772, keep_dims = var_1760, x = inputs_25_cast_fp16)[name = tensor("channels_mean_25_cast_fp16")]; + tensor zero_mean_25_cast_fp16 = sub(x = inputs_25_cast_fp16, y = channels_mean_25_cast_fp16)[name = tensor("zero_mean_25_cast_fp16")]; + tensor zero_mean_sq_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = zero_mean_25_cast_fp16)[name = tensor("zero_mean_sq_25_cast_fp16")]; + tensor var_1776 = const()[name = tensor("op_1776"), val = tensor([1])]; + tensor var_1777_cast_fp16 = reduce_mean(axes = var_1776, keep_dims = var_1760, x = zero_mean_sq_25_cast_fp16)[name = tensor("op_1777_cast_fp16")]; + tensor var_1778_to_fp16 = const()[name = tensor("op_1778_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1779_cast_fp16 = add(x = var_1777_cast_fp16, y = var_1778_to_fp16)[name = tensor("op_1779_cast_fp16")]; + tensor denom_25_epsilon_0 = const()[name = tensor("denom_25_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_25_cast_fp16 = rsqrt(epsilon = denom_25_epsilon_0, x = var_1779_cast_fp16)[name = tensor("denom_25_cast_fp16")]; + tensor out_25_cast_fp16 = mul(x = zero_mean_25_cast_fp16, y = denom_25_cast_fp16)[name = tensor("out_25_cast_fp16")]; + tensor obj_57_gamma_0_to_fp16 = const()[name = tensor("obj_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190816128)))]; + tensor obj_57_beta_0_to_fp16 = const()[name = tensor("obj_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190818752)))]; + tensor obj_57_epsilon_0_to_fp16 = const()[name = tensor("obj_57_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor("obj_57_cast_fp16")]; + tensor var_1797 = const()[name = tensor("op_1797"), val = tensor([1, 1])]; + tensor var_1799 = const()[name = tensor("op_1799"), val = tensor([1, 1])]; + tensor pretrained_out_81_pad_type_0 = const()[name = tensor("pretrained_out_81_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_81_pad_0 = const()[name = tensor("pretrained_out_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(190821376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191640640))), name = tensor("layers_4_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_4_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191640768)))]; + tensor pretrained_out_81_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1799, groups = var_1759, pad = pretrained_out_81_pad_0, pad_type = pretrained_out_81_pad_type_0, strides = var_1797, weight = layers_4_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = tensor("pretrained_out_81_cast_fp16")]; + tensor var_1803 = const()[name = tensor("op_1803"), val = tensor([1, 1])]; + tensor var_1805 = const()[name = tensor("op_1805"), val = tensor([1, 1])]; + tensor input_121_pad_type_0 = const()[name = tensor("input_121_pad_type_0"), val = tensor("custom")]; + tensor input_121_pad_0 = const()[name = tensor("input_121_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191643392)))]; + tensor input_121_cast_fp16 = conv(dilations = var_1805, groups = var_1759, pad = input_121_pad_0, pad_type = input_121_pad_type_0, strides = var_1803, weight = layers_4_self_attn_q_proj_loraA_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor var_1809 = const()[name = tensor("op_1809"), val = tensor([1, 1])]; + tensor var_1811 = const()[name = tensor("op_1811"), val = tensor([1, 1])]; + tensor lora_out_161_pad_type_0 = const()[name = tensor("lora_out_161_pad_type_0"), val = tensor("custom")]; + tensor lora_out_161_pad_0 = const()[name = tensor("lora_out_161_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_163_weight_0_to_fp16 = const()[name = tensor("lora_out_163_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191684416)))]; + tensor lora_out_163_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1811, groups = var_1759, pad = lora_out_161_pad_0, pad_type = lora_out_161_pad_type_0, strides = var_1809, weight = lora_out_163_weight_0_to_fp16, x = input_121_cast_fp16)[name = tensor("lora_out_163_cast_fp16")]; + tensor query_17_cast_fp16 = add(x = pretrained_out_81_cast_fp16, y = lora_out_163_cast_fp16)[name = tensor("query_17_cast_fp16")]; + tensor var_1821 = const()[name = tensor("op_1821"), val = tensor([1, 1])]; + tensor var_1823 = const()[name = tensor("op_1823"), val = tensor([1, 1])]; + tensor pretrained_out_83_pad_type_0 = const()[name = tensor("pretrained_out_83_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_83_pad_0 = const()[name = tensor("pretrained_out_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191725440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192544704))), name = tensor("layers_4_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_83_cast_fp16 = conv(dilations = var_1823, groups = var_1759, pad = pretrained_out_83_pad_0, pad_type = pretrained_out_83_pad_type_0, strides = var_1821, weight = layers_4_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = tensor("pretrained_out_83_cast_fp16")]; + tensor var_1827 = const()[name = tensor("op_1827"), val = tensor([1, 1])]; + tensor var_1829 = const()[name = tensor("op_1829"), val = tensor([1, 1])]; + tensor input_123_pad_type_0 = const()[name = tensor("input_123_pad_type_0"), val = tensor("custom")]; + tensor input_123_pad_0 = const()[name = tensor("input_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192544832)))]; + tensor input_123_cast_fp16 = conv(dilations = var_1829, groups = var_1759, pad = input_123_pad_0, pad_type = input_123_pad_type_0, strides = var_1827, weight = layers_4_self_attn_k_proj_loraA_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor var_1833 = const()[name = tensor("op_1833"), val = tensor([1, 1])]; + tensor var_1835 = const()[name = tensor("op_1835"), val = tensor([1, 1])]; + tensor lora_out_165_pad_type_0 = const()[name = tensor("lora_out_165_pad_type_0"), val = tensor("custom")]; + tensor lora_out_165_pad_0 = const()[name = tensor("lora_out_165_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_167_weight_0_to_fp16 = const()[name = tensor("lora_out_167_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192585856)))]; + tensor lora_out_167_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1835, groups = var_1759, pad = lora_out_165_pad_0, pad_type = lora_out_165_pad_type_0, strides = var_1833, weight = lora_out_167_weight_0_to_fp16, x = input_123_cast_fp16)[name = tensor("lora_out_167_cast_fp16")]; + tensor current_key_9_cast_fp16 = add(x = pretrained_out_83_cast_fp16, y = lora_out_167_cast_fp16)[name = tensor("current_key_9_cast_fp16")]; + tensor var_1846 = const()[name = tensor("op_1846"), val = tensor([1, 1])]; + tensor var_1848 = const()[name = tensor("op_1848"), val = tensor([1, 1])]; + tensor pretrained_out_85_pad_type_0 = const()[name = tensor("pretrained_out_85_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_85_pad_0 = const()[name = tensor("pretrained_out_85_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192626880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193446144))), name = tensor("layers_4_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_4_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193446272)))]; + tensor pretrained_out_85_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1848, groups = var_1759, pad = pretrained_out_85_pad_0, pad_type = pretrained_out_85_pad_type_0, strides = var_1846, weight = layers_4_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = tensor("pretrained_out_85_cast_fp16")]; + tensor var_1852 = const()[name = tensor("op_1852"), val = tensor([1, 1])]; + tensor var_1854 = const()[name = tensor("op_1854"), val = tensor([1, 1])]; + tensor input_125_pad_type_0 = const()[name = tensor("input_125_pad_type_0"), val = tensor("custom")]; + tensor input_125_pad_0 = const()[name = tensor("input_125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193448896)))]; + tensor input_125_cast_fp16 = conv(dilations = var_1854, groups = var_1759, pad = input_125_pad_0, pad_type = input_125_pad_type_0, strides = var_1852, weight = layers_4_self_attn_v_proj_loraA_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor var_1858 = const()[name = tensor("op_1858"), val = tensor([1, 1])]; + tensor var_1860 = const()[name = tensor("op_1860"), val = tensor([1, 1])]; + tensor lora_out_169_pad_type_0 = const()[name = tensor("lora_out_169_pad_type_0"), val = tensor("custom")]; + tensor lora_out_169_pad_0 = const()[name = tensor("lora_out_169_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_171_weight_0_to_fp16 = const()[name = tensor("lora_out_171_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193489920)))]; + tensor lora_out_171_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1860, groups = var_1759, pad = lora_out_169_pad_0, pad_type = lora_out_169_pad_type_0, strides = var_1858, weight = lora_out_171_weight_0_to_fp16, x = input_125_cast_fp16)[name = tensor("lora_out_171_cast_fp16")]; + tensor current_value_9_cast_fp16 = add(x = pretrained_out_85_cast_fp16, y = lora_out_171_cast_fp16)[name = tensor("current_value_9_cast_fp16")]; + tensor var_1870_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor var_1872_cast_fp16 = mul(x = var_103_cast_fp16_4, y = var_295_cast_fp16)[name = tensor("op_1872_cast_fp16")]; + tensor key_17_cast_fp16 = add(x = var_1870_cast_fp16, y = var_1872_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor var_1874_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_1874_cast_fp16")]; + tensor var_1876_cast_fp16 = mul(x = var_138_cast_fp16_4, y = var_295_cast_fp16)[name = tensor("op_1876_cast_fp16")]; + tensor value_17_cast_fp16 = add(x = var_1874_cast_fp16, y = var_1876_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_1879 = const()[name = tensor("op_1879"), val = tensor([1, 20, 64, -1])]; + tensor var_1880_cast_fp16 = reshape(shape = var_1879, x = query_17_cast_fp16)[name = tensor("op_1880_cast_fp16")]; + tensor var_1881_to_fp16 = const()[name = tensor("op_1881_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1882_cast_fp16 = mul(x = var_1880_cast_fp16, y = var_1881_to_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor var_1883 = const()[name = tensor("op_1883"), val = tensor([1, 20, 64, -1])]; + tensor var_1884_cast_fp16 = reshape(shape = var_1883, x = key_17_cast_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor mh_w_25_transpose_x_0 = const()[name = tensor("mh_w_25_transpose_x_0"), val = tensor(true)]; + tensor mh_w_25_transpose_y_0 = const()[name = tensor("mh_w_25_transpose_y_0"), val = tensor(false)]; + tensor mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1882_cast_fp16, y = var_1884_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_1892_cast_fp16 = softmax(axis = var_1752, x = mh_w_27_cast_fp16)[name = tensor("op_1892_cast_fp16")]; + tensor var_1893 = const()[name = tensor("op_1893"), val = tensor([1, 20, 64, -1])]; + tensor var_1894_cast_fp16 = reshape(shape = var_1893, x = value_17_cast_fp16)[name = tensor("op_1894_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(true)]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1894_cast_fp16, y = var_1892_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_1897 = const()[name = tensor("op_1897"), val = tensor([1, 1280, 1, -1])]; + tensor input_127_cast_fp16 = reshape(shape = var_1897, x = attn_17_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor var_1904 = const()[name = tensor("op_1904"), val = tensor([1, 1])]; + tensor var_1906 = const()[name = tensor("op_1906"), val = tensor([1, 1])]; + tensor pretrained_out_87_pad_type_0 = const()[name = tensor("pretrained_out_87_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_87_pad_0 = const()[name = tensor("pretrained_out_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193530944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194350208))), name = tensor("layers_4_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_4_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194350336)))]; + tensor pretrained_out_87_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1906, groups = var_1759, pad = pretrained_out_87_pad_0, pad_type = pretrained_out_87_pad_type_0, strides = var_1904, weight = layers_4_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_127_cast_fp16)[name = tensor("pretrained_out_87_cast_fp16")]; + tensor var_1910 = const()[name = tensor("op_1910"), val = tensor([1, 1])]; + tensor var_1912 = const()[name = tensor("op_1912"), val = tensor([1, 1])]; + tensor input_129_pad_type_0 = const()[name = tensor("input_129_pad_type_0"), val = tensor("custom")]; + tensor input_129_pad_0 = const()[name = tensor("input_129_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194352960)))]; + tensor input_129_cast_fp16 = conv(dilations = var_1912, groups = var_1759, pad = input_129_pad_0, pad_type = input_129_pad_type_0, strides = var_1910, weight = layers_4_self_attn_o_proj_loraA_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1, 1])]; + tensor var_1918 = const()[name = tensor("op_1918"), val = tensor([1, 1])]; + tensor lora_out_173_pad_type_0 = const()[name = tensor("lora_out_173_pad_type_0"), val = tensor("custom")]; + tensor lora_out_173_pad_0 = const()[name = tensor("lora_out_173_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_175_weight_0_to_fp16 = const()[name = tensor("lora_out_175_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194393984)))]; + tensor lora_out_175_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1918, groups = var_1759, pad = lora_out_173_pad_0, pad_type = lora_out_173_pad_type_0, strides = var_1916, weight = lora_out_175_weight_0_to_fp16, x = input_129_cast_fp16)[name = tensor("lora_out_175_cast_fp16")]; + tensor obj_63_cast_fp16 = add(x = pretrained_out_87_cast_fp16, y = lora_out_175_cast_fp16)[name = tensor("obj_63_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_63_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor var_1931 = const()[name = tensor("op_1931"), val = tensor([1])]; + tensor channels_mean_27_cast_fp16 = reduce_mean(axes = var_1931, keep_dims = var_1760, x = inputs_27_cast_fp16)[name = tensor("channels_mean_27_cast_fp16")]; + tensor zero_mean_27_cast_fp16 = sub(x = inputs_27_cast_fp16, y = channels_mean_27_cast_fp16)[name = tensor("zero_mean_27_cast_fp16")]; + tensor zero_mean_sq_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = zero_mean_27_cast_fp16)[name = tensor("zero_mean_sq_27_cast_fp16")]; + tensor var_1935 = const()[name = tensor("op_1935"), val = tensor([1])]; + tensor var_1936_cast_fp16 = reduce_mean(axes = var_1935, keep_dims = var_1760, x = zero_mean_sq_27_cast_fp16)[name = tensor("op_1936_cast_fp16")]; + tensor var_1937_to_fp16 = const()[name = tensor("op_1937_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1938_cast_fp16 = add(x = var_1936_cast_fp16, y = var_1937_to_fp16)[name = tensor("op_1938_cast_fp16")]; + tensor denom_27_epsilon_0 = const()[name = tensor("denom_27_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_27_cast_fp16 = rsqrt(epsilon = denom_27_epsilon_0, x = var_1938_cast_fp16)[name = tensor("denom_27_cast_fp16")]; + tensor out_27_cast_fp16 = mul(x = zero_mean_27_cast_fp16, y = denom_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; + tensor obj_65_gamma_0_to_fp16 = const()[name = tensor("obj_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194435008)))]; + tensor obj_65_beta_0_to_fp16 = const()[name = tensor("obj_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194437632)))]; + tensor obj_65_epsilon_0_to_fp16 = const()[name = tensor("obj_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor("obj_65_cast_fp16")]; + tensor var_1956 = const()[name = tensor("op_1956"), val = tensor([1, 1])]; + tensor var_1958 = const()[name = tensor("op_1958"), val = tensor([1, 1])]; + tensor pretrained_out_89_pad_type_0 = const()[name = tensor("pretrained_out_89_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_89_pad_0 = const()[name = tensor("pretrained_out_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194440256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195259520))), name = tensor("layers_4_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_4_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195259648)))]; + tensor pretrained_out_89_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1958, groups = var_1759, pad = pretrained_out_89_pad_0, pad_type = pretrained_out_89_pad_type_0, strides = var_1956, weight = layers_4_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_65_cast_fp16)[name = tensor("pretrained_out_89_cast_fp16")]; + tensor var_1962 = const()[name = tensor("op_1962"), val = tensor([1, 1])]; + tensor var_1964 = const()[name = tensor("op_1964"), val = tensor([1, 1])]; + tensor input_131_pad_type_0 = const()[name = tensor("input_131_pad_type_0"), val = tensor("custom")]; + tensor input_131_pad_0 = const()[name = tensor("input_131_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195262272)))]; + tensor input_131_cast_fp16 = conv(dilations = var_1964, groups = var_1759, pad = input_131_pad_0, pad_type = input_131_pad_type_0, strides = var_1962, weight = layers_4_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor var_1968 = const()[name = tensor("op_1968"), val = tensor([1, 1])]; + tensor var_1970 = const()[name = tensor("op_1970"), val = tensor([1, 1])]; + tensor lora_out_177_pad_type_0 = const()[name = tensor("lora_out_177_pad_type_0"), val = tensor("custom")]; + tensor lora_out_177_pad_0 = const()[name = tensor("lora_out_177_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_179_weight_0_to_fp16 = const()[name = tensor("lora_out_179_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195303296)))]; + tensor lora_out_179_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1970, groups = var_1759, pad = lora_out_177_pad_0, pad_type = lora_out_177_pad_type_0, strides = var_1968, weight = lora_out_179_weight_0_to_fp16, x = input_131_cast_fp16)[name = tensor("lora_out_179_cast_fp16")]; + tensor query_19_cast_fp16 = add(x = pretrained_out_89_cast_fp16, y = lora_out_179_cast_fp16)[name = tensor("query_19_cast_fp16")]; + tensor var_1980 = const()[name = tensor("op_1980"), val = tensor([1, 1])]; + tensor var_1982 = const()[name = tensor("op_1982"), val = tensor([1, 1])]; + tensor pretrained_out_91_pad_type_0 = const()[name = tensor("pretrained_out_91_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_91_pad_0 = const()[name = tensor("pretrained_out_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195344320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196163584))), name = tensor("layers_4_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_91_cast_fp16 = conv(dilations = var_1982, groups = var_1759, pad = pretrained_out_91_pad_0, pad_type = pretrained_out_91_pad_type_0, strides = var_1980, weight = layers_4_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_91_cast_fp16")]; + tensor var_1986 = const()[name = tensor("op_1986"), val = tensor([1, 1])]; + tensor var_1988 = const()[name = tensor("op_1988"), val = tensor([1, 1])]; + tensor input_133_pad_type_0 = const()[name = tensor("input_133_pad_type_0"), val = tensor("custom")]; + tensor input_133_pad_0 = const()[name = tensor("input_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196163712)))]; + tensor input_133_cast_fp16 = conv(dilations = var_1988, groups = var_1759, pad = input_133_pad_0, pad_type = input_133_pad_type_0, strides = var_1986, weight = layers_4_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_133_cast_fp16")]; + tensor var_1992 = const()[name = tensor("op_1992"), val = tensor([1, 1])]; + tensor var_1994 = const()[name = tensor("op_1994"), val = tensor([1, 1])]; + tensor lora_out_181_pad_type_0 = const()[name = tensor("lora_out_181_pad_type_0"), val = tensor("custom")]; + tensor lora_out_181_pad_0 = const()[name = tensor("lora_out_181_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_183_weight_0_to_fp16 = const()[name = tensor("lora_out_183_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196204736)))]; + tensor lora_out_183_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1994, groups = var_1759, pad = lora_out_181_pad_0, pad_type = lora_out_181_pad_type_0, strides = var_1992, weight = lora_out_183_weight_0_to_fp16, x = input_133_cast_fp16)[name = tensor("lora_out_183_cast_fp16")]; + tensor key_19_cast_fp16 = add(x = pretrained_out_91_cast_fp16, y = lora_out_183_cast_fp16)[name = tensor("key_19_cast_fp16")]; + tensor var_2005 = const()[name = tensor("op_2005"), val = tensor([1, 1])]; + tensor var_2007 = const()[name = tensor("op_2007"), val = tensor([1, 1])]; + tensor pretrained_out_93_pad_type_0 = const()[name = tensor("pretrained_out_93_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_93_pad_0 = const()[name = tensor("pretrained_out_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196245760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197065024))), name = tensor("layers_4_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_4_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197065152)))]; + tensor pretrained_out_93_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2007, groups = var_1759, pad = pretrained_out_93_pad_0, pad_type = pretrained_out_93_pad_type_0, strides = var_2005, weight = layers_4_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_93_cast_fp16")]; + tensor var_2011 = const()[name = tensor("op_2011"), val = tensor([1, 1])]; + tensor var_2013 = const()[name = tensor("op_2013"), val = tensor([1, 1])]; + tensor input_135_pad_type_0 = const()[name = tensor("input_135_pad_type_0"), val = tensor("custom")]; + tensor input_135_pad_0 = const()[name = tensor("input_135_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197067776)))]; + tensor input_135_cast_fp16 = conv(dilations = var_2013, groups = var_1759, pad = input_135_pad_0, pad_type = input_135_pad_type_0, strides = var_2011, weight = layers_4_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_135_cast_fp16")]; + tensor var_2017 = const()[name = tensor("op_2017"), val = tensor([1, 1])]; + tensor var_2019 = const()[name = tensor("op_2019"), val = tensor([1, 1])]; + tensor lora_out_185_pad_type_0 = const()[name = tensor("lora_out_185_pad_type_0"), val = tensor("custom")]; + tensor lora_out_185_pad_0 = const()[name = tensor("lora_out_185_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_187_weight_0_to_fp16 = const()[name = tensor("lora_out_187_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197108800)))]; + tensor lora_out_187_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2019, groups = var_1759, pad = lora_out_185_pad_0, pad_type = lora_out_185_pad_type_0, strides = var_2017, weight = lora_out_187_weight_0_to_fp16, x = input_135_cast_fp16)[name = tensor("lora_out_187_cast_fp16")]; + tensor value_19_cast_fp16 = add(x = pretrained_out_93_cast_fp16, y = lora_out_187_cast_fp16)[name = tensor("value_19_cast_fp16")]; + tensor var_2026 = const()[name = tensor("op_2026"), val = tensor([1, 20, 64, -1])]; + tensor var_2027_cast_fp16 = reshape(shape = var_2026, x = query_19_cast_fp16)[name = tensor("op_2027_cast_fp16")]; + tensor var_2028_to_fp16 = const()[name = tensor("op_2028_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2029_cast_fp16 = mul(x = var_2027_cast_fp16, y = var_2028_to_fp16)[name = tensor("op_2029_cast_fp16")]; + tensor var_2030 = const()[name = tensor("op_2030"), val = tensor([1, 20, 64, -1])]; + tensor var_2031_cast_fp16 = reshape(shape = var_2030, x = key_19_cast_fp16)[name = tensor("op_2031_cast_fp16")]; + tensor mh_w_29_transpose_x_0 = const()[name = tensor("mh_w_29_transpose_x_0"), val = tensor(true)]; + tensor mh_w_29_transpose_y_0 = const()[name = tensor("mh_w_29_transpose_y_0"), val = tensor(false)]; + tensor mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_2029_cast_fp16, y = var_2031_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor obj_69_cast_fp16 = softmax(axis = var_1752, x = mh_w_29_cast_fp16)[name = tensor("obj_69_cast_fp16")]; + tensor var_2035 = const()[name = tensor("op_2035"), val = tensor([1, 20, 64, -1])]; + tensor var_2036_cast_fp16 = reshape(shape = var_2035, x = value_19_cast_fp16)[name = tensor("op_2036_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(true)]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_2036_cast_fp16, y = obj_69_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2039 = const()[name = tensor("op_2039"), val = tensor([1, 1280, 1, -1])]; + tensor input_137_cast_fp16 = reshape(shape = var_2039, x = attn_19_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor var_2046 = const()[name = tensor("op_2046"), val = tensor([1, 1])]; + tensor var_2048 = const()[name = tensor("op_2048"), val = tensor([1, 1])]; + tensor pretrained_out_95_pad_type_0 = const()[name = tensor("pretrained_out_95_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_95_pad_0 = const()[name = tensor("pretrained_out_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197149824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197969088))), name = tensor("layers_4_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_4_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197969216)))]; + tensor pretrained_out_95_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2048, groups = var_1759, pad = pretrained_out_95_pad_0, pad_type = pretrained_out_95_pad_type_0, strides = var_2046, weight = layers_4_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_137_cast_fp16)[name = tensor("pretrained_out_95_cast_fp16")]; + tensor var_2052 = const()[name = tensor("op_2052"), val = tensor([1, 1])]; + tensor var_2054 = const()[name = tensor("op_2054"), val = tensor([1, 1])]; + tensor input_139_pad_type_0 = const()[name = tensor("input_139_pad_type_0"), val = tensor("custom")]; + tensor input_139_pad_0 = const()[name = tensor("input_139_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197971840)))]; + tensor input_139_cast_fp16 = conv(dilations = var_2054, groups = var_1759, pad = input_139_pad_0, pad_type = input_139_pad_type_0, strides = var_2052, weight = layers_4_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor var_2058 = const()[name = tensor("op_2058"), val = tensor([1, 1])]; + tensor var_2060 = const()[name = tensor("op_2060"), val = tensor([1, 1])]; + tensor lora_out_189_pad_type_0 = const()[name = tensor("lora_out_189_pad_type_0"), val = tensor("custom")]; + tensor lora_out_189_pad_0 = const()[name = tensor("lora_out_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_191_weight_0_to_fp16 = const()[name = tensor("lora_out_191_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198012864)))]; + tensor lora_out_191_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2060, groups = var_1759, pad = lora_out_189_pad_0, pad_type = lora_out_189_pad_type_0, strides = var_2058, weight = lora_out_191_weight_0_to_fp16, x = input_139_cast_fp16)[name = tensor("lora_out_191_cast_fp16")]; + tensor obj_67_cast_fp16 = add(x = pretrained_out_95_cast_fp16, y = lora_out_191_cast_fp16)[name = tensor("obj_67_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_67_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_2069 = const()[name = tensor("op_2069"), val = tensor([1])]; + tensor channels_mean_29_cast_fp16 = reduce_mean(axes = var_2069, keep_dims = var_1760, x = inputs_29_cast_fp16)[name = tensor("channels_mean_29_cast_fp16")]; + tensor zero_mean_29_cast_fp16 = sub(x = inputs_29_cast_fp16, y = channels_mean_29_cast_fp16)[name = tensor("zero_mean_29_cast_fp16")]; + tensor zero_mean_sq_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = zero_mean_29_cast_fp16)[name = tensor("zero_mean_sq_29_cast_fp16")]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1])]; + tensor var_2074_cast_fp16 = reduce_mean(axes = var_2073, keep_dims = var_1760, x = zero_mean_sq_29_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor var_2075_to_fp16 = const()[name = tensor("op_2075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2076_cast_fp16 = add(x = var_2074_cast_fp16, y = var_2075_to_fp16)[name = tensor("op_2076_cast_fp16")]; + tensor denom_29_epsilon_0 = const()[name = tensor("denom_29_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_29_cast_fp16 = rsqrt(epsilon = denom_29_epsilon_0, x = var_2076_cast_fp16)[name = tensor("denom_29_cast_fp16")]; + tensor out_29_cast_fp16 = mul(x = zero_mean_29_cast_fp16, y = denom_29_cast_fp16)[name = tensor("out_29_cast_fp16")]; + tensor input_141_gamma_0_to_fp16 = const()[name = tensor("input_141_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198053888)))]; + tensor input_141_beta_0_to_fp16 = const()[name = tensor("input_141_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198056512)))]; + tensor input_141_epsilon_0_to_fp16 = const()[name = tensor("input_141_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_141_cast_fp16 = batch_norm(beta = input_141_beta_0_to_fp16, epsilon = input_141_epsilon_0_to_fp16, gamma = input_141_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor var_2090 = const()[name = tensor("op_2090"), val = tensor([1, 1])]; + tensor var_2092 = const()[name = tensor("op_2092"), val = tensor([1, 1])]; + tensor pretrained_out_97_pad_type_0 = const()[name = tensor("pretrained_out_97_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_97_pad_0 = const()[name = tensor("pretrained_out_97_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198059136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201336000))), name = tensor("layers_4_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_4_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201336128)))]; + tensor pretrained_out_97_cast_fp16 = conv(bias = layers_4_fc1_pretrained_bias_to_fp16, dilations = var_2092, groups = var_1759, pad = pretrained_out_97_pad_0, pad_type = pretrained_out_97_pad_type_0, strides = var_2090, weight = layers_4_fc1_pretrained_weight_to_fp16_palettized, x = input_141_cast_fp16)[name = tensor("pretrained_out_97_cast_fp16")]; + tensor var_2096 = const()[name = tensor("op_2096"), val = tensor([1, 1])]; + tensor var_2098 = const()[name = tensor("op_2098"), val = tensor([1, 1])]; + tensor input_143_pad_type_0 = const()[name = tensor("input_143_pad_type_0"), val = tensor("custom")]; + tensor input_143_pad_0 = const()[name = tensor("input_143_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_4_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201346432)))]; + tensor input_143_cast_fp16 = conv(dilations = var_2098, groups = var_1759, pad = input_143_pad_0, pad_type = input_143_pad_type_0, strides = var_2096, weight = layers_4_fc1_loraA_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor var_2102 = const()[name = tensor("op_2102"), val = tensor([1, 1])]; + tensor var_2104 = const()[name = tensor("op_2104"), val = tensor([1, 1])]; + tensor lora_out_193_pad_type_0 = const()[name = tensor("lora_out_193_pad_type_0"), val = tensor("custom")]; + tensor lora_out_193_pad_0 = const()[name = tensor("lora_out_193_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_195_weight_0_to_fp16 = const()[name = tensor("lora_out_195_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201387456)))]; + tensor lora_out_195_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_2104, groups = var_1759, pad = lora_out_193_pad_0, pad_type = lora_out_193_pad_type_0, strides = var_2102, weight = lora_out_195_weight_0_to_fp16, x = input_143_cast_fp16)[name = tensor("lora_out_195_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = pretrained_out_97_cast_fp16, y = lora_out_195_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_mode_0 = const()[name = tensor("input_147_mode_0"), val = tensor("EXACT")]; + tensor input_147_cast_fp16 = gelu(mode = input_147_mode_0, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor var_2116 = const()[name = tensor("op_2116"), val = tensor([1, 1])]; + tensor var_2118 = const()[name = tensor("op_2118"), val = tensor([1, 1])]; + tensor pretrained_out_99_pad_type_0 = const()[name = tensor("pretrained_out_99_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_99_pad_0 = const()[name = tensor("pretrained_out_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201551360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204828224))), name = tensor("layers_4_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_4_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_4_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204828352)))]; + tensor pretrained_out_99_cast_fp16 = conv(bias = layers_4_fc2_pretrained_bias_to_fp16, dilations = var_2118, groups = var_1759, pad = pretrained_out_99_pad_0, pad_type = pretrained_out_99_pad_type_0, strides = var_2116, weight = layers_4_fc2_pretrained_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = tensor("pretrained_out_99_cast_fp16")]; + tensor var_2122 = const()[name = tensor("op_2122"), val = tensor([1, 1])]; + tensor var_2124 = const()[name = tensor("op_2124"), val = tensor([1, 1])]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("custom")]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_4_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_4_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204830976)))]; + tensor input_149_cast_fp16 = conv(dilations = var_2124, groups = var_1759, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = var_2122, weight = layers_4_fc2_loraA_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor var_2128 = const()[name = tensor("op_2128"), val = tensor([1, 1])]; + tensor var_2130 = const()[name = tensor("op_2130"), val = tensor([1, 1])]; + tensor lora_out_197_pad_type_0 = const()[name = tensor("lora_out_197_pad_type_0"), val = tensor("custom")]; + tensor lora_out_197_pad_0 = const()[name = tensor("lora_out_197_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_199_weight_0_to_fp16 = const()[name = tensor("lora_out_199_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204994880)))]; + tensor lora_out_199_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2130, groups = var_1759, pad = lora_out_197_pad_0, pad_type = lora_out_197_pad_type_0, strides = var_2128, weight = lora_out_199_weight_0_to_fp16, x = input_149_cast_fp16)[name = tensor("lora_out_199_cast_fp16")]; + tensor hidden_states_11_cast_fp16 = add(x = pretrained_out_99_cast_fp16, y = lora_out_199_cast_fp16)[name = tensor("hidden_states_11_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor var_2146 = const()[name = tensor("op_2146"), val = tensor(3)]; + tensor var_2153 = const()[name = tensor("op_2153"), val = tensor(1)]; + tensor var_2154 = const()[name = tensor("op_2154"), val = tensor(true)]; + tensor var_2166 = const()[name = tensor("op_2166"), val = tensor([1])]; + tensor channels_mean_31_cast_fp16 = reduce_mean(axes = var_2166, keep_dims = var_2154, x = inputs_31_cast_fp16)[name = tensor("channels_mean_31_cast_fp16")]; + tensor zero_mean_31_cast_fp16 = sub(x = inputs_31_cast_fp16, y = channels_mean_31_cast_fp16)[name = tensor("zero_mean_31_cast_fp16")]; + tensor zero_mean_sq_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = zero_mean_31_cast_fp16)[name = tensor("zero_mean_sq_31_cast_fp16")]; + tensor var_2170 = const()[name = tensor("op_2170"), val = tensor([1])]; + tensor var_2171_cast_fp16 = reduce_mean(axes = var_2170, keep_dims = var_2154, x = zero_mean_sq_31_cast_fp16)[name = tensor("op_2171_cast_fp16")]; + tensor var_2172_to_fp16 = const()[name = tensor("op_2172_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2173_cast_fp16 = add(x = var_2171_cast_fp16, y = var_2172_to_fp16)[name = tensor("op_2173_cast_fp16")]; + tensor denom_31_epsilon_0 = const()[name = tensor("denom_31_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_31_cast_fp16 = rsqrt(epsilon = denom_31_epsilon_0, x = var_2173_cast_fp16)[name = tensor("denom_31_cast_fp16")]; + tensor out_31_cast_fp16 = mul(x = zero_mean_31_cast_fp16, y = denom_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; + tensor obj_71_gamma_0_to_fp16 = const()[name = tensor("obj_71_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205035904)))]; + tensor obj_71_beta_0_to_fp16 = const()[name = tensor("obj_71_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205038528)))]; + tensor obj_71_epsilon_0_to_fp16 = const()[name = tensor("obj_71_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor("obj_71_cast_fp16")]; + tensor var_2191 = const()[name = tensor("op_2191"), val = tensor([1, 1])]; + tensor var_2193 = const()[name = tensor("op_2193"), val = tensor([1, 1])]; + tensor pretrained_out_101_pad_type_0 = const()[name = tensor("pretrained_out_101_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_101_pad_0 = const()[name = tensor("pretrained_out_101_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205041152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205860416))), name = tensor("layers_5_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_5_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205860544)))]; + tensor pretrained_out_101_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2193, groups = var_2153, pad = pretrained_out_101_pad_0, pad_type = pretrained_out_101_pad_type_0, strides = var_2191, weight = layers_5_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_71_cast_fp16)[name = tensor("pretrained_out_101_cast_fp16")]; + tensor var_2197 = const()[name = tensor("op_2197"), val = tensor([1, 1])]; + tensor var_2199 = const()[name = tensor("op_2199"), val = tensor([1, 1])]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("custom")]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205863168)))]; + tensor input_151_cast_fp16 = conv(dilations = var_2199, groups = var_2153, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = var_2197, weight = layers_5_self_attn_q_proj_loraA_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor var_2203 = const()[name = tensor("op_2203"), val = tensor([1, 1])]; + tensor var_2205 = const()[name = tensor("op_2205"), val = tensor([1, 1])]; + tensor lora_out_201_pad_type_0 = const()[name = tensor("lora_out_201_pad_type_0"), val = tensor("custom")]; + tensor lora_out_201_pad_0 = const()[name = tensor("lora_out_201_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_203_weight_0_to_fp16 = const()[name = tensor("lora_out_203_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205904192)))]; + tensor lora_out_203_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2205, groups = var_2153, pad = lora_out_201_pad_0, pad_type = lora_out_201_pad_type_0, strides = var_2203, weight = lora_out_203_weight_0_to_fp16, x = input_151_cast_fp16)[name = tensor("lora_out_203_cast_fp16")]; + tensor query_21_cast_fp16 = add(x = pretrained_out_101_cast_fp16, y = lora_out_203_cast_fp16)[name = tensor("query_21_cast_fp16")]; + tensor var_2215 = const()[name = tensor("op_2215"), val = tensor([1, 1])]; + tensor var_2217 = const()[name = tensor("op_2217"), val = tensor([1, 1])]; + tensor pretrained_out_103_pad_type_0 = const()[name = tensor("pretrained_out_103_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_103_pad_0 = const()[name = tensor("pretrained_out_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205945216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206764480))), name = tensor("layers_5_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_103_cast_fp16 = conv(dilations = var_2217, groups = var_2153, pad = pretrained_out_103_pad_0, pad_type = pretrained_out_103_pad_type_0, strides = var_2215, weight = layers_5_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_71_cast_fp16)[name = tensor("pretrained_out_103_cast_fp16")]; + tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1])]; + tensor var_2223 = const()[name = tensor("op_2223"), val = tensor([1, 1])]; + tensor input_153_pad_type_0 = const()[name = tensor("input_153_pad_type_0"), val = tensor("custom")]; + tensor input_153_pad_0 = const()[name = tensor("input_153_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206764608)))]; + tensor input_153_cast_fp16 = conv(dilations = var_2223, groups = var_2153, pad = input_153_pad_0, pad_type = input_153_pad_type_0, strides = var_2221, weight = layers_5_self_attn_k_proj_loraA_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor var_2227 = const()[name = tensor("op_2227"), val = tensor([1, 1])]; + tensor var_2229 = const()[name = tensor("op_2229"), val = tensor([1, 1])]; + tensor lora_out_205_pad_type_0 = const()[name = tensor("lora_out_205_pad_type_0"), val = tensor("custom")]; + tensor lora_out_205_pad_0 = const()[name = tensor("lora_out_205_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_207_weight_0_to_fp16 = const()[name = tensor("lora_out_207_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206805632)))]; + tensor lora_out_207_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2229, groups = var_2153, pad = lora_out_205_pad_0, pad_type = lora_out_205_pad_type_0, strides = var_2227, weight = lora_out_207_weight_0_to_fp16, x = input_153_cast_fp16)[name = tensor("lora_out_207_cast_fp16")]; + tensor current_key_11_cast_fp16 = add(x = pretrained_out_103_cast_fp16, y = lora_out_207_cast_fp16)[name = tensor("current_key_11_cast_fp16")]; + tensor var_2240 = const()[name = tensor("op_2240"), val = tensor([1, 1])]; + tensor var_2242 = const()[name = tensor("op_2242"), val = tensor([1, 1])]; + tensor pretrained_out_105_pad_type_0 = const()[name = tensor("pretrained_out_105_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_105_pad_0 = const()[name = tensor("pretrained_out_105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206846656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207665920))), name = tensor("layers_5_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_5_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207666048)))]; + tensor pretrained_out_105_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2242, groups = var_2153, pad = pretrained_out_105_pad_0, pad_type = pretrained_out_105_pad_type_0, strides = var_2240, weight = layers_5_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_71_cast_fp16)[name = tensor("pretrained_out_105_cast_fp16")]; + tensor var_2246 = const()[name = tensor("op_2246"), val = tensor([1, 1])]; + tensor var_2248 = const()[name = tensor("op_2248"), val = tensor([1, 1])]; + tensor input_155_pad_type_0 = const()[name = tensor("input_155_pad_type_0"), val = tensor("custom")]; + tensor input_155_pad_0 = const()[name = tensor("input_155_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207668672)))]; + tensor input_155_cast_fp16 = conv(dilations = var_2248, groups = var_2153, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = var_2246, weight = layers_5_self_attn_v_proj_loraA_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor var_2252 = const()[name = tensor("op_2252"), val = tensor([1, 1])]; + tensor var_2254 = const()[name = tensor("op_2254"), val = tensor([1, 1])]; + tensor lora_out_209_pad_type_0 = const()[name = tensor("lora_out_209_pad_type_0"), val = tensor("custom")]; + tensor lora_out_209_pad_0 = const()[name = tensor("lora_out_209_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_211_weight_0_to_fp16 = const()[name = tensor("lora_out_211_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207709696)))]; + tensor lora_out_211_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2254, groups = var_2153, pad = lora_out_209_pad_0, pad_type = lora_out_209_pad_type_0, strides = var_2252, weight = lora_out_211_weight_0_to_fp16, x = input_155_cast_fp16)[name = tensor("lora_out_211_cast_fp16")]; + tensor current_value_11_cast_fp16 = add(x = pretrained_out_105_cast_fp16, y = lora_out_211_cast_fp16)[name = tensor("current_value_11_cast_fp16")]; + tensor var_2264_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_2264_cast_fp16")]; + tensor var_2266_cast_fp16 = mul(x = var_103_cast_fp16_5, y = var_295_cast_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor key_21_cast_fp16 = add(x = var_2264_cast_fp16, y = var_2266_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor var_2268_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_2268_cast_fp16")]; + tensor var_2270_cast_fp16 = mul(x = var_138_cast_fp16_5, y = var_295_cast_fp16)[name = tensor("op_2270_cast_fp16")]; + tensor value_21_cast_fp16 = add(x = var_2268_cast_fp16, y = var_2270_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_2273 = const()[name = tensor("op_2273"), val = tensor([1, 20, 64, -1])]; + tensor var_2274_cast_fp16 = reshape(shape = var_2273, x = query_21_cast_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_to_fp16 = const()[name = tensor("op_2275_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2276_cast_fp16 = mul(x = var_2274_cast_fp16, y = var_2275_to_fp16)[name = tensor("op_2276_cast_fp16")]; + tensor var_2277 = const()[name = tensor("op_2277"), val = tensor([1, 20, 64, -1])]; + tensor var_2278_cast_fp16 = reshape(shape = var_2277, x = key_21_cast_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor mh_w_31_transpose_x_0 = const()[name = tensor("mh_w_31_transpose_x_0"), val = tensor(true)]; + tensor mh_w_31_transpose_y_0 = const()[name = tensor("mh_w_31_transpose_y_0"), val = tensor(false)]; + tensor mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_2276_cast_fp16, y = var_2278_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor var_2286_cast_fp16 = softmax(axis = var_2146, x = mh_w_33_cast_fp16)[name = tensor("op_2286_cast_fp16")]; + tensor var_2287 = const()[name = tensor("op_2287"), val = tensor([1, 20, 64, -1])]; + tensor var_2288_cast_fp16 = reshape(shape = var_2287, x = value_21_cast_fp16)[name = tensor("op_2288_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(true)]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_2288_cast_fp16, y = var_2286_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_2291 = const()[name = tensor("op_2291"), val = tensor([1, 1280, 1, -1])]; + tensor input_157_cast_fp16 = reshape(shape = var_2291, x = attn_21_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor var_2298 = const()[name = tensor("op_2298"), val = tensor([1, 1])]; + tensor var_2300 = const()[name = tensor("op_2300"), val = tensor([1, 1])]; + tensor pretrained_out_107_pad_type_0 = const()[name = tensor("pretrained_out_107_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_107_pad_0 = const()[name = tensor("pretrained_out_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207750720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208569984))), name = tensor("layers_5_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_5_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208570112)))]; + tensor pretrained_out_107_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2300, groups = var_2153, pad = pretrained_out_107_pad_0, pad_type = pretrained_out_107_pad_type_0, strides = var_2298, weight = layers_5_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_157_cast_fp16)[name = tensor("pretrained_out_107_cast_fp16")]; + tensor var_2304 = const()[name = tensor("op_2304"), val = tensor([1, 1])]; + tensor var_2306 = const()[name = tensor("op_2306"), val = tensor([1, 1])]; + tensor input_159_pad_type_0 = const()[name = tensor("input_159_pad_type_0"), val = tensor("custom")]; + tensor input_159_pad_0 = const()[name = tensor("input_159_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208572736)))]; + tensor input_159_cast_fp16 = conv(dilations = var_2306, groups = var_2153, pad = input_159_pad_0, pad_type = input_159_pad_type_0, strides = var_2304, weight = layers_5_self_attn_o_proj_loraA_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor var_2310 = const()[name = tensor("op_2310"), val = tensor([1, 1])]; + tensor var_2312 = const()[name = tensor("op_2312"), val = tensor([1, 1])]; + tensor lora_out_213_pad_type_0 = const()[name = tensor("lora_out_213_pad_type_0"), val = tensor("custom")]; + tensor lora_out_213_pad_0 = const()[name = tensor("lora_out_213_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_215_weight_0_to_fp16 = const()[name = tensor("lora_out_215_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208613760)))]; + tensor lora_out_215_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2312, groups = var_2153, pad = lora_out_213_pad_0, pad_type = lora_out_213_pad_type_0, strides = var_2310, weight = lora_out_215_weight_0_to_fp16, x = input_159_cast_fp16)[name = tensor("lora_out_215_cast_fp16")]; + tensor obj_77_cast_fp16 = add(x = pretrained_out_107_cast_fp16, y = lora_out_215_cast_fp16)[name = tensor("obj_77_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_77_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_2325 = const()[name = tensor("op_2325"), val = tensor([1])]; + tensor channels_mean_33_cast_fp16 = reduce_mean(axes = var_2325, keep_dims = var_2154, x = inputs_33_cast_fp16)[name = tensor("channels_mean_33_cast_fp16")]; + tensor zero_mean_33_cast_fp16 = sub(x = inputs_33_cast_fp16, y = channels_mean_33_cast_fp16)[name = tensor("zero_mean_33_cast_fp16")]; + tensor zero_mean_sq_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = zero_mean_33_cast_fp16)[name = tensor("zero_mean_sq_33_cast_fp16")]; + tensor var_2329 = const()[name = tensor("op_2329"), val = tensor([1])]; + tensor var_2330_cast_fp16 = reduce_mean(axes = var_2329, keep_dims = var_2154, x = zero_mean_sq_33_cast_fp16)[name = tensor("op_2330_cast_fp16")]; + tensor var_2331_to_fp16 = const()[name = tensor("op_2331_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2332_cast_fp16 = add(x = var_2330_cast_fp16, y = var_2331_to_fp16)[name = tensor("op_2332_cast_fp16")]; + tensor denom_33_epsilon_0 = const()[name = tensor("denom_33_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_33_cast_fp16 = rsqrt(epsilon = denom_33_epsilon_0, x = var_2332_cast_fp16)[name = tensor("denom_33_cast_fp16")]; + tensor out_33_cast_fp16 = mul(x = zero_mean_33_cast_fp16, y = denom_33_cast_fp16)[name = tensor("out_33_cast_fp16")]; + tensor obj_79_gamma_0_to_fp16 = const()[name = tensor("obj_79_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208654784)))]; + tensor obj_79_beta_0_to_fp16 = const()[name = tensor("obj_79_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208657408)))]; + tensor obj_79_epsilon_0_to_fp16 = const()[name = tensor("obj_79_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor("obj_79_cast_fp16")]; + tensor var_2350 = const()[name = tensor("op_2350"), val = tensor([1, 1])]; + tensor var_2352 = const()[name = tensor("op_2352"), val = tensor([1, 1])]; + tensor pretrained_out_109_pad_type_0 = const()[name = tensor("pretrained_out_109_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_109_pad_0 = const()[name = tensor("pretrained_out_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208660032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209479296))), name = tensor("layers_5_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_5_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209479424)))]; + tensor pretrained_out_109_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2352, groups = var_2153, pad = pretrained_out_109_pad_0, pad_type = pretrained_out_109_pad_type_0, strides = var_2350, weight = layers_5_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_79_cast_fp16)[name = tensor("pretrained_out_109_cast_fp16")]; + tensor var_2356 = const()[name = tensor("op_2356"), val = tensor([1, 1])]; + tensor var_2358 = const()[name = tensor("op_2358"), val = tensor([1, 1])]; + tensor input_161_pad_type_0 = const()[name = tensor("input_161_pad_type_0"), val = tensor("custom")]; + tensor input_161_pad_0 = const()[name = tensor("input_161_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209482048)))]; + tensor input_161_cast_fp16 = conv(dilations = var_2358, groups = var_2153, pad = input_161_pad_0, pad_type = input_161_pad_type_0, strides = var_2356, weight = layers_5_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor var_2362 = const()[name = tensor("op_2362"), val = tensor([1, 1])]; + tensor var_2364 = const()[name = tensor("op_2364"), val = tensor([1, 1])]; + tensor lora_out_217_pad_type_0 = const()[name = tensor("lora_out_217_pad_type_0"), val = tensor("custom")]; + tensor lora_out_217_pad_0 = const()[name = tensor("lora_out_217_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_219_weight_0_to_fp16 = const()[name = tensor("lora_out_219_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209523072)))]; + tensor lora_out_219_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2364, groups = var_2153, pad = lora_out_217_pad_0, pad_type = lora_out_217_pad_type_0, strides = var_2362, weight = lora_out_219_weight_0_to_fp16, x = input_161_cast_fp16)[name = tensor("lora_out_219_cast_fp16")]; + tensor query_23_cast_fp16 = add(x = pretrained_out_109_cast_fp16, y = lora_out_219_cast_fp16)[name = tensor("query_23_cast_fp16")]; + tensor var_2374 = const()[name = tensor("op_2374"), val = tensor([1, 1])]; + tensor var_2376 = const()[name = tensor("op_2376"), val = tensor([1, 1])]; + tensor pretrained_out_111_pad_type_0 = const()[name = tensor("pretrained_out_111_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_111_pad_0 = const()[name = tensor("pretrained_out_111_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209564096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210383360))), name = tensor("layers_5_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_111_cast_fp16 = conv(dilations = var_2376, groups = var_2153, pad = pretrained_out_111_pad_0, pad_type = pretrained_out_111_pad_type_0, strides = var_2374, weight = layers_5_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_111_cast_fp16")]; + tensor var_2380 = const()[name = tensor("op_2380"), val = tensor([1, 1])]; + tensor var_2382 = const()[name = tensor("op_2382"), val = tensor([1, 1])]; + tensor input_163_pad_type_0 = const()[name = tensor("input_163_pad_type_0"), val = tensor("custom")]; + tensor input_163_pad_0 = const()[name = tensor("input_163_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210383488)))]; + tensor input_163_cast_fp16 = conv(dilations = var_2382, groups = var_2153, pad = input_163_pad_0, pad_type = input_163_pad_type_0, strides = var_2380, weight = layers_5_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_163_cast_fp16")]; + tensor var_2386 = const()[name = tensor("op_2386"), val = tensor([1, 1])]; + tensor var_2388 = const()[name = tensor("op_2388"), val = tensor([1, 1])]; + tensor lora_out_221_pad_type_0 = const()[name = tensor("lora_out_221_pad_type_0"), val = tensor("custom")]; + tensor lora_out_221_pad_0 = const()[name = tensor("lora_out_221_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_223_weight_0_to_fp16 = const()[name = tensor("lora_out_223_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210424512)))]; + tensor lora_out_223_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2388, groups = var_2153, pad = lora_out_221_pad_0, pad_type = lora_out_221_pad_type_0, strides = var_2386, weight = lora_out_223_weight_0_to_fp16, x = input_163_cast_fp16)[name = tensor("lora_out_223_cast_fp16")]; + tensor key_23_cast_fp16 = add(x = pretrained_out_111_cast_fp16, y = lora_out_223_cast_fp16)[name = tensor("key_23_cast_fp16")]; + tensor var_2399 = const()[name = tensor("op_2399"), val = tensor([1, 1])]; + tensor var_2401 = const()[name = tensor("op_2401"), val = tensor([1, 1])]; + tensor pretrained_out_113_pad_type_0 = const()[name = tensor("pretrained_out_113_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_113_pad_0 = const()[name = tensor("pretrained_out_113_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210465536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211284800))), name = tensor("layers_5_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_5_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211284928)))]; + tensor pretrained_out_113_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2401, groups = var_2153, pad = pretrained_out_113_pad_0, pad_type = pretrained_out_113_pad_type_0, strides = var_2399, weight = layers_5_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_113_cast_fp16")]; + tensor var_2405 = const()[name = tensor("op_2405"), val = tensor([1, 1])]; + tensor var_2407 = const()[name = tensor("op_2407"), val = tensor([1, 1])]; + tensor input_165_pad_type_0 = const()[name = tensor("input_165_pad_type_0"), val = tensor("custom")]; + tensor input_165_pad_0 = const()[name = tensor("input_165_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211287552)))]; + tensor input_165_cast_fp16 = conv(dilations = var_2407, groups = var_2153, pad = input_165_pad_0, pad_type = input_165_pad_type_0, strides = var_2405, weight = layers_5_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_165_cast_fp16")]; + tensor var_2411 = const()[name = tensor("op_2411"), val = tensor([1, 1])]; + tensor var_2413 = const()[name = tensor("op_2413"), val = tensor([1, 1])]; + tensor lora_out_225_pad_type_0 = const()[name = tensor("lora_out_225_pad_type_0"), val = tensor("custom")]; + tensor lora_out_225_pad_0 = const()[name = tensor("lora_out_225_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_227_weight_0_to_fp16 = const()[name = tensor("lora_out_227_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211328576)))]; + tensor lora_out_227_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2413, groups = var_2153, pad = lora_out_225_pad_0, pad_type = lora_out_225_pad_type_0, strides = var_2411, weight = lora_out_227_weight_0_to_fp16, x = input_165_cast_fp16)[name = tensor("lora_out_227_cast_fp16")]; + tensor value_23_cast_fp16 = add(x = pretrained_out_113_cast_fp16, y = lora_out_227_cast_fp16)[name = tensor("value_23_cast_fp16")]; + tensor var_2420 = const()[name = tensor("op_2420"), val = tensor([1, 20, 64, -1])]; + tensor var_2421_cast_fp16 = reshape(shape = var_2420, x = query_23_cast_fp16)[name = tensor("op_2421_cast_fp16")]; + tensor var_2422_to_fp16 = const()[name = tensor("op_2422_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2423_cast_fp16 = mul(x = var_2421_cast_fp16, y = var_2422_to_fp16)[name = tensor("op_2423_cast_fp16")]; + tensor var_2424 = const()[name = tensor("op_2424"), val = tensor([1, 20, 64, -1])]; + tensor var_2425_cast_fp16 = reshape(shape = var_2424, x = key_23_cast_fp16)[name = tensor("op_2425_cast_fp16")]; + tensor mh_w_35_transpose_x_0 = const()[name = tensor("mh_w_35_transpose_x_0"), val = tensor(true)]; + tensor mh_w_35_transpose_y_0 = const()[name = tensor("mh_w_35_transpose_y_0"), val = tensor(false)]; + tensor mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_2423_cast_fp16, y = var_2425_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; + tensor obj_83_cast_fp16 = softmax(axis = var_2146, x = mh_w_35_cast_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor var_2429 = const()[name = tensor("op_2429"), val = tensor([1, 20, 64, -1])]; + tensor var_2430_cast_fp16 = reshape(shape = var_2429, x = value_23_cast_fp16)[name = tensor("op_2430_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(true)]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_2430_cast_fp16, y = obj_83_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_2433 = const()[name = tensor("op_2433"), val = tensor([1, 1280, 1, -1])]; + tensor input_167_cast_fp16 = reshape(shape = var_2433, x = attn_23_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor var_2440 = const()[name = tensor("op_2440"), val = tensor([1, 1])]; + tensor var_2442 = const()[name = tensor("op_2442"), val = tensor([1, 1])]; + tensor pretrained_out_115_pad_type_0 = const()[name = tensor("pretrained_out_115_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_115_pad_0 = const()[name = tensor("pretrained_out_115_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211369600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212188864))), name = tensor("layers_5_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_5_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212188992)))]; + tensor pretrained_out_115_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2442, groups = var_2153, pad = pretrained_out_115_pad_0, pad_type = pretrained_out_115_pad_type_0, strides = var_2440, weight = layers_5_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = tensor("pretrained_out_115_cast_fp16")]; + tensor var_2446 = const()[name = tensor("op_2446"), val = tensor([1, 1])]; + tensor var_2448 = const()[name = tensor("op_2448"), val = tensor([1, 1])]; + tensor input_169_pad_type_0 = const()[name = tensor("input_169_pad_type_0"), val = tensor("custom")]; + tensor input_169_pad_0 = const()[name = tensor("input_169_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212191616)))]; + tensor input_169_cast_fp16 = conv(dilations = var_2448, groups = var_2153, pad = input_169_pad_0, pad_type = input_169_pad_type_0, strides = var_2446, weight = layers_5_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor var_2452 = const()[name = tensor("op_2452"), val = tensor([1, 1])]; + tensor var_2454 = const()[name = tensor("op_2454"), val = tensor([1, 1])]; + tensor lora_out_229_pad_type_0 = const()[name = tensor("lora_out_229_pad_type_0"), val = tensor("custom")]; + tensor lora_out_229_pad_0 = const()[name = tensor("lora_out_229_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_231_weight_0_to_fp16 = const()[name = tensor("lora_out_231_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212232640)))]; + tensor lora_out_231_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2454, groups = var_2153, pad = lora_out_229_pad_0, pad_type = lora_out_229_pad_type_0, strides = var_2452, weight = lora_out_231_weight_0_to_fp16, x = input_169_cast_fp16)[name = tensor("lora_out_231_cast_fp16")]; + tensor obj_81_cast_fp16 = add(x = pretrained_out_115_cast_fp16, y = lora_out_231_cast_fp16)[name = tensor("obj_81_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_81_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor var_2463 = const()[name = tensor("op_2463"), val = tensor([1])]; + tensor channels_mean_35_cast_fp16 = reduce_mean(axes = var_2463, keep_dims = var_2154, x = inputs_35_cast_fp16)[name = tensor("channels_mean_35_cast_fp16")]; + tensor zero_mean_35_cast_fp16 = sub(x = inputs_35_cast_fp16, y = channels_mean_35_cast_fp16)[name = tensor("zero_mean_35_cast_fp16")]; + tensor zero_mean_sq_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = zero_mean_35_cast_fp16)[name = tensor("zero_mean_sq_35_cast_fp16")]; + tensor var_2467 = const()[name = tensor("op_2467"), val = tensor([1])]; + tensor var_2468_cast_fp16 = reduce_mean(axes = var_2467, keep_dims = var_2154, x = zero_mean_sq_35_cast_fp16)[name = tensor("op_2468_cast_fp16")]; + tensor var_2469_to_fp16 = const()[name = tensor("op_2469_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2470_cast_fp16 = add(x = var_2468_cast_fp16, y = var_2469_to_fp16)[name = tensor("op_2470_cast_fp16")]; + tensor denom_35_epsilon_0 = const()[name = tensor("denom_35_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_35_cast_fp16 = rsqrt(epsilon = denom_35_epsilon_0, x = var_2470_cast_fp16)[name = tensor("denom_35_cast_fp16")]; + tensor out_35_cast_fp16 = mul(x = zero_mean_35_cast_fp16, y = denom_35_cast_fp16)[name = tensor("out_35_cast_fp16")]; + tensor input_171_gamma_0_to_fp16 = const()[name = tensor("input_171_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212273664)))]; + tensor input_171_beta_0_to_fp16 = const()[name = tensor("input_171_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212276288)))]; + tensor input_171_epsilon_0_to_fp16 = const()[name = tensor("input_171_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = batch_norm(beta = input_171_beta_0_to_fp16, epsilon = input_171_epsilon_0_to_fp16, gamma = input_171_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor var_2484 = const()[name = tensor("op_2484"), val = tensor([1, 1])]; + tensor var_2486 = const()[name = tensor("op_2486"), val = tensor([1, 1])]; + tensor pretrained_out_117_pad_type_0 = const()[name = tensor("pretrained_out_117_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_117_pad_0 = const()[name = tensor("pretrained_out_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212278912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215555776))), name = tensor("layers_5_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_5_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215555904)))]; + tensor pretrained_out_117_cast_fp16 = conv(bias = layers_5_fc1_pretrained_bias_to_fp16, dilations = var_2486, groups = var_2153, pad = pretrained_out_117_pad_0, pad_type = pretrained_out_117_pad_type_0, strides = var_2484, weight = layers_5_fc1_pretrained_weight_to_fp16_palettized, x = input_171_cast_fp16)[name = tensor("pretrained_out_117_cast_fp16")]; + tensor var_2490 = const()[name = tensor("op_2490"), val = tensor([1, 1])]; + tensor var_2492 = const()[name = tensor("op_2492"), val = tensor([1, 1])]; + tensor input_173_pad_type_0 = const()[name = tensor("input_173_pad_type_0"), val = tensor("custom")]; + tensor input_173_pad_0 = const()[name = tensor("input_173_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_5_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215566208)))]; + tensor input_173_cast_fp16 = conv(dilations = var_2492, groups = var_2153, pad = input_173_pad_0, pad_type = input_173_pad_type_0, strides = var_2490, weight = layers_5_fc1_loraA_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor var_2496 = const()[name = tensor("op_2496"), val = tensor([1, 1])]; + tensor var_2498 = const()[name = tensor("op_2498"), val = tensor([1, 1])]; + tensor lora_out_233_pad_type_0 = const()[name = tensor("lora_out_233_pad_type_0"), val = tensor("custom")]; + tensor lora_out_233_pad_0 = const()[name = tensor("lora_out_233_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_235_weight_0_to_fp16 = const()[name = tensor("lora_out_235_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215607232)))]; + tensor lora_out_235_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_2498, groups = var_2153, pad = lora_out_233_pad_0, pad_type = lora_out_233_pad_type_0, strides = var_2496, weight = lora_out_235_weight_0_to_fp16, x = input_173_cast_fp16)[name = tensor("lora_out_235_cast_fp16")]; + tensor input_175_cast_fp16 = add(x = pretrained_out_117_cast_fp16, y = lora_out_235_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_mode_0 = const()[name = tensor("input_177_mode_0"), val = tensor("EXACT")]; + tensor input_177_cast_fp16 = gelu(mode = input_177_mode_0, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor var_2510 = const()[name = tensor("op_2510"), val = tensor([1, 1])]; + tensor var_2512 = const()[name = tensor("op_2512"), val = tensor([1, 1])]; + tensor pretrained_out_119_pad_type_0 = const()[name = tensor("pretrained_out_119_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_119_pad_0 = const()[name = tensor("pretrained_out_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215771136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219048000))), name = tensor("layers_5_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_5_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_5_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219048128)))]; + tensor pretrained_out_119_cast_fp16 = conv(bias = layers_5_fc2_pretrained_bias_to_fp16, dilations = var_2512, groups = var_2153, pad = pretrained_out_119_pad_0, pad_type = pretrained_out_119_pad_type_0, strides = var_2510, weight = layers_5_fc2_pretrained_weight_to_fp16_palettized, x = input_177_cast_fp16)[name = tensor("pretrained_out_119_cast_fp16")]; + tensor var_2516 = const()[name = tensor("op_2516"), val = tensor([1, 1])]; + tensor var_2518 = const()[name = tensor("op_2518"), val = tensor([1, 1])]; + tensor input_179_pad_type_0 = const()[name = tensor("input_179_pad_type_0"), val = tensor("custom")]; + tensor input_179_pad_0 = const()[name = tensor("input_179_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_5_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_5_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219050752)))]; + tensor input_179_cast_fp16 = conv(dilations = var_2518, groups = var_2153, pad = input_179_pad_0, pad_type = input_179_pad_type_0, strides = var_2516, weight = layers_5_fc2_loraA_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor var_2522 = const()[name = tensor("op_2522"), val = tensor([1, 1])]; + tensor var_2524 = const()[name = tensor("op_2524"), val = tensor([1, 1])]; + tensor lora_out_237_pad_type_0 = const()[name = tensor("lora_out_237_pad_type_0"), val = tensor("custom")]; + tensor lora_out_237_pad_0 = const()[name = tensor("lora_out_237_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_239_weight_0_to_fp16 = const()[name = tensor("lora_out_239_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219214656)))]; + tensor lora_out_239_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2524, groups = var_2153, pad = lora_out_237_pad_0, pad_type = lora_out_237_pad_type_0, strides = var_2522, weight = lora_out_239_weight_0_to_fp16, x = input_179_cast_fp16)[name = tensor("lora_out_239_cast_fp16")]; + tensor hidden_states_13_cast_fp16 = add(x = pretrained_out_119_cast_fp16, y = lora_out_239_cast_fp16)[name = tensor("hidden_states_13_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_2540 = const()[name = tensor("op_2540"), val = tensor(3)]; + tensor var_2547 = const()[name = tensor("op_2547"), val = tensor(1)]; + tensor var_2548 = const()[name = tensor("op_2548"), val = tensor(true)]; + tensor var_2560 = const()[name = tensor("op_2560"), val = tensor([1])]; + tensor channels_mean_37_cast_fp16 = reduce_mean(axes = var_2560, keep_dims = var_2548, x = inputs_37_cast_fp16)[name = tensor("channels_mean_37_cast_fp16")]; + tensor zero_mean_37_cast_fp16 = sub(x = inputs_37_cast_fp16, y = channels_mean_37_cast_fp16)[name = tensor("zero_mean_37_cast_fp16")]; + tensor zero_mean_sq_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = zero_mean_37_cast_fp16)[name = tensor("zero_mean_sq_37_cast_fp16")]; + tensor var_2564 = const()[name = tensor("op_2564"), val = tensor([1])]; + tensor var_2565_cast_fp16 = reduce_mean(axes = var_2564, keep_dims = var_2548, x = zero_mean_sq_37_cast_fp16)[name = tensor("op_2565_cast_fp16")]; + tensor var_2566_to_fp16 = const()[name = tensor("op_2566_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2567_cast_fp16 = add(x = var_2565_cast_fp16, y = var_2566_to_fp16)[name = tensor("op_2567_cast_fp16")]; + tensor denom_37_epsilon_0 = const()[name = tensor("denom_37_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_37_cast_fp16 = rsqrt(epsilon = denom_37_epsilon_0, x = var_2567_cast_fp16)[name = tensor("denom_37_cast_fp16")]; + tensor out_37_cast_fp16 = mul(x = zero_mean_37_cast_fp16, y = denom_37_cast_fp16)[name = tensor("out_37_cast_fp16")]; + tensor obj_85_gamma_0_to_fp16 = const()[name = tensor("obj_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219255680)))]; + tensor obj_85_beta_0_to_fp16 = const()[name = tensor("obj_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219258304)))]; + tensor obj_85_epsilon_0_to_fp16 = const()[name = tensor("obj_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor("obj_85_cast_fp16")]; + tensor var_2585 = const()[name = tensor("op_2585"), val = tensor([1, 1])]; + tensor var_2587 = const()[name = tensor("op_2587"), val = tensor([1, 1])]; + tensor pretrained_out_121_pad_type_0 = const()[name = tensor("pretrained_out_121_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_121_pad_0 = const()[name = tensor("pretrained_out_121_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219260928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220080192))), name = tensor("layers_6_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_6_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220080320)))]; + tensor pretrained_out_121_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2587, groups = var_2547, pad = pretrained_out_121_pad_0, pad_type = pretrained_out_121_pad_type_0, strides = var_2585, weight = layers_6_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor("pretrained_out_121_cast_fp16")]; + tensor var_2591 = const()[name = tensor("op_2591"), val = tensor([1, 1])]; + tensor var_2593 = const()[name = tensor("op_2593"), val = tensor([1, 1])]; + tensor input_181_pad_type_0 = const()[name = tensor("input_181_pad_type_0"), val = tensor("custom")]; + tensor input_181_pad_0 = const()[name = tensor("input_181_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220082944)))]; + tensor input_181_cast_fp16 = conv(dilations = var_2593, groups = var_2547, pad = input_181_pad_0, pad_type = input_181_pad_type_0, strides = var_2591, weight = layers_6_self_attn_q_proj_loraA_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor var_2597 = const()[name = tensor("op_2597"), val = tensor([1, 1])]; + tensor var_2599 = const()[name = tensor("op_2599"), val = tensor([1, 1])]; + tensor lora_out_241_pad_type_0 = const()[name = tensor("lora_out_241_pad_type_0"), val = tensor("custom")]; + tensor lora_out_241_pad_0 = const()[name = tensor("lora_out_241_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_243_weight_0_to_fp16 = const()[name = tensor("lora_out_243_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220123968)))]; + tensor lora_out_243_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2599, groups = var_2547, pad = lora_out_241_pad_0, pad_type = lora_out_241_pad_type_0, strides = var_2597, weight = lora_out_243_weight_0_to_fp16, x = input_181_cast_fp16)[name = tensor("lora_out_243_cast_fp16")]; + tensor query_25_cast_fp16 = add(x = pretrained_out_121_cast_fp16, y = lora_out_243_cast_fp16)[name = tensor("query_25_cast_fp16")]; + tensor var_2609 = const()[name = tensor("op_2609"), val = tensor([1, 1])]; + tensor var_2611 = const()[name = tensor("op_2611"), val = tensor([1, 1])]; + tensor pretrained_out_123_pad_type_0 = const()[name = tensor("pretrained_out_123_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_123_pad_0 = const()[name = tensor("pretrained_out_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220164992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220984256))), name = tensor("layers_6_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_123_cast_fp16 = conv(dilations = var_2611, groups = var_2547, pad = pretrained_out_123_pad_0, pad_type = pretrained_out_123_pad_type_0, strides = var_2609, weight = layers_6_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor("pretrained_out_123_cast_fp16")]; + tensor var_2615 = const()[name = tensor("op_2615"), val = tensor([1, 1])]; + tensor var_2617 = const()[name = tensor("op_2617"), val = tensor([1, 1])]; + tensor input_183_pad_type_0 = const()[name = tensor("input_183_pad_type_0"), val = tensor("custom")]; + tensor input_183_pad_0 = const()[name = tensor("input_183_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220984384)))]; + tensor input_183_cast_fp16 = conv(dilations = var_2617, groups = var_2547, pad = input_183_pad_0, pad_type = input_183_pad_type_0, strides = var_2615, weight = layers_6_self_attn_k_proj_loraA_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor var_2621 = const()[name = tensor("op_2621"), val = tensor([1, 1])]; + tensor var_2623 = const()[name = tensor("op_2623"), val = tensor([1, 1])]; + tensor lora_out_245_pad_type_0 = const()[name = tensor("lora_out_245_pad_type_0"), val = tensor("custom")]; + tensor lora_out_245_pad_0 = const()[name = tensor("lora_out_245_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_247_weight_0_to_fp16 = const()[name = tensor("lora_out_247_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221025408)))]; + tensor lora_out_247_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2623, groups = var_2547, pad = lora_out_245_pad_0, pad_type = lora_out_245_pad_type_0, strides = var_2621, weight = lora_out_247_weight_0_to_fp16, x = input_183_cast_fp16)[name = tensor("lora_out_247_cast_fp16")]; + tensor current_key_13_cast_fp16 = add(x = pretrained_out_123_cast_fp16, y = lora_out_247_cast_fp16)[name = tensor("current_key_13_cast_fp16")]; + tensor var_2634 = const()[name = tensor("op_2634"), val = tensor([1, 1])]; + tensor var_2636 = const()[name = tensor("op_2636"), val = tensor([1, 1])]; + tensor pretrained_out_125_pad_type_0 = const()[name = tensor("pretrained_out_125_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_125_pad_0 = const()[name = tensor("pretrained_out_125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221066432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221885696))), name = tensor("layers_6_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_6_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221885824)))]; + tensor pretrained_out_125_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2636, groups = var_2547, pad = pretrained_out_125_pad_0, pad_type = pretrained_out_125_pad_type_0, strides = var_2634, weight = layers_6_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor("pretrained_out_125_cast_fp16")]; + tensor var_2640 = const()[name = tensor("op_2640"), val = tensor([1, 1])]; + tensor var_2642 = const()[name = tensor("op_2642"), val = tensor([1, 1])]; + tensor input_185_pad_type_0 = const()[name = tensor("input_185_pad_type_0"), val = tensor("custom")]; + tensor input_185_pad_0 = const()[name = tensor("input_185_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221888448)))]; + tensor input_185_cast_fp16 = conv(dilations = var_2642, groups = var_2547, pad = input_185_pad_0, pad_type = input_185_pad_type_0, strides = var_2640, weight = layers_6_self_attn_v_proj_loraA_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor var_2646 = const()[name = tensor("op_2646"), val = tensor([1, 1])]; + tensor var_2648 = const()[name = tensor("op_2648"), val = tensor([1, 1])]; + tensor lora_out_249_pad_type_0 = const()[name = tensor("lora_out_249_pad_type_0"), val = tensor("custom")]; + tensor lora_out_249_pad_0 = const()[name = tensor("lora_out_249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_251_weight_0_to_fp16 = const()[name = tensor("lora_out_251_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221929472)))]; + tensor lora_out_251_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2648, groups = var_2547, pad = lora_out_249_pad_0, pad_type = lora_out_249_pad_type_0, strides = var_2646, weight = lora_out_251_weight_0_to_fp16, x = input_185_cast_fp16)[name = tensor("lora_out_251_cast_fp16")]; + tensor current_value_13_cast_fp16 = add(x = pretrained_out_125_cast_fp16, y = lora_out_251_cast_fp16)[name = tensor("current_value_13_cast_fp16")]; + tensor var_2658_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_2658_cast_fp16")]; + tensor var_2660_cast_fp16 = mul(x = var_103_cast_fp16_6, y = var_295_cast_fp16)[name = tensor("op_2660_cast_fp16")]; + tensor key_25_cast_fp16 = add(x = var_2658_cast_fp16, y = var_2660_cast_fp16)[name = tensor("key_25_cast_fp16")]; + tensor var_2662_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_2662_cast_fp16")]; + tensor var_2664_cast_fp16 = mul(x = var_138_cast_fp16_6, y = var_295_cast_fp16)[name = tensor("op_2664_cast_fp16")]; + tensor value_25_cast_fp16 = add(x = var_2662_cast_fp16, y = var_2664_cast_fp16)[name = tensor("value_25_cast_fp16")]; + tensor var_2667 = const()[name = tensor("op_2667"), val = tensor([1, 20, 64, -1])]; + tensor var_2668_cast_fp16 = reshape(shape = var_2667, x = query_25_cast_fp16)[name = tensor("op_2668_cast_fp16")]; + tensor var_2669_to_fp16 = const()[name = tensor("op_2669_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2670_cast_fp16 = mul(x = var_2668_cast_fp16, y = var_2669_to_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor var_2671 = const()[name = tensor("op_2671"), val = tensor([1, 20, 64, -1])]; + tensor var_2672_cast_fp16 = reshape(shape = var_2671, x = key_25_cast_fp16)[name = tensor("op_2672_cast_fp16")]; + tensor mh_w_37_transpose_x_0 = const()[name = tensor("mh_w_37_transpose_x_0"), val = tensor(true)]; + tensor mh_w_37_transpose_y_0 = const()[name = tensor("mh_w_37_transpose_y_0"), val = tensor(false)]; + tensor mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_2670_cast_fp16, y = var_2672_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; + tensor mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; + tensor var_2680_cast_fp16 = softmax(axis = var_2540, x = mh_w_39_cast_fp16)[name = tensor("op_2680_cast_fp16")]; + tensor var_2681 = const()[name = tensor("op_2681"), val = tensor([1, 20, 64, -1])]; + tensor var_2682_cast_fp16 = reshape(shape = var_2681, x = value_25_cast_fp16)[name = tensor("op_2682_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(true)]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_2682_cast_fp16, y = var_2680_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_2685 = const()[name = tensor("op_2685"), val = tensor([1, 1280, 1, -1])]; + tensor input_187_cast_fp16 = reshape(shape = var_2685, x = attn_25_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor var_2692 = const()[name = tensor("op_2692"), val = tensor([1, 1])]; + tensor var_2694 = const()[name = tensor("op_2694"), val = tensor([1, 1])]; + tensor pretrained_out_127_pad_type_0 = const()[name = tensor("pretrained_out_127_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_127_pad_0 = const()[name = tensor("pretrained_out_127_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221970496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222789760))), name = tensor("layers_6_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_6_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222789888)))]; + tensor pretrained_out_127_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2694, groups = var_2547, pad = pretrained_out_127_pad_0, pad_type = pretrained_out_127_pad_type_0, strides = var_2692, weight = layers_6_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = tensor("pretrained_out_127_cast_fp16")]; + tensor var_2698 = const()[name = tensor("op_2698"), val = tensor([1, 1])]; + tensor var_2700 = const()[name = tensor("op_2700"), val = tensor([1, 1])]; + tensor input_189_pad_type_0 = const()[name = tensor("input_189_pad_type_0"), val = tensor("custom")]; + tensor input_189_pad_0 = const()[name = tensor("input_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222792512)))]; + tensor input_189_cast_fp16 = conv(dilations = var_2700, groups = var_2547, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = var_2698, weight = layers_6_self_attn_o_proj_loraA_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor var_2704 = const()[name = tensor("op_2704"), val = tensor([1, 1])]; + tensor var_2706 = const()[name = tensor("op_2706"), val = tensor([1, 1])]; + tensor lora_out_253_pad_type_0 = const()[name = tensor("lora_out_253_pad_type_0"), val = tensor("custom")]; + tensor lora_out_253_pad_0 = const()[name = tensor("lora_out_253_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_255_weight_0_to_fp16 = const()[name = tensor("lora_out_255_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222833536)))]; + tensor lora_out_255_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2706, groups = var_2547, pad = lora_out_253_pad_0, pad_type = lora_out_253_pad_type_0, strides = var_2704, weight = lora_out_255_weight_0_to_fp16, x = input_189_cast_fp16)[name = tensor("lora_out_255_cast_fp16")]; + tensor obj_91_cast_fp16 = add(x = pretrained_out_127_cast_fp16, y = lora_out_255_cast_fp16)[name = tensor("obj_91_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_91_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor var_2719 = const()[name = tensor("op_2719"), val = tensor([1])]; + tensor channels_mean_39_cast_fp16 = reduce_mean(axes = var_2719, keep_dims = var_2548, x = inputs_39_cast_fp16)[name = tensor("channels_mean_39_cast_fp16")]; + tensor zero_mean_39_cast_fp16 = sub(x = inputs_39_cast_fp16, y = channels_mean_39_cast_fp16)[name = tensor("zero_mean_39_cast_fp16")]; + tensor zero_mean_sq_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = zero_mean_39_cast_fp16)[name = tensor("zero_mean_sq_39_cast_fp16")]; + tensor var_2723 = const()[name = tensor("op_2723"), val = tensor([1])]; + tensor var_2724_cast_fp16 = reduce_mean(axes = var_2723, keep_dims = var_2548, x = zero_mean_sq_39_cast_fp16)[name = tensor("op_2724_cast_fp16")]; + tensor var_2725_to_fp16 = const()[name = tensor("op_2725_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2726_cast_fp16 = add(x = var_2724_cast_fp16, y = var_2725_to_fp16)[name = tensor("op_2726_cast_fp16")]; + tensor denom_39_epsilon_0 = const()[name = tensor("denom_39_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_39_cast_fp16 = rsqrt(epsilon = denom_39_epsilon_0, x = var_2726_cast_fp16)[name = tensor("denom_39_cast_fp16")]; + tensor out_39_cast_fp16 = mul(x = zero_mean_39_cast_fp16, y = denom_39_cast_fp16)[name = tensor("out_39_cast_fp16")]; + tensor obj_93_gamma_0_to_fp16 = const()[name = tensor("obj_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222874560)))]; + tensor obj_93_beta_0_to_fp16 = const()[name = tensor("obj_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222877184)))]; + tensor obj_93_epsilon_0_to_fp16 = const()[name = tensor("obj_93_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor("obj_93_cast_fp16")]; + tensor var_2744 = const()[name = tensor("op_2744"), val = tensor([1, 1])]; + tensor var_2746 = const()[name = tensor("op_2746"), val = tensor([1, 1])]; + tensor pretrained_out_129_pad_type_0 = const()[name = tensor("pretrained_out_129_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_129_pad_0 = const()[name = tensor("pretrained_out_129_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222879808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223699072))), name = tensor("layers_6_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_6_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223699200)))]; + tensor pretrained_out_129_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2746, groups = var_2547, pad = pretrained_out_129_pad_0, pad_type = pretrained_out_129_pad_type_0, strides = var_2744, weight = layers_6_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_93_cast_fp16)[name = tensor("pretrained_out_129_cast_fp16")]; + tensor var_2750 = const()[name = tensor("op_2750"), val = tensor([1, 1])]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([1, 1])]; + tensor input_191_pad_type_0 = const()[name = tensor("input_191_pad_type_0"), val = tensor("custom")]; + tensor input_191_pad_0 = const()[name = tensor("input_191_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223701824)))]; + tensor input_191_cast_fp16 = conv(dilations = var_2752, groups = var_2547, pad = input_191_pad_0, pad_type = input_191_pad_type_0, strides = var_2750, weight = layers_6_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor var_2756 = const()[name = tensor("op_2756"), val = tensor([1, 1])]; + tensor var_2758 = const()[name = tensor("op_2758"), val = tensor([1, 1])]; + tensor lora_out_257_pad_type_0 = const()[name = tensor("lora_out_257_pad_type_0"), val = tensor("custom")]; + tensor lora_out_257_pad_0 = const()[name = tensor("lora_out_257_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_259_weight_0_to_fp16 = const()[name = tensor("lora_out_259_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223742848)))]; + tensor lora_out_259_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2758, groups = var_2547, pad = lora_out_257_pad_0, pad_type = lora_out_257_pad_type_0, strides = var_2756, weight = lora_out_259_weight_0_to_fp16, x = input_191_cast_fp16)[name = tensor("lora_out_259_cast_fp16")]; + tensor query_27_cast_fp16 = add(x = pretrained_out_129_cast_fp16, y = lora_out_259_cast_fp16)[name = tensor("query_27_cast_fp16")]; + tensor var_2768 = const()[name = tensor("op_2768"), val = tensor([1, 1])]; + tensor var_2770 = const()[name = tensor("op_2770"), val = tensor([1, 1])]; + tensor pretrained_out_131_pad_type_0 = const()[name = tensor("pretrained_out_131_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_131_pad_0 = const()[name = tensor("pretrained_out_131_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223783872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224603136))), name = tensor("layers_6_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_131_cast_fp16 = conv(dilations = var_2770, groups = var_2547, pad = pretrained_out_131_pad_0, pad_type = pretrained_out_131_pad_type_0, strides = var_2768, weight = layers_6_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_131_cast_fp16")]; + tensor var_2774 = const()[name = tensor("op_2774"), val = tensor([1, 1])]; + tensor var_2776 = const()[name = tensor("op_2776"), val = tensor([1, 1])]; + tensor input_193_pad_type_0 = const()[name = tensor("input_193_pad_type_0"), val = tensor("custom")]; + tensor input_193_pad_0 = const()[name = tensor("input_193_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224603264)))]; + tensor input_193_cast_fp16 = conv(dilations = var_2776, groups = var_2547, pad = input_193_pad_0, pad_type = input_193_pad_type_0, strides = var_2774, weight = layers_6_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_193_cast_fp16")]; + tensor var_2780 = const()[name = tensor("op_2780"), val = tensor([1, 1])]; + tensor var_2782 = const()[name = tensor("op_2782"), val = tensor([1, 1])]; + tensor lora_out_261_pad_type_0 = const()[name = tensor("lora_out_261_pad_type_0"), val = tensor("custom")]; + tensor lora_out_261_pad_0 = const()[name = tensor("lora_out_261_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_263_weight_0_to_fp16 = const()[name = tensor("lora_out_263_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224644288)))]; + tensor lora_out_263_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2782, groups = var_2547, pad = lora_out_261_pad_0, pad_type = lora_out_261_pad_type_0, strides = var_2780, weight = lora_out_263_weight_0_to_fp16, x = input_193_cast_fp16)[name = tensor("lora_out_263_cast_fp16")]; + tensor key_27_cast_fp16 = add(x = pretrained_out_131_cast_fp16, y = lora_out_263_cast_fp16)[name = tensor("key_27_cast_fp16")]; + tensor var_2793 = const()[name = tensor("op_2793"), val = tensor([1, 1])]; + tensor var_2795 = const()[name = tensor("op_2795"), val = tensor([1, 1])]; + tensor pretrained_out_133_pad_type_0 = const()[name = tensor("pretrained_out_133_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_133_pad_0 = const()[name = tensor("pretrained_out_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224685312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225504576))), name = tensor("layers_6_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_6_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225504704)))]; + tensor pretrained_out_133_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2795, groups = var_2547, pad = pretrained_out_133_pad_0, pad_type = pretrained_out_133_pad_type_0, strides = var_2793, weight = layers_6_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_133_cast_fp16")]; + tensor var_2799 = const()[name = tensor("op_2799"), val = tensor([1, 1])]; + tensor var_2801 = const()[name = tensor("op_2801"), val = tensor([1, 1])]; + tensor input_195_pad_type_0 = const()[name = tensor("input_195_pad_type_0"), val = tensor("custom")]; + tensor input_195_pad_0 = const()[name = tensor("input_195_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225507328)))]; + tensor input_195_cast_fp16 = conv(dilations = var_2801, groups = var_2547, pad = input_195_pad_0, pad_type = input_195_pad_type_0, strides = var_2799, weight = layers_6_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_195_cast_fp16")]; + tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([1, 1])]; + tensor var_2807 = const()[name = tensor("op_2807"), val = tensor([1, 1])]; + tensor lora_out_265_pad_type_0 = const()[name = tensor("lora_out_265_pad_type_0"), val = tensor("custom")]; + tensor lora_out_265_pad_0 = const()[name = tensor("lora_out_265_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_267_weight_0_to_fp16 = const()[name = tensor("lora_out_267_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225548352)))]; + tensor lora_out_267_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2807, groups = var_2547, pad = lora_out_265_pad_0, pad_type = lora_out_265_pad_type_0, strides = var_2805, weight = lora_out_267_weight_0_to_fp16, x = input_195_cast_fp16)[name = tensor("lora_out_267_cast_fp16")]; + tensor value_27_cast_fp16 = add(x = pretrained_out_133_cast_fp16, y = lora_out_267_cast_fp16)[name = tensor("value_27_cast_fp16")]; + tensor var_2814 = const()[name = tensor("op_2814"), val = tensor([1, 20, 64, -1])]; + tensor var_2815_cast_fp16 = reshape(shape = var_2814, x = query_27_cast_fp16)[name = tensor("op_2815_cast_fp16")]; + tensor var_2816_to_fp16 = const()[name = tensor("op_2816_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2817_cast_fp16 = mul(x = var_2815_cast_fp16, y = var_2816_to_fp16)[name = tensor("op_2817_cast_fp16")]; + tensor var_2818 = const()[name = tensor("op_2818"), val = tensor([1, 20, 64, -1])]; + tensor var_2819_cast_fp16 = reshape(shape = var_2818, x = key_27_cast_fp16)[name = tensor("op_2819_cast_fp16")]; + tensor mh_w_41_transpose_x_0 = const()[name = tensor("mh_w_41_transpose_x_0"), val = tensor(true)]; + tensor mh_w_41_transpose_y_0 = const()[name = tensor("mh_w_41_transpose_y_0"), val = tensor(false)]; + tensor mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_2817_cast_fp16, y = var_2819_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; + tensor obj_97_cast_fp16 = softmax(axis = var_2540, x = mh_w_41_cast_fp16)[name = tensor("obj_97_cast_fp16")]; + tensor var_2823 = const()[name = tensor("op_2823"), val = tensor([1, 20, 64, -1])]; + tensor var_2824_cast_fp16 = reshape(shape = var_2823, x = value_27_cast_fp16)[name = tensor("op_2824_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(true)]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_2824_cast_fp16, y = obj_97_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_2827 = const()[name = tensor("op_2827"), val = tensor([1, 1280, 1, -1])]; + tensor input_197_cast_fp16 = reshape(shape = var_2827, x = attn_27_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor var_2834 = const()[name = tensor("op_2834"), val = tensor([1, 1])]; + tensor var_2836 = const()[name = tensor("op_2836"), val = tensor([1, 1])]; + tensor pretrained_out_135_pad_type_0 = const()[name = tensor("pretrained_out_135_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_135_pad_0 = const()[name = tensor("pretrained_out_135_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(225589376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226408640))), name = tensor("layers_6_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_6_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226408768)))]; + tensor pretrained_out_135_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2836, groups = var_2547, pad = pretrained_out_135_pad_0, pad_type = pretrained_out_135_pad_type_0, strides = var_2834, weight = layers_6_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_197_cast_fp16)[name = tensor("pretrained_out_135_cast_fp16")]; + tensor var_2840 = const()[name = tensor("op_2840"), val = tensor([1, 1])]; + tensor var_2842 = const()[name = tensor("op_2842"), val = tensor([1, 1])]; + tensor input_199_pad_type_0 = const()[name = tensor("input_199_pad_type_0"), val = tensor("custom")]; + tensor input_199_pad_0 = const()[name = tensor("input_199_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226411392)))]; + tensor input_199_cast_fp16 = conv(dilations = var_2842, groups = var_2547, pad = input_199_pad_0, pad_type = input_199_pad_type_0, strides = var_2840, weight = layers_6_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor var_2846 = const()[name = tensor("op_2846"), val = tensor([1, 1])]; + tensor var_2848 = const()[name = tensor("op_2848"), val = tensor([1, 1])]; + tensor lora_out_269_pad_type_0 = const()[name = tensor("lora_out_269_pad_type_0"), val = tensor("custom")]; + tensor lora_out_269_pad_0 = const()[name = tensor("lora_out_269_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_271_weight_0_to_fp16 = const()[name = tensor("lora_out_271_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226452416)))]; + tensor lora_out_271_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2848, groups = var_2547, pad = lora_out_269_pad_0, pad_type = lora_out_269_pad_type_0, strides = var_2846, weight = lora_out_271_weight_0_to_fp16, x = input_199_cast_fp16)[name = tensor("lora_out_271_cast_fp16")]; + tensor obj_95_cast_fp16 = add(x = pretrained_out_135_cast_fp16, y = lora_out_271_cast_fp16)[name = tensor("obj_95_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_95_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_2857 = const()[name = tensor("op_2857"), val = tensor([1])]; + tensor channels_mean_41_cast_fp16 = reduce_mean(axes = var_2857, keep_dims = var_2548, x = inputs_41_cast_fp16)[name = tensor("channels_mean_41_cast_fp16")]; + tensor zero_mean_41_cast_fp16 = sub(x = inputs_41_cast_fp16, y = channels_mean_41_cast_fp16)[name = tensor("zero_mean_41_cast_fp16")]; + tensor zero_mean_sq_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = zero_mean_41_cast_fp16)[name = tensor("zero_mean_sq_41_cast_fp16")]; + tensor var_2861 = const()[name = tensor("op_2861"), val = tensor([1])]; + tensor var_2862_cast_fp16 = reduce_mean(axes = var_2861, keep_dims = var_2548, x = zero_mean_sq_41_cast_fp16)[name = tensor("op_2862_cast_fp16")]; + tensor var_2863_to_fp16 = const()[name = tensor("op_2863_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2864_cast_fp16 = add(x = var_2862_cast_fp16, y = var_2863_to_fp16)[name = tensor("op_2864_cast_fp16")]; + tensor denom_41_epsilon_0 = const()[name = tensor("denom_41_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_41_cast_fp16 = rsqrt(epsilon = denom_41_epsilon_0, x = var_2864_cast_fp16)[name = tensor("denom_41_cast_fp16")]; + tensor out_41_cast_fp16 = mul(x = zero_mean_41_cast_fp16, y = denom_41_cast_fp16)[name = tensor("out_41_cast_fp16")]; + tensor input_201_gamma_0_to_fp16 = const()[name = tensor("input_201_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226493440)))]; + tensor input_201_beta_0_to_fp16 = const()[name = tensor("input_201_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226496064)))]; + tensor input_201_epsilon_0_to_fp16 = const()[name = tensor("input_201_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_201_cast_fp16 = batch_norm(beta = input_201_beta_0_to_fp16, epsilon = input_201_epsilon_0_to_fp16, gamma = input_201_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor var_2878 = const()[name = tensor("op_2878"), val = tensor([1, 1])]; + tensor var_2880 = const()[name = tensor("op_2880"), val = tensor([1, 1])]; + tensor pretrained_out_137_pad_type_0 = const()[name = tensor("pretrained_out_137_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_137_pad_0 = const()[name = tensor("pretrained_out_137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226498688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229775552))), name = tensor("layers_6_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_6_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229775680)))]; + tensor pretrained_out_137_cast_fp16 = conv(bias = layers_6_fc1_pretrained_bias_to_fp16, dilations = var_2880, groups = var_2547, pad = pretrained_out_137_pad_0, pad_type = pretrained_out_137_pad_type_0, strides = var_2878, weight = layers_6_fc1_pretrained_weight_to_fp16_palettized, x = input_201_cast_fp16)[name = tensor("pretrained_out_137_cast_fp16")]; + tensor var_2884 = const()[name = tensor("op_2884"), val = tensor([1, 1])]; + tensor var_2886 = const()[name = tensor("op_2886"), val = tensor([1, 1])]; + tensor input_203_pad_type_0 = const()[name = tensor("input_203_pad_type_0"), val = tensor("custom")]; + tensor input_203_pad_0 = const()[name = tensor("input_203_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_6_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229785984)))]; + tensor input_203_cast_fp16 = conv(dilations = var_2886, groups = var_2547, pad = input_203_pad_0, pad_type = input_203_pad_type_0, strides = var_2884, weight = layers_6_fc1_loraA_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor var_2890 = const()[name = tensor("op_2890"), val = tensor([1, 1])]; + tensor var_2892 = const()[name = tensor("op_2892"), val = tensor([1, 1])]; + tensor lora_out_273_pad_type_0 = const()[name = tensor("lora_out_273_pad_type_0"), val = tensor("custom")]; + tensor lora_out_273_pad_0 = const()[name = tensor("lora_out_273_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_275_weight_0_to_fp16 = const()[name = tensor("lora_out_275_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229827008)))]; + tensor lora_out_275_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_2892, groups = var_2547, pad = lora_out_273_pad_0, pad_type = lora_out_273_pad_type_0, strides = var_2890, weight = lora_out_275_weight_0_to_fp16, x = input_203_cast_fp16)[name = tensor("lora_out_275_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = pretrained_out_137_cast_fp16, y = lora_out_275_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_mode_0 = const()[name = tensor("input_207_mode_0"), val = tensor("EXACT")]; + tensor input_207_cast_fp16 = gelu(mode = input_207_mode_0, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor var_2904 = const()[name = tensor("op_2904"), val = tensor([1, 1])]; + tensor var_2906 = const()[name = tensor("op_2906"), val = tensor([1, 1])]; + tensor pretrained_out_139_pad_type_0 = const()[name = tensor("pretrained_out_139_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_139_pad_0 = const()[name = tensor("pretrained_out_139_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229990912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233267776))), name = tensor("layers_6_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_6_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_6_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233267904)))]; + tensor pretrained_out_139_cast_fp16 = conv(bias = layers_6_fc2_pretrained_bias_to_fp16, dilations = var_2906, groups = var_2547, pad = pretrained_out_139_pad_0, pad_type = pretrained_out_139_pad_type_0, strides = var_2904, weight = layers_6_fc2_pretrained_weight_to_fp16_palettized, x = input_207_cast_fp16)[name = tensor("pretrained_out_139_cast_fp16")]; + tensor var_2910 = const()[name = tensor("op_2910"), val = tensor([1, 1])]; + tensor var_2912 = const()[name = tensor("op_2912"), val = tensor([1, 1])]; + tensor input_209_pad_type_0 = const()[name = tensor("input_209_pad_type_0"), val = tensor("custom")]; + tensor input_209_pad_0 = const()[name = tensor("input_209_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_6_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_6_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233270528)))]; + tensor input_209_cast_fp16 = conv(dilations = var_2912, groups = var_2547, pad = input_209_pad_0, pad_type = input_209_pad_type_0, strides = var_2910, weight = layers_6_fc2_loraA_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor var_2916 = const()[name = tensor("op_2916"), val = tensor([1, 1])]; + tensor var_2918 = const()[name = tensor("op_2918"), val = tensor([1, 1])]; + tensor lora_out_277_pad_type_0 = const()[name = tensor("lora_out_277_pad_type_0"), val = tensor("custom")]; + tensor lora_out_277_pad_0 = const()[name = tensor("lora_out_277_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_279_weight_0_to_fp16 = const()[name = tensor("lora_out_279_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233434432)))]; + tensor lora_out_279_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2918, groups = var_2547, pad = lora_out_277_pad_0, pad_type = lora_out_277_pad_type_0, strides = var_2916, weight = lora_out_279_weight_0_to_fp16, x = input_209_cast_fp16)[name = tensor("lora_out_279_cast_fp16")]; + tensor hidden_states_15_cast_fp16 = add(x = pretrained_out_139_cast_fp16, y = lora_out_279_cast_fp16)[name = tensor("hidden_states_15_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor var_2934 = const()[name = tensor("op_2934"), val = tensor(3)]; + tensor var_2941 = const()[name = tensor("op_2941"), val = tensor(1)]; + tensor var_2942 = const()[name = tensor("op_2942"), val = tensor(true)]; + tensor var_2954 = const()[name = tensor("op_2954"), val = tensor([1])]; + tensor channels_mean_43_cast_fp16 = reduce_mean(axes = var_2954, keep_dims = var_2942, x = inputs_43_cast_fp16)[name = tensor("channels_mean_43_cast_fp16")]; + tensor zero_mean_43_cast_fp16 = sub(x = inputs_43_cast_fp16, y = channels_mean_43_cast_fp16)[name = tensor("zero_mean_43_cast_fp16")]; + tensor zero_mean_sq_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = zero_mean_43_cast_fp16)[name = tensor("zero_mean_sq_43_cast_fp16")]; + tensor var_2958 = const()[name = tensor("op_2958"), val = tensor([1])]; + tensor var_2959_cast_fp16 = reduce_mean(axes = var_2958, keep_dims = var_2942, x = zero_mean_sq_43_cast_fp16)[name = tensor("op_2959_cast_fp16")]; + tensor var_2960_to_fp16 = const()[name = tensor("op_2960_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2961_cast_fp16 = add(x = var_2959_cast_fp16, y = var_2960_to_fp16)[name = tensor("op_2961_cast_fp16")]; + tensor denom_43_epsilon_0 = const()[name = tensor("denom_43_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_43_cast_fp16 = rsqrt(epsilon = denom_43_epsilon_0, x = var_2961_cast_fp16)[name = tensor("denom_43_cast_fp16")]; + tensor out_43_cast_fp16 = mul(x = zero_mean_43_cast_fp16, y = denom_43_cast_fp16)[name = tensor("out_43_cast_fp16")]; + tensor obj_99_gamma_0_to_fp16 = const()[name = tensor("obj_99_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233475456)))]; + tensor obj_99_beta_0_to_fp16 = const()[name = tensor("obj_99_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233478080)))]; + tensor obj_99_epsilon_0_to_fp16 = const()[name = tensor("obj_99_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_99_cast_fp16 = batch_norm(beta = obj_99_beta_0_to_fp16, epsilon = obj_99_epsilon_0_to_fp16, gamma = obj_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor("obj_99_cast_fp16")]; + tensor var_2979 = const()[name = tensor("op_2979"), val = tensor([1, 1])]; + tensor var_2981 = const()[name = tensor("op_2981"), val = tensor([1, 1])]; + tensor pretrained_out_141_pad_type_0 = const()[name = tensor("pretrained_out_141_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_141_pad_0 = const()[name = tensor("pretrained_out_141_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233480704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234299968))), name = tensor("layers_7_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_7_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234300096)))]; + tensor pretrained_out_141_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2981, groups = var_2941, pad = pretrained_out_141_pad_0, pad_type = pretrained_out_141_pad_type_0, strides = var_2979, weight = layers_7_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_99_cast_fp16)[name = tensor("pretrained_out_141_cast_fp16")]; + tensor var_2985 = const()[name = tensor("op_2985"), val = tensor([1, 1])]; + tensor var_2987 = const()[name = tensor("op_2987"), val = tensor([1, 1])]; + tensor input_211_pad_type_0 = const()[name = tensor("input_211_pad_type_0"), val = tensor("custom")]; + tensor input_211_pad_0 = const()[name = tensor("input_211_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234302720)))]; + tensor input_211_cast_fp16 = conv(dilations = var_2987, groups = var_2941, pad = input_211_pad_0, pad_type = input_211_pad_type_0, strides = var_2985, weight = layers_7_self_attn_q_proj_loraA_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor var_2991 = const()[name = tensor("op_2991"), val = tensor([1, 1])]; + tensor var_2993 = const()[name = tensor("op_2993"), val = tensor([1, 1])]; + tensor lora_out_281_pad_type_0 = const()[name = tensor("lora_out_281_pad_type_0"), val = tensor("custom")]; + tensor lora_out_281_pad_0 = const()[name = tensor("lora_out_281_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_283_weight_0_to_fp16 = const()[name = tensor("lora_out_283_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234343744)))]; + tensor lora_out_283_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2993, groups = var_2941, pad = lora_out_281_pad_0, pad_type = lora_out_281_pad_type_0, strides = var_2991, weight = lora_out_283_weight_0_to_fp16, x = input_211_cast_fp16)[name = tensor("lora_out_283_cast_fp16")]; + tensor query_29_cast_fp16 = add(x = pretrained_out_141_cast_fp16, y = lora_out_283_cast_fp16)[name = tensor("query_29_cast_fp16")]; + tensor var_3003 = const()[name = tensor("op_3003"), val = tensor([1, 1])]; + tensor var_3005 = const()[name = tensor("op_3005"), val = tensor([1, 1])]; + tensor pretrained_out_143_pad_type_0 = const()[name = tensor("pretrained_out_143_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_143_pad_0 = const()[name = tensor("pretrained_out_143_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234384768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235204032))), name = tensor("layers_7_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_143_cast_fp16 = conv(dilations = var_3005, groups = var_2941, pad = pretrained_out_143_pad_0, pad_type = pretrained_out_143_pad_type_0, strides = var_3003, weight = layers_7_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_99_cast_fp16)[name = tensor("pretrained_out_143_cast_fp16")]; + tensor var_3009 = const()[name = tensor("op_3009"), val = tensor([1, 1])]; + tensor var_3011 = const()[name = tensor("op_3011"), val = tensor([1, 1])]; + tensor input_213_pad_type_0 = const()[name = tensor("input_213_pad_type_0"), val = tensor("custom")]; + tensor input_213_pad_0 = const()[name = tensor("input_213_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235204160)))]; + tensor input_213_cast_fp16 = conv(dilations = var_3011, groups = var_2941, pad = input_213_pad_0, pad_type = input_213_pad_type_0, strides = var_3009, weight = layers_7_self_attn_k_proj_loraA_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor var_3015 = const()[name = tensor("op_3015"), val = tensor([1, 1])]; + tensor var_3017 = const()[name = tensor("op_3017"), val = tensor([1, 1])]; + tensor lora_out_285_pad_type_0 = const()[name = tensor("lora_out_285_pad_type_0"), val = tensor("custom")]; + tensor lora_out_285_pad_0 = const()[name = tensor("lora_out_285_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_287_weight_0_to_fp16 = const()[name = tensor("lora_out_287_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235245184)))]; + tensor lora_out_287_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3017, groups = var_2941, pad = lora_out_285_pad_0, pad_type = lora_out_285_pad_type_0, strides = var_3015, weight = lora_out_287_weight_0_to_fp16, x = input_213_cast_fp16)[name = tensor("lora_out_287_cast_fp16")]; + tensor current_key_15_cast_fp16 = add(x = pretrained_out_143_cast_fp16, y = lora_out_287_cast_fp16)[name = tensor("current_key_15_cast_fp16")]; + tensor var_3028 = const()[name = tensor("op_3028"), val = tensor([1, 1])]; + tensor var_3030 = const()[name = tensor("op_3030"), val = tensor([1, 1])]; + tensor pretrained_out_145_pad_type_0 = const()[name = tensor("pretrained_out_145_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_145_pad_0 = const()[name = tensor("pretrained_out_145_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235286208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236105472))), name = tensor("layers_7_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_7_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236105600)))]; + tensor pretrained_out_145_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3030, groups = var_2941, pad = pretrained_out_145_pad_0, pad_type = pretrained_out_145_pad_type_0, strides = var_3028, weight = layers_7_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_99_cast_fp16)[name = tensor("pretrained_out_145_cast_fp16")]; + tensor var_3034 = const()[name = tensor("op_3034"), val = tensor([1, 1])]; + tensor var_3036 = const()[name = tensor("op_3036"), val = tensor([1, 1])]; + tensor input_215_pad_type_0 = const()[name = tensor("input_215_pad_type_0"), val = tensor("custom")]; + tensor input_215_pad_0 = const()[name = tensor("input_215_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236108224)))]; + tensor input_215_cast_fp16 = conv(dilations = var_3036, groups = var_2941, pad = input_215_pad_0, pad_type = input_215_pad_type_0, strides = var_3034, weight = layers_7_self_attn_v_proj_loraA_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor var_3040 = const()[name = tensor("op_3040"), val = tensor([1, 1])]; + tensor var_3042 = const()[name = tensor("op_3042"), val = tensor([1, 1])]; + tensor lora_out_289_pad_type_0 = const()[name = tensor("lora_out_289_pad_type_0"), val = tensor("custom")]; + tensor lora_out_289_pad_0 = const()[name = tensor("lora_out_289_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_291_weight_0_to_fp16 = const()[name = tensor("lora_out_291_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236149248)))]; + tensor lora_out_291_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3042, groups = var_2941, pad = lora_out_289_pad_0, pad_type = lora_out_289_pad_type_0, strides = var_3040, weight = lora_out_291_weight_0_to_fp16, x = input_215_cast_fp16)[name = tensor("lora_out_291_cast_fp16")]; + tensor current_value_15_cast_fp16 = add(x = pretrained_out_145_cast_fp16, y = lora_out_291_cast_fp16)[name = tensor("current_value_15_cast_fp16")]; + tensor var_3052_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_3052_cast_fp16")]; + tensor var_3054_cast_fp16 = mul(x = var_103_cast_fp16_7, y = var_295_cast_fp16)[name = tensor("op_3054_cast_fp16")]; + tensor key_29_cast_fp16 = add(x = var_3052_cast_fp16, y = var_3054_cast_fp16)[name = tensor("key_29_cast_fp16")]; + tensor var_3056_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_3056_cast_fp16")]; + tensor var_3058_cast_fp16 = mul(x = var_138_cast_fp16_7, y = var_295_cast_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor value_29_cast_fp16 = add(x = var_3056_cast_fp16, y = var_3058_cast_fp16)[name = tensor("value_29_cast_fp16")]; + tensor var_3061 = const()[name = tensor("op_3061"), val = tensor([1, 20, 64, -1])]; + tensor var_3062_cast_fp16 = reshape(shape = var_3061, x = query_29_cast_fp16)[name = tensor("op_3062_cast_fp16")]; + tensor var_3063_to_fp16 = const()[name = tensor("op_3063_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3064_cast_fp16 = mul(x = var_3062_cast_fp16, y = var_3063_to_fp16)[name = tensor("op_3064_cast_fp16")]; + tensor var_3065 = const()[name = tensor("op_3065"), val = tensor([1, 20, 64, -1])]; + tensor var_3066_cast_fp16 = reshape(shape = var_3065, x = key_29_cast_fp16)[name = tensor("op_3066_cast_fp16")]; + tensor mh_w_43_transpose_x_0 = const()[name = tensor("mh_w_43_transpose_x_0"), val = tensor(true)]; + tensor mh_w_43_transpose_y_0 = const()[name = tensor("mh_w_43_transpose_y_0"), val = tensor(false)]; + tensor mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_3064_cast_fp16, y = var_3066_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; + tensor mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; + tensor var_3074_cast_fp16 = softmax(axis = var_2934, x = mh_w_45_cast_fp16)[name = tensor("op_3074_cast_fp16")]; + tensor var_3075 = const()[name = tensor("op_3075"), val = tensor([1, 20, 64, -1])]; + tensor var_3076_cast_fp16 = reshape(shape = var_3075, x = value_29_cast_fp16)[name = tensor("op_3076_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(true)]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_3076_cast_fp16, y = var_3074_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_3079 = const()[name = tensor("op_3079"), val = tensor([1, 1280, 1, -1])]; + tensor input_217_cast_fp16 = reshape(shape = var_3079, x = attn_29_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor var_3086 = const()[name = tensor("op_3086"), val = tensor([1, 1])]; + tensor var_3088 = const()[name = tensor("op_3088"), val = tensor([1, 1])]; + tensor pretrained_out_147_pad_type_0 = const()[name = tensor("pretrained_out_147_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_147_pad_0 = const()[name = tensor("pretrained_out_147_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236190272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237009536))), name = tensor("layers_7_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_7_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237009664)))]; + tensor pretrained_out_147_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3088, groups = var_2941, pad = pretrained_out_147_pad_0, pad_type = pretrained_out_147_pad_type_0, strides = var_3086, weight = layers_7_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_217_cast_fp16)[name = tensor("pretrained_out_147_cast_fp16")]; + tensor var_3092 = const()[name = tensor("op_3092"), val = tensor([1, 1])]; + tensor var_3094 = const()[name = tensor("op_3094"), val = tensor([1, 1])]; + tensor input_219_pad_type_0 = const()[name = tensor("input_219_pad_type_0"), val = tensor("custom")]; + tensor input_219_pad_0 = const()[name = tensor("input_219_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237012288)))]; + tensor input_219_cast_fp16 = conv(dilations = var_3094, groups = var_2941, pad = input_219_pad_0, pad_type = input_219_pad_type_0, strides = var_3092, weight = layers_7_self_attn_o_proj_loraA_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor var_3098 = const()[name = tensor("op_3098"), val = tensor([1, 1])]; + tensor var_3100 = const()[name = tensor("op_3100"), val = tensor([1, 1])]; + tensor lora_out_293_pad_type_0 = const()[name = tensor("lora_out_293_pad_type_0"), val = tensor("custom")]; + tensor lora_out_293_pad_0 = const()[name = tensor("lora_out_293_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_295_weight_0_to_fp16 = const()[name = tensor("lora_out_295_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237053312)))]; + tensor lora_out_295_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3100, groups = var_2941, pad = lora_out_293_pad_0, pad_type = lora_out_293_pad_type_0, strides = var_3098, weight = lora_out_295_weight_0_to_fp16, x = input_219_cast_fp16)[name = tensor("lora_out_295_cast_fp16")]; + tensor obj_105_cast_fp16 = add(x = pretrained_out_147_cast_fp16, y = lora_out_295_cast_fp16)[name = tensor("obj_105_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_105_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor var_3113 = const()[name = tensor("op_3113"), val = tensor([1])]; + tensor channels_mean_45_cast_fp16 = reduce_mean(axes = var_3113, keep_dims = var_2942, x = inputs_45_cast_fp16)[name = tensor("channels_mean_45_cast_fp16")]; + tensor zero_mean_45_cast_fp16 = sub(x = inputs_45_cast_fp16, y = channels_mean_45_cast_fp16)[name = tensor("zero_mean_45_cast_fp16")]; + tensor zero_mean_sq_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = zero_mean_45_cast_fp16)[name = tensor("zero_mean_sq_45_cast_fp16")]; + tensor var_3117 = const()[name = tensor("op_3117"), val = tensor([1])]; + tensor var_3118_cast_fp16 = reduce_mean(axes = var_3117, keep_dims = var_2942, x = zero_mean_sq_45_cast_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor var_3119_to_fp16 = const()[name = tensor("op_3119_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3120_cast_fp16 = add(x = var_3118_cast_fp16, y = var_3119_to_fp16)[name = tensor("op_3120_cast_fp16")]; + tensor denom_45_epsilon_0 = const()[name = tensor("denom_45_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_45_cast_fp16 = rsqrt(epsilon = denom_45_epsilon_0, x = var_3120_cast_fp16)[name = tensor("denom_45_cast_fp16")]; + tensor out_45_cast_fp16 = mul(x = zero_mean_45_cast_fp16, y = denom_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; + tensor obj_107_gamma_0_to_fp16 = const()[name = tensor("obj_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237094336)))]; + tensor obj_107_beta_0_to_fp16 = const()[name = tensor("obj_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237096960)))]; + tensor obj_107_epsilon_0_to_fp16 = const()[name = tensor("obj_107_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_107_cast_fp16 = batch_norm(beta = obj_107_beta_0_to_fp16, epsilon = obj_107_epsilon_0_to_fp16, gamma = obj_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor("obj_107_cast_fp16")]; + tensor var_3138 = const()[name = tensor("op_3138"), val = tensor([1, 1])]; + tensor var_3140 = const()[name = tensor("op_3140"), val = tensor([1, 1])]; + tensor pretrained_out_149_pad_type_0 = const()[name = tensor("pretrained_out_149_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_149_pad_0 = const()[name = tensor("pretrained_out_149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237099584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237918848))), name = tensor("layers_7_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_7_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237918976)))]; + tensor pretrained_out_149_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3140, groups = var_2941, pad = pretrained_out_149_pad_0, pad_type = pretrained_out_149_pad_type_0, strides = var_3138, weight = layers_7_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_107_cast_fp16)[name = tensor("pretrained_out_149_cast_fp16")]; + tensor var_3144 = const()[name = tensor("op_3144"), val = tensor([1, 1])]; + tensor var_3146 = const()[name = tensor("op_3146"), val = tensor([1, 1])]; + tensor input_221_pad_type_0 = const()[name = tensor("input_221_pad_type_0"), val = tensor("custom")]; + tensor input_221_pad_0 = const()[name = tensor("input_221_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237921600)))]; + tensor input_221_cast_fp16 = conv(dilations = var_3146, groups = var_2941, pad = input_221_pad_0, pad_type = input_221_pad_type_0, strides = var_3144, weight = layers_7_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_107_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor var_3150 = const()[name = tensor("op_3150"), val = tensor([1, 1])]; + tensor var_3152 = const()[name = tensor("op_3152"), val = tensor([1, 1])]; + tensor lora_out_297_pad_type_0 = const()[name = tensor("lora_out_297_pad_type_0"), val = tensor("custom")]; + tensor lora_out_297_pad_0 = const()[name = tensor("lora_out_297_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_299_weight_0_to_fp16 = const()[name = tensor("lora_out_299_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237962624)))]; + tensor lora_out_299_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3152, groups = var_2941, pad = lora_out_297_pad_0, pad_type = lora_out_297_pad_type_0, strides = var_3150, weight = lora_out_299_weight_0_to_fp16, x = input_221_cast_fp16)[name = tensor("lora_out_299_cast_fp16")]; + tensor query_31_cast_fp16 = add(x = pretrained_out_149_cast_fp16, y = lora_out_299_cast_fp16)[name = tensor("query_31_cast_fp16")]; + tensor var_3162 = const()[name = tensor("op_3162"), val = tensor([1, 1])]; + tensor var_3164 = const()[name = tensor("op_3164"), val = tensor([1, 1])]; + tensor pretrained_out_151_pad_type_0 = const()[name = tensor("pretrained_out_151_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_151_pad_0 = const()[name = tensor("pretrained_out_151_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238003648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238822912))), name = tensor("layers_7_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_151_cast_fp16 = conv(dilations = var_3164, groups = var_2941, pad = pretrained_out_151_pad_0, pad_type = pretrained_out_151_pad_type_0, strides = var_3162, weight = layers_7_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_151_cast_fp16")]; + tensor var_3168 = const()[name = tensor("op_3168"), val = tensor([1, 1])]; + tensor var_3170 = const()[name = tensor("op_3170"), val = tensor([1, 1])]; + tensor input_223_pad_type_0 = const()[name = tensor("input_223_pad_type_0"), val = tensor("custom")]; + tensor input_223_pad_0 = const()[name = tensor("input_223_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238823040)))]; + tensor input_223_cast_fp16 = conv(dilations = var_3170, groups = var_2941, pad = input_223_pad_0, pad_type = input_223_pad_type_0, strides = var_3168, weight = layers_7_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_223_cast_fp16")]; + tensor var_3174 = const()[name = tensor("op_3174"), val = tensor([1, 1])]; + tensor var_3176 = const()[name = tensor("op_3176"), val = tensor([1, 1])]; + tensor lora_out_301_pad_type_0 = const()[name = tensor("lora_out_301_pad_type_0"), val = tensor("custom")]; + tensor lora_out_301_pad_0 = const()[name = tensor("lora_out_301_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_303_weight_0_to_fp16 = const()[name = tensor("lora_out_303_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238864064)))]; + tensor lora_out_303_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3176, groups = var_2941, pad = lora_out_301_pad_0, pad_type = lora_out_301_pad_type_0, strides = var_3174, weight = lora_out_303_weight_0_to_fp16, x = input_223_cast_fp16)[name = tensor("lora_out_303_cast_fp16")]; + tensor key_31_cast_fp16 = add(x = pretrained_out_151_cast_fp16, y = lora_out_303_cast_fp16)[name = tensor("key_31_cast_fp16")]; + tensor var_3187 = const()[name = tensor("op_3187"), val = tensor([1, 1])]; + tensor var_3189 = const()[name = tensor("op_3189"), val = tensor([1, 1])]; + tensor pretrained_out_153_pad_type_0 = const()[name = tensor("pretrained_out_153_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_153_pad_0 = const()[name = tensor("pretrained_out_153_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238905088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239724352))), name = tensor("layers_7_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_7_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239724480)))]; + tensor pretrained_out_153_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3189, groups = var_2941, pad = pretrained_out_153_pad_0, pad_type = pretrained_out_153_pad_type_0, strides = var_3187, weight = layers_7_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_153_cast_fp16")]; + tensor var_3193 = const()[name = tensor("op_3193"), val = tensor([1, 1])]; + tensor var_3195 = const()[name = tensor("op_3195"), val = tensor([1, 1])]; + tensor input_225_pad_type_0 = const()[name = tensor("input_225_pad_type_0"), val = tensor("custom")]; + tensor input_225_pad_0 = const()[name = tensor("input_225_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239727104)))]; + tensor input_225_cast_fp16 = conv(dilations = var_3195, groups = var_2941, pad = input_225_pad_0, pad_type = input_225_pad_type_0, strides = var_3193, weight = layers_7_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_225_cast_fp16")]; + tensor var_3199 = const()[name = tensor("op_3199"), val = tensor([1, 1])]; + tensor var_3201 = const()[name = tensor("op_3201"), val = tensor([1, 1])]; + tensor lora_out_305_pad_type_0 = const()[name = tensor("lora_out_305_pad_type_0"), val = tensor("custom")]; + tensor lora_out_305_pad_0 = const()[name = tensor("lora_out_305_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_307_weight_0_to_fp16 = const()[name = tensor("lora_out_307_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239768128)))]; + tensor lora_out_307_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3201, groups = var_2941, pad = lora_out_305_pad_0, pad_type = lora_out_305_pad_type_0, strides = var_3199, weight = lora_out_307_weight_0_to_fp16, x = input_225_cast_fp16)[name = tensor("lora_out_307_cast_fp16")]; + tensor value_31_cast_fp16 = add(x = pretrained_out_153_cast_fp16, y = lora_out_307_cast_fp16)[name = tensor("value_31_cast_fp16")]; + tensor var_3208 = const()[name = tensor("op_3208"), val = tensor([1, 20, 64, -1])]; + tensor var_3209_cast_fp16 = reshape(shape = var_3208, x = query_31_cast_fp16)[name = tensor("op_3209_cast_fp16")]; + tensor var_3210_to_fp16 = const()[name = tensor("op_3210_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3211_cast_fp16 = mul(x = var_3209_cast_fp16, y = var_3210_to_fp16)[name = tensor("op_3211_cast_fp16")]; + tensor var_3212 = const()[name = tensor("op_3212"), val = tensor([1, 20, 64, -1])]; + tensor var_3213_cast_fp16 = reshape(shape = var_3212, x = key_31_cast_fp16)[name = tensor("op_3213_cast_fp16")]; + tensor mh_w_47_transpose_x_0 = const()[name = tensor("mh_w_47_transpose_x_0"), val = tensor(true)]; + tensor mh_w_47_transpose_y_0 = const()[name = tensor("mh_w_47_transpose_y_0"), val = tensor(false)]; + tensor mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_3211_cast_fp16, y = var_3213_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; + tensor obj_111_cast_fp16 = softmax(axis = var_2934, x = mh_w_47_cast_fp16)[name = tensor("obj_111_cast_fp16")]; + tensor var_3217 = const()[name = tensor("op_3217"), val = tensor([1, 20, 64, -1])]; + tensor var_3218_cast_fp16 = reshape(shape = var_3217, x = value_31_cast_fp16)[name = tensor("op_3218_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(true)]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_3218_cast_fp16, y = obj_111_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_3221 = const()[name = tensor("op_3221"), val = tensor([1, 1280, 1, -1])]; + tensor input_227_cast_fp16 = reshape(shape = var_3221, x = attn_31_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor var_3228 = const()[name = tensor("op_3228"), val = tensor([1, 1])]; + tensor var_3230 = const()[name = tensor("op_3230"), val = tensor([1, 1])]; + tensor pretrained_out_155_pad_type_0 = const()[name = tensor("pretrained_out_155_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_155_pad_0 = const()[name = tensor("pretrained_out_155_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239809152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240628416))), name = tensor("layers_7_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_7_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240628544)))]; + tensor pretrained_out_155_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3230, groups = var_2941, pad = pretrained_out_155_pad_0, pad_type = pretrained_out_155_pad_type_0, strides = var_3228, weight = layers_7_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_227_cast_fp16)[name = tensor("pretrained_out_155_cast_fp16")]; + tensor var_3234 = const()[name = tensor("op_3234"), val = tensor([1, 1])]; + tensor var_3236 = const()[name = tensor("op_3236"), val = tensor([1, 1])]; + tensor input_229_pad_type_0 = const()[name = tensor("input_229_pad_type_0"), val = tensor("custom")]; + tensor input_229_pad_0 = const()[name = tensor("input_229_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240631168)))]; + tensor input_229_cast_fp16 = conv(dilations = var_3236, groups = var_2941, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = var_3234, weight = layers_7_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor var_3240 = const()[name = tensor("op_3240"), val = tensor([1, 1])]; + tensor var_3242 = const()[name = tensor("op_3242"), val = tensor([1, 1])]; + tensor lora_out_309_pad_type_0 = const()[name = tensor("lora_out_309_pad_type_0"), val = tensor("custom")]; + tensor lora_out_309_pad_0 = const()[name = tensor("lora_out_309_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_311_weight_0_to_fp16 = const()[name = tensor("lora_out_311_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240672192)))]; + tensor lora_out_311_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3242, groups = var_2941, pad = lora_out_309_pad_0, pad_type = lora_out_309_pad_type_0, strides = var_3240, weight = lora_out_311_weight_0_to_fp16, x = input_229_cast_fp16)[name = tensor("lora_out_311_cast_fp16")]; + tensor obj_109_cast_fp16 = add(x = pretrained_out_155_cast_fp16, y = lora_out_311_cast_fp16)[name = tensor("obj_109_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_109_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor var_3251 = const()[name = tensor("op_3251"), val = tensor([1])]; + tensor channels_mean_47_cast_fp16 = reduce_mean(axes = var_3251, keep_dims = var_2942, x = inputs_47_cast_fp16)[name = tensor("channels_mean_47_cast_fp16")]; + tensor zero_mean_47_cast_fp16 = sub(x = inputs_47_cast_fp16, y = channels_mean_47_cast_fp16)[name = tensor("zero_mean_47_cast_fp16")]; + tensor zero_mean_sq_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = zero_mean_47_cast_fp16)[name = tensor("zero_mean_sq_47_cast_fp16")]; + tensor var_3255 = const()[name = tensor("op_3255"), val = tensor([1])]; + tensor var_3256_cast_fp16 = reduce_mean(axes = var_3255, keep_dims = var_2942, x = zero_mean_sq_47_cast_fp16)[name = tensor("op_3256_cast_fp16")]; + tensor var_3257_to_fp16 = const()[name = tensor("op_3257_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3258_cast_fp16 = add(x = var_3256_cast_fp16, y = var_3257_to_fp16)[name = tensor("op_3258_cast_fp16")]; + tensor denom_47_epsilon_0 = const()[name = tensor("denom_47_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_47_cast_fp16 = rsqrt(epsilon = denom_47_epsilon_0, x = var_3258_cast_fp16)[name = tensor("denom_47_cast_fp16")]; + tensor out_47_cast_fp16 = mul(x = zero_mean_47_cast_fp16, y = denom_47_cast_fp16)[name = tensor("out_47_cast_fp16")]; + tensor input_231_gamma_0_to_fp16 = const()[name = tensor("input_231_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240713216)))]; + tensor input_231_beta_0_to_fp16 = const()[name = tensor("input_231_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240715840)))]; + tensor input_231_epsilon_0_to_fp16 = const()[name = tensor("input_231_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_231_cast_fp16 = batch_norm(beta = input_231_beta_0_to_fp16, epsilon = input_231_epsilon_0_to_fp16, gamma = input_231_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor var_3272 = const()[name = tensor("op_3272"), val = tensor([1, 1])]; + tensor var_3274 = const()[name = tensor("op_3274"), val = tensor([1, 1])]; + tensor pretrained_out_157_pad_type_0 = const()[name = tensor("pretrained_out_157_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_157_pad_0 = const()[name = tensor("pretrained_out_157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240718464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243995328))), name = tensor("layers_7_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_7_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243995456)))]; + tensor pretrained_out_157_cast_fp16 = conv(bias = layers_7_fc1_pretrained_bias_to_fp16, dilations = var_3274, groups = var_2941, pad = pretrained_out_157_pad_0, pad_type = pretrained_out_157_pad_type_0, strides = var_3272, weight = layers_7_fc1_pretrained_weight_to_fp16_palettized, x = input_231_cast_fp16)[name = tensor("pretrained_out_157_cast_fp16")]; + tensor var_3278 = const()[name = tensor("op_3278"), val = tensor([1, 1])]; + tensor var_3280 = const()[name = tensor("op_3280"), val = tensor([1, 1])]; + tensor input_233_pad_type_0 = const()[name = tensor("input_233_pad_type_0"), val = tensor("custom")]; + tensor input_233_pad_0 = const()[name = tensor("input_233_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_7_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244005760)))]; + tensor input_233_cast_fp16 = conv(dilations = var_3280, groups = var_2941, pad = input_233_pad_0, pad_type = input_233_pad_type_0, strides = var_3278, weight = layers_7_fc1_loraA_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor var_3284 = const()[name = tensor("op_3284"), val = tensor([1, 1])]; + tensor var_3286 = const()[name = tensor("op_3286"), val = tensor([1, 1])]; + tensor lora_out_313_pad_type_0 = const()[name = tensor("lora_out_313_pad_type_0"), val = tensor("custom")]; + tensor lora_out_313_pad_0 = const()[name = tensor("lora_out_313_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_315_weight_0_to_fp16 = const()[name = tensor("lora_out_315_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244046784)))]; + tensor lora_out_315_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_3286, groups = var_2941, pad = lora_out_313_pad_0, pad_type = lora_out_313_pad_type_0, strides = var_3284, weight = lora_out_315_weight_0_to_fp16, x = input_233_cast_fp16)[name = tensor("lora_out_315_cast_fp16")]; + tensor input_235_cast_fp16 = add(x = pretrained_out_157_cast_fp16, y = lora_out_315_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_mode_0 = const()[name = tensor("input_237_mode_0"), val = tensor("EXACT")]; + tensor input_237_cast_fp16 = gelu(mode = input_237_mode_0, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor var_3298 = const()[name = tensor("op_3298"), val = tensor([1, 1])]; + tensor var_3300 = const()[name = tensor("op_3300"), val = tensor([1, 1])]; + tensor pretrained_out_159_pad_type_0 = const()[name = tensor("pretrained_out_159_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_159_pad_0 = const()[name = tensor("pretrained_out_159_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244210688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247487552))), name = tensor("layers_7_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_7_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_7_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247487680)))]; + tensor pretrained_out_159_cast_fp16 = conv(bias = layers_7_fc2_pretrained_bias_to_fp16, dilations = var_3300, groups = var_2941, pad = pretrained_out_159_pad_0, pad_type = pretrained_out_159_pad_type_0, strides = var_3298, weight = layers_7_fc2_pretrained_weight_to_fp16_palettized, x = input_237_cast_fp16)[name = tensor("pretrained_out_159_cast_fp16")]; + tensor var_3304 = const()[name = tensor("op_3304"), val = tensor([1, 1])]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 1])]; + tensor input_239_pad_type_0 = const()[name = tensor("input_239_pad_type_0"), val = tensor("custom")]; + tensor input_239_pad_0 = const()[name = tensor("input_239_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_7_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_7_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247490304)))]; + tensor input_239_cast_fp16 = conv(dilations = var_3306, groups = var_2941, pad = input_239_pad_0, pad_type = input_239_pad_type_0, strides = var_3304, weight = layers_7_fc2_loraA_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor var_3310 = const()[name = tensor("op_3310"), val = tensor([1, 1])]; + tensor var_3312 = const()[name = tensor("op_3312"), val = tensor([1, 1])]; + tensor lora_out_317_pad_type_0 = const()[name = tensor("lora_out_317_pad_type_0"), val = tensor("custom")]; + tensor lora_out_317_pad_0 = const()[name = tensor("lora_out_317_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_319_weight_0_to_fp16 = const()[name = tensor("lora_out_319_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247654208)))]; + tensor lora_out_319_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3312, groups = var_2941, pad = lora_out_317_pad_0, pad_type = lora_out_317_pad_type_0, strides = var_3310, weight = lora_out_319_weight_0_to_fp16, x = input_239_cast_fp16)[name = tensor("lora_out_319_cast_fp16")]; + tensor hidden_states_17_cast_fp16 = add(x = pretrained_out_159_cast_fp16, y = lora_out_319_cast_fp16)[name = tensor("hidden_states_17_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_3328 = const()[name = tensor("op_3328"), val = tensor(3)]; + tensor var_3335 = const()[name = tensor("op_3335"), val = tensor(1)]; + tensor var_3336 = const()[name = tensor("op_3336"), val = tensor(true)]; + tensor var_3348 = const()[name = tensor("op_3348"), val = tensor([1])]; + tensor channels_mean_49_cast_fp16 = reduce_mean(axes = var_3348, keep_dims = var_3336, x = inputs_49_cast_fp16)[name = tensor("channels_mean_49_cast_fp16")]; + tensor zero_mean_49_cast_fp16 = sub(x = inputs_49_cast_fp16, y = channels_mean_49_cast_fp16)[name = tensor("zero_mean_49_cast_fp16")]; + tensor zero_mean_sq_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = zero_mean_49_cast_fp16)[name = tensor("zero_mean_sq_49_cast_fp16")]; + tensor var_3352 = const()[name = tensor("op_3352"), val = tensor([1])]; + tensor var_3353_cast_fp16 = reduce_mean(axes = var_3352, keep_dims = var_3336, x = zero_mean_sq_49_cast_fp16)[name = tensor("op_3353_cast_fp16")]; + tensor var_3354_to_fp16 = const()[name = tensor("op_3354_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3355_cast_fp16 = add(x = var_3353_cast_fp16, y = var_3354_to_fp16)[name = tensor("op_3355_cast_fp16")]; + tensor denom_49_epsilon_0 = const()[name = tensor("denom_49_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_49_cast_fp16 = rsqrt(epsilon = denom_49_epsilon_0, x = var_3355_cast_fp16)[name = tensor("denom_49_cast_fp16")]; + tensor out_49_cast_fp16 = mul(x = zero_mean_49_cast_fp16, y = denom_49_cast_fp16)[name = tensor("out_49_cast_fp16")]; + tensor obj_113_gamma_0_to_fp16 = const()[name = tensor("obj_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247695232)))]; + tensor obj_113_beta_0_to_fp16 = const()[name = tensor("obj_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247697856)))]; + tensor obj_113_epsilon_0_to_fp16 = const()[name = tensor("obj_113_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor("obj_113_cast_fp16")]; + tensor var_3373 = const()[name = tensor("op_3373"), val = tensor([1, 1])]; + tensor var_3375 = const()[name = tensor("op_3375"), val = tensor([1, 1])]; + tensor pretrained_out_161_pad_type_0 = const()[name = tensor("pretrained_out_161_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_161_pad_0 = const()[name = tensor("pretrained_out_161_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247700480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248519744))), name = tensor("layers_8_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_8_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248519872)))]; + tensor pretrained_out_161_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3375, groups = var_3335, pad = pretrained_out_161_pad_0, pad_type = pretrained_out_161_pad_type_0, strides = var_3373, weight = layers_8_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_113_cast_fp16)[name = tensor("pretrained_out_161_cast_fp16")]; + tensor var_3379 = const()[name = tensor("op_3379"), val = tensor([1, 1])]; + tensor var_3381 = const()[name = tensor("op_3381"), val = tensor([1, 1])]; + tensor input_241_pad_type_0 = const()[name = tensor("input_241_pad_type_0"), val = tensor("custom")]; + tensor input_241_pad_0 = const()[name = tensor("input_241_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248522496)))]; + tensor input_241_cast_fp16 = conv(dilations = var_3381, groups = var_3335, pad = input_241_pad_0, pad_type = input_241_pad_type_0, strides = var_3379, weight = layers_8_self_attn_q_proj_loraA_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor var_3385 = const()[name = tensor("op_3385"), val = tensor([1, 1])]; + tensor var_3387 = const()[name = tensor("op_3387"), val = tensor([1, 1])]; + tensor lora_out_321_pad_type_0 = const()[name = tensor("lora_out_321_pad_type_0"), val = tensor("custom")]; + tensor lora_out_321_pad_0 = const()[name = tensor("lora_out_321_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_323_weight_0_to_fp16 = const()[name = tensor("lora_out_323_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248563520)))]; + tensor lora_out_323_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3387, groups = var_3335, pad = lora_out_321_pad_0, pad_type = lora_out_321_pad_type_0, strides = var_3385, weight = lora_out_323_weight_0_to_fp16, x = input_241_cast_fp16)[name = tensor("lora_out_323_cast_fp16")]; + tensor query_33_cast_fp16 = add(x = pretrained_out_161_cast_fp16, y = lora_out_323_cast_fp16)[name = tensor("query_33_cast_fp16")]; + tensor var_3397 = const()[name = tensor("op_3397"), val = tensor([1, 1])]; + tensor var_3399 = const()[name = tensor("op_3399"), val = tensor([1, 1])]; + tensor pretrained_out_163_pad_type_0 = const()[name = tensor("pretrained_out_163_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_163_pad_0 = const()[name = tensor("pretrained_out_163_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248604544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249423808))), name = tensor("layers_8_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_163_cast_fp16 = conv(dilations = var_3399, groups = var_3335, pad = pretrained_out_163_pad_0, pad_type = pretrained_out_163_pad_type_0, strides = var_3397, weight = layers_8_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_113_cast_fp16)[name = tensor("pretrained_out_163_cast_fp16")]; + tensor var_3403 = const()[name = tensor("op_3403"), val = tensor([1, 1])]; + tensor var_3405 = const()[name = tensor("op_3405"), val = tensor([1, 1])]; + tensor input_243_pad_type_0 = const()[name = tensor("input_243_pad_type_0"), val = tensor("custom")]; + tensor input_243_pad_0 = const()[name = tensor("input_243_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249423936)))]; + tensor input_243_cast_fp16 = conv(dilations = var_3405, groups = var_3335, pad = input_243_pad_0, pad_type = input_243_pad_type_0, strides = var_3403, weight = layers_8_self_attn_k_proj_loraA_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor var_3409 = const()[name = tensor("op_3409"), val = tensor([1, 1])]; + tensor var_3411 = const()[name = tensor("op_3411"), val = tensor([1, 1])]; + tensor lora_out_325_pad_type_0 = const()[name = tensor("lora_out_325_pad_type_0"), val = tensor("custom")]; + tensor lora_out_325_pad_0 = const()[name = tensor("lora_out_325_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_327_weight_0_to_fp16 = const()[name = tensor("lora_out_327_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249464960)))]; + tensor lora_out_327_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3411, groups = var_3335, pad = lora_out_325_pad_0, pad_type = lora_out_325_pad_type_0, strides = var_3409, weight = lora_out_327_weight_0_to_fp16, x = input_243_cast_fp16)[name = tensor("lora_out_327_cast_fp16")]; + tensor current_key_17_cast_fp16 = add(x = pretrained_out_163_cast_fp16, y = lora_out_327_cast_fp16)[name = tensor("current_key_17_cast_fp16")]; + tensor var_3422 = const()[name = tensor("op_3422"), val = tensor([1, 1])]; + tensor var_3424 = const()[name = tensor("op_3424"), val = tensor([1, 1])]; + tensor pretrained_out_165_pad_type_0 = const()[name = tensor("pretrained_out_165_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_165_pad_0 = const()[name = tensor("pretrained_out_165_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249505984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250325248))), name = tensor("layers_8_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_8_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250325376)))]; + tensor pretrained_out_165_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3424, groups = var_3335, pad = pretrained_out_165_pad_0, pad_type = pretrained_out_165_pad_type_0, strides = var_3422, weight = layers_8_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_113_cast_fp16)[name = tensor("pretrained_out_165_cast_fp16")]; + tensor var_3428 = const()[name = tensor("op_3428"), val = tensor([1, 1])]; + tensor var_3430 = const()[name = tensor("op_3430"), val = tensor([1, 1])]; + tensor input_245_pad_type_0 = const()[name = tensor("input_245_pad_type_0"), val = tensor("custom")]; + tensor input_245_pad_0 = const()[name = tensor("input_245_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250328000)))]; + tensor input_245_cast_fp16 = conv(dilations = var_3430, groups = var_3335, pad = input_245_pad_0, pad_type = input_245_pad_type_0, strides = var_3428, weight = layers_8_self_attn_v_proj_loraA_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor var_3434 = const()[name = tensor("op_3434"), val = tensor([1, 1])]; + tensor var_3436 = const()[name = tensor("op_3436"), val = tensor([1, 1])]; + tensor lora_out_329_pad_type_0 = const()[name = tensor("lora_out_329_pad_type_0"), val = tensor("custom")]; + tensor lora_out_329_pad_0 = const()[name = tensor("lora_out_329_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_331_weight_0_to_fp16 = const()[name = tensor("lora_out_331_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250369024)))]; + tensor lora_out_331_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3436, groups = var_3335, pad = lora_out_329_pad_0, pad_type = lora_out_329_pad_type_0, strides = var_3434, weight = lora_out_331_weight_0_to_fp16, x = input_245_cast_fp16)[name = tensor("lora_out_331_cast_fp16")]; + tensor current_value_17_cast_fp16 = add(x = pretrained_out_165_cast_fp16, y = lora_out_331_cast_fp16)[name = tensor("current_value_17_cast_fp16")]; + tensor var_3446_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_3446_cast_fp16")]; + tensor var_3448_cast_fp16 = mul(x = var_103_cast_fp16_8, y = var_295_cast_fp16)[name = tensor("op_3448_cast_fp16")]; + tensor key_33_cast_fp16 = add(x = var_3446_cast_fp16, y = var_3448_cast_fp16)[name = tensor("key_33_cast_fp16")]; + tensor var_3450_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_3450_cast_fp16")]; + tensor var_3452_cast_fp16 = mul(x = var_138_cast_fp16_8, y = var_295_cast_fp16)[name = tensor("op_3452_cast_fp16")]; + tensor value_33_cast_fp16 = add(x = var_3450_cast_fp16, y = var_3452_cast_fp16)[name = tensor("value_33_cast_fp16")]; + tensor var_3455 = const()[name = tensor("op_3455"), val = tensor([1, 20, 64, -1])]; + tensor var_3456_cast_fp16 = reshape(shape = var_3455, x = query_33_cast_fp16)[name = tensor("op_3456_cast_fp16")]; + tensor var_3457_to_fp16 = const()[name = tensor("op_3457_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3458_cast_fp16 = mul(x = var_3456_cast_fp16, y = var_3457_to_fp16)[name = tensor("op_3458_cast_fp16")]; + tensor var_3459 = const()[name = tensor("op_3459"), val = tensor([1, 20, 64, -1])]; + tensor var_3460_cast_fp16 = reshape(shape = var_3459, x = key_33_cast_fp16)[name = tensor("op_3460_cast_fp16")]; + tensor mh_w_49_transpose_x_0 = const()[name = tensor("mh_w_49_transpose_x_0"), val = tensor(true)]; + tensor mh_w_49_transpose_y_0 = const()[name = tensor("mh_w_49_transpose_y_0"), val = tensor(false)]; + tensor mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_3458_cast_fp16, y = var_3460_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; + tensor mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; + tensor var_3468_cast_fp16 = softmax(axis = var_3328, x = mh_w_51_cast_fp16)[name = tensor("op_3468_cast_fp16")]; + tensor var_3469 = const()[name = tensor("op_3469"), val = tensor([1, 20, 64, -1])]; + tensor var_3470_cast_fp16 = reshape(shape = var_3469, x = value_33_cast_fp16)[name = tensor("op_3470_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(true)]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_3470_cast_fp16, y = var_3468_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_3473 = const()[name = tensor("op_3473"), val = tensor([1, 1280, 1, -1])]; + tensor input_247_cast_fp16 = reshape(shape = var_3473, x = attn_33_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor var_3480 = const()[name = tensor("op_3480"), val = tensor([1, 1])]; + tensor var_3482 = const()[name = tensor("op_3482"), val = tensor([1, 1])]; + tensor pretrained_out_167_pad_type_0 = const()[name = tensor("pretrained_out_167_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_167_pad_0 = const()[name = tensor("pretrained_out_167_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250410048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251229312))), name = tensor("layers_8_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_8_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251229440)))]; + tensor pretrained_out_167_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3482, groups = var_3335, pad = pretrained_out_167_pad_0, pad_type = pretrained_out_167_pad_type_0, strides = var_3480, weight = layers_8_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = tensor("pretrained_out_167_cast_fp16")]; + tensor var_3486 = const()[name = tensor("op_3486"), val = tensor([1, 1])]; + tensor var_3488 = const()[name = tensor("op_3488"), val = tensor([1, 1])]; + tensor input_249_pad_type_0 = const()[name = tensor("input_249_pad_type_0"), val = tensor("custom")]; + tensor input_249_pad_0 = const()[name = tensor("input_249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251232064)))]; + tensor input_249_cast_fp16 = conv(dilations = var_3488, groups = var_3335, pad = input_249_pad_0, pad_type = input_249_pad_type_0, strides = var_3486, weight = layers_8_self_attn_o_proj_loraA_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor var_3492 = const()[name = tensor("op_3492"), val = tensor([1, 1])]; + tensor var_3494 = const()[name = tensor("op_3494"), val = tensor([1, 1])]; + tensor lora_out_333_pad_type_0 = const()[name = tensor("lora_out_333_pad_type_0"), val = tensor("custom")]; + tensor lora_out_333_pad_0 = const()[name = tensor("lora_out_333_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_335_weight_0_to_fp16 = const()[name = tensor("lora_out_335_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251273088)))]; + tensor lora_out_335_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3494, groups = var_3335, pad = lora_out_333_pad_0, pad_type = lora_out_333_pad_type_0, strides = var_3492, weight = lora_out_335_weight_0_to_fp16, x = input_249_cast_fp16)[name = tensor("lora_out_335_cast_fp16")]; + tensor obj_119_cast_fp16 = add(x = pretrained_out_167_cast_fp16, y = lora_out_335_cast_fp16)[name = tensor("obj_119_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_119_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor var_3507 = const()[name = tensor("op_3507"), val = tensor([1])]; + tensor channels_mean_51_cast_fp16 = reduce_mean(axes = var_3507, keep_dims = var_3336, x = inputs_51_cast_fp16)[name = tensor("channels_mean_51_cast_fp16")]; + tensor zero_mean_51_cast_fp16 = sub(x = inputs_51_cast_fp16, y = channels_mean_51_cast_fp16)[name = tensor("zero_mean_51_cast_fp16")]; + tensor zero_mean_sq_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = zero_mean_51_cast_fp16)[name = tensor("zero_mean_sq_51_cast_fp16")]; + tensor var_3511 = const()[name = tensor("op_3511"), val = tensor([1])]; + tensor var_3512_cast_fp16 = reduce_mean(axes = var_3511, keep_dims = var_3336, x = zero_mean_sq_51_cast_fp16)[name = tensor("op_3512_cast_fp16")]; + tensor var_3513_to_fp16 = const()[name = tensor("op_3513_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3514_cast_fp16 = add(x = var_3512_cast_fp16, y = var_3513_to_fp16)[name = tensor("op_3514_cast_fp16")]; + tensor denom_51_epsilon_0 = const()[name = tensor("denom_51_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_51_cast_fp16 = rsqrt(epsilon = denom_51_epsilon_0, x = var_3514_cast_fp16)[name = tensor("denom_51_cast_fp16")]; + tensor out_51_cast_fp16 = mul(x = zero_mean_51_cast_fp16, y = denom_51_cast_fp16)[name = tensor("out_51_cast_fp16")]; + tensor obj_121_gamma_0_to_fp16 = const()[name = tensor("obj_121_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251314112)))]; + tensor obj_121_beta_0_to_fp16 = const()[name = tensor("obj_121_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251316736)))]; + tensor obj_121_epsilon_0_to_fp16 = const()[name = tensor("obj_121_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor("obj_121_cast_fp16")]; + tensor var_3532 = const()[name = tensor("op_3532"), val = tensor([1, 1])]; + tensor var_3534 = const()[name = tensor("op_3534"), val = tensor([1, 1])]; + tensor pretrained_out_169_pad_type_0 = const()[name = tensor("pretrained_out_169_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_169_pad_0 = const()[name = tensor("pretrained_out_169_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251319360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252138624))), name = tensor("layers_8_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_8_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252138752)))]; + tensor pretrained_out_169_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3534, groups = var_3335, pad = pretrained_out_169_pad_0, pad_type = pretrained_out_169_pad_type_0, strides = var_3532, weight = layers_8_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor("pretrained_out_169_cast_fp16")]; + tensor var_3538 = const()[name = tensor("op_3538"), val = tensor([1, 1])]; + tensor var_3540 = const()[name = tensor("op_3540"), val = tensor([1, 1])]; + tensor input_251_pad_type_0 = const()[name = tensor("input_251_pad_type_0"), val = tensor("custom")]; + tensor input_251_pad_0 = const()[name = tensor("input_251_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252141376)))]; + tensor input_251_cast_fp16 = conv(dilations = var_3540, groups = var_3335, pad = input_251_pad_0, pad_type = input_251_pad_type_0, strides = var_3538, weight = layers_8_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor var_3544 = const()[name = tensor("op_3544"), val = tensor([1, 1])]; + tensor var_3546 = const()[name = tensor("op_3546"), val = tensor([1, 1])]; + tensor lora_out_337_pad_type_0 = const()[name = tensor("lora_out_337_pad_type_0"), val = tensor("custom")]; + tensor lora_out_337_pad_0 = const()[name = tensor("lora_out_337_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_339_weight_0_to_fp16 = const()[name = tensor("lora_out_339_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252182400)))]; + tensor lora_out_339_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3546, groups = var_3335, pad = lora_out_337_pad_0, pad_type = lora_out_337_pad_type_0, strides = var_3544, weight = lora_out_339_weight_0_to_fp16, x = input_251_cast_fp16)[name = tensor("lora_out_339_cast_fp16")]; + tensor query_35_cast_fp16 = add(x = pretrained_out_169_cast_fp16, y = lora_out_339_cast_fp16)[name = tensor("query_35_cast_fp16")]; + tensor var_3556 = const()[name = tensor("op_3556"), val = tensor([1, 1])]; + tensor var_3558 = const()[name = tensor("op_3558"), val = tensor([1, 1])]; + tensor pretrained_out_171_pad_type_0 = const()[name = tensor("pretrained_out_171_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_171_pad_0 = const()[name = tensor("pretrained_out_171_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252223424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253042688))), name = tensor("layers_8_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_171_cast_fp16 = conv(dilations = var_3558, groups = var_3335, pad = pretrained_out_171_pad_0, pad_type = pretrained_out_171_pad_type_0, strides = var_3556, weight = layers_8_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_171_cast_fp16")]; + tensor var_3562 = const()[name = tensor("op_3562"), val = tensor([1, 1])]; + tensor var_3564 = const()[name = tensor("op_3564"), val = tensor([1, 1])]; + tensor input_253_pad_type_0 = const()[name = tensor("input_253_pad_type_0"), val = tensor("custom")]; + tensor input_253_pad_0 = const()[name = tensor("input_253_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253042816)))]; + tensor input_253_cast_fp16 = conv(dilations = var_3564, groups = var_3335, pad = input_253_pad_0, pad_type = input_253_pad_type_0, strides = var_3562, weight = layers_8_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_253_cast_fp16")]; + tensor var_3568 = const()[name = tensor("op_3568"), val = tensor([1, 1])]; + tensor var_3570 = const()[name = tensor("op_3570"), val = tensor([1, 1])]; + tensor lora_out_341_pad_type_0 = const()[name = tensor("lora_out_341_pad_type_0"), val = tensor("custom")]; + tensor lora_out_341_pad_0 = const()[name = tensor("lora_out_341_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_343_weight_0_to_fp16 = const()[name = tensor("lora_out_343_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253083840)))]; + tensor lora_out_343_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3570, groups = var_3335, pad = lora_out_341_pad_0, pad_type = lora_out_341_pad_type_0, strides = var_3568, weight = lora_out_343_weight_0_to_fp16, x = input_253_cast_fp16)[name = tensor("lora_out_343_cast_fp16")]; + tensor key_35_cast_fp16 = add(x = pretrained_out_171_cast_fp16, y = lora_out_343_cast_fp16)[name = tensor("key_35_cast_fp16")]; + tensor var_3581 = const()[name = tensor("op_3581"), val = tensor([1, 1])]; + tensor var_3583 = const()[name = tensor("op_3583"), val = tensor([1, 1])]; + tensor pretrained_out_173_pad_type_0 = const()[name = tensor("pretrained_out_173_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_173_pad_0 = const()[name = tensor("pretrained_out_173_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253124864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253944128))), name = tensor("layers_8_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_8_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253944256)))]; + tensor pretrained_out_173_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3583, groups = var_3335, pad = pretrained_out_173_pad_0, pad_type = pretrained_out_173_pad_type_0, strides = var_3581, weight = layers_8_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_173_cast_fp16")]; + tensor var_3587 = const()[name = tensor("op_3587"), val = tensor([1, 1])]; + tensor var_3589 = const()[name = tensor("op_3589"), val = tensor([1, 1])]; + tensor input_255_pad_type_0 = const()[name = tensor("input_255_pad_type_0"), val = tensor("custom")]; + tensor input_255_pad_0 = const()[name = tensor("input_255_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253946880)))]; + tensor input_255_cast_fp16 = conv(dilations = var_3589, groups = var_3335, pad = input_255_pad_0, pad_type = input_255_pad_type_0, strides = var_3587, weight = layers_8_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_255_cast_fp16")]; + tensor var_3593 = const()[name = tensor("op_3593"), val = tensor([1, 1])]; + tensor var_3595 = const()[name = tensor("op_3595"), val = tensor([1, 1])]; + tensor lora_out_345_pad_type_0 = const()[name = tensor("lora_out_345_pad_type_0"), val = tensor("custom")]; + tensor lora_out_345_pad_0 = const()[name = tensor("lora_out_345_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_347_weight_0_to_fp16 = const()[name = tensor("lora_out_347_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(253987904)))]; + tensor lora_out_347_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3595, groups = var_3335, pad = lora_out_345_pad_0, pad_type = lora_out_345_pad_type_0, strides = var_3593, weight = lora_out_347_weight_0_to_fp16, x = input_255_cast_fp16)[name = tensor("lora_out_347_cast_fp16")]; + tensor value_35_cast_fp16 = add(x = pretrained_out_173_cast_fp16, y = lora_out_347_cast_fp16)[name = tensor("value_35_cast_fp16")]; + tensor var_3602 = const()[name = tensor("op_3602"), val = tensor([1, 20, 64, -1])]; + tensor var_3603_cast_fp16 = reshape(shape = var_3602, x = query_35_cast_fp16)[name = tensor("op_3603_cast_fp16")]; + tensor var_3604_to_fp16 = const()[name = tensor("op_3604_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3605_cast_fp16 = mul(x = var_3603_cast_fp16, y = var_3604_to_fp16)[name = tensor("op_3605_cast_fp16")]; + tensor var_3606 = const()[name = tensor("op_3606"), val = tensor([1, 20, 64, -1])]; + tensor var_3607_cast_fp16 = reshape(shape = var_3606, x = key_35_cast_fp16)[name = tensor("op_3607_cast_fp16")]; + tensor mh_w_53_transpose_x_0 = const()[name = tensor("mh_w_53_transpose_x_0"), val = tensor(true)]; + tensor mh_w_53_transpose_y_0 = const()[name = tensor("mh_w_53_transpose_y_0"), val = tensor(false)]; + tensor mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_3605_cast_fp16, y = var_3607_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; + tensor obj_125_cast_fp16 = softmax(axis = var_3328, x = mh_w_53_cast_fp16)[name = tensor("obj_125_cast_fp16")]; + tensor var_3611 = const()[name = tensor("op_3611"), val = tensor([1, 20, 64, -1])]; + tensor var_3612_cast_fp16 = reshape(shape = var_3611, x = value_35_cast_fp16)[name = tensor("op_3612_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(true)]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_3612_cast_fp16, y = obj_125_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_3615 = const()[name = tensor("op_3615"), val = tensor([1, 1280, 1, -1])]; + tensor input_257_cast_fp16 = reshape(shape = var_3615, x = attn_35_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor var_3622 = const()[name = tensor("op_3622"), val = tensor([1, 1])]; + tensor var_3624 = const()[name = tensor("op_3624"), val = tensor([1, 1])]; + tensor pretrained_out_175_pad_type_0 = const()[name = tensor("pretrained_out_175_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_175_pad_0 = const()[name = tensor("pretrained_out_175_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254028928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254848192))), name = tensor("layers_8_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_8_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254848320)))]; + tensor pretrained_out_175_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3624, groups = var_3335, pad = pretrained_out_175_pad_0, pad_type = pretrained_out_175_pad_type_0, strides = var_3622, weight = layers_8_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_257_cast_fp16)[name = tensor("pretrained_out_175_cast_fp16")]; + tensor var_3628 = const()[name = tensor("op_3628"), val = tensor([1, 1])]; + tensor var_3630 = const()[name = tensor("op_3630"), val = tensor([1, 1])]; + tensor input_259_pad_type_0 = const()[name = tensor("input_259_pad_type_0"), val = tensor("custom")]; + tensor input_259_pad_0 = const()[name = tensor("input_259_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254850944)))]; + tensor input_259_cast_fp16 = conv(dilations = var_3630, groups = var_3335, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = var_3628, weight = layers_8_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor var_3634 = const()[name = tensor("op_3634"), val = tensor([1, 1])]; + tensor var_3636 = const()[name = tensor("op_3636"), val = tensor([1, 1])]; + tensor lora_out_349_pad_type_0 = const()[name = tensor("lora_out_349_pad_type_0"), val = tensor("custom")]; + tensor lora_out_349_pad_0 = const()[name = tensor("lora_out_349_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_351_weight_0_to_fp16 = const()[name = tensor("lora_out_351_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254891968)))]; + tensor lora_out_351_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3636, groups = var_3335, pad = lora_out_349_pad_0, pad_type = lora_out_349_pad_type_0, strides = var_3634, weight = lora_out_351_weight_0_to_fp16, x = input_259_cast_fp16)[name = tensor("lora_out_351_cast_fp16")]; + tensor obj_123_cast_fp16 = add(x = pretrained_out_175_cast_fp16, y = lora_out_351_cast_fp16)[name = tensor("obj_123_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_123_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor var_3645 = const()[name = tensor("op_3645"), val = tensor([1])]; + tensor channels_mean_53_cast_fp16 = reduce_mean(axes = var_3645, keep_dims = var_3336, x = inputs_53_cast_fp16)[name = tensor("channels_mean_53_cast_fp16")]; + tensor zero_mean_53_cast_fp16 = sub(x = inputs_53_cast_fp16, y = channels_mean_53_cast_fp16)[name = tensor("zero_mean_53_cast_fp16")]; + tensor zero_mean_sq_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = zero_mean_53_cast_fp16)[name = tensor("zero_mean_sq_53_cast_fp16")]; + tensor var_3649 = const()[name = tensor("op_3649"), val = tensor([1])]; + tensor var_3650_cast_fp16 = reduce_mean(axes = var_3649, keep_dims = var_3336, x = zero_mean_sq_53_cast_fp16)[name = tensor("op_3650_cast_fp16")]; + tensor var_3651_to_fp16 = const()[name = tensor("op_3651_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3652_cast_fp16 = add(x = var_3650_cast_fp16, y = var_3651_to_fp16)[name = tensor("op_3652_cast_fp16")]; + tensor denom_53_epsilon_0 = const()[name = tensor("denom_53_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_53_cast_fp16 = rsqrt(epsilon = denom_53_epsilon_0, x = var_3652_cast_fp16)[name = tensor("denom_53_cast_fp16")]; + tensor out_53_cast_fp16 = mul(x = zero_mean_53_cast_fp16, y = denom_53_cast_fp16)[name = tensor("out_53_cast_fp16")]; + tensor input_261_gamma_0_to_fp16 = const()[name = tensor("input_261_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254932992)))]; + tensor input_261_beta_0_to_fp16 = const()[name = tensor("input_261_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254935616)))]; + tensor input_261_epsilon_0_to_fp16 = const()[name = tensor("input_261_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_261_cast_fp16 = batch_norm(beta = input_261_beta_0_to_fp16, epsilon = input_261_epsilon_0_to_fp16, gamma = input_261_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor var_3666 = const()[name = tensor("op_3666"), val = tensor([1, 1])]; + tensor var_3668 = const()[name = tensor("op_3668"), val = tensor([1, 1])]; + tensor pretrained_out_177_pad_type_0 = const()[name = tensor("pretrained_out_177_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_177_pad_0 = const()[name = tensor("pretrained_out_177_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254938240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258215104))), name = tensor("layers_8_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_8_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258215232)))]; + tensor pretrained_out_177_cast_fp16 = conv(bias = layers_8_fc1_pretrained_bias_to_fp16, dilations = var_3668, groups = var_3335, pad = pretrained_out_177_pad_0, pad_type = pretrained_out_177_pad_type_0, strides = var_3666, weight = layers_8_fc1_pretrained_weight_to_fp16_palettized, x = input_261_cast_fp16)[name = tensor("pretrained_out_177_cast_fp16")]; + tensor var_3672 = const()[name = tensor("op_3672"), val = tensor([1, 1])]; + tensor var_3674 = const()[name = tensor("op_3674"), val = tensor([1, 1])]; + tensor input_263_pad_type_0 = const()[name = tensor("input_263_pad_type_0"), val = tensor("custom")]; + tensor input_263_pad_0 = const()[name = tensor("input_263_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_8_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258225536)))]; + tensor input_263_cast_fp16 = conv(dilations = var_3674, groups = var_3335, pad = input_263_pad_0, pad_type = input_263_pad_type_0, strides = var_3672, weight = layers_8_fc1_loraA_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor var_3678 = const()[name = tensor("op_3678"), val = tensor([1, 1])]; + tensor var_3680 = const()[name = tensor("op_3680"), val = tensor([1, 1])]; + tensor lora_out_353_pad_type_0 = const()[name = tensor("lora_out_353_pad_type_0"), val = tensor("custom")]; + tensor lora_out_353_pad_0 = const()[name = tensor("lora_out_353_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_355_weight_0_to_fp16 = const()[name = tensor("lora_out_355_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258266560)))]; + tensor lora_out_355_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_3680, groups = var_3335, pad = lora_out_353_pad_0, pad_type = lora_out_353_pad_type_0, strides = var_3678, weight = lora_out_355_weight_0_to_fp16, x = input_263_cast_fp16)[name = tensor("lora_out_355_cast_fp16")]; + tensor input_265_cast_fp16 = add(x = pretrained_out_177_cast_fp16, y = lora_out_355_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_mode_0 = const()[name = tensor("input_267_mode_0"), val = tensor("EXACT")]; + tensor input_267_cast_fp16 = gelu(mode = input_267_mode_0, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor var_3692 = const()[name = tensor("op_3692"), val = tensor([1, 1])]; + tensor var_3694 = const()[name = tensor("op_3694"), val = tensor([1, 1])]; + tensor pretrained_out_179_pad_type_0 = const()[name = tensor("pretrained_out_179_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_179_pad_0 = const()[name = tensor("pretrained_out_179_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258430464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261707328))), name = tensor("layers_8_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_8_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_8_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261707456)))]; + tensor pretrained_out_179_cast_fp16 = conv(bias = layers_8_fc2_pretrained_bias_to_fp16, dilations = var_3694, groups = var_3335, pad = pretrained_out_179_pad_0, pad_type = pretrained_out_179_pad_type_0, strides = var_3692, weight = layers_8_fc2_pretrained_weight_to_fp16_palettized, x = input_267_cast_fp16)[name = tensor("pretrained_out_179_cast_fp16")]; + tensor var_3698 = const()[name = tensor("op_3698"), val = tensor([1, 1])]; + tensor var_3700 = const()[name = tensor("op_3700"), val = tensor([1, 1])]; + tensor input_269_pad_type_0 = const()[name = tensor("input_269_pad_type_0"), val = tensor("custom")]; + tensor input_269_pad_0 = const()[name = tensor("input_269_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_8_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_8_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261710080)))]; + tensor input_269_cast_fp16 = conv(dilations = var_3700, groups = var_3335, pad = input_269_pad_0, pad_type = input_269_pad_type_0, strides = var_3698, weight = layers_8_fc2_loraA_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor var_3704 = const()[name = tensor("op_3704"), val = tensor([1, 1])]; + tensor var_3706 = const()[name = tensor("op_3706"), val = tensor([1, 1])]; + tensor lora_out_357_pad_type_0 = const()[name = tensor("lora_out_357_pad_type_0"), val = tensor("custom")]; + tensor lora_out_357_pad_0 = const()[name = tensor("lora_out_357_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_359_weight_0_to_fp16 = const()[name = tensor("lora_out_359_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261873984)))]; + tensor lora_out_359_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3706, groups = var_3335, pad = lora_out_357_pad_0, pad_type = lora_out_357_pad_type_0, strides = var_3704, weight = lora_out_359_weight_0_to_fp16, x = input_269_cast_fp16)[name = tensor("lora_out_359_cast_fp16")]; + tensor hidden_states_19_cast_fp16 = add(x = pretrained_out_179_cast_fp16, y = lora_out_359_cast_fp16)[name = tensor("hidden_states_19_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor var_3722 = const()[name = tensor("op_3722"), val = tensor(3)]; + tensor var_3729 = const()[name = tensor("op_3729"), val = tensor(1)]; + tensor var_3730 = const()[name = tensor("op_3730"), val = tensor(true)]; + tensor var_3742 = const()[name = tensor("op_3742"), val = tensor([1])]; + tensor channels_mean_55_cast_fp16 = reduce_mean(axes = var_3742, keep_dims = var_3730, x = inputs_55_cast_fp16)[name = tensor("channels_mean_55_cast_fp16")]; + tensor zero_mean_55_cast_fp16 = sub(x = inputs_55_cast_fp16, y = channels_mean_55_cast_fp16)[name = tensor("zero_mean_55_cast_fp16")]; + tensor zero_mean_sq_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = zero_mean_55_cast_fp16)[name = tensor("zero_mean_sq_55_cast_fp16")]; + tensor var_3746 = const()[name = tensor("op_3746"), val = tensor([1])]; + tensor var_3747_cast_fp16 = reduce_mean(axes = var_3746, keep_dims = var_3730, x = zero_mean_sq_55_cast_fp16)[name = tensor("op_3747_cast_fp16")]; + tensor var_3748_to_fp16 = const()[name = tensor("op_3748_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3749_cast_fp16 = add(x = var_3747_cast_fp16, y = var_3748_to_fp16)[name = tensor("op_3749_cast_fp16")]; + tensor denom_55_epsilon_0 = const()[name = tensor("denom_55_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_55_cast_fp16 = rsqrt(epsilon = denom_55_epsilon_0, x = var_3749_cast_fp16)[name = tensor("denom_55_cast_fp16")]; + tensor out_55_cast_fp16 = mul(x = zero_mean_55_cast_fp16, y = denom_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; + tensor obj_127_gamma_0_to_fp16 = const()[name = tensor("obj_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261915008)))]; + tensor obj_127_beta_0_to_fp16 = const()[name = tensor("obj_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261917632)))]; + tensor obj_127_epsilon_0_to_fp16 = const()[name = tensor("obj_127_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_127_cast_fp16 = batch_norm(beta = obj_127_beta_0_to_fp16, epsilon = obj_127_epsilon_0_to_fp16, gamma = obj_127_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor("obj_127_cast_fp16")]; + tensor var_3767 = const()[name = tensor("op_3767"), val = tensor([1, 1])]; + tensor var_3769 = const()[name = tensor("op_3769"), val = tensor([1, 1])]; + tensor pretrained_out_181_pad_type_0 = const()[name = tensor("pretrained_out_181_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_181_pad_0 = const()[name = tensor("pretrained_out_181_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261920256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262739520))), name = tensor("layers_9_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_9_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262739648)))]; + tensor pretrained_out_181_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3769, groups = var_3729, pad = pretrained_out_181_pad_0, pad_type = pretrained_out_181_pad_type_0, strides = var_3767, weight = layers_9_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_127_cast_fp16)[name = tensor("pretrained_out_181_cast_fp16")]; + tensor var_3773 = const()[name = tensor("op_3773"), val = tensor([1, 1])]; + tensor var_3775 = const()[name = tensor("op_3775"), val = tensor([1, 1])]; + tensor input_271_pad_type_0 = const()[name = tensor("input_271_pad_type_0"), val = tensor("custom")]; + tensor input_271_pad_0 = const()[name = tensor("input_271_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262742272)))]; + tensor input_271_cast_fp16 = conv(dilations = var_3775, groups = var_3729, pad = input_271_pad_0, pad_type = input_271_pad_type_0, strides = var_3773, weight = layers_9_self_attn_q_proj_loraA_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor var_3779 = const()[name = tensor("op_3779"), val = tensor([1, 1])]; + tensor var_3781 = const()[name = tensor("op_3781"), val = tensor([1, 1])]; + tensor lora_out_361_pad_type_0 = const()[name = tensor("lora_out_361_pad_type_0"), val = tensor("custom")]; + tensor lora_out_361_pad_0 = const()[name = tensor("lora_out_361_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_363_weight_0_to_fp16 = const()[name = tensor("lora_out_363_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262783296)))]; + tensor lora_out_363_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3781, groups = var_3729, pad = lora_out_361_pad_0, pad_type = lora_out_361_pad_type_0, strides = var_3779, weight = lora_out_363_weight_0_to_fp16, x = input_271_cast_fp16)[name = tensor("lora_out_363_cast_fp16")]; + tensor query_37_cast_fp16 = add(x = pretrained_out_181_cast_fp16, y = lora_out_363_cast_fp16)[name = tensor("query_37_cast_fp16")]; + tensor var_3791 = const()[name = tensor("op_3791"), val = tensor([1, 1])]; + tensor var_3793 = const()[name = tensor("op_3793"), val = tensor([1, 1])]; + tensor pretrained_out_183_pad_type_0 = const()[name = tensor("pretrained_out_183_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_183_pad_0 = const()[name = tensor("pretrained_out_183_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262824320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263643584))), name = tensor("layers_9_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_183_cast_fp16 = conv(dilations = var_3793, groups = var_3729, pad = pretrained_out_183_pad_0, pad_type = pretrained_out_183_pad_type_0, strides = var_3791, weight = layers_9_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_127_cast_fp16)[name = tensor("pretrained_out_183_cast_fp16")]; + tensor var_3797 = const()[name = tensor("op_3797"), val = tensor([1, 1])]; + tensor var_3799 = const()[name = tensor("op_3799"), val = tensor([1, 1])]; + tensor input_273_pad_type_0 = const()[name = tensor("input_273_pad_type_0"), val = tensor("custom")]; + tensor input_273_pad_0 = const()[name = tensor("input_273_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263643712)))]; + tensor input_273_cast_fp16 = conv(dilations = var_3799, groups = var_3729, pad = input_273_pad_0, pad_type = input_273_pad_type_0, strides = var_3797, weight = layers_9_self_attn_k_proj_loraA_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor var_3803 = const()[name = tensor("op_3803"), val = tensor([1, 1])]; + tensor var_3805 = const()[name = tensor("op_3805"), val = tensor([1, 1])]; + tensor lora_out_365_pad_type_0 = const()[name = tensor("lora_out_365_pad_type_0"), val = tensor("custom")]; + tensor lora_out_365_pad_0 = const()[name = tensor("lora_out_365_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_367_weight_0_to_fp16 = const()[name = tensor("lora_out_367_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263684736)))]; + tensor lora_out_367_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3805, groups = var_3729, pad = lora_out_365_pad_0, pad_type = lora_out_365_pad_type_0, strides = var_3803, weight = lora_out_367_weight_0_to_fp16, x = input_273_cast_fp16)[name = tensor("lora_out_367_cast_fp16")]; + tensor current_key_19_cast_fp16 = add(x = pretrained_out_183_cast_fp16, y = lora_out_367_cast_fp16)[name = tensor("current_key_19_cast_fp16")]; + tensor var_3816 = const()[name = tensor("op_3816"), val = tensor([1, 1])]; + tensor var_3818 = const()[name = tensor("op_3818"), val = tensor([1, 1])]; + tensor pretrained_out_185_pad_type_0 = const()[name = tensor("pretrained_out_185_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_185_pad_0 = const()[name = tensor("pretrained_out_185_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(263725760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264545024))), name = tensor("layers_9_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_9_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264545152)))]; + tensor pretrained_out_185_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3818, groups = var_3729, pad = pretrained_out_185_pad_0, pad_type = pretrained_out_185_pad_type_0, strides = var_3816, weight = layers_9_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_127_cast_fp16)[name = tensor("pretrained_out_185_cast_fp16")]; + tensor var_3822 = const()[name = tensor("op_3822"), val = tensor([1, 1])]; + tensor var_3824 = const()[name = tensor("op_3824"), val = tensor([1, 1])]; + tensor input_275_pad_type_0 = const()[name = tensor("input_275_pad_type_0"), val = tensor("custom")]; + tensor input_275_pad_0 = const()[name = tensor("input_275_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264547776)))]; + tensor input_275_cast_fp16 = conv(dilations = var_3824, groups = var_3729, pad = input_275_pad_0, pad_type = input_275_pad_type_0, strides = var_3822, weight = layers_9_self_attn_v_proj_loraA_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor var_3828 = const()[name = tensor("op_3828"), val = tensor([1, 1])]; + tensor var_3830 = const()[name = tensor("op_3830"), val = tensor([1, 1])]; + tensor lora_out_369_pad_type_0 = const()[name = tensor("lora_out_369_pad_type_0"), val = tensor("custom")]; + tensor lora_out_369_pad_0 = const()[name = tensor("lora_out_369_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_371_weight_0_to_fp16 = const()[name = tensor("lora_out_371_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264588800)))]; + tensor lora_out_371_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3830, groups = var_3729, pad = lora_out_369_pad_0, pad_type = lora_out_369_pad_type_0, strides = var_3828, weight = lora_out_371_weight_0_to_fp16, x = input_275_cast_fp16)[name = tensor("lora_out_371_cast_fp16")]; + tensor current_value_19_cast_fp16 = add(x = pretrained_out_185_cast_fp16, y = lora_out_371_cast_fp16)[name = tensor("current_value_19_cast_fp16")]; + tensor var_3840_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_3840_cast_fp16")]; + tensor var_3842_cast_fp16 = mul(x = var_103_cast_fp16_9, y = var_295_cast_fp16)[name = tensor("op_3842_cast_fp16")]; + tensor key_37_cast_fp16 = add(x = var_3840_cast_fp16, y = var_3842_cast_fp16)[name = tensor("key_37_cast_fp16")]; + tensor var_3844_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_3844_cast_fp16")]; + tensor var_3846_cast_fp16 = mul(x = var_138_cast_fp16_9, y = var_295_cast_fp16)[name = tensor("op_3846_cast_fp16")]; + tensor value_37_cast_fp16 = add(x = var_3844_cast_fp16, y = var_3846_cast_fp16)[name = tensor("value_37_cast_fp16")]; + tensor var_3849 = const()[name = tensor("op_3849"), val = tensor([1, 20, 64, -1])]; + tensor var_3850_cast_fp16 = reshape(shape = var_3849, x = query_37_cast_fp16)[name = tensor("op_3850_cast_fp16")]; + tensor var_3851_to_fp16 = const()[name = tensor("op_3851_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3852_cast_fp16 = mul(x = var_3850_cast_fp16, y = var_3851_to_fp16)[name = tensor("op_3852_cast_fp16")]; + tensor var_3853 = const()[name = tensor("op_3853"), val = tensor([1, 20, 64, -1])]; + tensor var_3854_cast_fp16 = reshape(shape = var_3853, x = key_37_cast_fp16)[name = tensor("op_3854_cast_fp16")]; + tensor mh_w_55_transpose_x_0 = const()[name = tensor("mh_w_55_transpose_x_0"), val = tensor(true)]; + tensor mh_w_55_transpose_y_0 = const()[name = tensor("mh_w_55_transpose_y_0"), val = tensor(false)]; + tensor mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_3852_cast_fp16, y = var_3854_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; + tensor mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; + tensor var_3862_cast_fp16 = softmax(axis = var_3722, x = mh_w_57_cast_fp16)[name = tensor("op_3862_cast_fp16")]; + tensor var_3863 = const()[name = tensor("op_3863"), val = tensor([1, 20, 64, -1])]; + tensor var_3864_cast_fp16 = reshape(shape = var_3863, x = value_37_cast_fp16)[name = tensor("op_3864_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(true)]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_3864_cast_fp16, y = var_3862_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_3867 = const()[name = tensor("op_3867"), val = tensor([1, 1280, 1, -1])]; + tensor input_277_cast_fp16 = reshape(shape = var_3867, x = attn_37_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor var_3874 = const()[name = tensor("op_3874"), val = tensor([1, 1])]; + tensor var_3876 = const()[name = tensor("op_3876"), val = tensor([1, 1])]; + tensor pretrained_out_187_pad_type_0 = const()[name = tensor("pretrained_out_187_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_187_pad_0 = const()[name = tensor("pretrained_out_187_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264629824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265449088))), name = tensor("layers_9_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_9_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265449216)))]; + tensor pretrained_out_187_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3876, groups = var_3729, pad = pretrained_out_187_pad_0, pad_type = pretrained_out_187_pad_type_0, strides = var_3874, weight = layers_9_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = tensor("pretrained_out_187_cast_fp16")]; + tensor var_3880 = const()[name = tensor("op_3880"), val = tensor([1, 1])]; + tensor var_3882 = const()[name = tensor("op_3882"), val = tensor([1, 1])]; + tensor input_279_pad_type_0 = const()[name = tensor("input_279_pad_type_0"), val = tensor("custom")]; + tensor input_279_pad_0 = const()[name = tensor("input_279_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265451840)))]; + tensor input_279_cast_fp16 = conv(dilations = var_3882, groups = var_3729, pad = input_279_pad_0, pad_type = input_279_pad_type_0, strides = var_3880, weight = layers_9_self_attn_o_proj_loraA_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor var_3886 = const()[name = tensor("op_3886"), val = tensor([1, 1])]; + tensor var_3888 = const()[name = tensor("op_3888"), val = tensor([1, 1])]; + tensor lora_out_373_pad_type_0 = const()[name = tensor("lora_out_373_pad_type_0"), val = tensor("custom")]; + tensor lora_out_373_pad_0 = const()[name = tensor("lora_out_373_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_375_weight_0_to_fp16 = const()[name = tensor("lora_out_375_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265492864)))]; + tensor lora_out_375_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3888, groups = var_3729, pad = lora_out_373_pad_0, pad_type = lora_out_373_pad_type_0, strides = var_3886, weight = lora_out_375_weight_0_to_fp16, x = input_279_cast_fp16)[name = tensor("lora_out_375_cast_fp16")]; + tensor obj_133_cast_fp16 = add(x = pretrained_out_187_cast_fp16, y = lora_out_375_cast_fp16)[name = tensor("obj_133_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_133_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor var_3901 = const()[name = tensor("op_3901"), val = tensor([1])]; + tensor channels_mean_57_cast_fp16 = reduce_mean(axes = var_3901, keep_dims = var_3730, x = inputs_57_cast_fp16)[name = tensor("channels_mean_57_cast_fp16")]; + tensor zero_mean_57_cast_fp16 = sub(x = inputs_57_cast_fp16, y = channels_mean_57_cast_fp16)[name = tensor("zero_mean_57_cast_fp16")]; + tensor zero_mean_sq_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = zero_mean_57_cast_fp16)[name = tensor("zero_mean_sq_57_cast_fp16")]; + tensor var_3905 = const()[name = tensor("op_3905"), val = tensor([1])]; + tensor var_3906_cast_fp16 = reduce_mean(axes = var_3905, keep_dims = var_3730, x = zero_mean_sq_57_cast_fp16)[name = tensor("op_3906_cast_fp16")]; + tensor var_3907_to_fp16 = const()[name = tensor("op_3907_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3908_cast_fp16 = add(x = var_3906_cast_fp16, y = var_3907_to_fp16)[name = tensor("op_3908_cast_fp16")]; + tensor denom_57_epsilon_0 = const()[name = tensor("denom_57_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_57_cast_fp16 = rsqrt(epsilon = denom_57_epsilon_0, x = var_3908_cast_fp16)[name = tensor("denom_57_cast_fp16")]; + tensor out_57_cast_fp16 = mul(x = zero_mean_57_cast_fp16, y = denom_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; + tensor obj_135_gamma_0_to_fp16 = const()[name = tensor("obj_135_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265533888)))]; + tensor obj_135_beta_0_to_fp16 = const()[name = tensor("obj_135_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265536512)))]; + tensor obj_135_epsilon_0_to_fp16 = const()[name = tensor("obj_135_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_135_cast_fp16 = batch_norm(beta = obj_135_beta_0_to_fp16, epsilon = obj_135_epsilon_0_to_fp16, gamma = obj_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor("obj_135_cast_fp16")]; + tensor var_3926 = const()[name = tensor("op_3926"), val = tensor([1, 1])]; + tensor var_3928 = const()[name = tensor("op_3928"), val = tensor([1, 1])]; + tensor pretrained_out_189_pad_type_0 = const()[name = tensor("pretrained_out_189_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_189_pad_0 = const()[name = tensor("pretrained_out_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265539136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266358400))), name = tensor("layers_9_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_9_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266358528)))]; + tensor pretrained_out_189_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3928, groups = var_3729, pad = pretrained_out_189_pad_0, pad_type = pretrained_out_189_pad_type_0, strides = var_3926, weight = layers_9_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_135_cast_fp16)[name = tensor("pretrained_out_189_cast_fp16")]; + tensor var_3932 = const()[name = tensor("op_3932"), val = tensor([1, 1])]; + tensor var_3934 = const()[name = tensor("op_3934"), val = tensor([1, 1])]; + tensor input_281_pad_type_0 = const()[name = tensor("input_281_pad_type_0"), val = tensor("custom")]; + tensor input_281_pad_0 = const()[name = tensor("input_281_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266361152)))]; + tensor input_281_cast_fp16 = conv(dilations = var_3934, groups = var_3729, pad = input_281_pad_0, pad_type = input_281_pad_type_0, strides = var_3932, weight = layers_9_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_135_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor var_3938 = const()[name = tensor("op_3938"), val = tensor([1, 1])]; + tensor var_3940 = const()[name = tensor("op_3940"), val = tensor([1, 1])]; + tensor lora_out_377_pad_type_0 = const()[name = tensor("lora_out_377_pad_type_0"), val = tensor("custom")]; + tensor lora_out_377_pad_0 = const()[name = tensor("lora_out_377_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_379_weight_0_to_fp16 = const()[name = tensor("lora_out_379_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266402176)))]; + tensor lora_out_379_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3940, groups = var_3729, pad = lora_out_377_pad_0, pad_type = lora_out_377_pad_type_0, strides = var_3938, weight = lora_out_379_weight_0_to_fp16, x = input_281_cast_fp16)[name = tensor("lora_out_379_cast_fp16")]; + tensor query_39_cast_fp16 = add(x = pretrained_out_189_cast_fp16, y = lora_out_379_cast_fp16)[name = tensor("query_39_cast_fp16")]; + tensor var_3950 = const()[name = tensor("op_3950"), val = tensor([1, 1])]; + tensor var_3952 = const()[name = tensor("op_3952"), val = tensor([1, 1])]; + tensor pretrained_out_191_pad_type_0 = const()[name = tensor("pretrained_out_191_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_191_pad_0 = const()[name = tensor("pretrained_out_191_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266443200))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267262464))), name = tensor("layers_9_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_191_cast_fp16 = conv(dilations = var_3952, groups = var_3729, pad = pretrained_out_191_pad_0, pad_type = pretrained_out_191_pad_type_0, strides = var_3950, weight = layers_9_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_191_cast_fp16")]; + tensor var_3956 = const()[name = tensor("op_3956"), val = tensor([1, 1])]; + tensor var_3958 = const()[name = tensor("op_3958"), val = tensor([1, 1])]; + tensor input_283_pad_type_0 = const()[name = tensor("input_283_pad_type_0"), val = tensor("custom")]; + tensor input_283_pad_0 = const()[name = tensor("input_283_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267262592)))]; + tensor input_283_cast_fp16 = conv(dilations = var_3958, groups = var_3729, pad = input_283_pad_0, pad_type = input_283_pad_type_0, strides = var_3956, weight = layers_9_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_283_cast_fp16")]; + tensor var_3962 = const()[name = tensor("op_3962"), val = tensor([1, 1])]; + tensor var_3964 = const()[name = tensor("op_3964"), val = tensor([1, 1])]; + tensor lora_out_381_pad_type_0 = const()[name = tensor("lora_out_381_pad_type_0"), val = tensor("custom")]; + tensor lora_out_381_pad_0 = const()[name = tensor("lora_out_381_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_383_weight_0_to_fp16 = const()[name = tensor("lora_out_383_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267303616)))]; + tensor lora_out_383_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3964, groups = var_3729, pad = lora_out_381_pad_0, pad_type = lora_out_381_pad_type_0, strides = var_3962, weight = lora_out_383_weight_0_to_fp16, x = input_283_cast_fp16)[name = tensor("lora_out_383_cast_fp16")]; + tensor key_39_cast_fp16 = add(x = pretrained_out_191_cast_fp16, y = lora_out_383_cast_fp16)[name = tensor("key_39_cast_fp16")]; + tensor var_3975 = const()[name = tensor("op_3975"), val = tensor([1, 1])]; + tensor var_3977 = const()[name = tensor("op_3977"), val = tensor([1, 1])]; + tensor pretrained_out_193_pad_type_0 = const()[name = tensor("pretrained_out_193_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_193_pad_0 = const()[name = tensor("pretrained_out_193_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267344640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268163904))), name = tensor("layers_9_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_9_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268164032)))]; + tensor pretrained_out_193_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3977, groups = var_3729, pad = pretrained_out_193_pad_0, pad_type = pretrained_out_193_pad_type_0, strides = var_3975, weight = layers_9_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_193_cast_fp16")]; + tensor var_3981 = const()[name = tensor("op_3981"), val = tensor([1, 1])]; + tensor var_3983 = const()[name = tensor("op_3983"), val = tensor([1, 1])]; + tensor input_285_pad_type_0 = const()[name = tensor("input_285_pad_type_0"), val = tensor("custom")]; + tensor input_285_pad_0 = const()[name = tensor("input_285_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268166656)))]; + tensor input_285_cast_fp16 = conv(dilations = var_3983, groups = var_3729, pad = input_285_pad_0, pad_type = input_285_pad_type_0, strides = var_3981, weight = layers_9_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_285_cast_fp16")]; + tensor var_3987 = const()[name = tensor("op_3987"), val = tensor([1, 1])]; + tensor var_3989 = const()[name = tensor("op_3989"), val = tensor([1, 1])]; + tensor lora_out_385_pad_type_0 = const()[name = tensor("lora_out_385_pad_type_0"), val = tensor("custom")]; + tensor lora_out_385_pad_0 = const()[name = tensor("lora_out_385_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_387_weight_0_to_fp16 = const()[name = tensor("lora_out_387_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268207680)))]; + tensor lora_out_387_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3989, groups = var_3729, pad = lora_out_385_pad_0, pad_type = lora_out_385_pad_type_0, strides = var_3987, weight = lora_out_387_weight_0_to_fp16, x = input_285_cast_fp16)[name = tensor("lora_out_387_cast_fp16")]; + tensor value_39_cast_fp16 = add(x = pretrained_out_193_cast_fp16, y = lora_out_387_cast_fp16)[name = tensor("value_39_cast_fp16")]; + tensor var_3996 = const()[name = tensor("op_3996"), val = tensor([1, 20, 64, -1])]; + tensor var_3997_cast_fp16 = reshape(shape = var_3996, x = query_39_cast_fp16)[name = tensor("op_3997_cast_fp16")]; + tensor var_3998_to_fp16 = const()[name = tensor("op_3998_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3999_cast_fp16 = mul(x = var_3997_cast_fp16, y = var_3998_to_fp16)[name = tensor("op_3999_cast_fp16")]; + tensor var_4000 = const()[name = tensor("op_4000"), val = tensor([1, 20, 64, -1])]; + tensor var_4001_cast_fp16 = reshape(shape = var_4000, x = key_39_cast_fp16)[name = tensor("op_4001_cast_fp16")]; + tensor mh_w_59_transpose_x_0 = const()[name = tensor("mh_w_59_transpose_x_0"), val = tensor(true)]; + tensor mh_w_59_transpose_y_0 = const()[name = tensor("mh_w_59_transpose_y_0"), val = tensor(false)]; + tensor mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_3999_cast_fp16, y = var_4001_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; + tensor obj_139_cast_fp16 = softmax(axis = var_3722, x = mh_w_59_cast_fp16)[name = tensor("obj_139_cast_fp16")]; + tensor var_4005 = const()[name = tensor("op_4005"), val = tensor([1, 20, 64, -1])]; + tensor var_4006_cast_fp16 = reshape(shape = var_4005, x = value_39_cast_fp16)[name = tensor("op_4006_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(true)]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_4006_cast_fp16, y = obj_139_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_4009 = const()[name = tensor("op_4009"), val = tensor([1, 1280, 1, -1])]; + tensor input_287_cast_fp16 = reshape(shape = var_4009, x = attn_39_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor var_4016 = const()[name = tensor("op_4016"), val = tensor([1, 1])]; + tensor var_4018 = const()[name = tensor("op_4018"), val = tensor([1, 1])]; + tensor pretrained_out_195_pad_type_0 = const()[name = tensor("pretrained_out_195_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_195_pad_0 = const()[name = tensor("pretrained_out_195_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268248704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269067968))), name = tensor("layers_9_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_9_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269068096)))]; + tensor pretrained_out_195_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4018, groups = var_3729, pad = pretrained_out_195_pad_0, pad_type = pretrained_out_195_pad_type_0, strides = var_4016, weight = layers_9_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = tensor("pretrained_out_195_cast_fp16")]; + tensor var_4022 = const()[name = tensor("op_4022"), val = tensor([1, 1])]; + tensor var_4024 = const()[name = tensor("op_4024"), val = tensor([1, 1])]; + tensor input_289_pad_type_0 = const()[name = tensor("input_289_pad_type_0"), val = tensor("custom")]; + tensor input_289_pad_0 = const()[name = tensor("input_289_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269070720)))]; + tensor input_289_cast_fp16 = conv(dilations = var_4024, groups = var_3729, pad = input_289_pad_0, pad_type = input_289_pad_type_0, strides = var_4022, weight = layers_9_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor var_4028 = const()[name = tensor("op_4028"), val = tensor([1, 1])]; + tensor var_4030 = const()[name = tensor("op_4030"), val = tensor([1, 1])]; + tensor lora_out_389_pad_type_0 = const()[name = tensor("lora_out_389_pad_type_0"), val = tensor("custom")]; + tensor lora_out_389_pad_0 = const()[name = tensor("lora_out_389_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_391_weight_0_to_fp16 = const()[name = tensor("lora_out_391_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269111744)))]; + tensor lora_out_391_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4030, groups = var_3729, pad = lora_out_389_pad_0, pad_type = lora_out_389_pad_type_0, strides = var_4028, weight = lora_out_391_weight_0_to_fp16, x = input_289_cast_fp16)[name = tensor("lora_out_391_cast_fp16")]; + tensor obj_137_cast_fp16 = add(x = pretrained_out_195_cast_fp16, y = lora_out_391_cast_fp16)[name = tensor("obj_137_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_137_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor var_4039 = const()[name = tensor("op_4039"), val = tensor([1])]; + tensor channels_mean_59_cast_fp16 = reduce_mean(axes = var_4039, keep_dims = var_3730, x = inputs_59_cast_fp16)[name = tensor("channels_mean_59_cast_fp16")]; + tensor zero_mean_59_cast_fp16 = sub(x = inputs_59_cast_fp16, y = channels_mean_59_cast_fp16)[name = tensor("zero_mean_59_cast_fp16")]; + tensor zero_mean_sq_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = zero_mean_59_cast_fp16)[name = tensor("zero_mean_sq_59_cast_fp16")]; + tensor var_4043 = const()[name = tensor("op_4043"), val = tensor([1])]; + tensor var_4044_cast_fp16 = reduce_mean(axes = var_4043, keep_dims = var_3730, x = zero_mean_sq_59_cast_fp16)[name = tensor("op_4044_cast_fp16")]; + tensor var_4045_to_fp16 = const()[name = tensor("op_4045_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4046_cast_fp16 = add(x = var_4044_cast_fp16, y = var_4045_to_fp16)[name = tensor("op_4046_cast_fp16")]; + tensor denom_59_epsilon_0 = const()[name = tensor("denom_59_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_59_cast_fp16 = rsqrt(epsilon = denom_59_epsilon_0, x = var_4046_cast_fp16)[name = tensor("denom_59_cast_fp16")]; + tensor out_59_cast_fp16 = mul(x = zero_mean_59_cast_fp16, y = denom_59_cast_fp16)[name = tensor("out_59_cast_fp16")]; + tensor input_291_gamma_0_to_fp16 = const()[name = tensor("input_291_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269152768)))]; + tensor input_291_beta_0_to_fp16 = const()[name = tensor("input_291_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269155392)))]; + tensor input_291_epsilon_0_to_fp16 = const()[name = tensor("input_291_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_291_cast_fp16 = batch_norm(beta = input_291_beta_0_to_fp16, epsilon = input_291_epsilon_0_to_fp16, gamma = input_291_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor var_4060 = const()[name = tensor("op_4060"), val = tensor([1, 1])]; + tensor var_4062 = const()[name = tensor("op_4062"), val = tensor([1, 1])]; + tensor pretrained_out_197_pad_type_0 = const()[name = tensor("pretrained_out_197_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_197_pad_0 = const()[name = tensor("pretrained_out_197_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269158016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272434880))), name = tensor("layers_9_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_9_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272435008)))]; + tensor pretrained_out_197_cast_fp16 = conv(bias = layers_9_fc1_pretrained_bias_to_fp16, dilations = var_4062, groups = var_3729, pad = pretrained_out_197_pad_0, pad_type = pretrained_out_197_pad_type_0, strides = var_4060, weight = layers_9_fc1_pretrained_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = tensor("pretrained_out_197_cast_fp16")]; + tensor var_4066 = const()[name = tensor("op_4066"), val = tensor([1, 1])]; + tensor var_4068 = const()[name = tensor("op_4068"), val = tensor([1, 1])]; + tensor input_293_pad_type_0 = const()[name = tensor("input_293_pad_type_0"), val = tensor("custom")]; + tensor input_293_pad_0 = const()[name = tensor("input_293_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_9_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272445312)))]; + tensor input_293_cast_fp16 = conv(dilations = var_4068, groups = var_3729, pad = input_293_pad_0, pad_type = input_293_pad_type_0, strides = var_4066, weight = layers_9_fc1_loraA_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor var_4072 = const()[name = tensor("op_4072"), val = tensor([1, 1])]; + tensor var_4074 = const()[name = tensor("op_4074"), val = tensor([1, 1])]; + tensor lora_out_393_pad_type_0 = const()[name = tensor("lora_out_393_pad_type_0"), val = tensor("custom")]; + tensor lora_out_393_pad_0 = const()[name = tensor("lora_out_393_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_395_weight_0_to_fp16 = const()[name = tensor("lora_out_395_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272486336)))]; + tensor lora_out_395_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_4074, groups = var_3729, pad = lora_out_393_pad_0, pad_type = lora_out_393_pad_type_0, strides = var_4072, weight = lora_out_395_weight_0_to_fp16, x = input_293_cast_fp16)[name = tensor("lora_out_395_cast_fp16")]; + tensor input_295_cast_fp16 = add(x = pretrained_out_197_cast_fp16, y = lora_out_395_cast_fp16)[name = tensor("input_295_cast_fp16")]; + tensor input_297_mode_0 = const()[name = tensor("input_297_mode_0"), val = tensor("EXACT")]; + tensor input_297_cast_fp16 = gelu(mode = input_297_mode_0, x = input_295_cast_fp16)[name = tensor("input_297_cast_fp16")]; + tensor var_4086 = const()[name = tensor("op_4086"), val = tensor([1, 1])]; + tensor var_4088 = const()[name = tensor("op_4088"), val = tensor([1, 1])]; + tensor pretrained_out_199_pad_type_0 = const()[name = tensor("pretrained_out_199_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_199_pad_0 = const()[name = tensor("pretrained_out_199_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272650240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277565504))), name = tensor("layers_9_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_9_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_9_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277565696)))]; + tensor pretrained_out_199_cast_fp16 = conv(bias = layers_9_fc2_pretrained_bias_to_fp16, dilations = var_4088, groups = var_3729, pad = pretrained_out_199_pad_0, pad_type = pretrained_out_199_pad_type_0, strides = var_4086, weight = layers_9_fc2_pretrained_weight_to_fp16_palettized, x = input_297_cast_fp16)[name = tensor("pretrained_out_199_cast_fp16")]; + tensor var_4092 = const()[name = tensor("op_4092"), val = tensor([1, 1])]; + tensor var_4094 = const()[name = tensor("op_4094"), val = tensor([1, 1])]; + tensor input_299_pad_type_0 = const()[name = tensor("input_299_pad_type_0"), val = tensor("custom")]; + tensor input_299_pad_0 = const()[name = tensor("input_299_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_9_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_9_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277568320)))]; + tensor input_299_cast_fp16 = conv(dilations = var_4094, groups = var_3729, pad = input_299_pad_0, pad_type = input_299_pad_type_0, strides = var_4092, weight = layers_9_fc2_loraA_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("input_299_cast_fp16")]; + tensor var_4098 = const()[name = tensor("op_4098"), val = tensor([1, 1])]; + tensor var_4100 = const()[name = tensor("op_4100"), val = tensor([1, 1])]; + tensor lora_out_397_pad_type_0 = const()[name = tensor("lora_out_397_pad_type_0"), val = tensor("custom")]; + tensor lora_out_397_pad_0 = const()[name = tensor("lora_out_397_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_399_weight_0_to_fp16 = const()[name = tensor("lora_out_399_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277732224)))]; + tensor lora_out_399_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4100, groups = var_3729, pad = lora_out_397_pad_0, pad_type = lora_out_397_pad_type_0, strides = var_4098, weight = lora_out_399_weight_0_to_fp16, x = input_299_cast_fp16)[name = tensor("lora_out_399_cast_fp16")]; + tensor hidden_states_21_cast_fp16 = add(x = pretrained_out_199_cast_fp16, y = lora_out_399_cast_fp16)[name = tensor("hidden_states_21_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_4116 = const()[name = tensor("op_4116"), val = tensor(3)]; + tensor var_4123 = const()[name = tensor("op_4123"), val = tensor(1)]; + tensor var_4124 = const()[name = tensor("op_4124"), val = tensor(true)]; + tensor var_4136 = const()[name = tensor("op_4136"), val = tensor([1])]; + tensor channels_mean_61_cast_fp16 = reduce_mean(axes = var_4136, keep_dims = var_4124, x = inputs_61_cast_fp16)[name = tensor("channels_mean_61_cast_fp16")]; + tensor zero_mean_61_cast_fp16 = sub(x = inputs_61_cast_fp16, y = channels_mean_61_cast_fp16)[name = tensor("zero_mean_61_cast_fp16")]; + tensor zero_mean_sq_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = zero_mean_61_cast_fp16)[name = tensor("zero_mean_sq_61_cast_fp16")]; + tensor var_4140 = const()[name = tensor("op_4140"), val = tensor([1])]; + tensor var_4141_cast_fp16 = reduce_mean(axes = var_4140, keep_dims = var_4124, x = zero_mean_sq_61_cast_fp16)[name = tensor("op_4141_cast_fp16")]; + tensor var_4142_to_fp16 = const()[name = tensor("op_4142_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4143_cast_fp16 = add(x = var_4141_cast_fp16, y = var_4142_to_fp16)[name = tensor("op_4143_cast_fp16")]; + tensor denom_61_epsilon_0 = const()[name = tensor("denom_61_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_61_cast_fp16 = rsqrt(epsilon = denom_61_epsilon_0, x = var_4143_cast_fp16)[name = tensor("denom_61_cast_fp16")]; + tensor out_61_cast_fp16 = mul(x = zero_mean_61_cast_fp16, y = denom_61_cast_fp16)[name = tensor("out_61_cast_fp16")]; + tensor obj_141_gamma_0_to_fp16 = const()[name = tensor("obj_141_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277773248)))]; + tensor obj_141_beta_0_to_fp16 = const()[name = tensor("obj_141_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277775872)))]; + tensor obj_141_epsilon_0_to_fp16 = const()[name = tensor("obj_141_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_141_cast_fp16 = batch_norm(beta = obj_141_beta_0_to_fp16, epsilon = obj_141_epsilon_0_to_fp16, gamma = obj_141_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor("obj_141_cast_fp16")]; + tensor var_4161 = const()[name = tensor("op_4161"), val = tensor([1, 1])]; + tensor var_4163 = const()[name = tensor("op_4163"), val = tensor([1, 1])]; + tensor pretrained_out_201_pad_type_0 = const()[name = tensor("pretrained_out_201_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_201_pad_0 = const()[name = tensor("pretrained_out_201_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277778496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278597760))), name = tensor("layers_10_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_10_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278597888)))]; + tensor pretrained_out_201_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4163, groups = var_4123, pad = pretrained_out_201_pad_0, pad_type = pretrained_out_201_pad_type_0, strides = var_4161, weight = layers_10_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_141_cast_fp16)[name = tensor("pretrained_out_201_cast_fp16")]; + tensor var_4167 = const()[name = tensor("op_4167"), val = tensor([1, 1])]; + tensor var_4169 = const()[name = tensor("op_4169"), val = tensor([1, 1])]; + tensor input_301_pad_type_0 = const()[name = tensor("input_301_pad_type_0"), val = tensor("custom")]; + tensor input_301_pad_0 = const()[name = tensor("input_301_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278600512)))]; + tensor input_301_cast_fp16 = conv(dilations = var_4169, groups = var_4123, pad = input_301_pad_0, pad_type = input_301_pad_type_0, strides = var_4167, weight = layers_10_self_attn_q_proj_loraA_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor var_4173 = const()[name = tensor("op_4173"), val = tensor([1, 1])]; + tensor var_4175 = const()[name = tensor("op_4175"), val = tensor([1, 1])]; + tensor lora_out_401_pad_type_0 = const()[name = tensor("lora_out_401_pad_type_0"), val = tensor("custom")]; + tensor lora_out_401_pad_0 = const()[name = tensor("lora_out_401_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_403_weight_0_to_fp16 = const()[name = tensor("lora_out_403_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278641536)))]; + tensor lora_out_403_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4175, groups = var_4123, pad = lora_out_401_pad_0, pad_type = lora_out_401_pad_type_0, strides = var_4173, weight = lora_out_403_weight_0_to_fp16, x = input_301_cast_fp16)[name = tensor("lora_out_403_cast_fp16")]; + tensor query_41_cast_fp16 = add(x = pretrained_out_201_cast_fp16, y = lora_out_403_cast_fp16)[name = tensor("query_41_cast_fp16")]; + tensor var_4185 = const()[name = tensor("op_4185"), val = tensor([1, 1])]; + tensor var_4187 = const()[name = tensor("op_4187"), val = tensor([1, 1])]; + tensor pretrained_out_203_pad_type_0 = const()[name = tensor("pretrained_out_203_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_203_pad_0 = const()[name = tensor("pretrained_out_203_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278682560))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279501824))), name = tensor("layers_10_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_203_cast_fp16 = conv(dilations = var_4187, groups = var_4123, pad = pretrained_out_203_pad_0, pad_type = pretrained_out_203_pad_type_0, strides = var_4185, weight = layers_10_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_141_cast_fp16)[name = tensor("pretrained_out_203_cast_fp16")]; + tensor var_4191 = const()[name = tensor("op_4191"), val = tensor([1, 1])]; + tensor var_4193 = const()[name = tensor("op_4193"), val = tensor([1, 1])]; + tensor input_303_pad_type_0 = const()[name = tensor("input_303_pad_type_0"), val = tensor("custom")]; + tensor input_303_pad_0 = const()[name = tensor("input_303_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279501952)))]; + tensor input_303_cast_fp16 = conv(dilations = var_4193, groups = var_4123, pad = input_303_pad_0, pad_type = input_303_pad_type_0, strides = var_4191, weight = layers_10_self_attn_k_proj_loraA_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("input_303_cast_fp16")]; + tensor var_4197 = const()[name = tensor("op_4197"), val = tensor([1, 1])]; + tensor var_4199 = const()[name = tensor("op_4199"), val = tensor([1, 1])]; + tensor lora_out_405_pad_type_0 = const()[name = tensor("lora_out_405_pad_type_0"), val = tensor("custom")]; + tensor lora_out_405_pad_0 = const()[name = tensor("lora_out_405_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_407_weight_0_to_fp16 = const()[name = tensor("lora_out_407_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279542976)))]; + tensor lora_out_407_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4199, groups = var_4123, pad = lora_out_405_pad_0, pad_type = lora_out_405_pad_type_0, strides = var_4197, weight = lora_out_407_weight_0_to_fp16, x = input_303_cast_fp16)[name = tensor("lora_out_407_cast_fp16")]; + tensor current_key_21_cast_fp16 = add(x = pretrained_out_203_cast_fp16, y = lora_out_407_cast_fp16)[name = tensor("current_key_21_cast_fp16")]; + tensor var_4210 = const()[name = tensor("op_4210"), val = tensor([1, 1])]; + tensor var_4212 = const()[name = tensor("op_4212"), val = tensor([1, 1])]; + tensor pretrained_out_205_pad_type_0 = const()[name = tensor("pretrained_out_205_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_205_pad_0 = const()[name = tensor("pretrained_out_205_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279584000))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280403264))), name = tensor("layers_10_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_10_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280403392)))]; + tensor pretrained_out_205_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_4212, groups = var_4123, pad = pretrained_out_205_pad_0, pad_type = pretrained_out_205_pad_type_0, strides = var_4210, weight = layers_10_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_141_cast_fp16)[name = tensor("pretrained_out_205_cast_fp16")]; + tensor var_4216 = const()[name = tensor("op_4216"), val = tensor([1, 1])]; + tensor var_4218 = const()[name = tensor("op_4218"), val = tensor([1, 1])]; + tensor input_305_pad_type_0 = const()[name = tensor("input_305_pad_type_0"), val = tensor("custom")]; + tensor input_305_pad_0 = const()[name = tensor("input_305_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280406016)))]; + tensor input_305_cast_fp16 = conv(dilations = var_4218, groups = var_4123, pad = input_305_pad_0, pad_type = input_305_pad_type_0, strides = var_4216, weight = layers_10_self_attn_v_proj_loraA_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("input_305_cast_fp16")]; + tensor var_4222 = const()[name = tensor("op_4222"), val = tensor([1, 1])]; + tensor var_4224 = const()[name = tensor("op_4224"), val = tensor([1, 1])]; + tensor lora_out_409_pad_type_0 = const()[name = tensor("lora_out_409_pad_type_0"), val = tensor("custom")]; + tensor lora_out_409_pad_0 = const()[name = tensor("lora_out_409_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_411_weight_0_to_fp16 = const()[name = tensor("lora_out_411_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280447040)))]; + tensor lora_out_411_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4224, groups = var_4123, pad = lora_out_409_pad_0, pad_type = lora_out_409_pad_type_0, strides = var_4222, weight = lora_out_411_weight_0_to_fp16, x = input_305_cast_fp16)[name = tensor("lora_out_411_cast_fp16")]; + tensor current_value_21_cast_fp16 = add(x = pretrained_out_205_cast_fp16, y = lora_out_411_cast_fp16)[name = tensor("current_value_21_cast_fp16")]; + tensor var_4234_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_4234_cast_fp16")]; + tensor var_4236_cast_fp16 = mul(x = var_103_cast_fp16_10, y = var_295_cast_fp16)[name = tensor("op_4236_cast_fp16")]; + tensor key_41_cast_fp16 = add(x = var_4234_cast_fp16, y = var_4236_cast_fp16)[name = tensor("key_41_cast_fp16")]; + tensor var_4238_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_4238_cast_fp16")]; + tensor var_4240_cast_fp16 = mul(x = var_138_cast_fp16_10, y = var_295_cast_fp16)[name = tensor("op_4240_cast_fp16")]; + tensor value_41_cast_fp16 = add(x = var_4238_cast_fp16, y = var_4240_cast_fp16)[name = tensor("value_41_cast_fp16")]; + tensor var_4243 = const()[name = tensor("op_4243"), val = tensor([1, 20, 64, -1])]; + tensor var_4244_cast_fp16 = reshape(shape = var_4243, x = query_41_cast_fp16)[name = tensor("op_4244_cast_fp16")]; + tensor var_4245_to_fp16 = const()[name = tensor("op_4245_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4246_cast_fp16 = mul(x = var_4244_cast_fp16, y = var_4245_to_fp16)[name = tensor("op_4246_cast_fp16")]; + tensor var_4247 = const()[name = tensor("op_4247"), val = tensor([1, 20, 64, -1])]; + tensor var_4248_cast_fp16 = reshape(shape = var_4247, x = key_41_cast_fp16)[name = tensor("op_4248_cast_fp16")]; + tensor mh_w_61_transpose_x_0 = const()[name = tensor("mh_w_61_transpose_x_0"), val = tensor(true)]; + tensor mh_w_61_transpose_y_0 = const()[name = tensor("mh_w_61_transpose_y_0"), val = tensor(false)]; + tensor mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_4246_cast_fp16, y = var_4248_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; + tensor mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_63_cast_fp16")]; + tensor var_4256_cast_fp16 = softmax(axis = var_4116, x = mh_w_63_cast_fp16)[name = tensor("op_4256_cast_fp16")]; + tensor var_4257 = const()[name = tensor("op_4257"), val = tensor([1, 20, 64, -1])]; + tensor var_4258_cast_fp16 = reshape(shape = var_4257, x = value_41_cast_fp16)[name = tensor("op_4258_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(true)]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_4258_cast_fp16, y = var_4256_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_4261 = const()[name = tensor("op_4261"), val = tensor([1, 1280, 1, -1])]; + tensor input_307_cast_fp16 = reshape(shape = var_4261, x = attn_41_cast_fp16)[name = tensor("input_307_cast_fp16")]; + tensor var_4268 = const()[name = tensor("op_4268"), val = tensor([1, 1])]; + tensor var_4270 = const()[name = tensor("op_4270"), val = tensor([1, 1])]; + tensor pretrained_out_207_pad_type_0 = const()[name = tensor("pretrained_out_207_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_207_pad_0 = const()[name = tensor("pretrained_out_207_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280488064))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281307328))), name = tensor("layers_10_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_10_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281307456)))]; + tensor pretrained_out_207_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4270, groups = var_4123, pad = pretrained_out_207_pad_0, pad_type = pretrained_out_207_pad_type_0, strides = var_4268, weight = layers_10_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_307_cast_fp16)[name = tensor("pretrained_out_207_cast_fp16")]; + tensor var_4274 = const()[name = tensor("op_4274"), val = tensor([1, 1])]; + tensor var_4276 = const()[name = tensor("op_4276"), val = tensor([1, 1])]; + tensor input_309_pad_type_0 = const()[name = tensor("input_309_pad_type_0"), val = tensor("custom")]; + tensor input_309_pad_0 = const()[name = tensor("input_309_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281310080)))]; + tensor input_309_cast_fp16 = conv(dilations = var_4276, groups = var_4123, pad = input_309_pad_0, pad_type = input_309_pad_type_0, strides = var_4274, weight = layers_10_self_attn_o_proj_loraA_weight_to_fp16, x = input_307_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor var_4280 = const()[name = tensor("op_4280"), val = tensor([1, 1])]; + tensor var_4282 = const()[name = tensor("op_4282"), val = tensor([1, 1])]; + tensor lora_out_413_pad_type_0 = const()[name = tensor("lora_out_413_pad_type_0"), val = tensor("custom")]; + tensor lora_out_413_pad_0 = const()[name = tensor("lora_out_413_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_415_weight_0_to_fp16 = const()[name = tensor("lora_out_415_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281351104)))]; + tensor lora_out_415_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4282, groups = var_4123, pad = lora_out_413_pad_0, pad_type = lora_out_413_pad_type_0, strides = var_4280, weight = lora_out_415_weight_0_to_fp16, x = input_309_cast_fp16)[name = tensor("lora_out_415_cast_fp16")]; + tensor obj_147_cast_fp16 = add(x = pretrained_out_207_cast_fp16, y = lora_out_415_cast_fp16)[name = tensor("obj_147_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_147_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor var_4295 = const()[name = tensor("op_4295"), val = tensor([1])]; + tensor channels_mean_63_cast_fp16 = reduce_mean(axes = var_4295, keep_dims = var_4124, x = inputs_63_cast_fp16)[name = tensor("channels_mean_63_cast_fp16")]; + tensor zero_mean_63_cast_fp16 = sub(x = inputs_63_cast_fp16, y = channels_mean_63_cast_fp16)[name = tensor("zero_mean_63_cast_fp16")]; + tensor zero_mean_sq_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = zero_mean_63_cast_fp16)[name = tensor("zero_mean_sq_63_cast_fp16")]; + tensor var_4299 = const()[name = tensor("op_4299"), val = tensor([1])]; + tensor var_4300_cast_fp16 = reduce_mean(axes = var_4299, keep_dims = var_4124, x = zero_mean_sq_63_cast_fp16)[name = tensor("op_4300_cast_fp16")]; + tensor var_4301_to_fp16 = const()[name = tensor("op_4301_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4302_cast_fp16 = add(x = var_4300_cast_fp16, y = var_4301_to_fp16)[name = tensor("op_4302_cast_fp16")]; + tensor denom_63_epsilon_0 = const()[name = tensor("denom_63_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_63_cast_fp16 = rsqrt(epsilon = denom_63_epsilon_0, x = var_4302_cast_fp16)[name = tensor("denom_63_cast_fp16")]; + tensor out_63_cast_fp16 = mul(x = zero_mean_63_cast_fp16, y = denom_63_cast_fp16)[name = tensor("out_63_cast_fp16")]; + tensor obj_149_gamma_0_to_fp16 = const()[name = tensor("obj_149_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281392128)))]; + tensor obj_149_beta_0_to_fp16 = const()[name = tensor("obj_149_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281394752)))]; + tensor obj_149_epsilon_0_to_fp16 = const()[name = tensor("obj_149_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_149_cast_fp16 = batch_norm(beta = obj_149_beta_0_to_fp16, epsilon = obj_149_epsilon_0_to_fp16, gamma = obj_149_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor("obj_149_cast_fp16")]; + tensor var_4320 = const()[name = tensor("op_4320"), val = tensor([1, 1])]; + tensor var_4322 = const()[name = tensor("op_4322"), val = tensor([1, 1])]; + tensor pretrained_out_209_pad_type_0 = const()[name = tensor("pretrained_out_209_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_209_pad_0 = const()[name = tensor("pretrained_out_209_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281397376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282216640))), name = tensor("layers_10_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_10_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282216768)))]; + tensor pretrained_out_209_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4322, groups = var_4123, pad = pretrained_out_209_pad_0, pad_type = pretrained_out_209_pad_type_0, strides = var_4320, weight = layers_10_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_149_cast_fp16)[name = tensor("pretrained_out_209_cast_fp16")]; + tensor var_4326 = const()[name = tensor("op_4326"), val = tensor([1, 1])]; + tensor var_4328 = const()[name = tensor("op_4328"), val = tensor([1, 1])]; + tensor input_311_pad_type_0 = const()[name = tensor("input_311_pad_type_0"), val = tensor("custom")]; + tensor input_311_pad_0 = const()[name = tensor("input_311_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282219392)))]; + tensor input_311_cast_fp16 = conv(dilations = var_4328, groups = var_4123, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = var_4326, weight = layers_10_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_149_cast_fp16)[name = tensor("input_311_cast_fp16")]; + tensor var_4332 = const()[name = tensor("op_4332"), val = tensor([1, 1])]; + tensor var_4334 = const()[name = tensor("op_4334"), val = tensor([1, 1])]; + tensor lora_out_417_pad_type_0 = const()[name = tensor("lora_out_417_pad_type_0"), val = tensor("custom")]; + tensor lora_out_417_pad_0 = const()[name = tensor("lora_out_417_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_419_weight_0_to_fp16 = const()[name = tensor("lora_out_419_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282260416)))]; + tensor lora_out_419_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4334, groups = var_4123, pad = lora_out_417_pad_0, pad_type = lora_out_417_pad_type_0, strides = var_4332, weight = lora_out_419_weight_0_to_fp16, x = input_311_cast_fp16)[name = tensor("lora_out_419_cast_fp16")]; + tensor query_43_cast_fp16 = add(x = pretrained_out_209_cast_fp16, y = lora_out_419_cast_fp16)[name = tensor("query_43_cast_fp16")]; + tensor var_4344 = const()[name = tensor("op_4344"), val = tensor([1, 1])]; + tensor var_4346 = const()[name = tensor("op_4346"), val = tensor([1, 1])]; + tensor pretrained_out_211_pad_type_0 = const()[name = tensor("pretrained_out_211_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_211_pad_0 = const()[name = tensor("pretrained_out_211_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282301440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283120704))), name = tensor("layers_10_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_211_cast_fp16 = conv(dilations = var_4346, groups = var_4123, pad = pretrained_out_211_pad_0, pad_type = pretrained_out_211_pad_type_0, strides = var_4344, weight = layers_10_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_211_cast_fp16")]; + tensor var_4350 = const()[name = tensor("op_4350"), val = tensor([1, 1])]; + tensor var_4352 = const()[name = tensor("op_4352"), val = tensor([1, 1])]; + tensor input_313_pad_type_0 = const()[name = tensor("input_313_pad_type_0"), val = tensor("custom")]; + tensor input_313_pad_0 = const()[name = tensor("input_313_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283120832)))]; + tensor input_313_cast_fp16 = conv(dilations = var_4352, groups = var_4123, pad = input_313_pad_0, pad_type = input_313_pad_type_0, strides = var_4350, weight = layers_10_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_313_cast_fp16")]; + tensor var_4356 = const()[name = tensor("op_4356"), val = tensor([1, 1])]; + tensor var_4358 = const()[name = tensor("op_4358"), val = tensor([1, 1])]; + tensor lora_out_421_pad_type_0 = const()[name = tensor("lora_out_421_pad_type_0"), val = tensor("custom")]; + tensor lora_out_421_pad_0 = const()[name = tensor("lora_out_421_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_423_weight_0_to_fp16 = const()[name = tensor("lora_out_423_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283161856)))]; + tensor lora_out_423_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4358, groups = var_4123, pad = lora_out_421_pad_0, pad_type = lora_out_421_pad_type_0, strides = var_4356, weight = lora_out_423_weight_0_to_fp16, x = input_313_cast_fp16)[name = tensor("lora_out_423_cast_fp16")]; + tensor key_43_cast_fp16 = add(x = pretrained_out_211_cast_fp16, y = lora_out_423_cast_fp16)[name = tensor("key_43_cast_fp16")]; + tensor var_4369 = const()[name = tensor("op_4369"), val = tensor([1, 1])]; + tensor var_4371 = const()[name = tensor("op_4371"), val = tensor([1, 1])]; + tensor pretrained_out_213_pad_type_0 = const()[name = tensor("pretrained_out_213_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_213_pad_0 = const()[name = tensor("pretrained_out_213_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283202880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284022144))), name = tensor("layers_10_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_10_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284022272)))]; + tensor pretrained_out_213_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_4371, groups = var_4123, pad = pretrained_out_213_pad_0, pad_type = pretrained_out_213_pad_type_0, strides = var_4369, weight = layers_10_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_213_cast_fp16")]; + tensor var_4375 = const()[name = tensor("op_4375"), val = tensor([1, 1])]; + tensor var_4377 = const()[name = tensor("op_4377"), val = tensor([1, 1])]; + tensor input_315_pad_type_0 = const()[name = tensor("input_315_pad_type_0"), val = tensor("custom")]; + tensor input_315_pad_0 = const()[name = tensor("input_315_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284024896)))]; + tensor input_315_cast_fp16 = conv(dilations = var_4377, groups = var_4123, pad = input_315_pad_0, pad_type = input_315_pad_type_0, strides = var_4375, weight = layers_10_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_315_cast_fp16")]; + tensor var_4381 = const()[name = tensor("op_4381"), val = tensor([1, 1])]; + tensor var_4383 = const()[name = tensor("op_4383"), val = tensor([1, 1])]; + tensor lora_out_425_pad_type_0 = const()[name = tensor("lora_out_425_pad_type_0"), val = tensor("custom")]; + tensor lora_out_425_pad_0 = const()[name = tensor("lora_out_425_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_427_weight_0_to_fp16 = const()[name = tensor("lora_out_427_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284065920)))]; + tensor lora_out_427_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4383, groups = var_4123, pad = lora_out_425_pad_0, pad_type = lora_out_425_pad_type_0, strides = var_4381, weight = lora_out_427_weight_0_to_fp16, x = input_315_cast_fp16)[name = tensor("lora_out_427_cast_fp16")]; + tensor value_43_cast_fp16 = add(x = pretrained_out_213_cast_fp16, y = lora_out_427_cast_fp16)[name = tensor("value_43_cast_fp16")]; + tensor var_4390 = const()[name = tensor("op_4390"), val = tensor([1, 20, 64, -1])]; + tensor var_4391_cast_fp16 = reshape(shape = var_4390, x = query_43_cast_fp16)[name = tensor("op_4391_cast_fp16")]; + tensor var_4392_to_fp16 = const()[name = tensor("op_4392_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4393_cast_fp16 = mul(x = var_4391_cast_fp16, y = var_4392_to_fp16)[name = tensor("op_4393_cast_fp16")]; + tensor var_4394 = const()[name = tensor("op_4394"), val = tensor([1, 20, 64, -1])]; + tensor var_4395_cast_fp16 = reshape(shape = var_4394, x = key_43_cast_fp16)[name = tensor("op_4395_cast_fp16")]; + tensor mh_w_65_transpose_x_0 = const()[name = tensor("mh_w_65_transpose_x_0"), val = tensor(true)]; + tensor mh_w_65_transpose_y_0 = const()[name = tensor("mh_w_65_transpose_y_0"), val = tensor(false)]; + tensor mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_4393_cast_fp16, y = var_4395_cast_fp16)[name = tensor("mh_w_65_cast_fp16")]; + tensor obj_153_cast_fp16 = softmax(axis = var_4116, x = mh_w_65_cast_fp16)[name = tensor("obj_153_cast_fp16")]; + tensor var_4399 = const()[name = tensor("op_4399"), val = tensor([1, 20, 64, -1])]; + tensor var_4400_cast_fp16 = reshape(shape = var_4399, x = value_43_cast_fp16)[name = tensor("op_4400_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(true)]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_4400_cast_fp16, y = obj_153_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_4403 = const()[name = tensor("op_4403"), val = tensor([1, 1280, 1, -1])]; + tensor input_317_cast_fp16 = reshape(shape = var_4403, x = attn_43_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor var_4410 = const()[name = tensor("op_4410"), val = tensor([1, 1])]; + tensor var_4412 = const()[name = tensor("op_4412"), val = tensor([1, 1])]; + tensor pretrained_out_215_pad_type_0 = const()[name = tensor("pretrained_out_215_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_215_pad_0 = const()[name = tensor("pretrained_out_215_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284106944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284926208))), name = tensor("layers_10_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_10_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284926336)))]; + tensor pretrained_out_215_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4412, groups = var_4123, pad = pretrained_out_215_pad_0, pad_type = pretrained_out_215_pad_type_0, strides = var_4410, weight = layers_10_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_317_cast_fp16)[name = tensor("pretrained_out_215_cast_fp16")]; + tensor var_4416 = const()[name = tensor("op_4416"), val = tensor([1, 1])]; + tensor var_4418 = const()[name = tensor("op_4418"), val = tensor([1, 1])]; + tensor input_319_pad_type_0 = const()[name = tensor("input_319_pad_type_0"), val = tensor("custom")]; + tensor input_319_pad_0 = const()[name = tensor("input_319_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284928960)))]; + tensor input_319_cast_fp16 = conv(dilations = var_4418, groups = var_4123, pad = input_319_pad_0, pad_type = input_319_pad_type_0, strides = var_4416, weight = layers_10_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("input_319_cast_fp16")]; + tensor var_4422 = const()[name = tensor("op_4422"), val = tensor([1, 1])]; + tensor var_4424 = const()[name = tensor("op_4424"), val = tensor([1, 1])]; + tensor lora_out_429_pad_type_0 = const()[name = tensor("lora_out_429_pad_type_0"), val = tensor("custom")]; + tensor lora_out_429_pad_0 = const()[name = tensor("lora_out_429_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_431_weight_0_to_fp16 = const()[name = tensor("lora_out_431_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284969984)))]; + tensor lora_out_431_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4424, groups = var_4123, pad = lora_out_429_pad_0, pad_type = lora_out_429_pad_type_0, strides = var_4422, weight = lora_out_431_weight_0_to_fp16, x = input_319_cast_fp16)[name = tensor("lora_out_431_cast_fp16")]; + tensor obj_151_cast_fp16 = add(x = pretrained_out_215_cast_fp16, y = lora_out_431_cast_fp16)[name = tensor("obj_151_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_151_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor var_4436 = const()[name = tensor("op_4436"), val = tensor([1])]; + tensor channels_mean_65_cast_fp16 = reduce_mean(axes = var_4436, keep_dims = var_4124, x = inputs_65_cast_fp16)[name = tensor("channels_mean_65_cast_fp16")]; + tensor zero_mean_65_cast_fp16 = sub(x = inputs_65_cast_fp16, y = channels_mean_65_cast_fp16)[name = tensor("zero_mean_65_cast_fp16")]; + tensor zero_mean_sq_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = zero_mean_65_cast_fp16)[name = tensor("zero_mean_sq_65_cast_fp16")]; + tensor var_4440 = const()[name = tensor("op_4440"), val = tensor([1])]; + tensor var_4441_cast_fp16 = reduce_mean(axes = var_4440, keep_dims = var_4124, x = zero_mean_sq_65_cast_fp16)[name = tensor("op_4441_cast_fp16")]; + tensor var_4442_to_fp16 = const()[name = tensor("op_4442_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4443_cast_fp16 = add(x = var_4441_cast_fp16, y = var_4442_to_fp16)[name = tensor("op_4443_cast_fp16")]; + tensor denom_65_epsilon_0 = const()[name = tensor("denom_65_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_65_cast_fp16 = rsqrt(epsilon = denom_65_epsilon_0, x = var_4443_cast_fp16)[name = tensor("denom_65_cast_fp16")]; + tensor out_65_cast_fp16 = mul(x = zero_mean_65_cast_fp16, y = denom_65_cast_fp16)[name = tensor("out_65_cast_fp16")]; + tensor input_321_gamma_0_to_fp16 = const()[name = tensor("input_321_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285011008)))]; + tensor input_321_beta_0_to_fp16 = const()[name = tensor("input_321_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285013632)))]; + tensor input_321_epsilon_0_to_fp16 = const()[name = tensor("input_321_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_321_cast_fp16 = batch_norm(beta = input_321_beta_0_to_fp16, epsilon = input_321_epsilon_0_to_fp16, gamma = input_321_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor("input_321_cast_fp16")]; + tensor var_4457 = const()[name = tensor("op_4457"), val = tensor([1, 1])]; + tensor var_4459 = const()[name = tensor("op_4459"), val = tensor([1, 1])]; + tensor pretrained_out_217_pad_type_0 = const()[name = tensor("pretrained_out_217_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_217_pad_0 = const()[name = tensor("pretrained_out_217_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285016256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288293120))), name = tensor("layers_10_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_10_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288293248)))]; + tensor pretrained_out_217_cast_fp16 = conv(bias = layers_10_fc1_pretrained_bias_to_fp16, dilations = var_4459, groups = var_4123, pad = pretrained_out_217_pad_0, pad_type = pretrained_out_217_pad_type_0, strides = var_4457, weight = layers_10_fc1_pretrained_weight_to_fp16_palettized, x = input_321_cast_fp16)[name = tensor("pretrained_out_217_cast_fp16")]; + tensor var_4463 = const()[name = tensor("op_4463"), val = tensor([1, 1])]; + tensor var_4465 = const()[name = tensor("op_4465"), val = tensor([1, 1])]; + tensor input_323_pad_type_0 = const()[name = tensor("input_323_pad_type_0"), val = tensor("custom")]; + tensor input_323_pad_0 = const()[name = tensor("input_323_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_10_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288303552)))]; + tensor input_323_cast_fp16 = conv(dilations = var_4465, groups = var_4123, pad = input_323_pad_0, pad_type = input_323_pad_type_0, strides = var_4463, weight = layers_10_fc1_loraA_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("input_323_cast_fp16")]; + tensor var_4469 = const()[name = tensor("op_4469"), val = tensor([1, 1])]; + tensor var_4471 = const()[name = tensor("op_4471"), val = tensor([1, 1])]; + tensor lora_out_433_pad_type_0 = const()[name = tensor("lora_out_433_pad_type_0"), val = tensor("custom")]; + tensor lora_out_433_pad_0 = const()[name = tensor("lora_out_433_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_435_weight_0_to_fp16 = const()[name = tensor("lora_out_435_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288344576)))]; + tensor lora_out_435_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_4471, groups = var_4123, pad = lora_out_433_pad_0, pad_type = lora_out_433_pad_type_0, strides = var_4469, weight = lora_out_435_weight_0_to_fp16, x = input_323_cast_fp16)[name = tensor("lora_out_435_cast_fp16")]; + tensor input_325_cast_fp16 = add(x = pretrained_out_217_cast_fp16, y = lora_out_435_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor input_327_mode_0 = const()[name = tensor("input_327_mode_0"), val = tensor("EXACT")]; + tensor input_327_cast_fp16 = gelu(mode = input_327_mode_0, x = input_325_cast_fp16)[name = tensor("input_327_cast_fp16")]; + tensor var_4483 = const()[name = tensor("op_4483"), val = tensor([1, 1])]; + tensor var_4485 = const()[name = tensor("op_4485"), val = tensor([1, 1])]; + tensor pretrained_out_219_pad_type_0 = const()[name = tensor("pretrained_out_219_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_219_pad_0 = const()[name = tensor("pretrained_out_219_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288508480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291785344))), name = tensor("layers_10_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_10_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_10_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291785472)))]; + tensor pretrained_out_219_cast_fp16 = conv(bias = layers_10_fc2_pretrained_bias_to_fp16, dilations = var_4485, groups = var_4123, pad = pretrained_out_219_pad_0, pad_type = pretrained_out_219_pad_type_0, strides = var_4483, weight = layers_10_fc2_pretrained_weight_to_fp16_palettized, x = input_327_cast_fp16)[name = tensor("pretrained_out_219_cast_fp16")]; + tensor var_4489 = const()[name = tensor("op_4489"), val = tensor([1, 1])]; + tensor var_4491 = const()[name = tensor("op_4491"), val = tensor([1, 1])]; + tensor input_329_pad_type_0 = const()[name = tensor("input_329_pad_type_0"), val = tensor("custom")]; + tensor input_329_pad_0 = const()[name = tensor("input_329_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_10_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_10_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291788096)))]; + tensor input_329_cast_fp16 = conv(dilations = var_4491, groups = var_4123, pad = input_329_pad_0, pad_type = input_329_pad_type_0, strides = var_4489, weight = layers_10_fc2_loraA_weight_to_fp16, x = input_327_cast_fp16)[name = tensor("input_329_cast_fp16")]; + tensor var_4495 = const()[name = tensor("op_4495"), val = tensor([1, 1])]; + tensor var_4497 = const()[name = tensor("op_4497"), val = tensor([1, 1])]; + tensor lora_out_437_pad_type_0 = const()[name = tensor("lora_out_437_pad_type_0"), val = tensor("custom")]; + tensor lora_out_437_pad_0 = const()[name = tensor("lora_out_437_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_439_weight_0_to_fp16 = const()[name = tensor("lora_out_439_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291952000)))]; + tensor lora_out_439_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4497, groups = var_4123, pad = lora_out_437_pad_0, pad_type = lora_out_437_pad_type_0, strides = var_4495, weight = lora_out_439_weight_0_to_fp16, x = input_329_cast_fp16)[name = tensor("lora_out_439_cast_fp16")]; + tensor hidden_states_23_cast_fp16 = add(x = pretrained_out_219_cast_fp16, y = lora_out_439_cast_fp16)[name = tensor("hidden_states_23_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor var_4514 = const()[name = tensor("op_4514"), val = tensor(3)]; + tensor var_4521 = const()[name = tensor("op_4521"), val = tensor(1)]; + tensor var_4522 = const()[name = tensor("op_4522"), val = tensor(true)]; + tensor var_4534 = const()[name = tensor("op_4534"), val = tensor([1])]; + tensor channels_mean_67_cast_fp16 = reduce_mean(axes = var_4534, keep_dims = var_4522, x = inputs_67_cast_fp16)[name = tensor("channels_mean_67_cast_fp16")]; + tensor zero_mean_67_cast_fp16 = sub(x = inputs_67_cast_fp16, y = channels_mean_67_cast_fp16)[name = tensor("zero_mean_67_cast_fp16")]; + tensor zero_mean_sq_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = zero_mean_67_cast_fp16)[name = tensor("zero_mean_sq_67_cast_fp16")]; + tensor var_4538 = const()[name = tensor("op_4538"), val = tensor([1])]; + tensor var_4539_cast_fp16 = reduce_mean(axes = var_4538, keep_dims = var_4522, x = zero_mean_sq_67_cast_fp16)[name = tensor("op_4539_cast_fp16")]; + tensor var_4540_to_fp16 = const()[name = tensor("op_4540_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4541_cast_fp16 = add(x = var_4539_cast_fp16, y = var_4540_to_fp16)[name = tensor("op_4541_cast_fp16")]; + tensor denom_67_epsilon_0 = const()[name = tensor("denom_67_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_67_cast_fp16 = rsqrt(epsilon = denom_67_epsilon_0, x = var_4541_cast_fp16)[name = tensor("denom_67_cast_fp16")]; + tensor out_67_cast_fp16 = mul(x = zero_mean_67_cast_fp16, y = denom_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; + tensor obj_155_gamma_0_to_fp16 = const()[name = tensor("obj_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291993024)))]; + tensor obj_155_beta_0_to_fp16 = const()[name = tensor("obj_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291995648)))]; + tensor obj_155_epsilon_0_to_fp16 = const()[name = tensor("obj_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_155_cast_fp16 = batch_norm(beta = obj_155_beta_0_to_fp16, epsilon = obj_155_epsilon_0_to_fp16, gamma = obj_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor("obj_155_cast_fp16")]; + tensor var_4559 = const()[name = tensor("op_4559"), val = tensor([1, 1])]; + tensor var_4561 = const()[name = tensor("op_4561"), val = tensor([1, 1])]; + tensor pretrained_out_221_pad_type_0 = const()[name = tensor("pretrained_out_221_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_221_pad_0 = const()[name = tensor("pretrained_out_221_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291998272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292817536))), name = tensor("layers_11_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_11_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292817664)))]; + tensor pretrained_out_221_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4561, groups = var_4521, pad = pretrained_out_221_pad_0, pad_type = pretrained_out_221_pad_type_0, strides = var_4559, weight = layers_11_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_155_cast_fp16)[name = tensor("pretrained_out_221_cast_fp16")]; + tensor var_4565 = const()[name = tensor("op_4565"), val = tensor([1, 1])]; + tensor var_4567 = const()[name = tensor("op_4567"), val = tensor([1, 1])]; + tensor input_331_pad_type_0 = const()[name = tensor("input_331_pad_type_0"), val = tensor("custom")]; + tensor input_331_pad_0 = const()[name = tensor("input_331_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292820288)))]; + tensor input_331_cast_fp16 = conv(dilations = var_4567, groups = var_4521, pad = input_331_pad_0, pad_type = input_331_pad_type_0, strides = var_4565, weight = layers_11_self_attn_q_proj_loraA_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("input_331_cast_fp16")]; + tensor var_4571 = const()[name = tensor("op_4571"), val = tensor([1, 1])]; + tensor var_4573 = const()[name = tensor("op_4573"), val = tensor([1, 1])]; + tensor lora_out_441_pad_type_0 = const()[name = tensor("lora_out_441_pad_type_0"), val = tensor("custom")]; + tensor lora_out_441_pad_0 = const()[name = tensor("lora_out_441_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_443_weight_0_to_fp16 = const()[name = tensor("lora_out_443_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292861312)))]; + tensor lora_out_443_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4573, groups = var_4521, pad = lora_out_441_pad_0, pad_type = lora_out_441_pad_type_0, strides = var_4571, weight = lora_out_443_weight_0_to_fp16, x = input_331_cast_fp16)[name = tensor("lora_out_443_cast_fp16")]; + tensor query_45_cast_fp16 = add(x = pretrained_out_221_cast_fp16, y = lora_out_443_cast_fp16)[name = tensor("query_45_cast_fp16")]; + tensor var_4583 = const()[name = tensor("op_4583"), val = tensor([1, 1])]; + tensor var_4585 = const()[name = tensor("op_4585"), val = tensor([1, 1])]; + tensor pretrained_out_223_pad_type_0 = const()[name = tensor("pretrained_out_223_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_223_pad_0 = const()[name = tensor("pretrained_out_223_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292902336))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293721600))), name = tensor("layers_11_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_223_cast_fp16 = conv(dilations = var_4585, groups = var_4521, pad = pretrained_out_223_pad_0, pad_type = pretrained_out_223_pad_type_0, strides = var_4583, weight = layers_11_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_155_cast_fp16)[name = tensor("pretrained_out_223_cast_fp16")]; + tensor var_4589 = const()[name = tensor("op_4589"), val = tensor([1, 1])]; + tensor var_4591 = const()[name = tensor("op_4591"), val = tensor([1, 1])]; + tensor input_333_pad_type_0 = const()[name = tensor("input_333_pad_type_0"), val = tensor("custom")]; + tensor input_333_pad_0 = const()[name = tensor("input_333_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293721728)))]; + tensor input_333_cast_fp16 = conv(dilations = var_4591, groups = var_4521, pad = input_333_pad_0, pad_type = input_333_pad_type_0, strides = var_4589, weight = layers_11_self_attn_k_proj_loraA_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor var_4595 = const()[name = tensor("op_4595"), val = tensor([1, 1])]; + tensor var_4597 = const()[name = tensor("op_4597"), val = tensor([1, 1])]; + tensor lora_out_445_pad_type_0 = const()[name = tensor("lora_out_445_pad_type_0"), val = tensor("custom")]; + tensor lora_out_445_pad_0 = const()[name = tensor("lora_out_445_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_447_weight_0_to_fp16 = const()[name = tensor("lora_out_447_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293762752)))]; + tensor lora_out_447_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4597, groups = var_4521, pad = lora_out_445_pad_0, pad_type = lora_out_445_pad_type_0, strides = var_4595, weight = lora_out_447_weight_0_to_fp16, x = input_333_cast_fp16)[name = tensor("lora_out_447_cast_fp16")]; + tensor current_key_23_cast_fp16 = add(x = pretrained_out_223_cast_fp16, y = lora_out_447_cast_fp16)[name = tensor("current_key_23_cast_fp16")]; + tensor var_4608 = const()[name = tensor("op_4608"), val = tensor([1, 1])]; + tensor var_4610 = const()[name = tensor("op_4610"), val = tensor([1, 1])]; + tensor pretrained_out_225_pad_type_0 = const()[name = tensor("pretrained_out_225_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_225_pad_0 = const()[name = tensor("pretrained_out_225_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293803776))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294623040))), name = tensor("layers_11_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_11_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294623168)))]; + tensor pretrained_out_225_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_4610, groups = var_4521, pad = pretrained_out_225_pad_0, pad_type = pretrained_out_225_pad_type_0, strides = var_4608, weight = layers_11_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_155_cast_fp16)[name = tensor("pretrained_out_225_cast_fp16")]; + tensor var_4614 = const()[name = tensor("op_4614"), val = tensor([1, 1])]; + tensor var_4616 = const()[name = tensor("op_4616"), val = tensor([1, 1])]; + tensor input_335_pad_type_0 = const()[name = tensor("input_335_pad_type_0"), val = tensor("custom")]; + tensor input_335_pad_0 = const()[name = tensor("input_335_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294625792)))]; + tensor input_335_cast_fp16 = conv(dilations = var_4616, groups = var_4521, pad = input_335_pad_0, pad_type = input_335_pad_type_0, strides = var_4614, weight = layers_11_self_attn_v_proj_loraA_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("input_335_cast_fp16")]; + tensor var_4620 = const()[name = tensor("op_4620"), val = tensor([1, 1])]; + tensor var_4622 = const()[name = tensor("op_4622"), val = tensor([1, 1])]; + tensor lora_out_449_pad_type_0 = const()[name = tensor("lora_out_449_pad_type_0"), val = tensor("custom")]; + tensor lora_out_449_pad_0 = const()[name = tensor("lora_out_449_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_451_weight_0_to_fp16 = const()[name = tensor("lora_out_451_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294666816)))]; + tensor lora_out_451_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4622, groups = var_4521, pad = lora_out_449_pad_0, pad_type = lora_out_449_pad_type_0, strides = var_4620, weight = lora_out_451_weight_0_to_fp16, x = input_335_cast_fp16)[name = tensor("lora_out_451_cast_fp16")]; + tensor current_value_23_cast_fp16 = add(x = pretrained_out_225_cast_fp16, y = lora_out_451_cast_fp16)[name = tensor("current_value_23_cast_fp16")]; + tensor var_4632_cast_fp16 = mul(x = current_key_23_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_4632_cast_fp16")]; + tensor var_4634_cast_fp16 = mul(x = var_103_cast_fp16_11, y = var_295_cast_fp16)[name = tensor("op_4634_cast_fp16")]; + tensor key_45_cast_fp16 = add(x = var_4632_cast_fp16, y = var_4634_cast_fp16)[name = tensor("key_45_cast_fp16")]; + tensor var_4636_cast_fp16 = mul(x = current_value_23_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_4636_cast_fp16")]; + tensor var_4638_cast_fp16 = mul(x = var_138_cast_fp16_11, y = var_295_cast_fp16)[name = tensor("op_4638_cast_fp16")]; + tensor value_45_cast_fp16 = add(x = var_4636_cast_fp16, y = var_4638_cast_fp16)[name = tensor("value_45_cast_fp16")]; + tensor var_4641 = const()[name = tensor("op_4641"), val = tensor([1, 20, 64, -1])]; + tensor var_4642_cast_fp16 = reshape(shape = var_4641, x = query_45_cast_fp16)[name = tensor("op_4642_cast_fp16")]; + tensor var_4643_to_fp16 = const()[name = tensor("op_4643_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4644_cast_fp16 = mul(x = var_4642_cast_fp16, y = var_4643_to_fp16)[name = tensor("op_4644_cast_fp16")]; + tensor var_4645 = const()[name = tensor("op_4645"), val = tensor([1, 20, 64, -1])]; + tensor var_4646_cast_fp16 = reshape(shape = var_4645, x = key_45_cast_fp16)[name = tensor("op_4646_cast_fp16")]; + tensor mh_w_67_transpose_x_0 = const()[name = tensor("mh_w_67_transpose_x_0"), val = tensor(true)]; + tensor mh_w_67_transpose_y_0 = const()[name = tensor("mh_w_67_transpose_y_0"), val = tensor(false)]; + tensor mh_w_67_cast_fp16 = matmul(transpose_x = mh_w_67_transpose_x_0, transpose_y = mh_w_67_transpose_y_0, x = var_4644_cast_fp16, y = var_4646_cast_fp16)[name = tensor("mh_w_67_cast_fp16")]; + tensor mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_69_cast_fp16")]; + tensor var_4654_cast_fp16 = softmax(axis = var_4514, x = mh_w_69_cast_fp16)[name = tensor("op_4654_cast_fp16")]; + tensor var_4655 = const()[name = tensor("op_4655"), val = tensor([1, 20, 64, -1])]; + tensor var_4656_cast_fp16 = reshape(shape = var_4655, x = value_45_cast_fp16)[name = tensor("op_4656_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(true)]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_4656_cast_fp16, y = var_4654_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_4659 = const()[name = tensor("op_4659"), val = tensor([1, 1280, 1, -1])]; + tensor input_337_cast_fp16 = reshape(shape = var_4659, x = attn_45_cast_fp16)[name = tensor("input_337_cast_fp16")]; + tensor var_4666 = const()[name = tensor("op_4666"), val = tensor([1, 1])]; + tensor var_4668 = const()[name = tensor("op_4668"), val = tensor([1, 1])]; + tensor pretrained_out_227_pad_type_0 = const()[name = tensor("pretrained_out_227_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_227_pad_0 = const()[name = tensor("pretrained_out_227_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294707840))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295527104))), name = tensor("layers_11_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_11_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295527232)))]; + tensor pretrained_out_227_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4668, groups = var_4521, pad = pretrained_out_227_pad_0, pad_type = pretrained_out_227_pad_type_0, strides = var_4666, weight = layers_11_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_337_cast_fp16)[name = tensor("pretrained_out_227_cast_fp16")]; + tensor var_4672 = const()[name = tensor("op_4672"), val = tensor([1, 1])]; + tensor var_4674 = const()[name = tensor("op_4674"), val = tensor([1, 1])]; + tensor input_339_pad_type_0 = const()[name = tensor("input_339_pad_type_0"), val = tensor("custom")]; + tensor input_339_pad_0 = const()[name = tensor("input_339_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295529856)))]; + tensor input_339_cast_fp16 = conv(dilations = var_4674, groups = var_4521, pad = input_339_pad_0, pad_type = input_339_pad_type_0, strides = var_4672, weight = layers_11_self_attn_o_proj_loraA_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor var_4678 = const()[name = tensor("op_4678"), val = tensor([1, 1])]; + tensor var_4680 = const()[name = tensor("op_4680"), val = tensor([1, 1])]; + tensor lora_out_453_pad_type_0 = const()[name = tensor("lora_out_453_pad_type_0"), val = tensor("custom")]; + tensor lora_out_453_pad_0 = const()[name = tensor("lora_out_453_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_455_weight_0_to_fp16 = const()[name = tensor("lora_out_455_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295570880)))]; + tensor lora_out_455_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4680, groups = var_4521, pad = lora_out_453_pad_0, pad_type = lora_out_453_pad_type_0, strides = var_4678, weight = lora_out_455_weight_0_to_fp16, x = input_339_cast_fp16)[name = tensor("lora_out_455_cast_fp16")]; + tensor obj_161_cast_fp16 = add(x = pretrained_out_227_cast_fp16, y = lora_out_455_cast_fp16)[name = tensor("obj_161_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_161_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor var_4693 = const()[name = tensor("op_4693"), val = tensor([1])]; + tensor channels_mean_69_cast_fp16 = reduce_mean(axes = var_4693, keep_dims = var_4522, x = inputs_69_cast_fp16)[name = tensor("channels_mean_69_cast_fp16")]; + tensor zero_mean_69_cast_fp16 = sub(x = inputs_69_cast_fp16, y = channels_mean_69_cast_fp16)[name = tensor("zero_mean_69_cast_fp16")]; + tensor zero_mean_sq_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = zero_mean_69_cast_fp16)[name = tensor("zero_mean_sq_69_cast_fp16")]; + tensor var_4697 = const()[name = tensor("op_4697"), val = tensor([1])]; + tensor var_4698_cast_fp16 = reduce_mean(axes = var_4697, keep_dims = var_4522, x = zero_mean_sq_69_cast_fp16)[name = tensor("op_4698_cast_fp16")]; + tensor var_4699_to_fp16 = const()[name = tensor("op_4699_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4700_cast_fp16 = add(x = var_4698_cast_fp16, y = var_4699_to_fp16)[name = tensor("op_4700_cast_fp16")]; + tensor denom_69_epsilon_0 = const()[name = tensor("denom_69_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_69_cast_fp16 = rsqrt(epsilon = denom_69_epsilon_0, x = var_4700_cast_fp16)[name = tensor("denom_69_cast_fp16")]; + tensor out_69_cast_fp16 = mul(x = zero_mean_69_cast_fp16, y = denom_69_cast_fp16)[name = tensor("out_69_cast_fp16")]; + tensor obj_163_gamma_0_to_fp16 = const()[name = tensor("obj_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295611904)))]; + tensor obj_163_beta_0_to_fp16 = const()[name = tensor("obj_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295614528)))]; + tensor obj_163_epsilon_0_to_fp16 = const()[name = tensor("obj_163_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_163_cast_fp16 = batch_norm(beta = obj_163_beta_0_to_fp16, epsilon = obj_163_epsilon_0_to_fp16, gamma = obj_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor("obj_163_cast_fp16")]; + tensor var_4718 = const()[name = tensor("op_4718"), val = tensor([1, 1])]; + tensor var_4720 = const()[name = tensor("op_4720"), val = tensor([1, 1])]; + tensor pretrained_out_229_pad_type_0 = const()[name = tensor("pretrained_out_229_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_229_pad_0 = const()[name = tensor("pretrained_out_229_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295617152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296436416))), name = tensor("layers_11_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_11_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296436544)))]; + tensor pretrained_out_229_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4720, groups = var_4521, pad = pretrained_out_229_pad_0, pad_type = pretrained_out_229_pad_type_0, strides = var_4718, weight = layers_11_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_163_cast_fp16)[name = tensor("pretrained_out_229_cast_fp16")]; + tensor var_4724 = const()[name = tensor("op_4724"), val = tensor([1, 1])]; + tensor var_4726 = const()[name = tensor("op_4726"), val = tensor([1, 1])]; + tensor input_341_pad_type_0 = const()[name = tensor("input_341_pad_type_0"), val = tensor("custom")]; + tensor input_341_pad_0 = const()[name = tensor("input_341_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296439168)))]; + tensor input_341_cast_fp16 = conv(dilations = var_4726, groups = var_4521, pad = input_341_pad_0, pad_type = input_341_pad_type_0, strides = var_4724, weight = layers_11_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_163_cast_fp16)[name = tensor("input_341_cast_fp16")]; + tensor var_4730 = const()[name = tensor("op_4730"), val = tensor([1, 1])]; + tensor var_4732 = const()[name = tensor("op_4732"), val = tensor([1, 1])]; + tensor lora_out_457_pad_type_0 = const()[name = tensor("lora_out_457_pad_type_0"), val = tensor("custom")]; + tensor lora_out_457_pad_0 = const()[name = tensor("lora_out_457_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_459_weight_0_to_fp16 = const()[name = tensor("lora_out_459_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296480192)))]; + tensor lora_out_459_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4732, groups = var_4521, pad = lora_out_457_pad_0, pad_type = lora_out_457_pad_type_0, strides = var_4730, weight = lora_out_459_weight_0_to_fp16, x = input_341_cast_fp16)[name = tensor("lora_out_459_cast_fp16")]; + tensor query_47_cast_fp16 = add(x = pretrained_out_229_cast_fp16, y = lora_out_459_cast_fp16)[name = tensor("query_47_cast_fp16")]; + tensor var_4742 = const()[name = tensor("op_4742"), val = tensor([1, 1])]; + tensor var_4744 = const()[name = tensor("op_4744"), val = tensor([1, 1])]; + tensor pretrained_out_231_pad_type_0 = const()[name = tensor("pretrained_out_231_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_231_pad_0 = const()[name = tensor("pretrained_out_231_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296521216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297340480))), name = tensor("layers_11_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_231_cast_fp16 = conv(dilations = var_4744, groups = var_4521, pad = pretrained_out_231_pad_0, pad_type = pretrained_out_231_pad_type_0, strides = var_4742, weight = layers_11_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_231_cast_fp16")]; + tensor var_4748 = const()[name = tensor("op_4748"), val = tensor([1, 1])]; + tensor var_4750 = const()[name = tensor("op_4750"), val = tensor([1, 1])]; + tensor input_343_pad_type_0 = const()[name = tensor("input_343_pad_type_0"), val = tensor("custom")]; + tensor input_343_pad_0 = const()[name = tensor("input_343_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297340608)))]; + tensor input_343_cast_fp16 = conv(dilations = var_4750, groups = var_4521, pad = input_343_pad_0, pad_type = input_343_pad_type_0, strides = var_4748, weight = layers_11_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_343_cast_fp16")]; + tensor var_4754 = const()[name = tensor("op_4754"), val = tensor([1, 1])]; + tensor var_4756 = const()[name = tensor("op_4756"), val = tensor([1, 1])]; + tensor lora_out_461_pad_type_0 = const()[name = tensor("lora_out_461_pad_type_0"), val = tensor("custom")]; + tensor lora_out_461_pad_0 = const()[name = tensor("lora_out_461_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_463_weight_0_to_fp16 = const()[name = tensor("lora_out_463_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297381632)))]; + tensor lora_out_463_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4756, groups = var_4521, pad = lora_out_461_pad_0, pad_type = lora_out_461_pad_type_0, strides = var_4754, weight = lora_out_463_weight_0_to_fp16, x = input_343_cast_fp16)[name = tensor("lora_out_463_cast_fp16")]; + tensor key_47_cast_fp16 = add(x = pretrained_out_231_cast_fp16, y = lora_out_463_cast_fp16)[name = tensor("key_47_cast_fp16")]; + tensor var_4767 = const()[name = tensor("op_4767"), val = tensor([1, 1])]; + tensor var_4769 = const()[name = tensor("op_4769"), val = tensor([1, 1])]; + tensor pretrained_out_233_pad_type_0 = const()[name = tensor("pretrained_out_233_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_233_pad_0 = const()[name = tensor("pretrained_out_233_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297422656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298241920))), name = tensor("layers_11_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_11_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298242048)))]; + tensor pretrained_out_233_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_4769, groups = var_4521, pad = pretrained_out_233_pad_0, pad_type = pretrained_out_233_pad_type_0, strides = var_4767, weight = layers_11_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_233_cast_fp16")]; + tensor var_4773 = const()[name = tensor("op_4773"), val = tensor([1, 1])]; + tensor var_4775 = const()[name = tensor("op_4775"), val = tensor([1, 1])]; + tensor input_345_pad_type_0 = const()[name = tensor("input_345_pad_type_0"), val = tensor("custom")]; + tensor input_345_pad_0 = const()[name = tensor("input_345_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298244672)))]; + tensor input_345_cast_fp16 = conv(dilations = var_4775, groups = var_4521, pad = input_345_pad_0, pad_type = input_345_pad_type_0, strides = var_4773, weight = layers_11_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_345_cast_fp16")]; + tensor var_4779 = const()[name = tensor("op_4779"), val = tensor([1, 1])]; + tensor var_4781 = const()[name = tensor("op_4781"), val = tensor([1, 1])]; + tensor lora_out_465_pad_type_0 = const()[name = tensor("lora_out_465_pad_type_0"), val = tensor("custom")]; + tensor lora_out_465_pad_0 = const()[name = tensor("lora_out_465_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_467_weight_0_to_fp16 = const()[name = tensor("lora_out_467_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298285696)))]; + tensor lora_out_467_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4781, groups = var_4521, pad = lora_out_465_pad_0, pad_type = lora_out_465_pad_type_0, strides = var_4779, weight = lora_out_467_weight_0_to_fp16, x = input_345_cast_fp16)[name = tensor("lora_out_467_cast_fp16")]; + tensor value_47_cast_fp16 = add(x = pretrained_out_233_cast_fp16, y = lora_out_467_cast_fp16)[name = tensor("value_47_cast_fp16")]; + tensor var_4788 = const()[name = tensor("op_4788"), val = tensor([1, 20, 64, -1])]; + tensor var_4789_cast_fp16 = reshape(shape = var_4788, x = query_47_cast_fp16)[name = tensor("op_4789_cast_fp16")]; + tensor var_4790_to_fp16 = const()[name = tensor("op_4790_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4791_cast_fp16 = mul(x = var_4789_cast_fp16, y = var_4790_to_fp16)[name = tensor("op_4791_cast_fp16")]; + tensor var_4792 = const()[name = tensor("op_4792"), val = tensor([1, 20, 64, -1])]; + tensor var_4793_cast_fp16 = reshape(shape = var_4792, x = key_47_cast_fp16)[name = tensor("op_4793_cast_fp16")]; + tensor mh_w_71_transpose_x_0 = const()[name = tensor("mh_w_71_transpose_x_0"), val = tensor(true)]; + tensor mh_w_71_transpose_y_0 = const()[name = tensor("mh_w_71_transpose_y_0"), val = tensor(false)]; + tensor mh_w_71_cast_fp16 = matmul(transpose_x = mh_w_71_transpose_x_0, transpose_y = mh_w_71_transpose_y_0, x = var_4791_cast_fp16, y = var_4793_cast_fp16)[name = tensor("mh_w_71_cast_fp16")]; + tensor obj_167_cast_fp16 = softmax(axis = var_4514, x = mh_w_71_cast_fp16)[name = tensor("obj_167_cast_fp16")]; + tensor var_4797 = const()[name = tensor("op_4797"), val = tensor([1, 20, 64, -1])]; + tensor var_4798_cast_fp16 = reshape(shape = var_4797, x = value_47_cast_fp16)[name = tensor("op_4798_cast_fp16")]; + tensor attn_47_transpose_x_0 = const()[name = tensor("attn_47_transpose_x_0"), val = tensor(false)]; + tensor attn_47_transpose_y_0 = const()[name = tensor("attn_47_transpose_y_0"), val = tensor(true)]; + tensor attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_4798_cast_fp16, y = obj_167_cast_fp16)[name = tensor("attn_47_cast_fp16")]; + tensor var_4801 = const()[name = tensor("op_4801"), val = tensor([1, 1280, 1, -1])]; + tensor input_347_cast_fp16 = reshape(shape = var_4801, x = attn_47_cast_fp16)[name = tensor("input_347_cast_fp16")]; + tensor var_4808 = const()[name = tensor("op_4808"), val = tensor([1, 1])]; + tensor var_4810 = const()[name = tensor("op_4810"), val = tensor([1, 1])]; + tensor pretrained_out_235_pad_type_0 = const()[name = tensor("pretrained_out_235_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_235_pad_0 = const()[name = tensor("pretrained_out_235_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298326720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299145984))), name = tensor("layers_11_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_11_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299146112)))]; + tensor pretrained_out_235_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4810, groups = var_4521, pad = pretrained_out_235_pad_0, pad_type = pretrained_out_235_pad_type_0, strides = var_4808, weight = layers_11_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_347_cast_fp16)[name = tensor("pretrained_out_235_cast_fp16")]; + tensor var_4814 = const()[name = tensor("op_4814"), val = tensor([1, 1])]; + tensor var_4816 = const()[name = tensor("op_4816"), val = tensor([1, 1])]; + tensor input_349_pad_type_0 = const()[name = tensor("input_349_pad_type_0"), val = tensor("custom")]; + tensor input_349_pad_0 = const()[name = tensor("input_349_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299148736)))]; + tensor input_349_cast_fp16 = conv(dilations = var_4816, groups = var_4521, pad = input_349_pad_0, pad_type = input_349_pad_type_0, strides = var_4814, weight = layers_11_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_347_cast_fp16)[name = tensor("input_349_cast_fp16")]; + tensor var_4820 = const()[name = tensor("op_4820"), val = tensor([1, 1])]; + tensor var_4822 = const()[name = tensor("op_4822"), val = tensor([1, 1])]; + tensor lora_out_469_pad_type_0 = const()[name = tensor("lora_out_469_pad_type_0"), val = tensor("custom")]; + tensor lora_out_469_pad_0 = const()[name = tensor("lora_out_469_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_471_weight_0_to_fp16 = const()[name = tensor("lora_out_471_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299189760)))]; + tensor lora_out_471_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4822, groups = var_4521, pad = lora_out_469_pad_0, pad_type = lora_out_469_pad_type_0, strides = var_4820, weight = lora_out_471_weight_0_to_fp16, x = input_349_cast_fp16)[name = tensor("lora_out_471_cast_fp16")]; + tensor obj_165_cast_fp16 = add(x = pretrained_out_235_cast_fp16, y = lora_out_471_cast_fp16)[name = tensor("obj_165_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_165_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor var_4831 = const()[name = tensor("op_4831"), val = tensor([1])]; + tensor channels_mean_71_cast_fp16 = reduce_mean(axes = var_4831, keep_dims = var_4522, x = inputs_71_cast_fp16)[name = tensor("channels_mean_71_cast_fp16")]; + tensor zero_mean_71_cast_fp16 = sub(x = inputs_71_cast_fp16, y = channels_mean_71_cast_fp16)[name = tensor("zero_mean_71_cast_fp16")]; + tensor zero_mean_sq_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = zero_mean_71_cast_fp16)[name = tensor("zero_mean_sq_71_cast_fp16")]; + tensor var_4835 = const()[name = tensor("op_4835"), val = tensor([1])]; + tensor var_4836_cast_fp16 = reduce_mean(axes = var_4835, keep_dims = var_4522, x = zero_mean_sq_71_cast_fp16)[name = tensor("op_4836_cast_fp16")]; + tensor var_4837_to_fp16 = const()[name = tensor("op_4837_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4838_cast_fp16 = add(x = var_4836_cast_fp16, y = var_4837_to_fp16)[name = tensor("op_4838_cast_fp16")]; + tensor denom_71_epsilon_0 = const()[name = tensor("denom_71_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_71_cast_fp16 = rsqrt(epsilon = denom_71_epsilon_0, x = var_4838_cast_fp16)[name = tensor("denom_71_cast_fp16")]; + tensor out_71_cast_fp16 = mul(x = zero_mean_71_cast_fp16, y = denom_71_cast_fp16)[name = tensor("out_71_cast_fp16")]; + tensor input_351_gamma_0_to_fp16 = const()[name = tensor("input_351_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299230784)))]; + tensor input_351_beta_0_to_fp16 = const()[name = tensor("input_351_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299233408)))]; + tensor input_351_epsilon_0_to_fp16 = const()[name = tensor("input_351_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_351_cast_fp16 = batch_norm(beta = input_351_beta_0_to_fp16, epsilon = input_351_epsilon_0_to_fp16, gamma = input_351_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor("input_351_cast_fp16")]; + tensor var_4852 = const()[name = tensor("op_4852"), val = tensor([1, 1])]; + tensor var_4854 = const()[name = tensor("op_4854"), val = tensor([1, 1])]; + tensor pretrained_out_237_pad_type_0 = const()[name = tensor("pretrained_out_237_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_237_pad_0 = const()[name = tensor("pretrained_out_237_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299236032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302512896))), name = tensor("layers_11_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_11_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302513024)))]; + tensor pretrained_out_237_cast_fp16 = conv(bias = layers_11_fc1_pretrained_bias_to_fp16, dilations = var_4854, groups = var_4521, pad = pretrained_out_237_pad_0, pad_type = pretrained_out_237_pad_type_0, strides = var_4852, weight = layers_11_fc1_pretrained_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = tensor("pretrained_out_237_cast_fp16")]; + tensor var_4858 = const()[name = tensor("op_4858"), val = tensor([1, 1])]; + tensor var_4860 = const()[name = tensor("op_4860"), val = tensor([1, 1])]; + tensor input_353_pad_type_0 = const()[name = tensor("input_353_pad_type_0"), val = tensor("custom")]; + tensor input_353_pad_0 = const()[name = tensor("input_353_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_11_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302523328)))]; + tensor input_353_cast_fp16 = conv(dilations = var_4860, groups = var_4521, pad = input_353_pad_0, pad_type = input_353_pad_type_0, strides = var_4858, weight = layers_11_fc1_loraA_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor var_4864 = const()[name = tensor("op_4864"), val = tensor([1, 1])]; + tensor var_4866 = const()[name = tensor("op_4866"), val = tensor([1, 1])]; + tensor lora_out_473_pad_type_0 = const()[name = tensor("lora_out_473_pad_type_0"), val = tensor("custom")]; + tensor lora_out_473_pad_0 = const()[name = tensor("lora_out_473_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_475_weight_0_to_fp16 = const()[name = tensor("lora_out_475_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302564352)))]; + tensor lora_out_475_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_4866, groups = var_4521, pad = lora_out_473_pad_0, pad_type = lora_out_473_pad_type_0, strides = var_4864, weight = lora_out_475_weight_0_to_fp16, x = input_353_cast_fp16)[name = tensor("lora_out_475_cast_fp16")]; + tensor input_355_cast_fp16 = add(x = pretrained_out_237_cast_fp16, y = lora_out_475_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor input_357_mode_0 = const()[name = tensor("input_357_mode_0"), val = tensor("EXACT")]; + tensor input_357_cast_fp16 = gelu(mode = input_357_mode_0, x = input_355_cast_fp16)[name = tensor("input_357_cast_fp16")]; + tensor var_4878 = const()[name = tensor("op_4878"), val = tensor([1, 1])]; + tensor var_4880 = const()[name = tensor("op_4880"), val = tensor([1, 1])]; + tensor pretrained_out_239_pad_type_0 = const()[name = tensor("pretrained_out_239_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_239_pad_0 = const()[name = tensor("pretrained_out_239_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302728256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306005120))), name = tensor("layers_11_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_11_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_11_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306005248)))]; + tensor pretrained_out_239_cast_fp16 = conv(bias = layers_11_fc2_pretrained_bias_to_fp16, dilations = var_4880, groups = var_4521, pad = pretrained_out_239_pad_0, pad_type = pretrained_out_239_pad_type_0, strides = var_4878, weight = layers_11_fc2_pretrained_weight_to_fp16_palettized, x = input_357_cast_fp16)[name = tensor("pretrained_out_239_cast_fp16")]; + tensor var_4884 = const()[name = tensor("op_4884"), val = tensor([1, 1])]; + tensor var_4886 = const()[name = tensor("op_4886"), val = tensor([1, 1])]; + tensor input_359_pad_type_0 = const()[name = tensor("input_359_pad_type_0"), val = tensor("custom")]; + tensor input_359_pad_0 = const()[name = tensor("input_359_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_11_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_11_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306007872)))]; + tensor input_359_cast_fp16 = conv(dilations = var_4886, groups = var_4521, pad = input_359_pad_0, pad_type = input_359_pad_type_0, strides = var_4884, weight = layers_11_fc2_loraA_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("input_359_cast_fp16")]; + tensor var_4890 = const()[name = tensor("op_4890"), val = tensor([1, 1])]; + tensor var_4892 = const()[name = tensor("op_4892"), val = tensor([1, 1])]; + tensor lora_out_477_pad_type_0 = const()[name = tensor("lora_out_477_pad_type_0"), val = tensor("custom")]; + tensor lora_out_477_pad_0 = const()[name = tensor("lora_out_477_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_479_weight_0_to_fp16 = const()[name = tensor("lora_out_479_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306171776)))]; + tensor lora_out_479_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4892, groups = var_4521, pad = lora_out_477_pad_0, pad_type = lora_out_477_pad_type_0, strides = var_4890, weight = lora_out_479_weight_0_to_fp16, x = input_359_cast_fp16)[name = tensor("lora_out_479_cast_fp16")]; + tensor hidden_states_25_cast_fp16 = add(x = pretrained_out_239_cast_fp16, y = lora_out_479_cast_fp16)[name = tensor("hidden_states_25_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_4908 = const()[name = tensor("op_4908"), val = tensor(3)]; + tensor var_4915 = const()[name = tensor("op_4915"), val = tensor(1)]; + tensor var_4916 = const()[name = tensor("op_4916"), val = tensor(true)]; + tensor var_4928 = const()[name = tensor("op_4928"), val = tensor([1])]; + tensor channels_mean_73_cast_fp16 = reduce_mean(axes = var_4928, keep_dims = var_4916, x = inputs_73_cast_fp16)[name = tensor("channels_mean_73_cast_fp16")]; + tensor zero_mean_73_cast_fp16 = sub(x = inputs_73_cast_fp16, y = channels_mean_73_cast_fp16)[name = tensor("zero_mean_73_cast_fp16")]; + tensor zero_mean_sq_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = zero_mean_73_cast_fp16)[name = tensor("zero_mean_sq_73_cast_fp16")]; + tensor var_4932 = const()[name = tensor("op_4932"), val = tensor([1])]; + tensor var_4933_cast_fp16 = reduce_mean(axes = var_4932, keep_dims = var_4916, x = zero_mean_sq_73_cast_fp16)[name = tensor("op_4933_cast_fp16")]; + tensor var_4934_to_fp16 = const()[name = tensor("op_4934_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4935_cast_fp16 = add(x = var_4933_cast_fp16, y = var_4934_to_fp16)[name = tensor("op_4935_cast_fp16")]; + tensor denom_73_epsilon_0 = const()[name = tensor("denom_73_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_73_cast_fp16 = rsqrt(epsilon = denom_73_epsilon_0, x = var_4935_cast_fp16)[name = tensor("denom_73_cast_fp16")]; + tensor out_73_cast_fp16 = mul(x = zero_mean_73_cast_fp16, y = denom_73_cast_fp16)[name = tensor("out_73_cast_fp16")]; + tensor obj_169_gamma_0_to_fp16 = const()[name = tensor("obj_169_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306212800)))]; + tensor obj_169_beta_0_to_fp16 = const()[name = tensor("obj_169_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306215424)))]; + tensor obj_169_epsilon_0_to_fp16 = const()[name = tensor("obj_169_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_169_cast_fp16 = batch_norm(beta = obj_169_beta_0_to_fp16, epsilon = obj_169_epsilon_0_to_fp16, gamma = obj_169_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor("obj_169_cast_fp16")]; + tensor var_4953 = const()[name = tensor("op_4953"), val = tensor([1, 1])]; + tensor var_4955 = const()[name = tensor("op_4955"), val = tensor([1, 1])]; + tensor pretrained_out_241_pad_type_0 = const()[name = tensor("pretrained_out_241_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_241_pad_0 = const()[name = tensor("pretrained_out_241_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306218048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307037312))), name = tensor("layers_12_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_12_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307037440)))]; + tensor pretrained_out_241_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4955, groups = var_4915, pad = pretrained_out_241_pad_0, pad_type = pretrained_out_241_pad_type_0, strides = var_4953, weight = layers_12_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_169_cast_fp16)[name = tensor("pretrained_out_241_cast_fp16")]; + tensor var_4959 = const()[name = tensor("op_4959"), val = tensor([1, 1])]; + tensor var_4961 = const()[name = tensor("op_4961"), val = tensor([1, 1])]; + tensor input_361_pad_type_0 = const()[name = tensor("input_361_pad_type_0"), val = tensor("custom")]; + tensor input_361_pad_0 = const()[name = tensor("input_361_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307040064)))]; + tensor input_361_cast_fp16 = conv(dilations = var_4961, groups = var_4915, pad = input_361_pad_0, pad_type = input_361_pad_type_0, strides = var_4959, weight = layers_12_self_attn_q_proj_loraA_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor var_4965 = const()[name = tensor("op_4965"), val = tensor([1, 1])]; + tensor var_4967 = const()[name = tensor("op_4967"), val = tensor([1, 1])]; + tensor lora_out_481_pad_type_0 = const()[name = tensor("lora_out_481_pad_type_0"), val = tensor("custom")]; + tensor lora_out_481_pad_0 = const()[name = tensor("lora_out_481_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_483_weight_0_to_fp16 = const()[name = tensor("lora_out_483_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307081088)))]; + tensor lora_out_483_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4967, groups = var_4915, pad = lora_out_481_pad_0, pad_type = lora_out_481_pad_type_0, strides = var_4965, weight = lora_out_483_weight_0_to_fp16, x = input_361_cast_fp16)[name = tensor("lora_out_483_cast_fp16")]; + tensor query_49_cast_fp16 = add(x = pretrained_out_241_cast_fp16, y = lora_out_483_cast_fp16)[name = tensor("query_49_cast_fp16")]; + tensor var_4977 = const()[name = tensor("op_4977"), val = tensor([1, 1])]; + tensor var_4979 = const()[name = tensor("op_4979"), val = tensor([1, 1])]; + tensor pretrained_out_243_pad_type_0 = const()[name = tensor("pretrained_out_243_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_243_pad_0 = const()[name = tensor("pretrained_out_243_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307122112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307941376))), name = tensor("layers_12_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_243_cast_fp16 = conv(dilations = var_4979, groups = var_4915, pad = pretrained_out_243_pad_0, pad_type = pretrained_out_243_pad_type_0, strides = var_4977, weight = layers_12_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_169_cast_fp16)[name = tensor("pretrained_out_243_cast_fp16")]; + tensor var_4983 = const()[name = tensor("op_4983"), val = tensor([1, 1])]; + tensor var_4985 = const()[name = tensor("op_4985"), val = tensor([1, 1])]; + tensor input_363_pad_type_0 = const()[name = tensor("input_363_pad_type_0"), val = tensor("custom")]; + tensor input_363_pad_0 = const()[name = tensor("input_363_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307941504)))]; + tensor input_363_cast_fp16 = conv(dilations = var_4985, groups = var_4915, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = var_4983, weight = layers_12_self_attn_k_proj_loraA_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("input_363_cast_fp16")]; + tensor var_4989 = const()[name = tensor("op_4989"), val = tensor([1, 1])]; + tensor var_4991 = const()[name = tensor("op_4991"), val = tensor([1, 1])]; + tensor lora_out_485_pad_type_0 = const()[name = tensor("lora_out_485_pad_type_0"), val = tensor("custom")]; + tensor lora_out_485_pad_0 = const()[name = tensor("lora_out_485_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_487_weight_0_to_fp16 = const()[name = tensor("lora_out_487_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307982528)))]; + tensor lora_out_487_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4991, groups = var_4915, pad = lora_out_485_pad_0, pad_type = lora_out_485_pad_type_0, strides = var_4989, weight = lora_out_487_weight_0_to_fp16, x = input_363_cast_fp16)[name = tensor("lora_out_487_cast_fp16")]; + tensor current_key_25_cast_fp16 = add(x = pretrained_out_243_cast_fp16, y = lora_out_487_cast_fp16)[name = tensor("current_key_25_cast_fp16")]; + tensor var_5002 = const()[name = tensor("op_5002"), val = tensor([1, 1])]; + tensor var_5004 = const()[name = tensor("op_5004"), val = tensor([1, 1])]; + tensor pretrained_out_245_pad_type_0 = const()[name = tensor("pretrained_out_245_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_245_pad_0 = const()[name = tensor("pretrained_out_245_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308023552))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308842816))), name = tensor("layers_12_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_12_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308842944)))]; + tensor pretrained_out_245_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5004, groups = var_4915, pad = pretrained_out_245_pad_0, pad_type = pretrained_out_245_pad_type_0, strides = var_5002, weight = layers_12_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_169_cast_fp16)[name = tensor("pretrained_out_245_cast_fp16")]; + tensor var_5008 = const()[name = tensor("op_5008"), val = tensor([1, 1])]; + tensor var_5010 = const()[name = tensor("op_5010"), val = tensor([1, 1])]; + tensor input_365_pad_type_0 = const()[name = tensor("input_365_pad_type_0"), val = tensor("custom")]; + tensor input_365_pad_0 = const()[name = tensor("input_365_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308845568)))]; + tensor input_365_cast_fp16 = conv(dilations = var_5010, groups = var_4915, pad = input_365_pad_0, pad_type = input_365_pad_type_0, strides = var_5008, weight = layers_12_self_attn_v_proj_loraA_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("input_365_cast_fp16")]; + tensor var_5014 = const()[name = tensor("op_5014"), val = tensor([1, 1])]; + tensor var_5016 = const()[name = tensor("op_5016"), val = tensor([1, 1])]; + tensor lora_out_489_pad_type_0 = const()[name = tensor("lora_out_489_pad_type_0"), val = tensor("custom")]; + tensor lora_out_489_pad_0 = const()[name = tensor("lora_out_489_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_491_weight_0_to_fp16 = const()[name = tensor("lora_out_491_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308886592)))]; + tensor lora_out_491_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5016, groups = var_4915, pad = lora_out_489_pad_0, pad_type = lora_out_489_pad_type_0, strides = var_5014, weight = lora_out_491_weight_0_to_fp16, x = input_365_cast_fp16)[name = tensor("lora_out_491_cast_fp16")]; + tensor current_value_25_cast_fp16 = add(x = pretrained_out_245_cast_fp16, y = lora_out_491_cast_fp16)[name = tensor("current_value_25_cast_fp16")]; + tensor var_5026_cast_fp16 = mul(x = current_key_25_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_5026_cast_fp16")]; + tensor var_5028_cast_fp16 = mul(x = var_103_cast_fp16_12, y = var_295_cast_fp16)[name = tensor("op_5028_cast_fp16")]; + tensor key_49_cast_fp16 = add(x = var_5026_cast_fp16, y = var_5028_cast_fp16)[name = tensor("key_49_cast_fp16")]; + tensor var_5030_cast_fp16 = mul(x = current_value_25_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_5030_cast_fp16")]; + tensor var_5032_cast_fp16 = mul(x = var_138_cast_fp16_12, y = var_295_cast_fp16)[name = tensor("op_5032_cast_fp16")]; + tensor value_49_cast_fp16 = add(x = var_5030_cast_fp16, y = var_5032_cast_fp16)[name = tensor("value_49_cast_fp16")]; + tensor var_5035 = const()[name = tensor("op_5035"), val = tensor([1, 20, 64, -1])]; + tensor var_5036_cast_fp16 = reshape(shape = var_5035, x = query_49_cast_fp16)[name = tensor("op_5036_cast_fp16")]; + tensor var_5037_to_fp16 = const()[name = tensor("op_5037_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5038_cast_fp16 = mul(x = var_5036_cast_fp16, y = var_5037_to_fp16)[name = tensor("op_5038_cast_fp16")]; + tensor var_5039 = const()[name = tensor("op_5039"), val = tensor([1, 20, 64, -1])]; + tensor var_5040_cast_fp16 = reshape(shape = var_5039, x = key_49_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor mh_w_73_transpose_x_0 = const()[name = tensor("mh_w_73_transpose_x_0"), val = tensor(true)]; + tensor mh_w_73_transpose_y_0 = const()[name = tensor("mh_w_73_transpose_y_0"), val = tensor(false)]; + tensor mh_w_73_cast_fp16 = matmul(transpose_x = mh_w_73_transpose_x_0, transpose_y = mh_w_73_transpose_y_0, x = var_5038_cast_fp16, y = var_5040_cast_fp16)[name = tensor("mh_w_73_cast_fp16")]; + tensor mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_75_cast_fp16")]; + tensor var_5048_cast_fp16 = softmax(axis = var_4908, x = mh_w_75_cast_fp16)[name = tensor("op_5048_cast_fp16")]; + tensor var_5049 = const()[name = tensor("op_5049"), val = tensor([1, 20, 64, -1])]; + tensor var_5050_cast_fp16 = reshape(shape = var_5049, x = value_49_cast_fp16)[name = tensor("op_5050_cast_fp16")]; + tensor attn_49_transpose_x_0 = const()[name = tensor("attn_49_transpose_x_0"), val = tensor(false)]; + tensor attn_49_transpose_y_0 = const()[name = tensor("attn_49_transpose_y_0"), val = tensor(true)]; + tensor attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_5050_cast_fp16, y = var_5048_cast_fp16)[name = tensor("attn_49_cast_fp16")]; + tensor var_5053 = const()[name = tensor("op_5053"), val = tensor([1, 1280, 1, -1])]; + tensor input_367_cast_fp16 = reshape(shape = var_5053, x = attn_49_cast_fp16)[name = tensor("input_367_cast_fp16")]; + tensor var_5060 = const()[name = tensor("op_5060"), val = tensor([1, 1])]; + tensor var_5062 = const()[name = tensor("op_5062"), val = tensor([1, 1])]; + tensor pretrained_out_247_pad_type_0 = const()[name = tensor("pretrained_out_247_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_247_pad_0 = const()[name = tensor("pretrained_out_247_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308927616))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309746880))), name = tensor("layers_12_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_12_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309747008)))]; + tensor pretrained_out_247_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5062, groups = var_4915, pad = pretrained_out_247_pad_0, pad_type = pretrained_out_247_pad_type_0, strides = var_5060, weight = layers_12_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_367_cast_fp16)[name = tensor("pretrained_out_247_cast_fp16")]; + tensor var_5066 = const()[name = tensor("op_5066"), val = tensor([1, 1])]; + tensor var_5068 = const()[name = tensor("op_5068"), val = tensor([1, 1])]; + tensor input_369_pad_type_0 = const()[name = tensor("input_369_pad_type_0"), val = tensor("custom")]; + tensor input_369_pad_0 = const()[name = tensor("input_369_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309749632)))]; + tensor input_369_cast_fp16 = conv(dilations = var_5068, groups = var_4915, pad = input_369_pad_0, pad_type = input_369_pad_type_0, strides = var_5066, weight = layers_12_self_attn_o_proj_loraA_weight_to_fp16, x = input_367_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor var_5072 = const()[name = tensor("op_5072"), val = tensor([1, 1])]; + tensor var_5074 = const()[name = tensor("op_5074"), val = tensor([1, 1])]; + tensor lora_out_493_pad_type_0 = const()[name = tensor("lora_out_493_pad_type_0"), val = tensor("custom")]; + tensor lora_out_493_pad_0 = const()[name = tensor("lora_out_493_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_495_weight_0_to_fp16 = const()[name = tensor("lora_out_495_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309790656)))]; + tensor lora_out_495_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5074, groups = var_4915, pad = lora_out_493_pad_0, pad_type = lora_out_493_pad_type_0, strides = var_5072, weight = lora_out_495_weight_0_to_fp16, x = input_369_cast_fp16)[name = tensor("lora_out_495_cast_fp16")]; + tensor obj_175_cast_fp16 = add(x = pretrained_out_247_cast_fp16, y = lora_out_495_cast_fp16)[name = tensor("obj_175_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_175_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor var_5087 = const()[name = tensor("op_5087"), val = tensor([1])]; + tensor channels_mean_75_cast_fp16 = reduce_mean(axes = var_5087, keep_dims = var_4916, x = inputs_75_cast_fp16)[name = tensor("channels_mean_75_cast_fp16")]; + tensor zero_mean_75_cast_fp16 = sub(x = inputs_75_cast_fp16, y = channels_mean_75_cast_fp16)[name = tensor("zero_mean_75_cast_fp16")]; + tensor zero_mean_sq_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = zero_mean_75_cast_fp16)[name = tensor("zero_mean_sq_75_cast_fp16")]; + tensor var_5091 = const()[name = tensor("op_5091"), val = tensor([1])]; + tensor var_5092_cast_fp16 = reduce_mean(axes = var_5091, keep_dims = var_4916, x = zero_mean_sq_75_cast_fp16)[name = tensor("op_5092_cast_fp16")]; + tensor var_5093_to_fp16 = const()[name = tensor("op_5093_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5094_cast_fp16 = add(x = var_5092_cast_fp16, y = var_5093_to_fp16)[name = tensor("op_5094_cast_fp16")]; + tensor denom_75_epsilon_0 = const()[name = tensor("denom_75_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_75_cast_fp16 = rsqrt(epsilon = denom_75_epsilon_0, x = var_5094_cast_fp16)[name = tensor("denom_75_cast_fp16")]; + tensor out_75_cast_fp16 = mul(x = zero_mean_75_cast_fp16, y = denom_75_cast_fp16)[name = tensor("out_75_cast_fp16")]; + tensor obj_177_gamma_0_to_fp16 = const()[name = tensor("obj_177_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309831680)))]; + tensor obj_177_beta_0_to_fp16 = const()[name = tensor("obj_177_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309834304)))]; + tensor obj_177_epsilon_0_to_fp16 = const()[name = tensor("obj_177_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_177_cast_fp16 = batch_norm(beta = obj_177_beta_0_to_fp16, epsilon = obj_177_epsilon_0_to_fp16, gamma = obj_177_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor("obj_177_cast_fp16")]; + tensor var_5112 = const()[name = tensor("op_5112"), val = tensor([1, 1])]; + tensor var_5114 = const()[name = tensor("op_5114"), val = tensor([1, 1])]; + tensor pretrained_out_249_pad_type_0 = const()[name = tensor("pretrained_out_249_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_249_pad_0 = const()[name = tensor("pretrained_out_249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309836928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310656192))), name = tensor("layers_12_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_12_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310656320)))]; + tensor pretrained_out_249_cast_fp16 = conv(bias = layers_12_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5114, groups = var_4915, pad = pretrained_out_249_pad_0, pad_type = pretrained_out_249_pad_type_0, strides = var_5112, weight = layers_12_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_177_cast_fp16)[name = tensor("pretrained_out_249_cast_fp16")]; + tensor var_5118 = const()[name = tensor("op_5118"), val = tensor([1, 1])]; + tensor var_5120 = const()[name = tensor("op_5120"), val = tensor([1, 1])]; + tensor input_371_pad_type_0 = const()[name = tensor("input_371_pad_type_0"), val = tensor("custom")]; + tensor input_371_pad_0 = const()[name = tensor("input_371_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310658944)))]; + tensor input_371_cast_fp16 = conv(dilations = var_5120, groups = var_4915, pad = input_371_pad_0, pad_type = input_371_pad_type_0, strides = var_5118, weight = layers_12_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_177_cast_fp16)[name = tensor("input_371_cast_fp16")]; + tensor var_5124 = const()[name = tensor("op_5124"), val = tensor([1, 1])]; + tensor var_5126 = const()[name = tensor("op_5126"), val = tensor([1, 1])]; + tensor lora_out_497_pad_type_0 = const()[name = tensor("lora_out_497_pad_type_0"), val = tensor("custom")]; + tensor lora_out_497_pad_0 = const()[name = tensor("lora_out_497_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_499_weight_0_to_fp16 = const()[name = tensor("lora_out_499_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310699968)))]; + tensor lora_out_499_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5126, groups = var_4915, pad = lora_out_497_pad_0, pad_type = lora_out_497_pad_type_0, strides = var_5124, weight = lora_out_499_weight_0_to_fp16, x = input_371_cast_fp16)[name = tensor("lora_out_499_cast_fp16")]; + tensor query_51_cast_fp16 = add(x = pretrained_out_249_cast_fp16, y = lora_out_499_cast_fp16)[name = tensor("query_51_cast_fp16")]; + tensor var_5136 = const()[name = tensor("op_5136"), val = tensor([1, 1])]; + tensor var_5138 = const()[name = tensor("op_5138"), val = tensor([1, 1])]; + tensor pretrained_out_251_pad_type_0 = const()[name = tensor("pretrained_out_251_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_251_pad_0 = const()[name = tensor("pretrained_out_251_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310740992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311560256))), name = tensor("layers_12_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_251_cast_fp16 = conv(dilations = var_5138, groups = var_4915, pad = pretrained_out_251_pad_0, pad_type = pretrained_out_251_pad_type_0, strides = var_5136, weight = layers_12_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_251_cast_fp16")]; + tensor var_5142 = const()[name = tensor("op_5142"), val = tensor([1, 1])]; + tensor var_5144 = const()[name = tensor("op_5144"), val = tensor([1, 1])]; + tensor input_373_pad_type_0 = const()[name = tensor("input_373_pad_type_0"), val = tensor("custom")]; + tensor input_373_pad_0 = const()[name = tensor("input_373_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311560384)))]; + tensor input_373_cast_fp16 = conv(dilations = var_5144, groups = var_4915, pad = input_373_pad_0, pad_type = input_373_pad_type_0, strides = var_5142, weight = layers_12_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_373_cast_fp16")]; + tensor var_5148 = const()[name = tensor("op_5148"), val = tensor([1, 1])]; + tensor var_5150 = const()[name = tensor("op_5150"), val = tensor([1, 1])]; + tensor lora_out_501_pad_type_0 = const()[name = tensor("lora_out_501_pad_type_0"), val = tensor("custom")]; + tensor lora_out_501_pad_0 = const()[name = tensor("lora_out_501_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_503_weight_0_to_fp16 = const()[name = tensor("lora_out_503_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311601408)))]; + tensor lora_out_503_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5150, groups = var_4915, pad = lora_out_501_pad_0, pad_type = lora_out_501_pad_type_0, strides = var_5148, weight = lora_out_503_weight_0_to_fp16, x = input_373_cast_fp16)[name = tensor("lora_out_503_cast_fp16")]; + tensor key_51_cast_fp16 = add(x = pretrained_out_251_cast_fp16, y = lora_out_503_cast_fp16)[name = tensor("key_51_cast_fp16")]; + tensor var_5161 = const()[name = tensor("op_5161"), val = tensor([1, 1])]; + tensor var_5163 = const()[name = tensor("op_5163"), val = tensor([1, 1])]; + tensor pretrained_out_253_pad_type_0 = const()[name = tensor("pretrained_out_253_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_253_pad_0 = const()[name = tensor("pretrained_out_253_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311642432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312461696))), name = tensor("layers_12_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_12_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312461824)))]; + tensor pretrained_out_253_cast_fp16 = conv(bias = layers_12_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5163, groups = var_4915, pad = pretrained_out_253_pad_0, pad_type = pretrained_out_253_pad_type_0, strides = var_5161, weight = layers_12_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_253_cast_fp16")]; + tensor var_5167 = const()[name = tensor("op_5167"), val = tensor([1, 1])]; + tensor var_5169 = const()[name = tensor("op_5169"), val = tensor([1, 1])]; + tensor input_375_pad_type_0 = const()[name = tensor("input_375_pad_type_0"), val = tensor("custom")]; + tensor input_375_pad_0 = const()[name = tensor("input_375_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312464448)))]; + tensor input_375_cast_fp16 = conv(dilations = var_5169, groups = var_4915, pad = input_375_pad_0, pad_type = input_375_pad_type_0, strides = var_5167, weight = layers_12_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_375_cast_fp16")]; + tensor var_5173 = const()[name = tensor("op_5173"), val = tensor([1, 1])]; + tensor var_5175 = const()[name = tensor("op_5175"), val = tensor([1, 1])]; + tensor lora_out_505_pad_type_0 = const()[name = tensor("lora_out_505_pad_type_0"), val = tensor("custom")]; + tensor lora_out_505_pad_0 = const()[name = tensor("lora_out_505_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_507_weight_0_to_fp16 = const()[name = tensor("lora_out_507_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312505472)))]; + tensor lora_out_507_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5175, groups = var_4915, pad = lora_out_505_pad_0, pad_type = lora_out_505_pad_type_0, strides = var_5173, weight = lora_out_507_weight_0_to_fp16, x = input_375_cast_fp16)[name = tensor("lora_out_507_cast_fp16")]; + tensor value_51_cast_fp16 = add(x = pretrained_out_253_cast_fp16, y = lora_out_507_cast_fp16)[name = tensor("value_51_cast_fp16")]; + tensor var_5182 = const()[name = tensor("op_5182"), val = tensor([1, 20, 64, -1])]; + tensor var_5183_cast_fp16 = reshape(shape = var_5182, x = query_51_cast_fp16)[name = tensor("op_5183_cast_fp16")]; + tensor var_5184_to_fp16 = const()[name = tensor("op_5184_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5185_cast_fp16 = mul(x = var_5183_cast_fp16, y = var_5184_to_fp16)[name = tensor("op_5185_cast_fp16")]; + tensor var_5186 = const()[name = tensor("op_5186"), val = tensor([1, 20, 64, -1])]; + tensor var_5187_cast_fp16 = reshape(shape = var_5186, x = key_51_cast_fp16)[name = tensor("op_5187_cast_fp16")]; + tensor mh_w_77_transpose_x_0 = const()[name = tensor("mh_w_77_transpose_x_0"), val = tensor(true)]; + tensor mh_w_77_transpose_y_0 = const()[name = tensor("mh_w_77_transpose_y_0"), val = tensor(false)]; + tensor mh_w_77_cast_fp16 = matmul(transpose_x = mh_w_77_transpose_x_0, transpose_y = mh_w_77_transpose_y_0, x = var_5185_cast_fp16, y = var_5187_cast_fp16)[name = tensor("mh_w_77_cast_fp16")]; + tensor obj_181_cast_fp16 = softmax(axis = var_4908, x = mh_w_77_cast_fp16)[name = tensor("obj_181_cast_fp16")]; + tensor var_5191 = const()[name = tensor("op_5191"), val = tensor([1, 20, 64, -1])]; + tensor var_5192_cast_fp16 = reshape(shape = var_5191, x = value_51_cast_fp16)[name = tensor("op_5192_cast_fp16")]; + tensor attn_51_transpose_x_0 = const()[name = tensor("attn_51_transpose_x_0"), val = tensor(false)]; + tensor attn_51_transpose_y_0 = const()[name = tensor("attn_51_transpose_y_0"), val = tensor(true)]; + tensor attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_5192_cast_fp16, y = obj_181_cast_fp16)[name = tensor("attn_51_cast_fp16")]; + tensor var_5195 = const()[name = tensor("op_5195"), val = tensor([1, 1280, 1, -1])]; + tensor input_377_cast_fp16 = reshape(shape = var_5195, x = attn_51_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor var_5202 = const()[name = tensor("op_5202"), val = tensor([1, 1])]; + tensor var_5204 = const()[name = tensor("op_5204"), val = tensor([1, 1])]; + tensor pretrained_out_255_pad_type_0 = const()[name = tensor("pretrained_out_255_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_255_pad_0 = const()[name = tensor("pretrained_out_255_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312546496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313365760))), name = tensor("layers_12_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_12_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313365888)))]; + tensor pretrained_out_255_cast_fp16 = conv(bias = layers_12_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5204, groups = var_4915, pad = pretrained_out_255_pad_0, pad_type = pretrained_out_255_pad_type_0, strides = var_5202, weight = layers_12_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = tensor("pretrained_out_255_cast_fp16")]; + tensor var_5208 = const()[name = tensor("op_5208"), val = tensor([1, 1])]; + tensor var_5210 = const()[name = tensor("op_5210"), val = tensor([1, 1])]; + tensor input_379_pad_type_0 = const()[name = tensor("input_379_pad_type_0"), val = tensor("custom")]; + tensor input_379_pad_0 = const()[name = tensor("input_379_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313368512)))]; + tensor input_379_cast_fp16 = conv(dilations = var_5210, groups = var_4915, pad = input_379_pad_0, pad_type = input_379_pad_type_0, strides = var_5208, weight = layers_12_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("input_379_cast_fp16")]; + tensor var_5214 = const()[name = tensor("op_5214"), val = tensor([1, 1])]; + tensor var_5216 = const()[name = tensor("op_5216"), val = tensor([1, 1])]; + tensor lora_out_509_pad_type_0 = const()[name = tensor("lora_out_509_pad_type_0"), val = tensor("custom")]; + tensor lora_out_509_pad_0 = const()[name = tensor("lora_out_509_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_511_weight_0_to_fp16 = const()[name = tensor("lora_out_511_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313409536)))]; + tensor lora_out_511_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5216, groups = var_4915, pad = lora_out_509_pad_0, pad_type = lora_out_509_pad_type_0, strides = var_5214, weight = lora_out_511_weight_0_to_fp16, x = input_379_cast_fp16)[name = tensor("lora_out_511_cast_fp16")]; + tensor obj_179_cast_fp16 = add(x = pretrained_out_255_cast_fp16, y = lora_out_511_cast_fp16)[name = tensor("obj_179_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = obj_179_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor var_5225 = const()[name = tensor("op_5225"), val = tensor([1])]; + tensor channels_mean_77_cast_fp16 = reduce_mean(axes = var_5225, keep_dims = var_4916, x = inputs_77_cast_fp16)[name = tensor("channels_mean_77_cast_fp16")]; + tensor zero_mean_77_cast_fp16 = sub(x = inputs_77_cast_fp16, y = channels_mean_77_cast_fp16)[name = tensor("zero_mean_77_cast_fp16")]; + tensor zero_mean_sq_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = zero_mean_77_cast_fp16)[name = tensor("zero_mean_sq_77_cast_fp16")]; + tensor var_5229 = const()[name = tensor("op_5229"), val = tensor([1])]; + tensor var_5230_cast_fp16 = reduce_mean(axes = var_5229, keep_dims = var_4916, x = zero_mean_sq_77_cast_fp16)[name = tensor("op_5230_cast_fp16")]; + tensor var_5231_to_fp16 = const()[name = tensor("op_5231_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5232_cast_fp16 = add(x = var_5230_cast_fp16, y = var_5231_to_fp16)[name = tensor("op_5232_cast_fp16")]; + tensor denom_77_epsilon_0 = const()[name = tensor("denom_77_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_77_cast_fp16 = rsqrt(epsilon = denom_77_epsilon_0, x = var_5232_cast_fp16)[name = tensor("denom_77_cast_fp16")]; + tensor out_77_cast_fp16 = mul(x = zero_mean_77_cast_fp16, y = denom_77_cast_fp16)[name = tensor("out_77_cast_fp16")]; + tensor input_381_gamma_0_to_fp16 = const()[name = tensor("input_381_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313450560)))]; + tensor input_381_beta_0_to_fp16 = const()[name = tensor("input_381_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313453184)))]; + tensor input_381_epsilon_0_to_fp16 = const()[name = tensor("input_381_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_381_cast_fp16 = batch_norm(beta = input_381_beta_0_to_fp16, epsilon = input_381_epsilon_0_to_fp16, gamma = input_381_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor("input_381_cast_fp16")]; + tensor var_5246 = const()[name = tensor("op_5246"), val = tensor([1, 1])]; + tensor var_5248 = const()[name = tensor("op_5248"), val = tensor([1, 1])]; + tensor pretrained_out_257_pad_type_0 = const()[name = tensor("pretrained_out_257_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_257_pad_0 = const()[name = tensor("pretrained_out_257_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313455808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316732672))), name = tensor("layers_12_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_12_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316732800)))]; + tensor pretrained_out_257_cast_fp16 = conv(bias = layers_12_fc1_pretrained_bias_to_fp16, dilations = var_5248, groups = var_4915, pad = pretrained_out_257_pad_0, pad_type = pretrained_out_257_pad_type_0, strides = var_5246, weight = layers_12_fc1_pretrained_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = tensor("pretrained_out_257_cast_fp16")]; + tensor var_5252 = const()[name = tensor("op_5252"), val = tensor([1, 1])]; + tensor var_5254 = const()[name = tensor("op_5254"), val = tensor([1, 1])]; + tensor input_383_pad_type_0 = const()[name = tensor("input_383_pad_type_0"), val = tensor("custom")]; + tensor input_383_pad_0 = const()[name = tensor("input_383_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_12_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316743104)))]; + tensor input_383_cast_fp16 = conv(dilations = var_5254, groups = var_4915, pad = input_383_pad_0, pad_type = input_383_pad_type_0, strides = var_5252, weight = layers_12_fc1_loraA_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("input_383_cast_fp16")]; + tensor var_5258 = const()[name = tensor("op_5258"), val = tensor([1, 1])]; + tensor var_5260 = const()[name = tensor("op_5260"), val = tensor([1, 1])]; + tensor lora_out_513_pad_type_0 = const()[name = tensor("lora_out_513_pad_type_0"), val = tensor("custom")]; + tensor lora_out_513_pad_0 = const()[name = tensor("lora_out_513_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_515_weight_0_to_fp16 = const()[name = tensor("lora_out_515_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316784128)))]; + tensor lora_out_515_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_5260, groups = var_4915, pad = lora_out_513_pad_0, pad_type = lora_out_513_pad_type_0, strides = var_5258, weight = lora_out_515_weight_0_to_fp16, x = input_383_cast_fp16)[name = tensor("lora_out_515_cast_fp16")]; + tensor input_385_cast_fp16 = add(x = pretrained_out_257_cast_fp16, y = lora_out_515_cast_fp16)[name = tensor("input_385_cast_fp16")]; + tensor input_387_mode_0 = const()[name = tensor("input_387_mode_0"), val = tensor("EXACT")]; + tensor input_387_cast_fp16 = gelu(mode = input_387_mode_0, x = input_385_cast_fp16)[name = tensor("input_387_cast_fp16")]; + tensor var_5272 = const()[name = tensor("op_5272"), val = tensor([1, 1])]; + tensor var_5274 = const()[name = tensor("op_5274"), val = tensor([1, 1])]; + tensor pretrained_out_259_pad_type_0 = const()[name = tensor("pretrained_out_259_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_259_pad_0 = const()[name = tensor("pretrained_out_259_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316948032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320224896))), name = tensor("layers_12_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_12_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_12_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320225024)))]; + tensor pretrained_out_259_cast_fp16 = conv(bias = layers_12_fc2_pretrained_bias_to_fp16, dilations = var_5274, groups = var_4915, pad = pretrained_out_259_pad_0, pad_type = pretrained_out_259_pad_type_0, strides = var_5272, weight = layers_12_fc2_pretrained_weight_to_fp16_palettized, x = input_387_cast_fp16)[name = tensor("pretrained_out_259_cast_fp16")]; + tensor var_5278 = const()[name = tensor("op_5278"), val = tensor([1, 1])]; + tensor var_5280 = const()[name = tensor("op_5280"), val = tensor([1, 1])]; + tensor input_389_pad_type_0 = const()[name = tensor("input_389_pad_type_0"), val = tensor("custom")]; + tensor input_389_pad_0 = const()[name = tensor("input_389_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_12_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_12_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320227648)))]; + tensor input_389_cast_fp16 = conv(dilations = var_5280, groups = var_4915, pad = input_389_pad_0, pad_type = input_389_pad_type_0, strides = var_5278, weight = layers_12_fc2_loraA_weight_to_fp16, x = input_387_cast_fp16)[name = tensor("input_389_cast_fp16")]; + tensor var_5284 = const()[name = tensor("op_5284"), val = tensor([1, 1])]; + tensor var_5286 = const()[name = tensor("op_5286"), val = tensor([1, 1])]; + tensor lora_out_517_pad_type_0 = const()[name = tensor("lora_out_517_pad_type_0"), val = tensor("custom")]; + tensor lora_out_517_pad_0 = const()[name = tensor("lora_out_517_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_519_weight_0_to_fp16 = const()[name = tensor("lora_out_519_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320391552)))]; + tensor lora_out_519_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5286, groups = var_4915, pad = lora_out_517_pad_0, pad_type = lora_out_517_pad_type_0, strides = var_5284, weight = lora_out_519_weight_0_to_fp16, x = input_389_cast_fp16)[name = tensor("lora_out_519_cast_fp16")]; + tensor hidden_states_27_cast_fp16 = add(x = pretrained_out_259_cast_fp16, y = lora_out_519_cast_fp16)[name = tensor("hidden_states_27_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor var_5302 = const()[name = tensor("op_5302"), val = tensor(3)]; + tensor var_5309 = const()[name = tensor("op_5309"), val = tensor(1)]; + tensor var_5310 = const()[name = tensor("op_5310"), val = tensor(true)]; + tensor var_5322 = const()[name = tensor("op_5322"), val = tensor([1])]; + tensor channels_mean_79_cast_fp16 = reduce_mean(axes = var_5322, keep_dims = var_5310, x = inputs_79_cast_fp16)[name = tensor("channels_mean_79_cast_fp16")]; + tensor zero_mean_79_cast_fp16 = sub(x = inputs_79_cast_fp16, y = channels_mean_79_cast_fp16)[name = tensor("zero_mean_79_cast_fp16")]; + tensor zero_mean_sq_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = zero_mean_79_cast_fp16)[name = tensor("zero_mean_sq_79_cast_fp16")]; + tensor var_5326 = const()[name = tensor("op_5326"), val = tensor([1])]; + tensor var_5327_cast_fp16 = reduce_mean(axes = var_5326, keep_dims = var_5310, x = zero_mean_sq_79_cast_fp16)[name = tensor("op_5327_cast_fp16")]; + tensor var_5328_to_fp16 = const()[name = tensor("op_5328_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5329_cast_fp16 = add(x = var_5327_cast_fp16, y = var_5328_to_fp16)[name = tensor("op_5329_cast_fp16")]; + tensor denom_79_epsilon_0 = const()[name = tensor("denom_79_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_79_cast_fp16 = rsqrt(epsilon = denom_79_epsilon_0, x = var_5329_cast_fp16)[name = tensor("denom_79_cast_fp16")]; + tensor out_79_cast_fp16 = mul(x = zero_mean_79_cast_fp16, y = denom_79_cast_fp16)[name = tensor("out_79_cast_fp16")]; + tensor obj_183_gamma_0_to_fp16 = const()[name = tensor("obj_183_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320432576)))]; + tensor obj_183_beta_0_to_fp16 = const()[name = tensor("obj_183_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320435200)))]; + tensor obj_183_epsilon_0_to_fp16 = const()[name = tensor("obj_183_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_183_cast_fp16 = batch_norm(beta = obj_183_beta_0_to_fp16, epsilon = obj_183_epsilon_0_to_fp16, gamma = obj_183_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor("obj_183_cast_fp16")]; + tensor var_5347 = const()[name = tensor("op_5347"), val = tensor([1, 1])]; + tensor var_5349 = const()[name = tensor("op_5349"), val = tensor([1, 1])]; + tensor pretrained_out_261_pad_type_0 = const()[name = tensor("pretrained_out_261_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_261_pad_0 = const()[name = tensor("pretrained_out_261_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320437824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321257088))), name = tensor("layers_13_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_13_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321257216)))]; + tensor pretrained_out_261_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5349, groups = var_5309, pad = pretrained_out_261_pad_0, pad_type = pretrained_out_261_pad_type_0, strides = var_5347, weight = layers_13_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_183_cast_fp16)[name = tensor("pretrained_out_261_cast_fp16")]; + tensor var_5353 = const()[name = tensor("op_5353"), val = tensor([1, 1])]; + tensor var_5355 = const()[name = tensor("op_5355"), val = tensor([1, 1])]; + tensor input_391_pad_type_0 = const()[name = tensor("input_391_pad_type_0"), val = tensor("custom")]; + tensor input_391_pad_0 = const()[name = tensor("input_391_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321259840)))]; + tensor input_391_cast_fp16 = conv(dilations = var_5355, groups = var_5309, pad = input_391_pad_0, pad_type = input_391_pad_type_0, strides = var_5353, weight = layers_13_self_attn_q_proj_loraA_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("input_391_cast_fp16")]; + tensor var_5359 = const()[name = tensor("op_5359"), val = tensor([1, 1])]; + tensor var_5361 = const()[name = tensor("op_5361"), val = tensor([1, 1])]; + tensor lora_out_521_pad_type_0 = const()[name = tensor("lora_out_521_pad_type_0"), val = tensor("custom")]; + tensor lora_out_521_pad_0 = const()[name = tensor("lora_out_521_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_523_weight_0_to_fp16 = const()[name = tensor("lora_out_523_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321300864)))]; + tensor lora_out_523_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5361, groups = var_5309, pad = lora_out_521_pad_0, pad_type = lora_out_521_pad_type_0, strides = var_5359, weight = lora_out_523_weight_0_to_fp16, x = input_391_cast_fp16)[name = tensor("lora_out_523_cast_fp16")]; + tensor query_53_cast_fp16 = add(x = pretrained_out_261_cast_fp16, y = lora_out_523_cast_fp16)[name = tensor("query_53_cast_fp16")]; + tensor var_5371 = const()[name = tensor("op_5371"), val = tensor([1, 1])]; + tensor var_5373 = const()[name = tensor("op_5373"), val = tensor([1, 1])]; + tensor pretrained_out_263_pad_type_0 = const()[name = tensor("pretrained_out_263_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_263_pad_0 = const()[name = tensor("pretrained_out_263_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321341888))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322161152))), name = tensor("layers_13_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_263_cast_fp16 = conv(dilations = var_5373, groups = var_5309, pad = pretrained_out_263_pad_0, pad_type = pretrained_out_263_pad_type_0, strides = var_5371, weight = layers_13_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_183_cast_fp16)[name = tensor("pretrained_out_263_cast_fp16")]; + tensor var_5377 = const()[name = tensor("op_5377"), val = tensor([1, 1])]; + tensor var_5379 = const()[name = tensor("op_5379"), val = tensor([1, 1])]; + tensor input_393_pad_type_0 = const()[name = tensor("input_393_pad_type_0"), val = tensor("custom")]; + tensor input_393_pad_0 = const()[name = tensor("input_393_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322161280)))]; + tensor input_393_cast_fp16 = conv(dilations = var_5379, groups = var_5309, pad = input_393_pad_0, pad_type = input_393_pad_type_0, strides = var_5377, weight = layers_13_self_attn_k_proj_loraA_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor var_5383 = const()[name = tensor("op_5383"), val = tensor([1, 1])]; + tensor var_5385 = const()[name = tensor("op_5385"), val = tensor([1, 1])]; + tensor lora_out_525_pad_type_0 = const()[name = tensor("lora_out_525_pad_type_0"), val = tensor("custom")]; + tensor lora_out_525_pad_0 = const()[name = tensor("lora_out_525_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_527_weight_0_to_fp16 = const()[name = tensor("lora_out_527_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322202304)))]; + tensor lora_out_527_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5385, groups = var_5309, pad = lora_out_525_pad_0, pad_type = lora_out_525_pad_type_0, strides = var_5383, weight = lora_out_527_weight_0_to_fp16, x = input_393_cast_fp16)[name = tensor("lora_out_527_cast_fp16")]; + tensor current_key_27_cast_fp16 = add(x = pretrained_out_263_cast_fp16, y = lora_out_527_cast_fp16)[name = tensor("current_key_27_cast_fp16")]; + tensor var_5396 = const()[name = tensor("op_5396"), val = tensor([1, 1])]; + tensor var_5398 = const()[name = tensor("op_5398"), val = tensor([1, 1])]; + tensor pretrained_out_265_pad_type_0 = const()[name = tensor("pretrained_out_265_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_265_pad_0 = const()[name = tensor("pretrained_out_265_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322243328))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323062592))), name = tensor("layers_13_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_13_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323062720)))]; + tensor pretrained_out_265_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5398, groups = var_5309, pad = pretrained_out_265_pad_0, pad_type = pretrained_out_265_pad_type_0, strides = var_5396, weight = layers_13_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_183_cast_fp16)[name = tensor("pretrained_out_265_cast_fp16")]; + tensor var_5402 = const()[name = tensor("op_5402"), val = tensor([1, 1])]; + tensor var_5404 = const()[name = tensor("op_5404"), val = tensor([1, 1])]; + tensor input_395_pad_type_0 = const()[name = tensor("input_395_pad_type_0"), val = tensor("custom")]; + tensor input_395_pad_0 = const()[name = tensor("input_395_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323065344)))]; + tensor input_395_cast_fp16 = conv(dilations = var_5404, groups = var_5309, pad = input_395_pad_0, pad_type = input_395_pad_type_0, strides = var_5402, weight = layers_13_self_attn_v_proj_loraA_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("input_395_cast_fp16")]; + tensor var_5408 = const()[name = tensor("op_5408"), val = tensor([1, 1])]; + tensor var_5410 = const()[name = tensor("op_5410"), val = tensor([1, 1])]; + tensor lora_out_529_pad_type_0 = const()[name = tensor("lora_out_529_pad_type_0"), val = tensor("custom")]; + tensor lora_out_529_pad_0 = const()[name = tensor("lora_out_529_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_531_weight_0_to_fp16 = const()[name = tensor("lora_out_531_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323106368)))]; + tensor lora_out_531_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5410, groups = var_5309, pad = lora_out_529_pad_0, pad_type = lora_out_529_pad_type_0, strides = var_5408, weight = lora_out_531_weight_0_to_fp16, x = input_395_cast_fp16)[name = tensor("lora_out_531_cast_fp16")]; + tensor current_value_27_cast_fp16 = add(x = pretrained_out_265_cast_fp16, y = lora_out_531_cast_fp16)[name = tensor("current_value_27_cast_fp16")]; + tensor var_5420_cast_fp16 = mul(x = current_key_27_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_5420_cast_fp16")]; + tensor var_5422_cast_fp16 = mul(x = var_103_cast_fp16_13, y = var_295_cast_fp16)[name = tensor("op_5422_cast_fp16")]; + tensor key_53_cast_fp16 = add(x = var_5420_cast_fp16, y = var_5422_cast_fp16)[name = tensor("key_53_cast_fp16")]; + tensor var_5424_cast_fp16 = mul(x = current_value_27_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_5424_cast_fp16")]; + tensor var_5426_cast_fp16 = mul(x = var_138_cast_fp16_13, y = var_295_cast_fp16)[name = tensor("op_5426_cast_fp16")]; + tensor value_53_cast_fp16 = add(x = var_5424_cast_fp16, y = var_5426_cast_fp16)[name = tensor("value_53_cast_fp16")]; + tensor var_5429 = const()[name = tensor("op_5429"), val = tensor([1, 20, 64, -1])]; + tensor var_5430_cast_fp16 = reshape(shape = var_5429, x = query_53_cast_fp16)[name = tensor("op_5430_cast_fp16")]; + tensor var_5431_to_fp16 = const()[name = tensor("op_5431_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5432_cast_fp16 = mul(x = var_5430_cast_fp16, y = var_5431_to_fp16)[name = tensor("op_5432_cast_fp16")]; + tensor var_5433 = const()[name = tensor("op_5433"), val = tensor([1, 20, 64, -1])]; + tensor var_5434_cast_fp16 = reshape(shape = var_5433, x = key_53_cast_fp16)[name = tensor("op_5434_cast_fp16")]; + tensor mh_w_79_transpose_x_0 = const()[name = tensor("mh_w_79_transpose_x_0"), val = tensor(true)]; + tensor mh_w_79_transpose_y_0 = const()[name = tensor("mh_w_79_transpose_y_0"), val = tensor(false)]; + tensor mh_w_79_cast_fp16 = matmul(transpose_x = mh_w_79_transpose_x_0, transpose_y = mh_w_79_transpose_y_0, x = var_5432_cast_fp16, y = var_5434_cast_fp16)[name = tensor("mh_w_79_cast_fp16")]; + tensor mh_w_81_cast_fp16 = add(x = mh_w_79_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_81_cast_fp16")]; + tensor var_5442_cast_fp16 = softmax(axis = var_5302, x = mh_w_81_cast_fp16)[name = tensor("op_5442_cast_fp16")]; + tensor var_5443 = const()[name = tensor("op_5443"), val = tensor([1, 20, 64, -1])]; + tensor var_5444_cast_fp16 = reshape(shape = var_5443, x = value_53_cast_fp16)[name = tensor("op_5444_cast_fp16")]; + tensor attn_53_transpose_x_0 = const()[name = tensor("attn_53_transpose_x_0"), val = tensor(false)]; + tensor attn_53_transpose_y_0 = const()[name = tensor("attn_53_transpose_y_0"), val = tensor(true)]; + tensor attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_5444_cast_fp16, y = var_5442_cast_fp16)[name = tensor("attn_53_cast_fp16")]; + tensor var_5447 = const()[name = tensor("op_5447"), val = tensor([1, 1280, 1, -1])]; + tensor input_397_cast_fp16 = reshape(shape = var_5447, x = attn_53_cast_fp16)[name = tensor("input_397_cast_fp16")]; + tensor var_5454 = const()[name = tensor("op_5454"), val = tensor([1, 1])]; + tensor var_5456 = const()[name = tensor("op_5456"), val = tensor([1, 1])]; + tensor pretrained_out_267_pad_type_0 = const()[name = tensor("pretrained_out_267_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_267_pad_0 = const()[name = tensor("pretrained_out_267_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323147392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323966656))), name = tensor("layers_13_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_13_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323966784)))]; + tensor pretrained_out_267_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5456, groups = var_5309, pad = pretrained_out_267_pad_0, pad_type = pretrained_out_267_pad_type_0, strides = var_5454, weight = layers_13_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = tensor("pretrained_out_267_cast_fp16")]; + tensor var_5460 = const()[name = tensor("op_5460"), val = tensor([1, 1])]; + tensor var_5462 = const()[name = tensor("op_5462"), val = tensor([1, 1])]; + tensor input_399_pad_type_0 = const()[name = tensor("input_399_pad_type_0"), val = tensor("custom")]; + tensor input_399_pad_0 = const()[name = tensor("input_399_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323969408)))]; + tensor input_399_cast_fp16 = conv(dilations = var_5462, groups = var_5309, pad = input_399_pad_0, pad_type = input_399_pad_type_0, strides = var_5460, weight = layers_13_self_attn_o_proj_loraA_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("input_399_cast_fp16")]; + tensor var_5466 = const()[name = tensor("op_5466"), val = tensor([1, 1])]; + tensor var_5468 = const()[name = tensor("op_5468"), val = tensor([1, 1])]; + tensor lora_out_533_pad_type_0 = const()[name = tensor("lora_out_533_pad_type_0"), val = tensor("custom")]; + tensor lora_out_533_pad_0 = const()[name = tensor("lora_out_533_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_535_weight_0_to_fp16 = const()[name = tensor("lora_out_535_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324010432)))]; + tensor lora_out_535_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5468, groups = var_5309, pad = lora_out_533_pad_0, pad_type = lora_out_533_pad_type_0, strides = var_5466, weight = lora_out_535_weight_0_to_fp16, x = input_399_cast_fp16)[name = tensor("lora_out_535_cast_fp16")]; + tensor obj_189_cast_fp16 = add(x = pretrained_out_267_cast_fp16, y = lora_out_535_cast_fp16)[name = tensor("obj_189_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = obj_189_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor var_5481 = const()[name = tensor("op_5481"), val = tensor([1])]; + tensor channels_mean_81_cast_fp16 = reduce_mean(axes = var_5481, keep_dims = var_5310, x = inputs_81_cast_fp16)[name = tensor("channels_mean_81_cast_fp16")]; + tensor zero_mean_81_cast_fp16 = sub(x = inputs_81_cast_fp16, y = channels_mean_81_cast_fp16)[name = tensor("zero_mean_81_cast_fp16")]; + tensor zero_mean_sq_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = zero_mean_81_cast_fp16)[name = tensor("zero_mean_sq_81_cast_fp16")]; + tensor var_5485 = const()[name = tensor("op_5485"), val = tensor([1])]; + tensor var_5486_cast_fp16 = reduce_mean(axes = var_5485, keep_dims = var_5310, x = zero_mean_sq_81_cast_fp16)[name = tensor("op_5486_cast_fp16")]; + tensor var_5487_to_fp16 = const()[name = tensor("op_5487_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5488_cast_fp16 = add(x = var_5486_cast_fp16, y = var_5487_to_fp16)[name = tensor("op_5488_cast_fp16")]; + tensor denom_81_epsilon_0 = const()[name = tensor("denom_81_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_81_cast_fp16 = rsqrt(epsilon = denom_81_epsilon_0, x = var_5488_cast_fp16)[name = tensor("denom_81_cast_fp16")]; + tensor out_81_cast_fp16 = mul(x = zero_mean_81_cast_fp16, y = denom_81_cast_fp16)[name = tensor("out_81_cast_fp16")]; + tensor obj_191_gamma_0_to_fp16 = const()[name = tensor("obj_191_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324051456)))]; + tensor obj_191_beta_0_to_fp16 = const()[name = tensor("obj_191_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324054080)))]; + tensor obj_191_epsilon_0_to_fp16 = const()[name = tensor("obj_191_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_191_cast_fp16 = batch_norm(beta = obj_191_beta_0_to_fp16, epsilon = obj_191_epsilon_0_to_fp16, gamma = obj_191_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor("obj_191_cast_fp16")]; + tensor var_5506 = const()[name = tensor("op_5506"), val = tensor([1, 1])]; + tensor var_5508 = const()[name = tensor("op_5508"), val = tensor([1, 1])]; + tensor pretrained_out_269_pad_type_0 = const()[name = tensor("pretrained_out_269_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_269_pad_0 = const()[name = tensor("pretrained_out_269_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324056704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324875968))), name = tensor("layers_13_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_13_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324876096)))]; + tensor pretrained_out_269_cast_fp16 = conv(bias = layers_13_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5508, groups = var_5309, pad = pretrained_out_269_pad_0, pad_type = pretrained_out_269_pad_type_0, strides = var_5506, weight = layers_13_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_191_cast_fp16)[name = tensor("pretrained_out_269_cast_fp16")]; + tensor var_5512 = const()[name = tensor("op_5512"), val = tensor([1, 1])]; + tensor var_5514 = const()[name = tensor("op_5514"), val = tensor([1, 1])]; + tensor input_401_pad_type_0 = const()[name = tensor("input_401_pad_type_0"), val = tensor("custom")]; + tensor input_401_pad_0 = const()[name = tensor("input_401_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324878720)))]; + tensor input_401_cast_fp16 = conv(dilations = var_5514, groups = var_5309, pad = input_401_pad_0, pad_type = input_401_pad_type_0, strides = var_5512, weight = layers_13_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_191_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor var_5518 = const()[name = tensor("op_5518"), val = tensor([1, 1])]; + tensor var_5520 = const()[name = tensor("op_5520"), val = tensor([1, 1])]; + tensor lora_out_537_pad_type_0 = const()[name = tensor("lora_out_537_pad_type_0"), val = tensor("custom")]; + tensor lora_out_537_pad_0 = const()[name = tensor("lora_out_537_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_539_weight_0_to_fp16 = const()[name = tensor("lora_out_539_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324919744)))]; + tensor lora_out_539_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5520, groups = var_5309, pad = lora_out_537_pad_0, pad_type = lora_out_537_pad_type_0, strides = var_5518, weight = lora_out_539_weight_0_to_fp16, x = input_401_cast_fp16)[name = tensor("lora_out_539_cast_fp16")]; + tensor query_55_cast_fp16 = add(x = pretrained_out_269_cast_fp16, y = lora_out_539_cast_fp16)[name = tensor("query_55_cast_fp16")]; + tensor var_5530 = const()[name = tensor("op_5530"), val = tensor([1, 1])]; + tensor var_5532 = const()[name = tensor("op_5532"), val = tensor([1, 1])]; + tensor pretrained_out_271_pad_type_0 = const()[name = tensor("pretrained_out_271_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_271_pad_0 = const()[name = tensor("pretrained_out_271_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324960768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325780032))), name = tensor("layers_13_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_271_cast_fp16 = conv(dilations = var_5532, groups = var_5309, pad = pretrained_out_271_pad_0, pad_type = pretrained_out_271_pad_type_0, strides = var_5530, weight = layers_13_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_271_cast_fp16")]; + tensor var_5536 = const()[name = tensor("op_5536"), val = tensor([1, 1])]; + tensor var_5538 = const()[name = tensor("op_5538"), val = tensor([1, 1])]; + tensor input_403_pad_type_0 = const()[name = tensor("input_403_pad_type_0"), val = tensor("custom")]; + tensor input_403_pad_0 = const()[name = tensor("input_403_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325780160)))]; + tensor input_403_cast_fp16 = conv(dilations = var_5538, groups = var_5309, pad = input_403_pad_0, pad_type = input_403_pad_type_0, strides = var_5536, weight = layers_13_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_403_cast_fp16")]; + tensor var_5542 = const()[name = tensor("op_5542"), val = tensor([1, 1])]; + tensor var_5544 = const()[name = tensor("op_5544"), val = tensor([1, 1])]; + tensor lora_out_541_pad_type_0 = const()[name = tensor("lora_out_541_pad_type_0"), val = tensor("custom")]; + tensor lora_out_541_pad_0 = const()[name = tensor("lora_out_541_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_543_weight_0_to_fp16 = const()[name = tensor("lora_out_543_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325821184)))]; + tensor lora_out_543_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5544, groups = var_5309, pad = lora_out_541_pad_0, pad_type = lora_out_541_pad_type_0, strides = var_5542, weight = lora_out_543_weight_0_to_fp16, x = input_403_cast_fp16)[name = tensor("lora_out_543_cast_fp16")]; + tensor key_55_cast_fp16 = add(x = pretrained_out_271_cast_fp16, y = lora_out_543_cast_fp16)[name = tensor("key_55_cast_fp16")]; + tensor var_5555 = const()[name = tensor("op_5555"), val = tensor([1, 1])]; + tensor var_5557 = const()[name = tensor("op_5557"), val = tensor([1, 1])]; + tensor pretrained_out_273_pad_type_0 = const()[name = tensor("pretrained_out_273_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_273_pad_0 = const()[name = tensor("pretrained_out_273_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325862208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326681472))), name = tensor("layers_13_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_13_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326681600)))]; + tensor pretrained_out_273_cast_fp16 = conv(bias = layers_13_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5557, groups = var_5309, pad = pretrained_out_273_pad_0, pad_type = pretrained_out_273_pad_type_0, strides = var_5555, weight = layers_13_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_273_cast_fp16")]; + tensor var_5561 = const()[name = tensor("op_5561"), val = tensor([1, 1])]; + tensor var_5563 = const()[name = tensor("op_5563"), val = tensor([1, 1])]; + tensor input_405_pad_type_0 = const()[name = tensor("input_405_pad_type_0"), val = tensor("custom")]; + tensor input_405_pad_0 = const()[name = tensor("input_405_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326684224)))]; + tensor input_405_cast_fp16 = conv(dilations = var_5563, groups = var_5309, pad = input_405_pad_0, pad_type = input_405_pad_type_0, strides = var_5561, weight = layers_13_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_405_cast_fp16")]; + tensor var_5567 = const()[name = tensor("op_5567"), val = tensor([1, 1])]; + tensor var_5569 = const()[name = tensor("op_5569"), val = tensor([1, 1])]; + tensor lora_out_545_pad_type_0 = const()[name = tensor("lora_out_545_pad_type_0"), val = tensor("custom")]; + tensor lora_out_545_pad_0 = const()[name = tensor("lora_out_545_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_547_weight_0_to_fp16 = const()[name = tensor("lora_out_547_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326725248)))]; + tensor lora_out_547_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5569, groups = var_5309, pad = lora_out_545_pad_0, pad_type = lora_out_545_pad_type_0, strides = var_5567, weight = lora_out_547_weight_0_to_fp16, x = input_405_cast_fp16)[name = tensor("lora_out_547_cast_fp16")]; + tensor value_55_cast_fp16 = add(x = pretrained_out_273_cast_fp16, y = lora_out_547_cast_fp16)[name = tensor("value_55_cast_fp16")]; + tensor var_5576 = const()[name = tensor("op_5576"), val = tensor([1, 20, 64, -1])]; + tensor var_5577_cast_fp16 = reshape(shape = var_5576, x = query_55_cast_fp16)[name = tensor("op_5577_cast_fp16")]; + tensor var_5578_to_fp16 = const()[name = tensor("op_5578_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5579_cast_fp16 = mul(x = var_5577_cast_fp16, y = var_5578_to_fp16)[name = tensor("op_5579_cast_fp16")]; + tensor var_5580 = const()[name = tensor("op_5580"), val = tensor([1, 20, 64, -1])]; + tensor var_5581_cast_fp16 = reshape(shape = var_5580, x = key_55_cast_fp16)[name = tensor("op_5581_cast_fp16")]; + tensor mh_w_83_transpose_x_0 = const()[name = tensor("mh_w_83_transpose_x_0"), val = tensor(true)]; + tensor mh_w_83_transpose_y_0 = const()[name = tensor("mh_w_83_transpose_y_0"), val = tensor(false)]; + tensor mh_w_83_cast_fp16 = matmul(transpose_x = mh_w_83_transpose_x_0, transpose_y = mh_w_83_transpose_y_0, x = var_5579_cast_fp16, y = var_5581_cast_fp16)[name = tensor("mh_w_83_cast_fp16")]; + tensor obj_195_cast_fp16 = softmax(axis = var_5302, x = mh_w_83_cast_fp16)[name = tensor("obj_195_cast_fp16")]; + tensor var_5585 = const()[name = tensor("op_5585"), val = tensor([1, 20, 64, -1])]; + tensor var_5586_cast_fp16 = reshape(shape = var_5585, x = value_55_cast_fp16)[name = tensor("op_5586_cast_fp16")]; + tensor attn_55_transpose_x_0 = const()[name = tensor("attn_55_transpose_x_0"), val = tensor(false)]; + tensor attn_55_transpose_y_0 = const()[name = tensor("attn_55_transpose_y_0"), val = tensor(true)]; + tensor attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_5586_cast_fp16, y = obj_195_cast_fp16)[name = tensor("attn_55_cast_fp16")]; + tensor var_5589 = const()[name = tensor("op_5589"), val = tensor([1, 1280, 1, -1])]; + tensor input_407_cast_fp16 = reshape(shape = var_5589, x = attn_55_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor var_5596 = const()[name = tensor("op_5596"), val = tensor([1, 1])]; + tensor var_5598 = const()[name = tensor("op_5598"), val = tensor([1, 1])]; + tensor pretrained_out_275_pad_type_0 = const()[name = tensor("pretrained_out_275_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_275_pad_0 = const()[name = tensor("pretrained_out_275_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(326766272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327585536))), name = tensor("layers_13_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_13_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327585664)))]; + tensor pretrained_out_275_cast_fp16 = conv(bias = layers_13_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5598, groups = var_5309, pad = pretrained_out_275_pad_0, pad_type = pretrained_out_275_pad_type_0, strides = var_5596, weight = layers_13_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = tensor("pretrained_out_275_cast_fp16")]; + tensor var_5602 = const()[name = tensor("op_5602"), val = tensor([1, 1])]; + tensor var_5604 = const()[name = tensor("op_5604"), val = tensor([1, 1])]; + tensor input_409_pad_type_0 = const()[name = tensor("input_409_pad_type_0"), val = tensor("custom")]; + tensor input_409_pad_0 = const()[name = tensor("input_409_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327588288)))]; + tensor input_409_cast_fp16 = conv(dilations = var_5604, groups = var_5309, pad = input_409_pad_0, pad_type = input_409_pad_type_0, strides = var_5602, weight = layers_13_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("input_409_cast_fp16")]; + tensor var_5608 = const()[name = tensor("op_5608"), val = tensor([1, 1])]; + tensor var_5610 = const()[name = tensor("op_5610"), val = tensor([1, 1])]; + tensor lora_out_549_pad_type_0 = const()[name = tensor("lora_out_549_pad_type_0"), val = tensor("custom")]; + tensor lora_out_549_pad_0 = const()[name = tensor("lora_out_549_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_551_weight_0_to_fp16 = const()[name = tensor("lora_out_551_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327629312)))]; + tensor lora_out_551_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5610, groups = var_5309, pad = lora_out_549_pad_0, pad_type = lora_out_549_pad_type_0, strides = var_5608, weight = lora_out_551_weight_0_to_fp16, x = input_409_cast_fp16)[name = tensor("lora_out_551_cast_fp16")]; + tensor obj_193_cast_fp16 = add(x = pretrained_out_275_cast_fp16, y = lora_out_551_cast_fp16)[name = tensor("obj_193_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_193_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor var_5622 = const()[name = tensor("op_5622"), val = tensor([1])]; + tensor channels_mean_83_cast_fp16 = reduce_mean(axes = var_5622, keep_dims = var_5310, x = inputs_83_cast_fp16)[name = tensor("channels_mean_83_cast_fp16")]; + tensor zero_mean_83_cast_fp16 = sub(x = inputs_83_cast_fp16, y = channels_mean_83_cast_fp16)[name = tensor("zero_mean_83_cast_fp16")]; + tensor zero_mean_sq_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = zero_mean_83_cast_fp16)[name = tensor("zero_mean_sq_83_cast_fp16")]; + tensor var_5626 = const()[name = tensor("op_5626"), val = tensor([1])]; + tensor var_5627_cast_fp16 = reduce_mean(axes = var_5626, keep_dims = var_5310, x = zero_mean_sq_83_cast_fp16)[name = tensor("op_5627_cast_fp16")]; + tensor var_5628_to_fp16 = const()[name = tensor("op_5628_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5629_cast_fp16 = add(x = var_5627_cast_fp16, y = var_5628_to_fp16)[name = tensor("op_5629_cast_fp16")]; + tensor denom_83_epsilon_0 = const()[name = tensor("denom_83_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_83_cast_fp16 = rsqrt(epsilon = denom_83_epsilon_0, x = var_5629_cast_fp16)[name = tensor("denom_83_cast_fp16")]; + tensor out_83_cast_fp16 = mul(x = zero_mean_83_cast_fp16, y = denom_83_cast_fp16)[name = tensor("out_83_cast_fp16")]; + tensor input_411_gamma_0_to_fp16 = const()[name = tensor("input_411_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327670336)))]; + tensor input_411_beta_0_to_fp16 = const()[name = tensor("input_411_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327672960)))]; + tensor input_411_epsilon_0_to_fp16 = const()[name = tensor("input_411_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_411_cast_fp16 = batch_norm(beta = input_411_beta_0_to_fp16, epsilon = input_411_epsilon_0_to_fp16, gamma = input_411_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor("input_411_cast_fp16")]; + tensor var_5643 = const()[name = tensor("op_5643"), val = tensor([1, 1])]; + tensor var_5645 = const()[name = tensor("op_5645"), val = tensor([1, 1])]; + tensor pretrained_out_277_pad_type_0 = const()[name = tensor("pretrained_out_277_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_277_pad_0 = const()[name = tensor("pretrained_out_277_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327675584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(330952448))), name = tensor("layers_13_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_13_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(330952576)))]; + tensor pretrained_out_277_cast_fp16 = conv(bias = layers_13_fc1_pretrained_bias_to_fp16, dilations = var_5645, groups = var_5309, pad = pretrained_out_277_pad_0, pad_type = pretrained_out_277_pad_type_0, strides = var_5643, weight = layers_13_fc1_pretrained_weight_to_fp16_palettized, x = input_411_cast_fp16)[name = tensor("pretrained_out_277_cast_fp16")]; + tensor var_5649 = const()[name = tensor("op_5649"), val = tensor([1, 1])]; + tensor var_5651 = const()[name = tensor("op_5651"), val = tensor([1, 1])]; + tensor input_413_pad_type_0 = const()[name = tensor("input_413_pad_type_0"), val = tensor("custom")]; + tensor input_413_pad_0 = const()[name = tensor("input_413_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_13_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(330962880)))]; + tensor input_413_cast_fp16 = conv(dilations = var_5651, groups = var_5309, pad = input_413_pad_0, pad_type = input_413_pad_type_0, strides = var_5649, weight = layers_13_fc1_loraA_weight_to_fp16, x = input_411_cast_fp16)[name = tensor("input_413_cast_fp16")]; + tensor var_5655 = const()[name = tensor("op_5655"), val = tensor([1, 1])]; + tensor var_5657 = const()[name = tensor("op_5657"), val = tensor([1, 1])]; + tensor lora_out_553_pad_type_0 = const()[name = tensor("lora_out_553_pad_type_0"), val = tensor("custom")]; + tensor lora_out_553_pad_0 = const()[name = tensor("lora_out_553_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_555_weight_0_to_fp16 = const()[name = tensor("lora_out_555_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(331003904)))]; + tensor lora_out_555_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_5657, groups = var_5309, pad = lora_out_553_pad_0, pad_type = lora_out_553_pad_type_0, strides = var_5655, weight = lora_out_555_weight_0_to_fp16, x = input_413_cast_fp16)[name = tensor("lora_out_555_cast_fp16")]; + tensor input_415_cast_fp16 = add(x = pretrained_out_277_cast_fp16, y = lora_out_555_cast_fp16)[name = tensor("input_415_cast_fp16")]; + tensor input_417_mode_0 = const()[name = tensor("input_417_mode_0"), val = tensor("EXACT")]; + tensor input_417_cast_fp16 = gelu(mode = input_417_mode_0, x = input_415_cast_fp16)[name = tensor("input_417_cast_fp16")]; + tensor var_5669 = const()[name = tensor("op_5669"), val = tensor([1, 1])]; + tensor var_5671 = const()[name = tensor("op_5671"), val = tensor([1, 1])]; + tensor pretrained_out_279_pad_type_0 = const()[name = tensor("pretrained_out_279_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_279_pad_0 = const()[name = tensor("pretrained_out_279_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(331167808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334444672))), name = tensor("layers_13_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_13_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_13_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334444800)))]; + tensor pretrained_out_279_cast_fp16 = conv(bias = layers_13_fc2_pretrained_bias_to_fp16, dilations = var_5671, groups = var_5309, pad = pretrained_out_279_pad_0, pad_type = pretrained_out_279_pad_type_0, strides = var_5669, weight = layers_13_fc2_pretrained_weight_to_fp16_palettized, x = input_417_cast_fp16)[name = tensor("pretrained_out_279_cast_fp16")]; + tensor var_5675 = const()[name = tensor("op_5675"), val = tensor([1, 1])]; + tensor var_5677 = const()[name = tensor("op_5677"), val = tensor([1, 1])]; + tensor input_419_pad_type_0 = const()[name = tensor("input_419_pad_type_0"), val = tensor("custom")]; + tensor input_419_pad_0 = const()[name = tensor("input_419_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_13_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_13_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334447424)))]; + tensor input_419_cast_fp16 = conv(dilations = var_5677, groups = var_5309, pad = input_419_pad_0, pad_type = input_419_pad_type_0, strides = var_5675, weight = layers_13_fc2_loraA_weight_to_fp16, x = input_417_cast_fp16)[name = tensor("input_419_cast_fp16")]; + tensor var_5681 = const()[name = tensor("op_5681"), val = tensor([1, 1])]; + tensor var_5683 = const()[name = tensor("op_5683"), val = tensor([1, 1])]; + tensor lora_out_557_pad_type_0 = const()[name = tensor("lora_out_557_pad_type_0"), val = tensor("custom")]; + tensor lora_out_557_pad_0 = const()[name = tensor("lora_out_557_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_559_weight_0_to_fp16 = const()[name = tensor("lora_out_559_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334611328)))]; + tensor lora_out_559_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5683, groups = var_5309, pad = lora_out_557_pad_0, pad_type = lora_out_557_pad_type_0, strides = var_5681, weight = lora_out_559_weight_0_to_fp16, x = input_419_cast_fp16)[name = tensor("lora_out_559_cast_fp16")]; + tensor hidden_states_29_cast_fp16 = add(x = pretrained_out_279_cast_fp16, y = lora_out_559_cast_fp16)[name = tensor("hidden_states_29_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_5700 = const()[name = tensor("op_5700"), val = tensor(3)]; + tensor var_5707 = const()[name = tensor("op_5707"), val = tensor(1)]; + tensor var_5708 = const()[name = tensor("op_5708"), val = tensor(true)]; + tensor var_5720 = const()[name = tensor("op_5720"), val = tensor([1])]; + tensor channels_mean_85_cast_fp16 = reduce_mean(axes = var_5720, keep_dims = var_5708, x = inputs_85_cast_fp16)[name = tensor("channels_mean_85_cast_fp16")]; + tensor zero_mean_85_cast_fp16 = sub(x = inputs_85_cast_fp16, y = channels_mean_85_cast_fp16)[name = tensor("zero_mean_85_cast_fp16")]; + tensor zero_mean_sq_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = zero_mean_85_cast_fp16)[name = tensor("zero_mean_sq_85_cast_fp16")]; + tensor var_5724 = const()[name = tensor("op_5724"), val = tensor([1])]; + tensor var_5725_cast_fp16 = reduce_mean(axes = var_5724, keep_dims = var_5708, x = zero_mean_sq_85_cast_fp16)[name = tensor("op_5725_cast_fp16")]; + tensor var_5726_to_fp16 = const()[name = tensor("op_5726_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5727_cast_fp16 = add(x = var_5725_cast_fp16, y = var_5726_to_fp16)[name = tensor("op_5727_cast_fp16")]; + tensor denom_85_epsilon_0 = const()[name = tensor("denom_85_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_85_cast_fp16 = rsqrt(epsilon = denom_85_epsilon_0, x = var_5727_cast_fp16)[name = tensor("denom_85_cast_fp16")]; + tensor out_85_cast_fp16 = mul(x = zero_mean_85_cast_fp16, y = denom_85_cast_fp16)[name = tensor("out_85_cast_fp16")]; + tensor obj_197_gamma_0_to_fp16 = const()[name = tensor("obj_197_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334652352)))]; + tensor obj_197_beta_0_to_fp16 = const()[name = tensor("obj_197_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334654976)))]; + tensor obj_197_epsilon_0_to_fp16 = const()[name = tensor("obj_197_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_197_cast_fp16 = batch_norm(beta = obj_197_beta_0_to_fp16, epsilon = obj_197_epsilon_0_to_fp16, gamma = obj_197_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor("obj_197_cast_fp16")]; + tensor var_5745 = const()[name = tensor("op_5745"), val = tensor([1, 1])]; + tensor var_5747 = const()[name = tensor("op_5747"), val = tensor([1, 1])]; + tensor pretrained_out_281_pad_type_0 = const()[name = tensor("pretrained_out_281_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_281_pad_0 = const()[name = tensor("pretrained_out_281_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334657600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335476864))), name = tensor("layers_14_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_14_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335476992)))]; + tensor pretrained_out_281_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5747, groups = var_5707, pad = pretrained_out_281_pad_0, pad_type = pretrained_out_281_pad_type_0, strides = var_5745, weight = layers_14_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_197_cast_fp16)[name = tensor("pretrained_out_281_cast_fp16")]; + tensor var_5751 = const()[name = tensor("op_5751"), val = tensor([1, 1])]; + tensor var_5753 = const()[name = tensor("op_5753"), val = tensor([1, 1])]; + tensor input_421_pad_type_0 = const()[name = tensor("input_421_pad_type_0"), val = tensor("custom")]; + tensor input_421_pad_0 = const()[name = tensor("input_421_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335479616)))]; + tensor input_421_cast_fp16 = conv(dilations = var_5753, groups = var_5707, pad = input_421_pad_0, pad_type = input_421_pad_type_0, strides = var_5751, weight = layers_14_self_attn_q_proj_loraA_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor var_5757 = const()[name = tensor("op_5757"), val = tensor([1, 1])]; + tensor var_5759 = const()[name = tensor("op_5759"), val = tensor([1, 1])]; + tensor lora_out_561_pad_type_0 = const()[name = tensor("lora_out_561_pad_type_0"), val = tensor("custom")]; + tensor lora_out_561_pad_0 = const()[name = tensor("lora_out_561_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_563_weight_0_to_fp16 = const()[name = tensor("lora_out_563_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335520640)))]; + tensor lora_out_563_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5759, groups = var_5707, pad = lora_out_561_pad_0, pad_type = lora_out_561_pad_type_0, strides = var_5757, weight = lora_out_563_weight_0_to_fp16, x = input_421_cast_fp16)[name = tensor("lora_out_563_cast_fp16")]; + tensor query_57_cast_fp16 = add(x = pretrained_out_281_cast_fp16, y = lora_out_563_cast_fp16)[name = tensor("query_57_cast_fp16")]; + tensor var_5769 = const()[name = tensor("op_5769"), val = tensor([1, 1])]; + tensor var_5771 = const()[name = tensor("op_5771"), val = tensor([1, 1])]; + tensor pretrained_out_283_pad_type_0 = const()[name = tensor("pretrained_out_283_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_283_pad_0 = const()[name = tensor("pretrained_out_283_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335561664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336380928))), name = tensor("layers_14_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_283_cast_fp16 = conv(dilations = var_5771, groups = var_5707, pad = pretrained_out_283_pad_0, pad_type = pretrained_out_283_pad_type_0, strides = var_5769, weight = layers_14_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_197_cast_fp16)[name = tensor("pretrained_out_283_cast_fp16")]; + tensor var_5775 = const()[name = tensor("op_5775"), val = tensor([1, 1])]; + tensor var_5777 = const()[name = tensor("op_5777"), val = tensor([1, 1])]; + tensor input_423_pad_type_0 = const()[name = tensor("input_423_pad_type_0"), val = tensor("custom")]; + tensor input_423_pad_0 = const()[name = tensor("input_423_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336381056)))]; + tensor input_423_cast_fp16 = conv(dilations = var_5777, groups = var_5707, pad = input_423_pad_0, pad_type = input_423_pad_type_0, strides = var_5775, weight = layers_14_self_attn_k_proj_loraA_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor var_5781 = const()[name = tensor("op_5781"), val = tensor([1, 1])]; + tensor var_5783 = const()[name = tensor("op_5783"), val = tensor([1, 1])]; + tensor lora_out_565_pad_type_0 = const()[name = tensor("lora_out_565_pad_type_0"), val = tensor("custom")]; + tensor lora_out_565_pad_0 = const()[name = tensor("lora_out_565_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_567_weight_0_to_fp16 = const()[name = tensor("lora_out_567_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336422080)))]; + tensor lora_out_567_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5783, groups = var_5707, pad = lora_out_565_pad_0, pad_type = lora_out_565_pad_type_0, strides = var_5781, weight = lora_out_567_weight_0_to_fp16, x = input_423_cast_fp16)[name = tensor("lora_out_567_cast_fp16")]; + tensor current_key_29_cast_fp16 = add(x = pretrained_out_283_cast_fp16, y = lora_out_567_cast_fp16)[name = tensor("current_key_29_cast_fp16")]; + tensor var_5794 = const()[name = tensor("op_5794"), val = tensor([1, 1])]; + tensor var_5796 = const()[name = tensor("op_5796"), val = tensor([1, 1])]; + tensor pretrained_out_285_pad_type_0 = const()[name = tensor("pretrained_out_285_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_285_pad_0 = const()[name = tensor("pretrained_out_285_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336463104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337282368))), name = tensor("layers_14_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_14_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337282496)))]; + tensor pretrained_out_285_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5796, groups = var_5707, pad = pretrained_out_285_pad_0, pad_type = pretrained_out_285_pad_type_0, strides = var_5794, weight = layers_14_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_197_cast_fp16)[name = tensor("pretrained_out_285_cast_fp16")]; + tensor var_5800 = const()[name = tensor("op_5800"), val = tensor([1, 1])]; + tensor var_5802 = const()[name = tensor("op_5802"), val = tensor([1, 1])]; + tensor input_425_pad_type_0 = const()[name = tensor("input_425_pad_type_0"), val = tensor("custom")]; + tensor input_425_pad_0 = const()[name = tensor("input_425_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337285120)))]; + tensor input_425_cast_fp16 = conv(dilations = var_5802, groups = var_5707, pad = input_425_pad_0, pad_type = input_425_pad_type_0, strides = var_5800, weight = layers_14_self_attn_v_proj_loraA_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("input_425_cast_fp16")]; + tensor var_5806 = const()[name = tensor("op_5806"), val = tensor([1, 1])]; + tensor var_5808 = const()[name = tensor("op_5808"), val = tensor([1, 1])]; + tensor lora_out_569_pad_type_0 = const()[name = tensor("lora_out_569_pad_type_0"), val = tensor("custom")]; + tensor lora_out_569_pad_0 = const()[name = tensor("lora_out_569_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_571_weight_0_to_fp16 = const()[name = tensor("lora_out_571_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337326144)))]; + tensor lora_out_571_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5808, groups = var_5707, pad = lora_out_569_pad_0, pad_type = lora_out_569_pad_type_0, strides = var_5806, weight = lora_out_571_weight_0_to_fp16, x = input_425_cast_fp16)[name = tensor("lora_out_571_cast_fp16")]; + tensor current_value_29_cast_fp16 = add(x = pretrained_out_285_cast_fp16, y = lora_out_571_cast_fp16)[name = tensor("current_value_29_cast_fp16")]; + tensor var_5818_cast_fp16 = mul(x = current_key_29_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_5818_cast_fp16")]; + tensor var_5820_cast_fp16 = mul(x = var_103_cast_fp16_14, y = var_295_cast_fp16)[name = tensor("op_5820_cast_fp16")]; + tensor key_57_cast_fp16 = add(x = var_5818_cast_fp16, y = var_5820_cast_fp16)[name = tensor("key_57_cast_fp16")]; + tensor var_5822_cast_fp16 = mul(x = current_value_29_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_5822_cast_fp16")]; + tensor var_5824_cast_fp16 = mul(x = var_138_cast_fp16_14, y = var_295_cast_fp16)[name = tensor("op_5824_cast_fp16")]; + tensor value_57_cast_fp16 = add(x = var_5822_cast_fp16, y = var_5824_cast_fp16)[name = tensor("value_57_cast_fp16")]; + tensor var_5827 = const()[name = tensor("op_5827"), val = tensor([1, 20, 64, -1])]; + tensor var_5828_cast_fp16 = reshape(shape = var_5827, x = query_57_cast_fp16)[name = tensor("op_5828_cast_fp16")]; + tensor var_5829_to_fp16 = const()[name = tensor("op_5829_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5830_cast_fp16 = mul(x = var_5828_cast_fp16, y = var_5829_to_fp16)[name = tensor("op_5830_cast_fp16")]; + tensor var_5831 = const()[name = tensor("op_5831"), val = tensor([1, 20, 64, -1])]; + tensor var_5832_cast_fp16 = reshape(shape = var_5831, x = key_57_cast_fp16)[name = tensor("op_5832_cast_fp16")]; + tensor mh_w_85_transpose_x_0 = const()[name = tensor("mh_w_85_transpose_x_0"), val = tensor(true)]; + tensor mh_w_85_transpose_y_0 = const()[name = tensor("mh_w_85_transpose_y_0"), val = tensor(false)]; + tensor mh_w_85_cast_fp16 = matmul(transpose_x = mh_w_85_transpose_x_0, transpose_y = mh_w_85_transpose_y_0, x = var_5830_cast_fp16, y = var_5832_cast_fp16)[name = tensor("mh_w_85_cast_fp16")]; + tensor mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_87_cast_fp16")]; + tensor var_5840_cast_fp16 = softmax(axis = var_5700, x = mh_w_87_cast_fp16)[name = tensor("op_5840_cast_fp16")]; + tensor var_5841 = const()[name = tensor("op_5841"), val = tensor([1, 20, 64, -1])]; + tensor var_5842_cast_fp16 = reshape(shape = var_5841, x = value_57_cast_fp16)[name = tensor("op_5842_cast_fp16")]; + tensor attn_57_transpose_x_0 = const()[name = tensor("attn_57_transpose_x_0"), val = tensor(false)]; + tensor attn_57_transpose_y_0 = const()[name = tensor("attn_57_transpose_y_0"), val = tensor(true)]; + tensor attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_5842_cast_fp16, y = var_5840_cast_fp16)[name = tensor("attn_57_cast_fp16")]; + tensor var_5845 = const()[name = tensor("op_5845"), val = tensor([1, 1280, 1, -1])]; + tensor input_427_cast_fp16 = reshape(shape = var_5845, x = attn_57_cast_fp16)[name = tensor("input_427_cast_fp16")]; + tensor var_5852 = const()[name = tensor("op_5852"), val = tensor([1, 1])]; + tensor var_5854 = const()[name = tensor("op_5854"), val = tensor([1, 1])]; + tensor pretrained_out_287_pad_type_0 = const()[name = tensor("pretrained_out_287_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_287_pad_0 = const()[name = tensor("pretrained_out_287_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337367168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338186432))), name = tensor("layers_14_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_14_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338186560)))]; + tensor pretrained_out_287_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5854, groups = var_5707, pad = pretrained_out_287_pad_0, pad_type = pretrained_out_287_pad_type_0, strides = var_5852, weight = layers_14_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_427_cast_fp16)[name = tensor("pretrained_out_287_cast_fp16")]; + tensor var_5858 = const()[name = tensor("op_5858"), val = tensor([1, 1])]; + tensor var_5860 = const()[name = tensor("op_5860"), val = tensor([1, 1])]; + tensor input_429_pad_type_0 = const()[name = tensor("input_429_pad_type_0"), val = tensor("custom")]; + tensor input_429_pad_0 = const()[name = tensor("input_429_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338189184)))]; + tensor input_429_cast_fp16 = conv(dilations = var_5860, groups = var_5707, pad = input_429_pad_0, pad_type = input_429_pad_type_0, strides = var_5858, weight = layers_14_self_attn_o_proj_loraA_weight_to_fp16, x = input_427_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor var_5864 = const()[name = tensor("op_5864"), val = tensor([1, 1])]; + tensor var_5866 = const()[name = tensor("op_5866"), val = tensor([1, 1])]; + tensor lora_out_573_pad_type_0 = const()[name = tensor("lora_out_573_pad_type_0"), val = tensor("custom")]; + tensor lora_out_573_pad_0 = const()[name = tensor("lora_out_573_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_575_weight_0_to_fp16 = const()[name = tensor("lora_out_575_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338230208)))]; + tensor lora_out_575_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5866, groups = var_5707, pad = lora_out_573_pad_0, pad_type = lora_out_573_pad_type_0, strides = var_5864, weight = lora_out_575_weight_0_to_fp16, x = input_429_cast_fp16)[name = tensor("lora_out_575_cast_fp16")]; + tensor obj_203_cast_fp16 = add(x = pretrained_out_287_cast_fp16, y = lora_out_575_cast_fp16)[name = tensor("obj_203_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_203_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor var_5879 = const()[name = tensor("op_5879"), val = tensor([1])]; + tensor channels_mean_87_cast_fp16 = reduce_mean(axes = var_5879, keep_dims = var_5708, x = inputs_87_cast_fp16)[name = tensor("channels_mean_87_cast_fp16")]; + tensor zero_mean_87_cast_fp16 = sub(x = inputs_87_cast_fp16, y = channels_mean_87_cast_fp16)[name = tensor("zero_mean_87_cast_fp16")]; + tensor zero_mean_sq_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = zero_mean_87_cast_fp16)[name = tensor("zero_mean_sq_87_cast_fp16")]; + tensor var_5883 = const()[name = tensor("op_5883"), val = tensor([1])]; + tensor var_5884_cast_fp16 = reduce_mean(axes = var_5883, keep_dims = var_5708, x = zero_mean_sq_87_cast_fp16)[name = tensor("op_5884_cast_fp16")]; + tensor var_5885_to_fp16 = const()[name = tensor("op_5885_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5886_cast_fp16 = add(x = var_5884_cast_fp16, y = var_5885_to_fp16)[name = tensor("op_5886_cast_fp16")]; + tensor denom_87_epsilon_0 = const()[name = tensor("denom_87_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_87_cast_fp16 = rsqrt(epsilon = denom_87_epsilon_0, x = var_5886_cast_fp16)[name = tensor("denom_87_cast_fp16")]; + tensor out_87_cast_fp16 = mul(x = zero_mean_87_cast_fp16, y = denom_87_cast_fp16)[name = tensor("out_87_cast_fp16")]; + tensor obj_205_gamma_0_to_fp16 = const()[name = tensor("obj_205_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338271232)))]; + tensor obj_205_beta_0_to_fp16 = const()[name = tensor("obj_205_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338273856)))]; + tensor obj_205_epsilon_0_to_fp16 = const()[name = tensor("obj_205_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_205_cast_fp16 = batch_norm(beta = obj_205_beta_0_to_fp16, epsilon = obj_205_epsilon_0_to_fp16, gamma = obj_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor("obj_205_cast_fp16")]; + tensor var_5904 = const()[name = tensor("op_5904"), val = tensor([1, 1])]; + tensor var_5906 = const()[name = tensor("op_5906"), val = tensor([1, 1])]; + tensor pretrained_out_289_pad_type_0 = const()[name = tensor("pretrained_out_289_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_289_pad_0 = const()[name = tensor("pretrained_out_289_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338276480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339095744))), name = tensor("layers_14_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_14_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339095872)))]; + tensor pretrained_out_289_cast_fp16 = conv(bias = layers_14_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5906, groups = var_5707, pad = pretrained_out_289_pad_0, pad_type = pretrained_out_289_pad_type_0, strides = var_5904, weight = layers_14_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_205_cast_fp16)[name = tensor("pretrained_out_289_cast_fp16")]; + tensor var_5910 = const()[name = tensor("op_5910"), val = tensor([1, 1])]; + tensor var_5912 = const()[name = tensor("op_5912"), val = tensor([1, 1])]; + tensor input_431_pad_type_0 = const()[name = tensor("input_431_pad_type_0"), val = tensor("custom")]; + tensor input_431_pad_0 = const()[name = tensor("input_431_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339098496)))]; + tensor input_431_cast_fp16 = conv(dilations = var_5912, groups = var_5707, pad = input_431_pad_0, pad_type = input_431_pad_type_0, strides = var_5910, weight = layers_14_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_205_cast_fp16)[name = tensor("input_431_cast_fp16")]; + tensor var_5916 = const()[name = tensor("op_5916"), val = tensor([1, 1])]; + tensor var_5918 = const()[name = tensor("op_5918"), val = tensor([1, 1])]; + tensor lora_out_577_pad_type_0 = const()[name = tensor("lora_out_577_pad_type_0"), val = tensor("custom")]; + tensor lora_out_577_pad_0 = const()[name = tensor("lora_out_577_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_579_weight_0_to_fp16 = const()[name = tensor("lora_out_579_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339139520)))]; + tensor lora_out_579_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5918, groups = var_5707, pad = lora_out_577_pad_0, pad_type = lora_out_577_pad_type_0, strides = var_5916, weight = lora_out_579_weight_0_to_fp16, x = input_431_cast_fp16)[name = tensor("lora_out_579_cast_fp16")]; + tensor query_59_cast_fp16 = add(x = pretrained_out_289_cast_fp16, y = lora_out_579_cast_fp16)[name = tensor("query_59_cast_fp16")]; + tensor var_5928 = const()[name = tensor("op_5928"), val = tensor([1, 1])]; + tensor var_5930 = const()[name = tensor("op_5930"), val = tensor([1, 1])]; + tensor pretrained_out_291_pad_type_0 = const()[name = tensor("pretrained_out_291_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_291_pad_0 = const()[name = tensor("pretrained_out_291_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339180544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339999808))), name = tensor("layers_14_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_291_cast_fp16 = conv(dilations = var_5930, groups = var_5707, pad = pretrained_out_291_pad_0, pad_type = pretrained_out_291_pad_type_0, strides = var_5928, weight = layers_14_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_291_cast_fp16")]; + tensor var_5934 = const()[name = tensor("op_5934"), val = tensor([1, 1])]; + tensor var_5936 = const()[name = tensor("op_5936"), val = tensor([1, 1])]; + tensor input_433_pad_type_0 = const()[name = tensor("input_433_pad_type_0"), val = tensor("custom")]; + tensor input_433_pad_0 = const()[name = tensor("input_433_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339999936)))]; + tensor input_433_cast_fp16 = conv(dilations = var_5936, groups = var_5707, pad = input_433_pad_0, pad_type = input_433_pad_type_0, strides = var_5934, weight = layers_14_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_433_cast_fp16")]; + tensor var_5940 = const()[name = tensor("op_5940"), val = tensor([1, 1])]; + tensor var_5942 = const()[name = tensor("op_5942"), val = tensor([1, 1])]; + tensor lora_out_581_pad_type_0 = const()[name = tensor("lora_out_581_pad_type_0"), val = tensor("custom")]; + tensor lora_out_581_pad_0 = const()[name = tensor("lora_out_581_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_583_weight_0_to_fp16 = const()[name = tensor("lora_out_583_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340040960)))]; + tensor lora_out_583_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5942, groups = var_5707, pad = lora_out_581_pad_0, pad_type = lora_out_581_pad_type_0, strides = var_5940, weight = lora_out_583_weight_0_to_fp16, x = input_433_cast_fp16)[name = tensor("lora_out_583_cast_fp16")]; + tensor key_59_cast_fp16 = add(x = pretrained_out_291_cast_fp16, y = lora_out_583_cast_fp16)[name = tensor("key_59_cast_fp16")]; + tensor var_5953 = const()[name = tensor("op_5953"), val = tensor([1, 1])]; + tensor var_5955 = const()[name = tensor("op_5955"), val = tensor([1, 1])]; + tensor pretrained_out_293_pad_type_0 = const()[name = tensor("pretrained_out_293_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_293_pad_0 = const()[name = tensor("pretrained_out_293_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340081984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340901248))), name = tensor("layers_14_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_14_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340901376)))]; + tensor pretrained_out_293_cast_fp16 = conv(bias = layers_14_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5955, groups = var_5707, pad = pretrained_out_293_pad_0, pad_type = pretrained_out_293_pad_type_0, strides = var_5953, weight = layers_14_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_293_cast_fp16")]; + tensor var_5959 = const()[name = tensor("op_5959"), val = tensor([1, 1])]; + tensor var_5961 = const()[name = tensor("op_5961"), val = tensor([1, 1])]; + tensor input_435_pad_type_0 = const()[name = tensor("input_435_pad_type_0"), val = tensor("custom")]; + tensor input_435_pad_0 = const()[name = tensor("input_435_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340904000)))]; + tensor input_435_cast_fp16 = conv(dilations = var_5961, groups = var_5707, pad = input_435_pad_0, pad_type = input_435_pad_type_0, strides = var_5959, weight = layers_14_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_435_cast_fp16")]; + tensor var_5965 = const()[name = tensor("op_5965"), val = tensor([1, 1])]; + tensor var_5967 = const()[name = tensor("op_5967"), val = tensor([1, 1])]; + tensor lora_out_585_pad_type_0 = const()[name = tensor("lora_out_585_pad_type_0"), val = tensor("custom")]; + tensor lora_out_585_pad_0 = const()[name = tensor("lora_out_585_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_587_weight_0_to_fp16 = const()[name = tensor("lora_out_587_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340945024)))]; + tensor lora_out_587_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5967, groups = var_5707, pad = lora_out_585_pad_0, pad_type = lora_out_585_pad_type_0, strides = var_5965, weight = lora_out_587_weight_0_to_fp16, x = input_435_cast_fp16)[name = tensor("lora_out_587_cast_fp16")]; + tensor value_59_cast_fp16 = add(x = pretrained_out_293_cast_fp16, y = lora_out_587_cast_fp16)[name = tensor("value_59_cast_fp16")]; + tensor var_5974 = const()[name = tensor("op_5974"), val = tensor([1, 20, 64, -1])]; + tensor var_5975_cast_fp16 = reshape(shape = var_5974, x = query_59_cast_fp16)[name = tensor("op_5975_cast_fp16")]; + tensor var_5976_to_fp16 = const()[name = tensor("op_5976_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5977_cast_fp16 = mul(x = var_5975_cast_fp16, y = var_5976_to_fp16)[name = tensor("op_5977_cast_fp16")]; + tensor var_5978 = const()[name = tensor("op_5978"), val = tensor([1, 20, 64, -1])]; + tensor var_5979_cast_fp16 = reshape(shape = var_5978, x = key_59_cast_fp16)[name = tensor("op_5979_cast_fp16")]; + tensor mh_w_89_transpose_x_0 = const()[name = tensor("mh_w_89_transpose_x_0"), val = tensor(true)]; + tensor mh_w_89_transpose_y_0 = const()[name = tensor("mh_w_89_transpose_y_0"), val = tensor(false)]; + tensor mh_w_89_cast_fp16 = matmul(transpose_x = mh_w_89_transpose_x_0, transpose_y = mh_w_89_transpose_y_0, x = var_5977_cast_fp16, y = var_5979_cast_fp16)[name = tensor("mh_w_89_cast_fp16")]; + tensor obj_209_cast_fp16 = softmax(axis = var_5700, x = mh_w_89_cast_fp16)[name = tensor("obj_209_cast_fp16")]; + tensor var_5983 = const()[name = tensor("op_5983"), val = tensor([1, 20, 64, -1])]; + tensor var_5984_cast_fp16 = reshape(shape = var_5983, x = value_59_cast_fp16)[name = tensor("op_5984_cast_fp16")]; + tensor attn_59_transpose_x_0 = const()[name = tensor("attn_59_transpose_x_0"), val = tensor(false)]; + tensor attn_59_transpose_y_0 = const()[name = tensor("attn_59_transpose_y_0"), val = tensor(true)]; + tensor attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_5984_cast_fp16, y = obj_209_cast_fp16)[name = tensor("attn_59_cast_fp16")]; + tensor var_5987 = const()[name = tensor("op_5987"), val = tensor([1, 1280, 1, -1])]; + tensor input_437_cast_fp16 = reshape(shape = var_5987, x = attn_59_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor var_5994 = const()[name = tensor("op_5994"), val = tensor([1, 1])]; + tensor var_5996 = const()[name = tensor("op_5996"), val = tensor([1, 1])]; + tensor pretrained_out_295_pad_type_0 = const()[name = tensor("pretrained_out_295_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_295_pad_0 = const()[name = tensor("pretrained_out_295_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340986048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341805312))), name = tensor("layers_14_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_14_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341805440)))]; + tensor pretrained_out_295_cast_fp16 = conv(bias = layers_14_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5996, groups = var_5707, pad = pretrained_out_295_pad_0, pad_type = pretrained_out_295_pad_type_0, strides = var_5994, weight = layers_14_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_437_cast_fp16)[name = tensor("pretrained_out_295_cast_fp16")]; + tensor var_6000 = const()[name = tensor("op_6000"), val = tensor([1, 1])]; + tensor var_6002 = const()[name = tensor("op_6002"), val = tensor([1, 1])]; + tensor input_439_pad_type_0 = const()[name = tensor("input_439_pad_type_0"), val = tensor("custom")]; + tensor input_439_pad_0 = const()[name = tensor("input_439_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341808064)))]; + tensor input_439_cast_fp16 = conv(dilations = var_6002, groups = var_5707, pad = input_439_pad_0, pad_type = input_439_pad_type_0, strides = var_6000, weight = layers_14_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("input_439_cast_fp16")]; + tensor var_6006 = const()[name = tensor("op_6006"), val = tensor([1, 1])]; + tensor var_6008 = const()[name = tensor("op_6008"), val = tensor([1, 1])]; + tensor lora_out_589_pad_type_0 = const()[name = tensor("lora_out_589_pad_type_0"), val = tensor("custom")]; + tensor lora_out_589_pad_0 = const()[name = tensor("lora_out_589_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_591_weight_0_to_fp16 = const()[name = tensor("lora_out_591_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341849088)))]; + tensor lora_out_591_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6008, groups = var_5707, pad = lora_out_589_pad_0, pad_type = lora_out_589_pad_type_0, strides = var_6006, weight = lora_out_591_weight_0_to_fp16, x = input_439_cast_fp16)[name = tensor("lora_out_591_cast_fp16")]; + tensor obj_207_cast_fp16 = add(x = pretrained_out_295_cast_fp16, y = lora_out_591_cast_fp16)[name = tensor("obj_207_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = obj_207_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor var_6017 = const()[name = tensor("op_6017"), val = tensor([1])]; + tensor channels_mean_89_cast_fp16 = reduce_mean(axes = var_6017, keep_dims = var_5708, x = inputs_89_cast_fp16)[name = tensor("channels_mean_89_cast_fp16")]; + tensor zero_mean_89_cast_fp16 = sub(x = inputs_89_cast_fp16, y = channels_mean_89_cast_fp16)[name = tensor("zero_mean_89_cast_fp16")]; + tensor zero_mean_sq_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = zero_mean_89_cast_fp16)[name = tensor("zero_mean_sq_89_cast_fp16")]; + tensor var_6021 = const()[name = tensor("op_6021"), val = tensor([1])]; + tensor var_6022_cast_fp16 = reduce_mean(axes = var_6021, keep_dims = var_5708, x = zero_mean_sq_89_cast_fp16)[name = tensor("op_6022_cast_fp16")]; + tensor var_6023_to_fp16 = const()[name = tensor("op_6023_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6024_cast_fp16 = add(x = var_6022_cast_fp16, y = var_6023_to_fp16)[name = tensor("op_6024_cast_fp16")]; + tensor denom_89_epsilon_0 = const()[name = tensor("denom_89_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_89_cast_fp16 = rsqrt(epsilon = denom_89_epsilon_0, x = var_6024_cast_fp16)[name = tensor("denom_89_cast_fp16")]; + tensor out_89_cast_fp16 = mul(x = zero_mean_89_cast_fp16, y = denom_89_cast_fp16)[name = tensor("out_89_cast_fp16")]; + tensor input_441_gamma_0_to_fp16 = const()[name = tensor("input_441_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341890112)))]; + tensor input_441_beta_0_to_fp16 = const()[name = tensor("input_441_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341892736)))]; + tensor input_441_epsilon_0_to_fp16 = const()[name = tensor("input_441_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_441_cast_fp16 = batch_norm(beta = input_441_beta_0_to_fp16, epsilon = input_441_epsilon_0_to_fp16, gamma = input_441_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor("input_441_cast_fp16")]; + tensor var_6038 = const()[name = tensor("op_6038"), val = tensor([1, 1])]; + tensor var_6040 = const()[name = tensor("op_6040"), val = tensor([1, 1])]; + tensor pretrained_out_297_pad_type_0 = const()[name = tensor("pretrained_out_297_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_297_pad_0 = const()[name = tensor("pretrained_out_297_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(341895360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345172224))), name = tensor("layers_14_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_14_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345172352)))]; + tensor pretrained_out_297_cast_fp16 = conv(bias = layers_14_fc1_pretrained_bias_to_fp16, dilations = var_6040, groups = var_5707, pad = pretrained_out_297_pad_0, pad_type = pretrained_out_297_pad_type_0, strides = var_6038, weight = layers_14_fc1_pretrained_weight_to_fp16_palettized, x = input_441_cast_fp16)[name = tensor("pretrained_out_297_cast_fp16")]; + tensor var_6044 = const()[name = tensor("op_6044"), val = tensor([1, 1])]; + tensor var_6046 = const()[name = tensor("op_6046"), val = tensor([1, 1])]; + tensor input_443_pad_type_0 = const()[name = tensor("input_443_pad_type_0"), val = tensor("custom")]; + tensor input_443_pad_0 = const()[name = tensor("input_443_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_14_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345182656)))]; + tensor input_443_cast_fp16 = conv(dilations = var_6046, groups = var_5707, pad = input_443_pad_0, pad_type = input_443_pad_type_0, strides = var_6044, weight = layers_14_fc1_loraA_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("input_443_cast_fp16")]; + tensor var_6050 = const()[name = tensor("op_6050"), val = tensor([1, 1])]; + tensor var_6052 = const()[name = tensor("op_6052"), val = tensor([1, 1])]; + tensor lora_out_593_pad_type_0 = const()[name = tensor("lora_out_593_pad_type_0"), val = tensor("custom")]; + tensor lora_out_593_pad_0 = const()[name = tensor("lora_out_593_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_595_weight_0_to_fp16 = const()[name = tensor("lora_out_595_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345223680)))]; + tensor lora_out_595_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_6052, groups = var_5707, pad = lora_out_593_pad_0, pad_type = lora_out_593_pad_type_0, strides = var_6050, weight = lora_out_595_weight_0_to_fp16, x = input_443_cast_fp16)[name = tensor("lora_out_595_cast_fp16")]; + tensor input_445_cast_fp16 = add(x = pretrained_out_297_cast_fp16, y = lora_out_595_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor input_447_mode_0 = const()[name = tensor("input_447_mode_0"), val = tensor("EXACT")]; + tensor input_447_cast_fp16 = gelu(mode = input_447_mode_0, x = input_445_cast_fp16)[name = tensor("input_447_cast_fp16")]; + tensor var_6064 = const()[name = tensor("op_6064"), val = tensor([1, 1])]; + tensor var_6066 = const()[name = tensor("op_6066"), val = tensor([1, 1])]; + tensor pretrained_out_299_pad_type_0 = const()[name = tensor("pretrained_out_299_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_299_pad_0 = const()[name = tensor("pretrained_out_299_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345387584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348664448))), name = tensor("layers_14_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_14_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_14_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348664576)))]; + tensor pretrained_out_299_cast_fp16 = conv(bias = layers_14_fc2_pretrained_bias_to_fp16, dilations = var_6066, groups = var_5707, pad = pretrained_out_299_pad_0, pad_type = pretrained_out_299_pad_type_0, strides = var_6064, weight = layers_14_fc2_pretrained_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = tensor("pretrained_out_299_cast_fp16")]; + tensor var_6070 = const()[name = tensor("op_6070"), val = tensor([1, 1])]; + tensor var_6072 = const()[name = tensor("op_6072"), val = tensor([1, 1])]; + tensor input_449_pad_type_0 = const()[name = tensor("input_449_pad_type_0"), val = tensor("custom")]; + tensor input_449_pad_0 = const()[name = tensor("input_449_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_14_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_14_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348667200)))]; + tensor input_449_cast_fp16 = conv(dilations = var_6072, groups = var_5707, pad = input_449_pad_0, pad_type = input_449_pad_type_0, strides = var_6070, weight = layers_14_fc2_loraA_weight_to_fp16, x = input_447_cast_fp16)[name = tensor("input_449_cast_fp16")]; + tensor var_6076 = const()[name = tensor("op_6076"), val = tensor([1, 1])]; + tensor var_6078 = const()[name = tensor("op_6078"), val = tensor([1, 1])]; + tensor lora_out_597_pad_type_0 = const()[name = tensor("lora_out_597_pad_type_0"), val = tensor("custom")]; + tensor lora_out_597_pad_0 = const()[name = tensor("lora_out_597_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_599_weight_0_to_fp16 = const()[name = tensor("lora_out_599_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348831104)))]; + tensor lora_out_599_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6078, groups = var_5707, pad = lora_out_597_pad_0, pad_type = lora_out_597_pad_type_0, strides = var_6076, weight = lora_out_599_weight_0_to_fp16, x = input_449_cast_fp16)[name = tensor("lora_out_599_cast_fp16")]; + tensor hidden_states_31_cast_fp16 = add(x = pretrained_out_299_cast_fp16, y = lora_out_599_cast_fp16)[name = tensor("hidden_states_31_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor var_6094 = const()[name = tensor("op_6094"), val = tensor(3)]; + tensor var_6101 = const()[name = tensor("op_6101"), val = tensor(1)]; + tensor var_6102 = const()[name = tensor("op_6102"), val = tensor(true)]; + tensor var_6114 = const()[name = tensor("op_6114"), val = tensor([1])]; + tensor channels_mean_91_cast_fp16 = reduce_mean(axes = var_6114, keep_dims = var_6102, x = inputs_91_cast_fp16)[name = tensor("channels_mean_91_cast_fp16")]; + tensor zero_mean_91_cast_fp16 = sub(x = inputs_91_cast_fp16, y = channels_mean_91_cast_fp16)[name = tensor("zero_mean_91_cast_fp16")]; + tensor zero_mean_sq_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = zero_mean_91_cast_fp16)[name = tensor("zero_mean_sq_91_cast_fp16")]; + tensor var_6118 = const()[name = tensor("op_6118"), val = tensor([1])]; + tensor var_6119_cast_fp16 = reduce_mean(axes = var_6118, keep_dims = var_6102, x = zero_mean_sq_91_cast_fp16)[name = tensor("op_6119_cast_fp16")]; + tensor var_6120_to_fp16 = const()[name = tensor("op_6120_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6121_cast_fp16 = add(x = var_6119_cast_fp16, y = var_6120_to_fp16)[name = tensor("op_6121_cast_fp16")]; + tensor denom_91_epsilon_0 = const()[name = tensor("denom_91_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_91_cast_fp16 = rsqrt(epsilon = denom_91_epsilon_0, x = var_6121_cast_fp16)[name = tensor("denom_91_cast_fp16")]; + tensor out_91_cast_fp16 = mul(x = zero_mean_91_cast_fp16, y = denom_91_cast_fp16)[name = tensor("out_91_cast_fp16")]; + tensor obj_211_gamma_0_to_fp16 = const()[name = tensor("obj_211_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348872128)))]; + tensor obj_211_beta_0_to_fp16 = const()[name = tensor("obj_211_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348874752)))]; + tensor obj_211_epsilon_0_to_fp16 = const()[name = tensor("obj_211_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_211_cast_fp16 = batch_norm(beta = obj_211_beta_0_to_fp16, epsilon = obj_211_epsilon_0_to_fp16, gamma = obj_211_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor("obj_211_cast_fp16")]; + tensor var_6139 = const()[name = tensor("op_6139"), val = tensor([1, 1])]; + tensor var_6141 = const()[name = tensor("op_6141"), val = tensor([1, 1])]; + tensor pretrained_out_301_pad_type_0 = const()[name = tensor("pretrained_out_301_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_301_pad_0 = const()[name = tensor("pretrained_out_301_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348877376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349696640))), name = tensor("layers_15_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_15_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349696768)))]; + tensor pretrained_out_301_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6141, groups = var_6101, pad = pretrained_out_301_pad_0, pad_type = pretrained_out_301_pad_type_0, strides = var_6139, weight = layers_15_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_211_cast_fp16)[name = tensor("pretrained_out_301_cast_fp16")]; + tensor var_6145 = const()[name = tensor("op_6145"), val = tensor([1, 1])]; + tensor var_6147 = const()[name = tensor("op_6147"), val = tensor([1, 1])]; + tensor input_451_pad_type_0 = const()[name = tensor("input_451_pad_type_0"), val = tensor("custom")]; + tensor input_451_pad_0 = const()[name = tensor("input_451_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349699392)))]; + tensor input_451_cast_fp16 = conv(dilations = var_6147, groups = var_6101, pad = input_451_pad_0, pad_type = input_451_pad_type_0, strides = var_6145, weight = layers_15_self_attn_q_proj_loraA_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("input_451_cast_fp16")]; + tensor var_6151 = const()[name = tensor("op_6151"), val = tensor([1, 1])]; + tensor var_6153 = const()[name = tensor("op_6153"), val = tensor([1, 1])]; + tensor lora_out_601_pad_type_0 = const()[name = tensor("lora_out_601_pad_type_0"), val = tensor("custom")]; + tensor lora_out_601_pad_0 = const()[name = tensor("lora_out_601_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_603_weight_0_to_fp16 = const()[name = tensor("lora_out_603_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349740416)))]; + tensor lora_out_603_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6153, groups = var_6101, pad = lora_out_601_pad_0, pad_type = lora_out_601_pad_type_0, strides = var_6151, weight = lora_out_603_weight_0_to_fp16, x = input_451_cast_fp16)[name = tensor("lora_out_603_cast_fp16")]; + tensor query_61_cast_fp16 = add(x = pretrained_out_301_cast_fp16, y = lora_out_603_cast_fp16)[name = tensor("query_61_cast_fp16")]; + tensor var_6163 = const()[name = tensor("op_6163"), val = tensor([1, 1])]; + tensor var_6165 = const()[name = tensor("op_6165"), val = tensor([1, 1])]; + tensor pretrained_out_303_pad_type_0 = const()[name = tensor("pretrained_out_303_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_303_pad_0 = const()[name = tensor("pretrained_out_303_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(349781440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350600704))), name = tensor("layers_15_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_303_cast_fp16 = conv(dilations = var_6165, groups = var_6101, pad = pretrained_out_303_pad_0, pad_type = pretrained_out_303_pad_type_0, strides = var_6163, weight = layers_15_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_211_cast_fp16)[name = tensor("pretrained_out_303_cast_fp16")]; + tensor var_6169 = const()[name = tensor("op_6169"), val = tensor([1, 1])]; + tensor var_6171 = const()[name = tensor("op_6171"), val = tensor([1, 1])]; + tensor input_453_pad_type_0 = const()[name = tensor("input_453_pad_type_0"), val = tensor("custom")]; + tensor input_453_pad_0 = const()[name = tensor("input_453_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350600832)))]; + tensor input_453_cast_fp16 = conv(dilations = var_6171, groups = var_6101, pad = input_453_pad_0, pad_type = input_453_pad_type_0, strides = var_6169, weight = layers_15_self_attn_k_proj_loraA_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor var_6175 = const()[name = tensor("op_6175"), val = tensor([1, 1])]; + tensor var_6177 = const()[name = tensor("op_6177"), val = tensor([1, 1])]; + tensor lora_out_605_pad_type_0 = const()[name = tensor("lora_out_605_pad_type_0"), val = tensor("custom")]; + tensor lora_out_605_pad_0 = const()[name = tensor("lora_out_605_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_607_weight_0_to_fp16 = const()[name = tensor("lora_out_607_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350641856)))]; + tensor lora_out_607_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6177, groups = var_6101, pad = lora_out_605_pad_0, pad_type = lora_out_605_pad_type_0, strides = var_6175, weight = lora_out_607_weight_0_to_fp16, x = input_453_cast_fp16)[name = tensor("lora_out_607_cast_fp16")]; + tensor current_key_31_cast_fp16 = add(x = pretrained_out_303_cast_fp16, y = lora_out_607_cast_fp16)[name = tensor("current_key_31_cast_fp16")]; + tensor var_6188 = const()[name = tensor("op_6188"), val = tensor([1, 1])]; + tensor var_6190 = const()[name = tensor("op_6190"), val = tensor([1, 1])]; + tensor pretrained_out_305_pad_type_0 = const()[name = tensor("pretrained_out_305_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_305_pad_0 = const()[name = tensor("pretrained_out_305_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350682880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351502144))), name = tensor("layers_15_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_15_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351502272)))]; + tensor pretrained_out_305_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6190, groups = var_6101, pad = pretrained_out_305_pad_0, pad_type = pretrained_out_305_pad_type_0, strides = var_6188, weight = layers_15_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_211_cast_fp16)[name = tensor("pretrained_out_305_cast_fp16")]; + tensor var_6194 = const()[name = tensor("op_6194"), val = tensor([1, 1])]; + tensor var_6196 = const()[name = tensor("op_6196"), val = tensor([1, 1])]; + tensor input_455_pad_type_0 = const()[name = tensor("input_455_pad_type_0"), val = tensor("custom")]; + tensor input_455_pad_0 = const()[name = tensor("input_455_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351504896)))]; + tensor input_455_cast_fp16 = conv(dilations = var_6196, groups = var_6101, pad = input_455_pad_0, pad_type = input_455_pad_type_0, strides = var_6194, weight = layers_15_self_attn_v_proj_loraA_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("input_455_cast_fp16")]; + tensor var_6200 = const()[name = tensor("op_6200"), val = tensor([1, 1])]; + tensor var_6202 = const()[name = tensor("op_6202"), val = tensor([1, 1])]; + tensor lora_out_609_pad_type_0 = const()[name = tensor("lora_out_609_pad_type_0"), val = tensor("custom")]; + tensor lora_out_609_pad_0 = const()[name = tensor("lora_out_609_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_611_weight_0_to_fp16 = const()[name = tensor("lora_out_611_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351545920)))]; + tensor lora_out_611_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6202, groups = var_6101, pad = lora_out_609_pad_0, pad_type = lora_out_609_pad_type_0, strides = var_6200, weight = lora_out_611_weight_0_to_fp16, x = input_455_cast_fp16)[name = tensor("lora_out_611_cast_fp16")]; + tensor current_value_31_cast_fp16 = add(x = pretrained_out_305_cast_fp16, y = lora_out_611_cast_fp16)[name = tensor("current_value_31_cast_fp16")]; + tensor var_6212_cast_fp16 = mul(x = current_key_31_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_6212_cast_fp16")]; + tensor var_6214_cast_fp16 = mul(x = var_103_cast_fp16_15, y = var_295_cast_fp16)[name = tensor("op_6214_cast_fp16")]; + tensor key_61_cast_fp16 = add(x = var_6212_cast_fp16, y = var_6214_cast_fp16)[name = tensor("key_61_cast_fp16")]; + tensor var_6216_cast_fp16 = mul(x = current_value_31_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_6216_cast_fp16")]; + tensor var_6218_cast_fp16 = mul(x = var_138_cast_fp16_15, y = var_295_cast_fp16)[name = tensor("op_6218_cast_fp16")]; + tensor value_61_cast_fp16 = add(x = var_6216_cast_fp16, y = var_6218_cast_fp16)[name = tensor("value_61_cast_fp16")]; + tensor var_6221 = const()[name = tensor("op_6221"), val = tensor([1, 20, 64, -1])]; + tensor var_6222_cast_fp16 = reshape(shape = var_6221, x = query_61_cast_fp16)[name = tensor("op_6222_cast_fp16")]; + tensor var_6223_to_fp16 = const()[name = tensor("op_6223_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6224_cast_fp16 = mul(x = var_6222_cast_fp16, y = var_6223_to_fp16)[name = tensor("op_6224_cast_fp16")]; + tensor var_6225 = const()[name = tensor("op_6225"), val = tensor([1, 20, 64, -1])]; + tensor var_6226_cast_fp16 = reshape(shape = var_6225, x = key_61_cast_fp16)[name = tensor("op_6226_cast_fp16")]; + tensor mh_w_91_transpose_x_0 = const()[name = tensor("mh_w_91_transpose_x_0"), val = tensor(true)]; + tensor mh_w_91_transpose_y_0 = const()[name = tensor("mh_w_91_transpose_y_0"), val = tensor(false)]; + tensor mh_w_91_cast_fp16 = matmul(transpose_x = mh_w_91_transpose_x_0, transpose_y = mh_w_91_transpose_y_0, x = var_6224_cast_fp16, y = var_6226_cast_fp16)[name = tensor("mh_w_91_cast_fp16")]; + tensor mh_w_93_cast_fp16 = add(x = mh_w_91_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_93_cast_fp16")]; + tensor var_6234_cast_fp16 = softmax(axis = var_6094, x = mh_w_93_cast_fp16)[name = tensor("op_6234_cast_fp16")]; + tensor var_6235 = const()[name = tensor("op_6235"), val = tensor([1, 20, 64, -1])]; + tensor var_6236_cast_fp16 = reshape(shape = var_6235, x = value_61_cast_fp16)[name = tensor("op_6236_cast_fp16")]; + tensor attn_61_transpose_x_0 = const()[name = tensor("attn_61_transpose_x_0"), val = tensor(false)]; + tensor attn_61_transpose_y_0 = const()[name = tensor("attn_61_transpose_y_0"), val = tensor(true)]; + tensor attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_6236_cast_fp16, y = var_6234_cast_fp16)[name = tensor("attn_61_cast_fp16")]; + tensor var_6239 = const()[name = tensor("op_6239"), val = tensor([1, 1280, 1, -1])]; + tensor input_457_cast_fp16 = reshape(shape = var_6239, x = attn_61_cast_fp16)[name = tensor("input_457_cast_fp16")]; + tensor var_6246 = const()[name = tensor("op_6246"), val = tensor([1, 1])]; + tensor var_6248 = const()[name = tensor("op_6248"), val = tensor([1, 1])]; + tensor pretrained_out_307_pad_type_0 = const()[name = tensor("pretrained_out_307_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_307_pad_0 = const()[name = tensor("pretrained_out_307_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351586944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352406208))), name = tensor("layers_15_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_15_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352406336)))]; + tensor pretrained_out_307_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_6248, groups = var_6101, pad = pretrained_out_307_pad_0, pad_type = pretrained_out_307_pad_type_0, strides = var_6246, weight = layers_15_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_457_cast_fp16)[name = tensor("pretrained_out_307_cast_fp16")]; + tensor var_6252 = const()[name = tensor("op_6252"), val = tensor([1, 1])]; + tensor var_6254 = const()[name = tensor("op_6254"), val = tensor([1, 1])]; + tensor input_459_pad_type_0 = const()[name = tensor("input_459_pad_type_0"), val = tensor("custom")]; + tensor input_459_pad_0 = const()[name = tensor("input_459_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352408960)))]; + tensor input_459_cast_fp16 = conv(dilations = var_6254, groups = var_6101, pad = input_459_pad_0, pad_type = input_459_pad_type_0, strides = var_6252, weight = layers_15_self_attn_o_proj_loraA_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("input_459_cast_fp16")]; + tensor var_6258 = const()[name = tensor("op_6258"), val = tensor([1, 1])]; + tensor var_6260 = const()[name = tensor("op_6260"), val = tensor([1, 1])]; + tensor lora_out_613_pad_type_0 = const()[name = tensor("lora_out_613_pad_type_0"), val = tensor("custom")]; + tensor lora_out_613_pad_0 = const()[name = tensor("lora_out_613_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_615_weight_0_to_fp16 = const()[name = tensor("lora_out_615_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352449984)))]; + tensor lora_out_615_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6260, groups = var_6101, pad = lora_out_613_pad_0, pad_type = lora_out_613_pad_type_0, strides = var_6258, weight = lora_out_615_weight_0_to_fp16, x = input_459_cast_fp16)[name = tensor("lora_out_615_cast_fp16")]; + tensor obj_217_cast_fp16 = add(x = pretrained_out_307_cast_fp16, y = lora_out_615_cast_fp16)[name = tensor("obj_217_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = obj_217_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor var_6273 = const()[name = tensor("op_6273"), val = tensor([1])]; + tensor channels_mean_93_cast_fp16 = reduce_mean(axes = var_6273, keep_dims = var_6102, x = inputs_93_cast_fp16)[name = tensor("channels_mean_93_cast_fp16")]; + tensor zero_mean_93_cast_fp16 = sub(x = inputs_93_cast_fp16, y = channels_mean_93_cast_fp16)[name = tensor("zero_mean_93_cast_fp16")]; + tensor zero_mean_sq_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = zero_mean_93_cast_fp16)[name = tensor("zero_mean_sq_93_cast_fp16")]; + tensor var_6277 = const()[name = tensor("op_6277"), val = tensor([1])]; + tensor var_6278_cast_fp16 = reduce_mean(axes = var_6277, keep_dims = var_6102, x = zero_mean_sq_93_cast_fp16)[name = tensor("op_6278_cast_fp16")]; + tensor var_6279_to_fp16 = const()[name = tensor("op_6279_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6280_cast_fp16 = add(x = var_6278_cast_fp16, y = var_6279_to_fp16)[name = tensor("op_6280_cast_fp16")]; + tensor denom_93_epsilon_0 = const()[name = tensor("denom_93_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_93_cast_fp16 = rsqrt(epsilon = denom_93_epsilon_0, x = var_6280_cast_fp16)[name = tensor("denom_93_cast_fp16")]; + tensor out_93_cast_fp16 = mul(x = zero_mean_93_cast_fp16, y = denom_93_cast_fp16)[name = tensor("out_93_cast_fp16")]; + tensor obj_219_gamma_0_to_fp16 = const()[name = tensor("obj_219_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352491008)))]; + tensor obj_219_beta_0_to_fp16 = const()[name = tensor("obj_219_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352493632)))]; + tensor obj_219_epsilon_0_to_fp16 = const()[name = tensor("obj_219_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_219_cast_fp16 = batch_norm(beta = obj_219_beta_0_to_fp16, epsilon = obj_219_epsilon_0_to_fp16, gamma = obj_219_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor("obj_219_cast_fp16")]; + tensor var_6298 = const()[name = tensor("op_6298"), val = tensor([1, 1])]; + tensor var_6300 = const()[name = tensor("op_6300"), val = tensor([1, 1])]; + tensor pretrained_out_309_pad_type_0 = const()[name = tensor("pretrained_out_309_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_309_pad_0 = const()[name = tensor("pretrained_out_309_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352496256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353315520))), name = tensor("layers_15_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_15_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353315648)))]; + tensor pretrained_out_309_cast_fp16 = conv(bias = layers_15_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6300, groups = var_6101, pad = pretrained_out_309_pad_0, pad_type = pretrained_out_309_pad_type_0, strides = var_6298, weight = layers_15_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_219_cast_fp16)[name = tensor("pretrained_out_309_cast_fp16")]; + tensor var_6304 = const()[name = tensor("op_6304"), val = tensor([1, 1])]; + tensor var_6306 = const()[name = tensor("op_6306"), val = tensor([1, 1])]; + tensor input_461_pad_type_0 = const()[name = tensor("input_461_pad_type_0"), val = tensor("custom")]; + tensor input_461_pad_0 = const()[name = tensor("input_461_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353318272)))]; + tensor input_461_cast_fp16 = conv(dilations = var_6306, groups = var_6101, pad = input_461_pad_0, pad_type = input_461_pad_type_0, strides = var_6304, weight = layers_15_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_219_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor var_6310 = const()[name = tensor("op_6310"), val = tensor([1, 1])]; + tensor var_6312 = const()[name = tensor("op_6312"), val = tensor([1, 1])]; + tensor lora_out_617_pad_type_0 = const()[name = tensor("lora_out_617_pad_type_0"), val = tensor("custom")]; + tensor lora_out_617_pad_0 = const()[name = tensor("lora_out_617_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_619_weight_0_to_fp16 = const()[name = tensor("lora_out_619_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353359296)))]; + tensor lora_out_619_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6312, groups = var_6101, pad = lora_out_617_pad_0, pad_type = lora_out_617_pad_type_0, strides = var_6310, weight = lora_out_619_weight_0_to_fp16, x = input_461_cast_fp16)[name = tensor("lora_out_619_cast_fp16")]; + tensor query_63_cast_fp16 = add(x = pretrained_out_309_cast_fp16, y = lora_out_619_cast_fp16)[name = tensor("query_63_cast_fp16")]; + tensor var_6322 = const()[name = tensor("op_6322"), val = tensor([1, 1])]; + tensor var_6324 = const()[name = tensor("op_6324"), val = tensor([1, 1])]; + tensor pretrained_out_311_pad_type_0 = const()[name = tensor("pretrained_out_311_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_311_pad_0 = const()[name = tensor("pretrained_out_311_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353400320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354219584))), name = tensor("layers_15_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_311_cast_fp16 = conv(dilations = var_6324, groups = var_6101, pad = pretrained_out_311_pad_0, pad_type = pretrained_out_311_pad_type_0, strides = var_6322, weight = layers_15_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_311_cast_fp16")]; + tensor var_6328 = const()[name = tensor("op_6328"), val = tensor([1, 1])]; + tensor var_6330 = const()[name = tensor("op_6330"), val = tensor([1, 1])]; + tensor input_463_pad_type_0 = const()[name = tensor("input_463_pad_type_0"), val = tensor("custom")]; + tensor input_463_pad_0 = const()[name = tensor("input_463_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354219712)))]; + tensor input_463_cast_fp16 = conv(dilations = var_6330, groups = var_6101, pad = input_463_pad_0, pad_type = input_463_pad_type_0, strides = var_6328, weight = layers_15_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_463_cast_fp16")]; + tensor var_6334 = const()[name = tensor("op_6334"), val = tensor([1, 1])]; + tensor var_6336 = const()[name = tensor("op_6336"), val = tensor([1, 1])]; + tensor lora_out_621_pad_type_0 = const()[name = tensor("lora_out_621_pad_type_0"), val = tensor("custom")]; + tensor lora_out_621_pad_0 = const()[name = tensor("lora_out_621_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_623_weight_0_to_fp16 = const()[name = tensor("lora_out_623_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354260736)))]; + tensor lora_out_623_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6336, groups = var_6101, pad = lora_out_621_pad_0, pad_type = lora_out_621_pad_type_0, strides = var_6334, weight = lora_out_623_weight_0_to_fp16, x = input_463_cast_fp16)[name = tensor("lora_out_623_cast_fp16")]; + tensor key_63_cast_fp16 = add(x = pretrained_out_311_cast_fp16, y = lora_out_623_cast_fp16)[name = tensor("key_63_cast_fp16")]; + tensor var_6347 = const()[name = tensor("op_6347"), val = tensor([1, 1])]; + tensor var_6349 = const()[name = tensor("op_6349"), val = tensor([1, 1])]; + tensor pretrained_out_313_pad_type_0 = const()[name = tensor("pretrained_out_313_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_313_pad_0 = const()[name = tensor("pretrained_out_313_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354301760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355121024))), name = tensor("layers_15_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_15_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355121152)))]; + tensor pretrained_out_313_cast_fp16 = conv(bias = layers_15_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6349, groups = var_6101, pad = pretrained_out_313_pad_0, pad_type = pretrained_out_313_pad_type_0, strides = var_6347, weight = layers_15_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_313_cast_fp16")]; + tensor var_6353 = const()[name = tensor("op_6353"), val = tensor([1, 1])]; + tensor var_6355 = const()[name = tensor("op_6355"), val = tensor([1, 1])]; + tensor input_465_pad_type_0 = const()[name = tensor("input_465_pad_type_0"), val = tensor("custom")]; + tensor input_465_pad_0 = const()[name = tensor("input_465_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355123776)))]; + tensor input_465_cast_fp16 = conv(dilations = var_6355, groups = var_6101, pad = input_465_pad_0, pad_type = input_465_pad_type_0, strides = var_6353, weight = layers_15_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_465_cast_fp16")]; + tensor var_6359 = const()[name = tensor("op_6359"), val = tensor([1, 1])]; + tensor var_6361 = const()[name = tensor("op_6361"), val = tensor([1, 1])]; + tensor lora_out_625_pad_type_0 = const()[name = tensor("lora_out_625_pad_type_0"), val = tensor("custom")]; + tensor lora_out_625_pad_0 = const()[name = tensor("lora_out_625_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_627_weight_0_to_fp16 = const()[name = tensor("lora_out_627_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355164800)))]; + tensor lora_out_627_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6361, groups = var_6101, pad = lora_out_625_pad_0, pad_type = lora_out_625_pad_type_0, strides = var_6359, weight = lora_out_627_weight_0_to_fp16, x = input_465_cast_fp16)[name = tensor("lora_out_627_cast_fp16")]; + tensor value_63_cast_fp16 = add(x = pretrained_out_313_cast_fp16, y = lora_out_627_cast_fp16)[name = tensor("value_63_cast_fp16")]; + tensor var_6368 = const()[name = tensor("op_6368"), val = tensor([1, 20, 64, -1])]; + tensor var_6369_cast_fp16 = reshape(shape = var_6368, x = query_63_cast_fp16)[name = tensor("op_6369_cast_fp16")]; + tensor var_6370_to_fp16 = const()[name = tensor("op_6370_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6371_cast_fp16 = mul(x = var_6369_cast_fp16, y = var_6370_to_fp16)[name = tensor("op_6371_cast_fp16")]; + tensor var_6372 = const()[name = tensor("op_6372"), val = tensor([1, 20, 64, -1])]; + tensor var_6373_cast_fp16 = reshape(shape = var_6372, x = key_63_cast_fp16)[name = tensor("op_6373_cast_fp16")]; + tensor mh_w_95_transpose_x_0 = const()[name = tensor("mh_w_95_transpose_x_0"), val = tensor(true)]; + tensor mh_w_95_transpose_y_0 = const()[name = tensor("mh_w_95_transpose_y_0"), val = tensor(false)]; + tensor mh_w_95_cast_fp16 = matmul(transpose_x = mh_w_95_transpose_x_0, transpose_y = mh_w_95_transpose_y_0, x = var_6371_cast_fp16, y = var_6373_cast_fp16)[name = tensor("mh_w_95_cast_fp16")]; + tensor obj_223_cast_fp16 = softmax(axis = var_6094, x = mh_w_95_cast_fp16)[name = tensor("obj_223_cast_fp16")]; + tensor var_6377 = const()[name = tensor("op_6377"), val = tensor([1, 20, 64, -1])]; + tensor var_6378_cast_fp16 = reshape(shape = var_6377, x = value_63_cast_fp16)[name = tensor("op_6378_cast_fp16")]; + tensor attn_63_transpose_x_0 = const()[name = tensor("attn_63_transpose_x_0"), val = tensor(false)]; + tensor attn_63_transpose_y_0 = const()[name = tensor("attn_63_transpose_y_0"), val = tensor(true)]; + tensor attn_63_cast_fp16 = matmul(transpose_x = attn_63_transpose_x_0, transpose_y = attn_63_transpose_y_0, x = var_6378_cast_fp16, y = obj_223_cast_fp16)[name = tensor("attn_63_cast_fp16")]; + tensor var_6381 = const()[name = tensor("op_6381"), val = tensor([1, 1280, 1, -1])]; + tensor input_467_cast_fp16 = reshape(shape = var_6381, x = attn_63_cast_fp16)[name = tensor("input_467_cast_fp16")]; + tensor var_6388 = const()[name = tensor("op_6388"), val = tensor([1, 1])]; + tensor var_6390 = const()[name = tensor("op_6390"), val = tensor([1, 1])]; + tensor pretrained_out_315_pad_type_0 = const()[name = tensor("pretrained_out_315_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_315_pad_0 = const()[name = tensor("pretrained_out_315_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355205824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356025088))), name = tensor("layers_15_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_15_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356025216)))]; + tensor pretrained_out_315_cast_fp16 = conv(bias = layers_15_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_6390, groups = var_6101, pad = pretrained_out_315_pad_0, pad_type = pretrained_out_315_pad_type_0, strides = var_6388, weight = layers_15_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_467_cast_fp16)[name = tensor("pretrained_out_315_cast_fp16")]; + tensor var_6394 = const()[name = tensor("op_6394"), val = tensor([1, 1])]; + tensor var_6396 = const()[name = tensor("op_6396"), val = tensor([1, 1])]; + tensor input_469_pad_type_0 = const()[name = tensor("input_469_pad_type_0"), val = tensor("custom")]; + tensor input_469_pad_0 = const()[name = tensor("input_469_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356027840)))]; + tensor input_469_cast_fp16 = conv(dilations = var_6396, groups = var_6101, pad = input_469_pad_0, pad_type = input_469_pad_type_0, strides = var_6394, weight = layers_15_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_467_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor var_6400 = const()[name = tensor("op_6400"), val = tensor([1, 1])]; + tensor var_6402 = const()[name = tensor("op_6402"), val = tensor([1, 1])]; + tensor lora_out_629_pad_type_0 = const()[name = tensor("lora_out_629_pad_type_0"), val = tensor("custom")]; + tensor lora_out_629_pad_0 = const()[name = tensor("lora_out_629_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_631_weight_0_to_fp16 = const()[name = tensor("lora_out_631_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356068864)))]; + tensor lora_out_631_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6402, groups = var_6101, pad = lora_out_629_pad_0, pad_type = lora_out_629_pad_type_0, strides = var_6400, weight = lora_out_631_weight_0_to_fp16, x = input_469_cast_fp16)[name = tensor("lora_out_631_cast_fp16")]; + tensor obj_221_cast_fp16 = add(x = pretrained_out_315_cast_fp16, y = lora_out_631_cast_fp16)[name = tensor("obj_221_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_221_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor var_6411 = const()[name = tensor("op_6411"), val = tensor([1])]; + tensor channels_mean_95_cast_fp16 = reduce_mean(axes = var_6411, keep_dims = var_6102, x = inputs_95_cast_fp16)[name = tensor("channels_mean_95_cast_fp16")]; + tensor zero_mean_95_cast_fp16 = sub(x = inputs_95_cast_fp16, y = channels_mean_95_cast_fp16)[name = tensor("zero_mean_95_cast_fp16")]; + tensor zero_mean_sq_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = zero_mean_95_cast_fp16)[name = tensor("zero_mean_sq_95_cast_fp16")]; + tensor var_6415 = const()[name = tensor("op_6415"), val = tensor([1])]; + tensor var_6416_cast_fp16 = reduce_mean(axes = var_6415, keep_dims = var_6102, x = zero_mean_sq_95_cast_fp16)[name = tensor("op_6416_cast_fp16")]; + tensor var_6417_to_fp16 = const()[name = tensor("op_6417_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6418_cast_fp16 = add(x = var_6416_cast_fp16, y = var_6417_to_fp16)[name = tensor("op_6418_cast_fp16")]; + tensor denom_95_epsilon_0 = const()[name = tensor("denom_95_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_95_cast_fp16 = rsqrt(epsilon = denom_95_epsilon_0, x = var_6418_cast_fp16)[name = tensor("denom_95_cast_fp16")]; + tensor out_95_cast_fp16 = mul(x = zero_mean_95_cast_fp16, y = denom_95_cast_fp16)[name = tensor("out_95_cast_fp16")]; + tensor input_471_gamma_0_to_fp16 = const()[name = tensor("input_471_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356109888)))]; + tensor input_471_beta_0_to_fp16 = const()[name = tensor("input_471_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356112512)))]; + tensor input_471_epsilon_0_to_fp16 = const()[name = tensor("input_471_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_471_cast_fp16 = batch_norm(beta = input_471_beta_0_to_fp16, epsilon = input_471_epsilon_0_to_fp16, gamma = input_471_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor("input_471_cast_fp16")]; + tensor var_6432 = const()[name = tensor("op_6432"), val = tensor([1, 1])]; + tensor var_6434 = const()[name = tensor("op_6434"), val = tensor([1, 1])]; + tensor pretrained_out_317_pad_type_0 = const()[name = tensor("pretrained_out_317_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_317_pad_0 = const()[name = tensor("pretrained_out_317_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356115136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359392000))), name = tensor("layers_15_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_15_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359392128)))]; + tensor pretrained_out_317_cast_fp16 = conv(bias = layers_15_fc1_pretrained_bias_to_fp16, dilations = var_6434, groups = var_6101, pad = pretrained_out_317_pad_0, pad_type = pretrained_out_317_pad_type_0, strides = var_6432, weight = layers_15_fc1_pretrained_weight_to_fp16_palettized, x = input_471_cast_fp16)[name = tensor("pretrained_out_317_cast_fp16")]; + tensor var_6438 = const()[name = tensor("op_6438"), val = tensor([1, 1])]; + tensor var_6440 = const()[name = tensor("op_6440"), val = tensor([1, 1])]; + tensor input_473_pad_type_0 = const()[name = tensor("input_473_pad_type_0"), val = tensor("custom")]; + tensor input_473_pad_0 = const()[name = tensor("input_473_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_15_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359402432)))]; + tensor input_473_cast_fp16 = conv(dilations = var_6440, groups = var_6101, pad = input_473_pad_0, pad_type = input_473_pad_type_0, strides = var_6438, weight = layers_15_fc1_loraA_weight_to_fp16, x = input_471_cast_fp16)[name = tensor("input_473_cast_fp16")]; + tensor var_6444 = const()[name = tensor("op_6444"), val = tensor([1, 1])]; + tensor var_6446 = const()[name = tensor("op_6446"), val = tensor([1, 1])]; + tensor lora_out_633_pad_type_0 = const()[name = tensor("lora_out_633_pad_type_0"), val = tensor("custom")]; + tensor lora_out_633_pad_0 = const()[name = tensor("lora_out_633_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_635_weight_0_to_fp16 = const()[name = tensor("lora_out_635_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359443456)))]; + tensor lora_out_635_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_6446, groups = var_6101, pad = lora_out_633_pad_0, pad_type = lora_out_633_pad_type_0, strides = var_6444, weight = lora_out_635_weight_0_to_fp16, x = input_473_cast_fp16)[name = tensor("lora_out_635_cast_fp16")]; + tensor input_475_cast_fp16 = add(x = pretrained_out_317_cast_fp16, y = lora_out_635_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor input_477_mode_0 = const()[name = tensor("input_477_mode_0"), val = tensor("EXACT")]; + tensor input_477_cast_fp16 = gelu(mode = input_477_mode_0, x = input_475_cast_fp16)[name = tensor("input_477_cast_fp16")]; + tensor var_6458 = const()[name = tensor("op_6458"), val = tensor([1, 1])]; + tensor var_6460 = const()[name = tensor("op_6460"), val = tensor([1, 1])]; + tensor pretrained_out_319_pad_type_0 = const()[name = tensor("pretrained_out_319_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_319_pad_0 = const()[name = tensor("pretrained_out_319_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359607360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362884224))), name = tensor("layers_15_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_15_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_15_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362884352)))]; + tensor pretrained_out_319_cast_fp16 = conv(bias = layers_15_fc2_pretrained_bias_to_fp16, dilations = var_6460, groups = var_6101, pad = pretrained_out_319_pad_0, pad_type = pretrained_out_319_pad_type_0, strides = var_6458, weight = layers_15_fc2_pretrained_weight_to_fp16_palettized, x = input_477_cast_fp16)[name = tensor("pretrained_out_319_cast_fp16")]; + tensor var_6464 = const()[name = tensor("op_6464"), val = tensor([1, 1])]; + tensor var_6466 = const()[name = tensor("op_6466"), val = tensor([1, 1])]; + tensor input_479_pad_type_0 = const()[name = tensor("input_479_pad_type_0"), val = tensor("custom")]; + tensor input_479_pad_0 = const()[name = tensor("input_479_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_15_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_15_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362886976)))]; + tensor input_479_cast_fp16 = conv(dilations = var_6466, groups = var_6101, pad = input_479_pad_0, pad_type = input_479_pad_type_0, strides = var_6464, weight = layers_15_fc2_loraA_weight_to_fp16, x = input_477_cast_fp16)[name = tensor("input_479_cast_fp16")]; + tensor var_6470 = const()[name = tensor("op_6470"), val = tensor([1, 1])]; + tensor var_6472 = const()[name = tensor("op_6472"), val = tensor([1, 1])]; + tensor lora_out_637_pad_type_0 = const()[name = tensor("lora_out_637_pad_type_0"), val = tensor("custom")]; + tensor lora_out_637_pad_0 = const()[name = tensor("lora_out_637_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_639_weight_0_to_fp16 = const()[name = tensor("lora_out_639_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363050880)))]; + tensor lora_out_639_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6472, groups = var_6101, pad = lora_out_637_pad_0, pad_type = lora_out_637_pad_type_0, strides = var_6470, weight = lora_out_639_weight_0_to_fp16, x = input_479_cast_fp16)[name = tensor("lora_out_639_cast_fp16")]; + tensor hidden_states_33_cast_fp16 = add(x = pretrained_out_319_cast_fp16, y = lora_out_639_cast_fp16)[name = tensor("hidden_states_33_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; + tensor var_6488 = const()[name = tensor("op_6488"), val = tensor(3)]; + tensor var_6495 = const()[name = tensor("op_6495"), val = tensor(1)]; + tensor var_6496 = const()[name = tensor("op_6496"), val = tensor(true)]; + tensor var_6508 = const()[name = tensor("op_6508"), val = tensor([1])]; + tensor channels_mean_97_cast_fp16 = reduce_mean(axes = var_6508, keep_dims = var_6496, x = inputs_97_cast_fp16)[name = tensor("channels_mean_97_cast_fp16")]; + tensor zero_mean_97_cast_fp16 = sub(x = inputs_97_cast_fp16, y = channels_mean_97_cast_fp16)[name = tensor("zero_mean_97_cast_fp16")]; + tensor zero_mean_sq_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = zero_mean_97_cast_fp16)[name = tensor("zero_mean_sq_97_cast_fp16")]; + tensor var_6512 = const()[name = tensor("op_6512"), val = tensor([1])]; + tensor var_6513_cast_fp16 = reduce_mean(axes = var_6512, keep_dims = var_6496, x = zero_mean_sq_97_cast_fp16)[name = tensor("op_6513_cast_fp16")]; + tensor var_6514_to_fp16 = const()[name = tensor("op_6514_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6515_cast_fp16 = add(x = var_6513_cast_fp16, y = var_6514_to_fp16)[name = tensor("op_6515_cast_fp16")]; + tensor denom_97_epsilon_0 = const()[name = tensor("denom_97_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_97_cast_fp16 = rsqrt(epsilon = denom_97_epsilon_0, x = var_6515_cast_fp16)[name = tensor("denom_97_cast_fp16")]; + tensor out_97_cast_fp16 = mul(x = zero_mean_97_cast_fp16, y = denom_97_cast_fp16)[name = tensor("out_97_cast_fp16")]; + tensor obj_225_gamma_0_to_fp16 = const()[name = tensor("obj_225_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363091904)))]; + tensor obj_225_beta_0_to_fp16 = const()[name = tensor("obj_225_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363094528)))]; + tensor obj_225_epsilon_0_to_fp16 = const()[name = tensor("obj_225_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_225_cast_fp16 = batch_norm(beta = obj_225_beta_0_to_fp16, epsilon = obj_225_epsilon_0_to_fp16, gamma = obj_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor("obj_225_cast_fp16")]; + tensor var_6533 = const()[name = tensor("op_6533"), val = tensor([1, 1])]; + tensor var_6535 = const()[name = tensor("op_6535"), val = tensor([1, 1])]; + tensor pretrained_out_321_pad_type_0 = const()[name = tensor("pretrained_out_321_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_321_pad_0 = const()[name = tensor("pretrained_out_321_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363097152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363916416))), name = tensor("layers_16_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_16_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363916544)))]; + tensor pretrained_out_321_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6535, groups = var_6495, pad = pretrained_out_321_pad_0, pad_type = pretrained_out_321_pad_type_0, strides = var_6533, weight = layers_16_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_225_cast_fp16)[name = tensor("pretrained_out_321_cast_fp16")]; + tensor var_6539 = const()[name = tensor("op_6539"), val = tensor([1, 1])]; + tensor var_6541 = const()[name = tensor("op_6541"), val = tensor([1, 1])]; + tensor input_481_pad_type_0 = const()[name = tensor("input_481_pad_type_0"), val = tensor("custom")]; + tensor input_481_pad_0 = const()[name = tensor("input_481_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363919168)))]; + tensor input_481_cast_fp16 = conv(dilations = var_6541, groups = var_6495, pad = input_481_pad_0, pad_type = input_481_pad_type_0, strides = var_6539, weight = layers_16_self_attn_q_proj_loraA_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("input_481_cast_fp16")]; + tensor var_6545 = const()[name = tensor("op_6545"), val = tensor([1, 1])]; + tensor var_6547 = const()[name = tensor("op_6547"), val = tensor([1, 1])]; + tensor lora_out_641_pad_type_0 = const()[name = tensor("lora_out_641_pad_type_0"), val = tensor("custom")]; + tensor lora_out_641_pad_0 = const()[name = tensor("lora_out_641_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_643_weight_0_to_fp16 = const()[name = tensor("lora_out_643_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(363960192)))]; + tensor lora_out_643_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6547, groups = var_6495, pad = lora_out_641_pad_0, pad_type = lora_out_641_pad_type_0, strides = var_6545, weight = lora_out_643_weight_0_to_fp16, x = input_481_cast_fp16)[name = tensor("lora_out_643_cast_fp16")]; + tensor query_65_cast_fp16 = add(x = pretrained_out_321_cast_fp16, y = lora_out_643_cast_fp16)[name = tensor("query_65_cast_fp16")]; + tensor var_6557 = const()[name = tensor("op_6557"), val = tensor([1, 1])]; + tensor var_6559 = const()[name = tensor("op_6559"), val = tensor([1, 1])]; + tensor pretrained_out_323_pad_type_0 = const()[name = tensor("pretrained_out_323_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_323_pad_0 = const()[name = tensor("pretrained_out_323_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364001216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364820480))), name = tensor("layers_16_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_323_cast_fp16 = conv(dilations = var_6559, groups = var_6495, pad = pretrained_out_323_pad_0, pad_type = pretrained_out_323_pad_type_0, strides = var_6557, weight = layers_16_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_225_cast_fp16)[name = tensor("pretrained_out_323_cast_fp16")]; + tensor var_6563 = const()[name = tensor("op_6563"), val = tensor([1, 1])]; + tensor var_6565 = const()[name = tensor("op_6565"), val = tensor([1, 1])]; + tensor input_483_pad_type_0 = const()[name = tensor("input_483_pad_type_0"), val = tensor("custom")]; + tensor input_483_pad_0 = const()[name = tensor("input_483_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364820608)))]; + tensor input_483_cast_fp16 = conv(dilations = var_6565, groups = var_6495, pad = input_483_pad_0, pad_type = input_483_pad_type_0, strides = var_6563, weight = layers_16_self_attn_k_proj_loraA_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("input_483_cast_fp16")]; + tensor var_6569 = const()[name = tensor("op_6569"), val = tensor([1, 1])]; + tensor var_6571 = const()[name = tensor("op_6571"), val = tensor([1, 1])]; + tensor lora_out_645_pad_type_0 = const()[name = tensor("lora_out_645_pad_type_0"), val = tensor("custom")]; + tensor lora_out_645_pad_0 = const()[name = tensor("lora_out_645_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_647_weight_0_to_fp16 = const()[name = tensor("lora_out_647_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364861632)))]; + tensor lora_out_647_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6571, groups = var_6495, pad = lora_out_645_pad_0, pad_type = lora_out_645_pad_type_0, strides = var_6569, weight = lora_out_647_weight_0_to_fp16, x = input_483_cast_fp16)[name = tensor("lora_out_647_cast_fp16")]; + tensor current_key_33_cast_fp16 = add(x = pretrained_out_323_cast_fp16, y = lora_out_647_cast_fp16)[name = tensor("current_key_33_cast_fp16")]; + tensor var_6582 = const()[name = tensor("op_6582"), val = tensor([1, 1])]; + tensor var_6584 = const()[name = tensor("op_6584"), val = tensor([1, 1])]; + tensor pretrained_out_325_pad_type_0 = const()[name = tensor("pretrained_out_325_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_325_pad_0 = const()[name = tensor("pretrained_out_325_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(364902656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365721920))), name = tensor("layers_16_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_16_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365722048)))]; + tensor pretrained_out_325_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6584, groups = var_6495, pad = pretrained_out_325_pad_0, pad_type = pretrained_out_325_pad_type_0, strides = var_6582, weight = layers_16_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_225_cast_fp16)[name = tensor("pretrained_out_325_cast_fp16")]; + tensor var_6588 = const()[name = tensor("op_6588"), val = tensor([1, 1])]; + tensor var_6590 = const()[name = tensor("op_6590"), val = tensor([1, 1])]; + tensor input_485_pad_type_0 = const()[name = tensor("input_485_pad_type_0"), val = tensor("custom")]; + tensor input_485_pad_0 = const()[name = tensor("input_485_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365724672)))]; + tensor input_485_cast_fp16 = conv(dilations = var_6590, groups = var_6495, pad = input_485_pad_0, pad_type = input_485_pad_type_0, strides = var_6588, weight = layers_16_self_attn_v_proj_loraA_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("input_485_cast_fp16")]; + tensor var_6594 = const()[name = tensor("op_6594"), val = tensor([1, 1])]; + tensor var_6596 = const()[name = tensor("op_6596"), val = tensor([1, 1])]; + tensor lora_out_649_pad_type_0 = const()[name = tensor("lora_out_649_pad_type_0"), val = tensor("custom")]; + tensor lora_out_649_pad_0 = const()[name = tensor("lora_out_649_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_651_weight_0_to_fp16 = const()[name = tensor("lora_out_651_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365765696)))]; + tensor lora_out_651_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6596, groups = var_6495, pad = lora_out_649_pad_0, pad_type = lora_out_649_pad_type_0, strides = var_6594, weight = lora_out_651_weight_0_to_fp16, x = input_485_cast_fp16)[name = tensor("lora_out_651_cast_fp16")]; + tensor current_value_33_cast_fp16 = add(x = pretrained_out_325_cast_fp16, y = lora_out_651_cast_fp16)[name = tensor("current_value_33_cast_fp16")]; + tensor var_6606_cast_fp16 = mul(x = current_key_33_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_6606_cast_fp16")]; + tensor var_6608_cast_fp16 = mul(x = var_103_cast_fp16_16, y = var_295_cast_fp16)[name = tensor("op_6608_cast_fp16")]; + tensor key_65_cast_fp16 = add(x = var_6606_cast_fp16, y = var_6608_cast_fp16)[name = tensor("key_65_cast_fp16")]; + tensor var_6610_cast_fp16 = mul(x = current_value_33_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_6610_cast_fp16")]; + tensor var_6612_cast_fp16 = mul(x = var_138_cast_fp16_16, y = var_295_cast_fp16)[name = tensor("op_6612_cast_fp16")]; + tensor value_65_cast_fp16 = add(x = var_6610_cast_fp16, y = var_6612_cast_fp16)[name = tensor("value_65_cast_fp16")]; + tensor var_6615 = const()[name = tensor("op_6615"), val = tensor([1, 20, 64, -1])]; + tensor var_6616_cast_fp16 = reshape(shape = var_6615, x = query_65_cast_fp16)[name = tensor("op_6616_cast_fp16")]; + tensor var_6617_to_fp16 = const()[name = tensor("op_6617_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6618_cast_fp16 = mul(x = var_6616_cast_fp16, y = var_6617_to_fp16)[name = tensor("op_6618_cast_fp16")]; + tensor var_6619 = const()[name = tensor("op_6619"), val = tensor([1, 20, 64, -1])]; + tensor var_6620_cast_fp16 = reshape(shape = var_6619, x = key_65_cast_fp16)[name = tensor("op_6620_cast_fp16")]; + tensor mh_w_97_transpose_x_0 = const()[name = tensor("mh_w_97_transpose_x_0"), val = tensor(true)]; + tensor mh_w_97_transpose_y_0 = const()[name = tensor("mh_w_97_transpose_y_0"), val = tensor(false)]; + tensor mh_w_97_cast_fp16 = matmul(transpose_x = mh_w_97_transpose_x_0, transpose_y = mh_w_97_transpose_y_0, x = var_6618_cast_fp16, y = var_6620_cast_fp16)[name = tensor("mh_w_97_cast_fp16")]; + tensor mh_w_99_cast_fp16 = add(x = mh_w_97_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_99_cast_fp16")]; + tensor var_6628_cast_fp16 = softmax(axis = var_6488, x = mh_w_99_cast_fp16)[name = tensor("op_6628_cast_fp16")]; + tensor var_6629 = const()[name = tensor("op_6629"), val = tensor([1, 20, 64, -1])]; + tensor var_6630_cast_fp16 = reshape(shape = var_6629, x = value_65_cast_fp16)[name = tensor("op_6630_cast_fp16")]; + tensor attn_65_transpose_x_0 = const()[name = tensor("attn_65_transpose_x_0"), val = tensor(false)]; + tensor attn_65_transpose_y_0 = const()[name = tensor("attn_65_transpose_y_0"), val = tensor(true)]; + tensor attn_65_cast_fp16 = matmul(transpose_x = attn_65_transpose_x_0, transpose_y = attn_65_transpose_y_0, x = var_6630_cast_fp16, y = var_6628_cast_fp16)[name = tensor("attn_65_cast_fp16")]; + tensor var_6633 = const()[name = tensor("op_6633"), val = tensor([1, 1280, 1, -1])]; + tensor input_487_cast_fp16 = reshape(shape = var_6633, x = attn_65_cast_fp16)[name = tensor("input_487_cast_fp16")]; + tensor var_6640 = const()[name = tensor("op_6640"), val = tensor([1, 1])]; + tensor var_6642 = const()[name = tensor("op_6642"), val = tensor([1, 1])]; + tensor pretrained_out_327_pad_type_0 = const()[name = tensor("pretrained_out_327_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_327_pad_0 = const()[name = tensor("pretrained_out_327_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(365806720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366625984))), name = tensor("layers_16_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_16_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366626112)))]; + tensor pretrained_out_327_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_6642, groups = var_6495, pad = pretrained_out_327_pad_0, pad_type = pretrained_out_327_pad_type_0, strides = var_6640, weight = layers_16_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_487_cast_fp16)[name = tensor("pretrained_out_327_cast_fp16")]; + tensor var_6646 = const()[name = tensor("op_6646"), val = tensor([1, 1])]; + tensor var_6648 = const()[name = tensor("op_6648"), val = tensor([1, 1])]; + tensor input_489_pad_type_0 = const()[name = tensor("input_489_pad_type_0"), val = tensor("custom")]; + tensor input_489_pad_0 = const()[name = tensor("input_489_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366628736)))]; + tensor input_489_cast_fp16 = conv(dilations = var_6648, groups = var_6495, pad = input_489_pad_0, pad_type = input_489_pad_type_0, strides = var_6646, weight = layers_16_self_attn_o_proj_loraA_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor var_6652 = const()[name = tensor("op_6652"), val = tensor([1, 1])]; + tensor var_6654 = const()[name = tensor("op_6654"), val = tensor([1, 1])]; + tensor lora_out_653_pad_type_0 = const()[name = tensor("lora_out_653_pad_type_0"), val = tensor("custom")]; + tensor lora_out_653_pad_0 = const()[name = tensor("lora_out_653_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_655_weight_0_to_fp16 = const()[name = tensor("lora_out_655_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366669760)))]; + tensor lora_out_655_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6654, groups = var_6495, pad = lora_out_653_pad_0, pad_type = lora_out_653_pad_type_0, strides = var_6652, weight = lora_out_655_weight_0_to_fp16, x = input_489_cast_fp16)[name = tensor("lora_out_655_cast_fp16")]; + tensor obj_231_cast_fp16 = add(x = pretrained_out_327_cast_fp16, y = lora_out_655_cast_fp16)[name = tensor("obj_231_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_231_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; + tensor var_6667 = const()[name = tensor("op_6667"), val = tensor([1])]; + tensor channels_mean_99_cast_fp16 = reduce_mean(axes = var_6667, keep_dims = var_6496, x = inputs_99_cast_fp16)[name = tensor("channels_mean_99_cast_fp16")]; + tensor zero_mean_99_cast_fp16 = sub(x = inputs_99_cast_fp16, y = channels_mean_99_cast_fp16)[name = tensor("zero_mean_99_cast_fp16")]; + tensor zero_mean_sq_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = zero_mean_99_cast_fp16)[name = tensor("zero_mean_sq_99_cast_fp16")]; + tensor var_6671 = const()[name = tensor("op_6671"), val = tensor([1])]; + tensor var_6672_cast_fp16 = reduce_mean(axes = var_6671, keep_dims = var_6496, x = zero_mean_sq_99_cast_fp16)[name = tensor("op_6672_cast_fp16")]; + tensor var_6673_to_fp16 = const()[name = tensor("op_6673_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6674_cast_fp16 = add(x = var_6672_cast_fp16, y = var_6673_to_fp16)[name = tensor("op_6674_cast_fp16")]; + tensor denom_99_epsilon_0 = const()[name = tensor("denom_99_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_99_cast_fp16 = rsqrt(epsilon = denom_99_epsilon_0, x = var_6674_cast_fp16)[name = tensor("denom_99_cast_fp16")]; + tensor out_99_cast_fp16 = mul(x = zero_mean_99_cast_fp16, y = denom_99_cast_fp16)[name = tensor("out_99_cast_fp16")]; + tensor obj_233_gamma_0_to_fp16 = const()[name = tensor("obj_233_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366710784)))]; + tensor obj_233_beta_0_to_fp16 = const()[name = tensor("obj_233_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366713408)))]; + tensor obj_233_epsilon_0_to_fp16 = const()[name = tensor("obj_233_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_233_cast_fp16 = batch_norm(beta = obj_233_beta_0_to_fp16, epsilon = obj_233_epsilon_0_to_fp16, gamma = obj_233_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor("obj_233_cast_fp16")]; + tensor var_6692 = const()[name = tensor("op_6692"), val = tensor([1, 1])]; + tensor var_6694 = const()[name = tensor("op_6694"), val = tensor([1, 1])]; + tensor pretrained_out_329_pad_type_0 = const()[name = tensor("pretrained_out_329_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_329_pad_0 = const()[name = tensor("pretrained_out_329_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366716032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(367535296))), name = tensor("layers_16_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_16_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(367535424)))]; + tensor pretrained_out_329_cast_fp16 = conv(bias = layers_16_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6694, groups = var_6495, pad = pretrained_out_329_pad_0, pad_type = pretrained_out_329_pad_type_0, strides = var_6692, weight = layers_16_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_233_cast_fp16)[name = tensor("pretrained_out_329_cast_fp16")]; + tensor var_6698 = const()[name = tensor("op_6698"), val = tensor([1, 1])]; + tensor var_6700 = const()[name = tensor("op_6700"), val = tensor([1, 1])]; + tensor input_491_pad_type_0 = const()[name = tensor("input_491_pad_type_0"), val = tensor("custom")]; + tensor input_491_pad_0 = const()[name = tensor("input_491_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(367538048)))]; + tensor input_491_cast_fp16 = conv(dilations = var_6700, groups = var_6495, pad = input_491_pad_0, pad_type = input_491_pad_type_0, strides = var_6698, weight = layers_16_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_233_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor var_6704 = const()[name = tensor("op_6704"), val = tensor([1, 1])]; + tensor var_6706 = const()[name = tensor("op_6706"), val = tensor([1, 1])]; + tensor lora_out_657_pad_type_0 = const()[name = tensor("lora_out_657_pad_type_0"), val = tensor("custom")]; + tensor lora_out_657_pad_0 = const()[name = tensor("lora_out_657_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_659_weight_0_to_fp16 = const()[name = tensor("lora_out_659_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(367579072)))]; + tensor lora_out_659_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6706, groups = var_6495, pad = lora_out_657_pad_0, pad_type = lora_out_657_pad_type_0, strides = var_6704, weight = lora_out_659_weight_0_to_fp16, x = input_491_cast_fp16)[name = tensor("lora_out_659_cast_fp16")]; + tensor query_67_cast_fp16 = add(x = pretrained_out_329_cast_fp16, y = lora_out_659_cast_fp16)[name = tensor("query_67_cast_fp16")]; + tensor var_6716 = const()[name = tensor("op_6716"), val = tensor([1, 1])]; + tensor var_6718 = const()[name = tensor("op_6718"), val = tensor([1, 1])]; + tensor pretrained_out_331_pad_type_0 = const()[name = tensor("pretrained_out_331_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_331_pad_0 = const()[name = tensor("pretrained_out_331_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(367620096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368439360))), name = tensor("layers_16_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_331_cast_fp16 = conv(dilations = var_6718, groups = var_6495, pad = pretrained_out_331_pad_0, pad_type = pretrained_out_331_pad_type_0, strides = var_6716, weight = layers_16_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_331_cast_fp16")]; + tensor var_6722 = const()[name = tensor("op_6722"), val = tensor([1, 1])]; + tensor var_6724 = const()[name = tensor("op_6724"), val = tensor([1, 1])]; + tensor input_493_pad_type_0 = const()[name = tensor("input_493_pad_type_0"), val = tensor("custom")]; + tensor input_493_pad_0 = const()[name = tensor("input_493_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368439488)))]; + tensor input_493_cast_fp16 = conv(dilations = var_6724, groups = var_6495, pad = input_493_pad_0, pad_type = input_493_pad_type_0, strides = var_6722, weight = layers_16_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_493_cast_fp16")]; + tensor var_6728 = const()[name = tensor("op_6728"), val = tensor([1, 1])]; + tensor var_6730 = const()[name = tensor("op_6730"), val = tensor([1, 1])]; + tensor lora_out_661_pad_type_0 = const()[name = tensor("lora_out_661_pad_type_0"), val = tensor("custom")]; + tensor lora_out_661_pad_0 = const()[name = tensor("lora_out_661_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_663_weight_0_to_fp16 = const()[name = tensor("lora_out_663_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368480512)))]; + tensor lora_out_663_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6730, groups = var_6495, pad = lora_out_661_pad_0, pad_type = lora_out_661_pad_type_0, strides = var_6728, weight = lora_out_663_weight_0_to_fp16, x = input_493_cast_fp16)[name = tensor("lora_out_663_cast_fp16")]; + tensor key_67_cast_fp16 = add(x = pretrained_out_331_cast_fp16, y = lora_out_663_cast_fp16)[name = tensor("key_67_cast_fp16")]; + tensor var_6741 = const()[name = tensor("op_6741"), val = tensor([1, 1])]; + tensor var_6743 = const()[name = tensor("op_6743"), val = tensor([1, 1])]; + tensor pretrained_out_333_pad_type_0 = const()[name = tensor("pretrained_out_333_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_333_pad_0 = const()[name = tensor("pretrained_out_333_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(368521536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369340800))), name = tensor("layers_16_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_16_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369340928)))]; + tensor pretrained_out_333_cast_fp16 = conv(bias = layers_16_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6743, groups = var_6495, pad = pretrained_out_333_pad_0, pad_type = pretrained_out_333_pad_type_0, strides = var_6741, weight = layers_16_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_333_cast_fp16")]; + tensor var_6747 = const()[name = tensor("op_6747"), val = tensor([1, 1])]; + tensor var_6749 = const()[name = tensor("op_6749"), val = tensor([1, 1])]; + tensor input_495_pad_type_0 = const()[name = tensor("input_495_pad_type_0"), val = tensor("custom")]; + tensor input_495_pad_0 = const()[name = tensor("input_495_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369343552)))]; + tensor input_495_cast_fp16 = conv(dilations = var_6749, groups = var_6495, pad = input_495_pad_0, pad_type = input_495_pad_type_0, strides = var_6747, weight = layers_16_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_495_cast_fp16")]; + tensor var_6753 = const()[name = tensor("op_6753"), val = tensor([1, 1])]; + tensor var_6755 = const()[name = tensor("op_6755"), val = tensor([1, 1])]; + tensor lora_out_665_pad_type_0 = const()[name = tensor("lora_out_665_pad_type_0"), val = tensor("custom")]; + tensor lora_out_665_pad_0 = const()[name = tensor("lora_out_665_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_667_weight_0_to_fp16 = const()[name = tensor("lora_out_667_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369384576)))]; + tensor lora_out_667_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6755, groups = var_6495, pad = lora_out_665_pad_0, pad_type = lora_out_665_pad_type_0, strides = var_6753, weight = lora_out_667_weight_0_to_fp16, x = input_495_cast_fp16)[name = tensor("lora_out_667_cast_fp16")]; + tensor value_67_cast_fp16 = add(x = pretrained_out_333_cast_fp16, y = lora_out_667_cast_fp16)[name = tensor("value_67_cast_fp16")]; + tensor var_6762 = const()[name = tensor("op_6762"), val = tensor([1, 20, 64, -1])]; + tensor var_6763_cast_fp16 = reshape(shape = var_6762, x = query_67_cast_fp16)[name = tensor("op_6763_cast_fp16")]; + tensor var_6764_to_fp16 = const()[name = tensor("op_6764_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6765_cast_fp16 = mul(x = var_6763_cast_fp16, y = var_6764_to_fp16)[name = tensor("op_6765_cast_fp16")]; + tensor var_6766 = const()[name = tensor("op_6766"), val = tensor([1, 20, 64, -1])]; + tensor var_6767_cast_fp16 = reshape(shape = var_6766, x = key_67_cast_fp16)[name = tensor("op_6767_cast_fp16")]; + tensor mh_w_101_transpose_x_0 = const()[name = tensor("mh_w_101_transpose_x_0"), val = tensor(true)]; + tensor mh_w_101_transpose_y_0 = const()[name = tensor("mh_w_101_transpose_y_0"), val = tensor(false)]; + tensor mh_w_101_cast_fp16 = matmul(transpose_x = mh_w_101_transpose_x_0, transpose_y = mh_w_101_transpose_y_0, x = var_6765_cast_fp16, y = var_6767_cast_fp16)[name = tensor("mh_w_101_cast_fp16")]; + tensor obj_237_cast_fp16 = softmax(axis = var_6488, x = mh_w_101_cast_fp16)[name = tensor("obj_237_cast_fp16")]; + tensor var_6771 = const()[name = tensor("op_6771"), val = tensor([1, 20, 64, -1])]; + tensor var_6772_cast_fp16 = reshape(shape = var_6771, x = value_67_cast_fp16)[name = tensor("op_6772_cast_fp16")]; + tensor attn_67_transpose_x_0 = const()[name = tensor("attn_67_transpose_x_0"), val = tensor(false)]; + tensor attn_67_transpose_y_0 = const()[name = tensor("attn_67_transpose_y_0"), val = tensor(true)]; + tensor attn_67_cast_fp16 = matmul(transpose_x = attn_67_transpose_x_0, transpose_y = attn_67_transpose_y_0, x = var_6772_cast_fp16, y = obj_237_cast_fp16)[name = tensor("attn_67_cast_fp16")]; + tensor var_6775 = const()[name = tensor("op_6775"), val = tensor([1, 1280, 1, -1])]; + tensor input_497_cast_fp16 = reshape(shape = var_6775, x = attn_67_cast_fp16)[name = tensor("input_497_cast_fp16")]; + tensor var_6782 = const()[name = tensor("op_6782"), val = tensor([1, 1])]; + tensor var_6784 = const()[name = tensor("op_6784"), val = tensor([1, 1])]; + tensor pretrained_out_335_pad_type_0 = const()[name = tensor("pretrained_out_335_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_335_pad_0 = const()[name = tensor("pretrained_out_335_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369425600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370244864))), name = tensor("layers_16_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_16_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370244992)))]; + tensor pretrained_out_335_cast_fp16 = conv(bias = layers_16_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_6784, groups = var_6495, pad = pretrained_out_335_pad_0, pad_type = pretrained_out_335_pad_type_0, strides = var_6782, weight = layers_16_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_497_cast_fp16)[name = tensor("pretrained_out_335_cast_fp16")]; + tensor var_6788 = const()[name = tensor("op_6788"), val = tensor([1, 1])]; + tensor var_6790 = const()[name = tensor("op_6790"), val = tensor([1, 1])]; + tensor input_499_pad_type_0 = const()[name = tensor("input_499_pad_type_0"), val = tensor("custom")]; + tensor input_499_pad_0 = const()[name = tensor("input_499_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370247616)))]; + tensor input_499_cast_fp16 = conv(dilations = var_6790, groups = var_6495, pad = input_499_pad_0, pad_type = input_499_pad_type_0, strides = var_6788, weight = layers_16_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("input_499_cast_fp16")]; + tensor var_6794 = const()[name = tensor("op_6794"), val = tensor([1, 1])]; + tensor var_6796 = const()[name = tensor("op_6796"), val = tensor([1, 1])]; + tensor lora_out_669_pad_type_0 = const()[name = tensor("lora_out_669_pad_type_0"), val = tensor("custom")]; + tensor lora_out_669_pad_0 = const()[name = tensor("lora_out_669_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_671_weight_0_to_fp16 = const()[name = tensor("lora_out_671_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370288640)))]; + tensor lora_out_671_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6796, groups = var_6495, pad = lora_out_669_pad_0, pad_type = lora_out_669_pad_type_0, strides = var_6794, weight = lora_out_671_weight_0_to_fp16, x = input_499_cast_fp16)[name = tensor("lora_out_671_cast_fp16")]; + tensor obj_235_cast_fp16 = add(x = pretrained_out_335_cast_fp16, y = lora_out_671_cast_fp16)[name = tensor("obj_235_cast_fp16")]; + tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = obj_235_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor var_6808 = const()[name = tensor("op_6808"), val = tensor([1])]; + tensor channels_mean_101_cast_fp16 = reduce_mean(axes = var_6808, keep_dims = var_6496, x = inputs_101_cast_fp16)[name = tensor("channels_mean_101_cast_fp16")]; + tensor zero_mean_101_cast_fp16 = sub(x = inputs_101_cast_fp16, y = channels_mean_101_cast_fp16)[name = tensor("zero_mean_101_cast_fp16")]; + tensor zero_mean_sq_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = zero_mean_101_cast_fp16)[name = tensor("zero_mean_sq_101_cast_fp16")]; + tensor var_6812 = const()[name = tensor("op_6812"), val = tensor([1])]; + tensor var_6813_cast_fp16 = reduce_mean(axes = var_6812, keep_dims = var_6496, x = zero_mean_sq_101_cast_fp16)[name = tensor("op_6813_cast_fp16")]; + tensor var_6814_to_fp16 = const()[name = tensor("op_6814_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6815_cast_fp16 = add(x = var_6813_cast_fp16, y = var_6814_to_fp16)[name = tensor("op_6815_cast_fp16")]; + tensor denom_101_epsilon_0 = const()[name = tensor("denom_101_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_101_cast_fp16 = rsqrt(epsilon = denom_101_epsilon_0, x = var_6815_cast_fp16)[name = tensor("denom_101_cast_fp16")]; + tensor out_101_cast_fp16 = mul(x = zero_mean_101_cast_fp16, y = denom_101_cast_fp16)[name = tensor("out_101_cast_fp16")]; + tensor input_501_gamma_0_to_fp16 = const()[name = tensor("input_501_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370329664)))]; + tensor input_501_beta_0_to_fp16 = const()[name = tensor("input_501_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370332288)))]; + tensor input_501_epsilon_0_to_fp16 = const()[name = tensor("input_501_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_501_cast_fp16 = batch_norm(beta = input_501_beta_0_to_fp16, epsilon = input_501_epsilon_0_to_fp16, gamma = input_501_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor("input_501_cast_fp16")]; + tensor var_6829 = const()[name = tensor("op_6829"), val = tensor([1, 1])]; + tensor var_6831 = const()[name = tensor("op_6831"), val = tensor([1, 1])]; + tensor pretrained_out_337_pad_type_0 = const()[name = tensor("pretrained_out_337_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_337_pad_0 = const()[name = tensor("pretrained_out_337_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370334912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373611776))), name = tensor("layers_16_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_16_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373611904)))]; + tensor pretrained_out_337_cast_fp16 = conv(bias = layers_16_fc1_pretrained_bias_to_fp16, dilations = var_6831, groups = var_6495, pad = pretrained_out_337_pad_0, pad_type = pretrained_out_337_pad_type_0, strides = var_6829, weight = layers_16_fc1_pretrained_weight_to_fp16_palettized, x = input_501_cast_fp16)[name = tensor("pretrained_out_337_cast_fp16")]; + tensor var_6835 = const()[name = tensor("op_6835"), val = tensor([1, 1])]; + tensor var_6837 = const()[name = tensor("op_6837"), val = tensor([1, 1])]; + tensor input_503_pad_type_0 = const()[name = tensor("input_503_pad_type_0"), val = tensor("custom")]; + tensor input_503_pad_0 = const()[name = tensor("input_503_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_16_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373622208)))]; + tensor input_503_cast_fp16 = conv(dilations = var_6837, groups = var_6495, pad = input_503_pad_0, pad_type = input_503_pad_type_0, strides = var_6835, weight = layers_16_fc1_loraA_weight_to_fp16, x = input_501_cast_fp16)[name = tensor("input_503_cast_fp16")]; + tensor var_6841 = const()[name = tensor("op_6841"), val = tensor([1, 1])]; + tensor var_6843 = const()[name = tensor("op_6843"), val = tensor([1, 1])]; + tensor lora_out_673_pad_type_0 = const()[name = tensor("lora_out_673_pad_type_0"), val = tensor("custom")]; + tensor lora_out_673_pad_0 = const()[name = tensor("lora_out_673_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_675_weight_0_to_fp16 = const()[name = tensor("lora_out_675_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373663232)))]; + tensor lora_out_675_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_6843, groups = var_6495, pad = lora_out_673_pad_0, pad_type = lora_out_673_pad_type_0, strides = var_6841, weight = lora_out_675_weight_0_to_fp16, x = input_503_cast_fp16)[name = tensor("lora_out_675_cast_fp16")]; + tensor input_505_cast_fp16 = add(x = pretrained_out_337_cast_fp16, y = lora_out_675_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor input_507_mode_0 = const()[name = tensor("input_507_mode_0"), val = tensor("EXACT")]; + tensor input_507_cast_fp16 = gelu(mode = input_507_mode_0, x = input_505_cast_fp16)[name = tensor("input_507_cast_fp16")]; + tensor var_6855 = const()[name = tensor("op_6855"), val = tensor([1, 1])]; + tensor var_6857 = const()[name = tensor("op_6857"), val = tensor([1, 1])]; + tensor pretrained_out_339_pad_type_0 = const()[name = tensor("pretrained_out_339_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_339_pad_0 = const()[name = tensor("pretrained_out_339_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373827136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377104000))), name = tensor("layers_16_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_16_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_16_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377104128)))]; + tensor pretrained_out_339_cast_fp16 = conv(bias = layers_16_fc2_pretrained_bias_to_fp16, dilations = var_6857, groups = var_6495, pad = pretrained_out_339_pad_0, pad_type = pretrained_out_339_pad_type_0, strides = var_6855, weight = layers_16_fc2_pretrained_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = tensor("pretrained_out_339_cast_fp16")]; + tensor var_6861 = const()[name = tensor("op_6861"), val = tensor([1, 1])]; + tensor var_6863 = const()[name = tensor("op_6863"), val = tensor([1, 1])]; + tensor input_509_pad_type_0 = const()[name = tensor("input_509_pad_type_0"), val = tensor("custom")]; + tensor input_509_pad_0 = const()[name = tensor("input_509_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_16_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_16_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377106752)))]; + tensor input_509_cast_fp16 = conv(dilations = var_6863, groups = var_6495, pad = input_509_pad_0, pad_type = input_509_pad_type_0, strides = var_6861, weight = layers_16_fc2_loraA_weight_to_fp16, x = input_507_cast_fp16)[name = tensor("input_509_cast_fp16")]; + tensor var_6867 = const()[name = tensor("op_6867"), val = tensor([1, 1])]; + tensor var_6869 = const()[name = tensor("op_6869"), val = tensor([1, 1])]; + tensor lora_out_677_pad_type_0 = const()[name = tensor("lora_out_677_pad_type_0"), val = tensor("custom")]; + tensor lora_out_677_pad_0 = const()[name = tensor("lora_out_677_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_679_weight_0_to_fp16 = const()[name = tensor("lora_out_679_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377270656)))]; + tensor lora_out_679_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6869, groups = var_6495, pad = lora_out_677_pad_0, pad_type = lora_out_677_pad_type_0, strides = var_6867, weight = lora_out_679_weight_0_to_fp16, x = input_509_cast_fp16)[name = tensor("lora_out_679_cast_fp16")]; + tensor hidden_states_35_cast_fp16 = add(x = pretrained_out_339_cast_fp16, y = lora_out_679_cast_fp16)[name = tensor("hidden_states_35_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; + tensor var_6886 = const()[name = tensor("op_6886"), val = tensor(3)]; + tensor var_6893 = const()[name = tensor("op_6893"), val = tensor(1)]; + tensor var_6894 = const()[name = tensor("op_6894"), val = tensor(true)]; + tensor var_6906 = const()[name = tensor("op_6906"), val = tensor([1])]; + tensor channels_mean_103_cast_fp16 = reduce_mean(axes = var_6906, keep_dims = var_6894, x = inputs_103_cast_fp16)[name = tensor("channels_mean_103_cast_fp16")]; + tensor zero_mean_103_cast_fp16 = sub(x = inputs_103_cast_fp16, y = channels_mean_103_cast_fp16)[name = tensor("zero_mean_103_cast_fp16")]; + tensor zero_mean_sq_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = zero_mean_103_cast_fp16)[name = tensor("zero_mean_sq_103_cast_fp16")]; + tensor var_6910 = const()[name = tensor("op_6910"), val = tensor([1])]; + tensor var_6911_cast_fp16 = reduce_mean(axes = var_6910, keep_dims = var_6894, x = zero_mean_sq_103_cast_fp16)[name = tensor("op_6911_cast_fp16")]; + tensor var_6912_to_fp16 = const()[name = tensor("op_6912_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6913_cast_fp16 = add(x = var_6911_cast_fp16, y = var_6912_to_fp16)[name = tensor("op_6913_cast_fp16")]; + tensor denom_103_epsilon_0 = const()[name = tensor("denom_103_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_103_cast_fp16 = rsqrt(epsilon = denom_103_epsilon_0, x = var_6913_cast_fp16)[name = tensor("denom_103_cast_fp16")]; + tensor out_103_cast_fp16 = mul(x = zero_mean_103_cast_fp16, y = denom_103_cast_fp16)[name = tensor("out_103_cast_fp16")]; + tensor obj_239_gamma_0_to_fp16 = const()[name = tensor("obj_239_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377311680)))]; + tensor obj_239_beta_0_to_fp16 = const()[name = tensor("obj_239_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377314304)))]; + tensor obj_239_epsilon_0_to_fp16 = const()[name = tensor("obj_239_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_239_cast_fp16 = batch_norm(beta = obj_239_beta_0_to_fp16, epsilon = obj_239_epsilon_0_to_fp16, gamma = obj_239_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor("obj_239_cast_fp16")]; + tensor var_6931 = const()[name = tensor("op_6931"), val = tensor([1, 1])]; + tensor var_6933 = const()[name = tensor("op_6933"), val = tensor([1, 1])]; + tensor pretrained_out_341_pad_type_0 = const()[name = tensor("pretrained_out_341_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_341_pad_0 = const()[name = tensor("pretrained_out_341_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377316928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378136192))), name = tensor("layers_17_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_17_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378136320)))]; + tensor pretrained_out_341_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6933, groups = var_6893, pad = pretrained_out_341_pad_0, pad_type = pretrained_out_341_pad_type_0, strides = var_6931, weight = layers_17_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_239_cast_fp16)[name = tensor("pretrained_out_341_cast_fp16")]; + tensor var_6937 = const()[name = tensor("op_6937"), val = tensor([1, 1])]; + tensor var_6939 = const()[name = tensor("op_6939"), val = tensor([1, 1])]; + tensor input_511_pad_type_0 = const()[name = tensor("input_511_pad_type_0"), val = tensor("custom")]; + tensor input_511_pad_0 = const()[name = tensor("input_511_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378138944)))]; + tensor input_511_cast_fp16 = conv(dilations = var_6939, groups = var_6893, pad = input_511_pad_0, pad_type = input_511_pad_type_0, strides = var_6937, weight = layers_17_self_attn_q_proj_loraA_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("input_511_cast_fp16")]; + tensor var_6943 = const()[name = tensor("op_6943"), val = tensor([1, 1])]; + tensor var_6945 = const()[name = tensor("op_6945"), val = tensor([1, 1])]; + tensor lora_out_681_pad_type_0 = const()[name = tensor("lora_out_681_pad_type_0"), val = tensor("custom")]; + tensor lora_out_681_pad_0 = const()[name = tensor("lora_out_681_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_683_weight_0_to_fp16 = const()[name = tensor("lora_out_683_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378179968)))]; + tensor lora_out_683_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6945, groups = var_6893, pad = lora_out_681_pad_0, pad_type = lora_out_681_pad_type_0, strides = var_6943, weight = lora_out_683_weight_0_to_fp16, x = input_511_cast_fp16)[name = tensor("lora_out_683_cast_fp16")]; + tensor query_69_cast_fp16 = add(x = pretrained_out_341_cast_fp16, y = lora_out_683_cast_fp16)[name = tensor("query_69_cast_fp16")]; + tensor var_6955 = const()[name = tensor("op_6955"), val = tensor([1, 1])]; + tensor var_6957 = const()[name = tensor("op_6957"), val = tensor([1, 1])]; + tensor pretrained_out_343_pad_type_0 = const()[name = tensor("pretrained_out_343_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_343_pad_0 = const()[name = tensor("pretrained_out_343_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378220992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379040256))), name = tensor("layers_17_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_343_cast_fp16 = conv(dilations = var_6957, groups = var_6893, pad = pretrained_out_343_pad_0, pad_type = pretrained_out_343_pad_type_0, strides = var_6955, weight = layers_17_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_239_cast_fp16)[name = tensor("pretrained_out_343_cast_fp16")]; + tensor var_6961 = const()[name = tensor("op_6961"), val = tensor([1, 1])]; + tensor var_6963 = const()[name = tensor("op_6963"), val = tensor([1, 1])]; + tensor input_513_pad_type_0 = const()[name = tensor("input_513_pad_type_0"), val = tensor("custom")]; + tensor input_513_pad_0 = const()[name = tensor("input_513_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379040384)))]; + tensor input_513_cast_fp16 = conv(dilations = var_6963, groups = var_6893, pad = input_513_pad_0, pad_type = input_513_pad_type_0, strides = var_6961, weight = layers_17_self_attn_k_proj_loraA_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor var_6967 = const()[name = tensor("op_6967"), val = tensor([1, 1])]; + tensor var_6969 = const()[name = tensor("op_6969"), val = tensor([1, 1])]; + tensor lora_out_685_pad_type_0 = const()[name = tensor("lora_out_685_pad_type_0"), val = tensor("custom")]; + tensor lora_out_685_pad_0 = const()[name = tensor("lora_out_685_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_687_weight_0_to_fp16 = const()[name = tensor("lora_out_687_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379081408)))]; + tensor lora_out_687_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6969, groups = var_6893, pad = lora_out_685_pad_0, pad_type = lora_out_685_pad_type_0, strides = var_6967, weight = lora_out_687_weight_0_to_fp16, x = input_513_cast_fp16)[name = tensor("lora_out_687_cast_fp16")]; + tensor current_key_35_cast_fp16 = add(x = pretrained_out_343_cast_fp16, y = lora_out_687_cast_fp16)[name = tensor("current_key_35_cast_fp16")]; + tensor var_6980 = const()[name = tensor("op_6980"), val = tensor([1, 1])]; + tensor var_6982 = const()[name = tensor("op_6982"), val = tensor([1, 1])]; + tensor pretrained_out_345_pad_type_0 = const()[name = tensor("pretrained_out_345_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_345_pad_0 = const()[name = tensor("pretrained_out_345_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379122432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379941696))), name = tensor("layers_17_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_17_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379941824)))]; + tensor pretrained_out_345_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6982, groups = var_6893, pad = pretrained_out_345_pad_0, pad_type = pretrained_out_345_pad_type_0, strides = var_6980, weight = layers_17_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_239_cast_fp16)[name = tensor("pretrained_out_345_cast_fp16")]; + tensor var_6986 = const()[name = tensor("op_6986"), val = tensor([1, 1])]; + tensor var_6988 = const()[name = tensor("op_6988"), val = tensor([1, 1])]; + tensor input_515_pad_type_0 = const()[name = tensor("input_515_pad_type_0"), val = tensor("custom")]; + tensor input_515_pad_0 = const()[name = tensor("input_515_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379944448)))]; + tensor input_515_cast_fp16 = conv(dilations = var_6988, groups = var_6893, pad = input_515_pad_0, pad_type = input_515_pad_type_0, strides = var_6986, weight = layers_17_self_attn_v_proj_loraA_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("input_515_cast_fp16")]; + tensor var_6992 = const()[name = tensor("op_6992"), val = tensor([1, 1])]; + tensor var_6994 = const()[name = tensor("op_6994"), val = tensor([1, 1])]; + tensor lora_out_689_pad_type_0 = const()[name = tensor("lora_out_689_pad_type_0"), val = tensor("custom")]; + tensor lora_out_689_pad_0 = const()[name = tensor("lora_out_689_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_691_weight_0_to_fp16 = const()[name = tensor("lora_out_691_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(379985472)))]; + tensor lora_out_691_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6994, groups = var_6893, pad = lora_out_689_pad_0, pad_type = lora_out_689_pad_type_0, strides = var_6992, weight = lora_out_691_weight_0_to_fp16, x = input_515_cast_fp16)[name = tensor("lora_out_691_cast_fp16")]; + tensor current_value_35_cast_fp16 = add(x = pretrained_out_345_cast_fp16, y = lora_out_691_cast_fp16)[name = tensor("current_value_35_cast_fp16")]; + tensor var_7004_cast_fp16 = mul(x = current_key_35_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_7004_cast_fp16")]; + tensor var_7006_cast_fp16 = mul(x = var_103_cast_fp16_17, y = var_295_cast_fp16)[name = tensor("op_7006_cast_fp16")]; + tensor key_69_cast_fp16 = add(x = var_7004_cast_fp16, y = var_7006_cast_fp16)[name = tensor("key_69_cast_fp16")]; + tensor var_7008_cast_fp16 = mul(x = current_value_35_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_7008_cast_fp16")]; + tensor var_7010_cast_fp16 = mul(x = var_138_cast_fp16_17, y = var_295_cast_fp16)[name = tensor("op_7010_cast_fp16")]; + tensor value_69_cast_fp16 = add(x = var_7008_cast_fp16, y = var_7010_cast_fp16)[name = tensor("value_69_cast_fp16")]; + tensor var_7013 = const()[name = tensor("op_7013"), val = tensor([1, 20, 64, -1])]; + tensor var_7014_cast_fp16 = reshape(shape = var_7013, x = query_69_cast_fp16)[name = tensor("op_7014_cast_fp16")]; + tensor var_7015_to_fp16 = const()[name = tensor("op_7015_to_fp16"), val = tensor(0x1p-3)]; + tensor var_7016_cast_fp16 = mul(x = var_7014_cast_fp16, y = var_7015_to_fp16)[name = tensor("op_7016_cast_fp16")]; + tensor var_7017 = const()[name = tensor("op_7017"), val = tensor([1, 20, 64, -1])]; + tensor var_7018_cast_fp16 = reshape(shape = var_7017, x = key_69_cast_fp16)[name = tensor("op_7018_cast_fp16")]; + tensor mh_w_103_transpose_x_0 = const()[name = tensor("mh_w_103_transpose_x_0"), val = tensor(true)]; + tensor mh_w_103_transpose_y_0 = const()[name = tensor("mh_w_103_transpose_y_0"), val = tensor(false)]; + tensor mh_w_103_cast_fp16 = matmul(transpose_x = mh_w_103_transpose_x_0, transpose_y = mh_w_103_transpose_y_0, x = var_7016_cast_fp16, y = var_7018_cast_fp16)[name = tensor("mh_w_103_cast_fp16")]; + tensor mh_w_105_cast_fp16 = add(x = mh_w_103_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_105_cast_fp16")]; + tensor var_7026_cast_fp16 = softmax(axis = var_6886, x = mh_w_105_cast_fp16)[name = tensor("op_7026_cast_fp16")]; + tensor var_7027 = const()[name = tensor("op_7027"), val = tensor([1, 20, 64, -1])]; + tensor var_7028_cast_fp16 = reshape(shape = var_7027, x = value_69_cast_fp16)[name = tensor("op_7028_cast_fp16")]; + tensor attn_69_transpose_x_0 = const()[name = tensor("attn_69_transpose_x_0"), val = tensor(false)]; + tensor attn_69_transpose_y_0 = const()[name = tensor("attn_69_transpose_y_0"), val = tensor(true)]; + tensor attn_69_cast_fp16 = matmul(transpose_x = attn_69_transpose_x_0, transpose_y = attn_69_transpose_y_0, x = var_7028_cast_fp16, y = var_7026_cast_fp16)[name = tensor("attn_69_cast_fp16")]; + tensor var_7031 = const()[name = tensor("op_7031"), val = tensor([1, 1280, 1, -1])]; + tensor input_517_cast_fp16 = reshape(shape = var_7031, x = attn_69_cast_fp16)[name = tensor("input_517_cast_fp16")]; + tensor var_7038 = const()[name = tensor("op_7038"), val = tensor([1, 1])]; + tensor var_7040 = const()[name = tensor("op_7040"), val = tensor([1, 1])]; + tensor pretrained_out_347_pad_type_0 = const()[name = tensor("pretrained_out_347_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_347_pad_0 = const()[name = tensor("pretrained_out_347_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380026496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380845760))), name = tensor("layers_17_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_17_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380845888)))]; + tensor pretrained_out_347_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7040, groups = var_6893, pad = pretrained_out_347_pad_0, pad_type = pretrained_out_347_pad_type_0, strides = var_7038, weight = layers_17_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_517_cast_fp16)[name = tensor("pretrained_out_347_cast_fp16")]; + tensor var_7044 = const()[name = tensor("op_7044"), val = tensor([1, 1])]; + tensor var_7046 = const()[name = tensor("op_7046"), val = tensor([1, 1])]; + tensor input_519_pad_type_0 = const()[name = tensor("input_519_pad_type_0"), val = tensor("custom")]; + tensor input_519_pad_0 = const()[name = tensor("input_519_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380848512)))]; + tensor input_519_cast_fp16 = conv(dilations = var_7046, groups = var_6893, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = var_7044, weight = layers_17_self_attn_o_proj_loraA_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("input_519_cast_fp16")]; + tensor var_7050 = const()[name = tensor("op_7050"), val = tensor([1, 1])]; + tensor var_7052 = const()[name = tensor("op_7052"), val = tensor([1, 1])]; + tensor lora_out_693_pad_type_0 = const()[name = tensor("lora_out_693_pad_type_0"), val = tensor("custom")]; + tensor lora_out_693_pad_0 = const()[name = tensor("lora_out_693_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_695_weight_0_to_fp16 = const()[name = tensor("lora_out_695_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380889536)))]; + tensor lora_out_695_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7052, groups = var_6893, pad = lora_out_693_pad_0, pad_type = lora_out_693_pad_type_0, strides = var_7050, weight = lora_out_695_weight_0_to_fp16, x = input_519_cast_fp16)[name = tensor("lora_out_695_cast_fp16")]; + tensor obj_245_cast_fp16 = add(x = pretrained_out_347_cast_fp16, y = lora_out_695_cast_fp16)[name = tensor("obj_245_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = obj_245_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; + tensor var_7065 = const()[name = tensor("op_7065"), val = tensor([1])]; + tensor channels_mean_105_cast_fp16 = reduce_mean(axes = var_7065, keep_dims = var_6894, x = inputs_105_cast_fp16)[name = tensor("channels_mean_105_cast_fp16")]; + tensor zero_mean_105_cast_fp16 = sub(x = inputs_105_cast_fp16, y = channels_mean_105_cast_fp16)[name = tensor("zero_mean_105_cast_fp16")]; + tensor zero_mean_sq_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = zero_mean_105_cast_fp16)[name = tensor("zero_mean_sq_105_cast_fp16")]; + tensor var_7069 = const()[name = tensor("op_7069"), val = tensor([1])]; + tensor var_7070_cast_fp16 = reduce_mean(axes = var_7069, keep_dims = var_6894, x = zero_mean_sq_105_cast_fp16)[name = tensor("op_7070_cast_fp16")]; + tensor var_7071_to_fp16 = const()[name = tensor("op_7071_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7072_cast_fp16 = add(x = var_7070_cast_fp16, y = var_7071_to_fp16)[name = tensor("op_7072_cast_fp16")]; + tensor denom_105_epsilon_0 = const()[name = tensor("denom_105_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_105_cast_fp16 = rsqrt(epsilon = denom_105_epsilon_0, x = var_7072_cast_fp16)[name = tensor("denom_105_cast_fp16")]; + tensor out_105_cast_fp16 = mul(x = zero_mean_105_cast_fp16, y = denom_105_cast_fp16)[name = tensor("out_105_cast_fp16")]; + tensor obj_247_gamma_0_to_fp16 = const()[name = tensor("obj_247_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380930560)))]; + tensor obj_247_beta_0_to_fp16 = const()[name = tensor("obj_247_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380933184)))]; + tensor obj_247_epsilon_0_to_fp16 = const()[name = tensor("obj_247_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_247_cast_fp16 = batch_norm(beta = obj_247_beta_0_to_fp16, epsilon = obj_247_epsilon_0_to_fp16, gamma = obj_247_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor("obj_247_cast_fp16")]; + tensor var_7090 = const()[name = tensor("op_7090"), val = tensor([1, 1])]; + tensor var_7092 = const()[name = tensor("op_7092"), val = tensor([1, 1])]; + tensor pretrained_out_349_pad_type_0 = const()[name = tensor("pretrained_out_349_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_349_pad_0 = const()[name = tensor("pretrained_out_349_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380935808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381755072))), name = tensor("layers_17_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_17_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381755200)))]; + tensor pretrained_out_349_cast_fp16 = conv(bias = layers_17_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_7092, groups = var_6893, pad = pretrained_out_349_pad_0, pad_type = pretrained_out_349_pad_type_0, strides = var_7090, weight = layers_17_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_247_cast_fp16)[name = tensor("pretrained_out_349_cast_fp16")]; + tensor var_7096 = const()[name = tensor("op_7096"), val = tensor([1, 1])]; + tensor var_7098 = const()[name = tensor("op_7098"), val = tensor([1, 1])]; + tensor input_521_pad_type_0 = const()[name = tensor("input_521_pad_type_0"), val = tensor("custom")]; + tensor input_521_pad_0 = const()[name = tensor("input_521_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381757824)))]; + tensor input_521_cast_fp16 = conv(dilations = var_7098, groups = var_6893, pad = input_521_pad_0, pad_type = input_521_pad_type_0, strides = var_7096, weight = layers_17_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_247_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor var_7102 = const()[name = tensor("op_7102"), val = tensor([1, 1])]; + tensor var_7104 = const()[name = tensor("op_7104"), val = tensor([1, 1])]; + tensor lora_out_697_pad_type_0 = const()[name = tensor("lora_out_697_pad_type_0"), val = tensor("custom")]; + tensor lora_out_697_pad_0 = const()[name = tensor("lora_out_697_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_699_weight_0_to_fp16 = const()[name = tensor("lora_out_699_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381798848)))]; + tensor lora_out_699_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7104, groups = var_6893, pad = lora_out_697_pad_0, pad_type = lora_out_697_pad_type_0, strides = var_7102, weight = lora_out_699_weight_0_to_fp16, x = input_521_cast_fp16)[name = tensor("lora_out_699_cast_fp16")]; + tensor query_71_cast_fp16 = add(x = pretrained_out_349_cast_fp16, y = lora_out_699_cast_fp16)[name = tensor("query_71_cast_fp16")]; + tensor var_7114 = const()[name = tensor("op_7114"), val = tensor([1, 1])]; + tensor var_7116 = const()[name = tensor("op_7116"), val = tensor([1, 1])]; + tensor pretrained_out_351_pad_type_0 = const()[name = tensor("pretrained_out_351_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_351_pad_0 = const()[name = tensor("pretrained_out_351_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381839872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(382659136))), name = tensor("layers_17_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_351_cast_fp16 = conv(dilations = var_7116, groups = var_6893, pad = pretrained_out_351_pad_0, pad_type = pretrained_out_351_pad_type_0, strides = var_7114, weight = layers_17_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_351_cast_fp16")]; + tensor var_7120 = const()[name = tensor("op_7120"), val = tensor([1, 1])]; + tensor var_7122 = const()[name = tensor("op_7122"), val = tensor([1, 1])]; + tensor input_523_pad_type_0 = const()[name = tensor("input_523_pad_type_0"), val = tensor("custom")]; + tensor input_523_pad_0 = const()[name = tensor("input_523_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(382659264)))]; + tensor input_523_cast_fp16 = conv(dilations = var_7122, groups = var_6893, pad = input_523_pad_0, pad_type = input_523_pad_type_0, strides = var_7120, weight = layers_17_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_523_cast_fp16")]; + tensor var_7126 = const()[name = tensor("op_7126"), val = tensor([1, 1])]; + tensor var_7128 = const()[name = tensor("op_7128"), val = tensor([1, 1])]; + tensor lora_out_701_pad_type_0 = const()[name = tensor("lora_out_701_pad_type_0"), val = tensor("custom")]; + tensor lora_out_701_pad_0 = const()[name = tensor("lora_out_701_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_703_weight_0_to_fp16 = const()[name = tensor("lora_out_703_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(382700288)))]; + tensor lora_out_703_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7128, groups = var_6893, pad = lora_out_701_pad_0, pad_type = lora_out_701_pad_type_0, strides = var_7126, weight = lora_out_703_weight_0_to_fp16, x = input_523_cast_fp16)[name = tensor("lora_out_703_cast_fp16")]; + tensor key_71_cast_fp16 = add(x = pretrained_out_351_cast_fp16, y = lora_out_703_cast_fp16)[name = tensor("key_71_cast_fp16")]; + tensor var_7139 = const()[name = tensor("op_7139"), val = tensor([1, 1])]; + tensor var_7141 = const()[name = tensor("op_7141"), val = tensor([1, 1])]; + tensor pretrained_out_353_pad_type_0 = const()[name = tensor("pretrained_out_353_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_353_pad_0 = const()[name = tensor("pretrained_out_353_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(382741312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383560576))), name = tensor("layers_17_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_17_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383560704)))]; + tensor pretrained_out_353_cast_fp16 = conv(bias = layers_17_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_7141, groups = var_6893, pad = pretrained_out_353_pad_0, pad_type = pretrained_out_353_pad_type_0, strides = var_7139, weight = layers_17_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_353_cast_fp16")]; + tensor var_7145 = const()[name = tensor("op_7145"), val = tensor([1, 1])]; + tensor var_7147 = const()[name = tensor("op_7147"), val = tensor([1, 1])]; + tensor input_525_pad_type_0 = const()[name = tensor("input_525_pad_type_0"), val = tensor("custom")]; + tensor input_525_pad_0 = const()[name = tensor("input_525_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383563328)))]; + tensor input_525_cast_fp16 = conv(dilations = var_7147, groups = var_6893, pad = input_525_pad_0, pad_type = input_525_pad_type_0, strides = var_7145, weight = layers_17_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_525_cast_fp16")]; + tensor var_7151 = const()[name = tensor("op_7151"), val = tensor([1, 1])]; + tensor var_7153 = const()[name = tensor("op_7153"), val = tensor([1, 1])]; + tensor lora_out_705_pad_type_0 = const()[name = tensor("lora_out_705_pad_type_0"), val = tensor("custom")]; + tensor lora_out_705_pad_0 = const()[name = tensor("lora_out_705_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_707_weight_0_to_fp16 = const()[name = tensor("lora_out_707_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383604352)))]; + tensor lora_out_707_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7153, groups = var_6893, pad = lora_out_705_pad_0, pad_type = lora_out_705_pad_type_0, strides = var_7151, weight = lora_out_707_weight_0_to_fp16, x = input_525_cast_fp16)[name = tensor("lora_out_707_cast_fp16")]; + tensor value_71_cast_fp16 = add(x = pretrained_out_353_cast_fp16, y = lora_out_707_cast_fp16)[name = tensor("value_71_cast_fp16")]; + tensor var_7160 = const()[name = tensor("op_7160"), val = tensor([1, 20, 64, -1])]; + tensor var_7161_cast_fp16 = reshape(shape = var_7160, x = query_71_cast_fp16)[name = tensor("op_7161_cast_fp16")]; + tensor var_7162_to_fp16 = const()[name = tensor("op_7162_to_fp16"), val = tensor(0x1p-3)]; + tensor var_7163_cast_fp16 = mul(x = var_7161_cast_fp16, y = var_7162_to_fp16)[name = tensor("op_7163_cast_fp16")]; + tensor var_7164 = const()[name = tensor("op_7164"), val = tensor([1, 20, 64, -1])]; + tensor var_7165_cast_fp16 = reshape(shape = var_7164, x = key_71_cast_fp16)[name = tensor("op_7165_cast_fp16")]; + tensor mh_w_107_transpose_x_0 = const()[name = tensor("mh_w_107_transpose_x_0"), val = tensor(true)]; + tensor mh_w_107_transpose_y_0 = const()[name = tensor("mh_w_107_transpose_y_0"), val = tensor(false)]; + tensor mh_w_107_cast_fp16 = matmul(transpose_x = mh_w_107_transpose_x_0, transpose_y = mh_w_107_transpose_y_0, x = var_7163_cast_fp16, y = var_7165_cast_fp16)[name = tensor("mh_w_107_cast_fp16")]; + tensor obj_251_cast_fp16 = softmax(axis = var_6886, x = mh_w_107_cast_fp16)[name = tensor("obj_251_cast_fp16")]; + tensor var_7169 = const()[name = tensor("op_7169"), val = tensor([1, 20, 64, -1])]; + tensor var_7170_cast_fp16 = reshape(shape = var_7169, x = value_71_cast_fp16)[name = tensor("op_7170_cast_fp16")]; + tensor attn_71_transpose_x_0 = const()[name = tensor("attn_71_transpose_x_0"), val = tensor(false)]; + tensor attn_71_transpose_y_0 = const()[name = tensor("attn_71_transpose_y_0"), val = tensor(true)]; + tensor attn_71_cast_fp16 = matmul(transpose_x = attn_71_transpose_x_0, transpose_y = attn_71_transpose_y_0, x = var_7170_cast_fp16, y = obj_251_cast_fp16)[name = tensor("attn_71_cast_fp16")]; + tensor var_7173 = const()[name = tensor("op_7173"), val = tensor([1, 1280, 1, -1])]; + tensor input_527_cast_fp16 = reshape(shape = var_7173, x = attn_71_cast_fp16)[name = tensor("input_527_cast_fp16")]; + tensor var_7180 = const()[name = tensor("op_7180"), val = tensor([1, 1])]; + tensor var_7182 = const()[name = tensor("op_7182"), val = tensor([1, 1])]; + tensor pretrained_out_355_pad_type_0 = const()[name = tensor("pretrained_out_355_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_355_pad_0 = const()[name = tensor("pretrained_out_355_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383645376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384464640))), name = tensor("layers_17_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_17_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384464768)))]; + tensor pretrained_out_355_cast_fp16 = conv(bias = layers_17_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7182, groups = var_6893, pad = pretrained_out_355_pad_0, pad_type = pretrained_out_355_pad_type_0, strides = var_7180, weight = layers_17_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_527_cast_fp16)[name = tensor("pretrained_out_355_cast_fp16")]; + tensor var_7186 = const()[name = tensor("op_7186"), val = tensor([1, 1])]; + tensor var_7188 = const()[name = tensor("op_7188"), val = tensor([1, 1])]; + tensor input_529_pad_type_0 = const()[name = tensor("input_529_pad_type_0"), val = tensor("custom")]; + tensor input_529_pad_0 = const()[name = tensor("input_529_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384467392)))]; + tensor input_529_cast_fp16 = conv(dilations = var_7188, groups = var_6893, pad = input_529_pad_0, pad_type = input_529_pad_type_0, strides = var_7186, weight = layers_17_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_527_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor var_7192 = const()[name = tensor("op_7192"), val = tensor([1, 1])]; + tensor var_7194 = const()[name = tensor("op_7194"), val = tensor([1, 1])]; + tensor lora_out_709_pad_type_0 = const()[name = tensor("lora_out_709_pad_type_0"), val = tensor("custom")]; + tensor lora_out_709_pad_0 = const()[name = tensor("lora_out_709_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_711_weight_0_to_fp16 = const()[name = tensor("lora_out_711_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384508416)))]; + tensor lora_out_711_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7194, groups = var_6893, pad = lora_out_709_pad_0, pad_type = lora_out_709_pad_type_0, strides = var_7192, weight = lora_out_711_weight_0_to_fp16, x = input_529_cast_fp16)[name = tensor("lora_out_711_cast_fp16")]; + tensor obj_249_cast_fp16 = add(x = pretrained_out_355_cast_fp16, y = lora_out_711_cast_fp16)[name = tensor("obj_249_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_249_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; + tensor var_7206 = const()[name = tensor("op_7206"), val = tensor([1])]; + tensor channels_mean_107_cast_fp16 = reduce_mean(axes = var_7206, keep_dims = var_6894, x = inputs_107_cast_fp16)[name = tensor("channels_mean_107_cast_fp16")]; + tensor zero_mean_107_cast_fp16 = sub(x = inputs_107_cast_fp16, y = channels_mean_107_cast_fp16)[name = tensor("zero_mean_107_cast_fp16")]; + tensor zero_mean_sq_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = zero_mean_107_cast_fp16)[name = tensor("zero_mean_sq_107_cast_fp16")]; + tensor var_7210 = const()[name = tensor("op_7210"), val = tensor([1])]; + tensor var_7211_cast_fp16 = reduce_mean(axes = var_7210, keep_dims = var_6894, x = zero_mean_sq_107_cast_fp16)[name = tensor("op_7211_cast_fp16")]; + tensor var_7212_to_fp16 = const()[name = tensor("op_7212_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7213_cast_fp16 = add(x = var_7211_cast_fp16, y = var_7212_to_fp16)[name = tensor("op_7213_cast_fp16")]; + tensor denom_107_epsilon_0 = const()[name = tensor("denom_107_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_107_cast_fp16 = rsqrt(epsilon = denom_107_epsilon_0, x = var_7213_cast_fp16)[name = tensor("denom_107_cast_fp16")]; + tensor out_107_cast_fp16 = mul(x = zero_mean_107_cast_fp16, y = denom_107_cast_fp16)[name = tensor("out_107_cast_fp16")]; + tensor input_531_gamma_0_to_fp16 = const()[name = tensor("input_531_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384549440)))]; + tensor input_531_beta_0_to_fp16 = const()[name = tensor("input_531_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384552064)))]; + tensor input_531_epsilon_0_to_fp16 = const()[name = tensor("input_531_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_531_cast_fp16 = batch_norm(beta = input_531_beta_0_to_fp16, epsilon = input_531_epsilon_0_to_fp16, gamma = input_531_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor("input_531_cast_fp16")]; + tensor var_7227 = const()[name = tensor("op_7227"), val = tensor([1, 1])]; + tensor var_7229 = const()[name = tensor("op_7229"), val = tensor([1, 1])]; + tensor pretrained_out_357_pad_type_0 = const()[name = tensor("pretrained_out_357_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_357_pad_0 = const()[name = tensor("pretrained_out_357_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384554688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387831552))), name = tensor("layers_17_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_17_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387831680)))]; + tensor pretrained_out_357_cast_fp16 = conv(bias = layers_17_fc1_pretrained_bias_to_fp16, dilations = var_7229, groups = var_6893, pad = pretrained_out_357_pad_0, pad_type = pretrained_out_357_pad_type_0, strides = var_7227, weight = layers_17_fc1_pretrained_weight_to_fp16_palettized, x = input_531_cast_fp16)[name = tensor("pretrained_out_357_cast_fp16")]; + tensor var_7233 = const()[name = tensor("op_7233"), val = tensor([1, 1])]; + tensor var_7235 = const()[name = tensor("op_7235"), val = tensor([1, 1])]; + tensor input_533_pad_type_0 = const()[name = tensor("input_533_pad_type_0"), val = tensor("custom")]; + tensor input_533_pad_0 = const()[name = tensor("input_533_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_17_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387841984)))]; + tensor input_533_cast_fp16 = conv(dilations = var_7235, groups = var_6893, pad = input_533_pad_0, pad_type = input_533_pad_type_0, strides = var_7233, weight = layers_17_fc1_loraA_weight_to_fp16, x = input_531_cast_fp16)[name = tensor("input_533_cast_fp16")]; + tensor var_7239 = const()[name = tensor("op_7239"), val = tensor([1, 1])]; + tensor var_7241 = const()[name = tensor("op_7241"), val = tensor([1, 1])]; + tensor lora_out_713_pad_type_0 = const()[name = tensor("lora_out_713_pad_type_0"), val = tensor("custom")]; + tensor lora_out_713_pad_0 = const()[name = tensor("lora_out_713_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_715_weight_0_to_fp16 = const()[name = tensor("lora_out_715_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387883008)))]; + tensor lora_out_715_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_7241, groups = var_6893, pad = lora_out_713_pad_0, pad_type = lora_out_713_pad_type_0, strides = var_7239, weight = lora_out_715_weight_0_to_fp16, x = input_533_cast_fp16)[name = tensor("lora_out_715_cast_fp16")]; + tensor input_535_cast_fp16 = add(x = pretrained_out_357_cast_fp16, y = lora_out_715_cast_fp16)[name = tensor("input_535_cast_fp16")]; + tensor input_537_mode_0 = const()[name = tensor("input_537_mode_0"), val = tensor("EXACT")]; + tensor input_537_cast_fp16 = gelu(mode = input_537_mode_0, x = input_535_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor var_7253 = const()[name = tensor("op_7253"), val = tensor([1, 1])]; + tensor var_7255 = const()[name = tensor("op_7255"), val = tensor([1, 1])]; + tensor pretrained_out_359_pad_type_0 = const()[name = tensor("pretrained_out_359_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_359_pad_0 = const()[name = tensor("pretrained_out_359_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388046912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391323776))), name = tensor("layers_17_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_17_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_17_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391323904)))]; + tensor pretrained_out_359_cast_fp16 = conv(bias = layers_17_fc2_pretrained_bias_to_fp16, dilations = var_7255, groups = var_6893, pad = pretrained_out_359_pad_0, pad_type = pretrained_out_359_pad_type_0, strides = var_7253, weight = layers_17_fc2_pretrained_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = tensor("pretrained_out_359_cast_fp16")]; + tensor var_7259 = const()[name = tensor("op_7259"), val = tensor([1, 1])]; + tensor var_7261 = const()[name = tensor("op_7261"), val = tensor([1, 1])]; + tensor input_539_pad_type_0 = const()[name = tensor("input_539_pad_type_0"), val = tensor("custom")]; + tensor input_539_pad_0 = const()[name = tensor("input_539_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_17_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_17_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391326528)))]; + tensor input_539_cast_fp16 = conv(dilations = var_7261, groups = var_6893, pad = input_539_pad_0, pad_type = input_539_pad_type_0, strides = var_7259, weight = layers_17_fc2_loraA_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("input_539_cast_fp16")]; + tensor var_7265 = const()[name = tensor("op_7265"), val = tensor([1, 1])]; + tensor var_7267 = const()[name = tensor("op_7267"), val = tensor([1, 1])]; + tensor lora_out_717_pad_type_0 = const()[name = tensor("lora_out_717_pad_type_0"), val = tensor("custom")]; + tensor lora_out_717_pad_0 = const()[name = tensor("lora_out_717_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_719_weight_0_to_fp16 = const()[name = tensor("lora_out_719_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391490432)))]; + tensor lora_out_719_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7267, groups = var_6893, pad = lora_out_717_pad_0, pad_type = lora_out_717_pad_type_0, strides = var_7265, weight = lora_out_719_weight_0_to_fp16, x = input_539_cast_fp16)[name = tensor("lora_out_719_cast_fp16")]; + tensor hidden_states_37_cast_fp16 = add(x = pretrained_out_359_cast_fp16, y = lora_out_719_cast_fp16)[name = tensor("hidden_states_37_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; + tensor var_7284 = const()[name = tensor("op_7284"), val = tensor(3)]; + tensor var_7291 = const()[name = tensor("op_7291"), val = tensor(1)]; + tensor var_7292 = const()[name = tensor("op_7292"), val = tensor(true)]; + tensor var_7304 = const()[name = tensor("op_7304"), val = tensor([1])]; + tensor channels_mean_109_cast_fp16 = reduce_mean(axes = var_7304, keep_dims = var_7292, x = inputs_109_cast_fp16)[name = tensor("channels_mean_109_cast_fp16")]; + tensor zero_mean_109_cast_fp16 = sub(x = inputs_109_cast_fp16, y = channels_mean_109_cast_fp16)[name = tensor("zero_mean_109_cast_fp16")]; + tensor zero_mean_sq_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = zero_mean_109_cast_fp16)[name = tensor("zero_mean_sq_109_cast_fp16")]; + tensor var_7308 = const()[name = tensor("op_7308"), val = tensor([1])]; + tensor var_7309_cast_fp16 = reduce_mean(axes = var_7308, keep_dims = var_7292, x = zero_mean_sq_109_cast_fp16)[name = tensor("op_7309_cast_fp16")]; + tensor var_7310_to_fp16 = const()[name = tensor("op_7310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7311_cast_fp16 = add(x = var_7309_cast_fp16, y = var_7310_to_fp16)[name = tensor("op_7311_cast_fp16")]; + tensor denom_109_epsilon_0 = const()[name = tensor("denom_109_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_109_cast_fp16 = rsqrt(epsilon = denom_109_epsilon_0, x = var_7311_cast_fp16)[name = tensor("denom_109_cast_fp16")]; + tensor out_109_cast_fp16 = mul(x = zero_mean_109_cast_fp16, y = denom_109_cast_fp16)[name = tensor("out_109_cast_fp16")]; + tensor obj_253_gamma_0_to_fp16 = const()[name = tensor("obj_253_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391531456)))]; + tensor obj_253_beta_0_to_fp16 = const()[name = tensor("obj_253_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391534080)))]; + tensor obj_253_epsilon_0_to_fp16 = const()[name = tensor("obj_253_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_253_cast_fp16 = batch_norm(beta = obj_253_beta_0_to_fp16, epsilon = obj_253_epsilon_0_to_fp16, gamma = obj_253_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor("obj_253_cast_fp16")]; + tensor var_7329 = const()[name = tensor("op_7329"), val = tensor([1, 1])]; + tensor var_7331 = const()[name = tensor("op_7331"), val = tensor([1, 1])]; + tensor pretrained_out_361_pad_type_0 = const()[name = tensor("pretrained_out_361_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_361_pad_0 = const()[name = tensor("pretrained_out_361_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391536704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392355968))), name = tensor("layers_18_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_18_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392356096)))]; + tensor pretrained_out_361_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_7331, groups = var_7291, pad = pretrained_out_361_pad_0, pad_type = pretrained_out_361_pad_type_0, strides = var_7329, weight = layers_18_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_253_cast_fp16)[name = tensor("pretrained_out_361_cast_fp16")]; + tensor var_7335 = const()[name = tensor("op_7335"), val = tensor([1, 1])]; + tensor var_7337 = const()[name = tensor("op_7337"), val = tensor([1, 1])]; + tensor input_541_pad_type_0 = const()[name = tensor("input_541_pad_type_0"), val = tensor("custom")]; + tensor input_541_pad_0 = const()[name = tensor("input_541_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392358720)))]; + tensor input_541_cast_fp16 = conv(dilations = var_7337, groups = var_7291, pad = input_541_pad_0, pad_type = input_541_pad_type_0, strides = var_7335, weight = layers_18_self_attn_q_proj_loraA_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("input_541_cast_fp16")]; + tensor var_7341 = const()[name = tensor("op_7341"), val = tensor([1, 1])]; + tensor var_7343 = const()[name = tensor("op_7343"), val = tensor([1, 1])]; + tensor lora_out_721_pad_type_0 = const()[name = tensor("lora_out_721_pad_type_0"), val = tensor("custom")]; + tensor lora_out_721_pad_0 = const()[name = tensor("lora_out_721_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_723_weight_0_to_fp16 = const()[name = tensor("lora_out_723_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392399744)))]; + tensor lora_out_723_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7343, groups = var_7291, pad = lora_out_721_pad_0, pad_type = lora_out_721_pad_type_0, strides = var_7341, weight = lora_out_723_weight_0_to_fp16, x = input_541_cast_fp16)[name = tensor("lora_out_723_cast_fp16")]; + tensor query_73_cast_fp16 = add(x = pretrained_out_361_cast_fp16, y = lora_out_723_cast_fp16)[name = tensor("query_73_cast_fp16")]; + tensor var_7353 = const()[name = tensor("op_7353"), val = tensor([1, 1])]; + tensor var_7355 = const()[name = tensor("op_7355"), val = tensor([1, 1])]; + tensor pretrained_out_363_pad_type_0 = const()[name = tensor("pretrained_out_363_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_363_pad_0 = const()[name = tensor("pretrained_out_363_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392440768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393260032))), name = tensor("layers_18_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_363_cast_fp16 = conv(dilations = var_7355, groups = var_7291, pad = pretrained_out_363_pad_0, pad_type = pretrained_out_363_pad_type_0, strides = var_7353, weight = layers_18_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_253_cast_fp16)[name = tensor("pretrained_out_363_cast_fp16")]; + tensor var_7359 = const()[name = tensor("op_7359"), val = tensor([1, 1])]; + tensor var_7361 = const()[name = tensor("op_7361"), val = tensor([1, 1])]; + tensor input_543_pad_type_0 = const()[name = tensor("input_543_pad_type_0"), val = tensor("custom")]; + tensor input_543_pad_0 = const()[name = tensor("input_543_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393260160)))]; + tensor input_543_cast_fp16 = conv(dilations = var_7361, groups = var_7291, pad = input_543_pad_0, pad_type = input_543_pad_type_0, strides = var_7359, weight = layers_18_self_attn_k_proj_loraA_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("input_543_cast_fp16")]; + tensor var_7365 = const()[name = tensor("op_7365"), val = tensor([1, 1])]; + tensor var_7367 = const()[name = tensor("op_7367"), val = tensor([1, 1])]; + tensor lora_out_725_pad_type_0 = const()[name = tensor("lora_out_725_pad_type_0"), val = tensor("custom")]; + tensor lora_out_725_pad_0 = const()[name = tensor("lora_out_725_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_727_weight_0_to_fp16 = const()[name = tensor("lora_out_727_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393301184)))]; + tensor lora_out_727_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7367, groups = var_7291, pad = lora_out_725_pad_0, pad_type = lora_out_725_pad_type_0, strides = var_7365, weight = lora_out_727_weight_0_to_fp16, x = input_543_cast_fp16)[name = tensor("lora_out_727_cast_fp16")]; + tensor current_key_37_cast_fp16 = add(x = pretrained_out_363_cast_fp16, y = lora_out_727_cast_fp16)[name = tensor("current_key_37_cast_fp16")]; + tensor var_7378 = const()[name = tensor("op_7378"), val = tensor([1, 1])]; + tensor var_7380 = const()[name = tensor("op_7380"), val = tensor([1, 1])]; + tensor pretrained_out_365_pad_type_0 = const()[name = tensor("pretrained_out_365_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_365_pad_0 = const()[name = tensor("pretrained_out_365_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393342208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394161472))), name = tensor("layers_18_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_18_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394161600)))]; + tensor pretrained_out_365_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_7380, groups = var_7291, pad = pretrained_out_365_pad_0, pad_type = pretrained_out_365_pad_type_0, strides = var_7378, weight = layers_18_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_253_cast_fp16)[name = tensor("pretrained_out_365_cast_fp16")]; + tensor var_7384 = const()[name = tensor("op_7384"), val = tensor([1, 1])]; + tensor var_7386 = const()[name = tensor("op_7386"), val = tensor([1, 1])]; + tensor input_545_pad_type_0 = const()[name = tensor("input_545_pad_type_0"), val = tensor("custom")]; + tensor input_545_pad_0 = const()[name = tensor("input_545_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394164224)))]; + tensor input_545_cast_fp16 = conv(dilations = var_7386, groups = var_7291, pad = input_545_pad_0, pad_type = input_545_pad_type_0, strides = var_7384, weight = layers_18_self_attn_v_proj_loraA_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("input_545_cast_fp16")]; + tensor var_7390 = const()[name = tensor("op_7390"), val = tensor([1, 1])]; + tensor var_7392 = const()[name = tensor("op_7392"), val = tensor([1, 1])]; + tensor lora_out_729_pad_type_0 = const()[name = tensor("lora_out_729_pad_type_0"), val = tensor("custom")]; + tensor lora_out_729_pad_0 = const()[name = tensor("lora_out_729_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_731_weight_0_to_fp16 = const()[name = tensor("lora_out_731_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394205248)))]; + tensor lora_out_731_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7392, groups = var_7291, pad = lora_out_729_pad_0, pad_type = lora_out_729_pad_type_0, strides = var_7390, weight = lora_out_731_weight_0_to_fp16, x = input_545_cast_fp16)[name = tensor("lora_out_731_cast_fp16")]; + tensor current_value_37_cast_fp16 = add(x = pretrained_out_365_cast_fp16, y = lora_out_731_cast_fp16)[name = tensor("current_value_37_cast_fp16")]; + tensor var_7402_cast_fp16 = mul(x = current_key_37_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_7402_cast_fp16")]; + tensor var_7404_cast_fp16 = mul(x = var_103_cast_fp16_18, y = var_295_cast_fp16)[name = tensor("op_7404_cast_fp16")]; + tensor key_73_cast_fp16 = add(x = var_7402_cast_fp16, y = var_7404_cast_fp16)[name = tensor("key_73_cast_fp16")]; + tensor var_7406_cast_fp16 = mul(x = current_value_37_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_7406_cast_fp16")]; + tensor var_7408_cast_fp16 = mul(x = var_138_cast_fp16_18, y = var_295_cast_fp16)[name = tensor("op_7408_cast_fp16")]; + tensor value_73_cast_fp16 = add(x = var_7406_cast_fp16, y = var_7408_cast_fp16)[name = tensor("value_73_cast_fp16")]; + tensor var_7411 = const()[name = tensor("op_7411"), val = tensor([1, 20, 64, -1])]; + tensor var_7412_cast_fp16 = reshape(shape = var_7411, x = query_73_cast_fp16)[name = tensor("op_7412_cast_fp16")]; + tensor var_7413_to_fp16 = const()[name = tensor("op_7413_to_fp16"), val = tensor(0x1p-3)]; + tensor var_7414_cast_fp16 = mul(x = var_7412_cast_fp16, y = var_7413_to_fp16)[name = tensor("op_7414_cast_fp16")]; + tensor var_7415 = const()[name = tensor("op_7415"), val = tensor([1, 20, 64, -1])]; + tensor var_7416_cast_fp16 = reshape(shape = var_7415, x = key_73_cast_fp16)[name = tensor("op_7416_cast_fp16")]; + tensor mh_w_109_transpose_x_0 = const()[name = tensor("mh_w_109_transpose_x_0"), val = tensor(true)]; + tensor mh_w_109_transpose_y_0 = const()[name = tensor("mh_w_109_transpose_y_0"), val = tensor(false)]; + tensor mh_w_109_cast_fp16 = matmul(transpose_x = mh_w_109_transpose_x_0, transpose_y = mh_w_109_transpose_y_0, x = var_7414_cast_fp16, y = var_7416_cast_fp16)[name = tensor("mh_w_109_cast_fp16")]; + tensor mh_w_111_cast_fp16 = add(x = mh_w_109_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_111_cast_fp16")]; + tensor var_7424_cast_fp16 = softmax(axis = var_7284, x = mh_w_111_cast_fp16)[name = tensor("op_7424_cast_fp16")]; + tensor var_7425 = const()[name = tensor("op_7425"), val = tensor([1, 20, 64, -1])]; + tensor var_7426_cast_fp16 = reshape(shape = var_7425, x = value_73_cast_fp16)[name = tensor("op_7426_cast_fp16")]; + tensor attn_73_transpose_x_0 = const()[name = tensor("attn_73_transpose_x_0"), val = tensor(false)]; + tensor attn_73_transpose_y_0 = const()[name = tensor("attn_73_transpose_y_0"), val = tensor(true)]; + tensor attn_73_cast_fp16 = matmul(transpose_x = attn_73_transpose_x_0, transpose_y = attn_73_transpose_y_0, x = var_7426_cast_fp16, y = var_7424_cast_fp16)[name = tensor("attn_73_cast_fp16")]; + tensor var_7429 = const()[name = tensor("op_7429"), val = tensor([1, 1280, 1, -1])]; + tensor input_547_cast_fp16 = reshape(shape = var_7429, x = attn_73_cast_fp16)[name = tensor("input_547_cast_fp16")]; + tensor var_7436 = const()[name = tensor("op_7436"), val = tensor([1, 1])]; + tensor var_7438 = const()[name = tensor("op_7438"), val = tensor([1, 1])]; + tensor pretrained_out_367_pad_type_0 = const()[name = tensor("pretrained_out_367_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_367_pad_0 = const()[name = tensor("pretrained_out_367_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394246272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395065536))), name = tensor("layers_18_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_18_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395065664)))]; + tensor pretrained_out_367_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7438, groups = var_7291, pad = pretrained_out_367_pad_0, pad_type = pretrained_out_367_pad_type_0, strides = var_7436, weight = layers_18_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = tensor("pretrained_out_367_cast_fp16")]; + tensor var_7442 = const()[name = tensor("op_7442"), val = tensor([1, 1])]; + tensor var_7444 = const()[name = tensor("op_7444"), val = tensor([1, 1])]; + tensor input_549_pad_type_0 = const()[name = tensor("input_549_pad_type_0"), val = tensor("custom")]; + tensor input_549_pad_0 = const()[name = tensor("input_549_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395068288)))]; + tensor input_549_cast_fp16 = conv(dilations = var_7444, groups = var_7291, pad = input_549_pad_0, pad_type = input_549_pad_type_0, strides = var_7442, weight = layers_18_self_attn_o_proj_loraA_weight_to_fp16, x = input_547_cast_fp16)[name = tensor("input_549_cast_fp16")]; + tensor var_7448 = const()[name = tensor("op_7448"), val = tensor([1, 1])]; + tensor var_7450 = const()[name = tensor("op_7450"), val = tensor([1, 1])]; + tensor lora_out_733_pad_type_0 = const()[name = tensor("lora_out_733_pad_type_0"), val = tensor("custom")]; + tensor lora_out_733_pad_0 = const()[name = tensor("lora_out_733_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_735_weight_0_to_fp16 = const()[name = tensor("lora_out_735_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395109312)))]; + tensor lora_out_735_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7450, groups = var_7291, pad = lora_out_733_pad_0, pad_type = lora_out_733_pad_type_0, strides = var_7448, weight = lora_out_735_weight_0_to_fp16, x = input_549_cast_fp16)[name = tensor("lora_out_735_cast_fp16")]; + tensor obj_259_cast_fp16 = add(x = pretrained_out_367_cast_fp16, y = lora_out_735_cast_fp16)[name = tensor("obj_259_cast_fp16")]; + tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_259_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor var_7463 = const()[name = tensor("op_7463"), val = tensor([1])]; + tensor channels_mean_111_cast_fp16 = reduce_mean(axes = var_7463, keep_dims = var_7292, x = inputs_111_cast_fp16)[name = tensor("channels_mean_111_cast_fp16")]; + tensor zero_mean_111_cast_fp16 = sub(x = inputs_111_cast_fp16, y = channels_mean_111_cast_fp16)[name = tensor("zero_mean_111_cast_fp16")]; + tensor zero_mean_sq_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = zero_mean_111_cast_fp16)[name = tensor("zero_mean_sq_111_cast_fp16")]; + tensor var_7467 = const()[name = tensor("op_7467"), val = tensor([1])]; + tensor var_7468_cast_fp16 = reduce_mean(axes = var_7467, keep_dims = var_7292, x = zero_mean_sq_111_cast_fp16)[name = tensor("op_7468_cast_fp16")]; + tensor var_7469_to_fp16 = const()[name = tensor("op_7469_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7470_cast_fp16 = add(x = var_7468_cast_fp16, y = var_7469_to_fp16)[name = tensor("op_7470_cast_fp16")]; + tensor denom_111_epsilon_0 = const()[name = tensor("denom_111_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_111_cast_fp16 = rsqrt(epsilon = denom_111_epsilon_0, x = var_7470_cast_fp16)[name = tensor("denom_111_cast_fp16")]; + tensor out_111_cast_fp16 = mul(x = zero_mean_111_cast_fp16, y = denom_111_cast_fp16)[name = tensor("out_111_cast_fp16")]; + tensor obj_261_gamma_0_to_fp16 = const()[name = tensor("obj_261_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395150336)))]; + tensor obj_261_beta_0_to_fp16 = const()[name = tensor("obj_261_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395152960)))]; + tensor obj_261_epsilon_0_to_fp16 = const()[name = tensor("obj_261_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_261_cast_fp16 = batch_norm(beta = obj_261_beta_0_to_fp16, epsilon = obj_261_epsilon_0_to_fp16, gamma = obj_261_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor("obj_261_cast_fp16")]; + tensor var_7488 = const()[name = tensor("op_7488"), val = tensor([1, 1])]; + tensor var_7490 = const()[name = tensor("op_7490"), val = tensor([1, 1])]; + tensor pretrained_out_369_pad_type_0 = const()[name = tensor("pretrained_out_369_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_369_pad_0 = const()[name = tensor("pretrained_out_369_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395155584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395974848))), name = tensor("layers_18_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_18_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395974976)))]; + tensor pretrained_out_369_cast_fp16 = conv(bias = layers_18_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_7490, groups = var_7291, pad = pretrained_out_369_pad_0, pad_type = pretrained_out_369_pad_type_0, strides = var_7488, weight = layers_18_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_261_cast_fp16)[name = tensor("pretrained_out_369_cast_fp16")]; + tensor var_7494 = const()[name = tensor("op_7494"), val = tensor([1, 1])]; + tensor var_7496 = const()[name = tensor("op_7496"), val = tensor([1, 1])]; + tensor input_551_pad_type_0 = const()[name = tensor("input_551_pad_type_0"), val = tensor("custom")]; + tensor input_551_pad_0 = const()[name = tensor("input_551_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395977600)))]; + tensor input_551_cast_fp16 = conv(dilations = var_7496, groups = var_7291, pad = input_551_pad_0, pad_type = input_551_pad_type_0, strides = var_7494, weight = layers_18_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_261_cast_fp16)[name = tensor("input_551_cast_fp16")]; + tensor var_7500 = const()[name = tensor("op_7500"), val = tensor([1, 1])]; + tensor var_7502 = const()[name = tensor("op_7502"), val = tensor([1, 1])]; + tensor lora_out_737_pad_type_0 = const()[name = tensor("lora_out_737_pad_type_0"), val = tensor("custom")]; + tensor lora_out_737_pad_0 = const()[name = tensor("lora_out_737_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_739_weight_0_to_fp16 = const()[name = tensor("lora_out_739_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396018624)))]; + tensor lora_out_739_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7502, groups = var_7291, pad = lora_out_737_pad_0, pad_type = lora_out_737_pad_type_0, strides = var_7500, weight = lora_out_739_weight_0_to_fp16, x = input_551_cast_fp16)[name = tensor("lora_out_739_cast_fp16")]; + tensor query_75_cast_fp16 = add(x = pretrained_out_369_cast_fp16, y = lora_out_739_cast_fp16)[name = tensor("query_75_cast_fp16")]; + tensor var_7512 = const()[name = tensor("op_7512"), val = tensor([1, 1])]; + tensor var_7514 = const()[name = tensor("op_7514"), val = tensor([1, 1])]; + tensor pretrained_out_371_pad_type_0 = const()[name = tensor("pretrained_out_371_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_371_pad_0 = const()[name = tensor("pretrained_out_371_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396059648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396878912))), name = tensor("layers_18_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_371_cast_fp16 = conv(dilations = var_7514, groups = var_7291, pad = pretrained_out_371_pad_0, pad_type = pretrained_out_371_pad_type_0, strides = var_7512, weight = layers_18_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_371_cast_fp16")]; + tensor var_7518 = const()[name = tensor("op_7518"), val = tensor([1, 1])]; + tensor var_7520 = const()[name = tensor("op_7520"), val = tensor([1, 1])]; + tensor input_553_pad_type_0 = const()[name = tensor("input_553_pad_type_0"), val = tensor("custom")]; + tensor input_553_pad_0 = const()[name = tensor("input_553_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396879040)))]; + tensor input_553_cast_fp16 = conv(dilations = var_7520, groups = var_7291, pad = input_553_pad_0, pad_type = input_553_pad_type_0, strides = var_7518, weight = layers_18_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_553_cast_fp16")]; + tensor var_7524 = const()[name = tensor("op_7524"), val = tensor([1, 1])]; + tensor var_7526 = const()[name = tensor("op_7526"), val = tensor([1, 1])]; + tensor lora_out_741_pad_type_0 = const()[name = tensor("lora_out_741_pad_type_0"), val = tensor("custom")]; + tensor lora_out_741_pad_0 = const()[name = tensor("lora_out_741_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_743_weight_0_to_fp16 = const()[name = tensor("lora_out_743_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396920064)))]; + tensor lora_out_743_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7526, groups = var_7291, pad = lora_out_741_pad_0, pad_type = lora_out_741_pad_type_0, strides = var_7524, weight = lora_out_743_weight_0_to_fp16, x = input_553_cast_fp16)[name = tensor("lora_out_743_cast_fp16")]; + tensor key_75_cast_fp16 = add(x = pretrained_out_371_cast_fp16, y = lora_out_743_cast_fp16)[name = tensor("key_75_cast_fp16")]; + tensor var_7537 = const()[name = tensor("op_7537"), val = tensor([1, 1])]; + tensor var_7539 = const()[name = tensor("op_7539"), val = tensor([1, 1])]; + tensor pretrained_out_373_pad_type_0 = const()[name = tensor("pretrained_out_373_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_373_pad_0 = const()[name = tensor("pretrained_out_373_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(396961088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(397780352))), name = tensor("layers_18_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_18_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(397780480)))]; + tensor pretrained_out_373_cast_fp16 = conv(bias = layers_18_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_7539, groups = var_7291, pad = pretrained_out_373_pad_0, pad_type = pretrained_out_373_pad_type_0, strides = var_7537, weight = layers_18_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_373_cast_fp16")]; + tensor var_7543 = const()[name = tensor("op_7543"), val = tensor([1, 1])]; + tensor var_7545 = const()[name = tensor("op_7545"), val = tensor([1, 1])]; + tensor input_555_pad_type_0 = const()[name = tensor("input_555_pad_type_0"), val = tensor("custom")]; + tensor input_555_pad_0 = const()[name = tensor("input_555_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(397783104)))]; + tensor input_555_cast_fp16 = conv(dilations = var_7545, groups = var_7291, pad = input_555_pad_0, pad_type = input_555_pad_type_0, strides = var_7543, weight = layers_18_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_555_cast_fp16")]; + tensor var_7549 = const()[name = tensor("op_7549"), val = tensor([1, 1])]; + tensor var_7551 = const()[name = tensor("op_7551"), val = tensor([1, 1])]; + tensor lora_out_745_pad_type_0 = const()[name = tensor("lora_out_745_pad_type_0"), val = tensor("custom")]; + tensor lora_out_745_pad_0 = const()[name = tensor("lora_out_745_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_747_weight_0_to_fp16 = const()[name = tensor("lora_out_747_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(397824128)))]; + tensor lora_out_747_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7551, groups = var_7291, pad = lora_out_745_pad_0, pad_type = lora_out_745_pad_type_0, strides = var_7549, weight = lora_out_747_weight_0_to_fp16, x = input_555_cast_fp16)[name = tensor("lora_out_747_cast_fp16")]; + tensor value_75_cast_fp16 = add(x = pretrained_out_373_cast_fp16, y = lora_out_747_cast_fp16)[name = tensor("value_75_cast_fp16")]; + tensor var_7558 = const()[name = tensor("op_7558"), val = tensor([1, 20, 64, -1])]; + tensor var_7559_cast_fp16 = reshape(shape = var_7558, x = query_75_cast_fp16)[name = tensor("op_7559_cast_fp16")]; + tensor var_7560_to_fp16 = const()[name = tensor("op_7560_to_fp16"), val = tensor(0x1p-3)]; + tensor var_7561_cast_fp16 = mul(x = var_7559_cast_fp16, y = var_7560_to_fp16)[name = tensor("op_7561_cast_fp16")]; + tensor var_7562 = const()[name = tensor("op_7562"), val = tensor([1, 20, 64, -1])]; + tensor var_7563_cast_fp16 = reshape(shape = var_7562, x = key_75_cast_fp16)[name = tensor("op_7563_cast_fp16")]; + tensor mh_w_113_transpose_x_0 = const()[name = tensor("mh_w_113_transpose_x_0"), val = tensor(true)]; + tensor mh_w_113_transpose_y_0 = const()[name = tensor("mh_w_113_transpose_y_0"), val = tensor(false)]; + tensor mh_w_113_cast_fp16 = matmul(transpose_x = mh_w_113_transpose_x_0, transpose_y = mh_w_113_transpose_y_0, x = var_7561_cast_fp16, y = var_7563_cast_fp16)[name = tensor("mh_w_113_cast_fp16")]; + tensor