diff --git "a/openai_whisper-large-v3_947MB/AudioEncoder.mlmodelc/model.mil" "b/openai_whisper-large-v3_947MB/AudioEncoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/openai_whisper-large-v3_947MB/AudioEncoder.mlmodelc/model.mil" @@ -0,0 +1,5583 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3304.5.2"}, {"coremlc-version", "3304.6.2"}})] +{ + func main(tensor melspectrogram_features) { + tensor var_90 = const()[name = tensor("op_90"), val = tensor([1, 1])]; + tensor var_96 = const()[name = tensor("op_96"), val = tensor([1, 1])]; + tensor var_101 = const()[name = tensor("op_101"), val = tensor(1)]; + tensor var_106_pad_type_0 = const()[name = tensor("op_106_pad_type_0"), val = tensor("custom")]; + tensor var_106_pad_0 = const()[name = tensor("op_106_pad_0"), val = tensor([0, 0, 1, 1])]; + tensor var_81_to_fp16 = const()[name = tensor("op_81_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor var_87_to_fp16 = const()[name = tensor("op_87_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(983168)))]; + tensor var_106_cast_fp16 = conv(bias = var_87_to_fp16, dilations = var_96, groups = var_101, pad = var_106_pad_0, pad_type = var_106_pad_type_0, strides = var_90, weight = var_81_to_fp16, x = melspectrogram_features)[name = tensor("op_106_cast_fp16")]; + tensor hidden_states_1_mode_0 = const()[name = tensor("hidden_states_1_mode_0"), val = tensor("EXACT")]; + tensor hidden_states_1_cast_fp16 = gelu(mode = hidden_states_1_mode_0, x = var_106_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([2, 2])]; + tensor var_136 = const()[name = tensor("op_136"), val = tensor([1, 1])]; + tensor var_141 = const()[name = tensor("op_141"), val = tensor(1)]; + tensor var_146_pad_type_0 = const()[name = tensor("op_146_pad_type_0"), val = tensor("custom")]; + tensor var_146_pad_0 = const()[name = tensor("op_146_pad_0"), val = tensor([0, 0, 1, 1])]; + tensor var_121_to_fp16 = const()[name = tensor("op_121_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985792)))]; + tensor var_127_to_fp16 = const()[name = tensor("op_127_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10816256)))]; + tensor var_146_cast_fp16 = conv(bias = var_127_to_fp16, dilations = var_136, groups = var_141, pad = var_146_pad_0, pad_type = var_146_pad_type_0, strides = var_130, weight = var_121_to_fp16, x = hidden_states_1_cast_fp16)[name = tensor("op_146_cast_fp16")]; + tensor hidden_states_3_mode_0 = const()[name = tensor("hidden_states_3_mode_0"), val = tensor("EXACT")]; + tensor hidden_states_3_cast_fp16 = gelu(mode = hidden_states_3_mode_0, x = var_146_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; + tensor var_164_to_fp16 = const()[name = tensor("op_164_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10818880)))]; + tensor inputs_1_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = var_164_to_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_178 = const()[name = tensor("op_178"), val = tensor(3)]; + tensor var_180 = const()[name = tensor("op_180"), val = tensor(1)]; + tensor var_181 = const()[name = tensor("op_181"), val = tensor(true)]; + tensor var_191 = const()[name = tensor("op_191"), val = tensor([1])]; + tensor channels_mean_1_cast_fp16 = reduce_mean(axes = var_191, keep_dims = var_181, 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_195 = const()[name = tensor("op_195"), val = tensor([1])]; + tensor var_196_cast_fp16 = reduce_mean(axes = var_195, keep_dims = var_181, x = zero_mean_sq_1_cast_fp16)[name = tensor("op_196_cast_fp16")]; + tensor var_197_to_fp16 = const()[name = tensor("op_197_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_198_cast_fp16 = add(x = var_196_cast_fp16, y = var_197_to_fp16)[name = tensor("op_198_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_198_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(14658944)))]; + 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(14661568)))]; + 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(14664192)))]; + 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(14666816)))]; + 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_216 = const()[name = tensor("op_216"), val = tensor([1, 1])]; + tensor var_218 = const()[name = tensor("op_218"), 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(14669440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15488704))), 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(15488832)))]; + tensor pretrained_out_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_218, groups = var_180, pad = pretrained_out_1_pad_0, pad_type = pretrained_out_1_pad_type_0, strides = var_216, 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_222 = const()[name = tensor("op_222"), val = tensor([1, 1])]; + tensor var_224 = const()[name = tensor("op_224"), 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(15491456)))]; + tensor input_1_cast_fp16 = conv(dilations = var_224, groups = var_180, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = var_222, weight = layers_0_self_attn_q_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor var_228 = const()[name = tensor("op_228"), val = tensor([1, 1])]; + tensor var_230 = const()[name = tensor("op_230"), 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(15532480)))]; + tensor lora_out_3_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_230, groups = var_180, pad = lora_out_1_pad_0, pad_type = lora_out_1_pad_type_0, strides = var_228, 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_240 = const()[name = tensor("op_240"), val = tensor([1, 1])]; + tensor var_242 = const()[name = tensor("op_242"), 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(15573504))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16392768))), 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_242, groups = var_180, pad = pretrained_out_3_pad_0, pad_type = pretrained_out_3_pad_type_0, strides = var_240, 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_246 = const()[name = tensor("op_246"), val = tensor([1, 1])]; + tensor var_248 = const()[name = tensor("op_248"), 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(16392896)))]; + tensor input_3_cast_fp16 = conv(dilations = var_248, groups = var_180, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = var_246, weight = layers_0_self_attn_k_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor var_252 = const()[name = tensor("op_252"), val = tensor([1, 1])]; + tensor var_254 = const()[name = tensor("op_254"), 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(16433920)))]; + tensor lora_out_7_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_254, groups = var_180, pad = lora_out_5_pad_0, pad_type = lora_out_5_pad_type_0, strides = var_252, weight = lora_out_7_weight_0_to_fp16, x = input_3_cast_fp16)[name = tensor("lora_out_7_cast_fp16")]; + tensor key_1_cast_fp16 = add(x = pretrained_out_3_cast_fp16, y = lora_out_7_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_265 = const()[name = tensor("op_265"), val = tensor([1, 1])]; + tensor var_267 = const()[name = tensor("op_267"), 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(16474944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17294208))), 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(17294336)))]; + tensor pretrained_out_5_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_267, groups = var_180, pad = pretrained_out_5_pad_0, pad_type = pretrained_out_5_pad_type_0, strides = var_265, 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_271 = const()[name = tensor("op_271"), val = tensor([1, 1])]; + tensor var_273 = const()[name = tensor("op_273"), 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(17296960)))]; + tensor input_5_cast_fp16 = conv(dilations = var_273, groups = var_180, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = var_271, weight = layers_0_self_attn_v_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor var_277 = const()[name = tensor("op_277"), val = tensor([1, 1])]; + tensor var_279 = const()[name = tensor("op_279"), 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(17337984)))]; + tensor lora_out_11_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_279, groups = var_180, pad = lora_out_9_pad_0, pad_type = lora_out_9_pad_type_0, strides = var_277, weight = lora_out_11_weight_0_to_fp16, x = input_5_cast_fp16)[name = tensor("lora_out_11_cast_fp16")]; + tensor value_1_cast_fp16 = add(x = pretrained_out_5_cast_fp16, y = lora_out_11_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_286 = const()[name = tensor("op_286"), val = tensor([1, 20, 64, -1])]; + tensor var_287_cast_fp16 = reshape(shape = var_286, x = query_1_cast_fp16)[name = tensor("op_287_cast_fp16")]; + tensor var_288_to_fp16 = const()[name = tensor("op_288_to_fp16"), val = tensor(0x1p-3)]; + tensor var_289_cast_fp16 = mul(x = var_287_cast_fp16, y = var_288_to_fp16)[name = tensor("op_289_cast_fp16")]; + tensor var_290 = const()[name = tensor("op_290"), val = tensor([1, 20, 64, -1])]; + tensor var_291_cast_fp16 = reshape(shape = var_290, x = key_1_cast_fp16)[name = tensor("op_291_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_289_cast_fp16, y = var_291_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor var_294_cast_fp16 = softmax(axis = var_178, x = mh_w_1_cast_fp16)[name = tensor("op_294_cast_fp16")]; + tensor var_295 = const()[name = tensor("op_295"), val = tensor([1, 20, 64, -1])]; + tensor var_296_cast_fp16 = reshape(shape = var_295, x = value_1_cast_fp16)[name = tensor("op_296_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_296_cast_fp16, y = var_294_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_299 = const()[name = tensor("op_299"), val = tensor([1, 1280, 1, -1])]; + tensor input_7_cast_fp16 = reshape(shape = var_299, x = attn_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor var_306 = const()[name = tensor("op_306"), val = tensor([1, 1])]; + tensor var_308 = const()[name = tensor("op_308"), 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(17379008))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18198272))), 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(18198400)))]; + tensor pretrained_out_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_308, groups = var_180, pad = pretrained_out_7_pad_0, pad_type = pretrained_out_7_pad_type_0, strides = var_306, 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_312 = const()[name = tensor("op_312"), val = tensor([1, 1])]; + tensor var_314 = const()[name = tensor("op_314"), 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(18201024)))]; + tensor input_9_cast_fp16 = conv(dilations = var_314, groups = var_180, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = var_312, weight = layers_0_self_attn_o_proj_loraA_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor var_318 = const()[name = tensor("op_318"), val = tensor([1, 1])]; + tensor var_320 = const()[name = tensor("op_320"), 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(18242048)))]; + tensor lora_out_15_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_320, groups = var_180, pad = lora_out_13_pad_0, pad_type = lora_out_13_pad_type_0, strides = var_318, weight = lora_out_15_weight_0_to_fp16, x = input_9_cast_fp16)[name = tensor("lora_out_15_cast_fp16")]; + tensor obj_3_cast_fp16 = add(x = pretrained_out_7_cast_fp16, y = lora_out_15_cast_fp16)[name = tensor("obj_3_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_3_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor var_329 = const()[name = tensor("op_329"), val = tensor([1])]; + tensor channels_mean_3_cast_fp16 = reduce_mean(axes = var_329, keep_dims = var_181, 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_333 = const()[name = tensor("op_333"), val = tensor([1])]; + tensor var_334_cast_fp16 = reduce_mean(axes = var_333, keep_dims = var_181, x = zero_mean_sq_3_cast_fp16)[name = tensor("op_334_cast_fp16")]; + tensor var_335_to_fp16 = const()[name = tensor("op_335_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_336_cast_fp16 = add(x = var_334_cast_fp16, y = var_335_to_fp16)[name = tensor("op_336_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_336_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 input_11_gamma_0_to_fp16 = const()[name = tensor("input_11_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18283072)))]; + tensor input_11_beta_0_to_fp16 = const()[name = tensor("input_11_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18285696)))]; + tensor input_11_epsilon_0_to_fp16 = const()[name = tensor("input_11_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = batch_norm(beta = input_11_beta_0_to_fp16, epsilon = input_11_epsilon_0_to_fp16, gamma = input_11_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("input_11_cast_fp16")]; + tensor var_350 = const()[name = tensor("op_350"), val = tensor([1, 1])]; + tensor var_352 = const()[name = tensor("op_352"), 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_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18288320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21565184))), 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(21565312)))]; + tensor pretrained_out_9_cast_fp16 = conv(bias = layers_0_fc1_pretrained_bias_to_fp16, dilations = var_352, groups = var_180, pad = pretrained_out_9_pad_0, pad_type = pretrained_out_9_pad_type_0, strides = var_350, weight = layers_0_fc1_pretrained_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("pretrained_out_9_cast_fp16")]; + tensor var_356 = const()[name = tensor("op_356"), val = tensor([1, 1])]; + tensor var_358 = const()[name = tensor("op_358"), 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_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(21575616)))]; + tensor input_13_cast_fp16 = conv(dilations = var_358, groups = var_180, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = var_356, weight = layers_0_fc1_loraA_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor var_362 = const()[name = tensor("op_362"), val = tensor([1, 1])]; + tensor var_364 = const()[name = tensor("op_364"), 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(21616640)))]; + tensor lora_out_19_bias_0_to_fp16 = const()[name = tensor("lora_out_19_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21780544)))]; + tensor lora_out_19_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_364, groups = var_180, pad = lora_out_17_pad_0, pad_type = lora_out_17_pad_type_0, strides = var_362, weight = lora_out_19_weight_0_to_fp16, x = input_13_cast_fp16)[name = tensor("lora_out_19_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = pretrained_out_9_cast_fp16, y = lora_out_19_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_mode_0 = const()[name = tensor("input_17_mode_0"), val = tensor("EXACT")]; + tensor input_17_cast_fp16 = gelu(mode = input_17_mode_0, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 1])]; + tensor var_378 = const()[name = tensor("op_378"), 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_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21790848))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25067712))), 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(25067840)))]; + tensor pretrained_out_11_cast_fp16 = conv(bias = layers_0_fc2_pretrained_bias_to_fp16, dilations = var_378, groups = var_180, pad = pretrained_out_11_pad_0, pad_type = pretrained_out_11_pad_type_0, strides = var_376, weight = layers_0_fc2_pretrained_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = tensor("pretrained_out_11_cast_fp16")]; + tensor var_382 = const()[name = tensor("op_382"), val = tensor([1, 1])]; + tensor var_384 = const()[name = tensor("op_384"), 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_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(25070464)))]; + tensor input_19_cast_fp16 = conv(dilations = var_384, groups = var_180, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = var_382, weight = layers_0_fc2_loraA_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor var_388 = const()[name = tensor("op_388"), val = tensor([1, 1])]; + tensor var_390 = const()[name = tensor("op_390"), 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(25234368)))]; + tensor lora_out_23_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_390, groups = var_180, pad = lora_out_21_pad_0, pad_type = lora_out_21_pad_type_0, strides = var_388, weight = lora_out_23_weight_0_to_fp16, x = input_19_cast_fp16)[name = tensor("lora_out_23_cast_fp16")]; + tensor hidden_states_5_cast_fp16 = add(x = pretrained_out_11_cast_fp16, y = lora_out_23_cast_fp16)[name = tensor("hidden_states_5_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor(3)]; + tensor var_406 = const()[name = tensor("op_406"), val = tensor(1)]; + tensor var_407 = const()[name = tensor("op_407"), val = tensor(true)]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor([1])]; + tensor channels_mean_5_cast_fp16 = reduce_mean(axes = var_417, keep_dims = var_407, 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_421 = const()[name = tensor("op_421"), val = tensor([1])]; + tensor var_422_cast_fp16 = reduce_mean(axes = var_421, keep_dims = var_407, x = zero_mean_sq_5_cast_fp16)[name = tensor("op_422_cast_fp16")]; + tensor var_423_to_fp16 = const()[name = tensor("op_423_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_424_cast_fp16 = add(x = var_422_cast_fp16, y = var_423_to_fp16)[name = tensor("op_424_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_424_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 obj_5_gamma_0_to_fp16 = const()[name = tensor("obj_5_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25275392)))]; + tensor obj_5_beta_0_to_fp16 = const()[name = tensor("obj_5_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25278016)))]; + tensor obj_5_epsilon_0_to_fp16 = const()[name = tensor("obj_5_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_5_cast_fp16 = batch_norm(beta = obj_5_beta_0_to_fp16, epsilon = obj_5_epsilon_0_to_fp16, gamma = obj_5_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("obj_5_cast_fp16")]; + tensor var_442 = const()[name = tensor("op_442"), val = tensor([1, 1])]; + tensor var_444 = const()[name = tensor("op_444"), 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_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(25280640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26099904))), 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(26100032)))]; + tensor pretrained_out_13_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_444, groups = var_406, pad = pretrained_out_13_pad_0, pad_type = pretrained_out_13_pad_type_0, strides = var_442, weight = layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_5_cast_fp16)[name = tensor("pretrained_out_13_cast_fp16")]; + tensor var_448 = const()[name = tensor("op_448"), val = tensor([1, 1])]; + tensor var_450 = const()[name = tensor("op_450"), val = tensor([1, 1])]; + tensor input_21_pad_type_0 = const()[name = tensor("input_21_pad_type_0"), val = tensor("custom")]; + tensor input_21_pad_0 = const()[name = tensor("input_21_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(26102656)))]; + tensor input_21_cast_fp16 = conv(dilations = var_450, groups = var_406, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = var_448, weight = layers_1_self_attn_q_proj_loraA_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor var_454 = const()[name = tensor("op_454"), val = tensor([1, 1])]; + tensor var_456 = const()[name = tensor("op_456"), 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(26143680)))]; + tensor lora_out_27_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_456, groups = var_406, pad = lora_out_25_pad_0, pad_type = lora_out_25_pad_type_0, strides = var_454, weight = lora_out_27_weight_0_to_fp16, x = input_21_cast_fp16)[name = tensor("lora_out_27_cast_fp16")]; + tensor query_3_cast_fp16 = add(x = pretrained_out_13_cast_fp16, y = lora_out_27_cast_fp16)[name = tensor("query_3_cast_fp16")]; + tensor var_466 = const()[name = tensor("op_466"), val = tensor([1, 1])]; + tensor var_468 = const()[name = tensor("op_468"), 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_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(26184704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27003968))), name = tensor("layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_15_cast_fp16 = conv(dilations = var_468, groups = var_406, pad = pretrained_out_15_pad_0, pad_type = pretrained_out_15_pad_type_0, strides = var_466, weight = layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_5_cast_fp16)[name = tensor("pretrained_out_15_cast_fp16")]; + tensor var_472 = const()[name = tensor("op_472"), val = tensor([1, 1])]; + tensor var_474 = const()[name = tensor("op_474"), 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_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(27004096)))]; + tensor input_23_cast_fp16 = conv(dilations = var_474, groups = var_406, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = var_472, weight = layers_1_self_attn_k_proj_loraA_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor var_478 = const()[name = tensor("op_478"), val = tensor([1, 1])]; + tensor var_480 = const()[name = tensor("op_480"), 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(27045120)))]; + tensor lora_out_31_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_480, groups = var_406, pad = lora_out_29_pad_0, pad_type = lora_out_29_pad_type_0, strides = var_478, weight = lora_out_31_weight_0_to_fp16, x = input_23_cast_fp16)[name = tensor("lora_out_31_cast_fp16")]; + tensor key_3_cast_fp16 = add(x = pretrained_out_15_cast_fp16, y = lora_out_31_cast_fp16)[name = tensor("key_3_cast_fp16")]; + tensor var_491 = const()[name = tensor("op_491"), val = tensor([1, 1])]; + tensor var_493 = const()[name = tensor("op_493"), 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_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(27086144))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27905408))), 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(27905536)))]; + tensor pretrained_out_17_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_493, groups = var_406, pad = pretrained_out_17_pad_0, pad_type = pretrained_out_17_pad_type_0, strides = var_491, weight = layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_5_cast_fp16)[name = tensor("pretrained_out_17_cast_fp16")]; + tensor var_497 = const()[name = tensor("op_497"), val = tensor([1, 1])]; + tensor var_499 = const()[name = tensor("op_499"), val = tensor([1, 1])]; + tensor input_25_pad_type_0 = const()[name = tensor("input_25_pad_type_0"), val = tensor("custom")]; + tensor input_25_pad_0 = const()[name = tensor("input_25_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(27908160)))]; + tensor input_25_cast_fp16 = conv(dilations = var_499, groups = var_406, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = var_497, weight = layers_1_self_attn_v_proj_loraA_weight_to_fp16, x = obj_5_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor var_503 = const()[name = tensor("op_503"), val = tensor([1, 1])]; + tensor var_505 = const()[name = tensor("op_505"), 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(27949184)))]; + tensor lora_out_35_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_505, groups = var_406, pad = lora_out_33_pad_0, pad_type = lora_out_33_pad_type_0, strides = var_503, weight = lora_out_35_weight_0_to_fp16, x = input_25_cast_fp16)[name = tensor("lora_out_35_cast_fp16")]; + tensor value_3_cast_fp16 = add(x = pretrained_out_17_cast_fp16, y = lora_out_35_cast_fp16)[name = tensor("value_3_cast_fp16")]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor([1, 20, 64, -1])]; + tensor var_513_cast_fp16 = reshape(shape = var_512, x = query_3_cast_fp16)[name = tensor("op_513_cast_fp16")]; + tensor var_514_to_fp16 = const()[name = tensor("op_514_to_fp16"), val = tensor(0x1p-3)]; + tensor var_515_cast_fp16 = mul(x = var_513_cast_fp16, y = var_514_to_fp16)[name = tensor("op_515_cast_fp16")]; + tensor var_516 = const()[name = tensor("op_516"), val = tensor([1, 20, 64, -1])]; + tensor var_517_cast_fp16 = reshape(shape = var_516, x = key_3_cast_fp16)[name = tensor("op_517_cast_fp16")]; + tensor mh_w_3_transpose_x_0 = const()[name = tensor("mh_w_3_transpose_x_0"), val = tensor(true)]; + tensor mh_w_3_transpose_y_0 = const()[name = tensor("mh_w_3_transpose_y_0"), val = tensor(false)]; + tensor mh_w_3_cast_fp16 = matmul(transpose_x = mh_w_3_transpose_x_0, transpose_y = mh_w_3_transpose_y_0, x = var_515_cast_fp16, y = var_517_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_520_cast_fp16 = softmax(axis = var_404, x = mh_w_3_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor var_521 = const()[name = tensor("op_521"), val = tensor([1, 20, 64, -1])]; + tensor var_522_cast_fp16 = reshape(shape = var_521, x = value_3_cast_fp16)[name = tensor("op_522_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_522_cast_fp16, y = var_520_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_525 = const()[name = tensor("op_525"), val = tensor([1, 1280, 1, -1])]; + tensor input_27_cast_fp16 = reshape(shape = var_525, x = attn_3_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor var_532 = const()[name = tensor("op_532"), val = tensor([1, 1])]; + tensor var_534 = const()[name = tensor("op_534"), 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_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(27990208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28809472))), 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(28809600)))]; + tensor pretrained_out_19_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_534, groups = var_406, pad = pretrained_out_19_pad_0, pad_type = pretrained_out_19_pad_type_0, strides = var_532, weight = layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = tensor("pretrained_out_19_cast_fp16")]; + tensor var_538 = const()[name = tensor("op_538"), val = tensor([1, 1])]; + tensor var_540 = const()[name = tensor("op_540"), 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_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(28812224)))]; + tensor input_29_cast_fp16 = conv(dilations = var_540, groups = var_406, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = var_538, weight = layers_1_self_attn_o_proj_loraA_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor var_544 = const()[name = tensor("op_544"), val = tensor([1, 1])]; + tensor var_546 = const()[name = tensor("op_546"), 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(28853248)))]; + tensor lora_out_39_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_546, groups = var_406, pad = lora_out_37_pad_0, pad_type = lora_out_37_pad_type_0, strides = var_544, weight = lora_out_39_weight_0_to_fp16, x = input_29_cast_fp16)[name = tensor("lora_out_39_cast_fp16")]; + tensor obj_7_cast_fp16 = add(x = pretrained_out_19_cast_fp16, y = lora_out_39_cast_fp16)[name = tensor("obj_7_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor var_555 = const()[name = tensor("op_555"), val = tensor([1])]; + tensor channels_mean_7_cast_fp16 = reduce_mean(axes = var_555, keep_dims = var_407, 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_559 = const()[name = tensor("op_559"), val = tensor([1])]; + tensor var_560_cast_fp16 = reduce_mean(axes = var_559, keep_dims = var_407, x = zero_mean_sq_7_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor var_561_to_fp16 = const()[name = tensor("op_561_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_562_cast_fp16 = add(x = var_560_cast_fp16, y = var_561_to_fp16)[name = tensor("op_562_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_562_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 input_31_gamma_0_to_fp16 = const()[name = tensor("input_31_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28894272)))]; + tensor input_31_beta_0_to_fp16 = const()[name = tensor("input_31_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28896896)))]; + tensor input_31_epsilon_0_to_fp16 = const()[name = tensor("input_31_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_31_cast_fp16 = batch_norm(beta = input_31_beta_0_to_fp16, epsilon = input_31_epsilon_0_to_fp16, gamma = input_31_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("input_31_cast_fp16")]; + tensor var_576 = const()[name = tensor("op_576"), val = tensor([1, 1])]; + tensor var_578 = const()[name = tensor("op_578"), 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_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28899520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32176384))), 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(32176512)))]; + tensor pretrained_out_21_cast_fp16 = conv(bias = layers_1_fc1_pretrained_bias_to_fp16, dilations = var_578, groups = var_406, pad = pretrained_out_21_pad_0, pad_type = pretrained_out_21_pad_type_0, strides = var_576, weight = layers_1_fc1_pretrained_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = tensor("pretrained_out_21_cast_fp16")]; + tensor var_582 = const()[name = tensor("op_582"), val = tensor([1, 1])]; + tensor var_584 = const()[name = tensor("op_584"), 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_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(32186816)))]; + tensor input_33_cast_fp16 = conv(dilations = var_584, groups = var_406, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = var_582, weight = layers_1_fc1_loraA_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor var_588 = const()[name = tensor("op_588"), val = tensor([1, 1])]; + tensor var_590 = const()[name = tensor("op_590"), 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(32227840)))]; + tensor lora_out_43_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_590, groups = var_406, pad = lora_out_41_pad_0, pad_type = lora_out_41_pad_type_0, strides = var_588, weight = lora_out_43_weight_0_to_fp16, x = input_33_cast_fp16)[name = tensor("lora_out_43_cast_fp16")]; + tensor input_35_cast_fp16 = add(x = pretrained_out_21_cast_fp16, y = lora_out_43_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = input_35_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor var_602 = const()[name = tensor("op_602"), val = tensor([1, 1])]; + tensor var_604 = const()[name = tensor("op_604"), 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_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32391744))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35668608))), 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(35668736)))]; + tensor pretrained_out_23_cast_fp16 = conv(bias = layers_1_fc2_pretrained_bias_to_fp16, dilations = var_604, groups = var_406, pad = pretrained_out_23_pad_0, pad_type = pretrained_out_23_pad_type_0, strides = var_602, weight = layers_1_fc2_pretrained_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = tensor("pretrained_out_23_cast_fp16")]; + tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 1])]; + tensor var_610 = const()[name = tensor("op_610"), 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_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(35671360)))]; + tensor input_39_cast_fp16 = conv(dilations = var_610, groups = var_406, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = var_608, weight = layers_1_fc2_loraA_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor var_614 = const()[name = tensor("op_614"), val = tensor([1, 1])]; + tensor var_616 = const()[name = tensor("op_616"), 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(35835264)))]; + tensor lora_out_47_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_616, groups = var_406, pad = lora_out_45_pad_0, pad_type = lora_out_45_pad_type_0, strides = var_614, weight = lora_out_47_weight_0_to_fp16, x = input_39_cast_fp16)[name = tensor("lora_out_47_cast_fp16")]; + tensor hidden_states_7_cast_fp16 = add(x = pretrained_out_23_cast_fp16, y = lora_out_47_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor var_630 = const()[name = tensor("op_630"), val = tensor(3)]; + tensor var_632 = const()[name = tensor("op_632"), val = tensor(1)]; + tensor var_633 = const()[name = tensor("op_633"), val = tensor(true)]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1])]; + tensor channels_mean_9_cast_fp16 = reduce_mean(axes = var_643, keep_dims = var_633, 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_647 = const()[name = tensor("op_647"), val = tensor([1])]; + tensor var_648_cast_fp16 = reduce_mean(axes = var_647, keep_dims = var_633, x = zero_mean_sq_9_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor var_649_to_fp16 = const()[name = tensor("op_649_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_650_cast_fp16 = add(x = var_648_cast_fp16, y = var_649_to_fp16)[name = tensor("op_650_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_650_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_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(35876288)))]; + 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(35878912)))]; + 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_9_cast_fp16)[name = tensor("obj_9_cast_fp16")]; + tensor var_668 = const()[name = tensor("op_668"), val = tensor([1, 1])]; + tensor var_670 = const()[name = tensor("op_670"), 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_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(35881536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36700800))), 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(36700928)))]; + tensor pretrained_out_25_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_670, groups = var_632, pad = pretrained_out_25_pad_0, pad_type = pretrained_out_25_pad_type_0, strides = var_668, weight = layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor("pretrained_out_25_cast_fp16")]; + tensor var_674 = const()[name = tensor("op_674"), val = tensor([1, 1])]; + tensor var_676 = const()[name = tensor("op_676"), 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_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(36703552)))]; + tensor input_41_cast_fp16 = conv(dilations = var_676, groups = var_632, pad = input_41_pad_0, pad_type = input_41_pad_type_0, strides = var_674, weight = layers_2_self_attn_q_proj_loraA_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor var_680 = const()[name = tensor("op_680"), val = tensor([1, 1])]; + tensor var_682 = const()[name = tensor("op_682"), 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(36744576)))]; + tensor lora_out_51_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_682, groups = var_632, pad = lora_out_49_pad_0, pad_type = lora_out_49_pad_type_0, strides = var_680, weight = lora_out_51_weight_0_to_fp16, x = input_41_cast_fp16)[name = tensor("lora_out_51_cast_fp16")]; + tensor query_5_cast_fp16 = add(x = pretrained_out_25_cast_fp16, y = lora_out_51_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor var_692 = const()[name = tensor("op_692"), val = tensor([1, 1])]; + tensor var_694 = const()[name = tensor("op_694"), 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_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(36785600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37604864))), name = tensor("layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_27_cast_fp16 = conv(dilations = var_694, groups = var_632, pad = pretrained_out_27_pad_0, pad_type = pretrained_out_27_pad_type_0, strides = var_692, weight = layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor("pretrained_out_27_cast_fp16")]; + tensor var_698 = const()[name = tensor("op_698"), val = tensor([1, 1])]; + tensor var_700 = const()[name = tensor("op_700"), 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_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(37604992)))]; + tensor input_43_cast_fp16 = conv(dilations = var_700, groups = var_632, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = var_698, weight = layers_2_self_attn_k_proj_loraA_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor var_704 = const()[name = tensor("op_704"), val = tensor([1, 1])]; + tensor var_706 = const()[name = tensor("op_706"), 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(37646016)))]; + tensor lora_out_55_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_706, groups = var_632, pad = lora_out_53_pad_0, pad_type = lora_out_53_pad_type_0, strides = var_704, weight = lora_out_55_weight_0_to_fp16, x = input_43_cast_fp16)[name = tensor("lora_out_55_cast_fp16")]; + tensor key_5_cast_fp16 = add(x = pretrained_out_27_cast_fp16, y = lora_out_55_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_717 = const()[name = tensor("op_717"), val = tensor([1, 1])]; + tensor var_719 = const()[name = tensor("op_719"), 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_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(37687040))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38506304))), 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(38506432)))]; + tensor pretrained_out_29_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_719, groups = var_632, pad = pretrained_out_29_pad_0, pad_type = pretrained_out_29_pad_type_0, strides = var_717, weight = layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor("pretrained_out_29_cast_fp16")]; + tensor var_723 = const()[name = tensor("op_723"), val = tensor([1, 1])]; + tensor var_725 = const()[name = tensor("op_725"), 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_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(38509056)))]; + tensor input_45_cast_fp16 = conv(dilations = var_725, groups = var_632, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = var_723, weight = layers_2_self_attn_v_proj_loraA_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor var_729 = const()[name = tensor("op_729"), val = tensor([1, 1])]; + tensor var_731 = const()[name = tensor("op_731"), 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(38550080)))]; + tensor lora_out_59_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_731, groups = var_632, pad = lora_out_57_pad_0, pad_type = lora_out_57_pad_type_0, strides = var_729, weight = lora_out_59_weight_0_to_fp16, x = input_45_cast_fp16)[name = tensor("lora_out_59_cast_fp16")]; + tensor value_5_cast_fp16 = add(x = pretrained_out_29_cast_fp16, y = lora_out_59_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_738 = const()[name = tensor("op_738"), val = tensor([1, 20, 64, -1])]; + tensor var_739_cast_fp16 = reshape(shape = var_738, x = query_5_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1p-3)]; + tensor var_741_cast_fp16 = mul(x = var_739_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor var_742 = const()[name = tensor("op_742"), val = tensor([1, 20, 64, -1])]; + tensor var_743_cast_fp16 = reshape(shape = var_742, x = key_5_cast_fp16)[name = tensor("op_743_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_741_cast_fp16, y = var_743_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor var_746_cast_fp16 = softmax(axis = var_630, x = mh_w_5_cast_fp16)[name = tensor("op_746_cast_fp16")]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 20, 64, -1])]; + tensor var_748_cast_fp16 = reshape(shape = var_747, x = value_5_cast_fp16)[name = tensor("op_748_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_748_cast_fp16, y = var_746_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_751 = const()[name = tensor("op_751"), val = tensor([1, 1280, 1, -1])]; + tensor input_47_cast_fp16 = reshape(shape = var_751, x = attn_5_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; + tensor var_760 = const()[name = tensor("op_760"), 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_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(38591104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39410368))), 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(39410496)))]; + tensor pretrained_out_31_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_760, groups = var_632, pad = pretrained_out_31_pad_0, pad_type = pretrained_out_31_pad_type_0, strides = var_758, weight = layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_47_cast_fp16)[name = tensor("pretrained_out_31_cast_fp16")]; + tensor var_764 = const()[name = tensor("op_764"), val = tensor([1, 1])]; + tensor var_766 = const()[name = tensor("op_766"), 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_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(39413120)))]; + tensor input_49_cast_fp16 = conv(dilations = var_766, groups = var_632, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = var_764, weight = layers_2_self_attn_o_proj_loraA_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor var_770 = const()[name = tensor("op_770"), val = tensor([1, 1])]; + tensor var_772 = const()[name = tensor("op_772"), 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(39454144)))]; + tensor lora_out_63_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_772, groups = var_632, pad = lora_out_61_pad_0, pad_type = lora_out_61_pad_type_0, strides = var_770, weight = lora_out_63_weight_0_to_fp16, x = input_49_cast_fp16)[name = tensor("lora_out_63_cast_fp16")]; + tensor obj_11_cast_fp16 = add(x = pretrained_out_31_cast_fp16, y = lora_out_63_cast_fp16)[name = tensor("obj_11_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1])]; + tensor channels_mean_11_cast_fp16 = reduce_mean(axes = var_781, keep_dims = var_633, 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_785 = const()[name = tensor("op_785"), val = tensor([1])]; + tensor var_786_cast_fp16 = reduce_mean(axes = var_785, keep_dims = var_633, x = zero_mean_sq_11_cast_fp16)[name = tensor("op_786_cast_fp16")]; + tensor var_787_to_fp16 = const()[name = tensor("op_787_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_788_cast_fp16 = add(x = var_786_cast_fp16, y = var_787_to_fp16)[name = tensor("op_788_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_788_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(39495168)))]; + 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(39497792)))]; + 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_802 = const()[name = tensor("op_802"), val = tensor([1, 1])]; + tensor var_804 = const()[name = tensor("op_804"), 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_2_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39500416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42777280))), 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(42777408)))]; + tensor pretrained_out_33_cast_fp16 = conv(bias = layers_2_fc1_pretrained_bias_to_fp16, dilations = var_804, groups = var_632, pad = pretrained_out_33_pad_0, pad_type = pretrained_out_33_pad_type_0, strides = var_802, weight = layers_2_fc1_pretrained_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = tensor("pretrained_out_33_cast_fp16")]; + tensor var_808 = const()[name = tensor("op_808"), val = tensor([1, 1])]; + tensor var_810 = const()[name = tensor("op_810"), 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_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(42787712)))]; + tensor input_53_cast_fp16 = conv(dilations = var_810, groups = var_632, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = var_808, weight = layers_2_fc1_loraA_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor var_814 = const()[name = tensor("op_814"), val = tensor([1, 1])]; + tensor var_816 = const()[name = tensor("op_816"), 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(42828736)))]; + tensor lora_out_67_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_816, groups = var_632, pad = lora_out_65_pad_0, pad_type = lora_out_65_pad_type_0, strides = var_814, weight = lora_out_67_weight_0_to_fp16, x = input_53_cast_fp16)[name = tensor("lora_out_67_cast_fp16")]; + tensor input_55_cast_fp16 = add(x = pretrained_out_33_cast_fp16, y = lora_out_67_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_828 = const()[name = tensor("op_828"), val = tensor([1, 1])]; + tensor var_830 = const()[name = tensor("op_830"), 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_2_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42992640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46269504))), 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(46269632)))]; + tensor pretrained_out_35_cast_fp16 = conv(bias = layers_2_fc2_pretrained_bias_to_fp16, dilations = var_830, groups = var_632, pad = pretrained_out_35_pad_0, pad_type = pretrained_out_35_pad_type_0, strides = var_828, weight = layers_2_fc2_pretrained_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = tensor("pretrained_out_35_cast_fp16")]; + tensor var_834 = const()[name = tensor("op_834"), val = tensor([1, 1])]; + tensor var_836 = const()[name = tensor("op_836"), 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_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(46272256)))]; + tensor input_59_cast_fp16 = conv(dilations = var_836, groups = var_632, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = var_834, weight = layers_2_fc2_loraA_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor var_840 = const()[name = tensor("op_840"), val = tensor([1, 1])]; + tensor var_842 = const()[name = tensor("op_842"), 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(46436160)))]; + tensor lora_out_71_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_842, groups = var_632, pad = lora_out_69_pad_0, pad_type = lora_out_69_pad_type_0, strides = var_840, weight = lora_out_71_weight_0_to_fp16, x = input_59_cast_fp16)[name = tensor("lora_out_71_cast_fp16")]; + tensor hidden_states_9_cast_fp16 = add(x = pretrained_out_35_cast_fp16, y = lora_out_71_cast_fp16)[name = tensor("hidden_states_9_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_856 = const()[name = tensor("op_856"), val = tensor(3)]; + tensor var_858 = const()[name = tensor("op_858"), val = tensor(1)]; + tensor var_859 = const()[name = tensor("op_859"), val = tensor(true)]; + tensor var_869 = const()[name = tensor("op_869"), val = tensor([1])]; + tensor channels_mean_13_cast_fp16 = reduce_mean(axes = var_869, keep_dims = var_859, 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_873 = const()[name = tensor("op_873"), val = tensor([1])]; + tensor var_874_cast_fp16 = reduce_mean(axes = var_873, keep_dims = var_859, x = zero_mean_sq_13_cast_fp16)[name = tensor("op_874_cast_fp16")]; + tensor var_875_to_fp16 = const()[name = tensor("op_875_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_876_cast_fp16 = add(x = var_874_cast_fp16, y = var_875_to_fp16)[name = tensor("op_876_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_876_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_13_gamma_0_to_fp16 = const()[name = tensor("obj_13_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46477184)))]; + tensor obj_13_beta_0_to_fp16 = const()[name = tensor("obj_13_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46479808)))]; + tensor obj_13_epsilon_0_to_fp16 = const()[name = tensor("obj_13_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_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_13_cast_fp16")]; + tensor var_894 = const()[name = tensor("op_894"), val = tensor([1, 1])]; + tensor var_896 = const()[name = tensor("op_896"), 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_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(46482432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47301696))), 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(47301824)))]; + tensor pretrained_out_37_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_896, groups = var_858, pad = pretrained_out_37_pad_0, pad_type = pretrained_out_37_pad_type_0, strides = var_894, weight = layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor("pretrained_out_37_cast_fp16")]; + tensor var_900 = const()[name = tensor("op_900"), val = tensor([1, 1])]; + tensor var_902 = const()[name = tensor("op_902"), 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_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(47304448)))]; + tensor input_61_cast_fp16 = conv(dilations = var_902, groups = var_858, pad = input_61_pad_0, pad_type = input_61_pad_type_0, strides = var_900, weight = layers_3_self_attn_q_proj_loraA_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor var_906 = const()[name = tensor("op_906"), val = tensor([1, 1])]; + tensor var_908 = const()[name = tensor("op_908"), 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(47345472)))]; + tensor lora_out_75_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_908, groups = var_858, pad = lora_out_73_pad_0, pad_type = lora_out_73_pad_type_0, strides = var_906, weight = lora_out_75_weight_0_to_fp16, x = input_61_cast_fp16)[name = tensor("lora_out_75_cast_fp16")]; + tensor query_7_cast_fp16 = add(x = pretrained_out_37_cast_fp16, y = lora_out_75_cast_fp16)[name = tensor("query_7_cast_fp16")]; + tensor var_918 = const()[name = tensor("op_918"), val = tensor([1, 1])]; + tensor var_920 = const()[name = tensor("op_920"), 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_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(47386496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48205760))), name = tensor("layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_39_cast_fp16 = conv(dilations = var_920, groups = var_858, pad = pretrained_out_39_pad_0, pad_type = pretrained_out_39_pad_type_0, strides = var_918, weight = layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor("pretrained_out_39_cast_fp16")]; + tensor var_924 = const()[name = tensor("op_924"), val = tensor([1, 1])]; + tensor var_926 = const()[name = tensor("op_926"), 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_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(48205888)))]; + tensor input_63_cast_fp16 = conv(dilations = var_926, groups = var_858, pad = input_63_pad_0, pad_type = input_63_pad_type_0, strides = var_924, weight = layers_3_self_attn_k_proj_loraA_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor([1, 1])]; + tensor var_932 = const()[name = tensor("op_932"), 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(48246912)))]; + tensor lora_out_79_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_932, groups = var_858, pad = lora_out_77_pad_0, pad_type = lora_out_77_pad_type_0, strides = var_930, weight = lora_out_79_weight_0_to_fp16, x = input_63_cast_fp16)[name = tensor("lora_out_79_cast_fp16")]; + tensor key_7_cast_fp16 = add(x = pretrained_out_39_cast_fp16, y = lora_out_79_cast_fp16)[name = tensor("key_7_cast_fp16")]; + tensor var_943 = const()[name = tensor("op_943"), val = tensor([1, 1])]; + tensor var_945 = const()[name = tensor("op_945"), 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_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(48287936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49107200))), 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(49107328)))]; + tensor pretrained_out_41_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_945, groups = var_858, pad = pretrained_out_41_pad_0, pad_type = pretrained_out_41_pad_type_0, strides = var_943, weight = layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_13_cast_fp16)[name = tensor("pretrained_out_41_cast_fp16")]; + tensor var_949 = const()[name = tensor("op_949"), val = tensor([1, 1])]; + tensor var_951 = const()[name = tensor("op_951"), 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_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(49109952)))]; + tensor input_65_cast_fp16 = conv(dilations = var_951, groups = var_858, pad = input_65_pad_0, pad_type = input_65_pad_type_0, strides = var_949, weight = layers_3_self_attn_v_proj_loraA_weight_to_fp16, x = obj_13_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor var_955 = const()[name = tensor("op_955"), val = tensor([1, 1])]; + tensor var_957 = const()[name = tensor("op_957"), 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(49150976)))]; + tensor lora_out_83_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_957, groups = var_858, pad = lora_out_81_pad_0, pad_type = lora_out_81_pad_type_0, strides = var_955, weight = lora_out_83_weight_0_to_fp16, x = input_65_cast_fp16)[name = tensor("lora_out_83_cast_fp16")]; + tensor value_7_cast_fp16 = add(x = pretrained_out_41_cast_fp16, y = lora_out_83_cast_fp16)[name = tensor("value_7_cast_fp16")]; + tensor var_964 = const()[name = tensor("op_964"), val = tensor([1, 20, 64, -1])]; + tensor var_965_cast_fp16 = reshape(shape = var_964, x = query_7_cast_fp16)[name = tensor("op_965_cast_fp16")]; + tensor var_966_to_fp16 = const()[name = tensor("op_966_to_fp16"), val = tensor(0x1p-3)]; + tensor var_967_cast_fp16 = mul(x = var_965_cast_fp16, y = var_966_to_fp16)[name = tensor("op_967_cast_fp16")]; + tensor var_968 = const()[name = tensor("op_968"), val = tensor([1, 20, 64, -1])]; + tensor var_969_cast_fp16 = reshape(shape = var_968, x = key_7_cast_fp16)[name = tensor("op_969_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_967_cast_fp16, y = var_969_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor var_972_cast_fp16 = softmax(axis = var_856, x = mh_w_7_cast_fp16)[name = tensor("op_972_cast_fp16")]; + tensor var_973 = const()[name = tensor("op_973"), val = tensor([1, 20, 64, -1])]; + tensor var_974_cast_fp16 = reshape(shape = var_973, x = value_7_cast_fp16)[name = tensor("op_974_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_974_cast_fp16, y = var_972_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_977 = const()[name = tensor("op_977"), val = tensor([1, 1280, 1, -1])]; + tensor input_67_cast_fp16 = reshape(shape = var_977, x = attn_7_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor var_984 = const()[name = tensor("op_984"), val = tensor([1, 1])]; + tensor var_986 = const()[name = tensor("op_986"), 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_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(49192000))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50011264))), 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(50011392)))]; + tensor