obj_265_cast_fp16 = softmax(axis = var_7284, x = mh_w_113_cast_fp16)[name = tensor("obj_265_cast_fp16")]; + tensor var_7567 = const()[name = tensor("op_7567"), val = tensor([1, 20, 64, -1])]; + tensor var_7568_cast_fp16 = reshape(shape = var_7567, x = value_75_cast_fp16)[name = tensor("op_7568_cast_fp16")]; + tensor attn_75_transpose_x_0 = const()[name = tensor("attn_75_transpose_x_0"), val = tensor(false)]; + tensor attn_75_transpose_y_0 = const()[name = tensor("attn_75_transpose_y_0"), val = tensor(true)]; + tensor attn_75_cast_fp16 = matmul(transpose_x = attn_75_transpose_x_0, transpose_y = attn_75_transpose_y_0, x = var_7568_cast_fp16, y = obj_265_cast_fp16)[name = tensor("attn_75_cast_fp16")]; + tensor var_7571 = const()[name = tensor("op_7571"), val = tensor([1, 1280, 1, -1])]; + tensor input_557_cast_fp16 = reshape(shape = var_7571, x = attn_75_cast_fp16)[name = tensor("input_557_cast_fp16")]; + tensor var_7578 = const()[name = tensor("op_7578"), val = tensor([1, 1])]; + tensor var_7580 = const()[name = tensor("op_7580"), val = tensor([1, 1])]; + tensor pretrained_out_375_pad_type_0 = const()[name = tensor("pretrained_out_375_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_375_pad_0 = const()[name = tensor("pretrained_out_375_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(397865152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398684416))), name = tensor("layers_18_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_18_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398684544)))]; + tensor pretrained_out_375_cast_fp16 = conv(bias = layers_18_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7580, groups = var_7291, pad = pretrained_out_375_pad_0, pad_type = pretrained_out_375_pad_type_0, strides = var_7578, weight = layers_18_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_557_cast_fp16)[name = tensor("pretrained_out_375_cast_fp16")]; + tensor var_7584 = const()[name = tensor("op_7584"), val = tensor([1, 1])]; + tensor var_7586 = const()[name = tensor("op_7586"), val = tensor([1, 1])]; + tensor input_559_pad_type_0 = const()[name = tensor("input_559_pad_type_0"), val = tensor("custom")]; + tensor input_559_pad_0 = const()[name = tensor("input_559_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398687168)))]; + tensor input_559_cast_fp16 = conv(dilations = var_7586, groups = var_7291, pad = input_559_pad_0, pad_type = input_559_pad_type_0, strides = var_7584, weight = layers_18_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_557_cast_fp16)[name = tensor("input_559_cast_fp16")]; + tensor var_7590 = const()[name = tensor("op_7590"), val = tensor([1, 1])]; + tensor var_7592 = const()[name = tensor("op_7592"), val = tensor([1, 1])]; + tensor lora_out_749_pad_type_0 = const()[name = tensor("lora_out_749_pad_type_0"), val = tensor("custom")]; + tensor lora_out_749_pad_0 = const()[name = tensor("lora_out_749_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_751_weight_0_to_fp16 = const()[name = tensor("lora_out_751_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398728192)))]; + tensor lora_out_751_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7592, groups = var_7291, pad = lora_out_749_pad_0, pad_type = lora_out_749_pad_type_0, strides = var_7590, weight = lora_out_751_weight_0_to_fp16, x = input_559_cast_fp16)[name = tensor("lora_out_751_cast_fp16")]; + tensor obj_263_cast_fp16 = add(x = pretrained_out_375_cast_fp16, y = lora_out_751_cast_fp16)[name = tensor("obj_263_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = obj_263_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; + tensor var_7604 = const()[name = tensor("op_7604"), val = tensor([1])]; + tensor channels_mean_113_cast_fp16 = reduce_mean(axes = var_7604, keep_dims = var_7292, x = inputs_113_cast_fp16)[name = tensor("channels_mean_113_cast_fp16")]; + tensor zero_mean_113_cast_fp16 = sub(x = inputs_113_cast_fp16, y = channels_mean_113_cast_fp16)[name = tensor("zero_mean_113_cast_fp16")]; + tensor zero_mean_sq_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = zero_mean_113_cast_fp16)[name = tensor("zero_mean_sq_113_cast_fp16")]; + tensor var_7608 = const()[name = tensor("op_7608"), val = tensor([1])]; + tensor var_7609_cast_fp16 = reduce_mean(axes = var_7608, keep_dims = var_7292, x = zero_mean_sq_113_cast_fp16)[name = tensor("op_7609_cast_fp16")]; + tensor var_7610_to_fp16 = const()[name = tensor("op_7610_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7611_cast_fp16 = add(x = var_7609_cast_fp16, y = var_7610_to_fp16)[name = tensor("op_7611_cast_fp16")]; + tensor denom_113_epsilon_0 = const()[name = tensor("denom_113_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_113_cast_fp16 = rsqrt(epsilon = denom_113_epsilon_0, x = var_7611_cast_fp16)[name = tensor("denom_113_cast_fp16")]; + tensor out_113_cast_fp16 = mul(x = zero_mean_113_cast_fp16, y = denom_113_cast_fp16)[name = tensor("out_113_cast_fp16")]; + tensor input_561_gamma_0_to_fp16 = const()[name = tensor("input_561_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398769216)))]; + tensor input_561_beta_0_to_fp16 = const()[name = tensor("input_561_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398771840)))]; + tensor input_561_epsilon_0_to_fp16 = const()[name = tensor("input_561_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_561_cast_fp16 = batch_norm(beta = input_561_beta_0_to_fp16, epsilon = input_561_epsilon_0_to_fp16, gamma = input_561_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor("input_561_cast_fp16")]; + tensor var_7625 = const()[name = tensor("op_7625"), val = tensor([1, 1])]; + tensor var_7627 = const()[name = tensor("op_7627"), val = tensor([1, 1])]; + tensor pretrained_out_377_pad_type_0 = const()[name = tensor("pretrained_out_377_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_377_pad_0 = const()[name = tensor("pretrained_out_377_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(398774464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402051328))), name = tensor("layers_18_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_18_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402051456)))]; + tensor pretrained_out_377_cast_fp16 = conv(bias = layers_18_fc1_pretrained_bias_to_fp16, dilations = var_7627, groups = var_7291, pad = pretrained_out_377_pad_0, pad_type = pretrained_out_377_pad_type_0, strides = var_7625, weight = layers_18_fc1_pretrained_weight_to_fp16_palettized, x = input_561_cast_fp16)[name = tensor("pretrained_out_377_cast_fp16")]; + tensor var_7631 = const()[name = tensor("op_7631"), val = tensor([1, 1])]; + tensor var_7633 = const()[name = tensor("op_7633"), val = tensor([1, 1])]; + tensor input_563_pad_type_0 = const()[name = tensor("input_563_pad_type_0"), val = tensor("custom")]; + tensor input_563_pad_0 = const()[name = tensor("input_563_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_18_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402061760)))]; + tensor input_563_cast_fp16 = conv(dilations = var_7633, groups = var_7291, pad = input_563_pad_0, pad_type = input_563_pad_type_0, strides = var_7631, weight = layers_18_fc1_loraA_weight_to_fp16, x = input_561_cast_fp16)[name = tensor("input_563_cast_fp16")]; + tensor var_7637 = const()[name = tensor("op_7637"), val = tensor([1, 1])]; + tensor var_7639 = const()[name = tensor("op_7639"), val = tensor([1, 1])]; + tensor lora_out_753_pad_type_0 = const()[name = tensor("lora_out_753_pad_type_0"), val = tensor("custom")]; + tensor lora_out_753_pad_0 = const()[name = tensor("lora_out_753_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_755_weight_0_to_fp16 = const()[name = tensor("lora_out_755_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402102784)))]; + tensor lora_out_755_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_7639, groups = var_7291, pad = lora_out_753_pad_0, pad_type = lora_out_753_pad_type_0, strides = var_7637, weight = lora_out_755_weight_0_to_fp16, x = input_563_cast_fp16)[name = tensor("lora_out_755_cast_fp16")]; + tensor input_565_cast_fp16 = add(x = pretrained_out_377_cast_fp16, y = lora_out_755_cast_fp16)[name = tensor("input_565_cast_fp16")]; + tensor input_567_mode_0 = const()[name = tensor("input_567_mode_0"), val = tensor("EXACT")]; + tensor input_567_cast_fp16 = gelu(mode = input_567_mode_0, x = input_565_cast_fp16)[name = tensor("input_567_cast_fp16")]; + tensor var_7651 = const()[name = tensor("op_7651"), val = tensor([1, 1])]; + tensor var_7653 = const()[name = tensor("op_7653"), val = tensor([1, 1])]; + tensor pretrained_out_379_pad_type_0 = const()[name = tensor("pretrained_out_379_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_379_pad_0 = const()[name = tensor("pretrained_out_379_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402266688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405543552))), name = tensor("layers_18_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_18_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_18_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405543680)))]; + tensor pretrained_out_379_cast_fp16 = conv(bias = layers_18_fc2_pretrained_bias_to_fp16, dilations = var_7653, groups = var_7291, pad = pretrained_out_379_pad_0, pad_type = pretrained_out_379_pad_type_0, strides = var_7651, weight = layers_18_fc2_pretrained_weight_to_fp16_palettized, x = input_567_cast_fp16)[name = tensor("pretrained_out_379_cast_fp16")]; + tensor var_7657 = const()[name = tensor("op_7657"), val = tensor([1, 1])]; + tensor var_7659 = const()[name = tensor("op_7659"), val = tensor([1, 1])]; + tensor input_569_pad_type_0 = const()[name = tensor("input_569_pad_type_0"), val = tensor("custom")]; + tensor input_569_pad_0 = const()[name = tensor("input_569_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_18_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_18_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405546304)))]; + tensor input_569_cast_fp16 = conv(dilations = var_7659, groups = var_7291, pad = input_569_pad_0, pad_type = input_569_pad_type_0, strides = var_7657, weight = layers_18_fc2_loraA_weight_to_fp16, x = input_567_cast_fp16)[name = tensor("input_569_cast_fp16")]; + tensor var_7663 = const()[name = tensor("op_7663"), val = tensor([1, 1])]; + tensor var_7665 = const()[name = tensor("op_7665"), val = tensor([1, 1])]; + tensor lora_out_757_pad_type_0 = const()[name = tensor("lora_out_757_pad_type_0"), val = tensor("custom")]; + tensor lora_out_757_pad_0 = const()[name = tensor("lora_out_757_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_759_weight_0_to_fp16 = const()[name = tensor("lora_out_759_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405710208)))]; + tensor lora_out_759_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7665, groups = var_7291, pad = lora_out_757_pad_0, pad_type = lora_out_757_pad_type_0, strides = var_7663, weight = lora_out_759_weight_0_to_fp16, x = input_569_cast_fp16)[name = tensor("lora_out_759_cast_fp16")]; + tensor hidden_states_39_cast_fp16 = add(x = pretrained_out_379_cast_fp16, y = lora_out_759_cast_fp16)[name = tensor("hidden_states_39_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; + tensor var_7682 = const()[name = tensor("op_7682"), val = tensor(3)]; + tensor var_7689 = const()[name = tensor("op_7689"), val = tensor(1)]; + tensor var_7690 = const()[name = tensor("op_7690"), val = tensor(true)]; + tensor var_7702 = const()[name = tensor("op_7702"), val = tensor([1])]; + tensor channels_mean_115_cast_fp16 = reduce_mean(axes = var_7702, keep_dims = var_7690, x = inputs_115_cast_fp16)[name = tensor("channels_mean_115_cast_fp16")]; + tensor zero_mean_115_cast_fp16 = sub(x = inputs_115_cast_fp16, y = channels_mean_115_cast_fp16)[name = tensor("zero_mean_115_cast_fp16")]; + tensor zero_mean_sq_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = zero_mean_115_cast_fp16)[name = tensor("zero_mean_sq_115_cast_fp16")]; + tensor var_7706 = const()[name = tensor("op_7706"), val = tensor([1])]; + tensor var_7707_cast_fp16 = reduce_mean(axes = var_7706, keep_dims = var_7690, x = zero_mean_sq_115_cast_fp16)[name = tensor("op_7707_cast_fp16")]; + tensor var_7708_to_fp16 = const()[name = tensor("op_7708_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7709_cast_fp16 = add(x = var_7707_cast_fp16, y = var_7708_to_fp16)[name = tensor("op_7709_cast_fp16")]; + tensor denom_115_epsilon_0 = const()[name = tensor("denom_115_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_115_cast_fp16 = rsqrt(epsilon = denom_115_epsilon_0, x = var_7709_cast_fp16)[name = tensor("denom_115_cast_fp16")]; + tensor out_115_cast_fp16 = mul(x = zero_mean_115_cast_fp16, y = denom_115_cast_fp16)[name = tensor("out_115_cast_fp16")]; + tensor obj_267_gamma_0_to_fp16 = const()[name = tensor("obj_267_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405751232)))]; + tensor obj_267_beta_0_to_fp16 = const()[name = tensor("obj_267_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405753856)))]; + tensor obj_267_epsilon_0_to_fp16 = const()[name = tensor("obj_267_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_267_cast_fp16 = batch_norm(beta = obj_267_beta_0_to_fp16, epsilon = obj_267_epsilon_0_to_fp16, gamma = obj_267_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor("obj_267_cast_fp16")]; + tensor var_7727 = const()[name = tensor("op_7727"), val = tensor([1, 1])]; + tensor var_7729 = const()[name = tensor("op_7729"), val = tensor([1, 1])]; + tensor pretrained_out_381_pad_type_0 = const()[name = tensor("pretrained_out_381_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_381_pad_0 = const()[name = tensor("pretrained_out_381_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(405756480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406575744))), name = tensor("layers_19_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_19_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406575872)))]; + tensor pretrained_out_381_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_7729, groups = var_7689, pad = pretrained_out_381_pad_0, pad_type = pretrained_out_381_pad_type_0, strides = var_7727, weight = layers_19_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_267_cast_fp16)[name = tensor("pretrained_out_381_cast_fp16")]; + tensor var_7733 = const()[name = tensor("op_7733"), val = tensor([1, 1])]; + tensor var_7735 = const()[name = tensor("op_7735"), val = tensor([1, 1])]; + tensor input_571_pad_type_0 = const()[name = tensor("input_571_pad_type_0"), val = tensor("custom")]; + tensor input_571_pad_0 = const()[name = tensor("input_571_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406578496)))]; + tensor input_571_cast_fp16 = conv(dilations = var_7735, groups = var_7689, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = var_7733, weight = layers_19_self_attn_q_proj_loraA_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("input_571_cast_fp16")]; + tensor var_7739 = const()[name = tensor("op_7739"), val = tensor([1, 1])]; + tensor var_7741 = const()[name = tensor("op_7741"), val = tensor([1, 1])]; + tensor lora_out_761_pad_type_0 = const()[name = tensor("lora_out_761_pad_type_0"), val = tensor("custom")]; + tensor lora_out_761_pad_0 = const()[name = tensor("lora_out_761_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_763_weight_0_to_fp16 = const()[name = tensor("lora_out_763_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406619520)))]; + tensor lora_out_763_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7741, groups = var_7689, pad = lora_out_761_pad_0, pad_type = lora_out_761_pad_type_0, strides = var_7739, weight = lora_out_763_weight_0_to_fp16, x = input_571_cast_fp16)[name = tensor("lora_out_763_cast_fp16")]; + tensor query_77_cast_fp16 = add(x = pretrained_out_381_cast_fp16, y = lora_out_763_cast_fp16)[name = tensor("query_77_cast_fp16")]; + tensor var_7751 = const()[name = tensor("op_7751"), val = tensor([1, 1])]; + tensor var_7753 = const()[name = tensor("op_7753"), val = tensor([1, 1])]; + tensor pretrained_out_383_pad_type_0 = const()[name = tensor("pretrained_out_383_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_383_pad_0 = const()[name = tensor("pretrained_out_383_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(406660544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407479808))), name = tensor("layers_19_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_383_cast_fp16 = conv(dilations = var_7753, groups = var_7689, pad = pretrained_out_383_pad_0, pad_type = pretrained_out_383_pad_type_0, strides = var_7751, weight = layers_19_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_267_cast_fp16)[name = tensor("pretrained_out_383_cast_fp16")]; + tensor var_7757 = const()[name = tensor("op_7757"), val = tensor([1, 1])]; + tensor var_7759 = const()[name = tensor("op_7759"), val = tensor([1, 1])]; + tensor input_573_pad_type_0 = const()[name = tensor("input_573_pad_type_0"), val = tensor("custom")]; + tensor input_573_pad_0 = const()[name = tensor("input_573_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407479936)))]; + tensor input_573_cast_fp16 = conv(dilations = var_7759, groups = var_7689, pad = input_573_pad_0, pad_type = input_573_pad_type_0, strides = var_7757, weight = layers_19_self_attn_k_proj_loraA_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("input_573_cast_fp16")]; + tensor var_7763 = const()[name = tensor("op_7763"), val = tensor([1, 1])]; + tensor var_7765 = const()[name = tensor("op_7765"), val = tensor([1, 1])]; + tensor lora_out_765_pad_type_0 = const()[name = tensor("lora_out_765_pad_type_0"), val = tensor("custom")]; + tensor lora_out_765_pad_0 = const()[name = tensor("lora_out_765_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_767_weight_0_to_fp16 = const()[name = tensor("lora_out_767_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407520960)))]; + tensor lora_out_767_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7765, groups = var_7689, pad = lora_out_765_pad_0, pad_type = lora_out_765_pad_type_0, strides = var_7763, weight = lora_out_767_weight_0_to_fp16, x = input_573_cast_fp16)[name = tensor("lora_out_767_cast_fp16")]; + tensor current_key_39_cast_fp16 = add(x = pretrained_out_383_cast_fp16, y = lora_out_767_cast_fp16)[name = tensor("current_key_39_cast_fp16")]; + tensor var_7776 = const()[name = tensor("op_7776"), val = tensor([1, 1])]; + tensor var_7778 = const()[name = tensor("op_7778"), val = tensor([1, 1])]; + tensor pretrained_out_385_pad_type_0 = const()[name = tensor("pretrained_out_385_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_385_pad_0 = const()[name = tensor("pretrained_out_385_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407561984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408381248))), name = tensor("layers_19_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_19_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408381376)))]; + tensor pretrained_out_385_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_7778, groups = var_7689, pad = pretrained_out_385_pad_0, pad_type = pretrained_out_385_pad_type_0, strides = var_7776, weight = layers_19_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_267_cast_fp16)[name = tensor("pretrained_out_385_cast_fp16")]; + tensor var_7782 = const()[name = tensor("op_7782"), val = tensor([1, 1])]; + tensor var_7784 = const()[name = tensor("op_7784"), val = tensor([1, 1])]; + tensor input_575_pad_type_0 = const()[name = tensor("input_575_pad_type_0"), val = tensor("custom")]; + tensor input_575_pad_0 = const()[name = tensor("input_575_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408384000)))]; + tensor input_575_cast_fp16 = conv(dilations = var_7784, groups = var_7689, pad = input_575_pad_0, pad_type = input_575_pad_type_0, strides = var_7782, weight = layers_19_self_attn_v_proj_loraA_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("input_575_cast_fp16")]; + tensor var_7788 = const()[name = tensor("op_7788"), val = tensor([1, 1])]; + tensor var_7790 = const()[name = tensor("op_7790"), val = tensor([1, 1])]; + tensor lora_out_769_pad_type_0 = const()[name = tensor("lora_out_769_pad_type_0"), val = tensor("custom")]; + tensor lora_out_769_pad_0 = const()[name = tensor("lora_out_769_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_771_weight_0_to_fp16 = const()[name = tensor("lora_out_771_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408425024)))]; + tensor lora_out_771_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7790, groups = var_7689, pad = lora_out_769_pad_0, pad_type = lora_out_769_pad_type_0, strides = var_7788, weight = lora_out_771_weight_0_to_fp16, x = input_575_cast_fp16)[name = tensor("lora_out_771_cast_fp16")]; + tensor current_value_39_cast_fp16 = add(x = pretrained_out_385_cast_fp16, y = lora_out_771_cast_fp16)[name = tensor("current_value_39_cast_fp16")]; + tensor var_7800_cast_fp16 = mul(x = current_key_39_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_7800_cast_fp16")]; + tensor var_7802_cast_fp16 = mul(x = var_103_cast_fp16_19, y = var_295_cast_fp16)[name = tensor("op_7802_cast_fp16")]; + tensor key_77_cast_fp16 = add(x = var_7800_cast_fp16, y = var_7802_cast_fp16)[name = tensor("key_77_cast_fp16")]; + tensor var_7804_cast_fp16 = mul(x = current_value_39_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_7804_cast_fp16")]; + tensor var_7806_cast_fp16 = mul(x = var_138_cast_fp16_19, y = var_295_cast_fp16)[name = tensor("op_7806_cast_fp16")]; + tensor value_77_cast_fp16 = add(x = var_7804_cast_fp16, y = var_7806_cast_fp16)[name = tensor("value_77_cast_fp16")]; + tensor var_7809 = const()[name = tensor("op_7809"), val = tensor([1, 20, 64, -1])]; + tensor var_7810_cast_fp16 = reshape(shape = var_7809, x = query_77_cast_fp16)[name = tensor("op_7810_cast_fp16")]; + tensor var_7811_to_fp16 = const()[name = tensor("op_7811_to_fp16"), val = tensor(0x1p-3)]; + tensor var_7812_cast_fp16 = mul(x = var_7810_cast_fp16, y = var_7811_to_fp16)[name = tensor("op_7812_cast_fp16")]; + tensor var_7813 = const()[name = tensor("op_7813"), val = tensor([1, 20, 64, -1])]; + tensor var_7814_cast_fp16 = reshape(shape = var_7813, x = key_77_cast_fp16)[name = tensor("op_7814_cast_fp16")]; + tensor mh_w_115_transpose_x_0 = const()[name = tensor("mh_w_115_transpose_x_0"), val = tensor(true)]; + tensor mh_w_115_transpose_y_0 = const()[name = tensor("mh_w_115_transpose_y_0"), val = tensor(false)]; + tensor mh_w_115_cast_fp16 = matmul(transpose_x = mh_w_115_transpose_x_0, transpose_y = mh_w_115_transpose_y_0, x = var_7812_cast_fp16, y = var_7814_cast_fp16)[name = tensor("mh_w_115_cast_fp16")]; + tensor mh_w_117_cast_fp16 = add(x = mh_w_115_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_117_cast_fp16")]; + tensor var_7822_cast_fp16 = softmax(axis = var_7682, x = mh_w_117_cast_fp16)[name = tensor("op_7822_cast_fp16")]; + tensor var_7823 = const()[name = tensor("op_7823"), val = tensor([1, 20, 64, -1])]; + tensor var_7824_cast_fp16 = reshape(shape = var_7823, x = value_77_cast_fp16)[name = tensor("op_7824_cast_fp16")]; + tensor attn_77_transpose_x_0 = const()[name = tensor("attn_77_transpose_x_0"), val = tensor(false)]; + tensor attn_77_transpose_y_0 = const()[name = tensor("attn_77_transpose_y_0"), val = tensor(true)]; + tensor attn_77_cast_fp16 = matmul(transpose_x = attn_77_transpose_x_0, transpose_y = attn_77_transpose_y_0, x = var_7824_cast_fp16, y = var_7822_cast_fp16)[name = tensor("attn_77_cast_fp16")]; + tensor var_7827 = const()[name = tensor("op_7827"), val = tensor([1, 1280, 1, -1])]; + tensor input_577_cast_fp16 = reshape(shape = var_7827, x = attn_77_cast_fp16)[name = tensor("input_577_cast_fp16")]; + tensor var_7834 = const()[name = tensor("op_7834"), val = tensor([1, 1])]; + tensor var_7836 = const()[name = tensor("op_7836"), val = tensor([1, 1])]; + tensor pretrained_out_387_pad_type_0 = const()[name = tensor("pretrained_out_387_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_387_pad_0 = const()[name = tensor("pretrained_out_387_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408466048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409285312))), name = tensor("layers_19_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_19_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409285440)))]; + tensor pretrained_out_387_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7836, groups = var_7689, pad = pretrained_out_387_pad_0, pad_type = pretrained_out_387_pad_type_0, strides = var_7834, weight = layers_19_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_577_cast_fp16)[name = tensor("pretrained_out_387_cast_fp16")]; + tensor var_7840 = const()[name = tensor("op_7840"), val = tensor([1, 1])]; + tensor var_7842 = const()[name = tensor("op_7842"), val = tensor([1, 1])]; + tensor input_579_pad_type_0 = const()[name = tensor("input_579_pad_type_0"), val = tensor("custom")]; + tensor input_579_pad_0 = const()[name = tensor("input_579_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409288064)))]; + tensor input_579_cast_fp16 = conv(dilations = var_7842, groups = var_7689, pad = input_579_pad_0, pad_type = input_579_pad_type_0, strides = var_7840, weight = layers_19_self_attn_o_proj_loraA_weight_to_fp16, x = input_577_cast_fp16)[name = tensor("input_579_cast_fp16")]; + tensor var_7846 = const()[name = tensor("op_7846"), val = tensor([1, 1])]; + tensor var_7848 = const()[name = tensor("op_7848"), val = tensor([1, 1])]; + tensor lora_out_773_pad_type_0 = const()[name = tensor("lora_out_773_pad_type_0"), val = tensor("custom")]; + tensor lora_out_773_pad_0 = const()[name = tensor("lora_out_773_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_775_weight_0_to_fp16 = const()[name = tensor("lora_out_775_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409329088)))]; + tensor lora_out_775_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7848, groups = var_7689, pad = lora_out_773_pad_0, pad_type = lora_out_773_pad_type_0, strides = var_7846, weight = lora_out_775_weight_0_to_fp16, x = input_579_cast_fp16)[name = tensor("lora_out_775_cast_fp16")]; + tensor obj_273_cast_fp16 = add(x = pretrained_out_387_cast_fp16, y = lora_out_775_cast_fp16)[name = tensor("obj_273_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = obj_273_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; + tensor var_7861 = const()[name = tensor("op_7861"), val = tensor([1])]; + tensor channels_mean_117_cast_fp16 = reduce_mean(axes = var_7861, keep_dims = var_7690, x = inputs_117_cast_fp16)[name = tensor("channels_mean_117_cast_fp16")]; + tensor zero_mean_117_cast_fp16 = sub(x = inputs_117_cast_fp16, y = channels_mean_117_cast_fp16)[name = tensor("zero_mean_117_cast_fp16")]; + tensor zero_mean_sq_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = zero_mean_117_cast_fp16)[name = tensor("zero_mean_sq_117_cast_fp16")]; + tensor var_7865 = const()[name = tensor("op_7865"), val = tensor([1])]; + tensor var_7866_cast_fp16 = reduce_mean(axes = var_7865, keep_dims = var_7690, x = zero_mean_sq_117_cast_fp16)[name = tensor("op_7866_cast_fp16")]; + tensor var_7867_to_fp16 = const()[name = tensor("op_7867_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7868_cast_fp16 = add(x = var_7866_cast_fp16, y = var_7867_to_fp16)[name = tensor("op_7868_cast_fp16")]; + tensor denom_117_epsilon_0 = const()[name = tensor("denom_117_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_117_cast_fp16 = rsqrt(epsilon = denom_117_epsilon_0, x = var_7868_cast_fp16)[name = tensor("denom_117_cast_fp16")]; + tensor out_117_cast_fp16 = mul(x = zero_mean_117_cast_fp16, y = denom_117_cast_fp16)[name = tensor("out_117_cast_fp16")]; + tensor obj_275_gamma_0_to_fp16 = const()[name = tensor("obj_275_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409370112)))]; + tensor obj_275_beta_0_to_fp16 = const()[name = tensor("obj_275_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409372736)))]; + tensor obj_275_epsilon_0_to_fp16 = const()[name = tensor("obj_275_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_275_cast_fp16 = batch_norm(beta = obj_275_beta_0_to_fp16, epsilon = obj_275_epsilon_0_to_fp16, gamma = obj_275_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor("obj_275_cast_fp16")]; + tensor var_7886 = const()[name = tensor("op_7886"), val = tensor([1, 1])]; + tensor var_7888 = const()[name = tensor("op_7888"), val = tensor([1, 1])]; + tensor pretrained_out_389_pad_type_0 = const()[name = tensor("pretrained_out_389_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_389_pad_0 = const()[name = tensor("pretrained_out_389_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409375360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410194624))), name = tensor("layers_19_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_19_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410194752)))]; + tensor pretrained_out_389_cast_fp16 = conv(bias = layers_19_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_7888, groups = var_7689, pad = pretrained_out_389_pad_0, pad_type = pretrained_out_389_pad_type_0, strides = var_7886, weight = layers_19_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_275_cast_fp16)[name = tensor("pretrained_out_389_cast_fp16")]; + tensor var_7892 = const()[name = tensor("op_7892"), val = tensor([1, 1])]; + tensor var_7894 = const()[name = tensor("op_7894"), val = tensor([1, 1])]; + tensor input_581_pad_type_0 = const()[name = tensor("input_581_pad_type_0"), val = tensor("custom")]; + tensor input_581_pad_0 = const()[name = tensor("input_581_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410197376)))]; + tensor input_581_cast_fp16 = conv(dilations = var_7894, groups = var_7689, pad = input_581_pad_0, pad_type = input_581_pad_type_0, strides = var_7892, weight = layers_19_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_275_cast_fp16)[name = tensor("input_581_cast_fp16")]; + tensor var_7898 = const()[name = tensor("op_7898"), val = tensor([1, 1])]; + tensor var_7900 = const()[name = tensor("op_7900"), val = tensor([1, 1])]; + tensor lora_out_777_pad_type_0 = const()[name = tensor("lora_out_777_pad_type_0"), val = tensor("custom")]; + tensor lora_out_777_pad_0 = const()[name = tensor("lora_out_777_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_779_weight_0_to_fp16 = const()[name = tensor("lora_out_779_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410238400)))]; + tensor lora_out_779_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7900, groups = var_7689, pad = lora_out_777_pad_0, pad_type = lora_out_777_pad_type_0, strides = var_7898, weight = lora_out_779_weight_0_to_fp16, x = input_581_cast_fp16)[name = tensor("lora_out_779_cast_fp16")]; + tensor query_79_cast_fp16 = add(x = pretrained_out_389_cast_fp16, y = lora_out_779_cast_fp16)[name = tensor("query_79_cast_fp16")]; + tensor var_7910 = const()[name = tensor("op_7910"), val = tensor([1, 1])]; + tensor var_7912 = const()[name = tensor("op_7912"), val = tensor([1, 1])]; + tensor pretrained_out_391_pad_type_0 = const()[name = tensor("pretrained_out_391_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_391_pad_0 = const()[name = tensor("pretrained_out_391_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410279424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411098688))), name = tensor("layers_19_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_391_cast_fp16 = conv(dilations = var_7912, groups = var_7689, pad = pretrained_out_391_pad_0, pad_type = pretrained_out_391_pad_type_0, strides = var_7910, weight = layers_19_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_391_cast_fp16")]; + tensor var_7916 = const()[name = tensor("op_7916"), val = tensor([1, 1])]; + tensor var_7918 = const()[name = tensor("op_7918"), val = tensor([1, 1])]; + tensor input_583_pad_type_0 = const()[name = tensor("input_583_pad_type_0"), val = tensor("custom")]; + tensor input_583_pad_0 = const()[name = tensor("input_583_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411098816)))]; + tensor input_583_cast_fp16 = conv(dilations = var_7918, groups = var_7689, pad = input_583_pad_0, pad_type = input_583_pad_type_0, strides = var_7916, weight = layers_19_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_583_cast_fp16")]; + tensor var_7922 = const()[name = tensor("op_7922"), val = tensor([1, 1])]; + tensor var_7924 = const()[name = tensor("op_7924"), val = tensor([1, 1])]; + tensor lora_out_781_pad_type_0 = const()[name = tensor("lora_out_781_pad_type_0"), val = tensor("custom")]; + tensor lora_out_781_pad_0 = const()[name = tensor("lora_out_781_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_783_weight_0_to_fp16 = const()[name = tensor("lora_out_783_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411139840)))]; + tensor lora_out_783_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7924, groups = var_7689, pad = lora_out_781_pad_0, pad_type = lora_out_781_pad_type_0, strides = var_7922, weight = lora_out_783_weight_0_to_fp16, x = input_583_cast_fp16)[name = tensor("lora_out_783_cast_fp16")]; + tensor key_79_cast_fp16 = add(x = pretrained_out_391_cast_fp16, y = lora_out_783_cast_fp16)[name = tensor("key_79_cast_fp16")]; + tensor var_7935 = const()[name = tensor("op_7935"), val = tensor([1, 1])]; + tensor var_7937 = const()[name = tensor("op_7937"), val = tensor([1, 1])]; + tensor pretrained_out_393_pad_type_0 = const()[name = tensor("pretrained_out_393_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_393_pad_0 = const()[name = tensor("pretrained_out_393_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411180864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412000128))), name = tensor("layers_19_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_19_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412000256)))]; + tensor pretrained_out_393_cast_fp16 = conv(bias = layers_19_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_7937, groups = var_7689, pad = pretrained_out_393_pad_0, pad_type = pretrained_out_393_pad_type_0, strides = var_7935, weight = layers_19_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_393_cast_fp16")]; + tensor var_7941 = const()[name = tensor("op_7941"), val = tensor([1, 1])]; + tensor var_7943 = const()[name = tensor("op_7943"), val = tensor([1, 1])]; + tensor input_585_pad_type_0 = const()[name = tensor("input_585_pad_type_0"), val = tensor("custom")]; + tensor input_585_pad_0 = const()[name = tensor("input_585_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412002880)))]; + tensor input_585_cast_fp16 = conv(dilations = var_7943, groups = var_7689, pad = input_585_pad_0, pad_type = input_585_pad_type_0, strides = var_7941, weight = layers_19_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_585_cast_fp16")]; + tensor var_7947 = const()[name = tensor("op_7947"), val = tensor([1, 1])]; + tensor var_7949 = const()[name = tensor("op_7949"), val = tensor([1, 1])]; + tensor lora_out_785_pad_type_0 = const()[name = tensor("lora_out_785_pad_type_0"), val = tensor("custom")]; + tensor lora_out_785_pad_0 = const()[name = tensor("lora_out_785_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_787_weight_0_to_fp16 = const()[name = tensor("lora_out_787_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412043904)))]; + tensor lora_out_787_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7949, groups = var_7689, pad = lora_out_785_pad_0, pad_type = lora_out_785_pad_type_0, strides = var_7947, weight = lora_out_787_weight_0_to_fp16, x = input_585_cast_fp16)[name = tensor("lora_out_787_cast_fp16")]; + tensor value_79_cast_fp16 = add(x = pretrained_out_393_cast_fp16, y = lora_out_787_cast_fp16)[name = tensor("value_79_cast_fp16")]; + tensor var_7956 = const()[name = tensor("op_7956"), val = tensor([1, 20, 64, -1])]; + tensor var_7957_cast_fp16 = reshape(shape = var_7956, x = query_79_cast_fp16)[name = tensor("op_7957_cast_fp16")]; + tensor var_7958_to_fp16 = const()[name = tensor("op_7958_to_fp16"), val = tensor(0x1p-3)]; + tensor var_7959_cast_fp16 = mul(x = var_7957_cast_fp16, y = var_7958_to_fp16)[name = tensor("op_7959_cast_fp16")]; + tensor var_7960 = const()[name = tensor("op_7960"), val = tensor([1, 20, 64, -1])]; + tensor var_7961_cast_fp16 = reshape(shape = var_7960, x = key_79_cast_fp16)[name = tensor("op_7961_cast_fp16")]; + tensor mh_w_119_transpose_x_0 = const()[name = tensor("mh_w_119_transpose_x_0"), val = tensor(true)]; + tensor mh_w_119_transpose_y_0 = const()[name = tensor("mh_w_119_transpose_y_0"), val = tensor(false)]; + tensor mh_w_119_cast_fp16 = matmul(transpose_x = mh_w_119_transpose_x_0, transpose_y = mh_w_119_transpose_y_0, x = var_7959_cast_fp16, y = var_7961_cast_fp16)[name = tensor("mh_w_119_cast_fp16")]; + tensor obj_279_cast_fp16 = softmax(axis = var_7682, x = mh_w_119_cast_fp16)[name = tensor("obj_279_cast_fp16")]; + tensor var_7965 = const()[name = tensor("op_7965"), val = tensor([1, 20, 64, -1])]; + tensor var_7966_cast_fp16 = reshape(shape = var_7965, x = value_79_cast_fp16)[name = tensor("op_7966_cast_fp16")]; + tensor attn_79_transpose_x_0 = const()[name = tensor("attn_79_transpose_x_0"), val = tensor(false)]; + tensor attn_79_transpose_y_0 = const()[name = tensor("attn_79_transpose_y_0"), val = tensor(true)]; + tensor attn_79_cast_fp16 = matmul(transpose_x = attn_79_transpose_x_0, transpose_y = attn_79_transpose_y_0, x = var_7966_cast_fp16, y = obj_279_cast_fp16)[name = tensor("attn_79_cast_fp16")]; + tensor var_7969 = const()[name = tensor("op_7969"), val = tensor([1, 1280, 1, -1])]; + tensor input_587_cast_fp16 = reshape(shape = var_7969, x = attn_79_cast_fp16)[name = tensor("input_587_cast_fp16")]; + tensor var_7976 = const()[name = tensor("op_7976"), val = tensor([1, 1])]; + tensor var_7978 = const()[name = tensor("op_7978"), val = tensor([1, 1])]; + tensor pretrained_out_395_pad_type_0 = const()[name = tensor("pretrained_out_395_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_395_pad_0 = const()[name = tensor("pretrained_out_395_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412084928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412904192))), name = tensor("layers_19_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_19_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412904320)))]; + tensor pretrained_out_395_cast_fp16 = conv(bias = layers_19_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7978, groups = var_7689, pad = pretrained_out_395_pad_0, pad_type = pretrained_out_395_pad_type_0, strides = var_7976, weight = layers_19_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_587_cast_fp16)[name = tensor("pretrained_out_395_cast_fp16")]; + tensor var_7982 = const()[name = tensor("op_7982"), val = tensor([1, 1])]; + tensor var_7984 = const()[name = tensor("op_7984"), val = tensor([1, 1])]; + tensor input_589_pad_type_0 = const()[name = tensor("input_589_pad_type_0"), val = tensor("custom")]; + tensor input_589_pad_0 = const()[name = tensor("input_589_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412906944)))]; + tensor input_589_cast_fp16 = conv(dilations = var_7984, groups = var_7689, pad = input_589_pad_0, pad_type = input_589_pad_type_0, strides = var_7982, weight = layers_19_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_587_cast_fp16)[name = tensor("input_589_cast_fp16")]; + tensor var_7988 = const()[name = tensor("op_7988"), val = tensor([1, 1])]; + tensor var_7990 = const()[name = tensor("op_7990"), val = tensor([1, 1])]; + tensor lora_out_789_pad_type_0 = const()[name = tensor("lora_out_789_pad_type_0"), val = tensor("custom")]; + tensor lora_out_789_pad_0 = const()[name = tensor("lora_out_789_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_791_weight_0_to_fp16 = const()[name = tensor("lora_out_791_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412947968)))]; + tensor lora_out_791_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7990, groups = var_7689, pad = lora_out_789_pad_0, pad_type = lora_out_789_pad_type_0, strides = var_7988, weight = lora_out_791_weight_0_to_fp16, x = input_589_cast_fp16)[name = tensor("lora_out_791_cast_fp16")]; + tensor obj_277_cast_fp16 = add(x = pretrained_out_395_cast_fp16, y = lora_out_791_cast_fp16)[name = tensor("obj_277_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_277_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; + tensor var_8002 = const()[name = tensor("op_8002"), val = tensor([1])]; + tensor channels_mean_119_cast_fp16 = reduce_mean(axes = var_8002, keep_dims = var_7690, x = inputs_119_cast_fp16)[name = tensor("channels_mean_119_cast_fp16")]; + tensor zero_mean_119_cast_fp16 = sub(x = inputs_119_cast_fp16, y = channels_mean_119_cast_fp16)[name = tensor("zero_mean_119_cast_fp16")]; + tensor zero_mean_sq_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = zero_mean_119_cast_fp16)[name = tensor("zero_mean_sq_119_cast_fp16")]; + tensor var_8006 = const()[name = tensor("op_8006"), val = tensor([1])]; + tensor var_8007_cast_fp16 = reduce_mean(axes = var_8006, keep_dims = var_7690, x = zero_mean_sq_119_cast_fp16)[name = tensor("op_8007_cast_fp16")]; + tensor var_8008_to_fp16 = const()[name = tensor("op_8008_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_8009_cast_fp16 = add(x = var_8007_cast_fp16, y = var_8008_to_fp16)[name = tensor("op_8009_cast_fp16")]; + tensor denom_119_epsilon_0 = const()[name = tensor("denom_119_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_119_cast_fp16 = rsqrt(epsilon = denom_119_epsilon_0, x = var_8009_cast_fp16)[name = tensor("denom_119_cast_fp16")]; + tensor out_119_cast_fp16 = mul(x = zero_mean_119_cast_fp16, y = denom_119_cast_fp16)[name = tensor("out_119_cast_fp16")]; + tensor input_591_gamma_0_to_fp16 = const()[name = tensor("input_591_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412988992)))]; + tensor input_591_beta_0_to_fp16 = const()[name = tensor("input_591_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412991616)))]; + tensor input_591_epsilon_0_to_fp16 = const()[name = tensor("input_591_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_591_cast_fp16 = batch_norm(beta = input_591_beta_0_to_fp16, epsilon = input_591_epsilon_0_to_fp16, gamma = input_591_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor("input_591_cast_fp16")]; + tensor var_8023 = const()[name = tensor("op_8023"), val = tensor([1, 1])]; + tensor var_8025 = const()[name = tensor("op_8025"), val = tensor([1, 1])]; + tensor pretrained_out_397_pad_type_0 = const()[name = tensor("pretrained_out_397_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_397_pad_0 = const()[name = tensor("pretrained_out_397_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412994240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416271104))), name = tensor("layers_19_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_19_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416271232)))]; + tensor pretrained_out_397_cast_fp16 = conv(bias = layers_19_fc1_pretrained_bias_to_fp16, dilations = var_8025, groups = var_7689, pad = pretrained_out_397_pad_0, pad_type = pretrained_out_397_pad_type_0, strides = var_8023, weight = layers_19_fc1_pretrained_weight_to_fp16_palettized, x = input_591_cast_fp16)[name = tensor("pretrained_out_397_cast_fp16")]; + tensor var_8029 = const()[name = tensor("op_8029"), val = tensor([1, 1])]; + tensor var_8031 = const()[name = tensor("op_8031"), val = tensor([1, 1])]; + tensor input_593_pad_type_0 = const()[name = tensor("input_593_pad_type_0"), val = tensor("custom")]; + tensor input_593_pad_0 = const()[name = tensor("input_593_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_19_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416281536)))]; + tensor input_593_cast_fp16 = conv(dilations = var_8031, groups = var_7689, pad = input_593_pad_0, pad_type = input_593_pad_type_0, strides = var_8029, weight = layers_19_fc1_loraA_weight_to_fp16, x = input_591_cast_fp16)[name = tensor("input_593_cast_fp16")]; + tensor var_8035 = const()[name = tensor("op_8035"), val = tensor([1, 1])]; + tensor var_8037 = const()[name = tensor("op_8037"), val = tensor([1, 1])]; + tensor lora_out_793_pad_type_0 = const()[name = tensor("lora_out_793_pad_type_0"), val = tensor("custom")]; + tensor lora_out_793_pad_0 = const()[name = tensor("lora_out_793_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_795_weight_0_to_fp16 = const()[name = tensor("lora_out_795_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416322560)))]; + tensor lora_out_795_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_8037, groups = var_7689, pad = lora_out_793_pad_0, pad_type = lora_out_793_pad_type_0, strides = var_8035, weight = lora_out_795_weight_0_to_fp16, x = input_593_cast_fp16)[name = tensor("lora_out_795_cast_fp16")]; + tensor input_595_cast_fp16 = add(x = pretrained_out_397_cast_fp16, y = lora_out_795_cast_fp16)[name = tensor("input_595_cast_fp16")]; + tensor input_597_mode_0 = const()[name = tensor("input_597_mode_0"), val = tensor("EXACT")]; + tensor input_597_cast_fp16 = gelu(mode = input_597_mode_0, x = input_595_cast_fp16)[name = tensor("input_597_cast_fp16")]; + tensor var_8049 = const()[name = tensor("op_8049"), val = tensor([1, 1])]; + tensor var_8051 = const()[name = tensor("op_8051"), val = tensor([1, 1])]; + tensor pretrained_out_399_pad_type_0 = const()[name = tensor("pretrained_out_399_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_399_pad_0 = const()[name = tensor("pretrained_out_399_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416486464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419763328))), name = tensor("layers_19_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_19_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_19_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419763456)))]; + tensor pretrained_out_399_cast_fp16 = conv(bias = layers_19_fc2_pretrained_bias_to_fp16, dilations = var_8051, groups = var_7689, pad = pretrained_out_399_pad_0, pad_type = pretrained_out_399_pad_type_0, strides = var_8049, weight = layers_19_fc2_pretrained_weight_to_fp16_palettized, x = input_597_cast_fp16)[name = tensor("pretrained_out_399_cast_fp16")]; + tensor var_8055 = const()[name = tensor("op_8055"), val = tensor([1, 1])]; + tensor var_8057 = const()[name = tensor("op_8057"), val = tensor([1, 1])]; + tensor input_599_pad_type_0 = const()[name = tensor("input_599_pad_type_0"), val = tensor("custom")]; + tensor input_599_pad_0 = const()[name = tensor("input_599_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_19_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_19_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419766080)))]; + tensor input_599_cast_fp16 = conv(dilations = var_8057, groups = var_7689, pad = input_599_pad_0, pad_type = input_599_pad_type_0, strides = var_8055, weight = layers_19_fc2_loraA_weight_to_fp16, x = input_597_cast_fp16)[name = tensor("input_599_cast_fp16")]; + tensor var_8061 = const()[name = tensor("op_8061"), val = tensor([1, 1])]; + tensor var_8063 = const()[name = tensor("op_8063"), val = tensor([1, 1])]; + tensor lora_out_797_pad_type_0 = const()[name = tensor("lora_out_797_pad_type_0"), val = tensor("custom")]; + tensor lora_out_797_pad_0 = const()[name = tensor("lora_out_797_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_799_weight_0_to_fp16 = const()[name = tensor("lora_out_799_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419929984)))]; + tensor lora_out_799_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8063, groups = var_7689, pad = lora_out_797_pad_0, pad_type = lora_out_797_pad_type_0, strides = var_8061, weight = lora_out_799_weight_0_to_fp16, x = input_599_cast_fp16)[name = tensor("lora_out_799_cast_fp16")]; + tensor hidden_states_41_cast_fp16 = add(x = pretrained_out_399_cast_fp16, y = lora_out_799_cast_fp16)[name = tensor("hidden_states_41_cast_fp16")]; + tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_8080 = const()[name = tensor("op_8080"), val = tensor(3)]; + tensor var_8087 = const()[name = tensor("op_8087"), val = tensor(1)]; + tensor var_8088 = const()[name = tensor("op_8088"), val = tensor(true)]; + tensor var_8100 = const()[name = tensor("op_8100"), val = tensor([1])]; + tensor channels_mean_121_cast_fp16 = reduce_mean(axes = var_8100, keep_dims = var_8088, x = inputs_121_cast_fp16)[name = tensor("channels_mean_121_cast_fp16")]; + tensor zero_mean_121_cast_fp16 = sub(x = inputs_121_cast_fp16, y = channels_mean_121_cast_fp16)[name = tensor("zero_mean_121_cast_fp16")]; + tensor zero_mean_sq_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = zero_mean_121_cast_fp16)[name = tensor("zero_mean_sq_121_cast_fp16")]; + tensor var_8104 = const()[name = tensor("op_8104"), val = tensor([1])]; + tensor var_8105_cast_fp16 = reduce_mean(axes = var_8104, keep_dims = var_8088, x = zero_mean_sq_121_cast_fp16)[name = tensor("op_8105_cast_fp16")]; + tensor var_8106_to_fp16 = const()[name = tensor("op_8106_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_8107_cast_fp16 = add(x = var_8105_cast_fp16, y = var_8106_to_fp16)[name = tensor("op_8107_cast_fp16")]; + tensor denom_121_epsilon_0 = const()[name = tensor("denom_121_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_121_cast_fp16 = rsqrt(epsilon = denom_121_epsilon_0, x = var_8107_cast_fp16)[name = tensor("denom_121_cast_fp16")]; + tensor out_121_cast_fp16 = mul(x = zero_mean_121_cast_fp16, y = denom_121_cast_fp16)[name = tensor("out_121_cast_fp16")]; + tensor obj_281_gamma_0_to_fp16 = const()[name = tensor("obj_281_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419971008)))]; + tensor obj_281_beta_0_to_fp16 = const()[name = tensor("obj_281_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419973632)))]; + tensor obj_281_epsilon_0_to_fp16 = const()[name = tensor("obj_281_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_281_cast_fp16 = batch_norm(beta = obj_281_beta_0_to_fp16, epsilon = obj_281_epsilon_0_to_fp16, gamma = obj_281_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor("obj_281_cast_fp16")]; + tensor var_8125 = const()[name = tensor("op_8125"), val = tensor([1, 1])]; + tensor var_8127 = const()[name = tensor("op_8127"), val = tensor([1, 1])]; + tensor pretrained_out_401_pad_type_0 = const()[name = tensor("pretrained_out_401_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_401_pad_0 = const()[name = tensor("pretrained_out_401_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419976256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420795520))), name = tensor("layers_20_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_20_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420795648)))]; + tensor pretrained_out_401_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_8127, groups = var_8087, pad = pretrained_out_401_pad_0, pad_type = pretrained_out_401_pad_type_0, strides = var_8125, weight = layers_20_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_281_cast_fp16)[name = tensor("pretrained_out_401_cast_fp16")]; + tensor var_8131 = const()[name = tensor("op_8131"), val = tensor([1, 1])]; + tensor var_8133 = const()[name = tensor("op_8133"), val = tensor([1, 1])]; + tensor input_601_pad_type_0 = const()[name = tensor("input_601_pad_type_0"), val = tensor("custom")]; + tensor input_601_pad_0 = const()[name = tensor("input_601_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420798272)))]; + tensor input_601_cast_fp16 = conv(dilations = var_8133, groups = var_8087, pad = input_601_pad_0, pad_type = input_601_pad_type_0, strides = var_8131, weight = layers_20_self_attn_q_proj_loraA_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("input_601_cast_fp16")]; + tensor var_8137 = const()[name = tensor("op_8137"), val = tensor([1, 1])]; + tensor var_8139 = const()[name = tensor("op_8139"), val = tensor([1, 1])]; + tensor lora_out_801_pad_type_0 = const()[name = tensor("lora_out_801_pad_type_0"), val = tensor("custom")]; + tensor lora_out_801_pad_0 = const()[name = tensor("lora_out_801_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_803_weight_0_to_fp16 = const()[name = tensor("lora_out_803_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420839296)))]; + tensor lora_out_803_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8139, groups = var_8087, pad = lora_out_801_pad_0, pad_type = lora_out_801_pad_type_0, strides = var_8137, weight = lora_out_803_weight_0_to_fp16, x = input_601_cast_fp16)[name = tensor("lora_out_803_cast_fp16")]; + tensor query_81_cast_fp16 = add(x = pretrained_out_401_cast_fp16, y = lora_out_803_cast_fp16)[name = tensor("query_81_cast_fp16")]; + tensor var_8149 = const()[name = tensor("op_8149"), val = tensor([1, 1])]; + tensor var_8151 = const()[name = tensor("op_8151"), val = tensor([1, 1])]; + tensor pretrained_out_403_pad_type_0 = const()[name = tensor("pretrained_out_403_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_403_pad_0 = const()[name = tensor("pretrained_out_403_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420880320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421699584))), name = tensor("layers_20_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_403_cast_fp16 = conv(dilations = var_8151, groups = var_8087, pad = pretrained_out_403_pad_0, pad_type = pretrained_out_403_pad_type_0, strides = var_8149, weight = layers_20_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_281_cast_fp16)[name = tensor("pretrained_out_403_cast_fp16")]; + tensor var_8155 = const()[name = tensor("op_8155"), val = tensor([1, 1])]; + tensor var_8157 = const()[name = tensor("op_8157"), val = tensor([1, 1])]; + tensor input_603_pad_type_0 = const()[name = tensor("input_603_pad_type_0"), val = tensor("custom")]; + tensor input_603_pad_0 = const()[name = tensor("input_603_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421699712)))]; + tensor input_603_cast_fp16 = conv(dilations = var_8157, groups = var_8087, pad = input_603_pad_0, pad_type = input_603_pad_type_0, strides = var_8155, weight = layers_20_self_attn_k_proj_loraA_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("input_603_cast_fp16")]; + tensor var_8161 = const()[name = tensor("op_8161"), val = tensor([1, 1])]; + tensor var_8163 = const()[name = tensor("op_8163"), val = tensor([1, 1])]; + tensor lora_out_805_pad_type_0 = const()[name = tensor("lora_out_805_pad_type_0"), val = tensor("custom")]; + tensor lora_out_805_pad_0 = const()[name = tensor("lora_out_805_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_807_weight_0_to_fp16 = const()[name = tensor("lora_out_807_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421740736)))]; + tensor lora_out_807_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8163, groups = var_8087, pad = lora_out_805_pad_0, pad_type = lora_out_805_pad_type_0, strides = var_8161, weight = lora_out_807_weight_0_to_fp16, x = input_603_cast_fp16)[name = tensor("lora_out_807_cast_fp16")]; + tensor current_key_41_cast_fp16 = add(x = pretrained_out_403_cast_fp16, y = lora_out_807_cast_fp16)[name = tensor("current_key_41_cast_fp16")]; + tensor var_8174 = const()[name = tensor("op_8174"), val = tensor([1, 1])]; + tensor var_8176 = const()[name = tensor("op_8176"), val = tensor([1, 1])]; + tensor pretrained_out_405_pad_type_0 = const()[name = tensor("pretrained_out_405_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_405_pad_0 = const()[name = tensor("pretrained_out_405_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(421781760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422601024))), name = tensor("layers_20_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_20_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422601152)))]; + tensor pretrained_out_405_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_8176, groups = var_8087, pad = pretrained_out_405_pad_0, pad_type = pretrained_out_405_pad_type_0, strides = var_8174, weight = layers_20_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_281_cast_fp16)[name = tensor("pretrained_out_405_cast_fp16")]; + tensor var_8180 = const()[name = tensor("op_8180"), val = tensor([1, 1])]; + tensor var_8182 = const()[name = tensor("op_8182"), val = tensor([1, 1])]; + tensor input_605_pad_type_0 = const()[name = tensor("input_605_pad_type_0"), val = tensor("custom")]; + tensor input_605_pad_0 = const()[name = tensor("input_605_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422603776)))]; + tensor input_605_cast_fp16 = conv(dilations = var_8182, groups = var_8087, pad = input_605_pad_0, pad_type = input_605_pad_type_0, strides = var_8180, weight = layers_20_self_attn_v_proj_loraA_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("input_605_cast_fp16")]; + tensor var_8186 = const()[name = tensor("op_8186"), val = tensor([1, 1])]; + tensor var_8188 = const()[name = tensor("op_8188"), val = tensor([1, 1])]; + tensor lora_out_809_pad_type_0 = const()[name = tensor("lora_out_809_pad_type_0"), val = tensor("custom")]; + tensor lora_out_809_pad_0 = const()[name = tensor("lora_out_809_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_811_weight_0_to_fp16 = const()[name = tensor("lora_out_811_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422644800)))]; + tensor lora_out_811_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8188, groups = var_8087, pad = lora_out_809_pad_0, pad_type = lora_out_809_pad_type_0, strides = var_8186, weight = lora_out_811_weight_0_to_fp16, x = input_605_cast_fp16)[name = tensor("lora_out_811_cast_fp16")]; + tensor current_value_41_cast_fp16 = add(x = pretrained_out_405_cast_fp16, y = lora_out_811_cast_fp16)[name = tensor("current_value_41_cast_fp16")]; + tensor var_8198_cast_fp16 = mul(x = current_key_41_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_8198_cast_fp16")]; + tensor var_8200_cast_fp16 = mul(x = var_103_cast_fp16_20, y = var_295_cast_fp16)[name = tensor("op_8200_cast_fp16")]; + tensor key_81_cast_fp16 = add(x = var_8198_cast_fp16, y = var_8200_cast_fp16)[name = tensor("key_81_cast_fp16")]; + tensor var_8202_cast_fp16 = mul(x = current_value_41_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_8202_cast_fp16")]; + tensor var_8204_cast_fp16 = mul(x = var_138_cast_fp16_20, y = var_295_cast_fp16)[name = tensor("op_8204_cast_fp16")]; + tensor value_81_cast_fp16 = add(x = var_8202_cast_fp16, y = var_8204_cast_fp16)[name = tensor("value_81_cast_fp16")]; + tensor var_8207 = const()[name = tensor("op_8207"), val = tensor([1, 20, 64, -1])]; + tensor var_8208_cast_fp16 = reshape(shape = var_8207, x = query_81_cast_fp16)[name = tensor("op_8208_cast_fp16")]; + tensor var_8209_to_fp16 = const()[name = tensor("op_8209_to_fp16"), val = tensor(0x1p-3)]; + tensor var_8210_cast_fp16 = mul(x = var_8208_cast_fp16, y = var_8209_to_fp16)[name = tensor("op_8210_cast_fp16")]; + tensor var_8211 = const()[name = tensor("op_8211"), val = tensor([1, 20, 64, -1])]; + tensor var_8212_cast_fp16 = reshape(shape = var_8211, x = key_81_cast_fp16)[name = tensor("op_8212_cast_fp16")]; + tensor mh_w_121_transpose_x_0 = const()[name = tensor("mh_w_121_transpose_x_0"), val = tensor(true)]; + tensor mh_w_121_transpose_y_0 = const()[name = tensor("mh_w_121_transpose_y_0"), val = tensor(false)]; + tensor mh_w_121_cast_fp16 = matmul(transpose_x = mh_w_121_transpose_x_0, transpose_y = mh_w_121_transpose_y_0, x = var_8210_cast_fp16, y = var_8212_cast_fp16)[name = tensor("mh_w_121_cast_fp16")]; + tensor mh_w_123_cast_fp16 = add(x = mh_w_121_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_123_cast_fp16")]; + tensor var_8220_cast_fp16 = softmax(axis = var_8080, x = mh_w_123_cast_fp16)[name = tensor("op_8220_cast_fp16")]; + tensor var_8221 = const()[name = tensor("op_8221"), val = tensor([1, 20, 64, -1])]; + tensor var_8222_cast_fp16 = reshape(shape = var_8221, x = value_81_cast_fp16)[name = tensor("op_8222_cast_fp16")]; + tensor attn_81_transpose_x_0 = const()[name = tensor("attn_81_transpose_x_0"), val = tensor(false)]; + tensor attn_81_transpose_y_0 = const()[name = tensor("attn_81_transpose_y_0"), val = tensor(true)]; + tensor attn_81_cast_fp16 = matmul(transpose_x = attn_81_transpose_x_0, transpose_y = attn_81_transpose_y_0, x = var_8222_cast_fp16, y = var_8220_cast_fp16)[name = tensor("attn_81_cast_fp16")]; + tensor var_8225 = const()[name = tensor("op_8225"), val = tensor([1, 1280, 1, -1])]; + tensor input_607_cast_fp16 = reshape(shape = var_8225, x = attn_81_cast_fp16)[name = tensor("input_607_cast_fp16")]; + tensor var_8232 = const()[name = tensor("op_8232"), val = tensor([1, 1])]; + tensor var_8234 = const()[name = tensor("op_8234"), val = tensor([1, 1])]; + tensor pretrained_out_407_pad_type_0 = const()[name = tensor("pretrained_out_407_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_407_pad_0 = const()[name = tensor("pretrained_out_407_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422685824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423505088))), name = tensor("layers_20_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_20_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423505216)))]; + tensor pretrained_out_407_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_8234, groups = var_8087, pad = pretrained_out_407_pad_0, pad_type = pretrained_out_407_pad_type_0, strides = var_8232, weight = layers_20_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_607_cast_fp16)[name = tensor("pretrained_out_407_cast_fp16")]; + tensor var_8238 = const()[name = tensor("op_8238"), val = tensor([1, 1])]; + tensor var_8240 = const()[name = tensor("op_8240"), val = tensor([1, 1])]; + tensor input_609_pad_type_0 = const()[name = tensor("input_609_pad_type_0"), val = tensor("custom")]; + tensor input_609_pad_0 = const()[name = tensor("input_609_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423507840)))]; + tensor input_609_cast_fp16 = conv(dilations = var_8240, groups = var_8087, pad = input_609_pad_0, pad_type = input_609_pad_type_0, strides = var_8238, weight = layers_20_self_attn_o_proj_loraA_weight_to_fp16, x = input_607_cast_fp16)[name = tensor("input_609_cast_fp16")]; + tensor var_8244 = const()[name = tensor("op_8244"), val = tensor([1, 1])]; + tensor var_8246 = const()[name = tensor("op_8246"), val = tensor([1, 1])]; + tensor lora_out_813_pad_type_0 = const()[name = tensor("lora_out_813_pad_type_0"), val = tensor("custom")]; + tensor lora_out_813_pad_0 = const()[name = tensor("lora_out_813_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_815_weight_0_to_fp16 = const()[name = tensor("lora_out_815_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423548864)))]; + tensor lora_out_815_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8246, groups = var_8087, pad = lora_out_813_pad_0, pad_type = lora_out_813_pad_type_0, strides = var_8244, weight = lora_out_815_weight_0_to_fp16, x = input_609_cast_fp16)[name = tensor("lora_out_815_cast_fp16")]; + tensor obj_287_cast_fp16 = add(x = pretrained_out_407_cast_fp16, y = lora_out_815_cast_fp16)[name = tensor("obj_287_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_287_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; + tensor var_8259 = const()[name = tensor("op_8259"), val = tensor([1])]; + tensor channels_mean_123_cast_fp16 = reduce_mean(axes = var_8259, keep_dims = var_8088, x = inputs_123_cast_fp16)[name = tensor("channels_mean_123_cast_fp16")]; + tensor zero_mean_123_cast_fp16 = sub(x = inputs_123_cast_fp16, y = channels_mean_123_cast_fp16)[name = tensor("zero_mean_123_cast_fp16")]; + tensor zero_mean_sq_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = zero_mean_123_cast_fp16)[name = tensor("zero_mean_sq_123_cast_fp16")]; + tensor var_8263 = const()[name = tensor("op_8263"), val = tensor([1])]; + tensor var_8264_cast_fp16 = reduce_mean(axes = var_8263, keep_dims = var_8088, x = zero_mean_sq_123_cast_fp16)[name = tensor("op_8264_cast_fp16")]; + tensor var_8265_to_fp16 = const()[name = tensor("op_8265_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_8266_cast_fp16 = add(x = var_8264_cast_fp16, y = var_8265_to_fp16)[name = tensor("op_8266_cast_fp16")]; + tensor denom_123_epsilon_0 = const()[name = tensor("denom_123_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_123_cast_fp16 = rsqrt(epsilon = denom_123_epsilon_0, x = var_8266_cast_fp16)[name = tensor("denom_123_cast_fp16")]; + tensor out_123_cast_fp16 = mul(x = zero_mean_123_cast_fp16, y = denom_123_cast_fp16)[name = tensor("out_123_cast_fp16")]; + tensor obj_289_gamma_0_to_fp16 = const()[name = tensor("obj_289_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423589888)))]; + tensor obj_289_beta_0_to_fp16 = const()[name = tensor("obj_289_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423592512)))]; + tensor obj_289_epsilon_0_to_fp16 = const()[name = tensor("obj_289_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_289_cast_fp16 = batch_norm(beta = obj_289_beta_0_to_fp16, epsilon = obj_289_epsilon_0_to_fp16, gamma = obj_289_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor("obj_289_cast_fp16")]; + tensor var_8284 = const()[name = tensor("op_8284"), val = tensor([1, 1])]; + tensor var_8286 = const()[name = tensor("op_8286"), val = tensor([1, 1])]; + tensor pretrained_out_409_pad_type_0 = const()[name = tensor("pretrained_out_409_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_409_pad_0 = const()[name = tensor("pretrained_out_409_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(423595136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424414400))), name = tensor("layers_20_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_20_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424414528)))]; + tensor pretrained_out_409_cast_fp16 = conv(bias = layers_20_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_8286, groups = var_8087, pad = pretrained_out_409_pad_0, pad_type = pretrained_out_409_pad_type_0, strides = var_8284, weight = layers_20_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_289_cast_fp16)[name = tensor("pretrained_out_409_cast_fp16")]; + tensor var_8290 = const()[name = tensor("op_8290"), val = tensor([1, 1])]; + tensor var_8292 = const()[name = tensor("op_8292"), val = tensor([1, 1])]; + tensor input_611_pad_type_0 = const()[name = tensor("input_611_pad_type_0"), val = tensor("custom")]; + tensor input_611_pad_0 = const()[name = tensor("input_611_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424417152)))]; + tensor input_611_cast_fp16 = conv(dilations = var_8292, groups = var_8087, pad = input_611_pad_0, pad_type = input_611_pad_type_0, strides = var_8290, weight = layers_20_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_289_cast_fp16)[name = tensor("input_611_cast_fp16")]; + tensor var_8296 = const()[name = tensor("op_8296"), val = tensor([1, 1])]; + tensor var_8298 = const()[name = tensor("op_8298"), val = tensor([1, 1])]; + tensor lora_out_817_pad_type_0 = const()[name = tensor("lora_out_817_pad_type_0"), val = tensor("custom")]; + tensor lora_out_817_pad_0 = const()[name = tensor("lora_out_817_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_819_weight_0_to_fp16 = const()[name = tensor("lora_out_819_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424458176)))]; + tensor lora_out_819_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8298, groups = var_8087, pad = lora_out_817_pad_0, pad_type = lora_out_817_pad_type_0, strides = var_8296, weight = lora_out_819_weight_0_to_fp16, x = input_611_cast_fp16)[name = tensor("lora_out_819_cast_fp16")]; + tensor query_83_cast_fp16 = add(x = pretrained_out_409_cast_fp16, y = lora_out_819_cast_fp16)[name = tensor("query_83_cast_fp16")]; + tensor var_8308 = const()[name = tensor("op_8308"), val = tensor([1, 1])]; + tensor var_8310 = const()[name = tensor("op_8310"), val = tensor([1, 1])]; + tensor pretrained_out_411_pad_type_0 = const()[name = tensor("pretrained_out_411_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_411_pad_0 = const()[name = tensor("pretrained_out_411_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424499200))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425318464))), name = tensor("layers_20_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_411_cast_fp16 = conv(dilations = var_8310, groups = var_8087, pad = pretrained_out_411_pad_0, pad_type = pretrained_out_411_pad_type_0, strides = var_8308, weight = layers_20_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_411_cast_fp16")]; + tensor var_8314 = const()[name = tensor("op_8314"), val = tensor([1, 1])]; + tensor var_8316 = const()[name = tensor("op_8316"), val = tensor([1, 1])]; + tensor input_613_pad_type_0 = const()[name = tensor("input_613_pad_type_0"), val = tensor("custom")]; + tensor input_613_pad_0 = const()[name = tensor("input_613_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425318592)))]; + tensor input_613_cast_fp16 = conv(dilations = var_8316, groups = var_8087, pad = input_613_pad_0, pad_type = input_613_pad_type_0, strides = var_8314, weight = layers_20_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_613_cast_fp16")]; + tensor var_8320 = const()[name = tensor("op_8320"), val = tensor([1, 1])]; + tensor var_8322 = const()[name = tensor("op_8322"), val = tensor([1, 1])]; + tensor lora_out_821_pad_type_0 = const()[name = tensor("lora_out_821_pad_type_0"), val = tensor("custom")]; + tensor lora_out_821_pad_0 = const()[name = tensor("lora_out_821_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_823_weight_0_to_fp16 = const()[name = tensor("lora_out_823_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425359616)))]; + tensor lora_out_823_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8322, groups = var_8087, pad = lora_out_821_pad_0, pad_type = lora_out_821_pad_type_0, strides = var_8320, weight = lora_out_823_weight_0_to_fp16, x = input_613_cast_fp16)[name = tensor("lora_out_823_cast_fp16")]; + tensor key_83_cast_fp16 = add(x = pretrained_out_411_cast_fp16, y = lora_out_823_cast_fp16)[name = tensor("key_83_cast_fp16")]; + tensor var_8333 = const()[name = tensor("op_8333"), val = tensor([1, 1])]; + tensor var_8335 = const()[name = tensor("op_8335"), val = tensor([1, 1])]; + tensor pretrained_out_413_pad_type_0 = const()[name = tensor("pretrained_out_413_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_413_pad_0 = const()[name = tensor("pretrained_out_413_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425400640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426219904))), name = tensor("layers_20_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_20_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426220032)))]; + tensor pretrained_out_413_cast_fp16 = conv(bias = layers_20_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_8335, groups = var_8087, pad = pretrained_out_413_pad_0, pad_type = pretrained_out_413_pad_type_0, strides = var_8333, weight = layers_20_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_413_cast_fp16")]; + tensor var_8339 = const()[name = tensor("op_8339"), val = tensor([1, 1])]; + tensor var_8341 = const()[name = tensor("op_8341"), val = tensor([1, 1])]; + tensor input_615_pad_type_0 = const()[name = tensor("input_615_pad_type_0"), val = tensor("custom")]; + tensor input_615_pad_0 = const()[name = tensor("input_615_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426222656)))]; + tensor input_615_cast_fp16 = conv(dilations = var_8341, groups = var_8087, pad = input_615_pad_0, pad_type = input_615_pad_type_0, strides = var_8339, weight = layers_20_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_615_cast_fp16")]; + tensor var_8345 = const()[name = tensor("op_8345"), val = tensor([1, 1])]; + tensor var_8347 = const()[name = tensor("op_8347"), val = tensor([1, 1])]; + tensor lora_out_825_pad_type_0 = const()[name = tensor("lora_out_825_pad_type_0"), val = tensor("custom")]; + tensor lora_out_825_pad_0 = const()[name = tensor("lora_out_825_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_827_weight_0_to_fp16 = const()[name = tensor("lora_out_827_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426263680)))]; + tensor lora_out_827_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8347, groups = var_8087, pad = lora_out_825_pad_0, pad_type = lora_out_825_pad_type_0, strides = var_8345, weight = lora_out_827_weight_0_to_fp16, x = input_615_cast_fp16)[name = tensor("lora_out_827_cast_fp16")]; + tensor value_83_cast_fp16 = add(x = pretrained_out_413_cast_fp16, y = lora_out_827_cast_fp16)[name = tensor("value_83_cast_fp16")]; + tensor var_8354 = const()[name = tensor("op_8354"), val = tensor([1, 20, 64, -1])]; + tensor var_8355_cast_fp16 = reshape(shape = var_8354, x = query_83_cast_fp16)[name = tensor("op_8355_cast_fp16")]; + tensor var_8356_to_fp16 = const()[name = tensor("op_8356_to_fp16"), val = tensor(0x1p-3)]; + tensor var_8357_cast_fp16 = mul(x = var_8355_cast_fp16, y = var_8356_to_fp16)[name = tensor("op_8357_cast_fp16")]; + tensor var_8358 = const()[name = tensor("op_8358"), val = tensor([1, 20, 64, -1])]; + tensor var_8359_cast_fp16 = reshape(shape = var_8358, x = key_83_cast_fp16)[name = tensor("op_8359_cast_fp16")]; + tensor mh_w_125_transpose_x_0 = const()[name = tensor("mh_w_125_transpose_x_0"), val = tensor(true)]; + tensor mh_w_125_transpose_y_0 = const()[name = tensor("mh_w_125_transpose_y_0"), val = tensor(false)]; + tensor mh_w_125_cast_fp16 = matmul(transpose_x = mh_w_125_transpose_x_0, transpose_y = mh_w_125_transpose_y_0, x = var_8357_cast_fp16, y = var_8359_cast_fp16)[name = tensor("mh_w_125_cast_fp16")]; + tensor obj_293_cast_fp16 = softmax(axis = var_8080, x = mh_w_125_cast_fp16)[name = tensor("obj_293_cast_fp16")]; + tensor var_8363 = const()[name = tensor("op_8363"), val = tensor([1, 20, 64, -1])]; + tensor var_8364_cast_fp16 = reshape(shape = var_8363, x = value_83_cast_fp16)[name = tensor("op_8364_cast_fp16")]; + tensor attn_83_transpose_x_0 = const()[name = tensor("attn_83_transpose_x_0"), val = tensor(false)]; + tensor attn_83_transpose_y_0 = const()[name = tensor("attn_83_transpose_y_0"), val = tensor(true)]; + tensor attn_83_cast_fp16 = matmul(transpose_x = attn_83_transpose_x_0, transpose_y = attn_83_transpose_y_0, x = var_8364_cast_fp16, y = obj_293_cast_fp16)[name = tensor("attn_83_cast_fp16")]; + tensor var_8367 = const()[name = tensor("op_8367"), val = tensor([1, 1280, 1, -1])]; + tensor input_617_cast_fp16 = reshape(shape = var_8367, x = attn_83_cast_fp16)[name = tensor("input_617_cast_fp16")]; + tensor var_8374 = const()[name = tensor("op_8374"), val = tensor([1, 1])]; + tensor var_8376 = const()[name = tensor("op_8376"), val = tensor([1, 1])]; + tensor pretrained_out_415_pad_type_0 = const()[name = tensor("pretrained_out_415_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_415_pad_0 = const()[name = tensor("pretrained_out_415_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426304704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427123968))), name = tensor("layers_20_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_20_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427124096)))]; + tensor pretrained_out_415_cast_fp16 = conv(bias = layers_20_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_8376, groups = var_8087, pad = pretrained_out_415_pad_0, pad_type = pretrained_out_415_pad_type_0, strides = var_8374, weight = layers_20_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_617_cast_fp16)[name = tensor("pretrained_out_415_cast_fp16")]; + tensor var_8380 = const()[name = tensor("op_8380"), val = tensor([1, 1])]; + tensor var_8382 = const()[name = tensor("op_8382"), val = tensor([1, 1])]; + tensor input_619_pad_type_0 = const()[name = tensor("input_619_pad_type_0"), val = tensor("custom")]; + tensor input_619_pad_0 = const()[name = tensor("input_619_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427126720)))]; + tensor input_619_cast_fp16 = conv(dilations = var_8382, groups = var_8087, pad = input_619_pad_0, pad_type = input_619_pad_type_0, strides = var_8380, weight = layers_20_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_617_cast_fp16)[name = tensor("input_619_cast_fp16")]; + tensor var_8386 = const()[name = tensor("op_8386"), val = tensor([1, 1])]; + tensor var_8388 = const()[name = tensor("op_8388"), val = tensor([1, 1])]; + tensor lora_out_829_pad_type_0 = const()[name = tensor("lora_out_829_pad_type_0"), val = tensor("custom")]; + tensor lora_out_829_pad_0 = const()[name = tensor("lora_out_829_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_831_weight_0_to_fp16 = const()[name = tensor("lora_out_831_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427167744)))]; + tensor lora_out_831_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8388, groups = var_8087, pad = lora_out_829_pad_0, pad_type = lora_out_829_pad_type_0, strides = var_8386, weight = lora_out_831_weight_0_to_fp16, x = input_619_cast_fp16)[name = tensor("lora_out_831_cast_fp16")]; + tensor obj_291_cast_fp16 = add(x = pretrained_out_415_cast_fp16, y = lora_out_831_cast_fp16)[name = tensor("obj_291_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = obj_291_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; + tensor var_8397 = const()[name = tensor("op_8397"), val = tensor([1])]; + tensor channels_mean_125_cast_fp16 = reduce_mean(axes = var_8397, keep_dims = var_8088, x = inputs_125_cast_fp16)[name = tensor("channels_mean_125_cast_fp16")]; + tensor zero_mean_125_cast_fp16 = sub(x = inputs_125_cast_fp16, y = channels_mean_125_cast_fp16)[name = tensor("zero_mean_125_cast_fp16")]; + tensor zero_mean_sq_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = zero_mean_125_cast_fp16)[name = tensor("zero_mean_sq_125_cast_fp16")]; + tensor var_8401 = const()[name = tensor("op_8401"), val = tensor([1])]; + tensor var_8402_cast_fp16 = reduce_mean(axes = var_8401, keep_dims = var_8088, x = zero_mean_sq_125_cast_fp16)[name = tensor("op_8402_cast_fp16")]; + tensor var_8403_to_fp16 = const()[name = tensor("op_8403_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_8404_cast_fp16 = add(x = var_8402_cast_fp16, y = var_8403_to_fp16)[name = tensor("op_8404_cast_fp16")]; + tensor denom_125_epsilon_0 = const()[name = tensor("denom_125_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_125_cast_fp16 = rsqrt(epsilon = denom_125_epsilon_0, x = var_8404_cast_fp16)[name = tensor("denom_125_cast_fp16")]; + tensor out_125_cast_fp16 = mul(x = zero_mean_125_cast_fp16, y = denom_125_cast_fp16)[name = tensor("out_125_cast_fp16")]; + tensor input_621_gamma_0_to_fp16 = const()[name = tensor("input_621_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427208768)))]; + tensor input_621_beta_0_to_fp16 = const()[name = tensor("input_621_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427211392)))]; + tensor input_621_epsilon_0_to_fp16 = const()[name = tensor("input_621_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_621_cast_fp16 = batch_norm(beta = input_621_beta_0_to_fp16, epsilon = input_621_epsilon_0_to_fp16, gamma = input_621_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor("input_621_cast_fp16")]; + tensor var_8418 = const()[name = tensor("op_8418"), val = tensor([1, 1])]; + tensor var_8420 = const()[name = tensor("op_8420"), val = tensor([1, 1])]; + tensor pretrained_out_417_pad_type_0 = const()[name = tensor("pretrained_out_417_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_417_pad_0 = const()[name = tensor("pretrained_out_417_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427214016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430490880))), name = tensor("layers_20_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_20_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430491008)))]; + tensor pretrained_out_417_cast_fp16 = conv(bias = layers_20_fc1_pretrained_bias_to_fp16, dilations = var_8420, groups = var_8087, pad = pretrained_out_417_pad_0, pad_type = pretrained_out_417_pad_type_0, strides = var_8418, weight = layers_20_fc1_pretrained_weight_to_fp16_palettized, x = input_621_cast_fp16)[name = tensor("pretrained_out_417_cast_fp16")]; + tensor var_8424 = const()[name = tensor("op_8424"), val = tensor([1, 1])]; + tensor var_8426 = const()[name = tensor("op_8426"), val = tensor([1, 1])]; + tensor input_623_pad_type_0 = const()[name = tensor("input_623_pad_type_0"), val = tensor("custom")]; + tensor input_623_pad_0 = const()[name = tensor("input_623_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_20_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430501312)))]; + tensor input_623_cast_fp16 = conv(dilations = var_8426, groups = var_8087, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = var_8424, weight = layers_20_fc1_loraA_weight_to_fp16, x = input_621_cast_fp16)[name = tensor("input_623_cast_fp16")]; + tensor var_8430 = const()[name = tensor("op_8430"), val = tensor([1, 1])]; + tensor var_8432 = const()[name = tensor("op_8432"), val = tensor([1, 1])]; + tensor lora_out_833_pad_type_0 = const()[name = tensor("lora_out_833_pad_type_0"), val = tensor("custom")]; + tensor lora_out_833_pad_0 = const()[name = tensor("lora_out_833_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_835_weight_0_to_fp16 = const()[name = tensor("lora_out_835_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430542336)))]; + tensor lora_out_835_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_8432, groups = var_8087, pad = lora_out_833_pad_0, pad_type = lora_out_833_pad_type_0, strides = var_8430, weight = lora_out_835_weight_0_to_fp16, x = input_623_cast_fp16)[name = tensor("lora_out_835_cast_fp16")]; + tensor input_625_cast_fp16 = add(x = pretrained_out_417_cast_fp16, y = lora_out_835_cast_fp16)[name = tensor("input_625_cast_fp16")]; + tensor input_627_mode_0 = const()[name = tensor("input_627_mode_0"), val = tensor("EXACT")]; + tensor input_627_cast_fp16 = gelu(mode = input_627_mode_0, x = input_625_cast_fp16)[name = tensor("input_627_cast_fp16")]; + tensor var_8444 = const()[name = tensor("op_8444"), val = tensor([1, 1])]; + tensor var_8446 = const()[name = tensor("op_8446"), val = tensor([1, 1])]; + tensor pretrained_out_419_pad_type_0 = const()[name = tensor("pretrained_out_419_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_419_pad_0 = const()[name = tensor("pretrained_out_419_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(430706240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(433983104))), name = tensor("layers_20_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_20_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_20_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(433983232)))]; + tensor pretrained_out_419_cast_fp16 = conv(bias = layers_20_fc2_pretrained_bias_to_fp16, dilations = var_8446, groups = var_8087, pad = pretrained_out_419_pad_0, pad_type = pretrained_out_419_pad_type_0, strides = var_8444, weight = layers_20_fc2_pretrained_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = tensor("pretrained_out_419_cast_fp16")]; + tensor var_8450 = const()[name = tensor("op_8450"), val = tensor([1, 1])]; + tensor var_8452 = const()[name = tensor("op_8452"), val = tensor([1, 1])]; + tensor input_629_pad_type_0 = const()[name = tensor("input_629_pad_type_0"), val = tensor("custom")]; + tensor input_629_pad_0 = const()[name = tensor("input_629_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_20_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_20_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(433985856)))]; + tensor input_629_cast_fp16 = conv(dilations = var_8452, groups = var_8087, pad = input_629_pad_0, pad_type = input_629_pad_type_0, strides = var_8450, weight = layers_20_fc2_loraA_weight_to_fp16, x = input_627_cast_fp16)[name = tensor("input_629_cast_fp16")]; + tensor var_8456 = const()[name = tensor("op_8456"), val = tensor([1, 1])]; + tensor var_8458 = const()[name = tensor("op_8458"), val = tensor([1, 1])]; + tensor lora_out_837_pad_type_0 = const()[name = tensor("lora_out_837_pad_type_0"), val = tensor("custom")]; + tensor lora_out_837_pad_0 = const()[name = tensor("lora_out_837_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_839_weight_0_to_fp16 = const()[name = tensor("lora_out_839_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434149760)))]; + tensor lora_out_839_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8458, groups = var_8087, pad = lora_out_837_pad_0, pad_type = lora_out_837_pad_type_0, strides = var_8456, weight = lora_out_839_weight_0_to_fp16, x = input_629_cast_fp16)[name = tensor("lora_out_839_cast_fp16")]; + tensor hidden_states_43_cast_fp16 = add(x = pretrained_out_419_cast_fp16, y = lora_out_839_cast_fp16)[name = tensor("hidden_states_43_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; + tensor var_8474 = const()[name = tensor("op_8474"), val = tensor(3)]; + tensor var_8481 = const()[name = tensor("op_8481"), val = tensor(1)]; + tensor var_8482 = const()[name = tensor("op_8482"), val = tensor(true)]; + tensor var_8494 = const()[name = tensor("op_8494"), val = tensor([1])]; + tensor channels_mean_127_cast_fp16 = reduce_mean(axes = var_8494, keep_dims = var_8482, x = inputs_127_cast_fp16)[name = tensor("channels_mean_127_cast_fp16")]; + tensor zero_mean_127_cast_fp16 = sub(x = inputs_127_cast_fp16, y = channels_mean_127_cast_fp16)[name = tensor("zero_mean_127_cast_fp16")]; + tensor zero_mean_sq_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = zero_mean_127_cast_fp16)[name = tensor("zero_mean_sq_127_cast_fp16")]; + tensor var_8498 = const()[name = tensor("op_8498"), val = tensor([1])]; + tensor var_8499_cast_fp16 = reduce_mean(axes = var_8498, keep_dims = var_8482, x = zero_mean_sq_127_cast_fp16)[name = tensor("op_8499_cast_fp16")]; + tensor var_8500_to_fp16 = const()[name = tensor("op_8500_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_8501_cast_fp16 = add(x = var_8499_cast_fp16, y = var_8500_to_fp16)[name = tensor("op_8501_cast_fp16")]; + tensor denom_127_epsilon_0 = const()[name = tensor("denom_127_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_127_cast_fp16 = rsqrt(epsilon = denom_127_epsilon_0, x = var_8501_cast_fp16)[name = tensor("denom_127_cast_fp16")]; + tensor out_127_cast_fp16 = mul(x = zero_mean_127_cast_fp16, y = denom_127_cast_fp16)[name = tensor("out_127_cast_fp16")]; + tensor obj_295_gamma_0_to_fp16 = const()[name = tensor("obj_295_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434190784)))]; + tensor obj_295_beta_0_to_fp16 = const()[name = tensor("obj_295_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434193408)))]; + tensor obj_295_epsilon_0_to_fp16 = const()[name = tensor("obj_295_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_295_cast_fp16 = batch_norm(beta = obj_295_beta_0_to_fp16, epsilon = obj_295_epsilon_0_to_fp16, gamma = obj_295_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor("obj_295_cast_fp16")]; + tensor var_8519 = const()[name = tensor("op_8519"), val = tensor([1, 1])]; + tensor var_8521 = const()[name = tensor("op_8521"), val = tensor([1, 1])]; + tensor pretrained_out_421_pad_type_0 = const()[name = tensor("pretrained_out_421_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_421_pad_0 = const()[name = tensor("pretrained_out_421_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434196032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435015296))), name = tensor("layers_21_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_21_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435015424)))]; + tensor pretrained_out_421_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_8521, groups = var_8481, pad = pretrained_out_421_pad_0, pad_type = pretrained_out_421_pad_type_0, strides = var_8519, weight = layers_21_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_295_cast_fp16)[name = tensor("pretrained_out_421_cast_fp16")]; + tensor var_8525 = const()[name = tensor("op_8525"), val = tensor([1, 1])]; + tensor var_8527 = const()[name = tensor("op_8527"), val = tensor([1, 1])]; + tensor input_631_pad_type_0 = const()[name = tensor("input_631_pad_type_0"), val = tensor("custom")]; + tensor input_631_pad_0 = const()[name = tensor("input_631_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435018048)))]; + tensor input_631_cast_fp16 = conv(dilations = var_8527, groups = var_8481, pad = input_631_pad_0, pad_type = input_631_pad_type_0, strides = var_8525, weight = layers_21_self_attn_q_proj_loraA_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("input_631_cast_fp16")]; + tensor var_8531 = const()[name = tensor("op_8531"), val = tensor([1, 1])]; + tensor var_8533 = const()[name = tensor("op_8533"), val = tensor([1, 1])]; + tensor lora_out_841_pad_type_0 = const()[name = tensor("lora_out_841_pad_type_0"), val = tensor("custom")]; + tensor lora_out_841_pad_0 = const()[name = tensor("lora_out_841_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_843_weight_0_to_fp16 = const()[name = tensor("lora_out_843_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435059072)))]; + tensor lora_out_843_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8533, groups = var_8481, pad = lora_out_841_pad_0, pad_type = lora_out_841_pad_type_0, strides = var_8531, weight = lora_out_843_weight_0_to_fp16, x = input_631_cast_fp16)[name = tensor("lora_out_843_cast_fp16")]; + tensor query_85_cast_fp16 = add(x = pretrained_out_421_cast_fp16, y = lora_out_843_cast_fp16)[name = tensor("query_85_cast_fp16")]; + tensor var_8543 = const()[name = tensor("op_8543"), val = tensor([1, 1])]; + tensor var_8545 = const()[name = tensor("op_8545"), val = tensor([1, 1])]; + tensor pretrained_out_423_pad_type_0 = const()[name = tensor("pretrained_out_423_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_423_pad_0 = const()[name = tensor("pretrained_out_423_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435100096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435919360))), name = tensor("layers_21_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_423_cast_fp16 = conv(dilations = var_8545, groups = var_8481, pad = pretrained_out_423_pad_0, pad_type = pretrained_out_423_pad_type_0, strides = var_8543, weight = layers_21_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_295_cast_fp16)[name = tensor("pretrained_out_423_cast_fp16")]; + tensor var_8549 = const()[name = tensor("op_8549"), val = tensor([1, 1])]; + tensor var_8551 = const()[name = tensor("op_8551"), val = tensor([1, 1])]; + tensor input_633_pad_type_0 = const()[name = tensor("input_633_pad_type_0"), val = tensor("custom")]; + tensor input_633_pad_0 = const()[name = tensor("input_633_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435919488)))]; + tensor input_633_cast_fp16 = conv(dilations = var_8551, groups = var_8481, pad = input_633_pad_0, pad_type = input_633_pad_type_0, strides = var_8549, weight = layers_21_self_attn_k_proj_loraA_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("input_633_cast_fp16")]; + tensor var_8555 = const()[name = tensor("op_8555"), val = tensor([1, 1])]; + tensor var_8557 = const()[name = tensor("op_8557"), val = tensor([1, 1])]; + tensor lora_out_845_pad_type_0 = const()[name = tensor("lora_out_845_pad_type_0"), val = tensor("custom")]; + tensor lora_out_845_pad_0 = const()[name = tensor("lora_out_845_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_847_weight_0_to_fp16 = const()[name = tensor("lora_out_847_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435960512)))]; + tensor lora_out_847_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8557, groups = var_8481, pad = lora_out_845_pad_0, pad_type = lora_out_845_pad_type_0, strides = var_8555, weight = lora_out_847_weight_0_to_fp16, x = input_633_cast_fp16)[name = tensor("lora_out_847_cast_fp16")]; + tensor current_key_43_cast_fp16 = add(x = pretrained_out_423_cast_fp16, y = lora_out_847_cast_fp16)[name = tensor("current_key_43_cast_fp16")]; + tensor var_8568 = const()[name = tensor("op_8568"), val = tensor([1, 1])]; + tensor var_8570 = const()[name = tensor("op_8570"), val = tensor([1, 1])]; + tensor pretrained_out_425_pad_type_0 = const()[name = tensor("pretrained_out_425_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_425_pad_0 = const()[name = tensor("pretrained_out_425_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436001536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436820800))), name = tensor("layers_21_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_21_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436820928)))]; + tensor pretrained_out_425_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_8570, groups = var_8481, pad = pretrained_out_425_pad_0, pad_type = pretrained_out_425_pad_type_0, strides = var_8568, weight = layers_21_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_295_cast_fp16)[name = tensor("pretrained_out_425_cast_fp16")]; + tensor var_8574 = const()[name = tensor("op_8574"), val = tensor([1, 1])]; + tensor var_8576 = const()[name = tensor("op_8576"), val = tensor([1, 1])]; + tensor input_635_pad_type_0 = const()[name = tensor("input_635_pad_type_0"), val = tensor("custom")]; + tensor input_635_pad_0 = const()[name = tensor("input_635_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436823552)))]; + tensor input_635_cast_fp16 = conv(dilations = var_8576, groups = var_8481, pad = input_635_pad_0, pad_type = input_635_pad_type_0, strides = var_8574, weight = layers_21_self_attn_v_proj_loraA_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("input_635_cast_fp16")]; + tensor var_8580 = const()[name = tensor("op_8580"), val = tensor([1, 1])]; + tensor var_8582 = const()[name = tensor("op_8582"), val = tensor([1, 1])]; + tensor lora_out_849_pad_type_0 = const()[name = tensor("lora_out_849_pad_type_0"), val = tensor("custom")]; + tensor lora_out_849_pad_0 = const()[name = tensor("lora_out_849_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_851_weight_0_to_fp16 = const()[name = tensor("lora_out_851_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436864576)))]; + tensor lora_out_851_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8582, groups = var_8481, pad = lora_out_849_pad_0, pad_type = lora_out_849_pad_type_0, strides = var_8580, weight = lora_out_851_weight_0_to_fp16, x = input_635_cast_fp16)[name = tensor("lora_out_851_cast_fp16")]; + tensor current_value_43_cast_fp16 = add(x = pretrained_out_425_cast_fp16, y = lora_out_851_cast_fp16)[name = tensor("current_value_43_cast_fp16")]; + tensor var_8592_cast_fp16 = mul(x = current_key_43_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_8592_cast_fp16")]; + tensor var_8594_cast_fp16 = mul(x = var_103_cast_fp16_21, y = var_295_cast_fp16)[name = tensor("op_8594_cast_fp16")]; + tensor key_85_cast_fp16 = add(x = var_8592_cast_fp16, y = var_8594_cast_fp16)[name = tensor("key_85_cast_fp16")]; + tensor var_8596_cast_fp16 = mul(x = current_value_43_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_8596_cast_fp16")]; + tensor var_8598_cast_fp16 = mul(x = var_138_cast_fp16_21, y = var_295_cast_fp16)[name = tensor("op_8598_cast_fp16")]; + tensor value_85_cast_fp16 = add(x = var_8596_cast_fp16, y = var_8598_cast_fp16)[name = tensor("value_85_cast_fp16")]; + tensor var_8601 = const()[name = tensor("op_8601"), val = tensor([1, 20, 64, -1])]; + tensor var_8602_cast_fp16 = reshape(shape = var_8601, x = query_85_cast_fp16)[name = tensor("op_8602_cast_fp16")]; + tensor var_8603_to_fp16 = const()[name = tensor("op_8603_to_fp16"), val = tensor(0x1p-3)]; + tensor var_8604_cast_fp16 = mul(x = var_8602_cast_fp16, y = var_8603_to_fp16)[name = tensor("op_8604_cast_fp16")]; + tensor var_8605 = const()[name = tensor("op_8605"), val = tensor([1, 20, 64, -1])]; + tensor var_8606_cast_fp16 = reshape(shape = var_8605, x = key_85_cast_fp16)[name = tensor("op_8606_cast_fp16")]; + tensor mh_w_127_transpose_x_0 = const()[name = tensor("mh_w_127_transpose_x_0"), val = tensor(true)]; + tensor mh_w_127_transpose_y_0 = const()[name = tensor("mh_w_127_transpose_y_0"), val = tensor(false)]; + tensor mh_w_127_cast_fp16 = matmul(transpose_x = mh_w_127_transpose_x_0, transpose_y = mh_w_127_transpose_y_0, x = var_8604_cast_fp16, y = var_8606_cast_fp16)[name = tensor("mh_w_127_cast_fp16")]; + tensor mh_w_129_cast_fp16 = add(x = mh_w_127_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_129_cast_fp16")]; + tensor var_8614_cast_fp16 = softmax(axis = var_8474, x = mh_w_129_cast_fp16)[name = tensor("op_8614_cast_fp16")]; + tensor var_8615 = const()[name = tensor("op_8615"), val = tensor([1, 20, 64, -1])]; + tensor var_8616_cast_fp16 = reshape(shape = var_8615, x = value_85_cast_fp16)[name = tensor("op_8616_cast_fp16")]; + tensor attn_85_transpose_x_0 = const()[name = tensor("attn_85_transpose_x_0"), val = tensor(false)]; + tensor attn_85_transpose_y_0 = const()[name = tensor("attn_85_transpose_y_0"), val = tensor(true)]; + tensor attn_85_cast_fp16 = matmul(transpose_x = attn_85_transpose_x_0, transpose_y = attn_85_transpose_y_0, x = var_8616_cast_fp16, y = var_8614_cast_fp16)[name = tensor("attn_85_cast_fp16")]; + tensor var_8619 = const()[name = tensor("op_8619"), val = tensor([1, 1280, 1, -1])]; + tensor input_637_cast_fp16 = reshape(shape = var_8619, x = attn_85_cast_fp16)[name = tensor("input_637_cast_fp16")]; + tensor var_8626 = const()[name = tensor("op_8626"), val = tensor([1, 1])]; + tensor var_8628 = const()[name = tensor("op_8628"), val = tensor([1, 1])]; + tensor pretrained_out_427_pad_type_0 = const()[name = tensor("pretrained_out_427_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_427_pad_0 = const()[name = tensor("pretrained_out_427_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436905600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437724864))), name = tensor("layers_21_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_21_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437724992)))]; + tensor pretrained_out_427_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_8628, groups = var_8481, pad = pretrained_out_427_pad_0, pad_type = pretrained_out_427_pad_type_0, strides = var_8626, weight = layers_21_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = tensor("pretrained_out_427_cast_fp16")]; + tensor var_8632 = const()[name = tensor("op_8632"), val = tensor([1, 1])]; + tensor var_8634 = const()[name = tensor("op_8634"), val = tensor([1, 1])]; + tensor input_639_pad_type_0 = const()[name = tensor("input_639_pad_type_0"), val = tensor("custom")]; + tensor input_639_pad_0 = const()[name = tensor("input_639_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437727616)))]; + tensor input_639_cast_fp16 = conv(dilations = var_8634, groups = var_8481, pad = input_639_pad_0, pad_type = input_639_pad_type_0, strides = var_8632, weight = layers_21_self_attn_o_proj_loraA_weight_to_fp16, x = input_637_cast_fp16)[name = tensor("input_639_cast_fp16")]; + tensor var_8638 = const()[name = tensor("op_8638"), val = tensor([1, 1])]; + tensor var_8640 = const()[name = tensor("op_8640"), val = tensor([1, 1])]; + tensor lora_out_853_pad_type_0 = const()[name = tensor("lora_out_853_pad_type_0"), val = tensor("custom")]; + tensor lora_out_853_pad_0 = const()[name = tensor("lora_out_853_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_855_weight_0_to_fp16 = const()[name = tensor("lora_out_855_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437768640)))]; + tensor lora_out_855_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8640, groups = var_8481, pad = lora_out_853_pad_0, pad_type = lora_out_853_pad_type_0, strides = var_8638, weight = lora_out_855_weight_0_to_fp16, x = input_639_cast_fp16)[name = tensor("lora_out_855_cast_fp16")]; + tensor obj_301_cast_fp16 = add(x = pretrained_out_427_cast_fp16, y = lora_out_855_cast_fp16)[name = tensor("obj_301_cast_fp16")]; + tensor inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = obj_301_cast_fp16)[name = tensor("inputs_129_cast_fp16")]; + tensor var_8653 = const()[name = tensor("op_8653"), val = tensor([1])]; + tensor channels_mean_129_cast_fp16 = reduce_mean(axes = var_8653, keep_dims = var_8482, x = inputs_129_cast_fp16)[name = tensor("channels_mean_129_cast_fp16")]; + tensor zero_mean_129_cast_fp16 = sub(x = inputs_129_cast_fp16, y = channels_mean_129_cast_fp16)[name = tensor("zero_mean_129_cast_fp16")]; + tensor zero_mean_sq_129_cast_fp16 = mul(x = zero_mean_129_cast_fp16, y = zero_mean_129_cast_fp16)[name = tensor("zero_mean_sq_129_cast_fp16")]; + tensor var_8657 = const()[name = tensor("op_8657"), val = tensor([1])]; + tensor var_8658_cast_fp16 = reduce_mean(axes = var_8657, keep_dims = var_8482, x = zero_mean_sq_129_cast_fp16)[name = tensor("op_8658_cast_fp16")]; + tensor var_8659_to_fp16 = const()[name = tensor("op_8659_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_8660_cast_fp16 = add(x = var_8658_cast_fp16, y = var_8659_to_fp16)[name = tensor("op_8660_cast_fp16")]; + tensor denom_129_epsilon_0 = const()[name = tensor("denom_129_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_129_cast_fp16 = rsqrt(epsilon = denom_129_epsilon_0, x = var_8660_cast_fp16)[name = tensor("denom_129_cast_fp16")]; + tensor out_129_cast_fp16 = mul(x = zero_mean_129_cast_fp16, y = denom_129_cast_fp16)[name = tensor("out_129_cast_fp16")]; + tensor obj_303_gamma_0_to_fp16 = const()[name = tensor("obj_303_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437809664)))]; + tensor obj_303_beta_0_to_fp16 = const()[name = tensor("obj_303_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437812288)))]; + tensor obj_303_epsilon_0_to_fp16 = const()[name = tensor("obj_303_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_303_cast_fp16 = batch_norm(beta = obj_303_beta_0_to_fp16, epsilon = obj_303_epsilon_0_to_fp16, gamma = obj_303_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_129_cast_fp16)[name = tensor("obj_303_cast_fp16")]; + tensor var_8678 = const()[name = tensor("op_8678"), val = tensor([1, 1])]; + tensor var_8680 = const()[name = tensor("op_8680"), val = tensor([1, 1])]; + tensor pretrained_out_429_pad_type_0 = const()[name = tensor("pretrained_out_429_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_429_pad_0 = const()[name = tensor("pretrained_out_429_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437814912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438634176))), name = tensor("layers_21_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_21_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438634304)))]; + tensor pretrained_out_429_cast_fp16 = conv(bias = layers_21_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_8680, groups = var_8481, pad = pretrained_out_429_pad_0, pad_type = pretrained_out_429_pad_type_0, strides = var_8678, weight = layers_21_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_303_cast_fp16)[name = tensor("pretrained_out_429_cast_fp16")]; + tensor var_8684 = const()[name = tensor("op_8684"), val = tensor([1, 1])]; + tensor var_8686 = const()[name = tensor("op_8686"), val = tensor([1, 1])]; + tensor input_641_pad_type_0 = const()[name = tensor("input_641_pad_type_0"), val = tensor("custom")]; + tensor input_641_pad_0 = const()[name = tensor("input_641_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438636928)))]; + tensor input_641_cast_fp16 = conv(dilations = var_8686, groups = var_8481, pad = input_641_pad_0, pad_type = input_641_pad_type_0, strides = var_8684, weight = layers_21_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_303_cast_fp16)[name = tensor("input_641_cast_fp16")]; + tensor var_8690 = const()[name = tensor("op_8690"), val = tensor([1, 1])]; + tensor var_8692 = const()[name = tensor("op_8692"), val = tensor([1, 1])]; + tensor lora_out_857_pad_type_0 = const()[name = tensor("lora_out_857_pad_type_0"), val = tensor("custom")]; + tensor lora_out_857_pad_0 = const()[name = tensor("lora_out_857_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_859_weight_0_to_fp16 = const()[name = tensor("lora_out_859_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438677952)))]; + tensor lora_out_859_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8692, groups = var_8481, pad = lora_out_857_pad_0, pad_type = lora_out_857_pad_type_0, strides = var_8690, weight = lora_out_859_weight_0_to_fp16, x = input_641_cast_fp16)[name = tensor("lora_out_859_cast_fp16")]; + tensor query_87_cast_fp16 = add(x = pretrained_out_429_cast_fp16, y = lora_out_859_cast_fp16)[name = tensor("query_87_cast_fp16")]; + tensor var_8702 = const()[name = tensor("op_8702"), val = tensor([1, 1])]; + tensor var_8704 = const()[name = tensor("op_8704"), val = tensor([1, 1])]; + tensor pretrained_out_431_pad_type_0 = const()[name = tensor("pretrained_out_431_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_431_pad_0 = const()[name = tensor("pretrained_out_431_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438718976))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439538240))), name = tensor("layers_21_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_431_cast_fp16 = conv(dilations = var_8704, groups = var_8481, pad = pretrained_out_431_pad_0, pad_type = pretrained_out_431_pad_type_0, strides = var_8702, weight = layers_21_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_431_cast_fp16")]; + tensor var_8708 = const()[name = tensor("op_8708"), val = tensor([1, 1])]; + tensor var_8710 = const()[name = tensor("op_8710"), val = tensor([1, 1])]; + tensor input_643_pad_type_0 = const()[name = tensor("input_643_pad_type_0"), val = tensor("custom")]; + tensor input_643_pad_0 = const()[name = tensor("input_643_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439538368)))]; + tensor input_643_cast_fp16 = conv(dilations = var_8710, groups = var_8481, pad = input_643_pad_0, pad_type = input_643_pad_type_0, strides = var_8708, weight = layers_21_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_643_cast_fp16")]; + tensor var_8714 = const()[name = tensor("op_8714"), val = tensor([1, 1])]; + tensor var_8716 = const()[name = tensor("op_8716"), val = tensor([1, 1])]; + tensor lora_out_861_pad_type_0 = const()[name = tensor("lora_out_861_pad_type_0"), val = tensor("custom")]; + tensor lora_out_861_pad_0 = const()[name = tensor("lora_out_861_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_863_weight_0_to_fp16 = const()[name = tensor("lora_out_863_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439579392)))]; + tensor lora_out_863_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8716, groups = var_8481, pad = lora_out_861_pad_0, pad_type = lora_out_861_pad_type_0, strides = var_8714, weight = lora_out_863_weight_0_to_fp16, x = input_643_cast_fp16)[name = tensor("lora_out_863_cast_fp16")]; + tensor key_87_cast_fp16 = add(x = pretrained_out_431_cast_fp16, y = lora_out_863_cast_fp16)[name = tensor("key_87_cast_fp16")]; + tensor var_8727 = const()[name = tensor("op_8727"), val = tensor([1, 1])]; + tensor var_8729 = const()[name = tensor("op_8729"), val = tensor([1, 1])]; + tensor pretrained_out_433_pad_type_0 = const()[name = tensor("pretrained_out_433_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_433_pad_0 = const()[name = tensor("pretrained_out_433_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439620416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440439680))), name = tensor("layers_21_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_21_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440439808)))]; + tensor pretrained_out_433_cast_fp16 = conv(bias = layers_21_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_8729, groups = var_8481, pad = pretrained_out_433_pad_0, pad_type = pretrained_out_433_pad_type_0, strides = var_8727, weight = layers_21_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_433_cast_fp16")]; + tensor var_8733 = const()[name = tensor("op_8733"), val = tensor([1, 1])]; + tensor var_8735 = const()[name = tensor("op_8735"), val = tensor([1, 1])]; + tensor input_645_pad_type_0 = const()[name = tensor("input_645_pad_type_0"), val = tensor("custom")]; + tensor input_645_pad_0 = const()[name = tensor("input_645_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440442432)))]; + tensor input_645_cast_fp16 = conv(dilations = var_8735, groups = var_8481, pad = input_645_pad_0, pad_type = input_645_pad_type_0, strides = var_8733, weight = layers_21_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_645_cast_fp16")]; + tensor var_8739 = const()[name = tensor("op_8739"), val = tensor([1, 1])]; + tensor var_8741 = const()[name = tensor("op_8741"), val = tensor([1, 1])]; + tensor lora_out_865_pad_type_0 = const()[name = tensor("lora_out_865_pad_type_0"), val = tensor("custom")]; + tensor lora_out_865_pad_0 = const()[name = tensor("lora_out_865_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_867_weight_0_to_fp16 = const()[name = tensor("lora_out_867_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440483456)))]; + tensor lora_out_867_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8741, groups = var_8481, pad = lora_out_865_pad_0, pad_type = lora_out_865_pad_type_0, strides = var_8739, weight = lora_out_867_weight_0_to_fp16, x = input_645_cast_fp16)[name = tensor("lora_out_867_cast_fp16")]; + tensor value_87_cast_fp16 = add(x = pretrained_out_433_cast_fp16, y = lora_out_867_cast_fp16)[name = tensor("value_87_cast_fp16")]; + tensor var_8748 = const()[name = tensor("op_8748"), val = tensor([1, 20, 64, -1])]; + tensor var_8749_cast_fp16 = reshape(shape = var_8748, x = query_87_cast_fp16)[name = tensor("op_8749_cast_fp16")]; + tensor var_8750_to_fp16 = const()[name = tensor("op_8750_to_fp16"), val = tensor(0x1p-3)]; + tensor var_8751_cast_fp16 = mul(x = var_8749_cast_fp16, y = var_8750_to_fp16)[name = tensor("op_8751_cast_fp16")]; + tensor var_8752 = const()[name = tensor("op_8752"), val = tensor([1, 20, 64, -1])]; + tensor var_8753_cast_fp16 = reshape(shape = var_8752, x = key_87_cast_fp16)[name = tensor("op_8753_cast_fp16")]; + tensor mh_w_131_transpose_x_0 = const()[name = tensor("mh_w_131_transpose_x_0"), val = tensor(true)]; + tensor mh_w_131_transpose_y_0 = const()[name = tensor("mh_w_131_transpose_y_0"), val = tensor(false)]; + tensor mh_w_131_cast_fp16 = matmul(transpose_x = mh_w_131_transpose_x_0, transpose_y = mh_w_131_transpose_y_0, x = var_8751_cast_fp16, y = var_8753_cast_fp16)[name = tensor("mh_w_131_cast_fp16")]; + tensor obj_307_cast_fp16 = softmax(axis = var_8474, x = mh_w_131_cast_fp16)[name = tensor("obj_307_cast_fp16")]; + tensor var_8757 = const()[name = tensor("op_8757"), val = tensor([1, 20, 64, -1])]; + tensor var_8758_cast_fp16 = reshape(shape = var_8757, x = value_87_cast_fp16)[name = tensor("op_8758_cast_fp16")]; + tensor attn_87_transpose_x_0 = const()[name = tensor("attn_87_transpose_x_0"), val = tensor(false)]; + tensor attn_87_transpose_y_0 = const()[name = tensor("attn_87_transpose_y_0"), val = tensor(true)]; + tensor attn_87_cast_fp16 = matmul(transpose_x = attn_87_transpose_x_0, transpose_y = attn_87_transpose_y_0, x = var_8758_cast_fp16, y = obj_307_cast_fp16)[name = tensor("attn_87_cast_fp16")]; + tensor var_8761 = const()[name = tensor("op_8761"), val = tensor([1, 1280, 1, -1])]; + tensor input_647_cast_fp16 = reshape(shape = var_8761, x = attn_87_cast_fp16)[name = tensor("input_647_cast_fp16")]; + tensor var_8768 = const()[name = tensor("op_8768"), val = tensor([1, 1])]; + tensor var_8770 = const()[name = tensor("op_8770"), val = tensor([1, 1])]; + tensor pretrained_out_435_pad_type_0 = const()[name = tensor("pretrained_out_435_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_435_pad_0 = const()[name = tensor("pretrained_out_435_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(440524480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441343744))), name = tensor("layers_21_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_21_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441343872)))]; + tensor pretrained_out_435_cast_fp16 = conv(bias = layers_21_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_8770, groups = var_8481, pad = pretrained_out_435_pad_0, pad_type = pretrained_out_435_pad_type_0, strides = var_8768, weight = layers_21_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_647_cast_fp16)[name = tensor("pretrained_out_435_cast_fp16")]; + tensor var_8774 = const()[name = tensor("op_8774"), val = tensor([1, 1])]; + tensor var_8776 = const()[name = tensor("op_8776"), val = tensor([1, 1])]; + tensor input_649_pad_type_0 = const()[name = tensor("input_649_pad_type_0"), val = tensor("custom")]; + tensor input_649_pad_0 = const()[name = tensor("input_649_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441346496)))]; + tensor input_649_cast_fp16 = conv(dilations = var_8776, groups = var_8481, pad = input_649_pad_0, pad_type = input_649_pad_type_0, strides = var_8774, weight = layers_21_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_647_cast_fp16)[name = tensor("input_649_cast_fp16")]; + tensor var_8780 = const()[name = tensor("op_8780"), val = tensor([1, 1])]; + tensor var_8782 = const()[name = tensor("op_8782"), val = tensor([1, 1])]; + tensor lora_out_869_pad_type_0 = const()[name = tensor("lora_out_869_pad_type_0"), val = tensor("custom")]; + tensor lora_out_869_pad_0 = const()[name = tensor("lora_out_869_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_871_weight_0_to_fp16 = const()[name = tensor("lora_out_871_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441387520)))]; + tensor lora_out_871_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8782, groups = var_8481, pad = lora_out_869_pad_0, pad_type = lora_out_869_pad_type_0, strides = var_8780, weight = lora_out_871_weight_0_to_fp16, x = input_649_cast_fp16)[name = tensor("lora_out_871_cast_fp16")]; + tensor obj_305_cast_fp16 = add(x = pretrained_out_435_cast_fp16, y = lora_out_871_cast_fp16)[name = tensor("obj_305_cast_fp16")]; + tensor inputs_131_cast_fp16 = add(x = inputs_129_cast_fp16, y = obj_305_cast_fp16)[name = tensor("inputs_131_cast_fp16")]; + tensor var_8794 = const()[name = tensor("op_8794"), val = tensor([1])]; + tensor channels_mean_131_cast_fp16 = reduce_mean(axes = var_8794, keep_dims = var_8482, x = inputs_131_cast_fp16)[name = tensor("channels_mean_131_cast_fp16")]; + tensor zero_mean_131_cast_fp16 = sub(x = inputs_131_cast_fp16, y = channels_mean_131_cast_fp16)[name = tensor("zero_mean_131_cast_fp16")]; + tensor zero_mean_sq_131_cast_fp16 = mul(x = zero_mean_131_cast_fp16, y = zero_mean_131_cast_fp16)[name = tensor("zero_mean_sq_131_cast_fp16")]; + tensor var_8798 = const()[name = tensor("op_8798"), val = tensor([1])]; + tensor var_8799_cast_fp16 = reduce_mean(axes = var_8798, keep_dims = var_8482, x = zero_mean_sq_131_cast_fp16)[name = tensor("op_8799_cast_fp16")]; + tensor var_8800_to_fp16 = const()[name = tensor("op_8800_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_8801_cast_fp16 = add(x = var_8799_cast_fp16, y = var_8800_to_fp16)[name = tensor("op_8801_cast_fp16")]; + tensor denom_131_epsilon_0 = const()[name = tensor("denom_131_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_131_cast_fp16 = rsqrt(epsilon = denom_131_epsilon_0, x = var_8801_cast_fp16)[name = tensor("denom_131_cast_fp16")]; + tensor out_131_cast_fp16 = mul(x = zero_mean_131_cast_fp16, y = denom_131_cast_fp16)[name = tensor("out_131_cast_fp16")]; + tensor input_651_gamma_0_to_fp16 = const()[name = tensor("input_651_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441428544)))]; + tensor input_651_beta_0_to_fp16 = const()[name = tensor("input_651_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441431168)))]; + tensor input_651_epsilon_0_to_fp16 = const()[name = tensor("input_651_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_651_cast_fp16 = batch_norm(beta = input_651_beta_0_to_fp16, epsilon = input_651_epsilon_0_to_fp16, gamma = input_651_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_131_cast_fp16)[name = tensor("input_651_cast_fp16")]; + tensor var_8815 = const()[name = tensor("op_8815"), val = tensor([1, 1])]; + tensor var_8817 = const()[name = tensor("op_8817"), val = tensor([1, 1])]; + tensor pretrained_out_437_pad_type_0 = const()[name = tensor("pretrained_out_437_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_437_pad_0 = const()[name = tensor("pretrained_out_437_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441433792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444710656))), name = tensor("layers_21_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_21_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444710784)))]; + tensor pretrained_out_437_cast_fp16 = conv(bias = layers_21_fc1_pretrained_bias_to_fp16, dilations = var_8817, groups = var_8481, pad = pretrained_out_437_pad_0, pad_type = pretrained_out_437_pad_type_0, strides = var_8815, weight = layers_21_fc1_pretrained_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = tensor("pretrained_out_437_cast_fp16")]; + tensor var_8821 = const()[name = tensor("op_8821"), val = tensor([1, 1])]; + tensor var_8823 = const()[name = tensor("op_8823"), val = tensor([1, 1])]; + tensor input_653_pad_type_0 = const()[name = tensor("input_653_pad_type_0"), val = tensor("custom")]; + tensor input_653_pad_0 = const()[name = tensor("input_653_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_21_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444721088)))]; + tensor input_653_cast_fp16 = conv(dilations = var_8823, groups = var_8481, pad = input_653_pad_0, pad_type = input_653_pad_type_0, strides = var_8821, weight = layers_21_fc1_loraA_weight_to_fp16, x = input_651_cast_fp16)[name = tensor("input_653_cast_fp16")]; + tensor var_8827 = const()[name = tensor("op_8827"), val = tensor([1, 1])]; + tensor var_8829 = const()[name = tensor("op_8829"), val = tensor([1, 1])]; + tensor lora_out_873_pad_type_0 = const()[name = tensor("lora_out_873_pad_type_0"), val = tensor("custom")]; + tensor lora_out_873_pad_0 = const()[name = tensor("lora_out_873_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_875_weight_0_to_fp16 = const()[name = tensor("lora_out_875_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444762112)))]; + tensor lora_out_875_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_8829, groups = var_8481, pad = lora_out_873_pad_0, pad_type = lora_out_873_pad_type_0, strides = var_8827, weight = lora_out_875_weight_0_to_fp16, x = input_653_cast_fp16)[name = tensor("lora_out_875_cast_fp16")]; + tensor input_655_cast_fp16 = add(x = pretrained_out_437_cast_fp16, y = lora_out_875_cast_fp16)[name = tensor("input_655_cast_fp16")]; + tensor input_657_mode_0 = const()[name = tensor("input_657_mode_0"), val = tensor("EXACT")]; + tensor input_657_cast_fp16 = gelu(mode = input_657_mode_0, x = input_655_cast_fp16)[name = tensor("input_657_cast_fp16")]; + tensor var_8841 = const()[name = tensor("op_8841"), val = tensor([1, 1])]; + tensor var_8843 = const()[name = tensor("op_8843"), val = tensor([1, 1])]; + tensor pretrained_out_439_pad_type_0 = const()[name = tensor("pretrained_out_439_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_439_pad_0 = const()[name = tensor("pretrained_out_439_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444926016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448202880))), name = tensor("layers_21_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_21_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_21_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448203008)))]; + tensor pretrained_out_439_cast_fp16 = conv(bias = layers_21_fc2_pretrained_bias_to_fp16, dilations = var_8843, groups = var_8481, pad = pretrained_out_439_pad_0, pad_type = pretrained_out_439_pad_type_0, strides = var_8841, weight = layers_21_fc2_pretrained_weight_to_fp16_palettized, x = input_657_cast_fp16)[name = tensor("pretrained_out_439_cast_fp16")]; + tensor var_8847 = const()[name = tensor("op_8847"), val = tensor([1, 1])]; + tensor var_8849 = const()[name = tensor("op_8849"), val = tensor([1, 1])]; + tensor input_659_pad_type_0 = const()[name = tensor("input_659_pad_type_0"), val = tensor("custom")]; + tensor input_659_pad_0 = const()[name = tensor("input_659_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_21_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_21_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448205632)))]; + tensor input_659_cast_fp16 = conv(dilations = var_8849, groups = var_8481, pad = input_659_pad_0, pad_type = input_659_pad_type_0, strides = var_8847, weight = layers_21_fc2_loraA_weight_to_fp16, x = input_657_cast_fp16)[name = tensor("input_659_cast_fp16")]; + tensor var_8853 = const()[name = tensor("op_8853"), val = tensor([1, 1])]; + tensor var_8855 = const()[name = tensor("op_8855"), val = tensor([1, 1])]; + tensor lora_out_877_pad_type_0 = const()[name = tensor("lora_out_877_pad_type_0"), val = tensor("custom")]; + tensor lora_out_877_pad_0 = const()[name = tensor("lora_out_877_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_879_weight_0_to_fp16 = const()[name = tensor("lora_out_879_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448369536)))]; + tensor lora_out_879_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8855, groups = var_8481, pad = lora_out_877_pad_0, pad_type = lora_out_877_pad_type_0, strides = var_8853, weight = lora_out_879_weight_0_to_fp16, x = input_659_cast_fp16)[name = tensor("lora_out_879_cast_fp16")]; + tensor hidden_states_45_cast_fp16 = add(x = pretrained_out_439_cast_fp16, y = lora_out_879_cast_fp16)[name = tensor("hidden_states_45_cast_fp16")]; + tensor inputs_133_cast_fp16 = add(x = inputs_131_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor("inputs_133_cast_fp16")]; + tensor var_8872 = const()[name = tensor("op_8872"), val = tensor(3)]; + tensor var_8879 = const()[name = tensor("op_8879"), val = tensor(1)]; + tensor var_8880 = const()[name = tensor("op_8880"), val = tensor(true)]; + tensor var_8892 = const()[name = tensor("op_8892"), val = tensor([1])]; + tensor channels_mean_133_cast_fp16 = reduce_mean(axes = var_8892, keep_dims = var_8880, x = inputs_133_cast_fp16)[name = tensor("channels_mean_133_cast_fp16")]; + tensor zero_mean_133_cast_fp16 = sub(x = inputs_133_cast_fp16, y = channels_mean_133_cast_fp16)[name = tensor("zero_mean_133_cast_fp16")]; + tensor zero_mean_sq_133_cast_fp16 = mul(x = zero_mean_133_cast_fp16, y = zero_mean_133_cast_fp16)[name = tensor("zero_mean_sq_133_cast_fp16")]; + tensor var_8896 = const()[name = tensor("op_8896"), val = tensor([1])]; + tensor var_8897_cast_fp16 = reduce_mean(axes = var_8896, keep_dims = var_8880, x = zero_mean_sq_133_cast_fp16)[name = tensor("op_8897_cast_fp16")]; + tensor var_8898_to_fp16 = const()[name = tensor("op_8898_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_8899_cast_fp16 = add(x = var_8897_cast_fp16, y = var_8898_to_fp16)[name = tensor("op_8899_cast_fp16")]; + tensor denom_133_epsilon_0 = const()[name = tensor("denom_133_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_133_cast_fp16 = rsqrt(epsilon = denom_133_epsilon_0, x = var_8899_cast_fp16)[name = tensor("denom_133_cast_fp16")]; + tensor out_133_cast_fp16 = mul(x = zero_mean_133_cast_fp16, y = denom_133_cast_fp16)[name = tensor("out_133_cast_fp16")]; + tensor obj_309_gamma_0_to_fp16 = const()[name = tensor("obj_309_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448410560)))]; + tensor obj_309_beta_0_to_fp16 = const()[name = tensor("obj_309_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448413184)))]; + tensor obj_309_epsilon_0_to_fp16 = const()[name = tensor("obj_309_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_309_cast_fp16 = batch_norm(beta = obj_309_beta_0_to_fp16, epsilon = obj_309_epsilon_0_to_fp16, gamma = obj_309_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_133_cast_fp16)[name = tensor("obj_309_cast_fp16")]; + tensor var_8917 = const()[name = tensor("op_8917"), val = tensor([1, 1])]; + tensor var_8919 = const()[name = tensor("op_8919"), val = tensor([1, 1])]; + tensor pretrained_out_441_pad_type_0 = const()[name = tensor("pretrained_out_441_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_441_pad_0 = const()[name = tensor("pretrained_out_441_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448415808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449235072))), name = tensor("layers_22_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_22_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449235200)))]; + tensor pretrained_out_441_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_8919, groups = var_8879, pad = pretrained_out_441_pad_0, pad_type = pretrained_out_441_pad_type_0, strides = var_8917, weight = layers_22_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_309_cast_fp16)[name = tensor("pretrained_out_441_cast_fp16")]; + tensor var_8923 = const()[name = tensor("op_8923"), val = tensor([1, 1])]; + tensor var_8925 = const()[name = tensor("op_8925"), val = tensor([1, 1])]; + tensor input_661_pad_type_0 = const()[name = tensor("input_661_pad_type_0"), val = tensor("custom")]; + tensor input_661_pad_0 = const()[name = tensor("input_661_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449237824)))]; + tensor input_661_cast_fp16 = conv(dilations = var_8925, groups = var_8879, pad = input_661_pad_0, pad_type = input_661_pad_type_0, strides = var_8923, weight = layers_22_self_attn_q_proj_loraA_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("input_661_cast_fp16")]; + tensor var_8929 = const()[name = tensor("op_8929"), val = tensor([1, 1])]; + tensor var_8931 = const()[name = tensor("op_8931"), val = tensor([1, 1])]; + tensor lora_out_881_pad_type_0 = const()[name = tensor("lora_out_881_pad_type_0"), val = tensor("custom")]; + tensor lora_out_881_pad_0 = const()[name = tensor("lora_out_881_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_883_weight_0_to_fp16 = const()[name = tensor("lora_out_883_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449278848)))]; + tensor lora_out_883_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8931, groups = var_8879, pad = lora_out_881_pad_0, pad_type = lora_out_881_pad_type_0, strides = var_8929, weight = lora_out_883_weight_0_to_fp16, x = input_661_cast_fp16)[name = tensor("lora_out_883_cast_fp16")]; + tensor query_89_cast_fp16 = add(x = pretrained_out_441_cast_fp16, y = lora_out_883_cast_fp16)[name = tensor("query_89_cast_fp16")]; + tensor var_8941 = const()[name = tensor("op_8941"), val = tensor([1, 1])]; + tensor var_8943 = const()[name = tensor("op_8943"), val = tensor([1, 1])]; + tensor pretrained_out_443_pad_type_0 = const()[name = tensor("pretrained_out_443_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_443_pad_0 = const()[name = tensor("pretrained_out_443_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449319872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450139136))), name = tensor("layers_22_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_443_cast_fp16 = conv(dilations = var_8943, groups = var_8879, pad = pretrained_out_443_pad_0, pad_type = pretrained_out_443_pad_type_0, strides = var_8941, weight = layers_22_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_309_cast_fp16)[name = tensor("pretrained_out_443_cast_fp16")]; + tensor var_8947 = const()[name = tensor("op_8947"), val = tensor([1, 1])]; + tensor var_8949 = const()[name = tensor("op_8949"), val = tensor([1, 1])]; + tensor input_663_pad_type_0 = const()[name = tensor("input_663_pad_type_0"), val = tensor("custom")]; + tensor input_663_pad_0 = const()[name = tensor("input_663_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450139264)))]; + tensor input_663_cast_fp16 = conv(dilations = var_8949, groups = var_8879, pad = input_663_pad_0, pad_type = input_663_pad_type_0, strides = var_8947, weight = layers_22_self_attn_k_proj_loraA_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("input_663_cast_fp16")]; + tensor var_8953 = const()[name = tensor("op_8953"), val = tensor([1, 1])]; + tensor var_8955 = const()[name = tensor("op_8955"), val = tensor([1, 1])]; + tensor lora_out_885_pad_type_0 = const()[name = tensor("lora_out_885_pad_type_0"), val = tensor("custom")]; + tensor lora_out_885_pad_0 = const()[name = tensor("lora_out_885_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_887_weight_0_to_fp16 = const()[name = tensor("lora_out_887_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450180288)))]; + tensor lora_out_887_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8955, groups = var_8879, pad = lora_out_885_pad_0, pad_type = lora_out_885_pad_type_0, strides = var_8953, weight = lora_out_887_weight_0_to_fp16, x = input_663_cast_fp16)[name = tensor("lora_out_887_cast_fp16")]; + tensor current_key_45_cast_fp16 = add(x = pretrained_out_443_cast_fp16, y = lora_out_887_cast_fp16)[name = tensor("current_key_45_cast_fp16")]; + tensor var_8966 = const()[name = tensor("op_8966"), val = tensor([1, 1])]; + tensor var_8968 = const()[name = tensor("op_8968"), val = tensor([1, 1])]; + tensor pretrained_out_445_pad_type_0 = const()[name = tensor("pretrained_out_445_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_445_pad_0 = const()[name = tensor("pretrained_out_445_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450221312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451040576))), name = tensor("layers_22_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_22_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451040704)))]; + tensor pretrained_out_445_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_8968, groups = var_8879, pad = pretrained_out_445_pad_0, pad_type = pretrained_out_445_pad_type_0, strides = var_8966, weight = layers_22_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_309_cast_fp16)[name = tensor("pretrained_out_445_cast_fp16")]; + tensor var_8972 = const()[name = tensor("op_8972"), val = tensor([1, 1])]; + tensor var_8974 = const()[name = tensor("op_8974"), val = tensor([1, 1])]; + tensor input_665_pad_type_0 = const()[name = tensor("input_665_pad_type_0"), val = tensor("custom")]; + tensor input_665_pad_0 = const()[name = tensor("input_665_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451043328)))]; + tensor input_665_cast_fp16 = conv(dilations = var_8974, groups = var_8879, pad = input_665_pad_0, pad_type = input_665_pad_type_0, strides = var_8972, weight = layers_22_self_attn_v_proj_loraA_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("input_665_cast_fp16")]; + tensor var_8978 = const()[name = tensor("op_8978"), val = tensor([1, 1])]; + tensor var_8980 = const()[name = tensor("op_8980"), val = tensor([1, 1])]; + tensor lora_out_889_pad_type_0 = const()[name = tensor("lora_out_889_pad_type_0"), val = tensor("custom")]; + tensor lora_out_889_pad_0 = const()[name = tensor("lora_out_889_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_891_weight_0_to_fp16 = const()[name = tensor("lora_out_891_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451084352)))]; + tensor lora_out_891_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_8980, groups = var_8879, pad = lora_out_889_pad_0, pad_type = lora_out_889_pad_type_0, strides = var_8978, weight = lora_out_891_weight_0_to_fp16, x = input_665_cast_fp16)[name = tensor("lora_out_891_cast_fp16")]; + tensor current_value_45_cast_fp16 = add(x = pretrained_out_445_cast_fp16, y = lora_out_891_cast_fp16)[name = tensor("current_value_45_cast_fp16")]; + tensor var_8990_cast_fp16 = mul(x = current_key_45_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_8990_cast_fp16")]; + tensor var_8992_cast_fp16 = mul(x = var_103_cast_fp16_22, y = var_295_cast_fp16)[name = tensor("op_8992_cast_fp16")]; + tensor key_89_cast_fp16 = add(x = var_8990_cast_fp16, y = var_8992_cast_fp16)[name = tensor("key_89_cast_fp16")]; + tensor var_8994_cast_fp16 = mul(x = current_value_45_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_8994_cast_fp16")]; + tensor var_8996_cast_fp16 = mul(x = var_138_cast_fp16_22, y = var_295_cast_fp16)[name = tensor("op_8996_cast_fp16")]; + tensor value_89_cast_fp16 = add(x = var_8994_cast_fp16, y = var_8996_cast_fp16)[name = tensor("value_89_cast_fp16")]; + tensor var_8999 = const()[name = tensor("op_8999"), val = tensor([1, 20, 64, -1])]; + tensor var_9000_cast_fp16 = reshape(shape = var_8999, x = query_89_cast_fp16)[name = tensor("op_9000_cast_fp16")]; + tensor var_9001_to_fp16 = const()[name = tensor("op_9001_to_fp16"), val = tensor(0x1p-3)]; + tensor var_9002_cast_fp16 = mul(x = var_9000_cast_fp16, y = var_9001_to_fp16)[name = tensor("op_9002_cast_fp16")]; + tensor var_9003 = const()[name = tensor("op_9003"), val = tensor([1, 20, 64, -1])]; + tensor var_9004_cast_fp16 = reshape(shape = var_9003, x = key_89_cast_fp16)[name = tensor("op_9004_cast_fp16")]; + tensor mh_w_133_transpose_x_0 = const()[name = tensor("mh_w_133_transpose_x_0"), val = tensor(true)]; + tensor mh_w_133_transpose_y_0 = const()[name = tensor("mh_w_133_transpose_y_0"), val = tensor(false)]; + tensor mh_w_133_cast_fp16 = matmul(transpose_x = mh_w_133_transpose_x_0, transpose_y = mh_w_133_transpose_y_0, x = var_9002_cast_fp16, y = var_9004_cast_fp16)[name = tensor("mh_w_133_cast_fp16")]; + tensor mh_w_135_cast_fp16 = add(x = mh_w_133_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_135_cast_fp16")]; + tensor var_9012_cast_fp16 = softmax(axis = var_8872, x = mh_w_135_cast_fp16)[name = tensor("op_9012_cast_fp16")]; + tensor var_9013 = const()[name = tensor("op_9013"), val = tensor([1, 20, 64, -1])]; + tensor var_9014_cast_fp16 = reshape(shape = var_9013, x = value_89_cast_fp16)[name = tensor("op_9014_cast_fp16")]; + tensor attn_89_transpose_x_0 = const()[name = tensor("attn_89_transpose_x_0"), val = tensor(false)]; + tensor attn_89_transpose_y_0 = const()[name = tensor("attn_89_transpose_y_0"), val = tensor(true)]; + tensor attn_89_cast_fp16 = matmul(transpose_x = attn_89_transpose_x_0, transpose_y = attn_89_transpose_y_0, x = var_9014_cast_fp16, y = var_9012_cast_fp16)[name = tensor("attn_89_cast_fp16")]; + tensor var_9017 = const()[name = tensor("op_9017"), val = tensor([1, 1280, 1, -1])]; + tensor input_667_cast_fp16 = reshape(shape = var_9017, x = attn_89_cast_fp16)[name = tensor("input_667_cast_fp16")]; + tensor var_9024 = const()[name = tensor("op_9024"), val = tensor([1, 1])]; + tensor var_9026 = const()[name = tensor("op_9026"), val = tensor([1, 1])]; + tensor pretrained_out_447_pad_type_0 = const()[name = tensor("pretrained_out_447_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_447_pad_0 = const()[name = tensor("pretrained_out_447_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451125376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451944640))), name = tensor("layers_22_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_22_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451944768)))]; + tensor pretrained_out_447_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_9026, groups = var_8879, pad = pretrained_out_447_pad_0, pad_type = pretrained_out_447_pad_type_0, strides = var_9024, weight = layers_22_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = tensor("pretrained_out_447_cast_fp16")]; + tensor var_9030 = const()[name = tensor("op_9030"), val = tensor([1, 1])]; + tensor var_9032 = const()[name = tensor("op_9032"), val = tensor([1, 1])]; + tensor input_669_pad_type_0 = const()[name = tensor("input_669_pad_type_0"), val = tensor("custom")]; + tensor input_669_pad_0 = const()[name = tensor("input_669_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451947392)))]; + tensor input_669_cast_fp16 = conv(dilations = var_9032, groups = var_8879, pad = input_669_pad_0, pad_type = input_669_pad_type_0, strides = var_9030, weight = layers_22_self_attn_o_proj_loraA_weight_to_fp16, x = input_667_cast_fp16)[name = tensor("input_669_cast_fp16")]; + tensor var_9036 = const()[name = tensor("op_9036"), val = tensor([1, 1])]; + tensor var_9038 = const()[name = tensor("op_9038"), val = tensor([1, 1])]; + tensor lora_out_893_pad_type_0 = const()[name = tensor("lora_out_893_pad_type_0"), val = tensor("custom")]; + tensor lora_out_893_pad_0 = const()[name = tensor("lora_out_893_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_895_weight_0_to_fp16 = const()[name = tensor("lora_out_895_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451988416)))]; + tensor lora_out_895_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9038, groups = var_8879, pad = lora_out_893_pad_0, pad_type = lora_out_893_pad_type_0, strides = var_9036, weight = lora_out_895_weight_0_to_fp16, x = input_669_cast_fp16)[name = tensor("lora_out_895_cast_fp16")]; + tensor obj_315_cast_fp16 = add(x = pretrained_out_447_cast_fp16, y = lora_out_895_cast_fp16)[name = tensor("obj_315_cast_fp16")]; + tensor inputs_135_cast_fp16 = add(x = inputs_133_cast_fp16, y = obj_315_cast_fp16)[name = tensor("inputs_135_cast_fp16")]; + tensor var_9051 = const()[name = tensor("op_9051"), val = tensor([1])]; + tensor channels_mean_135_cast_fp16 = reduce_mean(axes = var_9051, keep_dims = var_8880, x = inputs_135_cast_fp16)[name = tensor("channels_mean_135_cast_fp16")]; + tensor zero_mean_135_cast_fp16 = sub(x = inputs_135_cast_fp16, y = channels_mean_135_cast_fp16)[name = tensor("zero_mean_135_cast_fp16")]; + tensor zero_mean_sq_135_cast_fp16 = mul(x = zero_mean_135_cast_fp16, y = zero_mean_135_cast_fp16)[name = tensor("zero_mean_sq_135_cast_fp16")]; + tensor var_9055 = const()[name = tensor("op_9055"), val = tensor([1])]; + tensor var_9056_cast_fp16 = reduce_mean(axes = var_9055, keep_dims = var_8880, x = zero_mean_sq_135_cast_fp16)[name = tensor("op_9056_cast_fp16")]; + tensor var_9057_to_fp16 = const()[name = tensor("op_9057_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_9058_cast_fp16 = add(x = var_9056_cast_fp16, y = var_9057_to_fp16)[name = tensor("op_9058_cast_fp16")]; + tensor denom_135_epsilon_0 = const()[name = tensor("denom_135_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_135_cast_fp16 = rsqrt(epsilon = denom_135_epsilon_0, x = var_9058_cast_fp16)[name = tensor("denom_135_cast_fp16")]; + tensor out_135_cast_fp16 = mul(x = zero_mean_135_cast_fp16, y = denom_135_cast_fp16)[name = tensor("out_135_cast_fp16")]; + tensor obj_317_gamma_0_to_fp16 = const()[name = tensor("obj_317_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452029440)))]; + tensor obj_317_beta_0_to_fp16 = const()[name = tensor("obj_317_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452032064)))]; + tensor obj_317_epsilon_0_to_fp16 = const()[name = tensor("obj_317_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_317_cast_fp16 = batch_norm(beta = obj_317_beta_0_to_fp16, epsilon = obj_317_epsilon_0_to_fp16, gamma = obj_317_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_135_cast_fp16)[name = tensor("obj_317_cast_fp16")]; + tensor var_9076 = const()[name = tensor("op_9076"), val = tensor([1, 1])]; + tensor var_9078 = const()[name = tensor("op_9078"), val = tensor([1, 1])]; + tensor pretrained_out_449_pad_type_0 = const()[name = tensor("pretrained_out_449_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_449_pad_0 = const()[name = tensor("pretrained_out_449_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452034688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452853952))), name = tensor("layers_22_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_22_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452854080)))]; + tensor pretrained_out_449_cast_fp16 = conv(bias = layers_22_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_9078, groups = var_8879, pad = pretrained_out_449_pad_0, pad_type = pretrained_out_449_pad_type_0, strides = var_9076, weight = layers_22_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_317_cast_fp16)[name = tensor("pretrained_out_449_cast_fp16")]; + tensor var_9082 = const()[name = tensor("op_9082"), val = tensor([1, 1])]; + tensor var_9084 = const()[name = tensor("op_9084"), val = tensor([1, 1])]; + tensor input_671_pad_type_0 = const()[name = tensor("input_671_pad_type_0"), val = tensor("custom")]; + tensor input_671_pad_0 = const()[name = tensor("input_671_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452856704)))]; + tensor input_671_cast_fp16 = conv(dilations = var_9084, groups = var_8879, pad = input_671_pad_0, pad_type = input_671_pad_type_0, strides = var_9082, weight = layers_22_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_317_cast_fp16)[name = tensor("input_671_cast_fp16")]; + tensor var_9088 = const()[name = tensor("op_9088"), val = tensor([1, 1])]; + tensor var_9090 = const()[name = tensor("op_9090"), val = tensor([1, 1])]; + tensor lora_out_897_pad_type_0 = const()[name = tensor("lora_out_897_pad_type_0"), val = tensor("custom")]; + tensor lora_out_897_pad_0 = const()[name = tensor("lora_out_897_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_899_weight_0_to_fp16 = const()[name = tensor("lora_out_899_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452897728)))]; + tensor lora_out_899_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9090, groups = var_8879, pad = lora_out_897_pad_0, pad_type = lora_out_897_pad_type_0, strides = var_9088, weight = lora_out_899_weight_0_to_fp16, x = input_671_cast_fp16)[name = tensor("lora_out_899_cast_fp16")]; + tensor query_91_cast_fp16 = add(x = pretrained_out_449_cast_fp16, y = lora_out_899_cast_fp16)[name = tensor("query_91_cast_fp16")]; + tensor var_9100 = const()[name = tensor("op_9100"), val = tensor([1, 1])]; + tensor var_9102 = const()[name = tensor("op_9102"), val = tensor([1, 1])]; + tensor pretrained_out_451_pad_type_0 = const()[name = tensor("pretrained_out_451_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_451_pad_0 = const()[name = tensor("pretrained_out_451_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(452938752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453758016))), name = tensor("layers_22_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_451_cast_fp16 = conv(dilations = var_9102, groups = var_8879, pad = pretrained_out_451_pad_0, pad_type = pretrained_out_451_pad_type_0, strides = var_9100, weight = layers_22_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_451_cast_fp16")]; + tensor var_9106 = const()[name = tensor("op_9106"), val = tensor([1, 1])]; + tensor var_9108 = const()[name = tensor("op_9108"), val = tensor([1, 1])]; + tensor input_673_pad_type_0 = const()[name = tensor("input_673_pad_type_0"), val = tensor("custom")]; + tensor input_673_pad_0 = const()[name = tensor("input_673_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453758144)))]; + tensor input_673_cast_fp16 = conv(dilations = var_9108, groups = var_8879, pad = input_673_pad_0, pad_type = input_673_pad_type_0, strides = var_9106, weight = layers_22_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_673_cast_fp16")]; + tensor var_9112 = const()[name = tensor("op_9112"), val = tensor([1, 1])]; + tensor var_9114 = const()[name = tensor("op_9114"), val = tensor([1, 1])]; + tensor lora_out_901_pad_type_0 = const()[name = tensor("lora_out_901_pad_type_0"), val = tensor("custom")]; + tensor lora_out_901_pad_0 = const()[name = tensor("lora_out_901_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_903_weight_0_to_fp16 = const()[name = tensor("lora_out_903_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453799168)))]; + tensor lora_out_903_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9114, groups = var_8879, pad = lora_out_901_pad_0, pad_type = lora_out_901_pad_type_0, strides = var_9112, weight = lora_out_903_weight_0_to_fp16, x = input_673_cast_fp16)[name = tensor("lora_out_903_cast_fp16")]; + tensor key_91_cast_fp16 = add(x = pretrained_out_451_cast_fp16, y = lora_out_903_cast_fp16)[name = tensor("key_91_cast_fp16")]; + tensor var_9125 = const()[name = tensor("op_9125"), val = tensor([1, 1])]; + tensor var_9127 = const()[name = tensor("op_9127"), val = tensor([1, 1])]; + tensor pretrained_out_453_pad_type_0 = const()[name = tensor("pretrained_out_453_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_453_pad_0 = const()[name = tensor("pretrained_out_453_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453840192))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454659456))), name = tensor("layers_22_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_22_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454659584)))]; + tensor pretrained_out_453_cast_fp16 = conv(bias = layers_22_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_9127, groups = var_8879, pad = pretrained_out_453_pad_0, pad_type = pretrained_out_453_pad_type_0, strides = var_9125, weight = layers_22_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_453_cast_fp16")]; + tensor var_9131 = const()[name = tensor("op_9131"), val = tensor([1, 1])]; + tensor var_9133 = const()[name = tensor("op_9133"), val = tensor([1, 1])]; + tensor input_675_pad_type_0 = const()[name = tensor("input_675_pad_type_0"), val = tensor("custom")]; + tensor input_675_pad_0 = const()[name = tensor("input_675_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454662208)))]; + tensor input_675_cast_fp16 = conv(dilations = var_9133, groups = var_8879, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = var_9131, weight = layers_22_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_675_cast_fp16")]; + tensor var_9137 = const()[name = tensor("op_9137"), val = tensor([1, 1])]; + tensor var_9139 = const()[name = tensor("op_9139"), val = tensor([1, 1])]; + tensor lora_out_905_pad_type_0 = const()[name = tensor("lora_out_905_pad_type_0"), val = tensor("custom")]; + tensor lora_out_905_pad_0 = const()[name = tensor("lora_out_905_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_907_weight_0_to_fp16 = const()[name = tensor("lora_out_907_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454703232)))]; + tensor lora_out_907_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9139, groups = var_8879, pad = lora_out_905_pad_0, pad_type = lora_out_905_pad_type_0, strides = var_9137, weight = lora_out_907_weight_0_to_fp16, x = input_675_cast_fp16)[name = tensor("lora_out_907_cast_fp16")]; + tensor value_91_cast_fp16 = add(x = pretrained_out_453_cast_fp16, y = lora_out_907_cast_fp16)[name = tensor("value_91_cast_fp16")]; + tensor var_9146 = const()[name = tensor("op_9146"), val = tensor([1, 20, 64, -1])]; + tensor var_9147_cast_fp16 = reshape(shape = var_9146, x = query_91_cast_fp16)[name = tensor("op_9147_cast_fp16")]; + tensor var_9148_to_fp16 = const()[name = tensor("op_9148_to_fp16"), val = tensor(0x1p-3)]; + tensor var_9149_cast_fp16 = mul(x = var_9147_cast_fp16, y = var_9148_to_fp16)[name = tensor("op_9149_cast_fp16")]; + tensor var_9150 = const()[name = tensor("op_9150"), val = tensor([1, 20, 64, -1])]; + tensor var_9151_cast_fp16 = reshape(shape = var_9150, x = key_91_cast_fp16)[name = tensor("op_9151_cast_fp16")]; + tensor mh_w_137_transpose_x_0 = const()[name = tensor("mh_w_137_transpose_x_0"), val = tensor(true)]; + tensor mh_w_137_transpose_y_0 = const()[name = tensor("mh_w_137_transpose_y_0"), val = tensor(false)]; + tensor mh_w_137_cast_fp16 = matmul(transpose_x = mh_w_137_transpose_x_0, transpose_y = mh_w_137_transpose_y_0, x = var_9149_cast_fp16, y = var_9151_cast_fp16)[name = tensor("mh_w_137_cast_fp16")]; + tensor obj_321_cast_fp16 = softmax(axis = var_8872, x = mh_w_137_cast_fp16)[name = tensor("obj_321_cast_fp16")]; + tensor var_9155 = const()[name = tensor("op_9155"), val = tensor([1, 20, 64, -1])]; + tensor var_9156_cast_fp16 = reshape(shape = var_9155, x = value_91_cast_fp16)[name = tensor("op_9156_cast_fp16")]; + tensor attn_91_transpose_x_0 = const()[name = tensor("attn_91_transpose_x_0"), val = tensor(false)]; + tensor attn_91_transpose_y_0 = const()[name = tensor("attn_91_transpose_y_0"), val = tensor(true)]; + tensor attn_91_cast_fp16 = matmul(transpose_x = attn_91_transpose_x_0, transpose_y = attn_91_transpose_y_0, x = var_9156_cast_fp16, y = obj_321_cast_fp16)[name = tensor("attn_91_cast_fp16")]; + tensor var_9159 = const()[name = tensor("op_9159"), val = tensor([1, 1280, 1, -1])]; + tensor input_677_cast_fp16 = reshape(shape = var_9159, x = attn_91_cast_fp16)[name = tensor("input_677_cast_fp16")]; + tensor var_9166 = const()[name = tensor("op_9166"), val = tensor([1, 1])]; + tensor var_9168 = const()[name = tensor("op_9168"), val = tensor([1, 1])]; + tensor pretrained_out_455_pad_type_0 = const()[name = tensor("pretrained_out_455_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_455_pad_0 = const()[name = tensor("pretrained_out_455_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454744256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455563520))), name = tensor("layers_22_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_22_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455563648)))]; + tensor pretrained_out_455_cast_fp16 = conv(bias = layers_22_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_9168, groups = var_8879, pad = pretrained_out_455_pad_0, pad_type = pretrained_out_455_pad_type_0, strides = var_9166, weight = layers_22_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_677_cast_fp16)[name = tensor("pretrained_out_455_cast_fp16")]; + tensor var_9172 = const()[name = tensor("op_9172"), val = tensor([1, 1])]; + tensor var_9174 = const()[name = tensor("op_9174"), val = tensor([1, 1])]; + tensor input_679_pad_type_0 = const()[name = tensor("input_679_pad_type_0"), val = tensor("custom")]; + tensor input_679_pad_0 = const()[name = tensor("input_679_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455566272)))]; + tensor input_679_cast_fp16 = conv(dilations = var_9174, groups = var_8879, pad = input_679_pad_0, pad_type = input_679_pad_type_0, strides = var_9172, weight = layers_22_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_677_cast_fp16)[name = tensor("input_679_cast_fp16")]; + tensor var_9178 = const()[name = tensor("op_9178"), val = tensor([1, 1])]; + tensor var_9180 = const()[name = tensor("op_9180"), val = tensor([1, 1])]; + tensor lora_out_909_pad_type_0 = const()[name = tensor("lora_out_909_pad_type_0"), val = tensor("custom")]; + tensor lora_out_909_pad_0 = const()[name = tensor("lora_out_909_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_911_weight_0_to_fp16 = const()[name = tensor("lora_out_911_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455607296)))]; + tensor lora_out_911_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9180, groups = var_8879, pad = lora_out_909_pad_0, pad_type = lora_out_909_pad_type_0, strides = var_9178, weight = lora_out_911_weight_0_to_fp16, x = input_679_cast_fp16)[name = tensor("lora_out_911_cast_fp16")]; + tensor obj_319_cast_fp16 = add(x = pretrained_out_455_cast_fp16, y = lora_out_911_cast_fp16)[name = tensor("obj_319_cast_fp16")]; + tensor inputs_137_cast_fp16 = add(x = inputs_135_cast_fp16, y = obj_319_cast_fp16)[name = tensor("inputs_137_cast_fp16")]; + tensor var_9192 = const()[name = tensor("op_9192"), val = tensor([1])]; + tensor channels_mean_137_cast_fp16 = reduce_mean(axes = var_9192, keep_dims = var_8880, x = inputs_137_cast_fp16)[name = tensor("channels_mean_137_cast_fp16")]; + tensor zero_mean_137_cast_fp16 = sub(x = inputs_137_cast_fp16, y = channels_mean_137_cast_fp16)[name = tensor("zero_mean_137_cast_fp16")]; + tensor zero_mean_sq_137_cast_fp16 = mul(x = zero_mean_137_cast_fp16, y = zero_mean_137_cast_fp16)[name = tensor("zero_mean_sq_137_cast_fp16")]; + tensor var_9196 = const()[name = tensor("op_9196"), val = tensor([1])]; + tensor var_9197_cast_fp16 = reduce_mean(axes = var_9196, keep_dims = var_8880, x = zero_mean_sq_137_cast_fp16)[name = tensor("op_9197_cast_fp16")]; + tensor var_9198_to_fp16 = const()[name = tensor("op_9198_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_9199_cast_fp16 = add(x = var_9197_cast_fp16, y = var_9198_to_fp16)[name = tensor("op_9199_cast_fp16")]; + tensor denom_137_epsilon_0 = const()[name = tensor("denom_137_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_137_cast_fp16 = rsqrt(epsilon = denom_137_epsilon_0, x = var_9199_cast_fp16)[name = tensor("denom_137_cast_fp16")]; + tensor out_137_cast_fp16 = mul(x = zero_mean_137_cast_fp16, y = denom_137_cast_fp16)[name = tensor("out_137_cast_fp16")]; + tensor input_681_gamma_0_to_fp16 = const()[name = tensor("input_681_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455648320)))]; + tensor input_681_beta_0_to_fp16 = const()[name = tensor("input_681_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455650944)))]; + tensor input_681_epsilon_0_to_fp16 = const()[name = tensor("input_681_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_681_cast_fp16 = batch_norm(beta = input_681_beta_0_to_fp16, epsilon = input_681_epsilon_0_to_fp16, gamma = input_681_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_137_cast_fp16)[name = tensor("input_681_cast_fp16")]; + tensor var_9213 = const()[name = tensor("op_9213"), val = tensor([1, 1])]; + tensor var_9215 = const()[name = tensor("op_9215"), val = tensor([1, 1])]; + tensor pretrained_out_457_pad_type_0 = const()[name = tensor("pretrained_out_457_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_457_pad_0 = const()[name = tensor("pretrained_out_457_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455653568))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458930432))), name = tensor("layers_22_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_22_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458930560)))]; + tensor pretrained_out_457_cast_fp16 = conv(bias = layers_22_fc1_pretrained_bias_to_fp16, dilations = var_9215, groups = var_8879, pad = pretrained_out_457_pad_0, pad_type = pretrained_out_457_pad_type_0, strides = var_9213, weight = layers_22_fc1_pretrained_weight_to_fp16_palettized, x = input_681_cast_fp16)[name = tensor("pretrained_out_457_cast_fp16")]; + tensor var_9219 = const()[name = tensor("op_9219"), val = tensor([1, 1])]; + tensor var_9221 = const()[name = tensor("op_9221"), val = tensor([1, 1])]; + tensor input_683_pad_type_0 = const()[name = tensor("input_683_pad_type_0"), val = tensor("custom")]; + tensor input_683_pad_0 = const()[name = tensor("input_683_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_22_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458940864)))]; + tensor input_683_cast_fp16 = conv(dilations = var_9221, groups = var_8879, pad = input_683_pad_0, pad_type = input_683_pad_type_0, strides = var_9219, weight = layers_22_fc1_loraA_weight_to_fp16, x = input_681_cast_fp16)[name = tensor("input_683_cast_fp16")]; + tensor var_9225 = const()[name = tensor("op_9225"), val = tensor([1, 1])]; + tensor var_9227 = const()[name = tensor("op_9227"), val = tensor([1, 1])]; + tensor lora_out_913_pad_type_0 = const()[name = tensor("lora_out_913_pad_type_0"), val = tensor("custom")]; + tensor lora_out_913_pad_0 = const()[name = tensor("lora_out_913_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_915_weight_0_to_fp16 = const()[name = tensor("lora_out_915_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(458981888)))]; + tensor lora_out_915_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_9227, groups = var_8879, pad = lora_out_913_pad_0, pad_type = lora_out_913_pad_type_0, strides = var_9225, weight = lora_out_915_weight_0_to_fp16, x = input_683_cast_fp16)[name = tensor("lora_out_915_cast_fp16")]; + tensor input_685_cast_fp16 = add(x = pretrained_out_457_cast_fp16, y = lora_out_915_cast_fp16)[name = tensor("input_685_cast_fp16")]; + tensor input_687_mode_0 = const()[name = tensor("input_687_mode_0"), val = tensor("EXACT")]; + tensor input_687_cast_fp16 = gelu(mode = input_687_mode_0, x = input_685_cast_fp16)[name = tensor("input_687_cast_fp16")]; + tensor var_9239 = const()[name = tensor("op_9239"), val = tensor([1, 1])]; + tensor var_9241 = const()[name = tensor("op_9241"), val = tensor([1, 1])]; + tensor pretrained_out_459_pad_type_0 = const()[name = tensor("pretrained_out_459_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_459_pad_0 = const()[name = tensor("pretrained_out_459_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459145792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462422656))), name = tensor("layers_22_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_22_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_22_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462422784)))]; + tensor pretrained_out_459_cast_fp16 = conv(bias = layers_22_fc2_pretrained_bias_to_fp16, dilations = var_9241, groups = var_8879, pad = pretrained_out_459_pad_0, pad_type = pretrained_out_459_pad_type_0, strides = var_9239, weight = layers_22_fc2_pretrained_weight_to_fp16_palettized, x = input_687_cast_fp16)[name = tensor("pretrained_out_459_cast_fp16")]; + tensor var_9245 = const()[name = tensor("op_9245"), val = tensor([1, 1])]; + tensor var_9247 = const()[name = tensor("op_9247"), val = tensor([1, 1])]; + tensor input_689_pad_type_0 = const()[name = tensor("input_689_pad_type_0"), val = tensor("custom")]; + tensor input_689_pad_0 = const()[name = tensor("input_689_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_22_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_22_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462425408)))]; + tensor input_689_cast_fp16 = conv(dilations = var_9247, groups = var_8879, pad = input_689_pad_0, pad_type = input_689_pad_type_0, strides = var_9245, weight = layers_22_fc2_loraA_weight_to_fp16, x = input_687_cast_fp16)[name = tensor("input_689_cast_fp16")]; + tensor var_9251 = const()[name = tensor("op_9251"), val = tensor([1, 1])]; + tensor var_9253 = const()[name = tensor("op_9253"), val = tensor([1, 1])]; + tensor lora_out_917_pad_type_0 = const()[name = tensor("lora_out_917_pad_type_0"), val = tensor("custom")]; + tensor lora_out_917_pad_0 = const()[name = tensor("lora_out_917_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_919_weight_0_to_fp16 = const()[name = tensor("lora_out_919_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462589312)))]; + tensor lora_out_919_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9253, groups = var_8879, pad = lora_out_917_pad_0, pad_type = lora_out_917_pad_type_0, strides = var_9251, weight = lora_out_919_weight_0_to_fp16, x = input_689_cast_fp16)[name = tensor("lora_out_919_cast_fp16")]; + tensor hidden_states_47_cast_fp16 = add(x = pretrained_out_459_cast_fp16, y = lora_out_919_cast_fp16)[name = tensor("hidden_states_47_cast_fp16")]; + tensor inputs_139_cast_fp16 = add(x = inputs_137_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor("inputs_139_cast_fp16")]; + tensor var_9270 = const()[name = tensor("op_9270"), val = tensor(3)]; + tensor var_9277 = const()[name = tensor("op_9277"), val = tensor(1)]; + tensor var_9278 = const()[name = tensor("op_9278"), val = tensor(true)]; + tensor var_9290 = const()[name = tensor("op_9290"), val = tensor([1])]; + tensor channels_mean_139_cast_fp16 = reduce_mean(axes = var_9290, keep_dims = var_9278, x = inputs_139_cast_fp16)[name = tensor("channels_mean_139_cast_fp16")]; + tensor zero_mean_139_cast_fp16 = sub(x = inputs_139_cast_fp16, y = channels_mean_139_cast_fp16)[name = tensor("zero_mean_139_cast_fp16")]; + tensor zero_mean_sq_139_cast_fp16 = mul(x = zero_mean_139_cast_fp16, y = zero_mean_139_cast_fp16)[name = tensor("zero_mean_sq_139_cast_fp16")]; + tensor var_9294 = const()[name = tensor("op_9294"), val = tensor([1])]; + tensor var_9295_cast_fp16 = reduce_mean(axes = var_9294, keep_dims = var_9278, x = zero_mean_sq_139_cast_fp16)[name = tensor("op_9295_cast_fp16")]; + tensor var_9296_to_fp16 = const()[name = tensor("op_9296_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_9297_cast_fp16 = add(x = var_9295_cast_fp16, y = var_9296_to_fp16)[name = tensor("op_9297_cast_fp16")]; + tensor denom_139_epsilon_0 = const()[name = tensor("denom_139_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_139_cast_fp16 = rsqrt(epsilon = denom_139_epsilon_0, x = var_9297_cast_fp16)[name = tensor("denom_139_cast_fp16")]; + tensor out_139_cast_fp16 = mul(x = zero_mean_139_cast_fp16, y = denom_139_cast_fp16)[name = tensor("out_139_cast_fp16")]; + tensor obj_323_gamma_0_to_fp16 = const()[name = tensor("obj_323_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462630336)))]; + tensor obj_323_beta_0_to_fp16 = const()[name = tensor("obj_323_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462632960)))]; + tensor obj_323_epsilon_0_to_fp16 = const()[name = tensor("obj_323_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_323_cast_fp16 = batch_norm(beta = obj_323_beta_0_to_fp16, epsilon = obj_323_epsilon_0_to_fp16, gamma = obj_323_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_139_cast_fp16)[name = tensor("obj_323_cast_fp16")]; + tensor var_9315 = const()[name = tensor("op_9315"), val = tensor([1, 1])]; + tensor var_9317 = const()[name = tensor("op_9317"), val = tensor([1, 1])]; + tensor pretrained_out_461_pad_type_0 = const()[name = tensor("pretrained_out_461_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_461_pad_0 = const()[name = tensor("pretrained_out_461_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462635584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463454848))), name = tensor("layers_23_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_23_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463454976)))]; + tensor pretrained_out_461_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_9317, groups = var_9277, pad = pretrained_out_461_pad_0, pad_type = pretrained_out_461_pad_type_0, strides = var_9315, weight = layers_23_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_323_cast_fp16)[name = tensor("pretrained_out_461_cast_fp16")]; + tensor var_9321 = const()[name = tensor("op_9321"), val = tensor([1, 1])]; + tensor var_9323 = const()[name = tensor("op_9323"), val = tensor([1, 1])]; + tensor input_691_pad_type_0 = const()[name = tensor("input_691_pad_type_0"), val = tensor("custom")]; + tensor input_691_pad_0 = const()[name = tensor("input_691_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463457600)))]; + tensor input_691_cast_fp16 = conv(dilations = var_9323, groups = var_9277, pad = input_691_pad_0, pad_type = input_691_pad_type_0, strides = var_9321, weight = layers_23_self_attn_q_proj_loraA_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("input_691_cast_fp16")]; + tensor var_9327 = const()[name = tensor("op_9327"), val = tensor([1, 1])]; + tensor var_9329 = const()[name = tensor("op_9329"), val = tensor([1, 1])]; + tensor lora_out_921_pad_type_0 = const()[name = tensor("lora_out_921_pad_type_0"), val = tensor("custom")]; + tensor lora_out_921_pad_0 = const()[name = tensor("lora_out_921_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_923_weight_0_to_fp16 = const()[name = tensor("lora_out_923_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463498624)))]; + tensor lora_out_923_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9329, groups = var_9277, pad = lora_out_921_pad_0, pad_type = lora_out_921_pad_type_0, strides = var_9327, weight = lora_out_923_weight_0_to_fp16, x = input_691_cast_fp16)[name = tensor("lora_out_923_cast_fp16")]; + tensor query_93_cast_fp16 = add(x = pretrained_out_461_cast_fp16, y = lora_out_923_cast_fp16)[name = tensor("query_93_cast_fp16")]; + tensor var_9339 = const()[name = tensor("op_9339"), val = tensor([1, 1])]; + tensor var_9341 = const()[name = tensor("op_9341"), val = tensor([1, 1])]; + tensor pretrained_out_463_pad_type_0 = const()[name = tensor("pretrained_out_463_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_463_pad_0 = const()[name = tensor("pretrained_out_463_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(463539648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464358912))), name = tensor("layers_23_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_463_cast_fp16 = conv(dilations = var_9341, groups = var_9277, pad = pretrained_out_463_pad_0, pad_type = pretrained_out_463_pad_type_0, strides = var_9339, weight = layers_23_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_323_cast_fp16)[name = tensor("pretrained_out_463_cast_fp16")]; + tensor var_9345 = const()[name = tensor("op_9345"), val = tensor([1, 1])]; + tensor var_9347 = const()[name = tensor("op_9347"), val = tensor([1, 1])]; + tensor input_693_pad_type_0 = const()[name = tensor("input_693_pad_type_0"), val = tensor("custom")]; + tensor input_693_pad_0 = const()[name = tensor("input_693_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464359040)))]; + tensor input_693_cast_fp16 = conv(dilations = var_9347, groups = var_9277, pad = input_693_pad_0, pad_type = input_693_pad_type_0, strides = var_9345, weight = layers_23_self_attn_k_proj_loraA_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("input_693_cast_fp16")]; + tensor var_9351 = const()[name = tensor("op_9351"), val = tensor([1, 1])]; + tensor var_9353 = const()[name = tensor("op_9353"), val = tensor([1, 1])]; + tensor lora_out_925_pad_type_0 = const()[name = tensor("lora_out_925_pad_type_0"), val = tensor("custom")]; + tensor lora_out_925_pad_0 = const()[name = tensor("lora_out_925_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_927_weight_0_to_fp16 = const()[name = tensor("lora_out_927_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464400064)))]; + tensor lora_out_927_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9353, groups = var_9277, pad = lora_out_925_pad_0, pad_type = lora_out_925_pad_type_0, strides = var_9351, weight = lora_out_927_weight_0_to_fp16, x = input_693_cast_fp16)[name = tensor("lora_out_927_cast_fp16")]; + tensor current_key_47_cast_fp16 = add(x = pretrained_out_463_cast_fp16, y = lora_out_927_cast_fp16)[name = tensor("current_key_47_cast_fp16")]; + tensor var_9364 = const()[name = tensor("op_9364"), val = tensor([1, 1])]; + tensor var_9366 = const()[name = tensor("op_9366"), val = tensor([1, 1])]; + tensor pretrained_out_465_pad_type_0 = const()[name = tensor("pretrained_out_465_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_465_pad_0 = const()[name = tensor("pretrained_out_465_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(464441088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465260352))), name = tensor("layers_23_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_23_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465260480)))]; + tensor pretrained_out_465_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_9366, groups = var_9277, pad = pretrained_out_465_pad_0, pad_type = pretrained_out_465_pad_type_0, strides = var_9364, weight = layers_23_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_323_cast_fp16)[name = tensor("pretrained_out_465_cast_fp16")]; + tensor var_9370 = const()[name = tensor("op_9370"), val = tensor([1, 1])]; + tensor var_9372 = const()[name = tensor("op_9372"), val = tensor([1, 1])]; + tensor input_695_pad_type_0 = const()[name = tensor("input_695_pad_type_0"), val = tensor("custom")]; + tensor input_695_pad_0 = const()[name = tensor("input_695_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465263104)))]; + tensor input_695_cast_fp16 = conv(dilations = var_9372, groups = var_9277, pad = input_695_pad_0, pad_type = input_695_pad_type_0, strides = var_9370, weight = layers_23_self_attn_v_proj_loraA_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("input_695_cast_fp16")]; + tensor var_9376 = const()[name = tensor("op_9376"), val = tensor([1, 1])]; + tensor var_9378 = const()[name = tensor("op_9378"), val = tensor([1, 1])]; + tensor lora_out_929_pad_type_0 = const()[name = tensor("lora_out_929_pad_type_0"), val = tensor("custom")]; + tensor lora_out_929_pad_0 = const()[name = tensor("lora_out_929_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_931_weight_0_to_fp16 = const()[name = tensor("lora_out_931_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465304128)))]; + tensor lora_out_931_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9378, groups = var_9277, pad = lora_out_929_pad_0, pad_type = lora_out_929_pad_type_0, strides = var_9376, weight = lora_out_931_weight_0_to_fp16, x = input_695_cast_fp16)[name = tensor("lora_out_931_cast_fp16")]; + tensor current_value_47_cast_fp16 = add(x = pretrained_out_465_cast_fp16, y = lora_out_931_cast_fp16)[name = tensor("current_value_47_cast_fp16")]; + tensor var_9388_cast_fp16 = mul(x = current_key_47_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_9388_cast_fp16")]; + tensor var_9390_cast_fp16 = mul(x = var_103_cast_fp16_23, y = var_295_cast_fp16)[name = tensor("op_9390_cast_fp16")]; + tensor key_93_cast_fp16 = add(x = var_9388_cast_fp16, y = var_9390_cast_fp16)[name = tensor("key_93_cast_fp16")]; + tensor var_9392_cast_fp16 = mul(x = current_value_47_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_9392_cast_fp16")]; + tensor var_9394_cast_fp16 = mul(x = var_138_cast_fp16_23, y = var_295_cast_fp16)[name = tensor("op_9394_cast_fp16")]; + tensor value_93_cast_fp16 = add(x = var_9392_cast_fp16, y = var_9394_cast_fp16)[name = tensor("value_93_cast_fp16")]; + tensor var_9397 = const()[name = tensor("op_9397"), val = tensor([1, 20, 64, -1])]; + tensor var_9398_cast_fp16 = reshape(shape = var_9397, x = query_93_cast_fp16)[name = tensor("op_9398_cast_fp16")]; + tensor var_9399_to_fp16 = const()[name = tensor("op_9399_to_fp16"), val = tensor(0x1p-3)]; + tensor var_9400_cast_fp16 = mul(x = var_9398_cast_fp16, y = var_9399_to_fp16)[name = tensor("op_9400_cast_fp16")]; + tensor var_9401 = const()[name = tensor("op_9401"), val = tensor([1, 20, 64, -1])]; + tensor var_9402_cast_fp16 = reshape(shape = var_9401, x = key_93_cast_fp16)[name = tensor("op_9402_cast_fp16")]; + tensor mh_w_139_transpose_x_0 = const()[name = tensor("mh_w_139_transpose_x_0"), val = tensor(true)]; + tensor mh_w_139_transpose_y_0 = const()[name = tensor("mh_w_139_transpose_y_0"), val = tensor(false)]; + tensor mh_w_139_cast_fp16 = matmul(transpose_x = mh_w_139_transpose_x_0, transpose_y = mh_w_139_transpose_y_0, x = var_9400_cast_fp16, y = var_9402_cast_fp16)[name = tensor("mh_w_139_cast_fp16")]; + tensor mh_w_141_cast_fp16 = add(x = mh_w_139_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_141_cast_fp16")]; + tensor var_9410_cast_fp16 = softmax(axis = var_9270, x = mh_w_141_cast_fp16)[name = tensor("op_9410_cast_fp16")]; + tensor var_9411 = const()[name = tensor("op_9411"), val = tensor([1, 20, 64, -1])]; + tensor var_9412_cast_fp16 = reshape(shape = var_9411, x = value_93_cast_fp16)[name = tensor("op_9412_cast_fp16")]; + tensor attn_93_transpose_x_0 = const()[name = tensor("attn_93_transpose_x_0"), val = tensor(false)]; + tensor attn_93_transpose_y_0 = const()[name = tensor("attn_93_transpose_y_0"), val = tensor(true)]; + tensor attn_93_cast_fp16 = matmul(transpose_x = attn_93_transpose_x_0, transpose_y = attn_93_transpose_y_0, x = var_9412_cast_fp16, y = var_9410_cast_fp16)[name = tensor("attn_93_cast_fp16")]; + tensor var_9415 = const()[name = tensor("op_9415"), val = tensor([1, 1280, 1, -1])]; + tensor input_697_cast_fp16 = reshape(shape = var_9415, x = attn_93_cast_fp16)[name = tensor("input_697_cast_fp16")]; + tensor var_9422 = const()[name = tensor("op_9422"), val = tensor([1, 1])]; + tensor var_9424 = const()[name = tensor("op_9424"), val = tensor([1, 1])]; + tensor pretrained_out_467_pad_type_0 = const()[name = tensor("pretrained_out_467_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_467_pad_0 = const()[name = tensor("pretrained_out_467_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(465345152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466164416))), name = tensor("layers_23_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_23_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466164544)))]; + tensor pretrained_out_467_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_9424, groups = var_9277, pad = pretrained_out_467_pad_0, pad_type = pretrained_out_467_pad_type_0, strides = var_9422, weight = layers_23_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_697_cast_fp16)[name = tensor("pretrained_out_467_cast_fp16")]; + tensor var_9428 = const()[name = tensor("op_9428"), val = tensor([1, 1])]; + tensor var_9430 = const()[name = tensor("op_9430"), val = tensor([1, 1])]; + tensor input_699_pad_type_0 = const()[name = tensor("input_699_pad_type_0"), val = tensor("custom")]; + tensor input_699_pad_0 = const()[name = tensor("input_699_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466167168)))]; + tensor input_699_cast_fp16 = conv(dilations = var_9430, groups = var_9277, pad = input_699_pad_0, pad_type = input_699_pad_type_0, strides = var_9428, weight = layers_23_self_attn_o_proj_loraA_weight_to_fp16, x = input_697_cast_fp16)[name = tensor("input_699_cast_fp16")]; + tensor var_9434 = const()[name = tensor("op_9434"), val = tensor([1, 1])]; + tensor var_9436 = const()[name = tensor("op_9436"), val = tensor([1, 1])]; + tensor lora_out_933_pad_type_0 = const()[name = tensor("lora_out_933_pad_type_0"), val = tensor("custom")]; + tensor lora_out_933_pad_0 = const()[name = tensor("lora_out_933_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_935_weight_0_to_fp16 = const()[name = tensor("lora_out_935_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466208192)))]; + tensor lora_out_935_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9436, groups = var_9277, pad = lora_out_933_pad_0, pad_type = lora_out_933_pad_type_0, strides = var_9434, weight = lora_out_935_weight_0_to_fp16, x = input_699_cast_fp16)[name = tensor("lora_out_935_cast_fp16")]; + tensor obj_329_cast_fp16 = add(x = pretrained_out_467_cast_fp16, y = lora_out_935_cast_fp16)[name = tensor("obj_329_cast_fp16")]; + tensor inputs_141_cast_fp16 = add(x = inputs_139_cast_fp16, y = obj_329_cast_fp16)[name = tensor("inputs_141_cast_fp16")]; + tensor var_9449 = const()[name = tensor("op_9449"), val = tensor([1])]; + tensor channels_mean_141_cast_fp16 = reduce_mean(axes = var_9449, keep_dims = var_9278, x = inputs_141_cast_fp16)[name = tensor("channels_mean_141_cast_fp16")]; + tensor zero_mean_141_cast_fp16 = sub(x = inputs_141_cast_fp16, y = channels_mean_141_cast_fp16)[name = tensor("zero_mean_141_cast_fp16")]; + tensor zero_mean_sq_141_cast_fp16 = mul(x = zero_mean_141_cast_fp16, y = zero_mean_141_cast_fp16)[name = tensor("zero_mean_sq_141_cast_fp16")]; + tensor var_9453 = const()[name = tensor("op_9453"), val = tensor([1])]; + tensor var_9454_cast_fp16 = reduce_mean(axes = var_9453, keep_dims = var_9278, x = zero_mean_sq_141_cast_fp16)[name = tensor("op_9454_cast_fp16")]; + tensor var_9455_to_fp16 = const()[name = tensor("op_9455_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_9456_cast_fp16 = add(x = var_9454_cast_fp16, y = var_9455_to_fp16)[name = tensor("op_9456_cast_fp16")]; + tensor denom_141_epsilon_0 = const()[name = tensor("denom_141_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_141_cast_fp16 = rsqrt(epsilon = denom_141_epsilon_0, x = var_9456_cast_fp16)[name = tensor("denom_141_cast_fp16")]; + tensor out_141_cast_fp16 = mul(x = zero_mean_141_cast_fp16, y = denom_141_cast_fp16)[name = tensor("out_141_cast_fp16")]; + tensor obj_331_gamma_0_to_fp16 = const()[name = tensor("obj_331_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466249216)))]; + tensor obj_331_beta_0_to_fp16 = const()[name = tensor("obj_331_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466251840)))]; + tensor obj_331_epsilon_0_to_fp16 = const()[name = tensor("obj_331_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_331_cast_fp16 = batch_norm(beta = obj_331_beta_0_to_fp16, epsilon = obj_331_epsilon_0_to_fp16, gamma = obj_331_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_141_cast_fp16)[name = tensor("obj_331_cast_fp16")]; + tensor var_9474 = const()[name = tensor("op_9474"), val = tensor([1, 1])]; + tensor var_9476 = const()[name = tensor("op_9476"), val = tensor([1, 1])]; + tensor pretrained_out_469_pad_type_0 = const()[name = tensor("pretrained_out_469_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_469_pad_0 = const()[name = tensor("pretrained_out_469_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(466254464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467073728))), name = tensor("layers_23_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_23_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467073856)))]; + tensor pretrained_out_469_cast_fp16 = conv(bias = layers_23_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_9476, groups = var_9277, pad = pretrained_out_469_pad_0, pad_type = pretrained_out_469_pad_type_0, strides = var_9474, weight = layers_23_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_331_cast_fp16)[name = tensor("pretrained_out_469_cast_fp16")]; + tensor var_9480 = const()[name = tensor("op_9480"), val = tensor([1, 1])]; + tensor var_9482 = const()[name = tensor("op_9482"), val = tensor([1, 1])]; + tensor input_701_pad_type_0 = const()[name = tensor("input_701_pad_type_0"), val = tensor("custom")]; + tensor input_701_pad_0 = const()[name = tensor("input_701_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467076480)))]; + tensor input_701_cast_fp16 = conv(dilations = var_9482, groups = var_9277, pad = input_701_pad_0, pad_type = input_701_pad_type_0, strides = var_9480, weight = layers_23_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_331_cast_fp16)[name = tensor("input_701_cast_fp16")]; + tensor var_9486 = const()[name = tensor("op_9486"), val = tensor([1, 1])]; + tensor var_9488 = const()[name = tensor("op_9488"), val = tensor([1, 1])]; + tensor lora_out_937_pad_type_0 = const()[name = tensor("lora_out_937_pad_type_0"), val = tensor("custom")]; + tensor lora_out_937_pad_0 = const()[name = tensor("lora_out_937_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_939_weight_0_to_fp16 = const()[name = tensor("lora_out_939_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467117504)))]; + tensor lora_out_939_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9488, groups = var_9277, pad = lora_out_937_pad_0, pad_type = lora_out_937_pad_type_0, strides = var_9486, weight = lora_out_939_weight_0_to_fp16, x = input_701_cast_fp16)[name = tensor("lora_out_939_cast_fp16")]; + tensor query_95_cast_fp16 = add(x = pretrained_out_469_cast_fp16, y = lora_out_939_cast_fp16)[name = tensor("query_95_cast_fp16")]; + tensor var_9498 = const()[name = tensor("op_9498"), val = tensor([1, 1])]; + tensor var_9500 = const()[name = tensor("op_9500"), val = tensor([1, 1])]; + tensor pretrained_out_471_pad_type_0 = const()[name = tensor("pretrained_out_471_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_471_pad_0 = const()[name = tensor("pretrained_out_471_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467158528))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467977792))), name = tensor("layers_23_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_471_cast_fp16 = conv(dilations = var_9500, groups = var_9277, pad = pretrained_out_471_pad_0, pad_type = pretrained_out_471_pad_type_0, strides = var_9498, weight = layers_23_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_471_cast_fp16")]; + tensor var_9504 = const()[name = tensor("op_9504"), val = tensor([1, 1])]; + tensor var_9506 = const()[name = tensor("op_9506"), val = tensor([1, 1])]; + tensor input_703_pad_type_0 = const()[name = tensor("input_703_pad_type_0"), val = tensor("custom")]; + tensor input_703_pad_0 = const()[name = tensor("input_703_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(467977920)))]; + tensor input_703_cast_fp16 = conv(dilations = var_9506, groups = var_9277, pad = input_703_pad_0, pad_type = input_703_pad_type_0, strides = var_9504, weight = layers_23_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_703_cast_fp16")]; + tensor var_9510 = const()[name = tensor("op_9510"), val = tensor([1, 1])]; + tensor var_9512 = const()[name = tensor("op_9512"), val = tensor([1, 1])]; + tensor lora_out_941_pad_type_0 = const()[name = tensor("lora_out_941_pad_type_0"), val = tensor("custom")]; + tensor lora_out_941_pad_0 = const()[name = tensor("lora_out_941_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_943_weight_0_to_fp16 = const()[name = tensor("lora_out_943_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468018944)))]; + tensor lora_out_943_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9512, groups = var_9277, pad = lora_out_941_pad_0, pad_type = lora_out_941_pad_type_0, strides = var_9510, weight = lora_out_943_weight_0_to_fp16, x = input_703_cast_fp16)[name = tensor("lora_out_943_cast_fp16")]; + tensor key_95_cast_fp16 = add(x = pretrained_out_471_cast_fp16, y = lora_out_943_cast_fp16)[name = tensor("key_95_cast_fp16")]; + tensor var_9523 = const()[name = tensor("op_9523"), val = tensor([1, 1])]; + tensor var_9525 = const()[name = tensor("op_9525"), val = tensor([1, 1])]; + tensor pretrained_out_473_pad_type_0 = const()[name = tensor("pretrained_out_473_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_473_pad_0 = const()[name = tensor("pretrained_out_473_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468059968))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468879232))), name = tensor("layers_23_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_23_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468879360)))]; + tensor pretrained_out_473_cast_fp16 = conv(bias = layers_23_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_9525, groups = var_9277, pad = pretrained_out_473_pad_0, pad_type = pretrained_out_473_pad_type_0, strides = var_9523, weight = layers_23_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_473_cast_fp16")]; + tensor var_9529 = const()[name = tensor("op_9529"), val = tensor([1, 1])]; + tensor var_9531 = const()[name = tensor("op_9531"), val = tensor([1, 1])]; + tensor input_705_pad_type_0 = const()[name = tensor("input_705_pad_type_0"), val = tensor("custom")]; + tensor input_705_pad_0 = const()[name = tensor("input_705_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468881984)))]; + tensor input_705_cast_fp16 = conv(dilations = var_9531, groups = var_9277, pad = input_705_pad_0, pad_type = input_705_pad_type_0, strides = var_9529, weight = layers_23_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_705_cast_fp16")]; + tensor var_9535 = const()[name = tensor("op_9535"), val = tensor([1, 1])]; + tensor var_9537 = const()[name = tensor("op_9537"), val = tensor([1, 1])]; + tensor lora_out_945_pad_type_0 = const()[name = tensor("lora_out_945_pad_type_0"), val = tensor("custom")]; + tensor lora_out_945_pad_0 = const()[name = tensor("lora_out_945_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_947_weight_0_to_fp16 = const()[name = tensor("lora_out_947_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468923008)))]; + tensor lora_out_947_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9537, groups = var_9277, pad = lora_out_945_pad_0, pad_type = lora_out_945_pad_type_0, strides = var_9535, weight = lora_out_947_weight_0_to_fp16, x = input_705_cast_fp16)[name = tensor("lora_out_947_cast_fp16")]; + tensor value_95_cast_fp16 = add(x = pretrained_out_473_cast_fp16, y = lora_out_947_cast_fp16)[name = tensor("value_95_cast_fp16")]; + tensor var_9544 = const()[name = tensor("op_9544"), val = tensor([1, 20, 64, -1])]; + tensor var_9545_cast_fp16 = reshape(shape = var_9544, x = query_95_cast_fp16)[name = tensor("op_9545_cast_fp16")]; + tensor var_9546_to_fp16 = const()[name = tensor("op_9546_to_fp16"), val = tensor(0x1p-3)]; + tensor var_9547_cast_fp16 = mul(x = var_9545_cast_fp16, y = var_9546_to_fp16)[name = tensor("op_9547_cast_fp16")]; + tensor var_9548 = const()[name = tensor("op_9548"), val = tensor([1, 20, 64, -1])]; + tensor var_9549_cast_fp16 = reshape(shape = var_9548, x = key_95_cast_fp16)[name = tensor("op_9549_cast_fp16")]; + tensor mh_w_143_transpose_x_0 = const()[name = tensor("mh_w_143_transpose_x_0"), val = tensor(true)]; + tensor mh_w_143_transpose_y_0 = const()[name = tensor("mh_w_143_transpose_y_0"), val = tensor(false)]; + tensor mh_w_143_cast_fp16 = matmul(transpose_x = mh_w_143_transpose_x_0, transpose_y = mh_w_143_transpose_y_0, x = var_9547_cast_fp16, y = var_9549_cast_fp16)[name = tensor("mh_w_143_cast_fp16")]; + tensor obj_335_cast_fp16 = softmax(axis = var_9270, x = mh_w_143_cast_fp16)[name = tensor("obj_335_cast_fp16")]; + tensor var_9553 = const()[name = tensor("op_9553"), val = tensor([1, 20, 64, -1])]; + tensor var_9554_cast_fp16 = reshape(shape = var_9553, x = value_95_cast_fp16)[name = tensor("op_9554_cast_fp16")]; + tensor attn_95_transpose_x_0 = const()[name = tensor("attn_95_transpose_x_0"), val = tensor(false)]; + tensor attn_95_transpose_y_0 = const()[name = tensor("attn_95_transpose_y_0"), val = tensor(true)]; + tensor attn_95_cast_fp16 = matmul(transpose_x = attn_95_transpose_x_0, transpose_y = attn_95_transpose_y_0, x = var_9554_cast_fp16, y = obj_335_cast_fp16)[name = tensor("attn_95_cast_fp16")]; + tensor var_9557 = const()[name = tensor("op_9557"), val = tensor([1, 1280, 1, -1])]; + tensor input_707_cast_fp16 = reshape(shape = var_9557, x = attn_95_cast_fp16)[name = tensor("input_707_cast_fp16")]; + tensor var_9564 = const()[name = tensor("op_9564"), val = tensor([1, 1])]; + tensor var_9566 = const()[name = tensor("op_9566"), val = tensor([1, 1])]; + tensor pretrained_out_475_pad_type_0 = const()[name = tensor("pretrained_out_475_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_475_pad_0 = const()[name = tensor("pretrained_out_475_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468964032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469783296))), name = tensor("layers_23_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_23_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469783424)))]; + tensor pretrained_out_475_cast_fp16 = conv(bias = layers_23_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_9566, groups = var_9277, pad = pretrained_out_475_pad_0, pad_type = pretrained_out_475_pad_type_0, strides = var_9564, weight = layers_23_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = tensor("pretrained_out_475_cast_fp16")]; + tensor var_9570 = const()[name = tensor("op_9570"), val = tensor([1, 1])]; + tensor var_9572 = const()[name = tensor("op_9572"), val = tensor([1, 1])]; + tensor input_709_pad_type_0 = const()[name = tensor("input_709_pad_type_0"), val = tensor("custom")]; + tensor input_709_pad_0 = const()[name = tensor("input_709_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469786048)))]; + tensor input_709_cast_fp16 = conv(dilations = var_9572, groups = var_9277, pad = input_709_pad_0, pad_type = input_709_pad_type_0, strides = var_9570, weight = layers_23_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_707_cast_fp16)[name = tensor("input_709_cast_fp16")]; + tensor var_9576 = const()[name = tensor("op_9576"), val = tensor([1, 1])]; + tensor var_9578 = const()[name = tensor("op_9578"), val = tensor([1, 1])]; + tensor lora_out_949_pad_type_0 = const()[name = tensor("lora_out_949_pad_type_0"), val = tensor("custom")]; + tensor lora_out_949_pad_0 = const()[name = tensor("lora_out_949_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_951_weight_0_to_fp16 = const()[name = tensor("lora_out_951_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469827072)))]; + tensor lora_out_951_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9578, groups = var_9277, pad = lora_out_949_pad_0, pad_type = lora_out_949_pad_type_0, strides = var_9576, weight = lora_out_951_weight_0_to_fp16, x = input_709_cast_fp16)[name = tensor("lora_out_951_cast_fp16")]; + tensor obj_333_cast_fp16 = add(x = pretrained_out_475_cast_fp16, y = lora_out_951_cast_fp16)[name = tensor("obj_333_cast_fp16")]; + tensor inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = obj_333_cast_fp16)[name = tensor("inputs_143_cast_fp16")]; + tensor var_9590 = const()[name = tensor("op_9590"), val = tensor([1])]; + tensor channels_mean_143_cast_fp16 = reduce_mean(axes = var_9590, keep_dims = var_9278, x = inputs_143_cast_fp16)[name = tensor("channels_mean_143_cast_fp16")]; + tensor zero_mean_143_cast_fp16 = sub(x = inputs_143_cast_fp16, y = channels_mean_143_cast_fp16)[name = tensor("zero_mean_143_cast_fp16")]; + tensor zero_mean_sq_143_cast_fp16 = mul(x = zero_mean_143_cast_fp16, y = zero_mean_143_cast_fp16)[name = tensor("zero_mean_sq_143_cast_fp16")]; + tensor var_9594 = const()[name = tensor("op_9594"), val = tensor([1])]; + tensor var_9595_cast_fp16 = reduce_mean(axes = var_9594, keep_dims = var_9278, x = zero_mean_sq_143_cast_fp16)[name = tensor("op_9595_cast_fp16")]; + tensor var_9596_to_fp16 = const()[name = tensor("op_9596_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_9597_cast_fp16 = add(x = var_9595_cast_fp16, y = var_9596_to_fp16)[name = tensor("op_9597_cast_fp16")]; + tensor denom_143_epsilon_0 = const()[name = tensor("denom_143_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_143_cast_fp16 = rsqrt(epsilon = denom_143_epsilon_0, x = var_9597_cast_fp16)[name = tensor("denom_143_cast_fp16")]; + tensor out_143_cast_fp16 = mul(x = zero_mean_143_cast_fp16, y = denom_143_cast_fp16)[name = tensor("out_143_cast_fp16")]; + tensor input_711_gamma_0_to_fp16 = const()[name = tensor("input_711_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469868096)))]; + tensor input_711_beta_0_to_fp16 = const()[name = tensor("input_711_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469870720)))]; + tensor input_711_epsilon_0_to_fp16 = const()[name = tensor("input_711_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_711_cast_fp16 = batch_norm(beta = input_711_beta_0_to_fp16, epsilon = input_711_epsilon_0_to_fp16, gamma = input_711_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_143_cast_fp16)[name = tensor("input_711_cast_fp16")]; + tensor var_9611 = const()[name = tensor("op_9611"), val = tensor([1, 1])]; + tensor var_9613 = const()[name = tensor("op_9613"), val = tensor([1, 1])]; + tensor pretrained_out_477_pad_type_0 = const()[name = tensor("pretrained_out_477_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_477_pad_0 = const()[name = tensor("pretrained_out_477_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469873344))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473150208))), name = tensor("layers_23_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_23_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473150336)))]; + tensor pretrained_out_477_cast_fp16 = conv(bias = layers_23_fc1_pretrained_bias_to_fp16, dilations = var_9613, groups = var_9277, pad = pretrained_out_477_pad_0, pad_type = pretrained_out_477_pad_type_0, strides = var_9611, weight = layers_23_fc1_pretrained_weight_to_fp16_palettized, x = input_711_cast_fp16)[name = tensor("pretrained_out_477_cast_fp16")]; + tensor var_9617 = const()[name = tensor("op_9617"), val = tensor([1, 1])]; + tensor var_9619 = const()[name = tensor("op_9619"), val = tensor([1, 1])]; + tensor input_713_pad_type_0 = const()[name = tensor("input_713_pad_type_0"), val = tensor("custom")]; + tensor input_713_pad_0 = const()[name = tensor("input_713_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_23_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473160640)))]; + tensor input_713_cast_fp16 = conv(dilations = var_9619, groups = var_9277, pad = input_713_pad_0, pad_type = input_713_pad_type_0, strides = var_9617, weight = layers_23_fc1_loraA_weight_to_fp16, x = input_711_cast_fp16)[name = tensor("input_713_cast_fp16")]; + tensor var_9623 = const()[name = tensor("op_9623"), val = tensor([1, 1])]; + tensor var_9625 = const()[name = tensor("op_9625"), val = tensor([1, 1])]; + tensor lora_out_953_pad_type_0 = const()[name = tensor("lora_out_953_pad_type_0"), val = tensor("custom")]; + tensor lora_out_953_pad_0 = const()[name = tensor("lora_out_953_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_955_weight_0_to_fp16 = const()[name = tensor("lora_out_955_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473201664)))]; + tensor lora_out_955_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_9625, groups = var_9277, pad = lora_out_953_pad_0, pad_type = lora_out_953_pad_type_0, strides = var_9623, weight = lora_out_955_weight_0_to_fp16, x = input_713_cast_fp16)[name = tensor("lora_out_955_cast_fp16")]; + tensor input_715_cast_fp16 = add(x = pretrained_out_477_cast_fp16, y = lora_out_955_cast_fp16)[name = tensor("input_715_cast_fp16")]; + tensor input_717_mode_0 = const()[name = tensor("input_717_mode_0"), val = tensor("EXACT")]; + tensor input_717_cast_fp16 = gelu(mode = input_717_mode_0, x = input_715_cast_fp16)[name = tensor("input_717_cast_fp16")]; + tensor var_9637 = const()[name = tensor("op_9637"), val = tensor([1, 1])]; + tensor var_9639 = const()[name = tensor("op_9639"), val = tensor([1, 1])]; + tensor pretrained_out_479_pad_type_0 = const()[name = tensor("pretrained_out_479_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_479_pad_0 = const()[name = tensor("pretrained_out_479_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(473365568))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476642432))), name = tensor("layers_23_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_23_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_23_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476642560)))]; + tensor pretrained_out_479_cast_fp16 = conv(bias = layers_23_fc2_pretrained_bias_to_fp16, dilations = var_9639, groups = var_9277, pad = pretrained_out_479_pad_0, pad_type = pretrained_out_479_pad_type_0, strides = var_9637, weight = layers_23_fc2_pretrained_weight_to_fp16_palettized, x = input_717_cast_fp16)[name = tensor("pretrained_out_479_cast_fp16")]; + tensor var_9643 = const()[name = tensor("op_9643"), val = tensor([1, 1])]; + tensor var_9645 = const()[name = tensor("op_9645"), val = tensor([1, 1])]; + tensor input_719_pad_type_0 = const()[name = tensor("input_719_pad_type_0"), val = tensor("custom")]; + tensor input_719_pad_0 = const()[name = tensor("input_719_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_23_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_23_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476645184)))]; + tensor input_719_cast_fp16 = conv(dilations = var_9645, groups = var_9277, pad = input_719_pad_0, pad_type = input_719_pad_type_0, strides = var_9643, weight = layers_23_fc2_loraA_weight_to_fp16, x = input_717_cast_fp16)[name = tensor("input_719_cast_fp16")]; + tensor var_9649 = const()[name = tensor("op_9649"), val = tensor([1, 1])]; + tensor var_9651 = const()[name = tensor("op_9651"), val = tensor([1, 1])]; + tensor lora_out_957_pad_type_0 = const()[name = tensor("lora_out_957_pad_type_0"), val = tensor("custom")]; + tensor lora_out_957_pad_0 = const()[name = tensor("lora_out_957_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_959_weight_0_to_fp16 = const()[name = tensor("lora_out_959_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476809088)))]; + tensor lora_out_959_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9651, groups = var_9277, pad = lora_out_957_pad_0, pad_type = lora_out_957_pad_type_0, strides = var_9649, weight = lora_out_959_weight_0_to_fp16, x = input_719_cast_fp16)[name = tensor("lora_out_959_cast_fp16")]; + tensor hidden_states_49_cast_fp16 = add(x = pretrained_out_479_cast_fp16, y = lora_out_959_cast_fp16)[name = tensor("hidden_states_49_cast_fp16")]; + tensor inputs_145_cast_fp16 = add(x = inputs_143_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor("inputs_145_cast_fp16")]; + tensor var_9668 = const()[name = tensor("op_9668"), val = tensor(3)]; + tensor var_9675 = const()[name = tensor("op_9675"), val = tensor(1)]; + tensor var_9676 = const()[name = tensor("op_9676"), val = tensor(true)]; + tensor var_9688 = const()[name = tensor("op_9688"), val = tensor([1])]; + tensor channels_mean_145_cast_fp16 = reduce_mean(axes = var_9688, keep_dims = var_9676, x = inputs_145_cast_fp16)[name = tensor("channels_mean_145_cast_fp16")]; + tensor zero_mean_145_cast_fp16 = sub(x = inputs_145_cast_fp16, y = channels_mean_145_cast_fp16)[name = tensor("zero_mean_145_cast_fp16")]; + tensor zero_mean_sq_145_cast_fp16 = mul(x = zero_mean_145_cast_fp16, y = zero_mean_145_cast_fp16)[name = tensor("zero_mean_sq_145_cast_fp16")]; + tensor var_9692 = const()[name = tensor("op_9692"), val = tensor([1])]; + tensor var_9693_cast_fp16 = reduce_mean(axes = var_9692, keep_dims = var_9676, x = zero_mean_sq_145_cast_fp16)[name = tensor("op_9693_cast_fp16")]; + tensor var_9694_to_fp16 = const()[name = tensor("op_9694_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_9695_cast_fp16 = add(x = var_9693_cast_fp16, y = var_9694_to_fp16)[name = tensor("op_9695_cast_fp16")]; + tensor denom_145_epsilon_0 = const()[name = tensor("denom_145_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_145_cast_fp16 = rsqrt(epsilon = denom_145_epsilon_0, x = var_9695_cast_fp16)[name = tensor("denom_145_cast_fp16")]; + tensor out_145_cast_fp16 = mul(x = zero_mean_145_cast_fp16, y = denom_145_cast_fp16)[name = tensor("out_145_cast_fp16")]; + tensor obj_337_gamma_0_to_fp16 = const()[name = tensor("obj_337_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476850112)))]; + tensor obj_337_beta_0_to_fp16 = const()[name = tensor("obj_337_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476852736)))]; + tensor obj_337_epsilon_0_to_fp16 = const()[name = tensor("obj_337_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_337_cast_fp16 = batch_norm(beta = obj_337_beta_0_to_fp16, epsilon = obj_337_epsilon_0_to_fp16, gamma = obj_337_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_145_cast_fp16)[name = tensor("obj_337_cast_fp16")]; + tensor var_9713 = const()[name = tensor("op_9713"), val = tensor([1, 1])]; + tensor var_9715 = const()[name = tensor("op_9715"), val = tensor([1, 1])]; + tensor pretrained_out_481_pad_type_0 = const()[name = tensor("pretrained_out_481_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_481_pad_0 = const()[name = tensor("pretrained_out_481_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476855360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477674624))), name = tensor("layers_24_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_24_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477674752)))]; + tensor pretrained_out_481_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_9715, groups = var_9675, pad = pretrained_out_481_pad_0, pad_type = pretrained_out_481_pad_type_0, strides = var_9713, weight = layers_24_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_337_cast_fp16)[name = tensor("pretrained_out_481_cast_fp16")]; + tensor var_9719 = const()[name = tensor("op_9719"), val = tensor([1, 1])]; + tensor var_9721 = const()[name = tensor("op_9721"), val = tensor([1, 1])]; + tensor input_721_pad_type_0 = const()[name = tensor("input_721_pad_type_0"), val = tensor("custom")]; + tensor input_721_pad_0 = const()[name = tensor("input_721_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477677376)))]; + tensor input_721_cast_fp16 = conv(dilations = var_9721, groups = var_9675, pad = input_721_pad_0, pad_type = input_721_pad_type_0, strides = var_9719, weight = layers_24_self_attn_q_proj_loraA_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("input_721_cast_fp16")]; + tensor var_9725 = const()[name = tensor("op_9725"), val = tensor([1, 1])]; + tensor var_9727 = const()[name = tensor("op_9727"), val = tensor([1, 1])]; + tensor lora_out_961_pad_type_0 = const()[name = tensor("lora_out_961_pad_type_0"), val = tensor("custom")]; + tensor lora_out_961_pad_0 = const()[name = tensor("lora_out_961_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_963_weight_0_to_fp16 = const()[name = tensor("lora_out_963_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477718400)))]; + tensor lora_out_963_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9727, groups = var_9675, pad = lora_out_961_pad_0, pad_type = lora_out_961_pad_type_0, strides = var_9725, weight = lora_out_963_weight_0_to_fp16, x = input_721_cast_fp16)[name = tensor("lora_out_963_cast_fp16")]; + tensor query_97_cast_fp16 = add(x = pretrained_out_481_cast_fp16, y = lora_out_963_cast_fp16)[name = tensor("query_97_cast_fp16")]; + tensor var_9737 = const()[name = tensor("op_9737"), val = tensor([1, 1])]; + tensor var_9739 = const()[name = tensor("op_9739"), val = tensor([1, 1])]; + tensor pretrained_out_483_pad_type_0 = const()[name = tensor("pretrained_out_483_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_483_pad_0 = const()[name = tensor("pretrained_out_483_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(477759424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478578688))), name = tensor("layers_24_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_483_cast_fp16 = conv(dilations = var_9739, groups = var_9675, pad = pretrained_out_483_pad_0, pad_type = pretrained_out_483_pad_type_0, strides = var_9737, weight = layers_24_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_337_cast_fp16)[name = tensor("pretrained_out_483_cast_fp16")]; + tensor var_9743 = const()[name = tensor("op_9743"), val = tensor([1, 1])]; + tensor var_9745 = const()[name = tensor("op_9745"), val = tensor([1, 1])]; + tensor input_723_pad_type_0 = const()[name = tensor("input_723_pad_type_0"), val = tensor("custom")]; + tensor input_723_pad_0 = const()[name = tensor("input_723_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478578816)))]; + tensor input_723_cast_fp16 = conv(dilations = var_9745, groups = var_9675, pad = input_723_pad_0, pad_type = input_723_pad_type_0, strides = var_9743, weight = layers_24_self_attn_k_proj_loraA_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("input_723_cast_fp16")]; + tensor var_9749 = const()[name = tensor("op_9749"), val = tensor([1, 1])]; + tensor var_9751 = const()[name = tensor("op_9751"), val = tensor([1, 1])]; + tensor lora_out_965_pad_type_0 = const()[name = tensor("lora_out_965_pad_type_0"), val = tensor("custom")]; + tensor lora_out_965_pad_0 = const()[name = tensor("lora_out_965_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_967_weight_0_to_fp16 = const()[name = tensor("lora_out_967_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478619840)))]; + tensor lora_out_967_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9751, groups = var_9675, pad = lora_out_965_pad_0, pad_type = lora_out_965_pad_type_0, strides = var_9749, weight = lora_out_967_weight_0_to_fp16, x = input_723_cast_fp16)[name = tensor("lora_out_967_cast_fp16")]; + tensor current_key_49_cast_fp16 = add(x = pretrained_out_483_cast_fp16, y = lora_out_967_cast_fp16)[name = tensor("current_key_49_cast_fp16")]; + tensor var_9762 = const()[name = tensor("op_9762"), val = tensor([1, 1])]; + tensor var_9764 = const()[name = tensor("op_9764"), val = tensor([1, 1])]; + tensor pretrained_out_485_pad_type_0 = const()[name = tensor("pretrained_out_485_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_485_pad_0 = const()[name = tensor("pretrained_out_485_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478660864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479480128))), name = tensor("layers_24_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_24_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479480256)))]; + tensor pretrained_out_485_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_9764, groups = var_9675, pad = pretrained_out_485_pad_0, pad_type = pretrained_out_485_pad_type_0, strides = var_9762, weight = layers_24_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_337_cast_fp16)[name = tensor("pretrained_out_485_cast_fp16")]; + tensor var_9768 = const()[name = tensor("op_9768"), val = tensor([1, 1])]; + tensor var_9770 = const()[name = tensor("op_9770"), val = tensor([1, 1])]; + tensor input_725_pad_type_0 = const()[name = tensor("input_725_pad_type_0"), val = tensor("custom")]; + tensor input_725_pad_0 = const()[name = tensor("input_725_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479482880)))]; + tensor input_725_cast_fp16 = conv(dilations = var_9770, groups = var_9675, pad = input_725_pad_0, pad_type = input_725_pad_type_0, strides = var_9768, weight = layers_24_self_attn_v_proj_loraA_weight_to_fp16, x = obj_337_cast_fp16)[name = tensor("input_725_cast_fp16")]; + tensor var_9774 = const()[name = tensor("op_9774"), val = tensor([1, 1])]; + tensor var_9776 = const()[name = tensor("op_9776"), val = tensor([1, 1])]; + tensor lora_out_969_pad_type_0 = const()[name = tensor("lora_out_969_pad_type_0"), val = tensor("custom")]; + tensor lora_out_969_pad_0 = const()[name = tensor("lora_out_969_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_971_weight_0_to_fp16 = const()[name = tensor("lora_out_971_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479523904)))]; + tensor lora_out_971_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9776, groups = var_9675, pad = lora_out_969_pad_0, pad_type = lora_out_969_pad_type_0, strides = var_9774, weight = lora_out_971_weight_0_to_fp16, x = input_725_cast_fp16)[name = tensor("lora_out_971_cast_fp16")]; + tensor current_value_49_cast_fp16 = add(x = pretrained_out_485_cast_fp16, y = lora_out_971_cast_fp16)[name = tensor("current_value_49_cast_fp16")]; + tensor var_9786_cast_fp16 = mul(x = current_key_49_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_9786_cast_fp16")]; + tensor var_9788_cast_fp16 = mul(x = var_103_cast_fp16_24, y = var_295_cast_fp16)[name = tensor("op_9788_cast_fp16")]; + tensor key_97_cast_fp16 = add(x = var_9786_cast_fp16, y = var_9788_cast_fp16)[name = tensor("key_97_cast_fp16")]; + tensor var_9790_cast_fp16 = mul(x = current_value_49_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_9790_cast_fp16")]; + tensor var_9792_cast_fp16 = mul(x = var_138_cast_fp16_24, y = var_295_cast_fp16)[name = tensor("op_9792_cast_fp16")]; + tensor value_97_cast_fp16 = add(x = var_9790_cast_fp16, y = var_9792_cast_fp16)[name = tensor("value_97_cast_fp16")]; + tensor var_9795 = const()[name = tensor("op_9795"), val = tensor([1, 20, 64, -1])]; + tensor var_9796_cast_fp16 = reshape(shape = var_9795, x = query_97_cast_fp16)[name = tensor("op_9796_cast_fp16")]; + tensor var_9797_to_fp16 = const()[name = tensor("op_9797_to_fp16"), val = tensor(0x1p-3)]; + tensor var_9798_cast_fp16 = mul(x = var_9796_cast_fp16, y = var_9797_to_fp16)[name = tensor("op_9798_cast_fp16")]; + tensor var_9799 = const()[name = tensor("op_9799"), val = tensor([1, 20, 64, -1])]; + tensor var_9800_cast_fp16 = reshape(shape = var_9799, x = key_97_cast_fp16)[name = tensor("op_9800_cast_fp16")]; + tensor mh_w_145_transpose_x_0 = const()[name = tensor("mh_w_145_transpose_x_0"), val = tensor(true)]; + tensor mh_w_145_transpose_y_0 = const()[name = tensor("mh_w_145_transpose_y_0"), val = tensor(false)]; + tensor mh_w_145_cast_fp16 = matmul(transpose_x = mh_w_145_transpose_x_0, transpose_y = mh_w_145_transpose_y_0, x = var_9798_cast_fp16, y = var_9800_cast_fp16)[name = tensor("mh_w_145_cast_fp16")]; + tensor mh_w_147_cast_fp16 = add(x = mh_w_145_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_147_cast_fp16")]; + tensor var_9808_cast_fp16 = softmax(axis = var_9668, x = mh_w_147_cast_fp16)[name = tensor("op_9808_cast_fp16")]; + tensor var_9809 = const()[name = tensor("op_9809"), val = tensor([1, 20, 64, -1])]; + tensor var_9810_cast_fp16 = reshape(shape = var_9809, x = value_97_cast_fp16)[name = tensor("op_9810_cast_fp16")]; + tensor attn_97_transpose_x_0 = const()[name = tensor("attn_97_transpose_x_0"), val = tensor(false)]; + tensor attn_97_transpose_y_0 = const()[name = tensor("attn_97_transpose_y_0"), val = tensor(true)]; + tensor attn_97_cast_fp16 = matmul(transpose_x = attn_97_transpose_x_0, transpose_y = attn_97_transpose_y_0, x = var_9810_cast_fp16, y = var_9808_cast_fp16)[name = tensor("attn_97_cast_fp16")]; + tensor var_9813 = const()[name = tensor("op_9813"), val = tensor([1, 1280, 1, -1])]; + tensor input_727_cast_fp16 = reshape(shape = var_9813, x = attn_97_cast_fp16)[name = tensor("input_727_cast_fp16")]; + tensor var_9820 = const()[name = tensor("op_9820"), val = tensor([1, 1])]; + tensor var_9822 = const()[name = tensor("op_9822"), val = tensor([1, 1])]; + tensor pretrained_out_487_pad_type_0 = const()[name = tensor("pretrained_out_487_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_487_pad_0 = const()[name = tensor("pretrained_out_487_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479564928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480384192))), name = tensor("layers_24_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_24_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480384320)))]; + tensor pretrained_out_487_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_9822, groups = var_9675, pad = pretrained_out_487_pad_0, pad_type = pretrained_out_487_pad_type_0, strides = var_9820, weight = layers_24_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_727_cast_fp16)[name = tensor("pretrained_out_487_cast_fp16")]; + tensor var_9826 = const()[name = tensor("op_9826"), val = tensor([1, 1])]; + tensor var_9828 = const()[name = tensor("op_9828"), val = tensor([1, 1])]; + tensor input_729_pad_type_0 = const()[name = tensor("input_729_pad_type_0"), val = tensor("custom")]; + tensor input_729_pad_0 = const()[name = tensor("input_729_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480386944)))]; + tensor input_729_cast_fp16 = conv(dilations = var_9828, groups = var_9675, pad = input_729_pad_0, pad_type = input_729_pad_type_0, strides = var_9826, weight = layers_24_self_attn_o_proj_loraA_weight_to_fp16, x = input_727_cast_fp16)[name = tensor("input_729_cast_fp16")]; + tensor var_9832 = const()[name = tensor("op_9832"), val = tensor([1, 1])]; + tensor var_9834 = const()[name = tensor("op_9834"), val = tensor([1, 1])]; + tensor lora_out_973_pad_type_0 = const()[name = tensor("lora_out_973_pad_type_0"), val = tensor("custom")]; + tensor lora_out_973_pad_0 = const()[name = tensor("lora_out_973_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_975_weight_0_to_fp16 = const()[name = tensor("lora_out_975_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480427968)))]; + tensor lora_out_975_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9834, groups = var_9675, pad = lora_out_973_pad_0, pad_type = lora_out_973_pad_type_0, strides = var_9832, weight = lora_out_975_weight_0_to_fp16, x = input_729_cast_fp16)[name = tensor("lora_out_975_cast_fp16")]; + tensor obj_343_cast_fp16 = add(x = pretrained_out_487_cast_fp16, y = lora_out_975_cast_fp16)[name = tensor("obj_343_cast_fp16")]; + tensor inputs_147_cast_fp16 = add(x = inputs_145_cast_fp16, y = obj_343_cast_fp16)[name = tensor("inputs_147_cast_fp16")]; + tensor var_9847 = const()[name = tensor("op_9847"), val = tensor([1])]; + tensor channels_mean_147_cast_fp16 = reduce_mean(axes = var_9847, keep_dims = var_9676, x = inputs_147_cast_fp16)[name = tensor("channels_mean_147_cast_fp16")]; + tensor zero_mean_147_cast_fp16 = sub(x = inputs_147_cast_fp16, y = channels_mean_147_cast_fp16)[name = tensor("zero_mean_147_cast_fp16")]; + tensor zero_mean_sq_147_cast_fp16 = mul(x = zero_mean_147_cast_fp16, y = zero_mean_147_cast_fp16)[name = tensor("zero_mean_sq_147_cast_fp16")]; + tensor var_9851 = const()[name = tensor("op_9851"), val = tensor([1])]; + tensor var_9852_cast_fp16 = reduce_mean(axes = var_9851, keep_dims = var_9676, x = zero_mean_sq_147_cast_fp16)[name = tensor("op_9852_cast_fp16")]; + tensor var_9853_to_fp16 = const()[name = tensor("op_9853_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_9854_cast_fp16 = add(x = var_9852_cast_fp16, y = var_9853_to_fp16)[name = tensor("op_9854_cast_fp16")]; + tensor denom_147_epsilon_0 = const()[name = tensor("denom_147_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_147_cast_fp16 = rsqrt(epsilon = denom_147_epsilon_0, x = var_9854_cast_fp16)[name = tensor("denom_147_cast_fp16")]; + tensor out_147_cast_fp16 = mul(x = zero_mean_147_cast_fp16, y = denom_147_cast_fp16)[name = tensor("out_147_cast_fp16")]; + tensor obj_345_gamma_0_to_fp16 = const()[name = tensor("obj_345_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480468992)))]; + tensor obj_345_beta_0_to_fp16 = const()[name = tensor("obj_345_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480471616)))]; + tensor obj_345_epsilon_0_to_fp16 = const()[name = tensor("obj_345_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_345_cast_fp16 = batch_norm(beta = obj_345_beta_0_to_fp16, epsilon = obj_345_epsilon_0_to_fp16, gamma = obj_345_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_147_cast_fp16)[name = tensor("obj_345_cast_fp16")]; + tensor var_9872 = const()[name = tensor("op_9872"), val = tensor([1, 1])]; + tensor var_9874 = const()[name = tensor("op_9874"), val = tensor([1, 1])]; + tensor pretrained_out_489_pad_type_0 = const()[name = tensor("pretrained_out_489_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_489_pad_0 = const()[name = tensor("pretrained_out_489_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480474240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481293504))), name = tensor("layers_24_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_24_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481293632)))]; + tensor pretrained_out_489_cast_fp16 = conv(bias = layers_24_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_9874, groups = var_9675, pad = pretrained_out_489_pad_0, pad_type = pretrained_out_489_pad_type_0, strides = var_9872, weight = layers_24_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_345_cast_fp16)[name = tensor("pretrained_out_489_cast_fp16")]; + tensor var_9878 = const()[name = tensor("op_9878"), val = tensor([1, 1])]; + tensor var_9880 = const()[name = tensor("op_9880"), val = tensor([1, 1])]; + tensor input_731_pad_type_0 = const()[name = tensor("input_731_pad_type_0"), val = tensor("custom")]; + tensor input_731_pad_0 = const()[name = tensor("input_731_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481296256)))]; + tensor input_731_cast_fp16 = conv(dilations = var_9880, groups = var_9675, pad = input_731_pad_0, pad_type = input_731_pad_type_0, strides = var_9878, weight = layers_24_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_345_cast_fp16)[name = tensor("input_731_cast_fp16")]; + tensor var_9884 = const()[name = tensor("op_9884"), val = tensor([1, 1])]; + tensor var_9886 = const()[name = tensor("op_9886"), val = tensor([1, 1])]; + tensor lora_out_977_pad_type_0 = const()[name = tensor("lora_out_977_pad_type_0"), val = tensor("custom")]; + tensor lora_out_977_pad_0 = const()[name = tensor("lora_out_977_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_979_weight_0_to_fp16 = const()[name = tensor("lora_out_979_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481337280)))]; + tensor lora_out_979_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9886, groups = var_9675, pad = lora_out_977_pad_0, pad_type = lora_out_977_pad_type_0, strides = var_9884, weight = lora_out_979_weight_0_to_fp16, x = input_731_cast_fp16)[name = tensor("lora_out_979_cast_fp16")]; + tensor query_99_cast_fp16 = add(x = pretrained_out_489_cast_fp16, y = lora_out_979_cast_fp16)[name = tensor("query_99_cast_fp16")]; + tensor var_9896 = const()[name = tensor("op_9896"), val = tensor([1, 1])]; + tensor var_9898 = const()[name = tensor("op_9898"), val = tensor([1, 1])]; + tensor pretrained_out_491_pad_type_0 = const()[name = tensor("pretrained_out_491_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_491_pad_0 = const()[name = tensor("pretrained_out_491_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481378304))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482197568))), name = tensor("layers_24_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_491_cast_fp16 = conv(dilations = var_9898, groups = var_9675, pad = pretrained_out_491_pad_0, pad_type = pretrained_out_491_pad_type_0, strides = var_9896, weight = layers_24_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_491_cast_fp16")]; + tensor var_9902 = const()[name = tensor("op_9902"), val = tensor([1, 1])]; + tensor var_9904 = const()[name = tensor("op_9904"), val = tensor([1, 1])]; + tensor input_733_pad_type_0 = const()[name = tensor("input_733_pad_type_0"), val = tensor("custom")]; + tensor input_733_pad_0 = const()[name = tensor("input_733_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482197696)))]; + tensor input_733_cast_fp16 = conv(dilations = var_9904, groups = var_9675, pad = input_733_pad_0, pad_type = input_733_pad_type_0, strides = var_9902, weight = layers_24_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_733_cast_fp16")]; + tensor var_9908 = const()[name = tensor("op_9908"), val = tensor([1, 1])]; + tensor var_9910 = const()[name = tensor("op_9910"), val = tensor([1, 1])]; + tensor lora_out_981_pad_type_0 = const()[name = tensor("lora_out_981_pad_type_0"), val = tensor("custom")]; + tensor lora_out_981_pad_0 = const()[name = tensor("lora_out_981_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_983_weight_0_to_fp16 = const()[name = tensor("lora_out_983_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482238720)))]; + tensor lora_out_983_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9910, groups = var_9675, pad = lora_out_981_pad_0, pad_type = lora_out_981_pad_type_0, strides = var_9908, weight = lora_out_983_weight_0_to_fp16, x = input_733_cast_fp16)[name = tensor("lora_out_983_cast_fp16")]; + tensor key_99_cast_fp16 = add(x = pretrained_out_491_cast_fp16, y = lora_out_983_cast_fp16)[name = tensor("key_99_cast_fp16")]; + tensor var_9921 = const()[name = tensor("op_9921"), val = tensor([1, 1])]; + tensor var_9923 = const()[name = tensor("op_9923"), val = tensor([1, 1])]; + tensor pretrained_out_493_pad_type_0 = const()[name = tensor("pretrained_out_493_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_493_pad_0 = const()[name = tensor("pretrained_out_493_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482279744))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483099008))), name = tensor("layers_24_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_24_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483099136)))]; + tensor pretrained_out_493_cast_fp16 = conv(bias = layers_24_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_9923, groups = var_9675, pad = pretrained_out_493_pad_0, pad_type = pretrained_out_493_pad_type_0, strides = var_9921, weight = layers_24_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_493_cast_fp16")]; + tensor var_9927 = const()[name = tensor("op_9927"), val = tensor([1, 1])]; + tensor var_9929 = const()[name = tensor("op_9929"), val = tensor([1, 1])]; + tensor input_735_pad_type_0 = const()[name = tensor("input_735_pad_type_0"), val = tensor("custom")]; + tensor input_735_pad_0 = const()[name = tensor("input_735_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483101760)))]; + tensor input_735_cast_fp16 = conv(dilations = var_9929, groups = var_9675, pad = input_735_pad_0, pad_type = input_735_pad_type_0, strides = var_9927, weight = layers_24_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_735_cast_fp16")]; + tensor var_9933 = const()[name = tensor("op_9933"), val = tensor([1, 1])]; + tensor var_9935 = const()[name = tensor("op_9935"), val = tensor([1, 1])]; + tensor lora_out_985_pad_type_0 = const()[name = tensor("lora_out_985_pad_type_0"), val = tensor("custom")]; + tensor lora_out_985_pad_0 = const()[name = tensor("lora_out_985_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_987_weight_0_to_fp16 = const()[name = tensor("lora_out_987_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483142784)))]; + tensor lora_out_987_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9935, groups = var_9675, pad = lora_out_985_pad_0, pad_type = lora_out_985_pad_type_0, strides = var_9933, weight = lora_out_987_weight_0_to_fp16, x = input_735_cast_fp16)[name = tensor("lora_out_987_cast_fp16")]; + tensor value_99_cast_fp16 = add(x = pretrained_out_493_cast_fp16, y = lora_out_987_cast_fp16)[name = tensor("value_99_cast_fp16")]; + tensor var_9942 = const()[name = tensor("op_9942"), val = tensor([1, 20, 64, -1])]; + tensor var_9943_cast_fp16 = reshape(shape = var_9942, x = query_99_cast_fp16)[name = tensor("op_9943_cast_fp16")]; + tensor var_9944_to_fp16 = const()[name = tensor("op_9944_to_fp16"), val = tensor(0x1p-3)]; + tensor var_9945_cast_fp16 = mul(x = var_9943_cast_fp16, y = var_9944_to_fp16)[name = tensor("op_9945_cast_fp16")]; + tensor var_9946 = const()[name = tensor("op_9946"), val = tensor([1, 20, 64, -1])]; + tensor var_9947_cast_fp16 = reshape(shape = var_9946, x = key_99_cast_fp16)[name = tensor("op_9947_cast_fp16")]; + tensor mh_w_149_transpose_x_0 = const()[name = tensor("mh_w_149_transpose_x_0"), val = tensor(true)]; + tensor mh_w_149_transpose_y_0 = const()[name = tensor("mh_w_149_transpose_y_0"), val = tensor(false)]; + tensor mh_w_149_cast_fp16 = matmul(transpose_x = mh_w_149_transpose_x_0, transpose_y = mh_w_149_transpose_y_0, x = var_9945_cast_fp16, y = var_9947_cast_fp16)[name = tensor("mh_w_149_cast_fp16")]; + tensor obj_349_cast_fp16 = softmax(axis = var_9668, x = mh_w_149_cast_fp16)[name = tensor("obj_349_cast_fp16")]; + tensor var_9951 = const()[name = tensor("op_9951"), val = tensor([1, 20, 64, -1])]; + tensor var_9952_cast_fp16 = reshape(shape = var_9951, x = value_99_cast_fp16)[name = tensor("op_9952_cast_fp16")]; + tensor attn_99_transpose_x_0 = const()[name = tensor("attn_99_transpose_x_0"), val = tensor(false)]; + tensor attn_99_transpose_y_0 = const()[name = tensor("attn_99_transpose_y_0"), val = tensor(true)]; + tensor attn_99_cast_fp16 = matmul(transpose_x = attn_99_transpose_x_0, transpose_y = attn_99_transpose_y_0, x = var_9952_cast_fp16, y = obj_349_cast_fp16)[name = tensor("attn_99_cast_fp16")]; + tensor var_9955 = const()[name = tensor("op_9955"), val = tensor([1, 1280, 1, -1])]; + tensor input_737_cast_fp16 = reshape(shape = var_9955, x = attn_99_cast_fp16)[name = tensor("input_737_cast_fp16")]; + tensor var_9962 = const()[name = tensor("op_9962"), val = tensor([1, 1])]; + tensor var_9964 = const()[name = tensor("op_9964"), val = tensor([1, 1])]; + tensor pretrained_out_495_pad_type_0 = const()[name = tensor("pretrained_out_495_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_495_pad_0 = const()[name = tensor("pretrained_out_495_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483183808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484003072))), name = tensor("layers_24_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_24_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484003200)))]; + tensor pretrained_out_495_cast_fp16 = conv(bias = layers_24_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_9964, groups = var_9675, pad = pretrained_out_495_pad_0, pad_type = pretrained_out_495_pad_type_0, strides = var_9962, weight = layers_24_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_737_cast_fp16)[name = tensor("pretrained_out_495_cast_fp16")]; + tensor var_9968 = const()[name = tensor("op_9968"), val = tensor([1, 1])]; + tensor var_9970 = const()[name = tensor("op_9970"), val = tensor([1, 1])]; + tensor input_739_pad_type_0 = const()[name = tensor("input_739_pad_type_0"), val = tensor("custom")]; + tensor input_739_pad_0 = const()[name = tensor("input_739_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_24_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484005824)))]; + tensor input_739_cast_fp16 = conv(dilations = var_9970, groups = var_9675, pad = input_739_pad_0, pad_type = input_739_pad_type_0, strides = var_9968, weight = layers_24_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_737_cast_fp16)[name = tensor("input_739_cast_fp16")]; + tensor var_9974 = const()[name = tensor("op_9974"), val = tensor([1, 1])]; + tensor var_9976 = const()[name = tensor("op_9976"), val = tensor([1, 1])]; + tensor lora_out_989_pad_type_0 = const()[name = tensor("lora_out_989_pad_type_0"), val = tensor("custom")]; + tensor lora_out_989_pad_0 = const()[name = tensor("lora_out_989_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_991_weight_0_to_fp16 = const()[name = tensor("lora_out_991_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484046848)))]; + tensor lora_out_991_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_9976, groups = var_9675, pad = lora_out_989_pad_0, pad_type = lora_out_989_pad_type_0, strides = var_9974, weight = lora_out_991_weight_0_to_fp16, x = input_739_cast_fp16)[name = tensor("lora_out_991_cast_fp16")]; + tensor obj_347_cast_fp16 = add(x = pretrained_out_495_cast_fp16, y = lora_out_991_cast_fp16)[name = tensor("obj_347_cast_fp16")]; + tensor inputs_149_cast_fp16 = add(x = inputs_147_cast_fp16, y = obj_347_cast_fp16)[name = tensor("inputs_149_cast_fp16")]; + tensor var_9985 = const()[name = tensor("op_9985"), val = tensor([1])]; + tensor channels_mean_149_cast_fp16 = reduce_mean(axes = var_9985, keep_dims = var_9676, x = inputs_149_cast_fp16)[name = tensor("channels_mean_149_cast_fp16")]; + tensor zero_mean_149_cast_fp16 = sub(x = inputs_149_cast_fp16, y = channels_mean_149_cast_fp16)[name = tensor("zero_mean_149_cast_fp16")]; + tensor zero_mean_sq_149_cast_fp16 = mul(x = zero_mean_149_cast_fp16, y = zero_mean_149_cast_fp16)[name = tensor("zero_mean_sq_149_cast_fp16")]; + tensor var_9989 = const()[name = tensor("op_9989"), val = tensor([1])]; + tensor var_9990_cast_fp16 = reduce_mean(axes = var_9989, keep_dims = var_9676, x = zero_mean_sq_149_cast_fp16)[name = tensor("op_9990_cast_fp16")]; + tensor var_9991_to_fp16 = const()[name = tensor("op_9991_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_9992_cast_fp16 = add(x = var_9990_cast_fp16, y = var_9991_to_fp16)[name = tensor("op_9992_cast_fp16")]; + tensor denom_149_epsilon_0 = const()[name = tensor("denom_149_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_149_cast_fp16 = rsqrt(epsilon = denom_149_epsilon_0, x = var_9992_cast_fp16)[name = tensor("denom_149_cast_fp16")]; + tensor out_149_cast_fp16 = mul(x = zero_mean_149_cast_fp16, y = denom_149_cast_fp16)[name = tensor("out_149_cast_fp16")]; + tensor input_741_gamma_0_to_fp16 = const()[name = tensor("input_741_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484087872)))]; + tensor input_741_beta_0_to_fp16 = const()[name = tensor("input_741_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484090496)))]; + tensor input_741_epsilon_0_to_fp16 = const()[name = tensor("input_741_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_741_cast_fp16 = batch_norm(beta = input_741_beta_0_to_fp16, epsilon = input_741_epsilon_0_to_fp16, gamma = input_741_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_149_cast_fp16)[name = tensor("input_741_cast_fp16")]; + tensor var_10006 = const()[name = tensor("op_10006"), val = tensor([1, 1])]; + tensor var_10008 = const()[name = tensor("op_10008"), val = tensor([1, 1])]; + tensor pretrained_out_497_pad_type_0 = const()[name = tensor("pretrained_out_497_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_497_pad_0 = const()[name = tensor("pretrained_out_497_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484093120))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487369984))), name = tensor("layers_24_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_24_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487370112)))]; + tensor pretrained_out_497_cast_fp16 = conv(bias = layers_24_fc1_pretrained_bias_to_fp16, dilations = var_10008, groups = var_9675, pad = pretrained_out_497_pad_0, pad_type = pretrained_out_497_pad_type_0, strides = var_10006, weight = layers_24_fc1_pretrained_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = tensor("pretrained_out_497_cast_fp16")]; + tensor var_10012 = const()[name = tensor("op_10012"), val = tensor([1, 1])]; + tensor var_10014 = const()[name = tensor("op_10014"), val = tensor([1, 1])]; + tensor input_743_pad_type_0 = const()[name = tensor("input_743_pad_type_0"), val = tensor("custom")]; + tensor input_743_pad_0 = const()[name = tensor("input_743_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_24_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487380416)))]; + tensor input_743_cast_fp16 = conv(dilations = var_10014, groups = var_9675, pad = input_743_pad_0, pad_type = input_743_pad_type_0, strides = var_10012, weight = layers_24_fc1_loraA_weight_to_fp16, x = input_741_cast_fp16)[name = tensor("input_743_cast_fp16")]; + tensor var_10018 = const()[name = tensor("op_10018"), val = tensor([1, 1])]; + tensor var_10020 = const()[name = tensor("op_10020"), val = tensor([1, 1])]; + tensor lora_out_993_pad_type_0 = const()[name = tensor("lora_out_993_pad_type_0"), val = tensor("custom")]; + tensor lora_out_993_pad_0 = const()[name = tensor("lora_out_993_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_995_weight_0_to_fp16 = const()[name = tensor("lora_out_995_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487421440)))]; + tensor lora_out_995_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_10020, groups = var_9675, pad = lora_out_993_pad_0, pad_type = lora_out_993_pad_type_0, strides = var_10018, weight = lora_out_995_weight_0_to_fp16, x = input_743_cast_fp16)[name = tensor("lora_out_995_cast_fp16")]; + tensor input_745_cast_fp16 = add(x = pretrained_out_497_cast_fp16, y = lora_out_995_cast_fp16)[name = tensor("input_745_cast_fp16")]; + tensor input_747_mode_0 = const()[name = tensor("input_747_mode_0"), val = tensor("EXACT")]; + tensor input_747_cast_fp16 = gelu(mode = input_747_mode_0, x = input_745_cast_fp16)[name = tensor("input_747_cast_fp16")]; + tensor var_10032 = const()[name = tensor("op_10032"), val = tensor([1, 1])]; + tensor var_10034 = const()[name = tensor("op_10034"), val = tensor([1, 1])]; + tensor pretrained_out_499_pad_type_0 = const()[name = tensor("pretrained_out_499_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_499_pad_0 = const()[name = tensor("pretrained_out_499_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487585344))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492500608))), name = tensor("layers_24_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_24_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_24_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492500800)))]; + tensor pretrained_out_499_cast_fp16 = conv(bias = layers_24_fc2_pretrained_bias_to_fp16, dilations = var_10034, groups = var_9675, pad = pretrained_out_499_pad_0, pad_type = pretrained_out_499_pad_type_0, strides = var_10032, weight = layers_24_fc2_pretrained_weight_to_fp16_palettized, x = input_747_cast_fp16)[name = tensor("pretrained_out_499_cast_fp16")]; + tensor var_10038 = const()[name = tensor("op_10038"), val = tensor([1, 1])]; + tensor var_10040 = const()[name = tensor("op_10040"), val = tensor([1, 1])]; + tensor input_749_pad_type_0 = const()[name = tensor("input_749_pad_type_0"), val = tensor("custom")]; + tensor input_749_pad_0 = const()[name = tensor("input_749_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_24_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_24_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492503424)))]; + tensor input_749_cast_fp16 = conv(dilations = var_10040, groups = var_9675, pad = input_749_pad_0, pad_type = input_749_pad_type_0, strides = var_10038, weight = layers_24_fc2_loraA_weight_to_fp16, x = input_747_cast_fp16)[name = tensor("input_749_cast_fp16")]; + tensor var_10044 = const()[name = tensor("op_10044"), val = tensor([1, 1])]; + tensor var_10046 = const()[name = tensor("op_10046"), val = tensor([1, 1])]; + tensor lora_out_997_pad_type_0 = const()[name = tensor("lora_out_997_pad_type_0"), val = tensor("custom")]; + tensor lora_out_997_pad_0 = const()[name = tensor("lora_out_997_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_999_weight_0_to_fp16 = const()[name = tensor("lora_out_999_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492667328)))]; + tensor lora_out_999_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10046, groups = var_9675, pad = lora_out_997_pad_0, pad_type = lora_out_997_pad_type_0, strides = var_10044, weight = lora_out_999_weight_0_to_fp16, x = input_749_cast_fp16)[name = tensor("lora_out_999_cast_fp16")]; + tensor hidden_states_51_cast_fp16 = add(x = pretrained_out_499_cast_fp16, y = lora_out_999_cast_fp16)[name = tensor("hidden_states_51_cast_fp16")]; + tensor inputs_151_cast_fp16 = add(x = inputs_149_cast_fp16, y = hidden_states_51_cast_fp16)[name = tensor("inputs_151_cast_fp16")]; + tensor var_10062 = const()[name = tensor("op_10062"), val = tensor(3)]; + tensor var_10069 = const()[name = tensor("op_10069"), val = tensor(1)]; + tensor var_10070 = const()[name = tensor("op_10070"), val = tensor(true)]; + tensor var_10082 = const()[name = tensor("op_10082"), val = tensor([1])]; + tensor channels_mean_151_cast_fp16 = reduce_mean(axes = var_10082, keep_dims = var_10070, x = inputs_151_cast_fp16)[name = tensor("channels_mean_151_cast_fp16")]; + tensor zero_mean_151_cast_fp16 = sub(x = inputs_151_cast_fp16, y = channels_mean_151_cast_fp16)[name = tensor("zero_mean_151_cast_fp16")]; + tensor zero_mean_sq_151_cast_fp16 = mul(x = zero_mean_151_cast_fp16, y = zero_mean_151_cast_fp16)[name = tensor("zero_mean_sq_151_cast_fp16")]; + tensor var_10086 = const()[name = tensor("op_10086"), val = tensor([1])]; + tensor var_10087_cast_fp16 = reduce_mean(axes = var_10086, keep_dims = var_10070, x = zero_mean_sq_151_cast_fp16)[name = tensor("op_10087_cast_fp16")]; + tensor var_10088_to_fp16 = const()[name = tensor("op_10088_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_10089_cast_fp16 = add(x = var_10087_cast_fp16, y = var_10088_to_fp16)[name = tensor("op_10089_cast_fp16")]; + tensor denom_151_epsilon_0 = const()[name = tensor("denom_151_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_151_cast_fp16 = rsqrt(epsilon = denom_151_epsilon_0, x = var_10089_cast_fp16)[name = tensor("denom_151_cast_fp16")]; + tensor out_151_cast_fp16 = mul(x = zero_mean_151_cast_fp16, y = denom_151_cast_fp16)[name = tensor("out_151_cast_fp16")]; + tensor obj_351_gamma_0_to_fp16 = const()[name = tensor("obj_351_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492708352)))]; + tensor obj_351_beta_0_to_fp16 = const()[name = tensor("obj_351_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492710976)))]; + tensor obj_351_epsilon_0_to_fp16 = const()[name = tensor("obj_351_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_351_cast_fp16 = batch_norm(beta = obj_351_beta_0_to_fp16, epsilon = obj_351_epsilon_0_to_fp16, gamma = obj_351_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_151_cast_fp16)[name = tensor("obj_351_cast_fp16")]; + tensor var_10107 = const()[name = tensor("op_10107"), val = tensor([1, 1])]; + tensor var_10109 = const()[name = tensor("op_10109"), val = tensor([1, 1])]; + tensor pretrained_out_501_pad_type_0 = const()[name = tensor("pretrained_out_501_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_501_pad_0 = const()[name = tensor("pretrained_out_501_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492713600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493532864))), name = tensor("layers_25_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_25_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493532992)))]; + tensor pretrained_out_501_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_10109, groups = var_10069, pad = pretrained_out_501_pad_0, pad_type = pretrained_out_501_pad_type_0, strides = var_10107, weight = layers_25_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_351_cast_fp16)[name = tensor("pretrained_out_501_cast_fp16")]; + tensor var_10113 = const()[name = tensor("op_10113"), val = tensor([1, 1])]; + tensor var_10115 = const()[name = tensor("op_10115"), val = tensor([1, 1])]; + tensor input_751_pad_type_0 = const()[name = tensor("input_751_pad_type_0"), val = tensor("custom")]; + tensor input_751_pad_0 = const()[name = tensor("input_751_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493535616)))]; + tensor input_751_cast_fp16 = conv(dilations = var_10115, groups = var_10069, pad = input_751_pad_0, pad_type = input_751_pad_type_0, strides = var_10113, weight = layers_25_self_attn_q_proj_loraA_weight_to_fp16, x = obj_351_cast_fp16)[name = tensor("input_751_cast_fp16")]; + tensor var_10119 = const()[name = tensor("op_10119"), val = tensor([1, 1])]; + tensor var_10121 = const()[name = tensor("op_10121"), val = tensor([1, 1])]; + tensor lora_out_1001_pad_type_0 = const()[name = tensor("lora_out_1001_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1001_pad_0 = const()[name = tensor("lora_out_1001_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1003_weight_0_to_fp16 = const()[name = tensor("lora_out_1003_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493576640)))]; + tensor lora_out_1003_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10121, groups = var_10069, pad = lora_out_1001_pad_0, pad_type = lora_out_1001_pad_type_0, strides = var_10119, weight = lora_out_1003_weight_0_to_fp16, x = input_751_cast_fp16)[name = tensor("lora_out_1003_cast_fp16")]; + tensor query_101_cast_fp16 = add(x = pretrained_out_501_cast_fp16, y = lora_out_1003_cast_fp16)[name = tensor("query_101_cast_fp16")]; + tensor var_10131 = const()[name = tensor("op_10131"), val = tensor([1, 1])]; + tensor var_10133 = const()[name = tensor("op_10133"), val = tensor([1, 1])]; + tensor pretrained_out_503_pad_type_0 = const()[name = tensor("pretrained_out_503_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_503_pad_0 = const()[name = tensor("pretrained_out_503_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493617664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494436928))), name = tensor("layers_25_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_503_cast_fp16 = conv(dilations = var_10133, groups = var_10069, pad = pretrained_out_503_pad_0, pad_type = pretrained_out_503_pad_type_0, strides = var_10131, weight = layers_25_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_351_cast_fp16)[name = tensor("pretrained_out_503_cast_fp16")]; + tensor var_10137 = const()[name = tensor("op_10137"), val = tensor([1, 1])]; + tensor var_10139 = const()[name = tensor("op_10139"), val = tensor([1, 1])]; + tensor input_753_pad_type_0 = const()[name = tensor("input_753_pad_type_0"), val = tensor("custom")]; + tensor input_753_pad_0 = const()[name = tensor("input_753_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494437056)))]; + tensor input_753_cast_fp16 = conv(dilations = var_10139, groups = var_10069, pad = input_753_pad_0, pad_type = input_753_pad_type_0, strides = var_10137, weight = layers_25_self_attn_k_proj_loraA_weight_to_fp16, x = obj_351_cast_fp16)[name = tensor("input_753_cast_fp16")]; + tensor var_10143 = const()[name = tensor("op_10143"), val = tensor([1, 1])]; + tensor var_10145 = const()[name = tensor("op_10145"), val = tensor([1, 1])]; + tensor lora_out_1005_pad_type_0 = const()[name = tensor("lora_out_1005_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1005_pad_0 = const()[name = tensor("lora_out_1005_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1007_weight_0_to_fp16 = const()[name = tensor("lora_out_1007_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494478080)))]; + tensor lora_out_1007_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10145, groups = var_10069, pad = lora_out_1005_pad_0, pad_type = lora_out_1005_pad_type_0, strides = var_10143, weight = lora_out_1007_weight_0_to_fp16, x = input_753_cast_fp16)[name = tensor("lora_out_1007_cast_fp16")]; + tensor current_key_51_cast_fp16 = add(x = pretrained_out_503_cast_fp16, y = lora_out_1007_cast_fp16)[name = tensor("current_key_51_cast_fp16")]; + tensor var_10156 = const()[name = tensor("op_10156"), val = tensor([1, 1])]; + tensor var_10158 = const()[name = tensor("op_10158"), val = tensor([1, 1])]; + tensor pretrained_out_505_pad_type_0 = const()[name = tensor("pretrained_out_505_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_505_pad_0 = const()[name = tensor("pretrained_out_505_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494519104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495338368))), name = tensor("layers_25_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_25_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495338496)))]; + tensor pretrained_out_505_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_10158, groups = var_10069, pad = pretrained_out_505_pad_0, pad_type = pretrained_out_505_pad_type_0, strides = var_10156, weight = layers_25_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_351_cast_fp16)[name = tensor("pretrained_out_505_cast_fp16")]; + tensor var_10162 = const()[name = tensor("op_10162"), val = tensor([1, 1])]; + tensor var_10164 = const()[name = tensor("op_10164"), val = tensor([1, 1])]; + tensor input_755_pad_type_0 = const()[name = tensor("input_755_pad_type_0"), val = tensor("custom")]; + tensor input_755_pad_0 = const()[name = tensor("input_755_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495341120)))]; + tensor input_755_cast_fp16 = conv(dilations = var_10164, groups = var_10069, pad = input_755_pad_0, pad_type = input_755_pad_type_0, strides = var_10162, weight = layers_25_self_attn_v_proj_loraA_weight_to_fp16, x = obj_351_cast_fp16)[name = tensor("input_755_cast_fp16")]; + tensor var_10168 = const()[name = tensor("op_10168"), val = tensor([1, 1])]; + tensor var_10170 = const()[name = tensor("op_10170"), val = tensor([1, 1])]; + tensor lora_out_1009_pad_type_0 = const()[name = tensor("lora_out_1009_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1009_pad_0 = const()[name = tensor("lora_out_1009_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1011_weight_0_to_fp16 = const()[name = tensor("lora_out_1011_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495382144)))]; + tensor lora_out_1011_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10170, groups = var_10069, pad = lora_out_1009_pad_0, pad_type = lora_out_1009_pad_type_0, strides = var_10168, weight = lora_out_1011_weight_0_to_fp16, x = input_755_cast_fp16)[name = tensor("lora_out_1011_cast_fp16")]; + tensor current_value_51_cast_fp16 = add(x = pretrained_out_505_cast_fp16, y = lora_out_1011_cast_fp16)[name = tensor("current_value_51_cast_fp16")]; + tensor var_10180_cast_fp16 = mul(x = current_key_51_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_10180_cast_fp16")]; + tensor var_10182_cast_fp16 = mul(x = var_103_cast_fp16_25, y = var_295_cast_fp16)[name = tensor("op_10182_cast_fp16")]; + tensor key_101_cast_fp16 = add(x = var_10180_cast_fp16, y = var_10182_cast_fp16)[name = tensor("key_101_cast_fp16")]; + tensor var_10184_cast_fp16 = mul(x = current_value_51_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_10184_cast_fp16")]; + tensor var_10186_cast_fp16 = mul(x = var_138_cast_fp16_25, y = var_295_cast_fp16)[name = tensor("op_10186_cast_fp16")]; + tensor value_101_cast_fp16 = add(x = var_10184_cast_fp16, y = var_10186_cast_fp16)[name = tensor("value_101_cast_fp16")]; + tensor var_10189 = const()[name = tensor("op_10189"), val = tensor([1, 20, 64, -1])]; + tensor var_10190_cast_fp16 = reshape(shape = var_10189, x = query_101_cast_fp16)[name = tensor("op_10190_cast_fp16")]; + tensor var_10191_to_fp16 = const()[name = tensor("op_10191_to_fp16"), val = tensor(0x1p-3)]; + tensor var_10192_cast_fp16 = mul(x = var_10190_cast_fp16, y = var_10191_to_fp16)[name = tensor("op_10192_cast_fp16")]; + tensor var_10193 = const()[name = tensor("op_10193"), val = tensor([1, 20, 64, -1])]; + tensor var_10194_cast_fp16 = reshape(shape = var_10193, x = key_101_cast_fp16)[name = tensor("op_10194_cast_fp16")]; + tensor mh_w_151_transpose_x_0 = const()[name = tensor("mh_w_151_transpose_x_0"), val = tensor(true)]; + tensor mh_w_151_transpose_y_0 = const()[name = tensor("mh_w_151_transpose_y_0"), val = tensor(false)]; + tensor mh_w_151_cast_fp16 = matmul(transpose_x = mh_w_151_transpose_x_0, transpose_y = mh_w_151_transpose_y_0, x = var_10192_cast_fp16, y = var_10194_cast_fp16)[name = tensor("mh_w_151_cast_fp16")]; + tensor mh_w_153_cast_fp16 = add(x = mh_w_151_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_153_cast_fp16")]; + tensor var_10202_cast_fp16 = softmax(axis = var_10062, x = mh_w_153_cast_fp16)[name = tensor("op_10202_cast_fp16")]; + tensor var_10203 = const()[name = tensor("op_10203"), val = tensor([1, 20, 64, -1])]; + tensor var_10204_cast_fp16 = reshape(shape = var_10203, x = value_101_cast_fp16)[name = tensor("op_10204_cast_fp16")]; + tensor attn_101_transpose_x_0 = const()[name = tensor("attn_101_transpose_x_0"), val = tensor(false)]; + tensor attn_101_transpose_y_0 = const()[name = tensor("attn_101_transpose_y_0"), val = tensor(true)]; + tensor attn_101_cast_fp16 = matmul(transpose_x = attn_101_transpose_x_0, transpose_y = attn_101_transpose_y_0, x = var_10204_cast_fp16, y = var_10202_cast_fp16)[name = tensor("attn_101_cast_fp16")]; + tensor var_10207 = const()[name = tensor("op_10207"), val = tensor([1, 1280, 1, -1])]; + tensor input_757_cast_fp16 = reshape(shape = var_10207, x = attn_101_cast_fp16)[name = tensor("input_757_cast_fp16")]; + tensor var_10214 = const()[name = tensor("op_10214"), val = tensor([1, 1])]; + tensor var_10216 = const()[name = tensor("op_10216"), val = tensor([1, 1])]; + tensor pretrained_out_507_pad_type_0 = const()[name = tensor("pretrained_out_507_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_507_pad_0 = const()[name = tensor("pretrained_out_507_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495423168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496242432))), name = tensor("layers_25_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_25_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496242560)))]; + tensor pretrained_out_507_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_10216, groups = var_10069, pad = pretrained_out_507_pad_0, pad_type = pretrained_out_507_pad_type_0, strides = var_10214, weight = layers_25_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_757_cast_fp16)[name = tensor("pretrained_out_507_cast_fp16")]; + tensor var_10220 = const()[name = tensor("op_10220"), val = tensor([1, 1])]; + tensor var_10222 = const()[name = tensor("op_10222"), val = tensor([1, 1])]; + tensor input_759_pad_type_0 = const()[name = tensor("input_759_pad_type_0"), val = tensor("custom")]; + tensor input_759_pad_0 = const()[name = tensor("input_759_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496245184)))]; + tensor input_759_cast_fp16 = conv(dilations = var_10222, groups = var_10069, pad = input_759_pad_0, pad_type = input_759_pad_type_0, strides = var_10220, weight = layers_25_self_attn_o_proj_loraA_weight_to_fp16, x = input_757_cast_fp16)[name = tensor("input_759_cast_fp16")]; + tensor var_10226 = const()[name = tensor("op_10226"), val = tensor([1, 1])]; + tensor var_10228 = const()[name = tensor("op_10228"), val = tensor([1, 1])]; + tensor lora_out_1013_pad_type_0 = const()[name = tensor("lora_out_1013_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1013_pad_0 = const()[name = tensor("lora_out_1013_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1015_weight_0_to_fp16 = const()[name = tensor("lora_out_1015_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496286208)))]; + tensor lora_out_1015_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10228, groups = var_10069, pad = lora_out_1013_pad_0, pad_type = lora_out_1013_pad_type_0, strides = var_10226, weight = lora_out_1015_weight_0_to_fp16, x = input_759_cast_fp16)[name = tensor("lora_out_1015_cast_fp16")]; + tensor obj_357_cast_fp16 = add(x = pretrained_out_507_cast_fp16, y = lora_out_1015_cast_fp16)[name = tensor("obj_357_cast_fp16")]; + tensor inputs_153_cast_fp16 = add(x = inputs_151_cast_fp16, y = obj_357_cast_fp16)[name = tensor("inputs_153_cast_fp16")]; + tensor var_10241 = const()[name = tensor("op_10241"), val = tensor([1])]; + tensor channels_mean_153_cast_fp16 = reduce_mean(axes = var_10241, keep_dims = var_10070, x = inputs_153_cast_fp16)[name = tensor("channels_mean_153_cast_fp16")]; + tensor zero_mean_153_cast_fp16 = sub(x = inputs_153_cast_fp16, y = channels_mean_153_cast_fp16)[name = tensor("zero_mean_153_cast_fp16")]; + tensor zero_mean_sq_153_cast_fp16 = mul(x = zero_mean_153_cast_fp16, y = zero_mean_153_cast_fp16)[name = tensor("zero_mean_sq_153_cast_fp16")]; + tensor var_10245 = const()[name = tensor("op_10245"), val = tensor([1])]; + tensor var_10246_cast_fp16 = reduce_mean(axes = var_10245, keep_dims = var_10070, x = zero_mean_sq_153_cast_fp16)[name = tensor("op_10246_cast_fp16")]; + tensor var_10247_to_fp16 = const()[name = tensor("op_10247_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_10248_cast_fp16 = add(x = var_10246_cast_fp16, y = var_10247_to_fp16)[name = tensor("op_10248_cast_fp16")]; + tensor denom_153_epsilon_0 = const()[name = tensor("denom_153_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_153_cast_fp16 = rsqrt(epsilon = denom_153_epsilon_0, x = var_10248_cast_fp16)[name = tensor("denom_153_cast_fp16")]; + tensor out_153_cast_fp16 = mul(x = zero_mean_153_cast_fp16, y = denom_153_cast_fp16)[name = tensor("out_153_cast_fp16")]; + tensor obj_359_gamma_0_to_fp16 = const()[name = tensor("obj_359_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496327232)))]; + tensor obj_359_beta_0_to_fp16 = const()[name = tensor("obj_359_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496329856)))]; + tensor obj_359_epsilon_0_to_fp16 = const()[name = tensor("obj_359_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_359_cast_fp16 = batch_norm(beta = obj_359_beta_0_to_fp16, epsilon = obj_359_epsilon_0_to_fp16, gamma = obj_359_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_153_cast_fp16)[name = tensor("obj_359_cast_fp16")]; + tensor var_10266 = const()[name = tensor("op_10266"), val = tensor([1, 1])]; + tensor var_10268 = const()[name = tensor("op_10268"), val = tensor([1, 1])]; + tensor pretrained_out_509_pad_type_0 = const()[name = tensor("pretrained_out_509_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_509_pad_0 = const()[name = tensor("pretrained_out_509_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496332480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497151744))), name = tensor("layers_25_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_25_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497151872)))]; + tensor pretrained_out_509_cast_fp16 = conv(bias = layers_25_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_10268, groups = var_10069, pad = pretrained_out_509_pad_0, pad_type = pretrained_out_509_pad_type_0, strides = var_10266, weight = layers_25_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_359_cast_fp16)[name = tensor("pretrained_out_509_cast_fp16")]; + tensor var_10272 = const()[name = tensor("op_10272"), val = tensor([1, 1])]; + tensor var_10274 = const()[name = tensor("op_10274"), val = tensor([1, 1])]; + tensor input_761_pad_type_0 = const()[name = tensor("input_761_pad_type_0"), val = tensor("custom")]; + tensor input_761_pad_0 = const()[name = tensor("input_761_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497154496)))]; + tensor input_761_cast_fp16 = conv(dilations = var_10274, groups = var_10069, pad = input_761_pad_0, pad_type = input_761_pad_type_0, strides = var_10272, weight = layers_25_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_359_cast_fp16)[name = tensor("input_761_cast_fp16")]; + tensor var_10278 = const()[name = tensor("op_10278"), val = tensor([1, 1])]; + tensor var_10280 = const()[name = tensor("op_10280"), val = tensor([1, 1])]; + tensor lora_out_1017_pad_type_0 = const()[name = tensor("lora_out_1017_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1017_pad_0 = const()[name = tensor("lora_out_1017_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1019_weight_0_to_fp16 = const()[name = tensor("lora_out_1019_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497195520)))]; + tensor lora_out_1019_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10280, groups = var_10069, pad = lora_out_1017_pad_0, pad_type = lora_out_1017_pad_type_0, strides = var_10278, weight = lora_out_1019_weight_0_to_fp16, x = input_761_cast_fp16)[name = tensor("lora_out_1019_cast_fp16")]; + tensor query_103_cast_fp16 = add(x = pretrained_out_509_cast_fp16, y = lora_out_1019_cast_fp16)[name = tensor("query_103_cast_fp16")]; + tensor var_10290 = const()[name = tensor("op_10290"), val = tensor([1, 1])]; + tensor var_10292 = const()[name = tensor("op_10292"), val = tensor([1, 1])]; + tensor pretrained_out_511_pad_type_0 = const()[name = tensor("pretrained_out_511_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_511_pad_0 = const()[name = tensor("pretrained_out_511_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(497236544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498055808))), name = tensor("layers_25_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_511_cast_fp16 = conv(dilations = var_10292, groups = var_10069, pad = pretrained_out_511_pad_0, pad_type = pretrained_out_511_pad_type_0, strides = var_10290, weight = layers_25_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_511_cast_fp16")]; + tensor var_10296 = const()[name = tensor("op_10296"), val = tensor([1, 1])]; + tensor var_10298 = const()[name = tensor("op_10298"), val = tensor([1, 1])]; + tensor input_763_pad_type_0 = const()[name = tensor("input_763_pad_type_0"), val = tensor("custom")]; + tensor input_763_pad_0 = const()[name = tensor("input_763_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498055936)))]; + tensor input_763_cast_fp16 = conv(dilations = var_10298, groups = var_10069, pad = input_763_pad_0, pad_type = input_763_pad_type_0, strides = var_10296, weight = layers_25_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_763_cast_fp16")]; + tensor var_10302 = const()[name = tensor("op_10302"), val = tensor([1, 1])]; + tensor var_10304 = const()[name = tensor("op_10304"), val = tensor([1, 1])]; + tensor lora_out_1021_pad_type_0 = const()[name = tensor("lora_out_1021_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1021_pad_0 = const()[name = tensor("lora_out_1021_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1023_weight_0_to_fp16 = const()[name = tensor("lora_out_1023_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498096960)))]; + tensor lora_out_1023_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10304, groups = var_10069, pad = lora_out_1021_pad_0, pad_type = lora_out_1021_pad_type_0, strides = var_10302, weight = lora_out_1023_weight_0_to_fp16, x = input_763_cast_fp16)[name = tensor("lora_out_1023_cast_fp16")]; + tensor key_103_cast_fp16 = add(x = pretrained_out_511_cast_fp16, y = lora_out_1023_cast_fp16)[name = tensor("key_103_cast_fp16")]; + tensor var_10315 = const()[name = tensor("op_10315"), val = tensor([1, 1])]; + tensor var_10317 = const()[name = tensor("op_10317"), val = tensor([1, 1])]; + tensor pretrained_out_513_pad_type_0 = const()[name = tensor("pretrained_out_513_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_513_pad_0 = const()[name = tensor("pretrained_out_513_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498137984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498957248))), name = tensor("layers_25_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_25_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498957376)))]; + tensor pretrained_out_513_cast_fp16 = conv(bias = layers_25_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_10317, groups = var_10069, pad = pretrained_out_513_pad_0, pad_type = pretrained_out_513_pad_type_0, strides = var_10315, weight = layers_25_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_513_cast_fp16")]; + tensor var_10321 = const()[name = tensor("op_10321"), val = tensor([1, 1])]; + tensor var_10323 = const()[name = tensor("op_10323"), val = tensor([1, 1])]; + tensor input_765_pad_type_0 = const()[name = tensor("input_765_pad_type_0"), val = tensor("custom")]; + tensor input_765_pad_0 = const()[name = tensor("input_765_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(498960000)))]; + tensor input_765_cast_fp16 = conv(dilations = var_10323, groups = var_10069, pad = input_765_pad_0, pad_type = input_765_pad_type_0, strides = var_10321, weight = layers_25_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_765_cast_fp16")]; + tensor var_10327 = const()[name = tensor("op_10327"), val = tensor([1, 1])]; + tensor var_10329 = const()[name = tensor("op_10329"), val = tensor([1, 1])]; + tensor lora_out_1025_pad_type_0 = const()[name = tensor("lora_out_1025_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1025_pad_0 = const()[name = tensor("lora_out_1025_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1027_weight_0_to_fp16 = const()[name = tensor("lora_out_1027_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499001024)))]; + tensor lora_out_1027_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10329, groups = var_10069, pad = lora_out_1025_pad_0, pad_type = lora_out_1025_pad_type_0, strides = var_10327, weight = lora_out_1027_weight_0_to_fp16, x = input_765_cast_fp16)[name = tensor("lora_out_1027_cast_fp16")]; + tensor value_103_cast_fp16 = add(x = pretrained_out_513_cast_fp16, y = lora_out_1027_cast_fp16)[name = tensor("value_103_cast_fp16")]; + tensor var_10336 = const()[name = tensor("op_10336"), val = tensor([1, 20, 64, -1])]; + tensor var_10337_cast_fp16 = reshape(shape = var_10336, x = query_103_cast_fp16)[name = tensor("op_10337_cast_fp16")]; + tensor var_10338_to_fp16 = const()[name = tensor("op_10338_to_fp16"), val = tensor(0x1p-3)]; + tensor var_10339_cast_fp16 = mul(x = var_10337_cast_fp16, y = var_10338_to_fp16)[name = tensor("op_10339_cast_fp16")]; + tensor var_10340 = const()[name = tensor("op_10340"), val = tensor([1, 20, 64, -1])]; + tensor var_10341_cast_fp16 = reshape(shape = var_10340, x = key_103_cast_fp16)[name = tensor("op_10341_cast_fp16")]; + tensor mh_w_155_transpose_x_0 = const()[name = tensor("mh_w_155_transpose_x_0"), val = tensor(true)]; + tensor mh_w_155_transpose_y_0 = const()[name = tensor("mh_w_155_transpose_y_0"), val = tensor(false)]; + tensor mh_w_155_cast_fp16 = matmul(transpose_x = mh_w_155_transpose_x_0, transpose_y = mh_w_155_transpose_y_0, x = var_10339_cast_fp16, y = var_10341_cast_fp16)[name = tensor("mh_w_155_cast_fp16")]; + tensor obj_363_cast_fp16 = softmax(axis = var_10062, x = mh_w_155_cast_fp16)[name = tensor("obj_363_cast_fp16")]; + tensor var_10345 = const()[name = tensor("op_10345"), val = tensor([1, 20, 64, -1])]; + tensor var_10346_cast_fp16 = reshape(shape = var_10345, x = value_103_cast_fp16)[name = tensor("op_10346_cast_fp16")]; + tensor attn_103_transpose_x_0 = const()[name = tensor("attn_103_transpose_x_0"), val = tensor(false)]; + tensor attn_103_transpose_y_0 = const()[name = tensor("attn_103_transpose_y_0"), val = tensor(true)]; + tensor attn_103_cast_fp16 = matmul(transpose_x = attn_103_transpose_x_0, transpose_y = attn_103_transpose_y_0, x = var_10346_cast_fp16, y = obj_363_cast_fp16)[name = tensor("attn_103_cast_fp16")]; + tensor var_10349 = const()[name = tensor("op_10349"), val = tensor([1, 1280, 1, -1])]; + tensor input_767_cast_fp16 = reshape(shape = var_10349, x = attn_103_cast_fp16)[name = tensor("input_767_cast_fp16")]; + tensor var_10356 = const()[name = tensor("op_10356"), val = tensor([1, 1])]; + tensor var_10358 = const()[name = tensor("op_10358"), val = tensor([1, 1])]; + tensor pretrained_out_515_pad_type_0 = const()[name = tensor("pretrained_out_515_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_515_pad_0 = const()[name = tensor("pretrained_out_515_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499042048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499861312))), name = tensor("layers_25_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_25_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499861440)))]; + tensor pretrained_out_515_cast_fp16 = conv(bias = layers_25_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_10358, groups = var_10069, pad = pretrained_out_515_pad_0, pad_type = pretrained_out_515_pad_type_0, strides = var_10356, weight = layers_25_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = tensor("pretrained_out_515_cast_fp16")]; + tensor var_10362 = const()[name = tensor("op_10362"), val = tensor([1, 1])]; + tensor var_10364 = const()[name = tensor("op_10364"), val = tensor([1, 1])]; + tensor input_769_pad_type_0 = const()[name = tensor("input_769_pad_type_0"), val = tensor("custom")]; + tensor input_769_pad_0 = const()[name = tensor("input_769_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_25_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499864064)))]; + tensor input_769_cast_fp16 = conv(dilations = var_10364, groups = var_10069, pad = input_769_pad_0, pad_type = input_769_pad_type_0, strides = var_10362, weight = layers_25_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_767_cast_fp16)[name = tensor("input_769_cast_fp16")]; + tensor var_10368 = const()[name = tensor("op_10368"), val = tensor([1, 1])]; + tensor var_10370 = const()[name = tensor("op_10370"), val = tensor([1, 1])]; + tensor lora_out_1029_pad_type_0 = const()[name = tensor("lora_out_1029_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1029_pad_0 = const()[name = tensor("lora_out_1029_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1031_weight_0_to_fp16 = const()[name = tensor("lora_out_1031_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499905088)))]; + tensor lora_out_1031_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10370, groups = var_10069, pad = lora_out_1029_pad_0, pad_type = lora_out_1029_pad_type_0, strides = var_10368, weight = lora_out_1031_weight_0_to_fp16, x = input_769_cast_fp16)[name = tensor("lora_out_1031_cast_fp16")]; + tensor obj_361_cast_fp16 = add(x = pretrained_out_515_cast_fp16, y = lora_out_1031_cast_fp16)[name = tensor("obj_361_cast_fp16")]; + tensor inputs_155_cast_fp16 = add(x = inputs_153_cast_fp16, y = obj_361_cast_fp16)[name = tensor("inputs_155_cast_fp16")]; + tensor var_10382 = const()[name = tensor("op_10382"), val = tensor([1])]; + tensor channels_mean_155_cast_fp16 = reduce_mean(axes = var_10382, keep_dims = var_10070, x = inputs_155_cast_fp16)[name = tensor("channels_mean_155_cast_fp16")]; + tensor zero_mean_155_cast_fp16 = sub(x = inputs_155_cast_fp16, y = channels_mean_155_cast_fp16)[name = tensor("zero_mean_155_cast_fp16")]; + tensor zero_mean_sq_155_cast_fp16 = mul(x = zero_mean_155_cast_fp16, y = zero_mean_155_cast_fp16)[name = tensor("zero_mean_sq_155_cast_fp16")]; + tensor var_10386 = const()[name = tensor("op_10386"), val = tensor([1])]; + tensor var_10387_cast_fp16 = reduce_mean(axes = var_10386, keep_dims = var_10070, x = zero_mean_sq_155_cast_fp16)[name = tensor("op_10387_cast_fp16")]; + tensor var_10388_to_fp16 = const()[name = tensor("op_10388_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_10389_cast_fp16 = add(x = var_10387_cast_fp16, y = var_10388_to_fp16)[name = tensor("op_10389_cast_fp16")]; + tensor denom_155_epsilon_0 = const()[name = tensor("denom_155_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_155_cast_fp16 = rsqrt(epsilon = denom_155_epsilon_0, x = var_10389_cast_fp16)[name = tensor("denom_155_cast_fp16")]; + tensor out_155_cast_fp16 = mul(x = zero_mean_155_cast_fp16, y = denom_155_cast_fp16)[name = tensor("out_155_cast_fp16")]; + tensor input_771_gamma_0_to_fp16 = const()[name = tensor("input_771_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499946112)))]; + tensor input_771_beta_0_to_fp16 = const()[name = tensor("input_771_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499948736)))]; + tensor input_771_epsilon_0_to_fp16 = const()[name = tensor("input_771_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_771_cast_fp16 = batch_norm(beta = input_771_beta_0_to_fp16, epsilon = input_771_epsilon_0_to_fp16, gamma = input_771_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_155_cast_fp16)[name = tensor("input_771_cast_fp16")]; + tensor var_10403 = const()[name = tensor("op_10403"), val = tensor([1, 1])]; + tensor var_10405 = const()[name = tensor("op_10405"), val = tensor([1, 1])]; + tensor pretrained_out_517_pad_type_0 = const()[name = tensor("pretrained_out_517_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_517_pad_0 = const()[name = tensor("pretrained_out_517_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(499951360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503228224))), name = tensor("layers_25_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_25_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503228352)))]; + tensor pretrained_out_517_cast_fp16 = conv(bias = layers_25_fc1_pretrained_bias_to_fp16, dilations = var_10405, groups = var_10069, pad = pretrained_out_517_pad_0, pad_type = pretrained_out_517_pad_type_0, strides = var_10403, weight = layers_25_fc1_pretrained_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = tensor("pretrained_out_517_cast_fp16")]; + tensor var_10409 = const()[name = tensor("op_10409"), val = tensor([1, 1])]; + tensor var_10411 = const()[name = tensor("op_10411"), val = tensor([1, 1])]; + tensor input_773_pad_type_0 = const()[name = tensor("input_773_pad_type_0"), val = tensor("custom")]; + tensor input_773_pad_0 = const()[name = tensor("input_773_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_25_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503238656)))]; + tensor input_773_cast_fp16 = conv(dilations = var_10411, groups = var_10069, pad = input_773_pad_0, pad_type = input_773_pad_type_0, strides = var_10409, weight = layers_25_fc1_loraA_weight_to_fp16, x = input_771_cast_fp16)[name = tensor("input_773_cast_fp16")]; + tensor var_10415 = const()[name = tensor("op_10415"), val = tensor([1, 1])]; + tensor var_10417 = const()[name = tensor("op_10417"), val = tensor([1, 1])]; + tensor lora_out_1033_pad_type_0 = const()[name = tensor("lora_out_1033_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1033_pad_0 = const()[name = tensor("lora_out_1033_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1035_weight_0_to_fp16 = const()[name = tensor("lora_out_1035_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503279680)))]; + tensor lora_out_1035_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_10417, groups = var_10069, pad = lora_out_1033_pad_0, pad_type = lora_out_1033_pad_type_0, strides = var_10415, weight = lora_out_1035_weight_0_to_fp16, x = input_773_cast_fp16)[name = tensor("lora_out_1035_cast_fp16")]; + tensor input_775_cast_fp16 = add(x = pretrained_out_517_cast_fp16, y = lora_out_1035_cast_fp16)[name = tensor("input_775_cast_fp16")]; + tensor input_777_mode_0 = const()[name = tensor("input_777_mode_0"), val = tensor("EXACT")]; + tensor input_777_cast_fp16 = gelu(mode = input_777_mode_0, x = input_775_cast_fp16)[name = tensor("input_777_cast_fp16")]; + tensor var_10429 = const()[name = tensor("op_10429"), val = tensor([1, 1])]; + tensor var_10431 = const()[name = tensor("op_10431"), val = tensor([1, 1])]; + tensor pretrained_out_519_pad_type_0 = const()[name = tensor("pretrained_out_519_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_519_pad_0 = const()[name = tensor("pretrained_out_519_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503443584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506720448))), name = tensor("layers_25_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_25_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_25_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506720576)))]; + tensor pretrained_out_519_cast_fp16 = conv(bias = layers_25_fc2_pretrained_bias_to_fp16, dilations = var_10431, groups = var_10069, pad = pretrained_out_519_pad_0, pad_type = pretrained_out_519_pad_type_0, strides = var_10429, weight = layers_25_fc2_pretrained_weight_to_fp16_palettized, x = input_777_cast_fp16)[name = tensor("pretrained_out_519_cast_fp16")]; + tensor var_10435 = const()[name = tensor("op_10435"), val = tensor([1, 1])]; + tensor var_10437 = const()[name = tensor("op_10437"), val = tensor([1, 1])]; + tensor input_779_pad_type_0 = const()[name = tensor("input_779_pad_type_0"), val = tensor("custom")]; + tensor input_779_pad_0 = const()[name = tensor("input_779_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_25_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_25_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506723200)))]; + tensor input_779_cast_fp16 = conv(dilations = var_10437, groups = var_10069, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = var_10435, weight = layers_25_fc2_loraA_weight_to_fp16, x = input_777_cast_fp16)[name = tensor("input_779_cast_fp16")]; + tensor var_10441 = const()[name = tensor("op_10441"), val = tensor([1, 1])]; + tensor var_10443 = const()[name = tensor("op_10443"), val = tensor([1, 1])]; + tensor lora_out_1037_pad_type_0 = const()[name = tensor("lora_out_1037_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1037_pad_0 = const()[name = tensor("lora_out_1037_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1039_weight_0_to_fp16 = const()[name = tensor("lora_out_1039_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506887104)))]; + tensor lora_out_1039_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10443, groups = var_10069, pad = lora_out_1037_pad_0, pad_type = lora_out_1037_pad_type_0, strides = var_10441, weight = lora_out_1039_weight_0_to_fp16, x = input_779_cast_fp16)[name = tensor("lora_out_1039_cast_fp16")]; + tensor hidden_states_53_cast_fp16 = add(x = pretrained_out_519_cast_fp16, y = lora_out_1039_cast_fp16)[name = tensor("hidden_states_53_cast_fp16")]; + tensor inputs_157_cast_fp16 = add(x = inputs_155_cast_fp16, y = hidden_states_53_cast_fp16)[name = tensor("inputs_157_cast_fp16")]; + tensor var_10460 = const()[name = tensor("op_10460"), val = tensor(3)]; + tensor var_10467 = const()[name = tensor("op_10467"), val = tensor(1)]; + tensor var_10468 = const()[name = tensor("op_10468"), val = tensor(true)]; + tensor var_10480 = const()[name = tensor("op_10480"), val = tensor([1])]; + tensor channels_mean_157_cast_fp16 = reduce_mean(axes = var_10480, keep_dims = var_10468, x = inputs_157_cast_fp16)[name = tensor("channels_mean_157_cast_fp16")]; + tensor zero_mean_157_cast_fp16 = sub(x = inputs_157_cast_fp16, y = channels_mean_157_cast_fp16)[name = tensor("zero_mean_157_cast_fp16")]; + tensor zero_mean_sq_157_cast_fp16 = mul(x = zero_mean_157_cast_fp16, y = zero_mean_157_cast_fp16)[name = tensor("zero_mean_sq_157_cast_fp16")]; + tensor var_10484 = const()[name = tensor("op_10484"), val = tensor([1])]; + tensor var_10485_cast_fp16 = reduce_mean(axes = var_10484, keep_dims = var_10468, x = zero_mean_sq_157_cast_fp16)[name = tensor("op_10485_cast_fp16")]; + tensor var_10486_to_fp16 = const()[name = tensor("op_10486_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_10487_cast_fp16 = add(x = var_10485_cast_fp16, y = var_10486_to_fp16)[name = tensor("op_10487_cast_fp16")]; + tensor denom_157_epsilon_0 = const()[name = tensor("denom_157_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_157_cast_fp16 = rsqrt(epsilon = denom_157_epsilon_0, x = var_10487_cast_fp16)[name = tensor("denom_157_cast_fp16")]; + tensor out_157_cast_fp16 = mul(x = zero_mean_157_cast_fp16, y = denom_157_cast_fp16)[name = tensor("out_157_cast_fp16")]; + tensor obj_365_gamma_0_to_fp16 = const()[name = tensor("obj_365_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506928128)))]; + tensor obj_365_beta_0_to_fp16 = const()[name = tensor("obj_365_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506930752)))]; + tensor obj_365_epsilon_0_to_fp16 = const()[name = tensor("obj_365_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_365_cast_fp16 = batch_norm(beta = obj_365_beta_0_to_fp16, epsilon = obj_365_epsilon_0_to_fp16, gamma = obj_365_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_157_cast_fp16)[name = tensor("obj_365_cast_fp16")]; + tensor var_10505 = const()[name = tensor("op_10505"), val = tensor([1, 1])]; + tensor var_10507 = const()[name = tensor("op_10507"), val = tensor([1, 1])]; + tensor pretrained_out_521_pad_type_0 = const()[name = tensor("pretrained_out_521_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_521_pad_0 = const()[name = tensor("pretrained_out_521_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(506933376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507752640))), name = tensor("layers_26_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_26_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507752768)))]; + tensor pretrained_out_521_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_10507, groups = var_10467, pad = pretrained_out_521_pad_0, pad_type = pretrained_out_521_pad_type_0, strides = var_10505, weight = layers_26_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_365_cast_fp16)[name = tensor("pretrained_out_521_cast_fp16")]; + tensor var_10511 = const()[name = tensor("op_10511"), val = tensor([1, 1])]; + tensor var_10513 = const()[name = tensor("op_10513"), val = tensor([1, 1])]; + tensor input_781_pad_type_0 = const()[name = tensor("input_781_pad_type_0"), val = tensor("custom")]; + tensor input_781_pad_0 = const()[name = tensor("input_781_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507755392)))]; + tensor input_781_cast_fp16 = conv(dilations = var_10513, groups = var_10467, pad = input_781_pad_0, pad_type = input_781_pad_type_0, strides = var_10511, weight = layers_26_self_attn_q_proj_loraA_weight_to_fp16, x = obj_365_cast_fp16)[name = tensor("input_781_cast_fp16")]; + tensor var_10517 = const()[name = tensor("op_10517"), val = tensor([1, 1])]; + tensor var_10519 = const()[name = tensor("op_10519"), val = tensor([1, 1])]; + tensor lora_out_1041_pad_type_0 = const()[name = tensor("lora_out_1041_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1041_pad_0 = const()[name = tensor("lora_out_1041_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1043_weight_0_to_fp16 = const()[name = tensor("lora_out_1043_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507796416)))]; + tensor lora_out_1043_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10519, groups = var_10467, pad = lora_out_1041_pad_0, pad_type = lora_out_1041_pad_type_0, strides = var_10517, weight = lora_out_1043_weight_0_to_fp16, x = input_781_cast_fp16)[name = tensor("lora_out_1043_cast_fp16")]; + tensor query_105_cast_fp16 = add(x = pretrained_out_521_cast_fp16, y = lora_out_1043_cast_fp16)[name = tensor("query_105_cast_fp16")]; + tensor var_10529 = const()[name = tensor("op_10529"), val = tensor([1, 1])]; + tensor var_10531 = const()[name = tensor("op_10531"), val = tensor([1, 1])]; + tensor pretrained_out_523_pad_type_0 = const()[name = tensor("pretrained_out_523_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_523_pad_0 = const()[name = tensor("pretrained_out_523_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(507837440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508656704))), name = tensor("layers_26_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_523_cast_fp16 = conv(dilations = var_10531, groups = var_10467, pad = pretrained_out_523_pad_0, pad_type = pretrained_out_523_pad_type_0, strides = var_10529, weight = layers_26_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_365_cast_fp16)[name = tensor("pretrained_out_523_cast_fp16")]; + tensor var_10535 = const()[name = tensor("op_10535"), val = tensor([1, 1])]; + tensor var_10537 = const()[name = tensor("op_10537"), val = tensor([1, 1])]; + tensor input_783_pad_type_0 = const()[name = tensor("input_783_pad_type_0"), val = tensor("custom")]; + tensor input_783_pad_0 = const()[name = tensor("input_783_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508656832)))]; + tensor input_783_cast_fp16 = conv(dilations = var_10537, groups = var_10467, pad = input_783_pad_0, pad_type = input_783_pad_type_0, strides = var_10535, weight = layers_26_self_attn_k_proj_loraA_weight_to_fp16, x = obj_365_cast_fp16)[name = tensor("input_783_cast_fp16")]; + tensor var_10541 = const()[name = tensor("op_10541"), val = tensor([1, 1])]; + tensor var_10543 = const()[name = tensor("op_10543"), val = tensor([1, 1])]; + tensor lora_out_1045_pad_type_0 = const()[name = tensor("lora_out_1045_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1045_pad_0 = const()[name = tensor("lora_out_1045_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1047_weight_0_to_fp16 = const()[name = tensor("lora_out_1047_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508697856)))]; + tensor lora_out_1047_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10543, groups = var_10467, pad = lora_out_1045_pad_0, pad_type = lora_out_1045_pad_type_0, strides = var_10541, weight = lora_out_1047_weight_0_to_fp16, x = input_783_cast_fp16)[name = tensor("lora_out_1047_cast_fp16")]; + tensor current_key_53_cast_fp16 = add(x = pretrained_out_523_cast_fp16, y = lora_out_1047_cast_fp16)[name = tensor("current_key_53_cast_fp16")]; + tensor var_10554 = const()[name = tensor("op_10554"), val = tensor([1, 1])]; + tensor var_10556 = const()[name = tensor("op_10556"), val = tensor([1, 1])]; + tensor pretrained_out_525_pad_type_0 = const()[name = tensor("pretrained_out_525_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_525_pad_0 = const()[name = tensor("pretrained_out_525_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(508738880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509558144))), name = tensor("layers_26_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_26_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509558272)))]; + tensor pretrained_out_525_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_10556, groups = var_10467, pad = pretrained_out_525_pad_0, pad_type = pretrained_out_525_pad_type_0, strides = var_10554, weight = layers_26_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_365_cast_fp16)[name = tensor("pretrained_out_525_cast_fp16")]; + tensor var_10560 = const()[name = tensor("op_10560"), val = tensor([1, 1])]; + tensor var_10562 = const()[name = tensor("op_10562"), val = tensor([1, 1])]; + tensor input_785_pad_type_0 = const()[name = tensor("input_785_pad_type_0"), val = tensor("custom")]; + tensor input_785_pad_0 = const()[name = tensor("input_785_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509560896)))]; + tensor input_785_cast_fp16 = conv(dilations = var_10562, groups = var_10467, pad = input_785_pad_0, pad_type = input_785_pad_type_0, strides = var_10560, weight = layers_26_self_attn_v_proj_loraA_weight_to_fp16, x = obj_365_cast_fp16)[name = tensor("input_785_cast_fp16")]; + tensor var_10566 = const()[name = tensor("op_10566"), val = tensor([1, 1])]; + tensor var_10568 = const()[name = tensor("op_10568"), val = tensor([1, 1])]; + tensor lora_out_1049_pad_type_0 = const()[name = tensor("lora_out_1049_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1049_pad_0 = const()[name = tensor("lora_out_1049_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1051_weight_0_to_fp16 = const()[name = tensor("lora_out_1051_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509601920)))]; + tensor lora_out_1051_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10568, groups = var_10467, pad = lora_out_1049_pad_0, pad_type = lora_out_1049_pad_type_0, strides = var_10566, weight = lora_out_1051_weight_0_to_fp16, x = input_785_cast_fp16)[name = tensor("lora_out_1051_cast_fp16")]; + tensor current_value_53_cast_fp16 = add(x = pretrained_out_525_cast_fp16, y = lora_out_1051_cast_fp16)[name = tensor("current_value_53_cast_fp16")]; + tensor var_10578_cast_fp16 = mul(x = current_key_53_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_10578_cast_fp16")]; + tensor var_10580_cast_fp16 = mul(x = var_103_cast_fp16_26, y = var_295_cast_fp16)[name = tensor("op_10580_cast_fp16")]; + tensor key_105_cast_fp16 = add(x = var_10578_cast_fp16, y = var_10580_cast_fp16)[name = tensor("key_105_cast_fp16")]; + tensor var_10582_cast_fp16 = mul(x = current_value_53_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_10582_cast_fp16")]; + tensor var_10584_cast_fp16 = mul(x = var_138_cast_fp16_26, y = var_295_cast_fp16)[name = tensor("op_10584_cast_fp16")]; + tensor value_105_cast_fp16 = add(x = var_10582_cast_fp16, y = var_10584_cast_fp16)[name = tensor("value_105_cast_fp16")]; + tensor var_10587 = const()[name = tensor("op_10587"), val = tensor([1, 20, 64, -1])]; + tensor var_10588_cast_fp16 = reshape(shape = var_10587, x = query_105_cast_fp16)[name = tensor("op_10588_cast_fp16")]; + tensor var_10589_to_fp16 = const()[name = tensor("op_10589_to_fp16"), val = tensor(0x1p-3)]; + tensor var_10590_cast_fp16 = mul(x = var_10588_cast_fp16, y = var_10589_to_fp16)[name = tensor("op_10590_cast_fp16")]; + tensor var_10591 = const()[name = tensor("op_10591"), val = tensor([1, 20, 64, -1])]; + tensor var_10592_cast_fp16 = reshape(shape = var_10591, x = key_105_cast_fp16)[name = tensor("op_10592_cast_fp16")]; + tensor mh_w_157_transpose_x_0 = const()[name = tensor("mh_w_157_transpose_x_0"), val = tensor(true)]; + tensor mh_w_157_transpose_y_0 = const()[name = tensor("mh_w_157_transpose_y_0"), val = tensor(false)]; + tensor mh_w_157_cast_fp16 = matmul(transpose_x = mh_w_157_transpose_x_0, transpose_y = mh_w_157_transpose_y_0, x = var_10590_cast_fp16, y = var_10592_cast_fp16)[name = tensor("mh_w_157_cast_fp16")]; + tensor mh_w_159_cast_fp16 = add(x = mh_w_157_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_159_cast_fp16")]; + tensor var_10600_cast_fp16 = softmax(axis = var_10460, x = mh_w_159_cast_fp16)[name = tensor("op_10600_cast_fp16")]; + tensor var_10601 = const()[name = tensor("op_10601"), val = tensor([1, 20, 64, -1])]; + tensor var_10602_cast_fp16 = reshape(shape = var_10601, x = value_105_cast_fp16)[name = tensor("op_10602_cast_fp16")]; + tensor attn_105_transpose_x_0 = const()[name = tensor("attn_105_transpose_x_0"), val = tensor(false)]; + tensor attn_105_transpose_y_0 = const()[name = tensor("attn_105_transpose_y_0"), val = tensor(true)]; + tensor attn_105_cast_fp16 = matmul(transpose_x = attn_105_transpose_x_0, transpose_y = attn_105_transpose_y_0, x = var_10602_cast_fp16, y = var_10600_cast_fp16)[name = tensor("attn_105_cast_fp16")]; + tensor var_10605 = const()[name = tensor("op_10605"), val = tensor([1, 1280, 1, -1])]; + tensor input_787_cast_fp16 = reshape(shape = var_10605, x = attn_105_cast_fp16)[name = tensor("input_787_cast_fp16")]; + tensor var_10612 = const()[name = tensor("op_10612"), val = tensor([1, 1])]; + tensor var_10614 = const()[name = tensor("op_10614"), val = tensor([1, 1])]; + tensor pretrained_out_527_pad_type_0 = const()[name = tensor("pretrained_out_527_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_527_pad_0 = const()[name = tensor("pretrained_out_527_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509642944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510462208))), name = tensor("layers_26_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_26_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510462336)))]; + tensor pretrained_out_527_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_10614, groups = var_10467, pad = pretrained_out_527_pad_0, pad_type = pretrained_out_527_pad_type_0, strides = var_10612, weight = layers_26_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_787_cast_fp16)[name = tensor("pretrained_out_527_cast_fp16")]; + tensor var_10618 = const()[name = tensor("op_10618"), val = tensor([1, 1])]; + tensor var_10620 = const()[name = tensor("op_10620"), val = tensor([1, 1])]; + tensor input_789_pad_type_0 = const()[name = tensor("input_789_pad_type_0"), val = tensor("custom")]; + tensor input_789_pad_0 = const()[name = tensor("input_789_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510464960)))]; + tensor input_789_cast_fp16 = conv(dilations = var_10620, groups = var_10467, pad = input_789_pad_0, pad_type = input_789_pad_type_0, strides = var_10618, weight = layers_26_self_attn_o_proj_loraA_weight_to_fp16, x = input_787_cast_fp16)[name = tensor("input_789_cast_fp16")]; + tensor var_10624 = const()[name = tensor("op_10624"), val = tensor([1, 1])]; + tensor var_10626 = const()[name = tensor("op_10626"), val = tensor([1, 1])]; + tensor lora_out_1053_pad_type_0 = const()[name = tensor("lora_out_1053_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1053_pad_0 = const()[name = tensor("lora_out_1053_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1055_weight_0_to_fp16 = const()[name = tensor("lora_out_1055_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510505984)))]; + tensor lora_out_1055_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10626, groups = var_10467, pad = lora_out_1053_pad_0, pad_type = lora_out_1053_pad_type_0, strides = var_10624, weight = lora_out_1055_weight_0_to_fp16, x = input_789_cast_fp16)[name = tensor("lora_out_1055_cast_fp16")]; + tensor obj_371_cast_fp16 = add(x = pretrained_out_527_cast_fp16, y = lora_out_1055_cast_fp16)[name = tensor("obj_371_cast_fp16")]; + tensor inputs_159_cast_fp16 = add(x = inputs_157_cast_fp16, y = obj_371_cast_fp16)[name = tensor("inputs_159_cast_fp16")]; + tensor var_10639 = const()[name = tensor("op_10639"), val = tensor([1])]; + tensor channels_mean_159_cast_fp16 = reduce_mean(axes = var_10639, keep_dims = var_10468, x = inputs_159_cast_fp16)[name = tensor("channels_mean_159_cast_fp16")]; + tensor zero_mean_159_cast_fp16 = sub(x = inputs_159_cast_fp16, y = channels_mean_159_cast_fp16)[name = tensor("zero_mean_159_cast_fp16")]; + tensor zero_mean_sq_159_cast_fp16 = mul(x = zero_mean_159_cast_fp16, y = zero_mean_159_cast_fp16)[name = tensor("zero_mean_sq_159_cast_fp16")]; + tensor var_10643 = const()[name = tensor("op_10643"), val = tensor([1])]; + tensor var_10644_cast_fp16 = reduce_mean(axes = var_10643, keep_dims = var_10468, x = zero_mean_sq_159_cast_fp16)[name = tensor("op_10644_cast_fp16")]; + tensor var_10645_to_fp16 = const()[name = tensor("op_10645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_10646_cast_fp16 = add(x = var_10644_cast_fp16, y = var_10645_to_fp16)[name = tensor("op_10646_cast_fp16")]; + tensor denom_159_epsilon_0 = const()[name = tensor("denom_159_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_159_cast_fp16 = rsqrt(epsilon = denom_159_epsilon_0, x = var_10646_cast_fp16)[name = tensor("denom_159_cast_fp16")]; + tensor out_159_cast_fp16 = mul(x = zero_mean_159_cast_fp16, y = denom_159_cast_fp16)[name = tensor("out_159_cast_fp16")]; + tensor obj_373_gamma_0_to_fp16 = const()[name = tensor("obj_373_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510547008)))]; + tensor obj_373_beta_0_to_fp16 = const()[name = tensor("obj_373_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510549632)))]; + tensor obj_373_epsilon_0_to_fp16 = const()[name = tensor("obj_373_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_373_cast_fp16 = batch_norm(beta = obj_373_beta_0_to_fp16, epsilon = obj_373_epsilon_0_to_fp16, gamma = obj_373_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_159_cast_fp16)[name = tensor("obj_373_cast_fp16")]; + tensor var_10664 = const()[name = tensor("op_10664"), val = tensor([1, 1])]; + tensor var_10666 = const()[name = tensor("op_10666"), val = tensor([1, 1])]; + tensor pretrained_out_529_pad_type_0 = const()[name = tensor("pretrained_out_529_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_529_pad_0 = const()[name = tensor("pretrained_out_529_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510552256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511371520))), name = tensor("layers_26_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_26_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511371648)))]; + tensor pretrained_out_529_cast_fp16 = conv(bias = layers_26_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_10666, groups = var_10467, pad = pretrained_out_529_pad_0, pad_type = pretrained_out_529_pad_type_0, strides = var_10664, weight = layers_26_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_373_cast_fp16)[name = tensor("pretrained_out_529_cast_fp16")]; + tensor var_10670 = const()[name = tensor("op_10670"), val = tensor([1, 1])]; + tensor var_10672 = const()[name = tensor("op_10672"), val = tensor([1, 1])]; + tensor input_791_pad_type_0 = const()[name = tensor("input_791_pad_type_0"), val = tensor("custom")]; + tensor input_791_pad_0 = const()[name = tensor("input_791_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511374272)))]; + tensor input_791_cast_fp16 = conv(dilations = var_10672, groups = var_10467, pad = input_791_pad_0, pad_type = input_791_pad_type_0, strides = var_10670, weight = layers_26_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_373_cast_fp16)[name = tensor("input_791_cast_fp16")]; + tensor var_10676 = const()[name = tensor("op_10676"), val = tensor([1, 1])]; + tensor var_10678 = const()[name = tensor("op_10678"), val = tensor([1, 1])]; + tensor lora_out_1057_pad_type_0 = const()[name = tensor("lora_out_1057_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1057_pad_0 = const()[name = tensor("lora_out_1057_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1059_weight_0_to_fp16 = const()[name = tensor("lora_out_1059_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511415296)))]; + tensor lora_out_1059_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10678, groups = var_10467, pad = lora_out_1057_pad_0, pad_type = lora_out_1057_pad_type_0, strides = var_10676, weight = lora_out_1059_weight_0_to_fp16, x = input_791_cast_fp16)[name = tensor("lora_out_1059_cast_fp16")]; + tensor query_107_cast_fp16 = add(x = pretrained_out_529_cast_fp16, y = lora_out_1059_cast_fp16)[name = tensor("query_107_cast_fp16")]; + tensor var_10688 = const()[name = tensor("op_10688"), val = tensor([1, 1])]; + tensor var_10690 = const()[name = tensor("op_10690"), val = tensor([1, 1])]; + tensor pretrained_out_531_pad_type_0 = const()[name = tensor("pretrained_out_531_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_531_pad_0 = const()[name = tensor("pretrained_out_531_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511456320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512275584))), name = tensor("layers_26_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_531_cast_fp16 = conv(dilations = var_10690, groups = var_10467, pad = pretrained_out_531_pad_0, pad_type = pretrained_out_531_pad_type_0, strides = var_10688, weight = layers_26_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_531_cast_fp16")]; + tensor var_10694 = const()[name = tensor("op_10694"), val = tensor([1, 1])]; + tensor var_10696 = const()[name = tensor("op_10696"), val = tensor([1, 1])]; + tensor input_793_pad_type_0 = const()[name = tensor("input_793_pad_type_0"), val = tensor("custom")]; + tensor input_793_pad_0 = const()[name = tensor("input_793_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512275712)))]; + tensor input_793_cast_fp16 = conv(dilations = var_10696, groups = var_10467, pad = input_793_pad_0, pad_type = input_793_pad_type_0, strides = var_10694, weight = layers_26_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_793_cast_fp16")]; + tensor var_10700 = const()[name = tensor("op_10700"), val = tensor([1, 1])]; + tensor var_10702 = const()[name = tensor("op_10702"), val = tensor([1, 1])]; + tensor lora_out_1061_pad_type_0 = const()[name = tensor("lora_out_1061_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1061_pad_0 = const()[name = tensor("lora_out_1061_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1063_weight_0_to_fp16 = const()[name = tensor("lora_out_1063_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512316736)))]; + tensor lora_out_1063_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10702, groups = var_10467, pad = lora_out_1061_pad_0, pad_type = lora_out_1061_pad_type_0, strides = var_10700, weight = lora_out_1063_weight_0_to_fp16, x = input_793_cast_fp16)[name = tensor("lora_out_1063_cast_fp16")]; + tensor key_107_cast_fp16 = add(x = pretrained_out_531_cast_fp16, y = lora_out_1063_cast_fp16)[name = tensor("key_107_cast_fp16")]; + tensor var_10713 = const()[name = tensor("op_10713"), val = tensor([1, 1])]; + tensor var_10715 = const()[name = tensor("op_10715"), val = tensor([1, 1])]; + tensor pretrained_out_533_pad_type_0 = const()[name = tensor("pretrained_out_533_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_533_pad_0 = const()[name = tensor("pretrained_out_533_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512357760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513177024))), name = tensor("layers_26_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_26_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513177152)))]; + tensor pretrained_out_533_cast_fp16 = conv(bias = layers_26_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_10715, groups = var_10467, pad = pretrained_out_533_pad_0, pad_type = pretrained_out_533_pad_type_0, strides = var_10713, weight = layers_26_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_533_cast_fp16")]; + tensor var_10719 = const()[name = tensor("op_10719"), val = tensor([1, 1])]; + tensor var_10721 = const()[name = tensor("op_10721"), val = tensor([1, 1])]; + tensor input_795_pad_type_0 = const()[name = tensor("input_795_pad_type_0"), val = tensor("custom")]; + tensor input_795_pad_0 = const()[name = tensor("input_795_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513179776)))]; + tensor input_795_cast_fp16 = conv(dilations = var_10721, groups = var_10467, pad = input_795_pad_0, pad_type = input_795_pad_type_0, strides = var_10719, weight = layers_26_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_795_cast_fp16")]; + tensor var_10725 = const()[name = tensor("op_10725"), val = tensor([1, 1])]; + tensor var_10727 = const()[name = tensor("op_10727"), val = tensor([1, 1])]; + tensor lora_out_1065_pad_type_0 = const()[name = tensor("lora_out_1065_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1065_pad_0 = const()[name = tensor("lora_out_1065_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1067_weight_0_to_fp16 = const()[name = tensor("lora_out_1067_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513220800)))]; + tensor lora_out_1067_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10727, groups = var_10467, pad = lora_out_1065_pad_0, pad_type = lora_out_1065_pad_type_0, strides = var_10725, weight = lora_out_1067_weight_0_to_fp16, x = input_795_cast_fp16)[name = tensor("lora_out_1067_cast_fp16")]; + tensor value_107_cast_fp16 = add(x = pretrained_out_533_cast_fp16, y = lora_out_1067_cast_fp16)[name = tensor("value_107_cast_fp16")]; + tensor var_10734 = const()[name = tensor("op_10734"), val = tensor([1, 20, 64, -1])]; + tensor var_10735_cast_fp16 = reshape(shape = var_10734, x = query_107_cast_fp16)[name = tensor("op_10735_cast_fp16")]; + tensor var_10736_to_fp16 = const()[name = tensor("op_10736_to_fp16"), val = tensor(0x1p-3)]; + tensor var_10737_cast_fp16 = mul(x = var_10735_cast_fp16, y = var_10736_to_fp16)[name = tensor("op_10737_cast_fp16")]; + tensor var_10738 = const()[name = tensor("op_10738"), val = tensor([1, 20, 64, -1])]; + tensor var_10739_cast_fp16 = reshape(shape = var_10738, x = key_107_cast_fp16)[name = tensor("op_10739_cast_fp16")]; + tensor mh_w_161_transpose_x_0 = const()[name = tensor("mh_w_161_transpose_x_0"), val = tensor(true)]; + tensor mh_w_161_transpose_y_0 = const()[name = tensor("mh_w_161_transpose_y_0"), val = tensor(false)]; + tensor mh_w_161_cast_fp16 = matmul(transpose_x = mh_w_161_transpose_x_0, transpose_y = mh_w_161_transpose_y_0, x = var_10737_cast_fp16, y = var_10739_cast_fp16)[name = tensor("mh_w_161_cast_fp16")]; + tensor obj_377_cast_fp16 = softmax(axis = var_10460, x = mh_w_161_cast_fp16)[name = tensor("obj_377_cast_fp16")]; + tensor var_10743 = const()[name = tensor("op_10743"), val = tensor([1, 20, 64, -1])]; + tensor var_10744_cast_fp16 = reshape(shape = var_10743, x = value_107_cast_fp16)[name = tensor("op_10744_cast_fp16")]; + tensor attn_107_transpose_x_0 = const()[name = tensor("attn_107_transpose_x_0"), val = tensor(false)]; + tensor attn_107_transpose_y_0 = const()[name = tensor("attn_107_transpose_y_0"), val = tensor(true)]; + tensor attn_107_cast_fp16 = matmul(transpose_x = attn_107_transpose_x_0, transpose_y = attn_107_transpose_y_0, x = var_10744_cast_fp16, y = obj_377_cast_fp16)[name = tensor("attn_107_cast_fp16")]; + tensor var_10747 = const()[name = tensor("op_10747"), val = tensor([1, 1280, 1, -1])]; + tensor input_797_cast_fp16 = reshape(shape = var_10747, x = attn_107_cast_fp16)[name = tensor("input_797_cast_fp16")]; + tensor var_10754 = const()[name = tensor("op_10754"), val = tensor([1, 1])]; + tensor var_10756 = const()[name = tensor("op_10756"), val = tensor([1, 1])]; + tensor pretrained_out_535_pad_type_0 = const()[name = tensor("pretrained_out_535_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_535_pad_0 = const()[name = tensor("pretrained_out_535_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513261824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514081088))), name = tensor("layers_26_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_26_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514081216)))]; + tensor pretrained_out_535_cast_fp16 = conv(bias = layers_26_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_10756, groups = var_10467, pad = pretrained_out_535_pad_0, pad_type = pretrained_out_535_pad_type_0, strides = var_10754, weight = layers_26_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = tensor("pretrained_out_535_cast_fp16")]; + tensor var_10760 = const()[name = tensor("op_10760"), val = tensor([1, 1])]; + tensor var_10762 = const()[name = tensor("op_10762"), val = tensor([1, 1])]; + tensor input_799_pad_type_0 = const()[name = tensor("input_799_pad_type_0"), val = tensor("custom")]; + tensor input_799_pad_0 = const()[name = tensor("input_799_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_26_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514083840)))]; + tensor input_799_cast_fp16 = conv(dilations = var_10762, groups = var_10467, pad = input_799_pad_0, pad_type = input_799_pad_type_0, strides = var_10760, weight = layers_26_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_797_cast_fp16)[name = tensor("input_799_cast_fp16")]; + tensor var_10766 = const()[name = tensor("op_10766"), val = tensor([1, 1])]; + tensor var_10768 = const()[name = tensor("op_10768"), val = tensor([1, 1])]; + tensor lora_out_1069_pad_type_0 = const()[name = tensor("lora_out_1069_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1069_pad_0 = const()[name = tensor("lora_out_1069_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1071_weight_0_to_fp16 = const()[name = tensor("lora_out_1071_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514124864)))]; + tensor lora_out_1071_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10768, groups = var_10467, pad = lora_out_1069_pad_0, pad_type = lora_out_1069_pad_type_0, strides = var_10766, weight = lora_out_1071_weight_0_to_fp16, x = input_799_cast_fp16)[name = tensor("lora_out_1071_cast_fp16")]; + tensor obj_375_cast_fp16 = add(x = pretrained_out_535_cast_fp16, y = lora_out_1071_cast_fp16)[name = tensor("obj_375_cast_fp16")]; + tensor inputs_161_cast_fp16 = add(x = inputs_159_cast_fp16, y = obj_375_cast_fp16)[name = tensor("inputs_161_cast_fp16")]; + tensor var_10780 = const()[name = tensor("op_10780"), val = tensor([1])]; + tensor channels_mean_161_cast_fp16 = reduce_mean(axes = var_10780, keep_dims = var_10468, x = inputs_161_cast_fp16)[name = tensor("channels_mean_161_cast_fp16")]; + tensor zero_mean_161_cast_fp16 = sub(x = inputs_161_cast_fp16, y = channels_mean_161_cast_fp16)[name = tensor("zero_mean_161_cast_fp16")]; + tensor zero_mean_sq_161_cast_fp16 = mul(x = zero_mean_161_cast_fp16, y = zero_mean_161_cast_fp16)[name = tensor("zero_mean_sq_161_cast_fp16")]; + tensor var_10784 = const()[name = tensor("op_10784"), val = tensor([1])]; + tensor var_10785_cast_fp16 = reduce_mean(axes = var_10784, keep_dims = var_10468, x = zero_mean_sq_161_cast_fp16)[name = tensor("op_10785_cast_fp16")]; + tensor var_10786_to_fp16 = const()[name = tensor("op_10786_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_10787_cast_fp16 = add(x = var_10785_cast_fp16, y = var_10786_to_fp16)[name = tensor("op_10787_cast_fp16")]; + tensor denom_161_epsilon_0 = const()[name = tensor("denom_161_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_161_cast_fp16 = rsqrt(epsilon = denom_161_epsilon_0, x = var_10787_cast_fp16)[name = tensor("denom_161_cast_fp16")]; + tensor out_161_cast_fp16 = mul(x = zero_mean_161_cast_fp16, y = denom_161_cast_fp16)[name = tensor("out_161_cast_fp16")]; + tensor input_801_gamma_0_to_fp16 = const()[name = tensor("input_801_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514165888)))]; + tensor input_801_beta_0_to_fp16 = const()[name = tensor("input_801_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514168512)))]; + tensor input_801_epsilon_0_to_fp16 = const()[name = tensor("input_801_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_801_cast_fp16 = batch_norm(beta = input_801_beta_0_to_fp16, epsilon = input_801_epsilon_0_to_fp16, gamma = input_801_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_161_cast_fp16)[name = tensor("input_801_cast_fp16")]; + tensor var_10801 = const()[name = tensor("op_10801"), val = tensor([1, 1])]; + tensor var_10803 = const()[name = tensor("op_10803"), val = tensor([1, 1])]; + tensor pretrained_out_537_pad_type_0 = const()[name = tensor("pretrained_out_537_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_537_pad_0 = const()[name = tensor("pretrained_out_537_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514171136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517448000))), name = tensor("layers_26_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_26_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517448128)))]; + tensor pretrained_out_537_cast_fp16 = conv(bias = layers_26_fc1_pretrained_bias_to_fp16, dilations = var_10803, groups = var_10467, pad = pretrained_out_537_pad_0, pad_type = pretrained_out_537_pad_type_0, strides = var_10801, weight = layers_26_fc1_pretrained_weight_to_fp16_palettized, x = input_801_cast_fp16)[name = tensor("pretrained_out_537_cast_fp16")]; + tensor var_10807 = const()[name = tensor("op_10807"), val = tensor([1, 1])]; + tensor var_10809 = const()[name = tensor("op_10809"), val = tensor([1, 1])]; + tensor input_803_pad_type_0 = const()[name = tensor("input_803_pad_type_0"), val = tensor("custom")]; + tensor input_803_pad_0 = const()[name = tensor("input_803_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_26_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517458432)))]; + tensor input_803_cast_fp16 = conv(dilations = var_10809, groups = var_10467, pad = input_803_pad_0, pad_type = input_803_pad_type_0, strides = var_10807, weight = layers_26_fc1_loraA_weight_to_fp16, x = input_801_cast_fp16)[name = tensor("input_803_cast_fp16")]; + tensor var_10813 = const()[name = tensor("op_10813"), val = tensor([1, 1])]; + tensor var_10815 = const()[name = tensor("op_10815"), val = tensor([1, 1])]; + tensor lora_out_1073_pad_type_0 = const()[name = tensor("lora_out_1073_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1073_pad_0 = const()[name = tensor("lora_out_1073_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1075_weight_0_to_fp16 = const()[name = tensor("lora_out_1075_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517499456)))]; + tensor lora_out_1075_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_10815, groups = var_10467, pad = lora_out_1073_pad_0, pad_type = lora_out_1073_pad_type_0, strides = var_10813, weight = lora_out_1075_weight_0_to_fp16, x = input_803_cast_fp16)[name = tensor("lora_out_1075_cast_fp16")]; + tensor input_805_cast_fp16 = add(x = pretrained_out_537_cast_fp16, y = lora_out_1075_cast_fp16)[name = tensor("input_805_cast_fp16")]; + tensor input_807_mode_0 = const()[name = tensor("input_807_mode_0"), val = tensor("EXACT")]; + tensor input_807_cast_fp16 = gelu(mode = input_807_mode_0, x = input_805_cast_fp16)[name = tensor("input_807_cast_fp16")]; + tensor var_10827 = const()[name = tensor("op_10827"), val = tensor([1, 1])]; + tensor var_10829 = const()[name = tensor("op_10829"), val = tensor([1, 1])]; + tensor pretrained_out_539_pad_type_0 = const()[name = tensor("pretrained_out_539_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_539_pad_0 = const()[name = tensor("pretrained_out_539_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(517663360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520940224))), name = tensor("layers_26_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_26_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_26_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520940352)))]; + tensor pretrained_out_539_cast_fp16 = conv(bias = layers_26_fc2_pretrained_bias_to_fp16, dilations = var_10829, groups = var_10467, pad = pretrained_out_539_pad_0, pad_type = pretrained_out_539_pad_type_0, strides = var_10827, weight = layers_26_fc2_pretrained_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = tensor("pretrained_out_539_cast_fp16")]; + tensor var_10833 = const()[name = tensor("op_10833"), val = tensor([1, 1])]; + tensor var_10835 = const()[name = tensor("op_10835"), val = tensor([1, 1])]; + tensor input_809_pad_type_0 = const()[name = tensor("input_809_pad_type_0"), val = tensor("custom")]; + tensor input_809_pad_0 = const()[name = tensor("input_809_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_26_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_26_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520942976)))]; + tensor input_809_cast_fp16 = conv(dilations = var_10835, groups = var_10467, pad = input_809_pad_0, pad_type = input_809_pad_type_0, strides = var_10833, weight = layers_26_fc2_loraA_weight_to_fp16, x = input_807_cast_fp16)[name = tensor("input_809_cast_fp16")]; + tensor var_10839 = const()[name = tensor("op_10839"), val = tensor([1, 1])]; + tensor var_10841 = const()[name = tensor("op_10841"), val = tensor([1, 1])]; + tensor lora_out_1077_pad_type_0 = const()[name = tensor("lora_out_1077_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1077_pad_0 = const()[name = tensor("lora_out_1077_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1079_weight_0_to_fp16 = const()[name = tensor("lora_out_1079_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521106880)))]; + tensor lora_out_1079_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10841, groups = var_10467, pad = lora_out_1077_pad_0, pad_type = lora_out_1077_pad_type_0, strides = var_10839, weight = lora_out_1079_weight_0_to_fp16, x = input_809_cast_fp16)[name = tensor("lora_out_1079_cast_fp16")]; + tensor hidden_states_55_cast_fp16 = add(x = pretrained_out_539_cast_fp16, y = lora_out_1079_cast_fp16)[name = tensor("hidden_states_55_cast_fp16")]; + tensor inputs_163_cast_fp16 = add(x = inputs_161_cast_fp16, y = hidden_states_55_cast_fp16)[name = tensor("inputs_163_cast_fp16")]; + tensor var_10858 = const()[name = tensor("op_10858"), val = tensor(3)]; + tensor var_10865 = const()[name = tensor("op_10865"), val = tensor(1)]; + tensor var_10866 = const()[name = tensor("op_10866"), val = tensor(true)]; + tensor var_10878 = const()[name = tensor("op_10878"), val = tensor([1])]; + tensor channels_mean_163_cast_fp16 = reduce_mean(axes = var_10878, keep_dims = var_10866, x = inputs_163_cast_fp16)[name = tensor("channels_mean_163_cast_fp16")]; + tensor zero_mean_163_cast_fp16 = sub(x = inputs_163_cast_fp16, y = channels_mean_163_cast_fp16)[name = tensor("zero_mean_163_cast_fp16")]; + tensor zero_mean_sq_163_cast_fp16 = mul(x = zero_mean_163_cast_fp16, y = zero_mean_163_cast_fp16)[name = tensor("zero_mean_sq_163_cast_fp16")]; + tensor var_10882 = const()[name = tensor("op_10882"), val = tensor([1])]; + tensor var_10883_cast_fp16 = reduce_mean(axes = var_10882, keep_dims = var_10866, x = zero_mean_sq_163_cast_fp16)[name = tensor("op_10883_cast_fp16")]; + tensor var_10884_to_fp16 = const()[name = tensor("op_10884_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_10885_cast_fp16 = add(x = var_10883_cast_fp16, y = var_10884_to_fp16)[name = tensor("op_10885_cast_fp16")]; + tensor denom_163_epsilon_0 = const()[name = tensor("denom_163_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_163_cast_fp16 = rsqrt(epsilon = denom_163_epsilon_0, x = var_10885_cast_fp16)[name = tensor("denom_163_cast_fp16")]; + tensor out_163_cast_fp16 = mul(x = zero_mean_163_cast_fp16, y = denom_163_cast_fp16)[name = tensor("out_163_cast_fp16")]; + tensor obj_379_gamma_0_to_fp16 = const()[name = tensor("obj_379_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521147904)))]; + tensor obj_379_beta_0_to_fp16 = const()[name = tensor("obj_379_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521150528)))]; + tensor obj_379_epsilon_0_to_fp16 = const()[name = tensor("obj_379_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_379_cast_fp16 = batch_norm(beta = obj_379_beta_0_to_fp16, epsilon = obj_379_epsilon_0_to_fp16, gamma = obj_379_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_163_cast_fp16)[name = tensor("obj_379_cast_fp16")]; + tensor var_10903 = const()[name = tensor("op_10903"), val = tensor([1, 1])]; + tensor var_10905 = const()[name = tensor("op_10905"), val = tensor([1, 1])]; + tensor pretrained_out_541_pad_type_0 = const()[name = tensor("pretrained_out_541_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_541_pad_0 = const()[name = tensor("pretrained_out_541_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521153152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521972416))), name = tensor("layers_27_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_27_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521972544)))]; + tensor pretrained_out_541_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_10905, groups = var_10865, pad = pretrained_out_541_pad_0, pad_type = pretrained_out_541_pad_type_0, strides = var_10903, weight = layers_27_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_379_cast_fp16)[name = tensor("pretrained_out_541_cast_fp16")]; + tensor var_10909 = const()[name = tensor("op_10909"), val = tensor([1, 1])]; + tensor var_10911 = const()[name = tensor("op_10911"), val = tensor([1, 1])]; + tensor input_811_pad_type_0 = const()[name = tensor("input_811_pad_type_0"), val = tensor("custom")]; + tensor input_811_pad_0 = const()[name = tensor("input_811_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521975168)))]; + tensor input_811_cast_fp16 = conv(dilations = var_10911, groups = var_10865, pad = input_811_pad_0, pad_type = input_811_pad_type_0, strides = var_10909, weight = layers_27_self_attn_q_proj_loraA_weight_to_fp16, x = obj_379_cast_fp16)[name = tensor("input_811_cast_fp16")]; + tensor var_10915 = const()[name = tensor("op_10915"), val = tensor([1, 1])]; + tensor var_10917 = const()[name = tensor("op_10917"), val = tensor([1, 1])]; + tensor lora_out_1081_pad_type_0 = const()[name = tensor("lora_out_1081_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1081_pad_0 = const()[name = tensor("lora_out_1081_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1083_weight_0_to_fp16 = const()[name = tensor("lora_out_1083_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522016192)))]; + tensor lora_out_1083_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10917, groups = var_10865, pad = lora_out_1081_pad_0, pad_type = lora_out_1081_pad_type_0, strides = var_10915, weight = lora_out_1083_weight_0_to_fp16, x = input_811_cast_fp16)[name = tensor("lora_out_1083_cast_fp16")]; + tensor query_109_cast_fp16 = add(x = pretrained_out_541_cast_fp16, y = lora_out_1083_cast_fp16)[name = tensor("query_109_cast_fp16")]; + tensor var_10927 = const()[name = tensor("op_10927"), val = tensor([1, 1])]; + tensor var_10929 = const()[name = tensor("op_10929"), val = tensor([1, 1])]; + tensor pretrained_out_543_pad_type_0 = const()[name = tensor("pretrained_out_543_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_543_pad_0 = const()[name = tensor("pretrained_out_543_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522057216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522876480))), name = tensor("layers_27_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_543_cast_fp16 = conv(dilations = var_10929, groups = var_10865, pad = pretrained_out_543_pad_0, pad_type = pretrained_out_543_pad_type_0, strides = var_10927, weight = layers_27_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_379_cast_fp16)[name = tensor("pretrained_out_543_cast_fp16")]; + tensor var_10933 = const()[name = tensor("op_10933"), val = tensor([1, 1])]; + tensor var_10935 = const()[name = tensor("op_10935"), val = tensor([1, 1])]; + tensor input_813_pad_type_0 = const()[name = tensor("input_813_pad_type_0"), val = tensor("custom")]; + tensor input_813_pad_0 = const()[name = tensor("input_813_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522876608)))]; + tensor input_813_cast_fp16 = conv(dilations = var_10935, groups = var_10865, pad = input_813_pad_0, pad_type = input_813_pad_type_0, strides = var_10933, weight = layers_27_self_attn_k_proj_loraA_weight_to_fp16, x = obj_379_cast_fp16)[name = tensor("input_813_cast_fp16")]; + tensor var_10939 = const()[name = tensor("op_10939"), val = tensor([1, 1])]; + tensor var_10941 = const()[name = tensor("op_10941"), val = tensor([1, 1])]; + tensor lora_out_1085_pad_type_0 = const()[name = tensor("lora_out_1085_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1085_pad_0 = const()[name = tensor("lora_out_1085_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1087_weight_0_to_fp16 = const()[name = tensor("lora_out_1087_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522917632)))]; + tensor lora_out_1087_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10941, groups = var_10865, pad = lora_out_1085_pad_0, pad_type = lora_out_1085_pad_type_0, strides = var_10939, weight = lora_out_1087_weight_0_to_fp16, x = input_813_cast_fp16)[name = tensor("lora_out_1087_cast_fp16")]; + tensor current_key_55_cast_fp16 = add(x = pretrained_out_543_cast_fp16, y = lora_out_1087_cast_fp16)[name = tensor("current_key_55_cast_fp16")]; + tensor var_10952 = const()[name = tensor("op_10952"), val = tensor([1, 1])]; + tensor var_10954 = const()[name = tensor("op_10954"), val = tensor([1, 1])]; + tensor pretrained_out_545_pad_type_0 = const()[name = tensor("pretrained_out_545_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_545_pad_0 = const()[name = tensor("pretrained_out_545_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522958656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523777920))), name = tensor("layers_27_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_27_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523778048)))]; + tensor pretrained_out_545_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_10954, groups = var_10865, pad = pretrained_out_545_pad_0, pad_type = pretrained_out_545_pad_type_0, strides = var_10952, weight = layers_27_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_379_cast_fp16)[name = tensor("pretrained_out_545_cast_fp16")]; + tensor var_10958 = const()[name = tensor("op_10958"), val = tensor([1, 1])]; + tensor var_10960 = const()[name = tensor("op_10960"), val = tensor([1, 1])]; + tensor input_815_pad_type_0 = const()[name = tensor("input_815_pad_type_0"), val = tensor("custom")]; + tensor input_815_pad_0 = const()[name = tensor("input_815_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523780672)))]; + tensor input_815_cast_fp16 = conv(dilations = var_10960, groups = var_10865, pad = input_815_pad_0, pad_type = input_815_pad_type_0, strides = var_10958, weight = layers_27_self_attn_v_proj_loraA_weight_to_fp16, x = obj_379_cast_fp16)[name = tensor("input_815_cast_fp16")]; + tensor var_10964 = const()[name = tensor("op_10964"), val = tensor([1, 1])]; + tensor var_10966 = const()[name = tensor("op_10966"), val = tensor([1, 1])]; + tensor lora_out_1089_pad_type_0 = const()[name = tensor("lora_out_1089_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1089_pad_0 = const()[name = tensor("lora_out_1089_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1091_weight_0_to_fp16 = const()[name = tensor("lora_out_1091_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523821696)))]; + tensor lora_out_1091_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_10966, groups = var_10865, pad = lora_out_1089_pad_0, pad_type = lora_out_1089_pad_type_0, strides = var_10964, weight = lora_out_1091_weight_0_to_fp16, x = input_815_cast_fp16)[name = tensor("lora_out_1091_cast_fp16")]; + tensor current_value_55_cast_fp16 = add(x = pretrained_out_545_cast_fp16, y = lora_out_1091_cast_fp16)[name = tensor("current_value_55_cast_fp16")]; + tensor var_10976_cast_fp16 = mul(x = current_key_55_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_10976_cast_fp16")]; + tensor var_10978_cast_fp16 = mul(x = var_103_cast_fp16_27, y = var_295_cast_fp16)[name = tensor("op_10978_cast_fp16")]; + tensor key_109_cast_fp16 = add(x = var_10976_cast_fp16, y = var_10978_cast_fp16)[name = tensor("key_109_cast_fp16")]; + tensor var_10980_cast_fp16 = mul(x = current_value_55_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_10980_cast_fp16")]; + tensor var_10982_cast_fp16 = mul(x = var_138_cast_fp16_27, y = var_295_cast_fp16)[name = tensor("op_10982_cast_fp16")]; + tensor value_109_cast_fp16 = add(x = var_10980_cast_fp16, y = var_10982_cast_fp16)[name = tensor("value_109_cast_fp16")]; + tensor var_10985 = const()[name = tensor("op_10985"), val = tensor([1, 20, 64, -1])]; + tensor var_10986_cast_fp16 = reshape(shape = var_10985, x = query_109_cast_fp16)[name = tensor("op_10986_cast_fp16")]; + tensor var_10987_to_fp16 = const()[name = tensor("op_10987_to_fp16"), val = tensor(0x1p-3)]; + tensor var_10988_cast_fp16 = mul(x = var_10986_cast_fp16, y = var_10987_to_fp16)[name = tensor("op_10988_cast_fp16")]; + tensor var_10989 = const()[name = tensor("op_10989"), val = tensor([1, 20, 64, -1])]; + tensor var_10990_cast_fp16 = reshape(shape = var_10989, x = key_109_cast_fp16)[name = tensor("op_10990_cast_fp16")]; + tensor mh_w_163_transpose_x_0 = const()[name = tensor("mh_w_163_transpose_x_0"), val = tensor(true)]; + tensor mh_w_163_transpose_y_0 = const()[name = tensor("mh_w_163_transpose_y_0"), val = tensor(false)]; + tensor mh_w_163_cast_fp16 = matmul(transpose_x = mh_w_163_transpose_x_0, transpose_y = mh_w_163_transpose_y_0, x = var_10988_cast_fp16, y = var_10990_cast_fp16)[name = tensor("mh_w_163_cast_fp16")]; + tensor mh_w_165_cast_fp16 = add(x = mh_w_163_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_165_cast_fp16")]; + tensor var_10998_cast_fp16 = softmax(axis = var_10858, x = mh_w_165_cast_fp16)[name = tensor("op_10998_cast_fp16")]; + tensor var_10999 = const()[name = tensor("op_10999"), val = tensor([1, 20, 64, -1])]; + tensor var_11000_cast_fp16 = reshape(shape = var_10999, x = value_109_cast_fp16)[name = tensor("op_11000_cast_fp16")]; + tensor attn_109_transpose_x_0 = const()[name = tensor("attn_109_transpose_x_0"), val = tensor(false)]; + tensor attn_109_transpose_y_0 = const()[name = tensor("attn_109_transpose_y_0"), val = tensor(true)]; + tensor attn_109_cast_fp16 = matmul(transpose_x = attn_109_transpose_x_0, transpose_y = attn_109_transpose_y_0, x = var_11000_cast_fp16, y = var_10998_cast_fp16)[name = tensor("attn_109_cast_fp16")]; + tensor var_11003 = const()[name = tensor("op_11003"), val = tensor([1, 1280, 1, -1])]; + tensor input_817_cast_fp16 = reshape(shape = var_11003, x = attn_109_cast_fp16)[name = tensor("input_817_cast_fp16")]; + tensor var_11010 = const()[name = tensor("op_11010"), val = tensor([1, 1])]; + tensor var_11012 = const()[name = tensor("op_11012"), val = tensor([1, 1])]; + tensor pretrained_out_547_pad_type_0 = const()[name = tensor("pretrained_out_547_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_547_pad_0 = const()[name = tensor("pretrained_out_547_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(523862720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524681984))), name = tensor("layers_27_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_27_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524682112)))]; + tensor pretrained_out_547_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_11012, groups = var_10865, pad = pretrained_out_547_pad_0, pad_type = pretrained_out_547_pad_type_0, strides = var_11010, weight = layers_27_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_817_cast_fp16)[name = tensor("pretrained_out_547_cast_fp16")]; + tensor var_11016 = const()[name = tensor("op_11016"), val = tensor([1, 1])]; + tensor var_11018 = const()[name = tensor("op_11018"), val = tensor([1, 1])]; + tensor input_819_pad_type_0 = const()[name = tensor("input_819_pad_type_0"), val = tensor("custom")]; + tensor input_819_pad_0 = const()[name = tensor("input_819_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524684736)))]; + tensor input_819_cast_fp16 = conv(dilations = var_11018, groups = var_10865, pad = input_819_pad_0, pad_type = input_819_pad_type_0, strides = var_11016, weight = layers_27_self_attn_o_proj_loraA_weight_to_fp16, x = input_817_cast_fp16)[name = tensor("input_819_cast_fp16")]; + tensor var_11022 = const()[name = tensor("op_11022"), val = tensor([1, 1])]; + tensor var_11024 = const()[name = tensor("op_11024"), val = tensor([1, 1])]; + tensor lora_out_1093_pad_type_0 = const()[name = tensor("lora_out_1093_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1093_pad_0 = const()[name = tensor("lora_out_1093_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1095_weight_0_to_fp16 = const()[name = tensor("lora_out_1095_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524725760)))]; + tensor lora_out_1095_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11024, groups = var_10865, pad = lora_out_1093_pad_0, pad_type = lora_out_1093_pad_type_0, strides = var_11022, weight = lora_out_1095_weight_0_to_fp16, x = input_819_cast_fp16)[name = tensor("lora_out_1095_cast_fp16")]; + tensor obj_385_cast_fp16 = add(x = pretrained_out_547_cast_fp16, y = lora_out_1095_cast_fp16)[name = tensor("obj_385_cast_fp16")]; + tensor inputs_165_cast_fp16 = add(x = inputs_163_cast_fp16, y = obj_385_cast_fp16)[name = tensor("inputs_165_cast_fp16")]; + tensor var_11037 = const()[name = tensor("op_11037"), val = tensor([1])]; + tensor channels_mean_165_cast_fp16 = reduce_mean(axes = var_11037, keep_dims = var_10866, x = inputs_165_cast_fp16)[name = tensor("channels_mean_165_cast_fp16")]; + tensor zero_mean_165_cast_fp16 = sub(x = inputs_165_cast_fp16, y = channels_mean_165_cast_fp16)[name = tensor("zero_mean_165_cast_fp16")]; + tensor zero_mean_sq_165_cast_fp16 = mul(x = zero_mean_165_cast_fp16, y = zero_mean_165_cast_fp16)[name = tensor("zero_mean_sq_165_cast_fp16")]; + tensor var_11041 = const()[name = tensor("op_11041"), val = tensor([1])]; + tensor var_11042_cast_fp16 = reduce_mean(axes = var_11041, keep_dims = var_10866, x = zero_mean_sq_165_cast_fp16)[name = tensor("op_11042_cast_fp16")]; + tensor var_11043_to_fp16 = const()[name = tensor("op_11043_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_11044_cast_fp16 = add(x = var_11042_cast_fp16, y = var_11043_to_fp16)[name = tensor("op_11044_cast_fp16")]; + tensor denom_165_epsilon_0 = const()[name = tensor("denom_165_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_165_cast_fp16 = rsqrt(epsilon = denom_165_epsilon_0, x = var_11044_cast_fp16)[name = tensor("denom_165_cast_fp16")]; + tensor out_165_cast_fp16 = mul(x = zero_mean_165_cast_fp16, y = denom_165_cast_fp16)[name = tensor("out_165_cast_fp16")]; + tensor obj_387_gamma_0_to_fp16 = const()[name = tensor("obj_387_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524766784)))]; + tensor obj_387_beta_0_to_fp16 = const()[name = tensor("obj_387_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524769408)))]; + tensor obj_387_epsilon_0_to_fp16 = const()[name = tensor("obj_387_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_387_cast_fp16 = batch_norm(beta = obj_387_beta_0_to_fp16, epsilon = obj_387_epsilon_0_to_fp16, gamma = obj_387_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_165_cast_fp16)[name = tensor("obj_387_cast_fp16")]; + tensor var_11062 = const()[name = tensor("op_11062"), val = tensor([1, 1])]; + tensor var_11064 = const()[name = tensor("op_11064"), val = tensor([1, 1])]; + tensor pretrained_out_549_pad_type_0 = const()[name = tensor("pretrained_out_549_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_549_pad_0 = const()[name = tensor("pretrained_out_549_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524772032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525591296))), name = tensor("layers_27_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_27_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525591424)))]; + tensor pretrained_out_549_cast_fp16 = conv(bias = layers_27_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_11064, groups = var_10865, pad = pretrained_out_549_pad_0, pad_type = pretrained_out_549_pad_type_0, strides = var_11062, weight = layers_27_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_387_cast_fp16)[name = tensor("pretrained_out_549_cast_fp16")]; + tensor var_11068 = const()[name = tensor("op_11068"), val = tensor([1, 1])]; + tensor var_11070 = const()[name = tensor("op_11070"), val = tensor([1, 1])]; + tensor input_821_pad_type_0 = const()[name = tensor("input_821_pad_type_0"), val = tensor("custom")]; + tensor input_821_pad_0 = const()[name = tensor("input_821_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525594048)))]; + tensor input_821_cast_fp16 = conv(dilations = var_11070, groups = var_10865, pad = input_821_pad_0, pad_type = input_821_pad_type_0, strides = var_11068, weight = layers_27_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_387_cast_fp16)[name = tensor("input_821_cast_fp16")]; + tensor var_11074 = const()[name = tensor("op_11074"), val = tensor([1, 1])]; + tensor var_11076 = const()[name = tensor("op_11076"), val = tensor([1, 1])]; + tensor lora_out_1097_pad_type_0 = const()[name = tensor("lora_out_1097_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1097_pad_0 = const()[name = tensor("lora_out_1097_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1099_weight_0_to_fp16 = const()[name = tensor("lora_out_1099_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525635072)))]; + tensor lora_out_1099_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11076, groups = var_10865, pad = lora_out_1097_pad_0, pad_type = lora_out_1097_pad_type_0, strides = var_11074, weight = lora_out_1099_weight_0_to_fp16, x = input_821_cast_fp16)[name = tensor("lora_out_1099_cast_fp16")]; + tensor query_111_cast_fp16 = add(x = pretrained_out_549_cast_fp16, y = lora_out_1099_cast_fp16)[name = tensor("query_111_cast_fp16")]; + tensor var_11086 = const()[name = tensor("op_11086"), val = tensor([1, 1])]; + tensor var_11088 = const()[name = tensor("op_11088"), val = tensor([1, 1])]; + tensor pretrained_out_551_pad_type_0 = const()[name = tensor("pretrained_out_551_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_551_pad_0 = const()[name = tensor("pretrained_out_551_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(525676096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526495360))), name = tensor("layers_27_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_551_cast_fp16 = conv(dilations = var_11088, groups = var_10865, pad = pretrained_out_551_pad_0, pad_type = pretrained_out_551_pad_type_0, strides = var_11086, weight = layers_27_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_551_cast_fp16")]; + tensor var_11092 = const()[name = tensor("op_11092"), val = tensor([1, 1])]; + tensor var_11094 = const()[name = tensor("op_11094"), val = tensor([1, 1])]; + tensor input_823_pad_type_0 = const()[name = tensor("input_823_pad_type_0"), val = tensor("custom")]; + tensor input_823_pad_0 = const()[name = tensor("input_823_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526495488)))]; + tensor input_823_cast_fp16 = conv(dilations = var_11094, groups = var_10865, pad = input_823_pad_0, pad_type = input_823_pad_type_0, strides = var_11092, weight = layers_27_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_823_cast_fp16")]; + tensor var_11098 = const()[name = tensor("op_11098"), val = tensor([1, 1])]; + tensor var_11100 = const()[name = tensor("op_11100"), val = tensor([1, 1])]; + tensor lora_out_1101_pad_type_0 = const()[name = tensor("lora_out_1101_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1101_pad_0 = const()[name = tensor("lora_out_1101_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1103_weight_0_to_fp16 = const()[name = tensor("lora_out_1103_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526536512)))]; + tensor lora_out_1103_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11100, groups = var_10865, pad = lora_out_1101_pad_0, pad_type = lora_out_1101_pad_type_0, strides = var_11098, weight = lora_out_1103_weight_0_to_fp16, x = input_823_cast_fp16)[name = tensor("lora_out_1103_cast_fp16")]; + tensor key_111_cast_fp16 = add(x = pretrained_out_551_cast_fp16, y = lora_out_1103_cast_fp16)[name = tensor("key_111_cast_fp16")]; + tensor var_11111 = const()[name = tensor("op_11111"), val = tensor([1, 1])]; + tensor var_11113 = const()[name = tensor("op_11113"), val = tensor([1, 1])]; + tensor pretrained_out_553_pad_type_0 = const()[name = tensor("pretrained_out_553_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_553_pad_0 = const()[name = tensor("pretrained_out_553_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(526577536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527396800))), name = tensor("layers_27_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_27_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527396928)))]; + tensor pretrained_out_553_cast_fp16 = conv(bias = layers_27_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_11113, groups = var_10865, pad = pretrained_out_553_pad_0, pad_type = pretrained_out_553_pad_type_0, strides = var_11111, weight = layers_27_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_553_cast_fp16")]; + tensor var_11117 = const()[name = tensor("op_11117"), val = tensor([1, 1])]; + tensor var_11119 = const()[name = tensor("op_11119"), val = tensor([1, 1])]; + tensor input_825_pad_type_0 = const()[name = tensor("input_825_pad_type_0"), val = tensor("custom")]; + tensor input_825_pad_0 = const()[name = tensor("input_825_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527399552)))]; + tensor input_825_cast_fp16 = conv(dilations = var_11119, groups = var_10865, pad = input_825_pad_0, pad_type = input_825_pad_type_0, strides = var_11117, weight = layers_27_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_825_cast_fp16")]; + tensor var_11123 = const()[name = tensor("op_11123"), val = tensor([1, 1])]; + tensor var_11125 = const()[name = tensor("op_11125"), val = tensor([1, 1])]; + tensor lora_out_1105_pad_type_0 = const()[name = tensor("lora_out_1105_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1105_pad_0 = const()[name = tensor("lora_out_1105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1107_weight_0_to_fp16 = const()[name = tensor("lora_out_1107_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527440576)))]; + tensor lora_out_1107_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11125, groups = var_10865, pad = lora_out_1105_pad_0, pad_type = lora_out_1105_pad_type_0, strides = var_11123, weight = lora_out_1107_weight_0_to_fp16, x = input_825_cast_fp16)[name = tensor("lora_out_1107_cast_fp16")]; + tensor value_111_cast_fp16 = add(x = pretrained_out_553_cast_fp16, y = lora_out_1107_cast_fp16)[name = tensor("value_111_cast_fp16")]; + tensor var_11132 = const()[name = tensor("op_11132"), val = tensor([1, 20, 64, -1])]; + tensor var_11133_cast_fp16 = reshape(shape = var_11132, x = query_111_cast_fp16)[name = tensor("op_11133_cast_fp16")]; + tensor var_11134_to_fp16 = const()[name = tensor("op_11134_to_fp16"), val = tensor(0x1p-3)]; + tensor var_11135_cast_fp16 = mul(x = var_11133_cast_fp16, y = var_11134_to_fp16)[name = tensor("op_11135_cast_fp16")]; + tensor var_11136 = const()[name = tensor("op_11136"), val = tensor([1, 20, 64, -1])]; + tensor var_11137_cast_fp16 = reshape(shape = var_11136, x = key_111_cast_fp16)[name = tensor("op_11137_cast_fp16")]; + tensor mh_w_167_transpose_x_0 = const()[name = tensor("mh_w_167_transpose_x_0"), val = tensor(true)]; + tensor mh_w_167_transpose_y_0 = const()[name = tensor("mh_w_167_transpose_y_0"), val = tensor(false)]; + tensor mh_w_167_cast_fp16 = matmul(transpose_x = mh_w_167_transpose_x_0, transpose_y = mh_w_167_transpose_y_0, x = var_11135_cast_fp16, y = var_11137_cast_fp16)[name = tensor("mh_w_167_cast_fp16")]; + tensor obj_391_cast_fp16 = softmax(axis = var_10858, x = mh_w_167_cast_fp16)[name = tensor("obj_391_cast_fp16")]; + tensor var_11141 = const()[name = tensor("op_11141"), val = tensor([1, 20, 64, -1])]; + tensor var_11142_cast_fp16 = reshape(shape = var_11141, x = value_111_cast_fp16)[name = tensor("op_11142_cast_fp16")]; + tensor attn_111_transpose_x_0 = const()[name = tensor("attn_111_transpose_x_0"), val = tensor(false)]; + tensor attn_111_transpose_y_0 = const()[name = tensor("attn_111_transpose_y_0"), val = tensor(true)]; + tensor attn_111_cast_fp16 = matmul(transpose_x = attn_111_transpose_x_0, transpose_y = attn_111_transpose_y_0, x = var_11142_cast_fp16, y = obj_391_cast_fp16)[name = tensor("attn_111_cast_fp16")]; + tensor var_11145 = const()[name = tensor("op_11145"), val = tensor([1, 1280, 1, -1])]; + tensor input_827_cast_fp16 = reshape(shape = var_11145, x = attn_111_cast_fp16)[name = tensor("input_827_cast_fp16")]; + tensor var_11152 = const()[name = tensor("op_11152"), val = tensor([1, 1])]; + tensor var_11154 = const()[name = tensor("op_11154"), val = tensor([1, 1])]; + tensor pretrained_out_555_pad_type_0 = const()[name = tensor("pretrained_out_555_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_555_pad_0 = const()[name = tensor("pretrained_out_555_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527481600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528300864))), name = tensor("layers_27_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_27_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528300992)))]; + tensor pretrained_out_555_cast_fp16 = conv(bias = layers_27_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_11154, groups = var_10865, pad = pretrained_out_555_pad_0, pad_type = pretrained_out_555_pad_type_0, strides = var_11152, weight = layers_27_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_827_cast_fp16)[name = tensor("pretrained_out_555_cast_fp16")]; + tensor var_11158 = const()[name = tensor("op_11158"), val = tensor([1, 1])]; + tensor var_11160 = const()[name = tensor("op_11160"), val = tensor([1, 1])]; + tensor input_829_pad_type_0 = const()[name = tensor("input_829_pad_type_0"), val = tensor("custom")]; + tensor input_829_pad_0 = const()[name = tensor("input_829_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_27_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528303616)))]; + tensor input_829_cast_fp16 = conv(dilations = var_11160, groups = var_10865, pad = input_829_pad_0, pad_type = input_829_pad_type_0, strides = var_11158, weight = layers_27_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_827_cast_fp16)[name = tensor("input_829_cast_fp16")]; + tensor var_11164 = const()[name = tensor("op_11164"), val = tensor([1, 1])]; + tensor var_11166 = const()[name = tensor("op_11166"), val = tensor([1, 1])]; + tensor lora_out_1109_pad_type_0 = const()[name = tensor("lora_out_1109_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1109_pad_0 = const()[name = tensor("lora_out_1109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1111_weight_0_to_fp16 = const()[name = tensor("lora_out_1111_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528344640)))]; + tensor lora_out_1111_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11166, groups = var_10865, pad = lora_out_1109_pad_0, pad_type = lora_out_1109_pad_type_0, strides = var_11164, weight = lora_out_1111_weight_0_to_fp16, x = input_829_cast_fp16)[name = tensor("lora_out_1111_cast_fp16")]; + tensor obj_389_cast_fp16 = add(x = pretrained_out_555_cast_fp16, y = lora_out_1111_cast_fp16)[name = tensor("obj_389_cast_fp16")]; + tensor inputs_167_cast_fp16 = add(x = inputs_165_cast_fp16, y = obj_389_cast_fp16)[name = tensor("inputs_167_cast_fp16")]; + tensor var_11178 = const()[name = tensor("op_11178"), val = tensor([1])]; + tensor channels_mean_167_cast_fp16 = reduce_mean(axes = var_11178, keep_dims = var_10866, x = inputs_167_cast_fp16)[name = tensor("channels_mean_167_cast_fp16")]; + tensor zero_mean_167_cast_fp16 = sub(x = inputs_167_cast_fp16, y = channels_mean_167_cast_fp16)[name = tensor("zero_mean_167_cast_fp16")]; + tensor zero_mean_sq_167_cast_fp16 = mul(x = zero_mean_167_cast_fp16, y = zero_mean_167_cast_fp16)[name = tensor("zero_mean_sq_167_cast_fp16")]; + tensor var_11182 = const()[name = tensor("op_11182"), val = tensor([1])]; + tensor var_11183_cast_fp16 = reduce_mean(axes = var_11182, keep_dims = var_10866, x = zero_mean_sq_167_cast_fp16)[name = tensor("op_11183_cast_fp16")]; + tensor var_11184_to_fp16 = const()[name = tensor("op_11184_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_11185_cast_fp16 = add(x = var_11183_cast_fp16, y = var_11184_to_fp16)[name = tensor("op_11185_cast_fp16")]; + tensor denom_167_epsilon_0 = const()[name = tensor("denom_167_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_167_cast_fp16 = rsqrt(epsilon = denom_167_epsilon_0, x = var_11185_cast_fp16)[name = tensor("denom_167_cast_fp16")]; + tensor out_167_cast_fp16 = mul(x = zero_mean_167_cast_fp16, y = denom_167_cast_fp16)[name = tensor("out_167_cast_fp16")]; + tensor input_831_gamma_0_to_fp16 = const()[name = tensor("input_831_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528385664)))]; + tensor input_831_beta_0_to_fp16 = const()[name = tensor("input_831_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528388288)))]; + tensor input_831_epsilon_0_to_fp16 = const()[name = tensor("input_831_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_831_cast_fp16 = batch_norm(beta = input_831_beta_0_to_fp16, epsilon = input_831_epsilon_0_to_fp16, gamma = input_831_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_167_cast_fp16)[name = tensor("input_831_cast_fp16")]; + tensor var_11199 = const()[name = tensor("op_11199"), val = tensor([1, 1])]; + tensor var_11201 = const()[name = tensor("op_11201"), val = tensor([1, 1])]; + tensor pretrained_out_557_pad_type_0 = const()[name = tensor("pretrained_out_557_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_557_pad_0 = const()[name = tensor("pretrained_out_557_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528390912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531667776))), name = tensor("layers_27_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_27_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531667904)))]; + tensor pretrained_out_557_cast_fp16 = conv(bias = layers_27_fc1_pretrained_bias_to_fp16, dilations = var_11201, groups = var_10865, pad = pretrained_out_557_pad_0, pad_type = pretrained_out_557_pad_type_0, strides = var_11199, weight = layers_27_fc1_pretrained_weight_to_fp16_palettized, x = input_831_cast_fp16)[name = tensor("pretrained_out_557_cast_fp16")]; + tensor var_11205 = const()[name = tensor("op_11205"), val = tensor([1, 1])]; + tensor var_11207 = const()[name = tensor("op_11207"), val = tensor([1, 1])]; + tensor input_833_pad_type_0 = const()[name = tensor("input_833_pad_type_0"), val = tensor("custom")]; + tensor input_833_pad_0 = const()[name = tensor("input_833_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_27_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531678208)))]; + tensor input_833_cast_fp16 = conv(dilations = var_11207, groups = var_10865, pad = input_833_pad_0, pad_type = input_833_pad_type_0, strides = var_11205, weight = layers_27_fc1_loraA_weight_to_fp16, x = input_831_cast_fp16)[name = tensor("input_833_cast_fp16")]; + tensor var_11211 = const()[name = tensor("op_11211"), val = tensor([1, 1])]; + tensor var_11213 = const()[name = tensor("op_11213"), val = tensor([1, 1])]; + tensor lora_out_1113_pad_type_0 = const()[name = tensor("lora_out_1113_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1113_pad_0 = const()[name = tensor("lora_out_1113_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1115_weight_0_to_fp16 = const()[name = tensor("lora_out_1115_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531719232)))]; + tensor lora_out_1115_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_11213, groups = var_10865, pad = lora_out_1113_pad_0, pad_type = lora_out_1113_pad_type_0, strides = var_11211, weight = lora_out_1115_weight_0_to_fp16, x = input_833_cast_fp16)[name = tensor("lora_out_1115_cast_fp16")]; + tensor input_835_cast_fp16 = add(x = pretrained_out_557_cast_fp16, y = lora_out_1115_cast_fp16)[name = tensor("input_835_cast_fp16")]; + tensor input_837_mode_0 = const()[name = tensor("input_837_mode_0"), val = tensor("EXACT")]; + tensor input_837_cast_fp16 = gelu(mode = input_837_mode_0, x = input_835_cast_fp16)[name = tensor("input_837_cast_fp16")]; + tensor var_11225 = const()[name = tensor("op_11225"), val = tensor([1, 1])]; + tensor var_11227 = const()[name = tensor("op_11227"), val = tensor([1, 1])]; + tensor pretrained_out_559_pad_type_0 = const()[name = tensor("pretrained_out_559_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_559_pad_0 = const()[name = tensor("pretrained_out_559_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(531883136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535160000))), name = tensor("layers_27_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_27_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_27_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535160128)))]; + tensor pretrained_out_559_cast_fp16 = conv(bias = layers_27_fc2_pretrained_bias_to_fp16, dilations = var_11227, groups = var_10865, pad = pretrained_out_559_pad_0, pad_type = pretrained_out_559_pad_type_0, strides = var_11225, weight = layers_27_fc2_pretrained_weight_to_fp16_palettized, x = input_837_cast_fp16)[name = tensor("pretrained_out_559_cast_fp16")]; + tensor var_11231 = const()[name = tensor("op_11231"), val = tensor([1, 1])]; + tensor var_11233 = const()[name = tensor("op_11233"), val = tensor([1, 1])]; + tensor input_839_pad_type_0 = const()[name = tensor("input_839_pad_type_0"), val = tensor("custom")]; + tensor input_839_pad_0 = const()[name = tensor("input_839_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_27_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_27_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535162752)))]; + tensor input_839_cast_fp16 = conv(dilations = var_11233, groups = var_10865, pad = input_839_pad_0, pad_type = input_839_pad_type_0, strides = var_11231, weight = layers_27_fc2_loraA_weight_to_fp16, x = input_837_cast_fp16)[name = tensor("input_839_cast_fp16")]; + tensor var_11237 = const()[name = tensor("op_11237"), val = tensor([1, 1])]; + tensor var_11239 = const()[name = tensor("op_11239"), val = tensor([1, 1])]; + tensor lora_out_1117_pad_type_0 = const()[name = tensor("lora_out_1117_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1117_pad_0 = const()[name = tensor("lora_out_1117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1119_weight_0_to_fp16 = const()[name = tensor("lora_out_1119_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535326656)))]; + tensor lora_out_1119_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11239, groups = var_10865, pad = lora_out_1117_pad_0, pad_type = lora_out_1117_pad_type_0, strides = var_11237, weight = lora_out_1119_weight_0_to_fp16, x = input_839_cast_fp16)[name = tensor("lora_out_1119_cast_fp16")]; + tensor hidden_states_57_cast_fp16 = add(x = pretrained_out_559_cast_fp16, y = lora_out_1119_cast_fp16)[name = tensor("hidden_states_57_cast_fp16")]; + tensor inputs_169_cast_fp16 = add(x = inputs_167_cast_fp16, y = hidden_states_57_cast_fp16)[name = tensor("inputs_169_cast_fp16")]; + tensor var_11256 = const()[name = tensor("op_11256"), val = tensor(3)]; + tensor var_11263 = const()[name = tensor("op_11263"), val = tensor(1)]; + tensor var_11264 = const()[name = tensor("op_11264"), val = tensor(true)]; + tensor var_11276 = const()[name = tensor("op_11276"), val = tensor([1])]; + tensor channels_mean_169_cast_fp16 = reduce_mean(axes = var_11276, keep_dims = var_11264, x = inputs_169_cast_fp16)[name = tensor("channels_mean_169_cast_fp16")]; + tensor zero_mean_169_cast_fp16 = sub(x = inputs_169_cast_fp16, y = channels_mean_169_cast_fp16)[name = tensor("zero_mean_169_cast_fp16")]; + tensor zero_mean_sq_169_cast_fp16 = mul(x = zero_mean_169_cast_fp16, y = zero_mean_169_cast_fp16)[name = tensor("zero_mean_sq_169_cast_fp16")]; + tensor var_11280 = const()[name = tensor("op_11280"), val = tensor([1])]; + tensor var_11281_cast_fp16 = reduce_mean(axes = var_11280, keep_dims = var_11264, x = zero_mean_sq_169_cast_fp16)[name = tensor("op_11281_cast_fp16")]; + tensor var_11282_to_fp16 = const()[name = tensor("op_11282_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_11283_cast_fp16 = add(x = var_11281_cast_fp16, y = var_11282_to_fp16)[name = tensor("op_11283_cast_fp16")]; + tensor denom_169_epsilon_0 = const()[name = tensor("denom_169_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_169_cast_fp16 = rsqrt(epsilon = denom_169_epsilon_0, x = var_11283_cast_fp16)[name = tensor("denom_169_cast_fp16")]; + tensor out_169_cast_fp16 = mul(x = zero_mean_169_cast_fp16, y = denom_169_cast_fp16)[name = tensor("out_169_cast_fp16")]; + tensor obj_393_gamma_0_to_fp16 = const()[name = tensor("obj_393_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535367680)))]; + tensor obj_393_beta_0_to_fp16 = const()[name = tensor("obj_393_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535370304)))]; + tensor obj_393_epsilon_0_to_fp16 = const()[name = tensor("obj_393_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_393_cast_fp16 = batch_norm(beta = obj_393_beta_0_to_fp16, epsilon = obj_393_epsilon_0_to_fp16, gamma = obj_393_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_169_cast_fp16)[name = tensor("obj_393_cast_fp16")]; + tensor var_11301 = const()[name = tensor("op_11301"), val = tensor([1, 1])]; + tensor var_11303 = const()[name = tensor("op_11303"), val = tensor([1, 1])]; + tensor pretrained_out_561_pad_type_0 = const()[name = tensor("pretrained_out_561_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_561_pad_0 = const()[name = tensor("pretrained_out_561_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535372928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536192192))), name = tensor("layers_28_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_28_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536192320)))]; + tensor pretrained_out_561_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_11303, groups = var_11263, pad = pretrained_out_561_pad_0, pad_type = pretrained_out_561_pad_type_0, strides = var_11301, weight = layers_28_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_393_cast_fp16)[name = tensor("pretrained_out_561_cast_fp16")]; + tensor var_11307 = const()[name = tensor("op_11307"), val = tensor([1, 1])]; + tensor var_11309 = const()[name = tensor("op_11309"), val = tensor([1, 1])]; + tensor input_841_pad_type_0 = const()[name = tensor("input_841_pad_type_0"), val = tensor("custom")]; + tensor input_841_pad_0 = const()[name = tensor("input_841_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536194944)))]; + tensor input_841_cast_fp16 = conv(dilations = var_11309, groups = var_11263, pad = input_841_pad_0, pad_type = input_841_pad_type_0, strides = var_11307, weight = layers_28_self_attn_q_proj_loraA_weight_to_fp16, x = obj_393_cast_fp16)[name = tensor("input_841_cast_fp16")]; + tensor var_11313 = const()[name = tensor("op_11313"), val = tensor([1, 1])]; + tensor var_11315 = const()[name = tensor("op_11315"), val = tensor([1, 1])]; + tensor lora_out_1121_pad_type_0 = const()[name = tensor("lora_out_1121_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1121_pad_0 = const()[name = tensor("lora_out_1121_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1123_weight_0_to_fp16 = const()[name = tensor("lora_out_1123_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536235968)))]; + tensor lora_out_1123_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11315, groups = var_11263, pad = lora_out_1121_pad_0, pad_type = lora_out_1121_pad_type_0, strides = var_11313, weight = lora_out_1123_weight_0_to_fp16, x = input_841_cast_fp16)[name = tensor("lora_out_1123_cast_fp16")]; + tensor query_113_cast_fp16 = add(x = pretrained_out_561_cast_fp16, y = lora_out_1123_cast_fp16)[name = tensor("query_113_cast_fp16")]; + tensor var_11325 = const()[name = tensor("op_11325"), val = tensor([1, 1])]; + tensor var_11327 = const()[name = tensor("op_11327"), val = tensor([1, 1])]; + tensor pretrained_out_563_pad_type_0 = const()[name = tensor("pretrained_out_563_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_563_pad_0 = const()[name = tensor("pretrained_out_563_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536276992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537096256))), name = tensor("layers_28_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_563_cast_fp16 = conv(dilations = var_11327, groups = var_11263, pad = pretrained_out_563_pad_0, pad_type = pretrained_out_563_pad_type_0, strides = var_11325, weight = layers_28_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_393_cast_fp16)[name = tensor("pretrained_out_563_cast_fp16")]; + tensor var_11331 = const()[name = tensor("op_11331"), val = tensor([1, 1])]; + tensor var_11333 = const()[name = tensor("op_11333"), val = tensor([1, 1])]; + tensor input_843_pad_type_0 = const()[name = tensor("input_843_pad_type_0"), val = tensor("custom")]; + tensor input_843_pad_0 = const()[name = tensor("input_843_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537096384)))]; + tensor input_843_cast_fp16 = conv(dilations = var_11333, groups = var_11263, pad = input_843_pad_0, pad_type = input_843_pad_type_0, strides = var_11331, weight = layers_28_self_attn_k_proj_loraA_weight_to_fp16, x = obj_393_cast_fp16)[name = tensor("input_843_cast_fp16")]; + tensor var_11337 = const()[name = tensor("op_11337"), val = tensor([1, 1])]; + tensor var_11339 = const()[name = tensor("op_11339"), val = tensor([1, 1])]; + tensor lora_out_1125_pad_type_0 = const()[name = tensor("lora_out_1125_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1125_pad_0 = const()[name = tensor("lora_out_1125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1127_weight_0_to_fp16 = const()[name = tensor("lora_out_1127_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537137408)))]; + tensor lora_out_1127_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11339, groups = var_11263, pad = lora_out_1125_pad_0, pad_type = lora_out_1125_pad_type_0, strides = var_11337, weight = lora_out_1127_weight_0_to_fp16, x = input_843_cast_fp16)[name = tensor("lora_out_1127_cast_fp16")]; + tensor current_key_57_cast_fp16 = add(x = pretrained_out_563_cast_fp16, y = lora_out_1127_cast_fp16)[name = tensor("current_key_57_cast_fp16")]; + tensor var_11350 = const()[name = tensor("op_11350"), val = tensor([1, 1])]; + tensor var_11352 = const()[name = tensor("op_11352"), val = tensor([1, 1])]; + tensor pretrained_out_565_pad_type_0 = const()[name = tensor("pretrained_out_565_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_565_pad_0 = const()[name = tensor("pretrained_out_565_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537178432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537997696))), name = tensor("layers_28_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_28_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537997824)))]; + tensor pretrained_out_565_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_11352, groups = var_11263, pad = pretrained_out_565_pad_0, pad_type = pretrained_out_565_pad_type_0, strides = var_11350, weight = layers_28_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_393_cast_fp16)[name = tensor("pretrained_out_565_cast_fp16")]; + tensor var_11356 = const()[name = tensor("op_11356"), val = tensor([1, 1])]; + tensor var_11358 = const()[name = tensor("op_11358"), val = tensor([1, 1])]; + tensor input_845_pad_type_0 = const()[name = tensor("input_845_pad_type_0"), val = tensor("custom")]; + tensor input_845_pad_0 = const()[name = tensor("input_845_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538000448)))]; + tensor input_845_cast_fp16 = conv(dilations = var_11358, groups = var_11263, pad = input_845_pad_0, pad_type = input_845_pad_type_0, strides = var_11356, weight = layers_28_self_attn_v_proj_loraA_weight_to_fp16, x = obj_393_cast_fp16)[name = tensor("input_845_cast_fp16")]; + tensor var_11362 = const()[name = tensor("op_11362"), val = tensor([1, 1])]; + tensor var_11364 = const()[name = tensor("op_11364"), val = tensor([1, 1])]; + tensor lora_out_1129_pad_type_0 = const()[name = tensor("lora_out_1129_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1129_pad_0 = const()[name = tensor("lora_out_1129_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1131_weight_0_to_fp16 = const()[name = tensor("lora_out_1131_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538041472)))]; + tensor lora_out_1131_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11364, groups = var_11263, pad = lora_out_1129_pad_0, pad_type = lora_out_1129_pad_type_0, strides = var_11362, weight = lora_out_1131_weight_0_to_fp16, x = input_845_cast_fp16)[name = tensor("lora_out_1131_cast_fp16")]; + tensor current_value_57_cast_fp16 = add(x = pretrained_out_565_cast_fp16, y = lora_out_1131_cast_fp16)[name = tensor("current_value_57_cast_fp16")]; + tensor var_11374_cast_fp16 = mul(x = current_key_57_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_11374_cast_fp16")]; + tensor var_11376_cast_fp16 = mul(x = var_103_cast_fp16_28, y = var_295_cast_fp16)[name = tensor("op_11376_cast_fp16")]; + tensor key_113_cast_fp16 = add(x = var_11374_cast_fp16, y = var_11376_cast_fp16)[name = tensor("key_113_cast_fp16")]; + tensor var_11378_cast_fp16 = mul(x = current_value_57_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_11378_cast_fp16")]; + tensor var_11380_cast_fp16 = mul(x = var_138_cast_fp16_28, y = var_295_cast_fp16)[name = tensor("op_11380_cast_fp16")]; + tensor value_113_cast_fp16 = add(x = var_11378_cast_fp16, y = var_11380_cast_fp16)[name = tensor("value_113_cast_fp16")]; + tensor var_11383 = const()[name = tensor("op_11383"), val = tensor([1, 20, 64, -1])]; + tensor var_11384_cast_fp16 = reshape(shape = var_11383, x = query_113_cast_fp16)[name = tensor("op_11384_cast_fp16")]; + tensor var_11385_to_fp16 = const()[name = tensor("op_11385_to_fp16"), val = tensor(0x1p-3)]; + tensor var_11386_cast_fp16 = mul(x = var_11384_cast_fp16, y = var_11385_to_fp16)[name = tensor("op_11386_cast_fp16")]; + tensor var_11387 = const()[name = tensor("op_11387"), val = tensor([1, 20, 64, -1])]; + tensor var_11388_cast_fp16 = reshape(shape = var_11387, x = key_113_cast_fp16)[name = tensor("op_11388_cast_fp16")]; + tensor mh_w_169_transpose_x_0 = const()[name = tensor("mh_w_169_transpose_x_0"), val = tensor(true)]; + tensor mh_w_169_transpose_y_0 = const()[name = tensor("mh_w_169_transpose_y_0"), val = tensor(false)]; + tensor mh_w_169_cast_fp16 = matmul(transpose_x = mh_w_169_transpose_x_0, transpose_y = mh_w_169_transpose_y_0, x = var_11386_cast_fp16, y = var_11388_cast_fp16)[name = tensor("mh_w_169_cast_fp16")]; + tensor mh_w_171_cast_fp16 = add(x = mh_w_169_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_171_cast_fp16")]; + tensor var_11396_cast_fp16 = softmax(axis = var_11256, x = mh_w_171_cast_fp16)[name = tensor("op_11396_cast_fp16")]; + tensor var_11397 = const()[name = tensor("op_11397"), val = tensor([1, 20, 64, -1])]; + tensor var_11398_cast_fp16 = reshape(shape = var_11397, x = value_113_cast_fp16)[name = tensor("op_11398_cast_fp16")]; + tensor attn_113_transpose_x_0 = const()[name = tensor("attn_113_transpose_x_0"), val = tensor(false)]; + tensor attn_113_transpose_y_0 = const()[name = tensor("attn_113_transpose_y_0"), val = tensor(true)]; + tensor attn_113_cast_fp16 = matmul(transpose_x = attn_113_transpose_x_0, transpose_y = attn_113_transpose_y_0, x = var_11398_cast_fp16, y = var_11396_cast_fp16)[name = tensor("attn_113_cast_fp16")]; + tensor var_11401 = const()[name = tensor("op_11401"), val = tensor([1, 1280, 1, -1])]; + tensor input_847_cast_fp16 = reshape(shape = var_11401, x = attn_113_cast_fp16)[name = tensor("input_847_cast_fp16")]; + tensor var_11408 = const()[name = tensor("op_11408"), val = tensor([1, 1])]; + tensor var_11410 = const()[name = tensor("op_11410"), val = tensor([1, 1])]; + tensor pretrained_out_567_pad_type_0 = const()[name = tensor("pretrained_out_567_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_567_pad_0 = const()[name = tensor("pretrained_out_567_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538082496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538901760))), name = tensor("layers_28_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_28_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538901888)))]; + tensor pretrained_out_567_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_11410, groups = var_11263, pad = pretrained_out_567_pad_0, pad_type = pretrained_out_567_pad_type_0, strides = var_11408, weight = layers_28_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_847_cast_fp16)[name = tensor("pretrained_out_567_cast_fp16")]; + tensor var_11414 = const()[name = tensor("op_11414"), val = tensor([1, 1])]; + tensor var_11416 = const()[name = tensor("op_11416"), val = tensor([1, 1])]; + tensor input_849_pad_type_0 = const()[name = tensor("input_849_pad_type_0"), val = tensor("custom")]; + tensor input_849_pad_0 = const()[name = tensor("input_849_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538904512)))]; + tensor input_849_cast_fp16 = conv(dilations = var_11416, groups = var_11263, pad = input_849_pad_0, pad_type = input_849_pad_type_0, strides = var_11414, weight = layers_28_self_attn_o_proj_loraA_weight_to_fp16, x = input_847_cast_fp16)[name = tensor("input_849_cast_fp16")]; + tensor var_11420 = const()[name = tensor("op_11420"), val = tensor([1, 1])]; + tensor var_11422 = const()[name = tensor("op_11422"), val = tensor([1, 1])]; + tensor lora_out_1133_pad_type_0 = const()[name = tensor("lora_out_1133_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1133_pad_0 = const()[name = tensor("lora_out_1133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1135_weight_0_to_fp16 = const()[name = tensor("lora_out_1135_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538945536)))]; + tensor lora_out_1135_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11422, groups = var_11263, pad = lora_out_1133_pad_0, pad_type = lora_out_1133_pad_type_0, strides = var_11420, weight = lora_out_1135_weight_0_to_fp16, x = input_849_cast_fp16)[name = tensor("lora_out_1135_cast_fp16")]; + tensor obj_399_cast_fp16 = add(x = pretrained_out_567_cast_fp16, y = lora_out_1135_cast_fp16)[name = tensor("obj_399_cast_fp16")]; + tensor inputs_171_cast_fp16 = add(x = inputs_169_cast_fp16, y = obj_399_cast_fp16)[name = tensor("inputs_171_cast_fp16")]; + tensor var_11435 = const()[name = tensor("op_11435"), val = tensor([1])]; + tensor channels_mean_171_cast_fp16 = reduce_mean(axes = var_11435, keep_dims = var_11264, x = inputs_171_cast_fp16)[name = tensor("channels_mean_171_cast_fp16")]; + tensor zero_mean_171_cast_fp16 = sub(x = inputs_171_cast_fp16, y = channels_mean_171_cast_fp16)[name = tensor("zero_mean_171_cast_fp16")]; + tensor zero_mean_sq_171_cast_fp16 = mul(x = zero_mean_171_cast_fp16, y = zero_mean_171_cast_fp16)[name = tensor("zero_mean_sq_171_cast_fp16")]; + tensor var_11439 = const()[name = tensor("op_11439"), val = tensor([1])]; + tensor var_11440_cast_fp16 = reduce_mean(axes = var_11439, keep_dims = var_11264, x = zero_mean_sq_171_cast_fp16)[name = tensor("op_11440_cast_fp16")]; + tensor var_11441_to_fp16 = const()[name = tensor("op_11441_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_11442_cast_fp16 = add(x = var_11440_cast_fp16, y = var_11441_to_fp16)[name = tensor("op_11442_cast_fp16")]; + tensor denom_171_epsilon_0 = const()[name = tensor("denom_171_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_171_cast_fp16 = rsqrt(epsilon = denom_171_epsilon_0, x = var_11442_cast_fp16)[name = tensor("denom_171_cast_fp16")]; + tensor out_171_cast_fp16 = mul(x = zero_mean_171_cast_fp16, y = denom_171_cast_fp16)[name = tensor("out_171_cast_fp16")]; + tensor obj_401_gamma_0_to_fp16 = const()[name = tensor("obj_401_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538986560)))]; + tensor obj_401_beta_0_to_fp16 = const()[name = tensor("obj_401_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538989184)))]; + tensor obj_401_epsilon_0_to_fp16 = const()[name = tensor("obj_401_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_401_cast_fp16 = batch_norm(beta = obj_401_beta_0_to_fp16, epsilon = obj_401_epsilon_0_to_fp16, gamma = obj_401_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_171_cast_fp16)[name = tensor("obj_401_cast_fp16")]; + tensor var_11460 = const()[name = tensor("op_11460"), val = tensor([1, 1])]; + tensor var_11462 = const()[name = tensor("op_11462"), val = tensor([1, 1])]; + tensor pretrained_out_569_pad_type_0 = const()[name = tensor("pretrained_out_569_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_569_pad_0 = const()[name = tensor("pretrained_out_569_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538991808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539811072))), name = tensor("layers_28_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_28_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539811200)))]; + tensor pretrained_out_569_cast_fp16 = conv(bias = layers_28_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_11462, groups = var_11263, pad = pretrained_out_569_pad_0, pad_type = pretrained_out_569_pad_type_0, strides = var_11460, weight = layers_28_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_401_cast_fp16)[name = tensor("pretrained_out_569_cast_fp16")]; + tensor var_11466 = const()[name = tensor("op_11466"), val = tensor([1, 1])]; + tensor var_11468 = const()[name = tensor("op_11468"), val = tensor([1, 1])]; + tensor input_851_pad_type_0 = const()[name = tensor("input_851_pad_type_0"), val = tensor("custom")]; + tensor input_851_pad_0 = const()[name = tensor("input_851_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539813824)))]; + tensor input_851_cast_fp16 = conv(dilations = var_11468, groups = var_11263, pad = input_851_pad_0, pad_type = input_851_pad_type_0, strides = var_11466, weight = layers_28_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_401_cast_fp16)[name = tensor("input_851_cast_fp16")]; + tensor var_11472 = const()[name = tensor("op_11472"), val = tensor([1, 1])]; + tensor var_11474 = const()[name = tensor("op_11474"), val = tensor([1, 1])]; + tensor lora_out_1137_pad_type_0 = const()[name = tensor("lora_out_1137_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1137_pad_0 = const()[name = tensor("lora_out_1137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1139_weight_0_to_fp16 = const()[name = tensor("lora_out_1139_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539854848)))]; + tensor lora_out_1139_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11474, groups = var_11263, pad = lora_out_1137_pad_0, pad_type = lora_out_1137_pad_type_0, strides = var_11472, weight = lora_out_1139_weight_0_to_fp16, x = input_851_cast_fp16)[name = tensor("lora_out_1139_cast_fp16")]; + tensor query_115_cast_fp16 = add(x = pretrained_out_569_cast_fp16, y = lora_out_1139_cast_fp16)[name = tensor("query_115_cast_fp16")]; + tensor var_11484 = const()[name = tensor("op_11484"), val = tensor([1, 1])]; + tensor var_11486 = const()[name = tensor("op_11486"), val = tensor([1, 1])]; + tensor pretrained_out_571_pad_type_0 = const()[name = tensor("pretrained_out_571_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_571_pad_0 = const()[name = tensor("pretrained_out_571_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(539895872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540715136))), name = tensor("layers_28_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_571_cast_fp16 = conv(dilations = var_11486, groups = var_11263, pad = pretrained_out_571_pad_0, pad_type = pretrained_out_571_pad_type_0, strides = var_11484, weight = layers_28_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_571_cast_fp16")]; + tensor var_11490 = const()[name = tensor("op_11490"), val = tensor([1, 1])]; + tensor var_11492 = const()[name = tensor("op_11492"), val = tensor([1, 1])]; + tensor input_853_pad_type_0 = const()[name = tensor("input_853_pad_type_0"), val = tensor("custom")]; + tensor input_853_pad_0 = const()[name = tensor("input_853_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540715264)))]; + tensor input_853_cast_fp16 = conv(dilations = var_11492, groups = var_11263, pad = input_853_pad_0, pad_type = input_853_pad_type_0, strides = var_11490, weight = layers_28_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_853_cast_fp16")]; + tensor var_11496 = const()[name = tensor("op_11496"), val = tensor([1, 1])]; + tensor var_11498 = const()[name = tensor("op_11498"), val = tensor([1, 1])]; + tensor lora_out_1141_pad_type_0 = const()[name = tensor("lora_out_1141_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1141_pad_0 = const()[name = tensor("lora_out_1141_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1143_weight_0_to_fp16 = const()[name = tensor("lora_out_1143_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540756288)))]; + tensor lora_out_1143_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11498, groups = var_11263, pad = lora_out_1141_pad_0, pad_type = lora_out_1141_pad_type_0, strides = var_11496, weight = lora_out_1143_weight_0_to_fp16, x = input_853_cast_fp16)[name = tensor("lora_out_1143_cast_fp16")]; + tensor key_115_cast_fp16 = add(x = pretrained_out_571_cast_fp16, y = lora_out_1143_cast_fp16)[name = tensor("key_115_cast_fp16")]; + tensor var_11509 = const()[name = tensor("op_11509"), val = tensor([1, 1])]; + tensor var_11511 = const()[name = tensor("op_11511"), val = tensor([1, 1])]; + tensor pretrained_out_573_pad_type_0 = const()[name = tensor("pretrained_out_573_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_573_pad_0 = const()[name = tensor("pretrained_out_573_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540797312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541616576))), name = tensor("layers_28_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_28_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541616704)))]; + tensor pretrained_out_573_cast_fp16 = conv(bias = layers_28_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_11511, groups = var_11263, pad = pretrained_out_573_pad_0, pad_type = pretrained_out_573_pad_type_0, strides = var_11509, weight = layers_28_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_573_cast_fp16")]; + tensor var_11515 = const()[name = tensor("op_11515"), val = tensor([1, 1])]; + tensor var_11517 = const()[name = tensor("op_11517"), val = tensor([1, 1])]; + tensor input_855_pad_type_0 = const()[name = tensor("input_855_pad_type_0"), val = tensor("custom")]; + tensor input_855_pad_0 = const()[name = tensor("input_855_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541619328)))]; + tensor input_855_cast_fp16 = conv(dilations = var_11517, groups = var_11263, pad = input_855_pad_0, pad_type = input_855_pad_type_0, strides = var_11515, weight = layers_28_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_855_cast_fp16")]; + tensor var_11521 = const()[name = tensor("op_11521"), val = tensor([1, 1])]; + tensor var_11523 = const()[name = tensor("op_11523"), val = tensor([1, 1])]; + tensor lora_out_1145_pad_type_0 = const()[name = tensor("lora_out_1145_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1145_pad_0 = const()[name = tensor("lora_out_1145_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1147_weight_0_to_fp16 = const()[name = tensor("lora_out_1147_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541660352)))]; + tensor lora_out_1147_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11523, groups = var_11263, pad = lora_out_1145_pad_0, pad_type = lora_out_1145_pad_type_0, strides = var_11521, weight = lora_out_1147_weight_0_to_fp16, x = input_855_cast_fp16)[name = tensor("lora_out_1147_cast_fp16")]; + tensor value_115_cast_fp16 = add(x = pretrained_out_573_cast_fp16, y = lora_out_1147_cast_fp16)[name = tensor("value_115_cast_fp16")]; + tensor var_11530 = const()[name = tensor("op_11530"), val = tensor([1, 20, 64, -1])]; + tensor var_11531_cast_fp16 = reshape(shape = var_11530, x = query_115_cast_fp16)[name = tensor("op_11531_cast_fp16")]; + tensor var_11532_to_fp16 = const()[name = tensor("op_11532_to_fp16"), val = tensor(0x1p-3)]; + tensor var_11533_cast_fp16 = mul(x = var_11531_cast_fp16, y = var_11532_to_fp16)[name = tensor("op_11533_cast_fp16")]; + tensor var_11534 = const()[name = tensor("op_11534"), val = tensor([1, 20, 64, -1])]; + tensor var_11535_cast_fp16 = reshape(shape = var_11534, x = key_115_cast_fp16)[name = tensor("op_11535_cast_fp16")]; + tensor mh_w_173_transpose_x_0 = const()[name = tensor("mh_w_173_transpose_x_0"), val = tensor(true)]; + tensor mh_w_173_transpose_y_0 = const()[name = tensor("mh_w_173_transpose_y_0"), val = tensor(false)]; + tensor mh_w_173_cast_fp16 = matmul(transpose_x = mh_w_173_transpose_x_0, transpose_y = mh_w_173_transpose_y_0, x = var_11533_cast_fp16, y = var_11535_cast_fp16)[name = tensor("mh_w_173_cast_fp16")]; + tensor obj_405_cast_fp16 = softmax(axis = var_11256, x = mh_w_173_cast_fp16)[name = tensor("obj_405_cast_fp16")]; + tensor var_11539 = const()[name = tensor("op_11539"), val = tensor([1, 20, 64, -1])]; + tensor var_11540_cast_fp16 = reshape(shape = var_11539, x = value_115_cast_fp16)[name = tensor("op_11540_cast_fp16")]; + tensor attn_115_transpose_x_0 = const()[name = tensor("attn_115_transpose_x_0"), val = tensor(false)]; + tensor attn_115_transpose_y_0 = const()[name = tensor("attn_115_transpose_y_0"), val = tensor(true)]; + tensor attn_115_cast_fp16 = matmul(transpose_x = attn_115_transpose_x_0, transpose_y = attn_115_transpose_y_0, x = var_11540_cast_fp16, y = obj_405_cast_fp16)[name = tensor("attn_115_cast_fp16")]; + tensor var_11543 = const()[name = tensor("op_11543"), val = tensor([1, 1280, 1, -1])]; + tensor input_857_cast_fp16 = reshape(shape = var_11543, x = attn_115_cast_fp16)[name = tensor("input_857_cast_fp16")]; + tensor var_11550 = const()[name = tensor("op_11550"), val = tensor([1, 1])]; + tensor var_11552 = const()[name = tensor("op_11552"), val = tensor([1, 1])]; + tensor pretrained_out_575_pad_type_0 = const()[name = tensor("pretrained_out_575_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_575_pad_0 = const()[name = tensor("pretrained_out_575_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(541701376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542520640))), name = tensor("layers_28_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_28_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542520768)))]; + tensor pretrained_out_575_cast_fp16 = conv(bias = layers_28_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_11552, groups = var_11263, pad = pretrained_out_575_pad_0, pad_type = pretrained_out_575_pad_type_0, strides = var_11550, weight = layers_28_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_857_cast_fp16)[name = tensor("pretrained_out_575_cast_fp16")]; + tensor var_11556 = const()[name = tensor("op_11556"), val = tensor([1, 1])]; + tensor var_11558 = const()[name = tensor("op_11558"), val = tensor([1, 1])]; + tensor input_859_pad_type_0 = const()[name = tensor("input_859_pad_type_0"), val = tensor("custom")]; + tensor input_859_pad_0 = const()[name = tensor("input_859_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_28_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542523392)))]; + tensor input_859_cast_fp16 = conv(dilations = var_11558, groups = var_11263, pad = input_859_pad_0, pad_type = input_859_pad_type_0, strides = var_11556, weight = layers_28_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_857_cast_fp16)[name = tensor("input_859_cast_fp16")]; + tensor var_11562 = const()[name = tensor("op_11562"), val = tensor([1, 1])]; + tensor var_11564 = const()[name = tensor("op_11564"), val = tensor([1, 1])]; + tensor lora_out_1149_pad_type_0 = const()[name = tensor("lora_out_1149_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1149_pad_0 = const()[name = tensor("lora_out_1149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1151_weight_0_to_fp16 = const()[name = tensor("lora_out_1151_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542564416)))]; + tensor lora_out_1151_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11564, groups = var_11263, pad = lora_out_1149_pad_0, pad_type = lora_out_1149_pad_type_0, strides = var_11562, weight = lora_out_1151_weight_0_to_fp16, x = input_859_cast_fp16)[name = tensor("lora_out_1151_cast_fp16")]; + tensor obj_403_cast_fp16 = add(x = pretrained_out_575_cast_fp16, y = lora_out_1151_cast_fp16)[name = tensor("obj_403_cast_fp16")]; + tensor inputs_173_cast_fp16 = add(x = inputs_171_cast_fp16, y = obj_403_cast_fp16)[name = tensor("inputs_173_cast_fp16")]; + tensor var_11573 = const()[name = tensor("op_11573"), val = tensor([1])]; + tensor channels_mean_173_cast_fp16 = reduce_mean(axes = var_11573, keep_dims = var_11264, x = inputs_173_cast_fp16)[name = tensor("channels_mean_173_cast_fp16")]; + tensor zero_mean_173_cast_fp16 = sub(x = inputs_173_cast_fp16, y = channels_mean_173_cast_fp16)[name = tensor("zero_mean_173_cast_fp16")]; + tensor zero_mean_sq_173_cast_fp16 = mul(x = zero_mean_173_cast_fp16, y = zero_mean_173_cast_fp16)[name = tensor("zero_mean_sq_173_cast_fp16")]; + tensor var_11577 = const()[name = tensor("op_11577"), val = tensor([1])]; + tensor var_11578_cast_fp16 = reduce_mean(axes = var_11577, keep_dims = var_11264, x = zero_mean_sq_173_cast_fp16)[name = tensor("op_11578_cast_fp16")]; + tensor var_11579_to_fp16 = const()[name = tensor("op_11579_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_11580_cast_fp16 = add(x = var_11578_cast_fp16, y = var_11579_to_fp16)[name = tensor("op_11580_cast_fp16")]; + tensor denom_173_epsilon_0 = const()[name = tensor("denom_173_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_173_cast_fp16 = rsqrt(epsilon = denom_173_epsilon_0, x = var_11580_cast_fp16)[name = tensor("denom_173_cast_fp16")]; + tensor out_173_cast_fp16 = mul(x = zero_mean_173_cast_fp16, y = denom_173_cast_fp16)[name = tensor("out_173_cast_fp16")]; + tensor input_861_gamma_0_to_fp16 = const()[name = tensor("input_861_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542605440)))]; + tensor input_861_beta_0_to_fp16 = const()[name = tensor("input_861_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542608064)))]; + tensor input_861_epsilon_0_to_fp16 = const()[name = tensor("input_861_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_861_cast_fp16 = batch_norm(beta = input_861_beta_0_to_fp16, epsilon = input_861_epsilon_0_to_fp16, gamma = input_861_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_173_cast_fp16)[name = tensor("input_861_cast_fp16")]; + tensor var_11594 = const()[name = tensor("op_11594"), val = tensor([1, 1])]; + tensor var_11596 = const()[name = tensor("op_11596"), val = tensor([1, 1])]; + tensor pretrained_out_577_pad_type_0 = const()[name = tensor("pretrained_out_577_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_577_pad_0 = const()[name = tensor("pretrained_out_577_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(542610688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545887552))), name = tensor("layers_28_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_28_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545887680)))]; + tensor pretrained_out_577_cast_fp16 = conv(bias = layers_28_fc1_pretrained_bias_to_fp16, dilations = var_11596, groups = var_11263, pad = pretrained_out_577_pad_0, pad_type = pretrained_out_577_pad_type_0, strides = var_11594, weight = layers_28_fc1_pretrained_weight_to_fp16_palettized, x = input_861_cast_fp16)[name = tensor("pretrained_out_577_cast_fp16")]; + tensor var_11600 = const()[name = tensor("op_11600"), val = tensor([1, 1])]; + tensor var_11602 = const()[name = tensor("op_11602"), val = tensor([1, 1])]; + tensor input_863_pad_type_0 = const()[name = tensor("input_863_pad_type_0"), val = tensor("custom")]; + tensor input_863_pad_0 = const()[name = tensor("input_863_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_28_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545897984)))]; + tensor input_863_cast_fp16 = conv(dilations = var_11602, groups = var_11263, pad = input_863_pad_0, pad_type = input_863_pad_type_0, strides = var_11600, weight = layers_28_fc1_loraA_weight_to_fp16, x = input_861_cast_fp16)[name = tensor("input_863_cast_fp16")]; + tensor var_11606 = const()[name = tensor("op_11606"), val = tensor([1, 1])]; + tensor var_11608 = const()[name = tensor("op_11608"), val = tensor([1, 1])]; + tensor lora_out_1153_pad_type_0 = const()[name = tensor("lora_out_1153_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1153_pad_0 = const()[name = tensor("lora_out_1153_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1155_weight_0_to_fp16 = const()[name = tensor("lora_out_1155_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545939008)))]; + tensor lora_out_1155_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_11608, groups = var_11263, pad = lora_out_1153_pad_0, pad_type = lora_out_1153_pad_type_0, strides = var_11606, weight = lora_out_1155_weight_0_to_fp16, x = input_863_cast_fp16)[name = tensor("lora_out_1155_cast_fp16")]; + tensor input_865_cast_fp16 = add(x = pretrained_out_577_cast_fp16, y = lora_out_1155_cast_fp16)[name = tensor("input_865_cast_fp16")]; + tensor input_867_mode_0 = const()[name = tensor("input_867_mode_0"), val = tensor("EXACT")]; + tensor input_867_cast_fp16 = gelu(mode = input_867_mode_0, x = input_865_cast_fp16)[name = tensor("input_867_cast_fp16")]; + tensor var_11620 = const()[name = tensor("op_11620"), val = tensor([1, 1])]; + tensor var_11622 = const()[name = tensor("op_11622"), val = tensor([1, 1])]; + tensor pretrained_out_579_pad_type_0 = const()[name = tensor("pretrained_out_579_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_579_pad_0 = const()[name = tensor("pretrained_out_579_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546102912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549379776))), name = tensor("layers_28_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_28_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_28_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549379904)))]; + tensor pretrained_out_579_cast_fp16 = conv(bias = layers_28_fc2_pretrained_bias_to_fp16, dilations = var_11622, groups = var_11263, pad = pretrained_out_579_pad_0, pad_type = pretrained_out_579_pad_type_0, strides = var_11620, weight = layers_28_fc2_pretrained_weight_to_fp16_palettized, x = input_867_cast_fp16)[name = tensor("pretrained_out_579_cast_fp16")]; + tensor var_11626 = const()[name = tensor("op_11626"), val = tensor([1, 1])]; + tensor var_11628 = const()[name = tensor("op_11628"), val = tensor([1, 1])]; + tensor input_869_pad_type_0 = const()[name = tensor("input_869_pad_type_0"), val = tensor("custom")]; + tensor input_869_pad_0 = const()[name = tensor("input_869_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_28_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_28_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549382528)))]; + tensor input_869_cast_fp16 = conv(dilations = var_11628, groups = var_11263, pad = input_869_pad_0, pad_type = input_869_pad_type_0, strides = var_11626, weight = layers_28_fc2_loraA_weight_to_fp16, x = input_867_cast_fp16)[name = tensor("input_869_cast_fp16")]; + tensor var_11632 = const()[name = tensor("op_11632"), val = tensor([1, 1])]; + tensor var_11634 = const()[name = tensor("op_11634"), val = tensor([1, 1])]; + tensor lora_out_1157_pad_type_0 = const()[name = tensor("lora_out_1157_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1157_pad_0 = const()[name = tensor("lora_out_1157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1159_weight_0_to_fp16 = const()[name = tensor("lora_out_1159_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549546432)))]; + tensor lora_out_1159_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11634, groups = var_11263, pad = lora_out_1157_pad_0, pad_type = lora_out_1157_pad_type_0, strides = var_11632, weight = lora_out_1159_weight_0_to_fp16, x = input_869_cast_fp16)[name = tensor("lora_out_1159_cast_fp16")]; + tensor hidden_states_59_cast_fp16 = add(x = pretrained_out_579_cast_fp16, y = lora_out_1159_cast_fp16)[name = tensor("hidden_states_59_cast_fp16")]; + tensor inputs_175_cast_fp16 = add(x = inputs_173_cast_fp16, y = hidden_states_59_cast_fp16)[name = tensor("inputs_175_cast_fp16")]; + tensor var_11650 = const()[name = tensor("op_11650"), val = tensor(3)]; + tensor var_11657 = const()[name = tensor("op_11657"), val = tensor(1)]; + tensor var_11658 = const()[name = tensor("op_11658"), val = tensor(true)]; + tensor var_11670 = const()[name = tensor("op_11670"), val = tensor([1])]; + tensor channels_mean_175_cast_fp16 = reduce_mean(axes = var_11670, keep_dims = var_11658, x = inputs_175_cast_fp16)[name = tensor("channels_mean_175_cast_fp16")]; + tensor zero_mean_175_cast_fp16 = sub(x = inputs_175_cast_fp16, y = channels_mean_175_cast_fp16)[name = tensor("zero_mean_175_cast_fp16")]; + tensor zero_mean_sq_175_cast_fp16 = mul(x = zero_mean_175_cast_fp16, y = zero_mean_175_cast_fp16)[name = tensor("zero_mean_sq_175_cast_fp16")]; + tensor var_11674 = const()[name = tensor("op_11674"), val = tensor([1])]; + tensor var_11675_cast_fp16 = reduce_mean(axes = var_11674, keep_dims = var_11658, x = zero_mean_sq_175_cast_fp16)[name = tensor("op_11675_cast_fp16")]; + tensor var_11676_to_fp16 = const()[name = tensor("op_11676_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_11677_cast_fp16 = add(x = var_11675_cast_fp16, y = var_11676_to_fp16)[name = tensor("op_11677_cast_fp16")]; + tensor denom_175_epsilon_0 = const()[name = tensor("denom_175_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_175_cast_fp16 = rsqrt(epsilon = denom_175_epsilon_0, x = var_11677_cast_fp16)[name = tensor("denom_175_cast_fp16")]; + tensor out_175_cast_fp16 = mul(x = zero_mean_175_cast_fp16, y = denom_175_cast_fp16)[name = tensor("out_175_cast_fp16")]; + tensor obj_407_gamma_0_to_fp16 = const()[name = tensor("obj_407_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549587456)))]; + tensor obj_407_beta_0_to_fp16 = const()[name = tensor("obj_407_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549590080)))]; + tensor obj_407_epsilon_0_to_fp16 = const()[name = tensor("obj_407_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_407_cast_fp16 = batch_norm(beta = obj_407_beta_0_to_fp16, epsilon = obj_407_epsilon_0_to_fp16, gamma = obj_407_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_175_cast_fp16)[name = tensor("obj_407_cast_fp16")]; + tensor var_11695 = const()[name = tensor("op_11695"), val = tensor([1, 1])]; + tensor var_11697 = const()[name = tensor("op_11697"), val = tensor([1, 1])]; + tensor pretrained_out_581_pad_type_0 = const()[name = tensor("pretrained_out_581_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_581_pad_0 = const()[name = tensor("pretrained_out_581_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(549592704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550411968))), name = tensor("layers_29_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_29_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550412096)))]; + tensor pretrained_out_581_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_11697, groups = var_11657, pad = pretrained_out_581_pad_0, pad_type = pretrained_out_581_pad_type_0, strides = var_11695, weight = layers_29_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_407_cast_fp16)[name = tensor("pretrained_out_581_cast_fp16")]; + tensor var_11701 = const()[name = tensor("op_11701"), val = tensor([1, 1])]; + tensor var_11703 = const()[name = tensor("op_11703"), val = tensor([1, 1])]; + tensor input_871_pad_type_0 = const()[name = tensor("input_871_pad_type_0"), val = tensor("custom")]; + tensor input_871_pad_0 = const()[name = tensor("input_871_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550414720)))]; + tensor input_871_cast_fp16 = conv(dilations = var_11703, groups = var_11657, pad = input_871_pad_0, pad_type = input_871_pad_type_0, strides = var_11701, weight = layers_29_self_attn_q_proj_loraA_weight_to_fp16, x = obj_407_cast_fp16)[name = tensor("input_871_cast_fp16")]; + tensor var_11707 = const()[name = tensor("op_11707"), val = tensor([1, 1])]; + tensor var_11709 = const()[name = tensor("op_11709"), val = tensor([1, 1])]; + tensor lora_out_1161_pad_type_0 = const()[name = tensor("lora_out_1161_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1161_pad_0 = const()[name = tensor("lora_out_1161_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1163_weight_0_to_fp16 = const()[name = tensor("lora_out_1163_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550455744)))]; + tensor lora_out_1163_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11709, groups = var_11657, pad = lora_out_1161_pad_0, pad_type = lora_out_1161_pad_type_0, strides = var_11707, weight = lora_out_1163_weight_0_to_fp16, x = input_871_cast_fp16)[name = tensor("lora_out_1163_cast_fp16")]; + tensor query_117_cast_fp16 = add(x = pretrained_out_581_cast_fp16, y = lora_out_1163_cast_fp16)[name = tensor("query_117_cast_fp16")]; + tensor var_11719 = const()[name = tensor("op_11719"), val = tensor([1, 1])]; + tensor var_11721 = const()[name = tensor("op_11721"), val = tensor([1, 1])]; + tensor pretrained_out_583_pad_type_0 = const()[name = tensor("pretrained_out_583_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_583_pad_0 = const()[name = tensor("pretrained_out_583_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550496768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(551316032))), name = tensor("layers_29_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_583_cast_fp16 = conv(dilations = var_11721, groups = var_11657, pad = pretrained_out_583_pad_0, pad_type = pretrained_out_583_pad_type_0, strides = var_11719, weight = layers_29_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_407_cast_fp16)[name = tensor("pretrained_out_583_cast_fp16")]; + tensor var_11725 = const()[name = tensor("op_11725"), val = tensor([1, 1])]; + tensor var_11727 = const()[name = tensor("op_11727"), val = tensor([1, 1])]; + tensor input_873_pad_type_0 = const()[name = tensor("input_873_pad_type_0"), val = tensor("custom")]; + tensor input_873_pad_0 = const()[name = tensor("input_873_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(551316160)))]; + tensor input_873_cast_fp16 = conv(dilations = var_11727, groups = var_11657, pad = input_873_pad_0, pad_type = input_873_pad_type_0, strides = var_11725, weight = layers_29_self_attn_k_proj_loraA_weight_to_fp16, x = obj_407_cast_fp16)[name = tensor("input_873_cast_fp16")]; + tensor var_11731 = const()[name = tensor("op_11731"), val = tensor([1, 1])]; + tensor var_11733 = const()[name = tensor("op_11733"), val = tensor([1, 1])]; + tensor lora_out_1165_pad_type_0 = const()[name = tensor("lora_out_1165_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1165_pad_0 = const()[name = tensor("lora_out_1165_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1167_weight_0_to_fp16 = const()[name = tensor("lora_out_1167_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(551357184)))]; + tensor lora_out_1167_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11733, groups = var_11657, pad = lora_out_1165_pad_0, pad_type = lora_out_1165_pad_type_0, strides = var_11731, weight = lora_out_1167_weight_0_to_fp16, x = input_873_cast_fp16)[name = tensor("lora_out_1167_cast_fp16")]; + tensor current_key_59_cast_fp16 = add(x = pretrained_out_583_cast_fp16, y = lora_out_1167_cast_fp16)[name = tensor("current_key_59_cast_fp16")]; + tensor var_11744 = const()[name = tensor("op_11744"), val = tensor([1, 1])]; + tensor var_11746 = const()[name = tensor("op_11746"), val = tensor([1, 1])]; + tensor pretrained_out_585_pad_type_0 = const()[name = tensor("pretrained_out_585_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_585_pad_0 = const()[name = tensor("pretrained_out_585_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(551398208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552217472))), name = tensor("layers_29_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_29_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552217600)))]; + tensor pretrained_out_585_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_11746, groups = var_11657, pad = pretrained_out_585_pad_0, pad_type = pretrained_out_585_pad_type_0, strides = var_11744, weight = layers_29_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_407_cast_fp16)[name = tensor("pretrained_out_585_cast_fp16")]; + tensor var_11750 = const()[name = tensor("op_11750"), val = tensor([1, 1])]; + tensor var_11752 = const()[name = tensor("op_11752"), val = tensor([1, 1])]; + tensor input_875_pad_type_0 = const()[name = tensor("input_875_pad_type_0"), val = tensor("custom")]; + tensor input_875_pad_0 = const()[name = tensor("input_875_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552220224)))]; + tensor input_875_cast_fp16 = conv(dilations = var_11752, groups = var_11657, pad = input_875_pad_0, pad_type = input_875_pad_type_0, strides = var_11750, weight = layers_29_self_attn_v_proj_loraA_weight_to_fp16, x = obj_407_cast_fp16)[name = tensor("input_875_cast_fp16")]; + tensor var_11756 = const()[name = tensor("op_11756"), val = tensor([1, 1])]; + tensor var_11758 = const()[name = tensor("op_11758"), val = tensor([1, 1])]; + tensor lora_out_1169_pad_type_0 = const()[name = tensor("lora_out_1169_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1169_pad_0 = const()[name = tensor("lora_out_1169_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1171_weight_0_to_fp16 = const()[name = tensor("lora_out_1171_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552261248)))]; + tensor lora_out_1171_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11758, groups = var_11657, pad = lora_out_1169_pad_0, pad_type = lora_out_1169_pad_type_0, strides = var_11756, weight = lora_out_1171_weight_0_to_fp16, x = input_875_cast_fp16)[name = tensor("lora_out_1171_cast_fp16")]; + tensor current_value_59_cast_fp16 = add(x = pretrained_out_585_cast_fp16, y = lora_out_1171_cast_fp16)[name = tensor("current_value_59_cast_fp16")]; + tensor var_11768_cast_fp16 = mul(x = current_key_59_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_11768_cast_fp16")]; + tensor var_11770_cast_fp16 = mul(x = var_103_cast_fp16_29, y = var_295_cast_fp16)[name = tensor("op_11770_cast_fp16")]; + tensor key_117_cast_fp16 = add(x = var_11768_cast_fp16, y = var_11770_cast_fp16)[name = tensor("key_117_cast_fp16")]; + tensor var_11772_cast_fp16 = mul(x = current_value_59_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_11772_cast_fp16")]; + tensor var_11774_cast_fp16 = mul(x = var_138_cast_fp16_29, y = var_295_cast_fp16)[name = tensor("op_11774_cast_fp16")]; + tensor value_117_cast_fp16 = add(x = var_11772_cast_fp16, y = var_11774_cast_fp16)[name = tensor("value_117_cast_fp16")]; + tensor var_11777 = const()[name = tensor("op_11777"), val = tensor([1, 20, 64, -1])]; + tensor var_11778_cast_fp16 = reshape(shape = var_11777, x = query_117_cast_fp16)[name = tensor("op_11778_cast_fp16")]; + tensor var_11779_to_fp16 = const()[name = tensor("op_11779_to_fp16"), val = tensor(0x1p-3)]; + tensor var_11780_cast_fp16 = mul(x = var_11778_cast_fp16, y = var_11779_to_fp16)[name = tensor("op_11780_cast_fp16")]; + tensor var_11781 = const()[name = tensor("op_11781"), val = tensor([1, 20, 64, -1])]; + tensor var_11782_cast_fp16 = reshape(shape = var_11781, x = key_117_cast_fp16)[name = tensor("op_11782_cast_fp16")]; + tensor mh_w_175_transpose_x_0 = const()[name = tensor("mh_w_175_transpose_x_0"), val = tensor(true)]; + tensor mh_w_175_transpose_y_0 = const()[name = tensor("mh_w_175_transpose_y_0"), val = tensor(false)]; + tensor mh_w_175_cast_fp16 = matmul(transpose_x = mh_w_175_transpose_x_0, transpose_y = mh_w_175_transpose_y_0, x = var_11780_cast_fp16, y = var_11782_cast_fp16)[name = tensor("mh_w_175_cast_fp16")]; + tensor mh_w_177_cast_fp16 = add(x = mh_w_175_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_177_cast_fp16")]; + tensor var_11790_cast_fp16 = softmax(axis = var_11650, x = mh_w_177_cast_fp16)[name = tensor("op_11790_cast_fp16")]; + tensor var_11791 = const()[name = tensor("op_11791"), val = tensor([1, 20, 64, -1])]; + tensor var_11792_cast_fp16 = reshape(shape = var_11791, x = value_117_cast_fp16)[name = tensor("op_11792_cast_fp16")]; + tensor attn_117_transpose_x_0 = const()[name = tensor("attn_117_transpose_x_0"), val = tensor(false)]; + tensor attn_117_transpose_y_0 = const()[name = tensor("attn_117_transpose_y_0"), val = tensor(true)]; + tensor attn_117_cast_fp16 = matmul(transpose_x = attn_117_transpose_x_0, transpose_y = attn_117_transpose_y_0, x = var_11792_cast_fp16, y = var_11790_cast_fp16)[name = tensor("attn_117_cast_fp16")]; + tensor var_11795 = const()[name = tensor("op_11795"), val = tensor([1, 1280, 1, -1])]; + tensor input_877_cast_fp16 = reshape(shape = var_11795, x = attn_117_cast_fp16)[name = tensor("input_877_cast_fp16")]; + tensor var_11802 = const()[name = tensor("op_11802"), val = tensor([1, 1])]; + tensor var_11804 = const()[name = tensor("op_11804"), val = tensor([1, 1])]; + tensor pretrained_out_587_pad_type_0 = const()[name = tensor("pretrained_out_587_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_587_pad_0 = const()[name = tensor("pretrained_out_587_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552302272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553121536))), name = tensor("layers_29_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_29_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553121664)))]; + tensor pretrained_out_587_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_11804, groups = var_11657, pad = pretrained_out_587_pad_0, pad_type = pretrained_out_587_pad_type_0, strides = var_11802, weight = layers_29_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_877_cast_fp16)[name = tensor("pretrained_out_587_cast_fp16")]; + tensor var_11808 = const()[name = tensor("op_11808"), val = tensor([1, 1])]; + tensor var_11810 = const()[name = tensor("op_11810"), val = tensor([1, 1])]; + tensor input_879_pad_type_0 = const()[name = tensor("input_879_pad_type_0"), val = tensor("custom")]; + tensor input_879_pad_0 = const()[name = tensor("input_879_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553124288)))]; + tensor input_879_cast_fp16 = conv(dilations = var_11810, groups = var_11657, pad = input_879_pad_0, pad_type = input_879_pad_type_0, strides = var_11808, weight = layers_29_self_attn_o_proj_loraA_weight_to_fp16, x = input_877_cast_fp16)[name = tensor("input_879_cast_fp16")]; + tensor var_11814 = const()[name = tensor("op_11814"), val = tensor([1, 1])]; + tensor var_11816 = const()[name = tensor("op_11816"), val = tensor([1, 1])]; + tensor lora_out_1173_pad_type_0 = const()[name = tensor("lora_out_1173_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1173_pad_0 = const()[name = tensor("lora_out_1173_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1175_weight_0_to_fp16 = const()[name = tensor("lora_out_1175_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553165312)))]; + tensor lora_out_1175_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11816, groups = var_11657, pad = lora_out_1173_pad_0, pad_type = lora_out_1173_pad_type_0, strides = var_11814, weight = lora_out_1175_weight_0_to_fp16, x = input_879_cast_fp16)[name = tensor("lora_out_1175_cast_fp16")]; + tensor obj_413_cast_fp16 = add(x = pretrained_out_587_cast_fp16, y = lora_out_1175_cast_fp16)[name = tensor("obj_413_cast_fp16")]; + tensor inputs_177_cast_fp16 = add(x = inputs_175_cast_fp16, y = obj_413_cast_fp16)[name = tensor("inputs_177_cast_fp16")]; + tensor var_11829 = const()[name = tensor("op_11829"), val = tensor([1])]; + tensor channels_mean_177_cast_fp16 = reduce_mean(axes = var_11829, keep_dims = var_11658, x = inputs_177_cast_fp16)[name = tensor("channels_mean_177_cast_fp16")]; + tensor zero_mean_177_cast_fp16 = sub(x = inputs_177_cast_fp16, y = channels_mean_177_cast_fp16)[name = tensor("zero_mean_177_cast_fp16")]; + tensor zero_mean_sq_177_cast_fp16 = mul(x = zero_mean_177_cast_fp16, y = zero_mean_177_cast_fp16)[name = tensor("zero_mean_sq_177_cast_fp16")]; + tensor var_11833 = const()[name = tensor("op_11833"), val = tensor([1])]; + tensor var_11834_cast_fp16 = reduce_mean(axes = var_11833, keep_dims = var_11658, x = zero_mean_sq_177_cast_fp16)[name = tensor("op_11834_cast_fp16")]; + tensor var_11835_to_fp16 = const()[name = tensor("op_11835_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_11836_cast_fp16 = add(x = var_11834_cast_fp16, y = var_11835_to_fp16)[name = tensor("op_11836_cast_fp16")]; + tensor denom_177_epsilon_0 = const()[name = tensor("denom_177_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_177_cast_fp16 = rsqrt(epsilon = denom_177_epsilon_0, x = var_11836_cast_fp16)[name = tensor("denom_177_cast_fp16")]; + tensor out_177_cast_fp16 = mul(x = zero_mean_177_cast_fp16, y = denom_177_cast_fp16)[name = tensor("out_177_cast_fp16")]; + tensor obj_415_gamma_0_to_fp16 = const()[name = tensor("obj_415_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553206336)))]; + tensor obj_415_beta_0_to_fp16 = const()[name = tensor("obj_415_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553208960)))]; + tensor obj_415_epsilon_0_to_fp16 = const()[name = tensor("obj_415_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_415_cast_fp16 = batch_norm(beta = obj_415_beta_0_to_fp16, epsilon = obj_415_epsilon_0_to_fp16, gamma = obj_415_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_177_cast_fp16)[name = tensor("obj_415_cast_fp16")]; + tensor var_11854 = const()[name = tensor("op_11854"), val = tensor([1, 1])]; + tensor var_11856 = const()[name = tensor("op_11856"), val = tensor([1, 1])]; + tensor pretrained_out_589_pad_type_0 = const()[name = tensor("pretrained_out_589_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_589_pad_0 = const()[name = tensor("pretrained_out_589_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553211584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554030848))), name = tensor("layers_29_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_29_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554030976)))]; + tensor pretrained_out_589_cast_fp16 = conv(bias = layers_29_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_11856, groups = var_11657, pad = pretrained_out_589_pad_0, pad_type = pretrained_out_589_pad_type_0, strides = var_11854, weight = layers_29_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_415_cast_fp16)[name = tensor("pretrained_out_589_cast_fp16")]; + tensor var_11860 = const()[name = tensor("op_11860"), val = tensor([1, 1])]; + tensor var_11862 = const()[name = tensor("op_11862"), val = tensor([1, 1])]; + tensor input_881_pad_type_0 = const()[name = tensor("input_881_pad_type_0"), val = tensor("custom")]; + tensor input_881_pad_0 = const()[name = tensor("input_881_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554033600)))]; + tensor input_881_cast_fp16 = conv(dilations = var_11862, groups = var_11657, pad = input_881_pad_0, pad_type = input_881_pad_type_0, strides = var_11860, weight = layers_29_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_415_cast_fp16)[name = tensor("input_881_cast_fp16")]; + tensor var_11866 = const()[name = tensor("op_11866"), val = tensor([1, 1])]; + tensor var_11868 = const()[name = tensor("op_11868"), val = tensor([1, 1])]; + tensor lora_out_1177_pad_type_0 = const()[name = tensor("lora_out_1177_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1177_pad_0 = const()[name = tensor("lora_out_1177_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1179_weight_0_to_fp16 = const()[name = tensor("lora_out_1179_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554074624)))]; + tensor lora_out_1179_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11868, groups = var_11657, pad = lora_out_1177_pad_0, pad_type = lora_out_1177_pad_type_0, strides = var_11866, weight = lora_out_1179_weight_0_to_fp16, x = input_881_cast_fp16)[name = tensor("lora_out_1179_cast_fp16")]; + tensor query_119_cast_fp16 = add(x = pretrained_out_589_cast_fp16, y = lora_out_1179_cast_fp16)[name = tensor("query_119_cast_fp16")]; + tensor var_11878 = const()[name = tensor("op_11878"), val = tensor([1, 1])]; + tensor var_11880 = const()[name = tensor("op_11880"), val = tensor([1, 1])]; + tensor pretrained_out_591_pad_type_0 = const()[name = tensor("pretrained_out_591_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_591_pad_0 = const()[name = tensor("pretrained_out_591_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554115648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554934912))), name = tensor("layers_29_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_591_cast_fp16 = conv(dilations = var_11880, groups = var_11657, pad = pretrained_out_591_pad_0, pad_type = pretrained_out_591_pad_type_0, strides = var_11878, weight = layers_29_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_591_cast_fp16")]; + tensor var_11884 = const()[name = tensor("op_11884"), val = tensor([1, 1])]; + tensor var_11886 = const()[name = tensor("op_11886"), val = tensor([1, 1])]; + tensor input_883_pad_type_0 = const()[name = tensor("input_883_pad_type_0"), val = tensor("custom")]; + tensor input_883_pad_0 = const()[name = tensor("input_883_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554935040)))]; + tensor input_883_cast_fp16 = conv(dilations = var_11886, groups = var_11657, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = var_11884, weight = layers_29_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_883_cast_fp16")]; + tensor var_11890 = const()[name = tensor("op_11890"), val = tensor([1, 1])]; + tensor var_11892 = const()[name = tensor("op_11892"), val = tensor([1, 1])]; + tensor lora_out_1181_pad_type_0 = const()[name = tensor("lora_out_1181_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1181_pad_0 = const()[name = tensor("lora_out_1181_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1183_weight_0_to_fp16 = const()[name = tensor("lora_out_1183_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554976064)))]; + tensor lora_out_1183_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11892, groups = var_11657, pad = lora_out_1181_pad_0, pad_type = lora_out_1181_pad_type_0, strides = var_11890, weight = lora_out_1183_weight_0_to_fp16, x = input_883_cast_fp16)[name = tensor("lora_out_1183_cast_fp16")]; + tensor key_119_cast_fp16 = add(x = pretrained_out_591_cast_fp16, y = lora_out_1183_cast_fp16)[name = tensor("key_119_cast_fp16")]; + tensor var_11903 = const()[name = tensor("op_11903"), val = tensor([1, 1])]; + tensor var_11905 = const()[name = tensor("op_11905"), val = tensor([1, 1])]; + tensor pretrained_out_593_pad_type_0 = const()[name = tensor("pretrained_out_593_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_593_pad_0 = const()[name = tensor("pretrained_out_593_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555017088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555836352))), name = tensor("layers_29_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_29_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555836480)))]; + tensor pretrained_out_593_cast_fp16 = conv(bias = layers_29_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_11905, groups = var_11657, pad = pretrained_out_593_pad_0, pad_type = pretrained_out_593_pad_type_0, strides = var_11903, weight = layers_29_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_593_cast_fp16")]; + tensor var_11909 = const()[name = tensor("op_11909"), val = tensor([1, 1])]; + tensor var_11911 = const()[name = tensor("op_11911"), val = tensor([1, 1])]; + tensor input_885_pad_type_0 = const()[name = tensor("input_885_pad_type_0"), val = tensor("custom")]; + tensor input_885_pad_0 = const()[name = tensor("input_885_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555839104)))]; + tensor input_885_cast_fp16 = conv(dilations = var_11911, groups = var_11657, pad = input_885_pad_0, pad_type = input_885_pad_type_0, strides = var_11909, weight = layers_29_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_885_cast_fp16")]; + tensor var_11915 = const()[name = tensor("op_11915"), val = tensor([1, 1])]; + tensor var_11917 = const()[name = tensor("op_11917"), val = tensor([1, 1])]; + tensor lora_out_1185_pad_type_0 = const()[name = tensor("lora_out_1185_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1185_pad_0 = const()[name = tensor("lora_out_1185_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1187_weight_0_to_fp16 = const()[name = tensor("lora_out_1187_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555880128)))]; + tensor lora_out_1187_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11917, groups = var_11657, pad = lora_out_1185_pad_0, pad_type = lora_out_1185_pad_type_0, strides = var_11915, weight = lora_out_1187_weight_0_to_fp16, x = input_885_cast_fp16)[name = tensor("lora_out_1187_cast_fp16")]; + tensor value_119_cast_fp16 = add(x = pretrained_out_593_cast_fp16, y = lora_out_1187_cast_fp16)[name = tensor("value_119_cast_fp16")]; + tensor var_11924 = const()[name = tensor("op_11924"), val = tensor([1, 20, 64, -1])]; + tensor var_11925_cast_fp16 = reshape(shape = var_11924, x = query_119_cast_fp16)[name = tensor("op_11925_cast_fp16")]; + tensor var_11926_to_fp16 = const()[name = tensor("op_11926_to_fp16"), val = tensor(0x1p-3)]; + tensor var_11927_cast_fp16 = mul(x = var_11925_cast_fp16, y = var_11926_to_fp16)[name = tensor("op_11927_cast_fp16")]; + tensor var_11928 = const()[name = tensor("op_11928"), val = tensor([1, 20, 64, -1])]; + tensor var_11929_cast_fp16 = reshape(shape = var_11928, x = key_119_cast_fp16)[name = tensor("op_11929_cast_fp16")]; + tensor mh_w_179_transpose_x_0 = const()[name = tensor("mh_w_179_transpose_x_0"), val = tensor(true)]; + tensor mh_w_179_transpose_y_0 = const()[name = tensor("mh_w_179_transpose_y_0"), val = tensor(false)]; + tensor mh_w_179_cast_fp16 = matmul(transpose_x = mh_w_179_transpose_x_0, transpose_y = mh_w_179_transpose_y_0, x = var_11927_cast_fp16, y = var_11929_cast_fp16)[name = tensor("mh_w_179_cast_fp16")]; + tensor obj_419_cast_fp16 = softmax(axis = var_11650, x = mh_w_179_cast_fp16)[name = tensor("obj_419_cast_fp16")]; + tensor var_11933 = const()[name = tensor("op_11933"), val = tensor([1, 20, 64, -1])]; + tensor var_11934_cast_fp16 = reshape(shape = var_11933, x = value_119_cast_fp16)[name = tensor("op_11934_cast_fp16")]; + tensor attn_119_transpose_x_0 = const()[name = tensor("attn_119_transpose_x_0"), val = tensor(false)]; + tensor attn_119_transpose_y_0 = const()[name = tensor("attn_119_transpose_y_0"), val = tensor(true)]; + tensor attn_119_cast_fp16 = matmul(transpose_x = attn_119_transpose_x_0, transpose_y = attn_119_transpose_y_0, x = var_11934_cast_fp16, y = obj_419_cast_fp16)[name = tensor("attn_119_cast_fp16")]; + tensor var_11937 = const()[name = tensor("op_11937"), val = tensor([1, 1280, 1, -1])]; + tensor input_887_cast_fp16 = reshape(shape = var_11937, x = attn_119_cast_fp16)[name = tensor("input_887_cast_fp16")]; + tensor var_11944 = const()[name = tensor("op_11944"), val = tensor([1, 1])]; + tensor var_11946 = const()[name = tensor("op_11946"), val = tensor([1, 1])]; + tensor pretrained_out_595_pad_type_0 = const()[name = tensor("pretrained_out_595_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_595_pad_0 = const()[name = tensor("pretrained_out_595_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555921152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556740416))), name = tensor("layers_29_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_29_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556740544)))]; + tensor pretrained_out_595_cast_fp16 = conv(bias = layers_29_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_11946, groups = var_11657, pad = pretrained_out_595_pad_0, pad_type = pretrained_out_595_pad_type_0, strides = var_11944, weight = layers_29_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_887_cast_fp16)[name = tensor("pretrained_out_595_cast_fp16")]; + tensor var_11950 = const()[name = tensor("op_11950"), val = tensor([1, 1])]; + tensor var_11952 = const()[name = tensor("op_11952"), val = tensor([1, 1])]; + tensor input_889_pad_type_0 = const()[name = tensor("input_889_pad_type_0"), val = tensor("custom")]; + tensor input_889_pad_0 = const()[name = tensor("input_889_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_29_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556743168)))]; + tensor input_889_cast_fp16 = conv(dilations = var_11952, groups = var_11657, pad = input_889_pad_0, pad_type = input_889_pad_type_0, strides = var_11950, weight = layers_29_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_887_cast_fp16)[name = tensor("input_889_cast_fp16")]; + tensor var_11956 = const()[name = tensor("op_11956"), val = tensor([1, 1])]; + tensor var_11958 = const()[name = tensor("op_11958"), val = tensor([1, 1])]; + tensor lora_out_1189_pad_type_0 = const()[name = tensor("lora_out_1189_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1189_pad_0 = const()[name = tensor("lora_out_1189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1191_weight_0_to_fp16 = const()[name = tensor("lora_out_1191_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556784192)))]; + tensor lora_out_1191_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_11958, groups = var_11657, pad = lora_out_1189_pad_0, pad_type = lora_out_1189_pad_type_0, strides = var_11956, weight = lora_out_1191_weight_0_to_fp16, x = input_889_cast_fp16)[name = tensor("lora_out_1191_cast_fp16")]; + tensor obj_417_cast_fp16 = add(x = pretrained_out_595_cast_fp16, y = lora_out_1191_cast_fp16)[name = tensor("obj_417_cast_fp16")]; + tensor inputs_179_cast_fp16 = add(x = inputs_177_cast_fp16, y = obj_417_cast_fp16)[name = tensor("inputs_179_cast_fp16")]; + tensor var_11967 = const()[name = tensor("op_11967"), val = tensor([1])]; + tensor channels_mean_179_cast_fp16 = reduce_mean(axes = var_11967, keep_dims = var_11658, x = inputs_179_cast_fp16)[name = tensor("channels_mean_179_cast_fp16")]; + tensor zero_mean_179_cast_fp16 = sub(x = inputs_179_cast_fp16, y = channels_mean_179_cast_fp16)[name = tensor("zero_mean_179_cast_fp16")]; + tensor zero_mean_sq_179_cast_fp16 = mul(x = zero_mean_179_cast_fp16, y = zero_mean_179_cast_fp16)[name = tensor("zero_mean_sq_179_cast_fp16")]; + tensor var_11971 = const()[name = tensor("op_11971"), val = tensor([1])]; + tensor var_11972_cast_fp16 = reduce_mean(axes = var_11971, keep_dims = var_11658, x = zero_mean_sq_179_cast_fp16)[name = tensor("op_11972_cast_fp16")]; + tensor var_11973_to_fp16 = const()[name = tensor("op_11973_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_11974_cast_fp16 = add(x = var_11972_cast_fp16, y = var_11973_to_fp16)[name = tensor("op_11974_cast_fp16")]; + tensor denom_179_epsilon_0 = const()[name = tensor("denom_179_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_179_cast_fp16 = rsqrt(epsilon = denom_179_epsilon_0, x = var_11974_cast_fp16)[name = tensor("denom_179_cast_fp16")]; + tensor out_179_cast_fp16 = mul(x = zero_mean_179_cast_fp16, y = denom_179_cast_fp16)[name = tensor("out_179_cast_fp16")]; + tensor input_891_gamma_0_to_fp16 = const()[name = tensor("input_891_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556825216)))]; + tensor input_891_beta_0_to_fp16 = const()[name = tensor("input_891_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556827840)))]; + tensor input_891_epsilon_0_to_fp16 = const()[name = tensor("input_891_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_891_cast_fp16 = batch_norm(beta = input_891_beta_0_to_fp16, epsilon = input_891_epsilon_0_to_fp16, gamma = input_891_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_179_cast_fp16)[name = tensor("input_891_cast_fp16")]; + tensor var_11988 = const()[name = tensor("op_11988"), val = tensor([1, 1])]; + tensor var_11990 = const()[name = tensor("op_11990"), val = tensor([1, 1])]; + tensor pretrained_out_597_pad_type_0 = const()[name = tensor("pretrained_out_597_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_597_pad_0 = const()[name = tensor("pretrained_out_597_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556830464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560107328))), name = tensor("layers_29_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_29_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560107456)))]; + tensor pretrained_out_597_cast_fp16 = conv(bias = layers_29_fc1_pretrained_bias_to_fp16, dilations = var_11990, groups = var_11657, pad = pretrained_out_597_pad_0, pad_type = pretrained_out_597_pad_type_0, strides = var_11988, weight = layers_29_fc1_pretrained_weight_to_fp16_palettized, x = input_891_cast_fp16)[name = tensor("pretrained_out_597_cast_fp16")]; + tensor var_11994 = const()[name = tensor("op_11994"), val = tensor([1, 1])]; + tensor var_11996 = const()[name = tensor("op_11996"), val = tensor([1, 1])]; + tensor input_893_pad_type_0 = const()[name = tensor("input_893_pad_type_0"), val = tensor("custom")]; + tensor input_893_pad_0 = const()[name = tensor("input_893_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_29_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560117760)))]; + tensor input_893_cast_fp16 = conv(dilations = var_11996, groups = var_11657, pad = input_893_pad_0, pad_type = input_893_pad_type_0, strides = var_11994, weight = layers_29_fc1_loraA_weight_to_fp16, x = input_891_cast_fp16)[name = tensor("input_893_cast_fp16")]; + tensor var_12000 = const()[name = tensor("op_12000"), val = tensor([1, 1])]; + tensor var_12002 = const()[name = tensor("op_12002"), val = tensor([1, 1])]; + tensor lora_out_1193_pad_type_0 = const()[name = tensor("lora_out_1193_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1193_pad_0 = const()[name = tensor("lora_out_1193_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1195_weight_0_to_fp16 = const()[name = tensor("lora_out_1195_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560158784)))]; + tensor lora_out_1195_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_12002, groups = var_11657, pad = lora_out_1193_pad_0, pad_type = lora_out_1193_pad_type_0, strides = var_12000, weight = lora_out_1195_weight_0_to_fp16, x = input_893_cast_fp16)[name = tensor("lora_out_1195_cast_fp16")]; + tensor input_895_cast_fp16 = add(x = pretrained_out_597_cast_fp16, y = lora_out_1195_cast_fp16)[name = tensor("input_895_cast_fp16")]; + tensor input_897_mode_0 = const()[name = tensor("input_897_mode_0"), val = tensor("EXACT")]; + tensor input_897_cast_fp16 = gelu(mode = input_897_mode_0, x = input_895_cast_fp16)[name = tensor("input_897_cast_fp16")]; + tensor var_12014 = const()[name = tensor("op_12014"), val = tensor([1, 1])]; + tensor var_12016 = const()[name = tensor("op_12016"), val = tensor([1, 1])]; + tensor pretrained_out_599_pad_type_0 = const()[name = tensor("pretrained_out_599_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_599_pad_0 = const()[name = tensor("pretrained_out_599_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560322688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563599552))), name = tensor("layers_29_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_29_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_29_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563599680)))]; + tensor pretrained_out_599_cast_fp16 = conv(bias = layers_29_fc2_pretrained_bias_to_fp16, dilations = var_12016, groups = var_11657, pad = pretrained_out_599_pad_0, pad_type = pretrained_out_599_pad_type_0, strides = var_12014, weight = layers_29_fc2_pretrained_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = tensor("pretrained_out_599_cast_fp16")]; + tensor var_12020 = const()[name = tensor("op_12020"), val = tensor([1, 1])]; + tensor var_12022 = const()[name = tensor("op_12022"), val = tensor([1, 1])]; + tensor input_899_pad_type_0 = const()[name = tensor("input_899_pad_type_0"), val = tensor("custom")]; + tensor input_899_pad_0 = const()[name = tensor("input_899_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_29_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_29_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563602304)))]; + tensor input_899_cast_fp16 = conv(dilations = var_12022, groups = var_11657, pad = input_899_pad_0, pad_type = input_899_pad_type_0, strides = var_12020, weight = layers_29_fc2_loraA_weight_to_fp16, x = input_897_cast_fp16)[name = tensor("input_899_cast_fp16")]; + tensor var_12026 = const()[name = tensor("op_12026"), val = tensor([1, 1])]; + tensor var_12028 = const()[name = tensor("op_12028"), val = tensor([1, 1])]; + tensor lora_out_1197_pad_type_0 = const()[name = tensor("lora_out_1197_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1197_pad_0 = const()[name = tensor("lora_out_1197_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1199_weight_0_to_fp16 = const()[name = tensor("lora_out_1199_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563766208)))]; + tensor lora_out_1199_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12028, groups = var_11657, pad = lora_out_1197_pad_0, pad_type = lora_out_1197_pad_type_0, strides = var_12026, weight = lora_out_1199_weight_0_to_fp16, x = input_899_cast_fp16)[name = tensor("lora_out_1199_cast_fp16")]; + tensor hidden_states_61_cast_fp16 = add(x = pretrained_out_599_cast_fp16, y = lora_out_1199_cast_fp16)[name = tensor("hidden_states_61_cast_fp16")]; + tensor inputs_181_cast_fp16 = add(x = inputs_179_cast_fp16, y = hidden_states_61_cast_fp16)[name = tensor("inputs_181_cast_fp16")]; + tensor var_12044 = const()[name = tensor("op_12044"), val = tensor(3)]; + tensor var_12051 = const()[name = tensor("op_12051"), val = tensor(1)]; + tensor var_12052 = const()[name = tensor("op_12052"), val = tensor(true)]; + tensor var_12064 = const()[name = tensor("op_12064"), val = tensor([1])]; + tensor channels_mean_181_cast_fp16 = reduce_mean(axes = var_12064, keep_dims = var_12052, x = inputs_181_cast_fp16)[name = tensor("channels_mean_181_cast_fp16")]; + tensor zero_mean_181_cast_fp16 = sub(x = inputs_181_cast_fp16, y = channels_mean_181_cast_fp16)[name = tensor("zero_mean_181_cast_fp16")]; + tensor zero_mean_sq_181_cast_fp16 = mul(x = zero_mean_181_cast_fp16, y = zero_mean_181_cast_fp16)[name = tensor("zero_mean_sq_181_cast_fp16")]; + tensor var_12068 = const()[name = tensor("op_12068"), val = tensor([1])]; + tensor var_12069_cast_fp16 = reduce_mean(axes = var_12068, keep_dims = var_12052, x = zero_mean_sq_181_cast_fp16)[name = tensor("op_12069_cast_fp16")]; + tensor var_12070_to_fp16 = const()[name = tensor("op_12070_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_12071_cast_fp16 = add(x = var_12069_cast_fp16, y = var_12070_to_fp16)[name = tensor("op_12071_cast_fp16")]; + tensor denom_181_epsilon_0 = const()[name = tensor("denom_181_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_181_cast_fp16 = rsqrt(epsilon = denom_181_epsilon_0, x = var_12071_cast_fp16)[name = tensor("denom_181_cast_fp16")]; + tensor out_181_cast_fp16 = mul(x = zero_mean_181_cast_fp16, y = denom_181_cast_fp16)[name = tensor("out_181_cast_fp16")]; + tensor obj_421_gamma_0_to_fp16 = const()[name = tensor("obj_421_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563807232)))]; + tensor obj_421_beta_0_to_fp16 = const()[name = tensor("obj_421_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563809856)))]; + tensor obj_421_epsilon_0_to_fp16 = const()[name = tensor("obj_421_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_421_cast_fp16 = batch_norm(beta = obj_421_beta_0_to_fp16, epsilon = obj_421_epsilon_0_to_fp16, gamma = obj_421_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_181_cast_fp16)[name = tensor("obj_421_cast_fp16")]; + tensor var_12089 = const()[name = tensor("op_12089"), val = tensor([1, 1])]; + tensor var_12091 = const()[name = tensor("op_12091"), val = tensor([1, 1])]; + tensor pretrained_out_601_pad_type_0 = const()[name = tensor("pretrained_out_601_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_601_pad_0 = const()[name = tensor("pretrained_out_601_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563812480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564631744))), name = tensor("layers_30_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_30_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564631872)))]; + tensor pretrained_out_601_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_12091, groups = var_12051, pad = pretrained_out_601_pad_0, pad_type = pretrained_out_601_pad_type_0, strides = var_12089, weight = layers_30_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_421_cast_fp16)[name = tensor("pretrained_out_601_cast_fp16")]; + tensor var_12095 = const()[name = tensor("op_12095"), val = tensor([1, 1])]; + tensor var_12097 = const()[name = tensor("op_12097"), val = tensor([1, 1])]; + tensor input_901_pad_type_0 = const()[name = tensor("input_901_pad_type_0"), val = tensor("custom")]; + tensor input_901_pad_0 = const()[name = tensor("input_901_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564634496)))]; + tensor input_901_cast_fp16 = conv(dilations = var_12097, groups = var_12051, pad = input_901_pad_0, pad_type = input_901_pad_type_0, strides = var_12095, weight = layers_30_self_attn_q_proj_loraA_weight_to_fp16, x = obj_421_cast_fp16)[name = tensor("input_901_cast_fp16")]; + tensor var_12101 = const()[name = tensor("op_12101"), val = tensor([1, 1])]; + tensor var_12103 = const()[name = tensor("op_12103"), val = tensor([1, 1])]; + tensor lora_out_1201_pad_type_0 = const()[name = tensor("lora_out_1201_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1201_pad_0 = const()[name = tensor("lora_out_1201_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1203_weight_0_to_fp16 = const()[name = tensor("lora_out_1203_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564675520)))]; + tensor lora_out_1203_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12103, groups = var_12051, pad = lora_out_1201_pad_0, pad_type = lora_out_1201_pad_type_0, strides = var_12101, weight = lora_out_1203_weight_0_to_fp16, x = input_901_cast_fp16)[name = tensor("lora_out_1203_cast_fp16")]; + tensor query_121_cast_fp16 = add(x = pretrained_out_601_cast_fp16, y = lora_out_1203_cast_fp16)[name = tensor("query_121_cast_fp16")]; + tensor var_12113 = const()[name = tensor("op_12113"), val = tensor([1, 1])]; + tensor var_12115 = const()[name = tensor("op_12115"), val = tensor([1, 1])]; + tensor pretrained_out_603_pad_type_0 = const()[name = tensor("pretrained_out_603_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_603_pad_0 = const()[name = tensor("pretrained_out_603_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(564716544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565535808))), name = tensor("layers_30_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_603_cast_fp16 = conv(dilations = var_12115, groups = var_12051, pad = pretrained_out_603_pad_0, pad_type = pretrained_out_603_pad_type_0, strides = var_12113, weight = layers_30_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_421_cast_fp16)[name = tensor("pretrained_out_603_cast_fp16")]; + tensor var_12119 = const()[name = tensor("op_12119"), val = tensor([1, 1])]; + tensor var_12121 = const()[name = tensor("op_12121"), val = tensor([1, 1])]; + tensor input_903_pad_type_0 = const()[name = tensor("input_903_pad_type_0"), val = tensor("custom")]; + tensor input_903_pad_0 = const()[name = tensor("input_903_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565535936)))]; + tensor input_903_cast_fp16 = conv(dilations = var_12121, groups = var_12051, pad = input_903_pad_0, pad_type = input_903_pad_type_0, strides = var_12119, weight = layers_30_self_attn_k_proj_loraA_weight_to_fp16, x = obj_421_cast_fp16)[name = tensor("input_903_cast_fp16")]; + tensor var_12125 = const()[name = tensor("op_12125"), val = tensor([1, 1])]; + tensor var_12127 = const()[name = tensor("op_12127"), val = tensor([1, 1])]; + tensor lora_out_1205_pad_type_0 = const()[name = tensor("lora_out_1205_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1205_pad_0 = const()[name = tensor("lora_out_1205_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1207_weight_0_to_fp16 = const()[name = tensor("lora_out_1207_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565576960)))]; + tensor lora_out_1207_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12127, groups = var_12051, pad = lora_out_1205_pad_0, pad_type = lora_out_1205_pad_type_0, strides = var_12125, weight = lora_out_1207_weight_0_to_fp16, x = input_903_cast_fp16)[name = tensor("lora_out_1207_cast_fp16")]; + tensor current_key_61_cast_fp16 = add(x = pretrained_out_603_cast_fp16, y = lora_out_1207_cast_fp16)[name = tensor("current_key_61_cast_fp16")]; + tensor var_12138 = const()[name = tensor("op_12138"), val = tensor([1, 1])]; + tensor var_12140 = const()[name = tensor("op_12140"), val = tensor([1, 1])]; + tensor pretrained_out_605_pad_type_0 = const()[name = tensor("pretrained_out_605_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_605_pad_0 = const()[name = tensor("pretrained_out_605_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(565617984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566437248))), name = tensor("layers_30_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_30_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566437376)))]; + tensor pretrained_out_605_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_12140, groups = var_12051, pad = pretrained_out_605_pad_0, pad_type = pretrained_out_605_pad_type_0, strides = var_12138, weight = layers_30_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_421_cast_fp16)[name = tensor("pretrained_out_605_cast_fp16")]; + tensor var_12144 = const()[name = tensor("op_12144"), val = tensor([1, 1])]; + tensor var_12146 = const()[name = tensor("op_12146"), val = tensor([1, 1])]; + tensor input_905_pad_type_0 = const()[name = tensor("input_905_pad_type_0"), val = tensor("custom")]; + tensor input_905_pad_0 = const()[name = tensor("input_905_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566440000)))]; + tensor input_905_cast_fp16 = conv(dilations = var_12146, groups = var_12051, pad = input_905_pad_0, pad_type = input_905_pad_type_0, strides = var_12144, weight = layers_30_self_attn_v_proj_loraA_weight_to_fp16, x = obj_421_cast_fp16)[name = tensor("input_905_cast_fp16")]; + tensor var_12150 = const()[name = tensor("op_12150"), val = tensor([1, 1])]; + tensor var_12152 = const()[name = tensor("op_12152"), val = tensor([1, 1])]; + tensor lora_out_1209_pad_type_0 = const()[name = tensor("lora_out_1209_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1209_pad_0 = const()[name = tensor("lora_out_1209_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1211_weight_0_to_fp16 = const()[name = tensor("lora_out_1211_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566481024)))]; + tensor lora_out_1211_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12152, groups = var_12051, pad = lora_out_1209_pad_0, pad_type = lora_out_1209_pad_type_0, strides = var_12150, weight = lora_out_1211_weight_0_to_fp16, x = input_905_cast_fp16)[name = tensor("lora_out_1211_cast_fp16")]; + tensor current_value_61_cast_fp16 = add(x = pretrained_out_605_cast_fp16, y = lora_out_1211_cast_fp16)[name = tensor("current_value_61_cast_fp16")]; + tensor var_12162_cast_fp16 = mul(x = current_key_61_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_12162_cast_fp16")]; + tensor var_12164_cast_fp16 = mul(x = var_103_cast_fp16_30, y = var_295_cast_fp16)[name = tensor("op_12164_cast_fp16")]; + tensor key_121_cast_fp16 = add(x = var_12162_cast_fp16, y = var_12164_cast_fp16)[name = tensor("key_121_cast_fp16")]; + tensor var_12166_cast_fp16 = mul(x = current_value_61_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_12166_cast_fp16")]; + tensor var_12168_cast_fp16 = mul(x = var_138_cast_fp16_30, y = var_295_cast_fp16)[name = tensor("op_12168_cast_fp16")]; + tensor value_121_cast_fp16 = add(x = var_12166_cast_fp16, y = var_12168_cast_fp16)[name = tensor("value_121_cast_fp16")]; + tensor var_12171 = const()[name = tensor("op_12171"), val = tensor([1, 20, 64, -1])]; + tensor var_12172_cast_fp16 = reshape(shape = var_12171, x = query_121_cast_fp16)[name = tensor("op_12172_cast_fp16")]; + tensor var_12173_to_fp16 = const()[name = tensor("op_12173_to_fp16"), val = tensor(0x1p-3)]; + tensor var_12174_cast_fp16 = mul(x = var_12172_cast_fp16, y = var_12173_to_fp16)[name = tensor("op_12174_cast_fp16")]; + tensor var_12175 = const()[name = tensor("op_12175"), val = tensor([1, 20, 64, -1])]; + tensor var_12176_cast_fp16 = reshape(shape = var_12175, x = key_121_cast_fp16)[name = tensor("op_12176_cast_fp16")]; + tensor mh_w_181_transpose_x_0 = const()[name = tensor("mh_w_181_transpose_x_0"), val = tensor(true)]; + tensor mh_w_181_transpose_y_0 = const()[name = tensor("mh_w_181_transpose_y_0"), val = tensor(false)]; + tensor mh_w_181_cast_fp16 = matmul(transpose_x = mh_w_181_transpose_x_0, transpose_y = mh_w_181_transpose_y_0, x = var_12174_cast_fp16, y = var_12176_cast_fp16)[name = tensor("mh_w_181_cast_fp16")]; + tensor mh_w_183_cast_fp16 = add(x = mh_w_181_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_183_cast_fp16")]; + tensor var_12184_cast_fp16 = softmax(axis = var_12044, x = mh_w_183_cast_fp16)[name = tensor("op_12184_cast_fp16")]; + tensor var_12185 = const()[name = tensor("op_12185"), val = tensor([1, 20, 64, -1])]; + tensor var_12186_cast_fp16 = reshape(shape = var_12185, x = value_121_cast_fp16)[name = tensor("op_12186_cast_fp16")]; + tensor attn_121_transpose_x_0 = const()[name = tensor("attn_121_transpose_x_0"), val = tensor(false)]; + tensor attn_121_transpose_y_0 = const()[name = tensor("attn_121_transpose_y_0"), val = tensor(true)]; + tensor attn_121_cast_fp16 = matmul(transpose_x = attn_121_transpose_x_0, transpose_y = attn_121_transpose_y_0, x = var_12186_cast_fp16, y = var_12184_cast_fp16)[name = tensor("attn_121_cast_fp16")]; + tensor var_12189 = const()[name = tensor("op_12189"), val = tensor([1, 1280, 1, -1])]; + tensor input_907_cast_fp16 = reshape(shape = var_12189, x = attn_121_cast_fp16)[name = tensor("input_907_cast_fp16")]; + tensor var_12196 = const()[name = tensor("op_12196"), val = tensor([1, 1])]; + tensor var_12198 = const()[name = tensor("op_12198"), val = tensor([1, 1])]; + tensor pretrained_out_607_pad_type_0 = const()[name = tensor("pretrained_out_607_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_607_pad_0 = const()[name = tensor("pretrained_out_607_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(566522048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567341312))), name = tensor("layers_30_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_30_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567341440)))]; + tensor pretrained_out_607_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_12198, groups = var_12051, pad = pretrained_out_607_pad_0, pad_type = pretrained_out_607_pad_type_0, strides = var_12196, weight = layers_30_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_907_cast_fp16)[name = tensor("pretrained_out_607_cast_fp16")]; + tensor var_12202 = const()[name = tensor("op_12202"), val = tensor([1, 1])]; + tensor var_12204 = const()[name = tensor("op_12204"), val = tensor([1, 1])]; + tensor input_909_pad_type_0 = const()[name = tensor("input_909_pad_type_0"), val = tensor("custom")]; + tensor input_909_pad_0 = const()[name = tensor("input_909_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567344064)))]; + tensor input_909_cast_fp16 = conv(dilations = var_12204, groups = var_12051, pad = input_909_pad_0, pad_type = input_909_pad_type_0, strides = var_12202, weight = layers_30_self_attn_o_proj_loraA_weight_to_fp16, x = input_907_cast_fp16)[name = tensor("input_909_cast_fp16")]; + tensor var_12208 = const()[name = tensor("op_12208"), val = tensor([1, 1])]; + tensor var_12210 = const()[name = tensor("op_12210"), val = tensor([1, 1])]; + tensor lora_out_1213_pad_type_0 = const()[name = tensor("lora_out_1213_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1213_pad_0 = const()[name = tensor("lora_out_1213_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1215_weight_0_to_fp16 = const()[name = tensor("lora_out_1215_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567385088)))]; + tensor lora_out_1215_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12210, groups = var_12051, pad = lora_out_1213_pad_0, pad_type = lora_out_1213_pad_type_0, strides = var_12208, weight = lora_out_1215_weight_0_to_fp16, x = input_909_cast_fp16)[name = tensor("lora_out_1215_cast_fp16")]; + tensor obj_427_cast_fp16 = add(x = pretrained_out_607_cast_fp16, y = lora_out_1215_cast_fp16)[name = tensor("obj_427_cast_fp16")]; + tensor inputs_183_cast_fp16 = add(x = inputs_181_cast_fp16, y = obj_427_cast_fp16)[name = tensor("inputs_183_cast_fp16")]; + tensor var_12223 = const()[name = tensor("op_12223"), val = tensor([1])]; + tensor channels_mean_183_cast_fp16 = reduce_mean(axes = var_12223, keep_dims = var_12052, x = inputs_183_cast_fp16)[name = tensor("channels_mean_183_cast_fp16")]; + tensor zero_mean_183_cast_fp16 = sub(x = inputs_183_cast_fp16, y = channels_mean_183_cast_fp16)[name = tensor("zero_mean_183_cast_fp16")]; + tensor zero_mean_sq_183_cast_fp16 = mul(x = zero_mean_183_cast_fp16, y = zero_mean_183_cast_fp16)[name = tensor("zero_mean_sq_183_cast_fp16")]; + tensor var_12227 = const()[name = tensor("op_12227"), val = tensor([1])]; + tensor var_12228_cast_fp16 = reduce_mean(axes = var_12227, keep_dims = var_12052, x = zero_mean_sq_183_cast_fp16)[name = tensor("op_12228_cast_fp16")]; + tensor var_12229_to_fp16 = const()[name = tensor("op_12229_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_12230_cast_fp16 = add(x = var_12228_cast_fp16, y = var_12229_to_fp16)[name = tensor("op_12230_cast_fp16")]; + tensor denom_183_epsilon_0 = const()[name = tensor("denom_183_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_183_cast_fp16 = rsqrt(epsilon = denom_183_epsilon_0, x = var_12230_cast_fp16)[name = tensor("denom_183_cast_fp16")]; + tensor out_183_cast_fp16 = mul(x = zero_mean_183_cast_fp16, y = denom_183_cast_fp16)[name = tensor("out_183_cast_fp16")]; + tensor obj_429_gamma_0_to_fp16 = const()[name = tensor("obj_429_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567426112)))]; + tensor obj_429_beta_0_to_fp16 = const()[name = tensor("obj_429_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567428736)))]; + tensor obj_429_epsilon_0_to_fp16 = const()[name = tensor("obj_429_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_429_cast_fp16 = batch_norm(beta = obj_429_beta_0_to_fp16, epsilon = obj_429_epsilon_0_to_fp16, gamma = obj_429_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_183_cast_fp16)[name = tensor("obj_429_cast_fp16")]; + tensor var_12248 = const()[name = tensor("op_12248"), val = tensor([1, 1])]; + tensor var_12250 = const()[name = tensor("op_12250"), val = tensor([1, 1])]; + tensor pretrained_out_609_pad_type_0 = const()[name = tensor("pretrained_out_609_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_609_pad_0 = const()[name = tensor("pretrained_out_609_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(567431360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568250624))), name = tensor("layers_30_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_30_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568250752)))]; + tensor pretrained_out_609_cast_fp16 = conv(bias = layers_30_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_12250, groups = var_12051, pad = pretrained_out_609_pad_0, pad_type = pretrained_out_609_pad_type_0, strides = var_12248, weight = layers_30_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_429_cast_fp16)[name = tensor("pretrained_out_609_cast_fp16")]; + tensor var_12254 = const()[name = tensor("op_12254"), val = tensor([1, 1])]; + tensor var_12256 = const()[name = tensor("op_12256"), val = tensor([1, 1])]; + tensor input_911_pad_type_0 = const()[name = tensor("input_911_pad_type_0"), val = tensor("custom")]; + tensor input_911_pad_0 = const()[name = tensor("input_911_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568253376)))]; + tensor input_911_cast_fp16 = conv(dilations = var_12256, groups = var_12051, pad = input_911_pad_0, pad_type = input_911_pad_type_0, strides = var_12254, weight = layers_30_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_429_cast_fp16)[name = tensor("input_911_cast_fp16")]; + tensor var_12260 = const()[name = tensor("op_12260"), val = tensor([1, 1])]; + tensor var_12262 = const()[name = tensor("op_12262"), val = tensor([1, 1])]; + tensor lora_out_1217_pad_type_0 = const()[name = tensor("lora_out_1217_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1217_pad_0 = const()[name = tensor("lora_out_1217_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1219_weight_0_to_fp16 = const()[name = tensor("lora_out_1219_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568294400)))]; + tensor lora_out_1219_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12262, groups = var_12051, pad = lora_out_1217_pad_0, pad_type = lora_out_1217_pad_type_0, strides = var_12260, weight = lora_out_1219_weight_0_to_fp16, x = input_911_cast_fp16)[name = tensor("lora_out_1219_cast_fp16")]; + tensor query_123_cast_fp16 = add(x = pretrained_out_609_cast_fp16, y = lora_out_1219_cast_fp16)[name = tensor("query_123_cast_fp16")]; + tensor var_12272 = const()[name = tensor("op_12272"), val = tensor([1, 1])]; + tensor var_12274 = const()[name = tensor("op_12274"), val = tensor([1, 1])]; + tensor pretrained_out_611_pad_type_0 = const()[name = tensor("pretrained_out_611_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_611_pad_0 = const()[name = tensor("pretrained_out_611_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(568335424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569154688))), name = tensor("layers_30_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_611_cast_fp16 = conv(dilations = var_12274, groups = var_12051, pad = pretrained_out_611_pad_0, pad_type = pretrained_out_611_pad_type_0, strides = var_12272, weight = layers_30_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_611_cast_fp16")]; + tensor var_12278 = const()[name = tensor("op_12278"), val = tensor([1, 1])]; + tensor var_12280 = const()[name = tensor("op_12280"), val = tensor([1, 1])]; + tensor input_913_pad_type_0 = const()[name = tensor("input_913_pad_type_0"), val = tensor("custom")]; + tensor input_913_pad_0 = const()[name = tensor("input_913_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569154816)))]; + tensor input_913_cast_fp16 = conv(dilations = var_12280, groups = var_12051, pad = input_913_pad_0, pad_type = input_913_pad_type_0, strides = var_12278, weight = layers_30_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_913_cast_fp16")]; + tensor var_12284 = const()[name = tensor("op_12284"), val = tensor([1, 1])]; + tensor var_12286 = const()[name = tensor("op_12286"), val = tensor([1, 1])]; + tensor lora_out_1221_pad_type_0 = const()[name = tensor("lora_out_1221_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1221_pad_0 = const()[name = tensor("lora_out_1221_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1223_weight_0_to_fp16 = const()[name = tensor("lora_out_1223_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569195840)))]; + tensor lora_out_1223_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12286, groups = var_12051, pad = lora_out_1221_pad_0, pad_type = lora_out_1221_pad_type_0, strides = var_12284, weight = lora_out_1223_weight_0_to_fp16, x = input_913_cast_fp16)[name = tensor("lora_out_1223_cast_fp16")]; + tensor key_123_cast_fp16 = add(x = pretrained_out_611_cast_fp16, y = lora_out_1223_cast_fp16)[name = tensor("key_123_cast_fp16")]; + tensor var_12297 = const()[name = tensor("op_12297"), val = tensor([1, 1])]; + tensor var_12299 = const()[name = tensor("op_12299"), val = tensor([1, 1])]; + tensor pretrained_out_613_pad_type_0 = const()[name = tensor("pretrained_out_613_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_613_pad_0 = const()[name = tensor("pretrained_out_613_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569236864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570056128))), name = tensor("layers_30_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_30_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570056256)))]; + tensor pretrained_out_613_cast_fp16 = conv(bias = layers_30_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_12299, groups = var_12051, pad = pretrained_out_613_pad_0, pad_type = pretrained_out_613_pad_type_0, strides = var_12297, weight = layers_30_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_613_cast_fp16")]; + tensor var_12303 = const()[name = tensor("op_12303"), val = tensor([1, 1])]; + tensor var_12305 = const()[name = tensor("op_12305"), val = tensor([1, 1])]; + tensor input_915_pad_type_0 = const()[name = tensor("input_915_pad_type_0"), val = tensor("custom")]; + tensor input_915_pad_0 = const()[name = tensor("input_915_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570058880)))]; + tensor input_915_cast_fp16 = conv(dilations = var_12305, groups = var_12051, pad = input_915_pad_0, pad_type = input_915_pad_type_0, strides = var_12303, weight = layers_30_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_915_cast_fp16")]; + tensor var_12309 = const()[name = tensor("op_12309"), val = tensor([1, 1])]; + tensor var_12311 = const()[name = tensor("op_12311"), val = tensor([1, 1])]; + tensor lora_out_1225_pad_type_0 = const()[name = tensor("lora_out_1225_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1225_pad_0 = const()[name = tensor("lora_out_1225_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1227_weight_0_to_fp16 = const()[name = tensor("lora_out_1227_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570099904)))]; + tensor lora_out_1227_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12311, groups = var_12051, pad = lora_out_1225_pad_0, pad_type = lora_out_1225_pad_type_0, strides = var_12309, weight = lora_out_1227_weight_0_to_fp16, x = input_915_cast_fp16)[name = tensor("lora_out_1227_cast_fp16")]; + tensor value_123_cast_fp16 = add(x = pretrained_out_613_cast_fp16, y = lora_out_1227_cast_fp16)[name = tensor("value_123_cast_fp16")]; + tensor var_12318 = const()[name = tensor("op_12318"), val = tensor([1, 20, 64, -1])]; + tensor var_12319_cast_fp16 = reshape(shape = var_12318, x = query_123_cast_fp16)[name = tensor("op_12319_cast_fp16")]; + tensor var_12320_to_fp16 = const()[name = tensor("op_12320_to_fp16"), val = tensor(0x1p-3)]; + tensor var_12321_cast_fp16 = mul(x = var_12319_cast_fp16, y = var_12320_to_fp16)[name = tensor("op_12321_cast_fp16")]; + tensor var_12322 = const()[name = tensor("op_12322"), val = tensor([1, 20, 64, -1])]; + tensor var_12323_cast_fp16 = reshape(shape = var_12322, x = key_123_cast_fp16)[name = tensor("op_12323_cast_fp16")]; + tensor mh_w_185_transpose_x_0 = const()[name = tensor("mh_w_185_transpose_x_0"), val = tensor(true)]; + tensor mh_w_185_transpose_y_0 = const()[name = tensor("mh_w_185_transpose_y_0"), val = tensor(false)]; + tensor mh_w_185_cast_fp16 = matmul(transpose_x = mh_w_185_transpose_x_0, transpose_y = mh_w_185_transpose_y_0, x = var_12321_cast_fp16, y = var_12323_cast_fp16)[name = tensor("mh_w_185_cast_fp16")]; + tensor obj_433_cast_fp16 = softmax(axis = var_12044, x = mh_w_185_cast_fp16)[name = tensor("obj_433_cast_fp16")]; + tensor var_12327 = const()[name = tensor("op_12327"), val = tensor([1, 20, 64, -1])]; + tensor var_12328_cast_fp16 = reshape(shape = var_12327, x = value_123_cast_fp16)[name = tensor("op_12328_cast_fp16")]; + tensor attn_123_transpose_x_0 = const()[name = tensor("attn_123_transpose_x_0"), val = tensor(false)]; + tensor attn_123_transpose_y_0 = const()[name = tensor("attn_123_transpose_y_0"), val = tensor(true)]; + tensor attn_123_cast_fp16 = matmul(transpose_x = attn_123_transpose_x_0, transpose_y = attn_123_transpose_y_0, x = var_12328_cast_fp16, y = obj_433_cast_fp16)[name = tensor("attn_123_cast_fp16")]; + tensor var_12331 = const()[name = tensor("op_12331"), val = tensor([1, 1280, 1, -1])]; + tensor input_917_cast_fp16 = reshape(shape = var_12331, x = attn_123_cast_fp16)[name = tensor("input_917_cast_fp16")]; + tensor var_12338 = const()[name = tensor("op_12338"), val = tensor([1, 1])]; + tensor var_12340 = const()[name = tensor("op_12340"), val = tensor([1, 1])]; + tensor pretrained_out_615_pad_type_0 = const()[name = tensor("pretrained_out_615_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_615_pad_0 = const()[name = tensor("pretrained_out_615_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570140928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570960192))), name = tensor("layers_30_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_30_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570960320)))]; + tensor pretrained_out_615_cast_fp16 = conv(bias = layers_30_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_12340, groups = var_12051, pad = pretrained_out_615_pad_0, pad_type = pretrained_out_615_pad_type_0, strides = var_12338, weight = layers_30_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_917_cast_fp16)[name = tensor("pretrained_out_615_cast_fp16")]; + tensor var_12344 = const()[name = tensor("op_12344"), val = tensor([1, 1])]; + tensor var_12346 = const()[name = tensor("op_12346"), val = tensor([1, 1])]; + tensor input_919_pad_type_0 = const()[name = tensor("input_919_pad_type_0"), val = tensor("custom")]; + tensor input_919_pad_0 = const()[name = tensor("input_919_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_30_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570962944)))]; + tensor input_919_cast_fp16 = conv(dilations = var_12346, groups = var_12051, pad = input_919_pad_0, pad_type = input_919_pad_type_0, strides = var_12344, weight = layers_30_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_917_cast_fp16)[name = tensor("input_919_cast_fp16")]; + tensor var_12350 = const()[name = tensor("op_12350"), val = tensor([1, 1])]; + tensor var_12352 = const()[name = tensor("op_12352"), val = tensor([1, 1])]; + tensor lora_out_1229_pad_type_0 = const()[name = tensor("lora_out_1229_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1229_pad_0 = const()[name = tensor("lora_out_1229_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1231_weight_0_to_fp16 = const()[name = tensor("lora_out_1231_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571003968)))]; + tensor lora_out_1231_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12352, groups = var_12051, pad = lora_out_1229_pad_0, pad_type = lora_out_1229_pad_type_0, strides = var_12350, weight = lora_out_1231_weight_0_to_fp16, x = input_919_cast_fp16)[name = tensor("lora_out_1231_cast_fp16")]; + tensor obj_431_cast_fp16 = add(x = pretrained_out_615_cast_fp16, y = lora_out_1231_cast_fp16)[name = tensor("obj_431_cast_fp16")]; + tensor inputs_185_cast_fp16 = add(x = inputs_183_cast_fp16, y = obj_431_cast_fp16)[name = tensor("inputs_185_cast_fp16")]; + tensor var_12361 = const()[name = tensor("op_12361"), val = tensor([1])]; + tensor channels_mean_185_cast_fp16 = reduce_mean(axes = var_12361, keep_dims = var_12052, x = inputs_185_cast_fp16)[name = tensor("channels_mean_185_cast_fp16")]; + tensor zero_mean_185_cast_fp16 = sub(x = inputs_185_cast_fp16, y = channels_mean_185_cast_fp16)[name = tensor("zero_mean_185_cast_fp16")]; + tensor zero_mean_sq_185_cast_fp16 = mul(x = zero_mean_185_cast_fp16, y = zero_mean_185_cast_fp16)[name = tensor("zero_mean_sq_185_cast_fp16")]; + tensor var_12365 = const()[name = tensor("op_12365"), val = tensor([1])]; + tensor var_12366_cast_fp16 = reduce_mean(axes = var_12365, keep_dims = var_12052, x = zero_mean_sq_185_cast_fp16)[name = tensor("op_12366_cast_fp16")]; + tensor var_12367_to_fp16 = const()[name = tensor("op_12367_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_12368_cast_fp16 = add(x = var_12366_cast_fp16, y = var_12367_to_fp16)[name = tensor("op_12368_cast_fp16")]; + tensor denom_185_epsilon_0 = const()[name = tensor("denom_185_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_185_cast_fp16 = rsqrt(epsilon = denom_185_epsilon_0, x = var_12368_cast_fp16)[name = tensor("denom_185_cast_fp16")]; + tensor out_185_cast_fp16 = mul(x = zero_mean_185_cast_fp16, y = denom_185_cast_fp16)[name = tensor("out_185_cast_fp16")]; + tensor input_921_gamma_0_to_fp16 = const()[name = tensor("input_921_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571044992)))]; + tensor input_921_beta_0_to_fp16 = const()[name = tensor("input_921_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571047616)))]; + tensor input_921_epsilon_0_to_fp16 = const()[name = tensor("input_921_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_921_cast_fp16 = batch_norm(beta = input_921_beta_0_to_fp16, epsilon = input_921_epsilon_0_to_fp16, gamma = input_921_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_185_cast_fp16)[name = tensor("input_921_cast_fp16")]; + tensor var_12382 = const()[name = tensor("op_12382"), val = tensor([1, 1])]; + tensor var_12384 = const()[name = tensor("op_12384"), val = tensor([1, 1])]; + tensor pretrained_out_617_pad_type_0 = const()[name = tensor("pretrained_out_617_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_617_pad_0 = const()[name = tensor("pretrained_out_617_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571050240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(574327104))), name = tensor("layers_30_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_30_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(574327232)))]; + tensor pretrained_out_617_cast_fp16 = conv(bias = layers_30_fc1_pretrained_bias_to_fp16, dilations = var_12384, groups = var_12051, pad = pretrained_out_617_pad_0, pad_type = pretrained_out_617_pad_type_0, strides = var_12382, weight = layers_30_fc1_pretrained_weight_to_fp16_palettized, x = input_921_cast_fp16)[name = tensor("pretrained_out_617_cast_fp16")]; + tensor var_12388 = const()[name = tensor("op_12388"), val = tensor([1, 1])]; + tensor var_12390 = const()[name = tensor("op_12390"), val = tensor([1, 1])]; + tensor input_923_pad_type_0 = const()[name = tensor("input_923_pad_type_0"), val = tensor("custom")]; + tensor input_923_pad_0 = const()[name = tensor("input_923_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_30_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(574337536)))]; + tensor input_923_cast_fp16 = conv(dilations = var_12390, groups = var_12051, pad = input_923_pad_0, pad_type = input_923_pad_type_0, strides = var_12388, weight = layers_30_fc1_loraA_weight_to_fp16, x = input_921_cast_fp16)[name = tensor("input_923_cast_fp16")]; + tensor var_12394 = const()[name = tensor("op_12394"), val = tensor([1, 1])]; + tensor var_12396 = const()[name = tensor("op_12396"), val = tensor([1, 1])]; + tensor lora_out_1233_pad_type_0 = const()[name = tensor("lora_out_1233_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1233_pad_0 = const()[name = tensor("lora_out_1233_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1235_weight_0_to_fp16 = const()[name = tensor("lora_out_1235_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(574378560)))]; + tensor lora_out_1235_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_12396, groups = var_12051, pad = lora_out_1233_pad_0, pad_type = lora_out_1233_pad_type_0, strides = var_12394, weight = lora_out_1235_weight_0_to_fp16, x = input_923_cast_fp16)[name = tensor("lora_out_1235_cast_fp16")]; + tensor input_925_cast_fp16 = add(x = pretrained_out_617_cast_fp16, y = lora_out_1235_cast_fp16)[name = tensor("input_925_cast_fp16")]; + tensor input_927_mode_0 = const()[name = tensor("input_927_mode_0"), val = tensor("EXACT")]; + tensor input_927_cast_fp16 = gelu(mode = input_927_mode_0, x = input_925_cast_fp16)[name = tensor("input_927_cast_fp16")]; + tensor var_12408 = const()[name = tensor("op_12408"), val = tensor([1, 1])]; + tensor var_12410 = const()[name = tensor("op_12410"), val = tensor([1, 1])]; + tensor pretrained_out_619_pad_type_0 = const()[name = tensor("pretrained_out_619_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_619_pad_0 = const()[name = tensor("pretrained_out_619_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(574542464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577819328))), name = tensor("layers_30_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_30_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_30_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577819456)))]; + tensor pretrained_out_619_cast_fp16 = conv(bias = layers_30_fc2_pretrained_bias_to_fp16, dilations = var_12410, groups = var_12051, pad = pretrained_out_619_pad_0, pad_type = pretrained_out_619_pad_type_0, strides = var_12408, weight = layers_30_fc2_pretrained_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = tensor("pretrained_out_619_cast_fp16")]; + tensor var_12414 = const()[name = tensor("op_12414"), val = tensor([1, 1])]; + tensor var_12416 = const()[name = tensor("op_12416"), val = tensor([1, 1])]; + tensor input_929_pad_type_0 = const()[name = tensor("input_929_pad_type_0"), val = tensor("custom")]; + tensor input_929_pad_0 = const()[name = tensor("input_929_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_30_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_30_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577822080)))]; + tensor input_929_cast_fp16 = conv(dilations = var_12416, groups = var_12051, pad = input_929_pad_0, pad_type = input_929_pad_type_0, strides = var_12414, weight = layers_30_fc2_loraA_weight_to_fp16, x = input_927_cast_fp16)[name = tensor("input_929_cast_fp16")]; + tensor var_12420 = const()[name = tensor("op_12420"), val = tensor([1, 1])]; + tensor var_12422 = const()[name = tensor("op_12422"), val = tensor([1, 1])]; + tensor lora_out_1237_pad_type_0 = const()[name = tensor("lora_out_1237_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1237_pad_0 = const()[name = tensor("lora_out_1237_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1239_weight_0_to_fp16 = const()[name = tensor("lora_out_1239_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577985984)))]; + tensor lora_out_1239_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12422, groups = var_12051, pad = lora_out_1237_pad_0, pad_type = lora_out_1237_pad_type_0, strides = var_12420, weight = lora_out_1239_weight_0_to_fp16, x = input_929_cast_fp16)[name = tensor("lora_out_1239_cast_fp16")]; + tensor hidden_states_63_cast_fp16 = add(x = pretrained_out_619_cast_fp16, y = lora_out_1239_cast_fp16)[name = tensor("hidden_states_63_cast_fp16")]; + tensor inputs_187_cast_fp16 = add(x = inputs_185_cast_fp16, y = hidden_states_63_cast_fp16)[name = tensor("inputs_187_cast_fp16")]; + tensor var_12438 = const()[name = tensor("op_12438"), val = tensor(3)]; + tensor var_12445 = const()[name = tensor("op_12445"), val = tensor(1)]; + tensor var_12446 = const()[name = tensor("op_12446"), val = tensor(true)]; + tensor var_12458 = const()[name = tensor("op_12458"), val = tensor([1])]; + tensor channels_mean_187_cast_fp16 = reduce_mean(axes = var_12458, keep_dims = var_12446, x = inputs_187_cast_fp16)[name = tensor("channels_mean_187_cast_fp16")]; + tensor zero_mean_187_cast_fp16 = sub(x = inputs_187_cast_fp16, y = channels_mean_187_cast_fp16)[name = tensor("zero_mean_187_cast_fp16")]; + tensor zero_mean_sq_187_cast_fp16 = mul(x = zero_mean_187_cast_fp16, y = zero_mean_187_cast_fp16)[name = tensor("zero_mean_sq_187_cast_fp16")]; + tensor var_12462 = const()[name = tensor("op_12462"), val = tensor([1])]; + tensor var_12463_cast_fp16 = reduce_mean(axes = var_12462, keep_dims = var_12446, x = zero_mean_sq_187_cast_fp16)[name = tensor("op_12463_cast_fp16")]; + tensor var_12464_to_fp16 = const()[name = tensor("op_12464_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_12465_cast_fp16 = add(x = var_12463_cast_fp16, y = var_12464_to_fp16)[name = tensor("op_12465_cast_fp16")]; + tensor denom_187_epsilon_0 = const()[name = tensor("denom_187_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_187_cast_fp16 = rsqrt(epsilon = denom_187_epsilon_0, x = var_12465_cast_fp16)[name = tensor("denom_187_cast_fp16")]; + tensor out_187_cast_fp16 = mul(x = zero_mean_187_cast_fp16, y = denom_187_cast_fp16)[name = tensor("out_187_cast_fp16")]; + tensor obj_435_gamma_0_to_fp16 = const()[name = tensor("obj_435_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578027008)))]; + tensor obj_435_beta_0_to_fp16 = const()[name = tensor("obj_435_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578029632)))]; + tensor obj_435_epsilon_0_to_fp16 = const()[name = tensor("obj_435_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_435_cast_fp16 = batch_norm(beta = obj_435_beta_0_to_fp16, epsilon = obj_435_epsilon_0_to_fp16, gamma = obj_435_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_187_cast_fp16)[name = tensor("obj_435_cast_fp16")]; + tensor var_12483 = const()[name = tensor("op_12483"), val = tensor([1, 1])]; + tensor var_12485 = const()[name = tensor("op_12485"), val = tensor([1, 1])]; + tensor pretrained_out_621_pad_type_0 = const()[name = tensor("pretrained_out_621_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_621_pad_0 = const()[name = tensor("pretrained_out_621_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578032256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578851520))), name = tensor("layers_31_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_31_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578851648)))]; + tensor pretrained_out_621_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_12485, groups = var_12445, pad = pretrained_out_621_pad_0, pad_type = pretrained_out_621_pad_type_0, strides = var_12483, weight = layers_31_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_435_cast_fp16)[name = tensor("pretrained_out_621_cast_fp16")]; + tensor var_12489 = const()[name = tensor("op_12489"), val = tensor([1, 1])]; + tensor var_12491 = const()[name = tensor("op_12491"), val = tensor([1, 1])]; + tensor input_931_pad_type_0 = const()[name = tensor("input_931_pad_type_0"), val = tensor("custom")]; + tensor input_931_pad_0 = const()[name = tensor("input_931_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578854272)))]; + tensor input_931_cast_fp16 = conv(dilations = var_12491, groups = var_12445, pad = input_931_pad_0, pad_type = input_931_pad_type_0, strides = var_12489, weight = layers_31_self_attn_q_proj_loraA_weight_to_fp16, x = obj_435_cast_fp16)[name = tensor("input_931_cast_fp16")]; + tensor var_12495 = const()[name = tensor("op_12495"), val = tensor([1, 1])]; + tensor var_12497 = const()[name = tensor("op_12497"), val = tensor([1, 1])]; + tensor lora_out_1241_pad_type_0 = const()[name = tensor("lora_out_1241_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1241_pad_0 = const()[name = tensor("lora_out_1241_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1243_weight_0_to_fp16 = const()[name = tensor("lora_out_1243_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578895296)))]; + tensor lora_out_1243_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12497, groups = var_12445, pad = lora_out_1241_pad_0, pad_type = lora_out_1241_pad_type_0, strides = var_12495, weight = lora_out_1243_weight_0_to_fp16, x = input_931_cast_fp16)[name = tensor("lora_out_1243_cast_fp16")]; + tensor query_125_cast_fp16 = add(x = pretrained_out_621_cast_fp16, y = lora_out_1243_cast_fp16)[name = tensor("query_125_cast_fp16")]; + tensor var_12507 = const()[name = tensor("op_12507"), val = tensor([1, 1])]; + tensor var_12509 = const()[name = tensor("op_12509"), val = tensor([1, 1])]; + tensor pretrained_out_623_pad_type_0 = const()[name = tensor("pretrained_out_623_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_623_pad_0 = const()[name = tensor("pretrained_out_623_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(578936320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579755584))), name = tensor("layers_31_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_623_cast_fp16 = conv(dilations = var_12509, groups = var_12445, pad = pretrained_out_623_pad_0, pad_type = pretrained_out_623_pad_type_0, strides = var_12507, weight = layers_31_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_435_cast_fp16)[name = tensor("pretrained_out_623_cast_fp16")]; + tensor var_12513 = const()[name = tensor("op_12513"), val = tensor([1, 1])]; + tensor var_12515 = const()[name = tensor("op_12515"), val = tensor([1, 1])]; + tensor input_933_pad_type_0 = const()[name = tensor("input_933_pad_type_0"), val = tensor("custom")]; + tensor input_933_pad_0 = const()[name = tensor("input_933_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579755712)))]; + tensor input_933_cast_fp16 = conv(dilations = var_12515, groups = var_12445, pad = input_933_pad_0, pad_type = input_933_pad_type_0, strides = var_12513, weight = layers_31_self_attn_k_proj_loraA_weight_to_fp16, x = obj_435_cast_fp16)[name = tensor("input_933_cast_fp16")]; + tensor var_12519 = const()[name = tensor("op_12519"), val = tensor([1, 1])]; + tensor var_12521 = const()[name = tensor("op_12521"), val = tensor([1, 1])]; + tensor lora_out_1245_pad_type_0 = const()[name = tensor("lora_out_1245_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1245_pad_0 = const()[name = tensor("lora_out_1245_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1247_weight_0_to_fp16 = const()[name = tensor("lora_out_1247_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579796736)))]; + tensor lora_out_1247_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12521, groups = var_12445, pad = lora_out_1245_pad_0, pad_type = lora_out_1245_pad_type_0, strides = var_12519, weight = lora_out_1247_weight_0_to_fp16, x = input_933_cast_fp16)[name = tensor("lora_out_1247_cast_fp16")]; + tensor current_key_cast_fp16 = add(x = pretrained_out_623_cast_fp16, y = lora_out_1247_cast_fp16)[name = tensor("current_key_cast_fp16")]; + tensor var_12532 = const()[name = tensor("op_12532"), val = tensor([1, 1])]; + tensor var_12534 = const()[name = tensor("op_12534"), val = tensor([1, 1])]; + tensor pretrained_out_625_pad_type_0 = const()[name = tensor("pretrained_out_625_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_625_pad_0 = const()[name = tensor("pretrained_out_625_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579837760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580657024))), name = tensor("layers_31_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_31_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580657152)))]; + tensor pretrained_out_625_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_12534, groups = var_12445, pad = pretrained_out_625_pad_0, pad_type = pretrained_out_625_pad_type_0, strides = var_12532, weight = layers_31_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_435_cast_fp16)[name = tensor("pretrained_out_625_cast_fp16")]; + tensor var_12538 = const()[name = tensor("op_12538"), val = tensor([1, 1])]; + tensor var_12540 = const()[name = tensor("op_12540"), val = tensor([1, 1])]; + tensor input_935_pad_type_0 = const()[name = tensor("input_935_pad_type_0"), val = tensor("custom")]; + tensor input_935_pad_0 = const()[name = tensor("input_935_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580659776)))]; + tensor input_935_cast_fp16 = conv(dilations = var_12540, groups = var_12445, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = var_12538, weight = layers_31_self_attn_v_proj_loraA_weight_to_fp16, x = obj_435_cast_fp16)[name = tensor("input_935_cast_fp16")]; + tensor var_12544 = const()[name = tensor("op_12544"), val = tensor([1, 1])]; + tensor var_12546 = const()[name = tensor("op_12546"), val = tensor([1, 1])]; + tensor lora_out_1249_pad_type_0 = const()[name = tensor("lora_out_1249_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1249_pad_0 = const()[name = tensor("lora_out_1249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1251_weight_0_to_fp16 = const()[name = tensor("lora_out_1251_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580700800)))]; + tensor lora_out_1251_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12546, groups = var_12445, pad = lora_out_1249_pad_0, pad_type = lora_out_1249_pad_type_0, strides = var_12544, weight = lora_out_1251_weight_0_to_fp16, x = input_935_cast_fp16)[name = tensor("lora_out_1251_cast_fp16")]; + tensor current_value_cast_fp16 = add(x = pretrained_out_625_cast_fp16, y = lora_out_1251_cast_fp16)[name = tensor("current_value_cast_fp16")]; + tensor var_12556_cast_fp16 = mul(x = current_key_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_12556_cast_fp16")]; + tensor var_12558_cast_fp16 = mul(x = var_103_cast_fp16_31, y = var_295_cast_fp16)[name = tensor("op_12558_cast_fp16")]; + tensor key_125_cast_fp16 = add(x = var_12556_cast_fp16, y = var_12558_cast_fp16)[name = tensor("key_125_cast_fp16")]; + tensor var_12560_cast_fp16 = mul(x = current_value_cast_fp16, y = var_292_cast_fp16)[name = tensor("op_12560_cast_fp16")]; + tensor var_12562_cast_fp16 = mul(x = var_138_cast_fp16_31, y = var_295_cast_fp16)[name = tensor("op_12562_cast_fp16")]; + tensor value_125_cast_fp16 = add(x = var_12560_cast_fp16, y = var_12562_cast_fp16)[name = tensor("value_125_cast_fp16")]; + tensor var_12565 = const()[name = tensor("op_12565"), val = tensor([1, 20, 64, -1])]; + tensor var_12566_cast_fp16 = reshape(shape = var_12565, x = query_125_cast_fp16)[name = tensor("op_12566_cast_fp16")]; + tensor var_12567_to_fp16 = const()[name = tensor("op_12567_to_fp16"), val = tensor(0x1p-3)]; + tensor var_12568_cast_fp16 = mul(x = var_12566_cast_fp16, y = var_12567_to_fp16)[name = tensor("op_12568_cast_fp16")]; + tensor var_12569 = const()[name = tensor("op_12569"), val = tensor([1, 20, 64, -1])]; + tensor var_12570_cast_fp16 = reshape(shape = var_12569, x = key_125_cast_fp16)[name = tensor("op_12570_cast_fp16")]; + tensor mh_w_187_transpose_x_0 = const()[name = tensor("mh_w_187_transpose_x_0"), val = tensor(true)]; + tensor mh_w_187_transpose_y_0 = const()[name = tensor("mh_w_187_transpose_y_0"), val = tensor(false)]; + tensor mh_w_187_cast_fp16 = matmul(transpose_x = mh_w_187_transpose_x_0, transpose_y = mh_w_187_transpose_y_0, x = var_12568_cast_fp16, y = var_12570_cast_fp16)[name = tensor("mh_w_187_cast_fp16")]; + tensor mh_w_189_cast_fp16 = add(x = mh_w_187_cast_fp16, y = var_313_cast_fp16)[name = tensor("mh_w_189_cast_fp16")]; + tensor var_12578_cast_fp16 = softmax(axis = var_12438, x = mh_w_189_cast_fp16)[name = tensor("op_12578_cast_fp16")]; + tensor var_12579 = const()[name = tensor("op_12579"), val = tensor([1, 20, 64, -1])]; + tensor var_12580_cast_fp16 = reshape(shape = var_12579, x = value_125_cast_fp16)[name = tensor("op_12580_cast_fp16")]; + tensor attn_125_transpose_x_0 = const()[name = tensor("attn_125_transpose_x_0"), val = tensor(false)]; + tensor attn_125_transpose_y_0 = const()[name = tensor("attn_125_transpose_y_0"), val = tensor(true)]; + tensor attn_125_cast_fp16 = matmul(transpose_x = attn_125_transpose_x_0, transpose_y = attn_125_transpose_y_0, x = var_12580_cast_fp16, y = var_12578_cast_fp16)[name = tensor("attn_125_cast_fp16")]; + tensor var_12583 = const()[name = tensor("op_12583"), val = tensor([1, 1280, 1, -1])]; + tensor input_937_cast_fp16 = reshape(shape = var_12583, x = attn_125_cast_fp16)[name = tensor("input_937_cast_fp16")]; + tensor var_12590 = const()[name = tensor("op_12590"), val = tensor([1, 1])]; + tensor var_12592 = const()[name = tensor("op_12592"), val = tensor([1, 1])]; + tensor pretrained_out_627_pad_type_0 = const()[name = tensor("pretrained_out_627_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_627_pad_0 = const()[name = tensor("pretrained_out_627_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580741824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581561088))), name = tensor("layers_31_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_31_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581561216)))]; + tensor pretrained_out_627_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_12592, groups = var_12445, pad = pretrained_out_627_pad_0, pad_type = pretrained_out_627_pad_type_0, strides = var_12590, weight = layers_31_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_937_cast_fp16)[name = tensor("pretrained_out_627_cast_fp16")]; + tensor var_12596 = const()[name = tensor("op_12596"), val = tensor([1, 1])]; + tensor var_12598 = const()[name = tensor("op_12598"), val = tensor([1, 1])]; + tensor input_939_pad_type_0 = const()[name = tensor("input_939_pad_type_0"), val = tensor("custom")]; + tensor input_939_pad_0 = const()[name = tensor("input_939_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581563840)))]; + tensor input_939_cast_fp16 = conv(dilations = var_12598, groups = var_12445, pad = input_939_pad_0, pad_type = input_939_pad_type_0, strides = var_12596, weight = layers_31_self_attn_o_proj_loraA_weight_to_fp16, x = input_937_cast_fp16)[name = tensor("input_939_cast_fp16")]; + tensor var_12602 = const()[name = tensor("op_12602"), val = tensor([1, 1])]; + tensor var_12604 = const()[name = tensor("op_12604"), val = tensor([1, 1])]; + tensor lora_out_1253_pad_type_0 = const()[name = tensor("lora_out_1253_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1253_pad_0 = const()[name = tensor("lora_out_1253_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1255_weight_0_to_fp16 = const()[name = tensor("lora_out_1255_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581604864)))]; + tensor lora_out_1255_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12604, groups = var_12445, pad = lora_out_1253_pad_0, pad_type = lora_out_1253_pad_type_0, strides = var_12602, weight = lora_out_1255_weight_0_to_fp16, x = input_939_cast_fp16)[name = tensor("lora_out_1255_cast_fp16")]; + tensor obj_441_cast_fp16 = add(x = pretrained_out_627_cast_fp16, y = lora_out_1255_cast_fp16)[name = tensor("obj_441_cast_fp16")]; + tensor inputs_189_cast_fp16 = add(x = inputs_187_cast_fp16, y = obj_441_cast_fp16)[name = tensor("inputs_189_cast_fp16")]; + tensor var_12617 = const()[name = tensor("op_12617"), val = tensor([1])]; + tensor channels_mean_189_cast_fp16 = reduce_mean(axes = var_12617, keep_dims = var_12446, x = inputs_189_cast_fp16)[name = tensor("channels_mean_189_cast_fp16")]; + tensor zero_mean_189_cast_fp16 = sub(x = inputs_189_cast_fp16, y = channels_mean_189_cast_fp16)[name = tensor("zero_mean_189_cast_fp16")]; + tensor zero_mean_sq_189_cast_fp16 = mul(x = zero_mean_189_cast_fp16, y = zero_mean_189_cast_fp16)[name = tensor("zero_mean_sq_189_cast_fp16")]; + tensor var_12621 = const()[name = tensor("op_12621"), val = tensor([1])]; + tensor var_12622_cast_fp16 = reduce_mean(axes = var_12621, keep_dims = var_12446, x = zero_mean_sq_189_cast_fp16)[name = tensor("op_12622_cast_fp16")]; + tensor var_12623_to_fp16 = const()[name = tensor("op_12623_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_12624_cast_fp16 = add(x = var_12622_cast_fp16, y = var_12623_to_fp16)[name = tensor("op_12624_cast_fp16")]; + tensor denom_189_epsilon_0 = const()[name = tensor("denom_189_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_189_cast_fp16 = rsqrt(epsilon = denom_189_epsilon_0, x = var_12624_cast_fp16)[name = tensor("denom_189_cast_fp16")]; + tensor out_189_cast_fp16 = mul(x = zero_mean_189_cast_fp16, y = denom_189_cast_fp16)[name = tensor("out_189_cast_fp16")]; + tensor obj_443_gamma_0_to_fp16 = const()[name = tensor("obj_443_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581645888)))]; + tensor obj_443_beta_0_to_fp16 = const()[name = tensor("obj_443_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581648512)))]; + tensor obj_443_epsilon_0_to_fp16 = const()[name = tensor("obj_443_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_443_cast_fp16 = batch_norm(beta = obj_443_beta_0_to_fp16, epsilon = obj_443_epsilon_0_to_fp16, gamma = obj_443_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_189_cast_fp16)[name = tensor("obj_443_cast_fp16")]; + tensor var_12642 = const()[name = tensor("op_12642"), val = tensor([1, 1])]; + tensor var_12644 = const()[name = tensor("op_12644"), val = tensor([1, 1])]; + tensor pretrained_out_629_pad_type_0 = const()[name = tensor("pretrained_out_629_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_629_pad_0 = const()[name = tensor("pretrained_out_629_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581651136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582470400))), name = tensor("layers_31_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_31_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582470528)))]; + tensor pretrained_out_629_cast_fp16 = conv(bias = layers_31_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = var_12644, groups = var_12445, pad = pretrained_out_629_pad_0, pad_type = pretrained_out_629_pad_type_0, strides = var_12642, weight = layers_31_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_443_cast_fp16)[name = tensor("pretrained_out_629_cast_fp16")]; + tensor var_12648 = const()[name = tensor("op_12648"), val = tensor([1, 1])]; + tensor var_12650 = const()[name = tensor("op_12650"), val = tensor([1, 1])]; + tensor input_941_pad_type_0 = const()[name = tensor("input_941_pad_type_0"), val = tensor("custom")]; + tensor input_941_pad_0 = const()[name = tensor("input_941_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582473152)))]; + tensor input_941_cast_fp16 = conv(dilations = var_12650, groups = var_12445, pad = input_941_pad_0, pad_type = input_941_pad_type_0, strides = var_12648, weight = layers_31_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_443_cast_fp16)[name = tensor("input_941_cast_fp16")]; + tensor var_12654 = const()[name = tensor("op_12654"), val = tensor([1, 1])]; + tensor var_12656 = const()[name = tensor("op_12656"), val = tensor([1, 1])]; + tensor lora_out_1257_pad_type_0 = const()[name = tensor("lora_out_1257_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1257_pad_0 = const()[name = tensor("lora_out_1257_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1259_weight_0_to_fp16 = const()[name = tensor("lora_out_1259_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582514176)))]; + tensor lora_out_1259_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12656, groups = var_12445, pad = lora_out_1257_pad_0, pad_type = lora_out_1257_pad_type_0, strides = var_12654, weight = lora_out_1259_weight_0_to_fp16, x = input_941_cast_fp16)[name = tensor("lora_out_1259_cast_fp16")]; + tensor query_cast_fp16 = add(x = pretrained_out_629_cast_fp16, y = lora_out_1259_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor var_12666 = const()[name = tensor("op_12666"), val = tensor([1, 1])]; + tensor var_12668 = const()[name = tensor("op_12668"), val = tensor([1, 1])]; + tensor pretrained_out_631_pad_type_0 = const()[name = tensor("pretrained_out_631_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_631_pad_0 = const()[name = tensor("pretrained_out_631_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582555200))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(583374464))), name = tensor("layers_31_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_631_cast_fp16 = conv(dilations = var_12668, groups = var_12445, pad = pretrained_out_631_pad_0, pad_type = pretrained_out_631_pad_type_0, strides = var_12666, weight = layers_31_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_631_cast_fp16")]; + tensor var_12672 = const()[name = tensor("op_12672"), val = tensor([1, 1])]; + tensor var_12674 = const()[name = tensor("op_12674"), val = tensor([1, 1])]; + tensor input_943_pad_type_0 = const()[name = tensor("input_943_pad_type_0"), val = tensor("custom")]; + tensor input_943_pad_0 = const()[name = tensor("input_943_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(583374592)))]; + tensor input_943_cast_fp16 = conv(dilations = var_12674, groups = var_12445, pad = input_943_pad_0, pad_type = input_943_pad_type_0, strides = var_12672, weight = layers_31_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_943_cast_fp16")]; + tensor var_12678 = const()[name = tensor("op_12678"), val = tensor([1, 1])]; + tensor var_12680 = const()[name = tensor("op_12680"), val = tensor([1, 1])]; + tensor lora_out_1261_pad_type_0 = const()[name = tensor("lora_out_1261_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1261_pad_0 = const()[name = tensor("lora_out_1261_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1263_weight_0_to_fp16 = const()[name = tensor("lora_out_1263_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(583415616)))]; + tensor lora_out_1263_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12680, groups = var_12445, pad = lora_out_1261_pad_0, pad_type = lora_out_1261_pad_type_0, strides = var_12678, weight = lora_out_1263_weight_0_to_fp16, x = input_943_cast_fp16)[name = tensor("lora_out_1263_cast_fp16")]; + tensor key_cast_fp16 = add(x = pretrained_out_631_cast_fp16, y = lora_out_1263_cast_fp16)[name = tensor("key_cast_fp16")]; + tensor var_12691 = const()[name = tensor("op_12691"), val = tensor([1, 1])]; + tensor var_12693 = const()[name = tensor("op_12693"), val = tensor([1, 1])]; + tensor pretrained_out_633_pad_type_0 = const()[name = tensor("pretrained_out_633_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_633_pad_0 = const()[name = tensor("pretrained_out_633_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(583456640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584275904))), name = tensor("layers_31_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_31_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584276032)))]; + tensor pretrained_out_633_cast_fp16 = conv(bias = layers_31_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = var_12693, groups = var_12445, pad = pretrained_out_633_pad_0, pad_type = pretrained_out_633_pad_type_0, strides = var_12691, weight = layers_31_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_633_cast_fp16")]; + tensor var_12697 = const()[name = tensor("op_12697"), val = tensor([1, 1])]; + tensor var_12699 = const()[name = tensor("op_12699"), val = tensor([1, 1])]; + tensor input_945_pad_type_0 = const()[name = tensor("input_945_pad_type_0"), val = tensor("custom")]; + tensor input_945_pad_0 = const()[name = tensor("input_945_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584278656)))]; + tensor input_945_cast_fp16 = conv(dilations = var_12699, groups = var_12445, pad = input_945_pad_0, pad_type = input_945_pad_type_0, strides = var_12697, weight = layers_31_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_945_cast_fp16")]; + tensor var_12703 = const()[name = tensor("op_12703"), val = tensor([1, 1])]; + tensor var_12705 = const()[name = tensor("op_12705"), val = tensor([1, 1])]; + tensor lora_out_1265_pad_type_0 = const()[name = tensor("lora_out_1265_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1265_pad_0 = const()[name = tensor("lora_out_1265_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1267_weight_0_to_fp16 = const()[name = tensor("lora_out_1267_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584319680)))]; + tensor lora_out_1267_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12705, groups = var_12445, pad = lora_out_1265_pad_0, pad_type = lora_out_1265_pad_type_0, strides = var_12703, weight = lora_out_1267_weight_0_to_fp16, x = input_945_cast_fp16)[name = tensor("lora_out_1267_cast_fp16")]; + tensor value_cast_fp16 = add(x = pretrained_out_633_cast_fp16, y = lora_out_1267_cast_fp16)[name = tensor("value_cast_fp16")]; + tensor var_12712 = const()[name = tensor("op_12712"), val = tensor([1, 20, 64, -1])]; + tensor var_12713_cast_fp16 = reshape(shape = var_12712, x = query_cast_fp16)[name = tensor("op_12713_cast_fp16")]; + tensor var_12714_to_fp16 = const()[name = tensor("op_12714_to_fp16"), val = tensor(0x1p-3)]; + tensor var_12715_cast_fp16 = mul(x = var_12713_cast_fp16, y = var_12714_to_fp16)[name = tensor("op_12715_cast_fp16")]; + tensor var_12716 = const()[name = tensor("op_12716"), val = tensor([1, 20, 64, -1])]; + tensor var_12717_cast_fp16 = reshape(shape = var_12716, x = key_cast_fp16)[name = tensor("op_12717_cast_fp16")]; + tensor mh_w_transpose_x_0 = const()[name = tensor("mh_w_transpose_x_0"), val = tensor(true)]; + tensor mh_w_transpose_y_0 = const()[name = tensor("mh_w_transpose_y_0"), val = tensor(false)]; + tensor mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_12715_cast_fp16, y = var_12717_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor obj_447_cast_fp16 = softmax(axis = var_12438, x = mh_w_cast_fp16)[name = tensor("obj_447_cast_fp16")]; + tensor var_12721 = const()[name = tensor("op_12721"), val = tensor([1, 20, 64, -1])]; + tensor var_12722_cast_fp16 = reshape(shape = var_12721, x = value_cast_fp16)[name = tensor("op_12722_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(true)]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_12722_cast_fp16, y = obj_447_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_12725 = const()[name = tensor("op_12725"), val = tensor([1, 1280, 1, -1])]; + tensor input_947_cast_fp16 = reshape(shape = var_12725, x = attn_cast_fp16)[name = tensor("input_947_cast_fp16")]; + tensor var_12732 = const()[name = tensor("op_12732"), val = tensor([1, 1])]; + tensor var_12734 = const()[name = tensor("op_12734"), val = tensor([1, 1])]; + tensor pretrained_out_635_pad_type_0 = const()[name = tensor("pretrained_out_635_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_635_pad_0 = const()[name = tensor("pretrained_out_635_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584360704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585179968))), name = tensor("layers_31_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor layers_31_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585180096)))]; + tensor pretrained_out_635_cast_fp16 = conv(bias = layers_31_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = var_12734, groups = var_12445, pad = pretrained_out_635_pad_0, pad_type = pretrained_out_635_pad_type_0, strides = var_12732, weight = layers_31_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_947_cast_fp16)[name = tensor("pretrained_out_635_cast_fp16")]; + tensor var_12738 = const()[name = tensor("op_12738"), val = tensor([1, 1])]; + tensor var_12740 = const()[name = tensor("op_12740"), val = tensor([1, 1])]; + tensor input_949_pad_type_0 = const()[name = tensor("input_949_pad_type_0"), val = tensor("custom")]; + tensor input_949_pad_0 = const()[name = tensor("input_949_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_31_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585182720)))]; + tensor input_949_cast_fp16 = conv(dilations = var_12740, groups = var_12445, pad = input_949_pad_0, pad_type = input_949_pad_type_0, strides = var_12738, weight = layers_31_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_947_cast_fp16)[name = tensor("input_949_cast_fp16")]; + tensor var_12744 = const()[name = tensor("op_12744"), val = tensor([1, 1])]; + tensor var_12746 = const()[name = tensor("op_12746"), val = tensor([1, 1])]; + tensor lora_out_1269_pad_type_0 = const()[name = tensor("lora_out_1269_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1269_pad_0 = const()[name = tensor("lora_out_1269_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1271_weight_0_to_fp16 = const()[name = tensor("lora_out_1271_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585223744)))]; + tensor lora_out_1271_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12746, groups = var_12445, pad = lora_out_1269_pad_0, pad_type = lora_out_1269_pad_type_0, strides = var_12744, weight = lora_out_1271_weight_0_to_fp16, x = input_949_cast_fp16)[name = tensor("lora_out_1271_cast_fp16")]; + tensor obj_445_cast_fp16 = add(x = pretrained_out_635_cast_fp16, y = lora_out_1271_cast_fp16)[name = tensor("obj_445_cast_fp16")]; + tensor inputs_191_cast_fp16 = add(x = inputs_189_cast_fp16, y = obj_445_cast_fp16)[name = tensor("inputs_191_cast_fp16")]; + tensor var_12755 = const()[name = tensor("op_12755"), val = tensor([1])]; + tensor channels_mean_191_cast_fp16 = reduce_mean(axes = var_12755, keep_dims = var_12446, x = inputs_191_cast_fp16)[name = tensor("channels_mean_191_cast_fp16")]; + tensor zero_mean_191_cast_fp16 = sub(x = inputs_191_cast_fp16, y = channels_mean_191_cast_fp16)[name = tensor("zero_mean_191_cast_fp16")]; + tensor zero_mean_sq_191_cast_fp16 = mul(x = zero_mean_191_cast_fp16, y = zero_mean_191_cast_fp16)[name = tensor("zero_mean_sq_191_cast_fp16")]; + tensor var_12759 = const()[name = tensor("op_12759"), val = tensor([1])]; + tensor var_12760_cast_fp16 = reduce_mean(axes = var_12759, keep_dims = var_12446, x = zero_mean_sq_191_cast_fp16)[name = tensor("op_12760_cast_fp16")]; + tensor var_12761_to_fp16 = const()[name = tensor("op_12761_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_12762_cast_fp16 = add(x = var_12760_cast_fp16, y = var_12761_to_fp16)[name = tensor("op_12762_cast_fp16")]; + tensor denom_191_epsilon_0 = const()[name = tensor("denom_191_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_191_cast_fp16 = rsqrt(epsilon = denom_191_epsilon_0, x = var_12762_cast_fp16)[name = tensor("denom_191_cast_fp16")]; + tensor out_191_cast_fp16 = mul(x = zero_mean_191_cast_fp16, y = denom_191_cast_fp16)[name = tensor("out_191_cast_fp16")]; + tensor input_951_gamma_0_to_fp16 = const()[name = tensor("input_951_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585264768)))]; + tensor input_951_beta_0_to_fp16 = const()[name = tensor("input_951_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585267392)))]; + tensor input_951_epsilon_0_to_fp16 = const()[name = tensor("input_951_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_951_cast_fp16 = batch_norm(beta = input_951_beta_0_to_fp16, epsilon = input_951_epsilon_0_to_fp16, gamma = input_951_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_191_cast_fp16)[name = tensor("input_951_cast_fp16")]; + tensor var_12776 = const()[name = tensor("op_12776"), val = tensor([1, 1])]; + tensor var_12778 = const()[name = tensor("op_12778"), val = tensor([1, 1])]; + tensor pretrained_out_637_pad_type_0 = const()[name = tensor("pretrained_out_637_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_637_pad_0 = const()[name = tensor("pretrained_out_637_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585270016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588546880))), name = tensor("layers_31_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; + tensor layers_31_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588547008)))]; + tensor pretrained_out_637_cast_fp16 = conv(bias = layers_31_fc1_pretrained_bias_to_fp16, dilations = var_12778, groups = var_12445, pad = pretrained_out_637_pad_0, pad_type = pretrained_out_637_pad_type_0, strides = var_12776, weight = layers_31_fc1_pretrained_weight_to_fp16_palettized, x = input_951_cast_fp16)[name = tensor("pretrained_out_637_cast_fp16")]; + tensor var_12782 = const()[name = tensor("op_12782"), val = tensor([1, 1])]; + tensor var_12784 = const()[name = tensor("op_12784"), val = tensor([1, 1])]; + tensor input_953_pad_type_0 = const()[name = tensor("input_953_pad_type_0"), val = tensor("custom")]; + tensor input_953_pad_0 = const()[name = tensor("input_953_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_31_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588557312)))]; + tensor input_953_cast_fp16 = conv(dilations = var_12784, groups = var_12445, pad = input_953_pad_0, pad_type = input_953_pad_type_0, strides = var_12782, weight = layers_31_fc1_loraA_weight_to_fp16, x = input_951_cast_fp16)[name = tensor("input_953_cast_fp16")]; + tensor var_12788 = const()[name = tensor("op_12788"), val = tensor([1, 1])]; + tensor var_12790 = const()[name = tensor("op_12790"), val = tensor([1, 1])]; + tensor lora_out_1273_pad_type_0 = const()[name = tensor("lora_out_1273_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1273_pad_0 = const()[name = tensor("lora_out_1273_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_1275_weight_0_to_fp16 = const()[name = tensor("lora_out_1275_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588598336)))]; + tensor lora_out_1275_cast_fp16 = conv(bias = lora_out_35_bias_0_to_fp16, dilations = var_12790, groups = var_12445, pad = lora_out_1273_pad_0, pad_type = lora_out_1273_pad_type_0, strides = var_12788, weight = lora_out_1275_weight_0_to_fp16, x = input_953_cast_fp16)[name = tensor("lora_out_1275_cast_fp16")]; + tensor input_955_cast_fp16 = add(x = pretrained_out_637_cast_fp16, y = lora_out_1275_cast_fp16)[name = tensor("input_955_cast_fp16")]; + tensor input_957_mode_0 = const()[name = tensor("input_957_mode_0"), val = tensor("EXACT")]; + tensor input_957_cast_fp16 = gelu(mode = input_957_mode_0, x = input_955_cast_fp16)[name = tensor("input_957_cast_fp16")]; + tensor var_12802 = const()[name = tensor("op_12802"), val = tensor([1, 1])]; + tensor var_12804 = const()[name = tensor("op_12804"), val = tensor([1, 1])]; + tensor pretrained_out_pad_type_0 = const()[name = tensor("pretrained_out_pad_type_0"), val = tensor("custom")]; + tensor pretrained_out_pad_0 = const()[name = tensor("pretrained_out_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588762240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592039104))), name = tensor("layers_31_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; + tensor layers_31_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_31_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592039232)))]; + tensor pretrained_out_cast_fp16 = conv(bias = layers_31_fc2_pretrained_bias_to_fp16, dilations = var_12804, groups = var_12445, pad = pretrained_out_pad_0, pad_type = pretrained_out_pad_type_0, strides = var_12802, weight = layers_31_fc2_pretrained_weight_to_fp16_palettized, x = input_957_cast_fp16)[name = tensor("pretrained_out_cast_fp16")]; + tensor var_12808 = const()[name = tensor("op_12808"), val = tensor([1, 1])]; + tensor var_12810 = const()[name = tensor("op_12810"), val = tensor([1, 1])]; + tensor input_pad_type_0 = const()[name = tensor("input_pad_type_0"), val = tensor("custom")]; + tensor input_pad_0 = const()[name = tensor("input_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor layers_31_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_31_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592041856)))]; + tensor input_cast_fp16 = conv(dilations = var_12810, groups = var_12445, pad = input_pad_0, pad_type = input_pad_type_0, strides = var_12808, weight = layers_31_fc2_loraA_weight_to_fp16, x = input_957_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_12814 = const()[name = tensor("op_12814"), val = tensor([1, 1])]; + tensor var_12816 = const()[name = tensor("op_12816"), val = tensor([1, 1])]; + tensor lora_out_1277_pad_type_0 = const()[name = tensor("lora_out_1277_pad_type_0"), val = tensor("custom")]; + tensor lora_out_1277_pad_0 = const()[name = tensor("lora_out_1277_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor lora_out_weight_0_to_fp16 = const()[name = tensor("lora_out_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592205760)))]; + tensor lora_out_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_12816, groups = var_12445, pad = lora_out_1277_pad_0, pad_type = lora_out_1277_pad_type_0, strides = var_12814, weight = lora_out_weight_0_to_fp16, x = input_cast_fp16)[name = tensor("lora_out_cast_fp16")]; + tensor hidden_states_65_cast_fp16 = add(x = pretrained_out_cast_fp16, y = lora_out_cast_fp16)[name = tensor("hidden_states_65_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_191_cast_fp16, y = hidden_states_65_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor var_12829 = const()[name = tensor("op_12829"), val = tensor(true)]; + tensor var_12833 = const()[name = tensor("op_12833"), val = tensor([1])]; + tensor channels_mean_cast_fp16 = reduce_mean(axes = var_12833, keep_dims = var_12829, x = inputs_cast_fp16)[name = tensor("channels_mean_cast_fp16")]; + tensor zero_mean_cast_fp16 = sub(x = inputs_cast_fp16, y = channels_mean_cast_fp16)[name = tensor("zero_mean_cast_fp16")]; + tensor zero_mean_sq_cast_fp16 = mul(x = zero_mean_cast_fp16, y = zero_mean_cast_fp16)[name = tensor("zero_mean_sq_cast_fp16")]; + tensor var_12837 = const()[name = tensor("op_12837"), val = tensor([1])]; + tensor var_12838_cast_fp16 = reduce_mean(axes = var_12837, keep_dims = var_12829, x = zero_mean_sq_cast_fp16)[name = tensor("op_12838_cast_fp16")]; + tensor var_12839_to_fp16 = const()[name = tensor("op_12839_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_12840_cast_fp16 = add(x = var_12838_cast_fp16, y = var_12839_to_fp16)[name = tensor("op_12840_cast_fp16")]; + tensor denom_epsilon_0 = const()[name = tensor("denom_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor denom_cast_fp16 = rsqrt(epsilon = denom_epsilon_0, x = var_12840_cast_fp16)[name = tensor("denom_cast_fp16")]; + tensor out_cast_fp16 = mul(x = zero_mean_cast_fp16, y = denom_cast_fp16)[name = tensor("out_cast_fp16")]; + tensor hidden_states_gamma_0_to_fp16 = const()[name = tensor("hidden_states_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592246784)))]; + tensor hidden_states_beta_0_to_fp16 = const()[name = tensor("hidden_states_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592249408)))]; + tensor hidden_states_epsilon_0_to_fp16 = const()[name = tensor("hidden_states_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor("hidden_states_cast_fp16")]; + tensor var_12850_axes_0 = const()[name = tensor("op_12850_axes_0"), val = tensor([2])]; + tensor var_12850_cast_fp16 = squeeze(axes = var_12850_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_12850_cast_fp16")]; + tensor var_12853_perm_0 = const()[name = tensor("op_12853_perm_0"), val = tensor([0, 2, 1])]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592252032)))]; + tensor transpose_0 = transpose(perm = var_12853_perm_0, x = var_12850_cast_fp16)[name = tensor("transpose_0")]; + tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = transpose_0)[name = tensor("linear_0_cast_fp16")]; + tensor var_12857 = const()[name = tensor("op_12857"), val = tensor(1)]; + tensor obj_451_interleave_0 = const()[name = tensor("obj_451_interleave_0"), val = tensor(false)]; + tensor key_cache_updates = concat(axis = var_12857, interleave = obj_451_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_7_cast_fp16, current_key_9_cast_fp16, current_key_11_cast_fp16, current_key_13_cast_fp16, current_key_15_cast_fp16, current_key_17_cast_fp16, current_key_19_cast_fp16, current_key_21_cast_fp16, current_key_23_cast_fp16, current_key_25_cast_fp16, current_key_27_cast_fp16, current_key_29_cast_fp16, current_key_31_cast_fp16, current_key_33_cast_fp16, current_key_35_cast_fp16, current_key_37_cast_fp16, current_key_39_cast_fp16, current_key_41_cast_fp16, current_key_43_cast_fp16, current_key_45_cast_fp16, current_key_47_cast_fp16, current_key_49_cast_fp16, current_key_51_cast_fp16, current_key_53_cast_fp16, current_key_55_cast_fp16, current_key_57_cast_fp16, current_key_59_cast_fp16, current_key_61_cast_fp16, current_key_cast_fp16))[name = tensor("obj_451_cast_fp16")]; + tensor var_12860 = const()[name = tensor("op_12860"), val = tensor(1)]; + tensor obj_453_interleave_0 = const()[name = tensor("obj_453_interleave_0"), val = tensor(false)]; + tensor value_cache_updates = concat(axis = var_12860, interleave = obj_453_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_7_cast_fp16, current_value_9_cast_fp16, current_value_11_cast_fp16, current_value_13_cast_fp16, current_value_15_cast_fp16, current_value_17_cast_fp16, current_value_19_cast_fp16, current_value_21_cast_fp16, current_value_23_cast_fp16, current_value_25_cast_fp16, current_value_27_cast_fp16, current_value_29_cast_fp16, current_value_31_cast_fp16, current_value_33_cast_fp16, current_value_35_cast_fp16, current_value_37_cast_fp16, current_value_39_cast_fp16, current_value_41_cast_fp16, current_value_43_cast_fp16, current_value_45_cast_fp16, current_value_47_cast_fp16, current_value_49_cast_fp16, current_value_51_cast_fp16, current_value_53_cast_fp16, current_value_55_cast_fp16, current_value_57_cast_fp16, current_value_59_cast_fp16, current_value_61_cast_fp16, current_value_cast_fp16))[name = tensor("obj_453_cast_fp16")]; + tensor var_12871_begin_0 = const()[name = tensor("op_12871_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_12871_end_0 = const()[name = tensor("op_12871_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_12871_end_mask_0 = const()[name = tensor("op_12871_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_12871_cast_fp16 = slice_by_index(begin = var_12871_begin_0, end = var_12871_end_0, end_mask = var_12871_end_mask_0, x = obj_153_cast_fp16)[name = tensor("op_12871_cast_fp16")]; + tensor var_12874_begin_0 = const()[name = tensor("op_12874_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_12874_end_0 = const()[name = tensor("op_12874_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_12874_end_mask_0 = const()[name = tensor("op_12874_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_12874_squeeze_mask_0 = const()[name = tensor("op_12874_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_12874_cast_fp16 = slice_by_index(begin = var_12874_begin_0, end = var_12874_end_0, end_mask = var_12874_end_mask_0, squeeze_mask = var_12874_squeeze_mask_0, x = var_12871_cast_fp16)[name = tensor("op_12874_cast_fp16")]; + tensor var_12889_begin_0 = const()[name = tensor("op_12889_begin_0"), val = tensor([0, 17, 0, 0])]; + tensor var_12889_end_0 = const()[name = tensor("op_12889_end_0"), val = tensor([1, 18, 1, 1500])]; + tensor var_12889_end_mask_0 = const()[name = tensor("op_12889_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_12889_cast_fp16 = slice_by_index(begin = var_12889_begin_0, end = var_12889_end_0, end_mask = var_12889_end_mask_0, x = obj_195_cast_fp16)[name = tensor("op_12889_cast_fp16")]; + tensor var_12892_begin_0 = const()[name = tensor("op_12892_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_12892_end_0 = const()[name = tensor("op_12892_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_12892_end_mask_0 = const()[name = tensor("op_12892_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_12892_squeeze_mask_0 = const()[name = tensor("op_12892_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_12892_cast_fp16 = slice_by_index(begin = var_12892_begin_0, end = var_12892_end_0, end_mask = var_12892_end_mask_0, squeeze_mask = var_12892_squeeze_mask_0, x = var_12889_cast_fp16)[name = tensor("op_12892_cast_fp16")]; + tensor var_12907_begin_0 = const()[name = tensor("op_12907_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_12907_end_0 = const()[name = tensor("op_12907_end_0"), val = tensor([1, 12, 1, 1500])]; + tensor var_12907_end_mask_0 = const()[name = tensor("op_12907_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_12907_cast_fp16 = slice_by_index(begin = var_12907_begin_0, end = var_12907_end_0, end_mask = var_12907_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_12907_cast_fp16")]; + tensor var_12910_begin_0 = const()[name = tensor("op_12910_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_12910_end_0 = const()[name = tensor("op_12910_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_12910_end_mask_0 = const()[name = tensor("op_12910_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_12910_squeeze_mask_0 = const()[name = tensor("op_12910_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_12910_cast_fp16 = slice_by_index(begin = var_12910_begin_0, end = var_12910_end_0, end_mask = var_12910_end_mask_0, squeeze_mask = var_12910_squeeze_mask_0, x = var_12907_cast_fp16)[name = tensor("op_12910_cast_fp16")]; + tensor var_12925_begin_0 = const()[name = tensor("op_12925_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_12925_end_0 = const()[name = tensor("op_12925_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_12925_end_mask_0 = const()[name = tensor("op_12925_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_12925_cast_fp16 = slice_by_index(begin = var_12925_begin_0, end = var_12925_end_0, end_mask = var_12925_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_12925_cast_fp16")]; + tensor var_12928_begin_0 = const()[name = tensor("op_12928_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_12928_end_0 = const()[name = tensor("op_12928_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_12928_end_mask_0 = const()[name = tensor("op_12928_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_12928_squeeze_mask_0 = const()[name = tensor("op_12928_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_12928_cast_fp16 = slice_by_index(begin = var_12928_begin_0, end = var_12928_end_0, end_mask = var_12928_end_mask_0, squeeze_mask = var_12928_squeeze_mask_0, x = var_12925_cast_fp16)[name = tensor("op_12928_cast_fp16")]; + tensor var_12943_begin_0 = const()[name = tensor("op_12943_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_12943_end_0 = const()[name = tensor("op_12943_end_0"), val = tensor([1, 14, 1, 1500])]; + tensor var_12943_end_mask_0 = const()[name = tensor("op_12943_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_12943_cast_fp16 = slice_by_index(begin = var_12943_begin_0, end = var_12943_end_0, end_mask = var_12943_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_12943_cast_fp16")]; + tensor var_12946_begin_0 = const()[name = tensor("op_12946_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_12946_end_0 = const()[name = tensor("op_12946_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_12946_end_mask_0 = const()[name = tensor("op_12946_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_12946_squeeze_mask_0 = const()[name = tensor("op_12946_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_12946_cast_fp16 = slice_by_index(begin = var_12946_begin_0, end = var_12946_end_0, end_mask = var_12946_end_mask_0, squeeze_mask = var_12946_squeeze_mask_0, x = var_12943_cast_fp16)[name = tensor("op_12946_cast_fp16")]; + tensor var_12961_begin_0 = const()[name = tensor("op_12961_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_12961_end_0 = const()[name = tensor("op_12961_end_0"), val = tensor([1, 16, 1, 1500])]; + tensor var_12961_end_mask_0 = const()[name = tensor("op_12961_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_12961_cast_fp16 = slice_by_index(begin = var_12961_begin_0, end = var_12961_end_0, end_mask = var_12961_end_mask_0, x = obj_251_cast_fp16)[name = tensor("op_12961_cast_fp16")]; + tensor var_12964_begin_0 = const()[name = tensor("op_12964_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_12964_end_0 = const()[name = tensor("op_12964_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_12964_end_mask_0 = const()[name = tensor("op_12964_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_12964_squeeze_mask_0 = const()[name = tensor("op_12964_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_12964_cast_fp16 = slice_by_index(begin = var_12964_begin_0, end = var_12964_end_0, end_mask = var_12964_end_mask_0, squeeze_mask = var_12964_squeeze_mask_0, x = var_12961_cast_fp16)[name = tensor("op_12964_cast_fp16")]; + tensor var_12979_begin_0 = const()[name = tensor("op_12979_begin_0"), val = tensor([0, 16, 0, 0])]; + tensor var_12979_end_0 = const()[name = tensor("op_12979_end_0"), val = tensor([1, 17, 1, 1500])]; + tensor var_12979_end_mask_0 = const()[name = tensor("op_12979_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_12979_cast_fp16 = slice_by_index(begin = var_12979_begin_0, end = var_12979_end_0, end_mask = var_12979_end_mask_0, x = obj_251_cast_fp16)[name = tensor("op_12979_cast_fp16")]; + tensor var_12982_begin_0 = const()[name = tensor("op_12982_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_12982_end_0 = const()[name = tensor("op_12982_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_12982_end_mask_0 = const()[name = tensor("op_12982_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_12982_squeeze_mask_0 = const()[name = tensor("op_12982_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_12982_cast_fp16 = slice_by_index(begin = var_12982_begin_0, end = var_12982_end_0, end_mask = var_12982_end_mask_0, squeeze_mask = var_12982_squeeze_mask_0, x = var_12979_cast_fp16)[name = tensor("op_12982_cast_fp16")]; + tensor var_12997_begin_0 = const()[name = tensor("op_12997_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_12997_end_0 = const()[name = tensor("op_12997_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_12997_end_mask_0 = const()[name = tensor("op_12997_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_12997_cast_fp16 = slice_by_index(begin = var_12997_begin_0, end = var_12997_end_0, end_mask = var_12997_end_mask_0, x = obj_265_cast_fp16)[name = tensor("op_12997_cast_fp16")]; + tensor var_13000_begin_0 = const()[name = tensor("op_13000_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13000_end_0 = const()[name = tensor("op_13000_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13000_end_mask_0 = const()[name = tensor("op_13000_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13000_squeeze_mask_0 = const()[name = tensor("op_13000_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13000_cast_fp16 = slice_by_index(begin = var_13000_begin_0, end = var_13000_end_0, end_mask = var_13000_end_mask_0, squeeze_mask = var_13000_squeeze_mask_0, x = var_12997_cast_fp16)[name = tensor("op_13000_cast_fp16")]; + tensor var_13015_begin_0 = const()[name = tensor("op_13015_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_13015_end_0 = const()[name = tensor("op_13015_end_0"), val = tensor([1, 12, 1, 1500])]; + tensor var_13015_end_mask_0 = const()[name = tensor("op_13015_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13015_cast_fp16 = slice_by_index(begin = var_13015_begin_0, end = var_13015_end_0, end_mask = var_13015_end_mask_0, x = obj_265_cast_fp16)[name = tensor("op_13015_cast_fp16")]; + tensor var_13018_begin_0 = const()[name = tensor("op_13018_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13018_end_0 = const()[name = tensor("op_13018_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13018_end_mask_0 = const()[name = tensor("op_13018_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13018_squeeze_mask_0 = const()[name = tensor("op_13018_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13018_cast_fp16 = slice_by_index(begin = var_13018_begin_0, end = var_13018_end_0, end_mask = var_13018_end_mask_0, squeeze_mask = var_13018_squeeze_mask_0, x = var_13015_cast_fp16)[name = tensor("op_13018_cast_fp16")]; + tensor var_13033_begin_0 = const()[name = tensor("op_13033_begin_0"), val = tensor([0, 19, 0, 0])]; + tensor var_13033_end_0 = const()[name = tensor("op_13033_end_0"), val = tensor([1, 20, 1, 1500])]; + tensor var_13033_end_mask_0 = const()[name = tensor("op_13033_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13033_cast_fp16 = slice_by_index(begin = var_13033_begin_0, end = var_13033_end_0, end_mask = var_13033_end_mask_0, x = obj_265_cast_fp16)[name = tensor("op_13033_cast_fp16")]; + tensor var_13036_begin_0 = const()[name = tensor("op_13036_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13036_end_0 = const()[name = tensor("op_13036_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13036_end_mask_0 = const()[name = tensor("op_13036_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13036_squeeze_mask_0 = const()[name = tensor("op_13036_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13036_cast_fp16 = slice_by_index(begin = var_13036_begin_0, end = var_13036_end_0, end_mask = var_13036_end_mask_0, squeeze_mask = var_13036_squeeze_mask_0, x = var_13033_cast_fp16)[name = tensor("op_13036_cast_fp16")]; + tensor var_13051_begin_0 = const()[name = tensor("op_13051_begin_0"), val = tensor([0, 11, 0, 0])]; + tensor var_13051_end_0 = const()[name = tensor("op_13051_end_0"), val = tensor([1, 12, 1, 1500])]; + tensor var_13051_end_mask_0 = const()[name = tensor("op_13051_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13051_cast_fp16 = slice_by_index(begin = var_13051_begin_0, end = var_13051_end_0, end_mask = var_13051_end_mask_0, x = obj_279_cast_fp16)[name = tensor("op_13051_cast_fp16")]; + tensor var_13054_begin_0 = const()[name = tensor("op_13054_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13054_end_0 = const()[name = tensor("op_13054_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13054_end_mask_0 = const()[name = tensor("op_13054_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13054_squeeze_mask_0 = const()[name = tensor("op_13054_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13054_cast_fp16 = slice_by_index(begin = var_13054_begin_0, end = var_13054_end_0, end_mask = var_13054_end_mask_0, squeeze_mask = var_13054_squeeze_mask_0, x = var_13051_cast_fp16)[name = tensor("op_13054_cast_fp16")]; + tensor var_13069_begin_0 = const()[name = tensor("op_13069_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor var_13069_end_0 = const()[name = tensor("op_13069_end_0"), val = tensor([1, 3, 1, 1500])]; + tensor var_13069_end_mask_0 = const()[name = tensor("op_13069_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13069_cast_fp16 = slice_by_index(begin = var_13069_begin_0, end = var_13069_end_0, end_mask = var_13069_end_mask_0, x = obj_307_cast_fp16)[name = tensor("op_13069_cast_fp16")]; + tensor var_13072_begin_0 = const()[name = tensor("op_13072_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13072_end_0 = const()[name = tensor("op_13072_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13072_end_mask_0 = const()[name = tensor("op_13072_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13072_squeeze_mask_0 = const()[name = tensor("op_13072_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13072_cast_fp16 = slice_by_index(begin = var_13072_begin_0, end = var_13072_end_0, end_mask = var_13072_end_mask_0, squeeze_mask = var_13072_squeeze_mask_0, x = var_13069_cast_fp16)[name = tensor("op_13072_cast_fp16")]; + tensor var_13087_begin_0 = const()[name = tensor("op_13087_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_13087_end_0 = const()[name = tensor("op_13087_end_0"), val = tensor([1, 4, 1, 1500])]; + tensor var_13087_end_mask_0 = const()[name = tensor("op_13087_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13087_cast_fp16 = slice_by_index(begin = var_13087_begin_0, end = var_13087_end_0, end_mask = var_13087_end_mask_0, x = obj_307_cast_fp16)[name = tensor("op_13087_cast_fp16")]; + tensor var_13090_begin_0 = const()[name = tensor("op_13090_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13090_end_0 = const()[name = tensor("op_13090_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13090_end_mask_0 = const()[name = tensor("op_13090_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13090_squeeze_mask_0 = const()[name = tensor("op_13090_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13090_cast_fp16 = slice_by_index(begin = var_13090_begin_0, end = var_13090_end_0, end_mask = var_13090_end_mask_0, squeeze_mask = var_13090_squeeze_mask_0, x = var_13087_cast_fp16)[name = tensor("op_13090_cast_fp16")]; + tensor var_13105_begin_0 = const()[name = tensor("op_13105_begin_0"), val = tensor([0, 3, 0, 0])]; + tensor var_13105_end_0 = const()[name = tensor("op_13105_end_0"), val = tensor([1, 4, 1, 1500])]; + tensor var_13105_end_mask_0 = const()[name = tensor("op_13105_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13105_cast_fp16 = slice_by_index(begin = var_13105_begin_0, end = var_13105_end_0, end_mask = var_13105_end_mask_0, x = obj_321_cast_fp16)[name = tensor("op_13105_cast_fp16")]; + tensor var_13108_begin_0 = const()[name = tensor("op_13108_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13108_end_0 = const()[name = tensor("op_13108_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13108_end_mask_0 = const()[name = tensor("op_13108_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13108_squeeze_mask_0 = const()[name = tensor("op_13108_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13108_cast_fp16 = slice_by_index(begin = var_13108_begin_0, end = var_13108_end_0, end_mask = var_13108_end_mask_0, squeeze_mask = var_13108_squeeze_mask_0, x = var_13105_cast_fp16)[name = tensor("op_13108_cast_fp16")]; + tensor var_13123_begin_0 = const()[name = tensor("op_13123_begin_0"), val = tensor([0, 9, 0, 0])]; + tensor var_13123_end_0 = const()[name = tensor("op_13123_end_0"), val = tensor([1, 10, 1, 1500])]; + tensor var_13123_end_mask_0 = const()[name = tensor("op_13123_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13123_cast_fp16 = slice_by_index(begin = var_13123_begin_0, end = var_13123_end_0, end_mask = var_13123_end_mask_0, x = obj_321_cast_fp16)[name = tensor("op_13123_cast_fp16")]; + tensor var_13126_begin_0 = const()[name = tensor("op_13126_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13126_end_0 = const()[name = tensor("op_13126_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13126_end_mask_0 = const()[name = tensor("op_13126_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13126_squeeze_mask_0 = const()[name = tensor("op_13126_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13126_cast_fp16 = slice_by_index(begin = var_13126_begin_0, end = var_13126_end_0, end_mask = var_13126_end_mask_0, squeeze_mask = var_13126_squeeze_mask_0, x = var_13123_cast_fp16)[name = tensor("op_13126_cast_fp16")]; + tensor var_13141_begin_0 = const()[name = tensor("op_13141_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_13141_end_0 = const()[name = tensor("op_13141_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_13141_end_mask_0 = const()[name = tensor("op_13141_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13141_cast_fp16 = slice_by_index(begin = var_13141_begin_0, end = var_13141_end_0, end_mask = var_13141_end_mask_0, x = obj_321_cast_fp16)[name = tensor("op_13141_cast_fp16")]; + tensor var_13144_begin_0 = const()[name = tensor("op_13144_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13144_end_0 = const()[name = tensor("op_13144_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13144_end_mask_0 = const()[name = tensor("op_13144_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13144_squeeze_mask_0 = const()[name = tensor("op_13144_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13144_cast_fp16 = slice_by_index(begin = var_13144_begin_0, end = var_13144_end_0, end_mask = var_13144_end_mask_0, squeeze_mask = var_13144_squeeze_mask_0, x = var_13141_cast_fp16)[name = tensor("op_13144_cast_fp16")]; + tensor var_13159_begin_0 = const()[name = tensor("op_13159_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_13159_end_0 = const()[name = tensor("op_13159_end_0"), val = tensor([1, 6, 1, 1500])]; + tensor var_13159_end_mask_0 = const()[name = tensor("op_13159_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13159_cast_fp16 = slice_by_index(begin = var_13159_begin_0, end = var_13159_end_0, end_mask = var_13159_end_mask_0, x = obj_335_cast_fp16)[name = tensor("op_13159_cast_fp16")]; + tensor var_13162_begin_0 = const()[name = tensor("op_13162_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13162_end_0 = const()[name = tensor("op_13162_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13162_end_mask_0 = const()[name = tensor("op_13162_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13162_squeeze_mask_0 = const()[name = tensor("op_13162_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13162_cast_fp16 = slice_by_index(begin = var_13162_begin_0, end = var_13162_end_0, end_mask = var_13162_end_mask_0, squeeze_mask = var_13162_squeeze_mask_0, x = var_13159_cast_fp16)[name = tensor("op_13162_cast_fp16")]; + tensor var_13177_begin_0 = const()[name = tensor("op_13177_begin_0"), val = tensor([0, 7, 0, 0])]; + tensor var_13177_end_0 = const()[name = tensor("op_13177_end_0"), val = tensor([1, 8, 1, 1500])]; + tensor var_13177_end_mask_0 = const()[name = tensor("op_13177_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13177_cast_fp16 = slice_by_index(begin = var_13177_begin_0, end = var_13177_end_0, end_mask = var_13177_end_mask_0, x = obj_335_cast_fp16)[name = tensor("op_13177_cast_fp16")]; + tensor var_13180_begin_0 = const()[name = tensor("op_13180_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13180_end_0 = const()[name = tensor("op_13180_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13180_end_mask_0 = const()[name = tensor("op_13180_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13180_squeeze_mask_0 = const()[name = tensor("op_13180_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13180_cast_fp16 = slice_by_index(begin = var_13180_begin_0, end = var_13180_end_0, end_mask = var_13180_end_mask_0, squeeze_mask = var_13180_squeeze_mask_0, x = var_13177_cast_fp16)[name = tensor("op_13180_cast_fp16")]; + tensor var_13195_begin_0 = const()[name = tensor("op_13195_begin_0"), val = tensor([0, 13, 0, 0])]; + tensor var_13195_end_0 = const()[name = tensor("op_13195_end_0"), val = tensor([1, 14, 1, 1500])]; + tensor var_13195_end_mask_0 = const()[name = tensor("op_13195_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13195_cast_fp16 = slice_by_index(begin = var_13195_begin_0, end = var_13195_end_0, end_mask = var_13195_end_mask_0, x = obj_335_cast_fp16)[name = tensor("op_13195_cast_fp16")]; + tensor var_13198_begin_0 = const()[name = tensor("op_13198_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13198_end_0 = const()[name = tensor("op_13198_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13198_end_mask_0 = const()[name = tensor("op_13198_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13198_squeeze_mask_0 = const()[name = tensor("op_13198_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13198_cast_fp16 = slice_by_index(begin = var_13198_begin_0, end = var_13198_end_0, end_mask = var_13198_end_mask_0, squeeze_mask = var_13198_squeeze_mask_0, x = var_13195_cast_fp16)[name = tensor("op_13198_cast_fp16")]; + tensor var_13213_begin_0 = const()[name = tensor("op_13213_begin_0"), val = tensor([0, 5, 0, 0])]; + tensor var_13213_end_0 = const()[name = tensor("op_13213_end_0"), val = tensor([1, 6, 1, 1500])]; + tensor var_13213_end_mask_0 = const()[name = tensor("op_13213_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13213_cast_fp16 = slice_by_index(begin = var_13213_begin_0, end = var_13213_end_0, end_mask = var_13213_end_mask_0, x = obj_363_cast_fp16)[name = tensor("op_13213_cast_fp16")]; + tensor var_13216_begin_0 = const()[name = tensor("op_13216_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13216_end_0 = const()[name = tensor("op_13216_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13216_end_mask_0 = const()[name = tensor("op_13216_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13216_squeeze_mask_0 = const()[name = tensor("op_13216_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13216_cast_fp16 = slice_by_index(begin = var_13216_begin_0, end = var_13216_end_0, end_mask = var_13216_end_mask_0, squeeze_mask = var_13216_squeeze_mask_0, x = var_13213_cast_fp16)[name = tensor("op_13216_cast_fp16")]; + tensor var_13231_begin_0 = const()[name = tensor("op_13231_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_13231_end_0 = const()[name = tensor("op_13231_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_13231_end_mask_0 = const()[name = tensor("op_13231_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13231_cast_fp16 = slice_by_index(begin = var_13231_begin_0, end = var_13231_end_0, end_mask = var_13231_end_mask_0, x = obj_377_cast_fp16)[name = tensor("op_13231_cast_fp16")]; + tensor var_13234_begin_0 = const()[name = tensor("op_13234_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13234_end_0 = const()[name = tensor("op_13234_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13234_end_mask_0 = const()[name = tensor("op_13234_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13234_squeeze_mask_0 = const()[name = tensor("op_13234_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13234_cast_fp16 = slice_by_index(begin = var_13234_begin_0, end = var_13234_end_0, end_mask = var_13234_end_mask_0, squeeze_mask = var_13234_squeeze_mask_0, x = var_13231_cast_fp16)[name = tensor("op_13234_cast_fp16")]; + tensor var_13249_begin_0 = const()[name = tensor("op_13249_begin_0"), val = tensor([0, 12, 0, 0])]; + tensor var_13249_end_0 = const()[name = tensor("op_13249_end_0"), val = tensor([1, 13, 1, 1500])]; + tensor var_13249_end_mask_0 = const()[name = tensor("op_13249_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13249_cast_fp16 = slice_by_index(begin = var_13249_begin_0, end = var_13249_end_0, end_mask = var_13249_end_mask_0, x = obj_377_cast_fp16)[name = tensor("op_13249_cast_fp16")]; + tensor var_13252_begin_0 = const()[name = tensor("op_13252_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13252_end_0 = const()[name = tensor("op_13252_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13252_end_mask_0 = const()[name = tensor("op_13252_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13252_squeeze_mask_0 = const()[name = tensor("op_13252_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13252_cast_fp16 = slice_by_index(begin = var_13252_begin_0, end = var_13252_end_0, end_mask = var_13252_end_mask_0, squeeze_mask = var_13252_squeeze_mask_0, x = var_13249_cast_fp16)[name = tensor("op_13252_cast_fp16")]; + tensor var_13267_begin_0 = const()[name = tensor("op_13267_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_13267_end_0 = const()[name = tensor("op_13267_end_0"), val = tensor([1, 16, 1, 1500])]; + tensor var_13267_end_mask_0 = const()[name = tensor("op_13267_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_13267_cast_fp16 = slice_by_index(begin = var_13267_begin_0, end = var_13267_end_0, end_mask = var_13267_end_mask_0, x = obj_391_cast_fp16)[name = tensor("op_13267_cast_fp16")]; + tensor var_13270_begin_0 = const()[name = tensor("op_13270_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_13270_end_0 = const()[name = tensor("op_13270_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_13270_end_mask_0 = const()[name = tensor("op_13270_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_13270_squeeze_mask_0 = const()[name = tensor("op_13270_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_13270_cast_fp16 = slice_by_index(begin = var_13270_begin_0, end = var_13270_end_0, end_mask = var_13270_end_mask_0, squeeze_mask = var_13270_squeeze_mask_0, x = var_13267_cast_fp16)[name = tensor("op_13270_cast_fp16")]; + tensor var_13277 = const()[name = tensor("op_13277"), val = tensor(1)]; + tensor var_13278_interleave_0 = const()[name = tensor("op_13278_interleave_0"), val = tensor(false)]; + tensor var_13278_cast_fp16 = concat(axis = var_13277, interleave = var_13278_interleave_0, values = (var_12874_cast_fp16, var_12892_cast_fp16, var_12910_cast_fp16, var_12928_cast_fp16, var_12946_cast_fp16, var_12964_cast_fp16, var_12982_cast_fp16, var_13000_cast_fp16, var_13018_cast_fp16, var_13036_cast_fp16, var_13054_cast_fp16, var_13072_cast_fp16, var_13090_cast_fp16, var_13108_cast_fp16, var_13126_cast_fp16, var_13144_cast_fp16, var_13162_cast_fp16, var_13180_cast_fp16, var_13198_cast_fp16, var_13216_cast_fp16, var_13234_cast_fp16, var_13252_cast_fp16, var_13270_cast_fp16))[name = tensor("op_13278_cast_fp16")]; + tensor var_13280 = const()[name = tensor("op_13280"), val = tensor([1])]; + tensor var_13281 = const()[name = tensor("op_13281"), val = tensor(false)]; + tensor alignment_heads_weights = reduce_mean(axes = var_13280, keep_dims = var_13281, x = var_13278_cast_fp16)[name = tensor("obj_cast_fp16")]; + } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); +} \ No newline at end of file