pretrained_out_43_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_986, groups = var_858, pad = pretrained_out_43_pad_0, pad_type = pretrained_out_43_pad_type_0, strides = var_984, weight = layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = tensor("pretrained_out_43_cast_fp16")]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 1])]; + tensor var_992 = const()[name = tensor("op_992"), 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_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(50014016)))]; + tensor input_69_cast_fp16 = conv(dilations = var_992, groups = var_858, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = var_990, weight = layers_3_self_attn_o_proj_loraA_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor var_996 = const()[name = tensor("op_996"), val = tensor([1, 1])]; + tensor var_998 = const()[name = tensor("op_998"), 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(50055040)))]; + tensor lora_out_87_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_998, groups = var_858, pad = lora_out_85_pad_0, pad_type = lora_out_85_pad_type_0, strides = var_996, weight = lora_out_87_weight_0_to_fp16, x = input_69_cast_fp16)[name = tensor("lora_out_87_cast_fp16")]; + tensor obj_15_cast_fp16 = add(x = pretrained_out_43_cast_fp16, y = lora_out_87_cast_fp16)[name = tensor("obj_15_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_15_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor var_1007 = const()[name = tensor("op_1007"), val = tensor([1])]; + tensor channels_mean_15_cast_fp16 = reduce_mean(axes = var_1007, keep_dims = var_859, 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_1011 = const()[name = tensor("op_1011"), val = tensor([1])]; + tensor var_1012_cast_fp16 = reduce_mean(axes = var_1011, keep_dims = var_859, x = zero_mean_sq_15_cast_fp16)[name = tensor("op_1012_cast_fp16")]; + tensor var_1013_to_fp16 = const()[name = tensor("op_1013_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1014_cast_fp16 = add(x = var_1012_cast_fp16, y = var_1013_to_fp16)[name = tensor("op_1014_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_1014_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 input_71_gamma_0_to_fp16 = const()[name = tensor("input_71_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50096064)))]; + tensor input_71_beta_0_to_fp16 = const()[name = tensor("input_71_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50098688)))]; + tensor input_71_epsilon_0_to_fp16 = const()[name = tensor("input_71_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = batch_norm(beta = input_71_beta_0_to_fp16, epsilon = input_71_epsilon_0_to_fp16, gamma = input_71_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("input_71_cast_fp16")]; + tensor var_1028 = const()[name = tensor("op_1028"), val = tensor([1, 1])]; + tensor var_1030 = const()[name = tensor("op_1030"), 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_3_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50101312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53378176))), 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(53378304)))]; + tensor pretrained_out_45_cast_fp16 = conv(bias = layers_3_fc1_pretrained_bias_to_fp16, dilations = var_1030, groups = var_858, pad = pretrained_out_45_pad_0, pad_type = pretrained_out_45_pad_type_0, strides = var_1028, weight = layers_3_fc1_pretrained_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = tensor("pretrained_out_45_cast_fp16")]; + tensor var_1034 = const()[name = tensor("op_1034"), val = tensor([1, 1])]; + tensor var_1036 = const()[name = tensor("op_1036"), 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_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(53388608)))]; + tensor input_73_cast_fp16 = conv(dilations = var_1036, groups = var_858, pad = input_73_pad_0, pad_type = input_73_pad_type_0, strides = var_1034, weight = layers_3_fc1_loraA_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor var_1040 = const()[name = tensor("op_1040"), val = tensor([1, 1])]; + tensor var_1042 = const()[name = tensor("op_1042"), 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(53429632)))]; + tensor lora_out_91_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_1042, groups = var_858, pad = lora_out_89_pad_0, pad_type = lora_out_89_pad_type_0, strides = var_1040, weight = lora_out_91_weight_0_to_fp16, x = input_73_cast_fp16)[name = tensor("lora_out_91_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = pretrained_out_45_cast_fp16, y = lora_out_91_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_mode_0 = const()[name = tensor("input_77_mode_0"), val = tensor("EXACT")]; + tensor input_77_cast_fp16 = gelu(mode = input_77_mode_0, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor var_1054 = const()[name = tensor("op_1054"), val = tensor([1, 1])]; + tensor var_1056 = const()[name = tensor("op_1056"), 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_3_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53593536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56870400))), 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(56870528)))]; + tensor pretrained_out_47_cast_fp16 = conv(bias = layers_3_fc2_pretrained_bias_to_fp16, dilations = var_1056, groups = var_858, pad = pretrained_out_47_pad_0, pad_type = pretrained_out_47_pad_type_0, strides = var_1054, weight = layers_3_fc2_pretrained_weight_to_fp16_palettized, x = input_77_cast_fp16)[name = tensor("pretrained_out_47_cast_fp16")]; + tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1])]; + tensor var_1062 = const()[name = tensor("op_1062"), 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_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(56873152)))]; + tensor input_79_cast_fp16 = conv(dilations = var_1062, groups = var_858, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = var_1060, weight = layers_3_fc2_loraA_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor var_1066 = const()[name = tensor("op_1066"), val = tensor([1, 1])]; + tensor var_1068 = const()[name = tensor("op_1068"), 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(57037056)))]; + tensor lora_out_95_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1068, groups = var_858, pad = lora_out_93_pad_0, pad_type = lora_out_93_pad_type_0, strides = var_1066, weight = lora_out_95_weight_0_to_fp16, x = input_79_cast_fp16)[name = tensor("lora_out_95_cast_fp16")]; + tensor hidden_states_11_cast_fp16 = add(x = pretrained_out_47_cast_fp16, y = lora_out_95_cast_fp16)[name = tensor("hidden_states_11_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor var_1082 = const()[name = tensor("op_1082"), val = tensor(3)]; + tensor var_1084 = const()[name = tensor("op_1084"), val = tensor(1)]; + tensor var_1085 = const()[name = tensor("op_1085"), val = tensor(true)]; + tensor var_1095 = const()[name = tensor("op_1095"), val = tensor([1])]; + tensor channels_mean_17_cast_fp16 = reduce_mean(axes = var_1095, keep_dims = var_1085, 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_1099 = const()[name = tensor("op_1099"), val = tensor([1])]; + tensor var_1100_cast_fp16 = reduce_mean(axes = var_1099, keep_dims = var_1085, x = zero_mean_sq_17_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor var_1101_to_fp16 = const()[name = tensor("op_1101_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1102_cast_fp16 = add(x = var_1100_cast_fp16, y = var_1101_to_fp16)[name = tensor("op_1102_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_1102_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 obj_17_gamma_0_to_fp16 = const()[name = tensor("obj_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57078080)))]; + tensor obj_17_beta_0_to_fp16 = const()[name = tensor("obj_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57080704)))]; + tensor obj_17_epsilon_0_to_fp16 = const()[name = tensor("obj_17_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_17_cast_fp16 = batch_norm(beta = obj_17_beta_0_to_fp16, epsilon = obj_17_epsilon_0_to_fp16, gamma = obj_17_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("obj_17_cast_fp16")]; + tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([1, 1])]; + tensor var_1122 = const()[name = tensor("op_1122"), 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_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(57083328))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57902592))), 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(57902720)))]; + tensor pretrained_out_49_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1122, groups = var_1084, pad = pretrained_out_49_pad_0, pad_type = pretrained_out_49_pad_type_0, strides = var_1120, weight = layers_4_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_17_cast_fp16)[name = tensor("pretrained_out_49_cast_fp16")]; + tensor var_1126 = const()[name = tensor("op_1126"), val = tensor([1, 1])]; + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 1])]; + tensor input_81_pad_type_0 = const()[name = tensor("input_81_pad_type_0"), val = tensor("custom")]; + tensor input_81_pad_0 = const()[name = tensor("input_81_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(57905344)))]; + tensor input_81_cast_fp16 = conv(dilations = var_1128, groups = var_1084, pad = input_81_pad_0, pad_type = input_81_pad_type_0, strides = var_1126, weight = layers_4_self_attn_q_proj_loraA_weight_to_fp16, x = obj_17_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor var_1132 = const()[name = tensor("op_1132"), val = tensor([1, 1])]; + tensor var_1134 = const()[name = tensor("op_1134"), 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(57946368)))]; + tensor lora_out_99_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1134, groups = var_1084, pad = lora_out_97_pad_0, pad_type = lora_out_97_pad_type_0, strides = var_1132, weight = lora_out_99_weight_0_to_fp16, x = input_81_cast_fp16)[name = tensor("lora_out_99_cast_fp16")]; + tensor query_9_cast_fp16 = add(x = pretrained_out_49_cast_fp16, y = lora_out_99_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor var_1144 = const()[name = tensor("op_1144"), val = tensor([1, 1])]; + tensor var_1146 = const()[name = tensor("op_1146"), 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_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(57987392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58806656))), name = tensor("layers_4_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_51_cast_fp16 = conv(dilations = var_1146, groups = var_1084, pad = pretrained_out_51_pad_0, pad_type = pretrained_out_51_pad_type_0, strides = var_1144, weight = layers_4_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_17_cast_fp16)[name = tensor("pretrained_out_51_cast_fp16")]; + tensor var_1150 = const()[name = tensor("op_1150"), val = tensor([1, 1])]; + tensor var_1152 = const()[name = tensor("op_1152"), 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_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(58806784)))]; + tensor input_83_cast_fp16 = conv(dilations = var_1152, groups = var_1084, pad = input_83_pad_0, pad_type = input_83_pad_type_0, strides = var_1150, weight = layers_4_self_attn_k_proj_loraA_weight_to_fp16, x = obj_17_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1])]; + tensor var_1158 = const()[name = tensor("op_1158"), 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(58847808)))]; + tensor lora_out_103_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1158, groups = var_1084, pad = lora_out_101_pad_0, pad_type = lora_out_101_pad_type_0, strides = var_1156, weight = lora_out_103_weight_0_to_fp16, x = input_83_cast_fp16)[name = tensor("lora_out_103_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = pretrained_out_51_cast_fp16, y = lora_out_103_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_1169 = const()[name = tensor("op_1169"), val = tensor([1, 1])]; + tensor var_1171 = const()[name = tensor("op_1171"), 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_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(58888832))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59708096))), 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(59708224)))]; + tensor pretrained_out_53_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1171, groups = var_1084, pad = pretrained_out_53_pad_0, pad_type = pretrained_out_53_pad_type_0, strides = var_1169, weight = layers_4_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_17_cast_fp16)[name = tensor("pretrained_out_53_cast_fp16")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 1])]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([1, 1])]; + tensor input_85_pad_type_0 = const()[name = tensor("input_85_pad_type_0"), val = tensor("custom")]; + tensor input_85_pad_0 = const()[name = tensor("input_85_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(59710848)))]; + tensor input_85_cast_fp16 = conv(dilations = var_1177, groups = var_1084, pad = input_85_pad_0, pad_type = input_85_pad_type_0, strides = var_1175, weight = layers_4_self_attn_v_proj_loraA_weight_to_fp16, x = obj_17_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor var_1181 = const()[name = tensor("op_1181"), val = tensor([1, 1])]; + tensor var_1183 = const()[name = tensor("op_1183"), 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(59751872)))]; + tensor lora_out_107_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1183, groups = var_1084, pad = lora_out_105_pad_0, pad_type = lora_out_105_pad_type_0, strides = var_1181, weight = lora_out_107_weight_0_to_fp16, x = input_85_cast_fp16)[name = tensor("lora_out_107_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = pretrained_out_53_cast_fp16, y = lora_out_107_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 20, 64, -1])]; + tensor var_1191_cast_fp16 = reshape(shape = var_1190, x = query_9_cast_fp16)[name = tensor("op_1191_cast_fp16")]; + tensor var_1192_to_fp16 = const()[name = tensor("op_1192_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1193_cast_fp16 = mul(x = var_1191_cast_fp16, y = var_1192_to_fp16)[name = tensor("op_1193_cast_fp16")]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([1, 20, 64, -1])]; + tensor var_1195_cast_fp16 = reshape(shape = var_1194, x = key_9_cast_fp16)[name = tensor("op_1195_cast_fp16")]; + tensor mh_w_9_transpose_x_0 = const()[name = tensor("mh_w_9_transpose_x_0"), val = tensor(true)]; + tensor mh_w_9_transpose_y_0 = const()[name = tensor("mh_w_9_transpose_y_0"), val = tensor(false)]; + tensor mh_w_9_cast_fp16 = matmul(transpose_x = mh_w_9_transpose_x_0, transpose_y = mh_w_9_transpose_y_0, x = var_1193_cast_fp16, y = var_1195_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_1198_cast_fp16 = softmax(axis = var_1082, x = mh_w_9_cast_fp16)[name = tensor("op_1198_cast_fp16")]; + tensor var_1199 = const()[name = tensor("op_1199"), val = tensor([1, 20, 64, -1])]; + tensor var_1200_cast_fp16 = reshape(shape = var_1199, x = value_9_cast_fp16)[name = tensor("op_1200_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_1200_cast_fp16, y = var_1198_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1203 = const()[name = tensor("op_1203"), val = tensor([1, 1280, 1, -1])]; + tensor input_87_cast_fp16 = reshape(shape = var_1203, x = attn_9_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor var_1210 = const()[name = tensor("op_1210"), val = tensor([1, 1])]; + tensor var_1212 = const()[name = tensor("op_1212"), 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_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(59792896))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60612160))), 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(60612288)))]; + tensor pretrained_out_55_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1212, groups = var_1084, pad = pretrained_out_55_pad_0, pad_type = pretrained_out_55_pad_type_0, strides = var_1210, weight = layers_4_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = tensor("pretrained_out_55_cast_fp16")]; + tensor var_1216 = const()[name = tensor("op_1216"), val = tensor([1, 1])]; + tensor var_1218 = const()[name = tensor("op_1218"), 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_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(60614912)))]; + tensor input_89_cast_fp16 = conv(dilations = var_1218, groups = var_1084, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = var_1216, weight = layers_4_self_attn_o_proj_loraA_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor var_1222 = const()[name = tensor("op_1222"), val = tensor([1, 1])]; + tensor var_1224 = const()[name = tensor("op_1224"), 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(60655936)))]; + tensor lora_out_111_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1224, groups = var_1084, pad = lora_out_109_pad_0, pad_type = lora_out_109_pad_type_0, strides = var_1222, weight = lora_out_111_weight_0_to_fp16, x = input_89_cast_fp16)[name = tensor("lora_out_111_cast_fp16")]; + tensor obj_19_cast_fp16 = add(x = pretrained_out_55_cast_fp16, y = lora_out_111_cast_fp16)[name = tensor("obj_19_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = obj_19_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor var_1233 = const()[name = tensor("op_1233"), val = tensor([1])]; + tensor channels_mean_19_cast_fp16 = reduce_mean(axes = var_1233, keep_dims = var_1085, 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_1237 = const()[name = tensor("op_1237"), val = tensor([1])]; + tensor var_1238_cast_fp16 = reduce_mean(axes = var_1237, keep_dims = var_1085, x = zero_mean_sq_19_cast_fp16)[name = tensor("op_1238_cast_fp16")]; + tensor var_1239_to_fp16 = const()[name = tensor("op_1239_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1240_cast_fp16 = add(x = var_1238_cast_fp16, y = var_1239_to_fp16)[name = tensor("op_1240_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_1240_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 input_91_gamma_0_to_fp16 = const()[name = tensor("input_91_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60696960)))]; + tensor input_91_beta_0_to_fp16 = const()[name = tensor("input_91_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60699584)))]; + tensor input_91_epsilon_0_to_fp16 = const()[name = tensor("input_91_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_91_cast_fp16 = batch_norm(beta = input_91_beta_0_to_fp16, epsilon = input_91_epsilon_0_to_fp16, gamma = input_91_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("input_91_cast_fp16")]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 1])]; + tensor var_1256 = const()[name = tensor("op_1256"), 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_4_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60702208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63979072))), 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(63979200)))]; + tensor pretrained_out_57_cast_fp16 = conv(bias = layers_4_fc1_pretrained_bias_to_fp16, dilations = var_1256, groups = var_1084, pad = pretrained_out_57_pad_0, pad_type = pretrained_out_57_pad_type_0, strides = var_1254, weight = layers_4_fc1_pretrained_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = tensor("pretrained_out_57_cast_fp16")]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([1, 1])]; + tensor var_1262 = const()[name = tensor("op_1262"), 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_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(63989504)))]; + tensor input_93_cast_fp16 = conv(dilations = var_1262, groups = var_1084, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = var_1260, weight = layers_4_fc1_loraA_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor var_1266 = const()[name = tensor("op_1266"), val = tensor([1, 1])]; + tensor var_1268 = const()[name = tensor("op_1268"), 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(64030528)))]; + tensor lora_out_115_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_1268, groups = var_1084, pad = lora_out_113_pad_0, pad_type = lora_out_113_pad_type_0, strides = var_1266, weight = lora_out_115_weight_0_to_fp16, x = input_93_cast_fp16)[name = tensor("lora_out_115_cast_fp16")]; + tensor input_95_cast_fp16 = add(x = pretrained_out_57_cast_fp16, y = lora_out_115_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor input_97_mode_0 = const()[name = tensor("input_97_mode_0"), val = tensor("EXACT")]; + tensor input_97_cast_fp16 = gelu(mode = input_97_mode_0, x = input_95_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([1, 1])]; + tensor var_1282 = const()[name = tensor("op_1282"), 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_4_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64194432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67471296))), 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(67471424)))]; + tensor pretrained_out_59_cast_fp16 = conv(bias = layers_4_fc2_pretrained_bias_to_fp16, dilations = var_1282, groups = var_1084, pad = pretrained_out_59_pad_0, pad_type = pretrained_out_59_pad_type_0, strides = var_1280, weight = layers_4_fc2_pretrained_weight_to_fp16_palettized, x = input_97_cast_fp16)[name = tensor("pretrained_out_59_cast_fp16")]; + tensor var_1286 = const()[name = tensor("op_1286"), val = tensor([1, 1])]; + tensor var_1288 = const()[name = tensor("op_1288"), 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_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(67474048)))]; + tensor input_99_cast_fp16 = conv(dilations = var_1288, groups = var_1084, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = var_1286, weight = layers_4_fc2_loraA_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor var_1292 = const()[name = tensor("op_1292"), val = tensor([1, 1])]; + tensor var_1294 = const()[name = tensor("op_1294"), 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(67637952)))]; + tensor lora_out_119_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1294, groups = var_1084, pad = lora_out_117_pad_0, pad_type = lora_out_117_pad_type_0, strides = var_1292, weight = lora_out_119_weight_0_to_fp16, x = input_99_cast_fp16)[name = tensor("lora_out_119_cast_fp16")]; + tensor hidden_states_13_cast_fp16 = add(x = pretrained_out_59_cast_fp16, y = lora_out_119_cast_fp16)[name = tensor("hidden_states_13_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_1308 = const()[name = tensor("op_1308"), val = tensor(3)]; + tensor var_1310 = const()[name = tensor("op_1310"), val = tensor(1)]; + tensor var_1311 = const()[name = tensor("op_1311"), val = tensor(true)]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor([1])]; + tensor channels_mean_21_cast_fp16 = reduce_mean(axes = var_1321, keep_dims = var_1311, 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_1325 = const()[name = tensor("op_1325"), val = tensor([1])]; + tensor var_1326_cast_fp16 = reduce_mean(axes = var_1325, keep_dims = var_1311, x = zero_mean_sq_21_cast_fp16)[name = tensor("op_1326_cast_fp16")]; + tensor var_1327_to_fp16 = const()[name = tensor("op_1327_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1328_cast_fp16 = add(x = var_1326_cast_fp16, y = var_1327_to_fp16)[name = tensor("op_1328_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_1328_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_21_gamma_0_to_fp16 = const()[name = tensor("obj_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67678976)))]; + tensor obj_21_beta_0_to_fp16 = const()[name = tensor("obj_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67681600)))]; + tensor obj_21_epsilon_0_to_fp16 = const()[name = tensor("obj_21_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_21_cast_fp16 = batch_norm(beta = obj_21_beta_0_to_fp16, epsilon = obj_21_epsilon_0_to_fp16, gamma = obj_21_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_21_cast_fp16")]; + tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([1, 1])]; + tensor var_1348 = const()[name = tensor("op_1348"), 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_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(67684224))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68503488))), 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(68503616)))]; + tensor pretrained_out_61_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1348, groups = var_1310, pad = pretrained_out_61_pad_0, pad_type = pretrained_out_61_pad_type_0, strides = var_1346, weight = layers_5_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = tensor("pretrained_out_61_cast_fp16")]; + tensor var_1352 = const()[name = tensor("op_1352"), val = tensor([1, 1])]; + tensor var_1354 = const()[name = tensor("op_1354"), 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_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(68506240)))]; + tensor input_101_cast_fp16 = conv(dilations = var_1354, groups = var_1310, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = var_1352, weight = layers_5_self_attn_q_proj_loraA_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([1, 1])]; + tensor var_1360 = const()[name = tensor("op_1360"), 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(68547264)))]; + tensor lora_out_123_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1360, groups = var_1310, pad = lora_out_121_pad_0, pad_type = lora_out_121_pad_type_0, strides = var_1358, weight = lora_out_123_weight_0_to_fp16, x = input_101_cast_fp16)[name = tensor("lora_out_123_cast_fp16")]; + tensor query_11_cast_fp16 = add(x = pretrained_out_61_cast_fp16, y = lora_out_123_cast_fp16)[name = tensor("query_11_cast_fp16")]; + tensor var_1370 = const()[name = tensor("op_1370"), val = tensor([1, 1])]; + tensor var_1372 = const()[name = tensor("op_1372"), 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_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(68588288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69407552))), name = tensor("layers_5_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_63_cast_fp16 = conv(dilations = var_1372, groups = var_1310, pad = pretrained_out_63_pad_0, pad_type = pretrained_out_63_pad_type_0, strides = var_1370, weight = layers_5_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = tensor("pretrained_out_63_cast_fp16")]; + tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([1, 1])]; + tensor var_1378 = const()[name = tensor("op_1378"), 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_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(69407680)))]; + tensor input_103_cast_fp16 = conv(dilations = var_1378, groups = var_1310, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = var_1376, weight = layers_5_self_attn_k_proj_loraA_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor([1, 1])]; + tensor var_1384 = const()[name = tensor("op_1384"), 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(69448704)))]; + tensor lora_out_127_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1384, groups = var_1310, pad = lora_out_125_pad_0, pad_type = lora_out_125_pad_type_0, strides = var_1382, weight = lora_out_127_weight_0_to_fp16, x = input_103_cast_fp16)[name = tensor("lora_out_127_cast_fp16")]; + tensor key_11_cast_fp16 = add(x = pretrained_out_63_cast_fp16, y = lora_out_127_cast_fp16)[name = tensor("key_11_cast_fp16")]; + tensor var_1395 = const()[name = tensor("op_1395"), val = tensor([1, 1])]; + tensor var_1397 = const()[name = tensor("op_1397"), 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_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(69489728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70308992))), 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(70309120)))]; + tensor pretrained_out_65_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1397, groups = var_1310, pad = pretrained_out_65_pad_0, pad_type = pretrained_out_65_pad_type_0, strides = var_1395, weight = layers_5_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_21_cast_fp16)[name = tensor("pretrained_out_65_cast_fp16")]; + tensor var_1401 = const()[name = tensor("op_1401"), val = tensor([1, 1])]; + tensor var_1403 = const()[name = tensor("op_1403"), 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_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(70311744)))]; + tensor input_105_cast_fp16 = conv(dilations = var_1403, groups = var_1310, pad = input_105_pad_0, pad_type = input_105_pad_type_0, strides = var_1401, weight = layers_5_self_attn_v_proj_loraA_weight_to_fp16, x = obj_21_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor var_1407 = const()[name = tensor("op_1407"), val = tensor([1, 1])]; + tensor var_1409 = const()[name = tensor("op_1409"), 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(70352768)))]; + tensor lora_out_131_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1409, groups = var_1310, pad = lora_out_129_pad_0, pad_type = lora_out_129_pad_type_0, strides = var_1407, weight = lora_out_131_weight_0_to_fp16, x = input_105_cast_fp16)[name = tensor("lora_out_131_cast_fp16")]; + tensor value_11_cast_fp16 = add(x = pretrained_out_65_cast_fp16, y = lora_out_131_cast_fp16)[name = tensor("value_11_cast_fp16")]; + tensor var_1416 = const()[name = tensor("op_1416"), val = tensor([1, 20, 64, -1])]; + tensor var_1417_cast_fp16 = reshape(shape = var_1416, x = query_11_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor var_1418_to_fp16 = const()[name = tensor("op_1418_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1419_cast_fp16 = mul(x = var_1417_cast_fp16, y = var_1418_to_fp16)[name = tensor("op_1419_cast_fp16")]; + tensor var_1420 = const()[name = tensor("op_1420"), val = tensor([1, 20, 64, -1])]; + tensor var_1421_cast_fp16 = reshape(shape = var_1420, x = key_11_cast_fp16)[name = tensor("op_1421_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_1419_cast_fp16, y = var_1421_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor var_1424_cast_fp16 = softmax(axis = var_1308, x = mh_w_11_cast_fp16)[name = tensor("op_1424_cast_fp16")]; + tensor var_1425 = const()[name = tensor("op_1425"), val = tensor([1, 20, 64, -1])]; + tensor var_1426_cast_fp16 = reshape(shape = var_1425, x = value_11_cast_fp16)[name = tensor("op_1426_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_1426_cast_fp16, y = var_1424_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1429 = const()[name = tensor("op_1429"), val = tensor([1, 1280, 1, -1])]; + tensor input_107_cast_fp16 = reshape(shape = var_1429, x = attn_11_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor var_1436 = const()[name = tensor("op_1436"), val = tensor([1, 1])]; + tensor var_1438 = const()[name = tensor("op_1438"), 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_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(70393792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71213056))), 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(71213184)))]; + tensor pretrained_out_67_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1438, groups = var_1310, pad = pretrained_out_67_pad_0, pad_type = pretrained_out_67_pad_type_0, strides = var_1436, weight = layers_5_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_107_cast_fp16)[name = tensor("pretrained_out_67_cast_fp16")]; + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([1, 1])]; + tensor var_1444 = const()[name = tensor("op_1444"), 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_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(71215808)))]; + tensor input_109_cast_fp16 = conv(dilations = var_1444, groups = var_1310, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = var_1442, weight = layers_5_self_attn_o_proj_loraA_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([1, 1])]; + tensor var_1450 = const()[name = tensor("op_1450"), 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(71256832)))]; + tensor lora_out_135_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1450, groups = var_1310, pad = lora_out_133_pad_0, pad_type = lora_out_133_pad_type_0, strides = var_1448, weight = lora_out_135_weight_0_to_fp16, x = input_109_cast_fp16)[name = tensor("lora_out_135_cast_fp16")]; + tensor obj_23_cast_fp16 = add(x = pretrained_out_67_cast_fp16, y = lora_out_135_cast_fp16)[name = tensor("obj_23_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_23_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([1])]; + tensor channels_mean_23_cast_fp16 = reduce_mean(axes = var_1459, keep_dims = var_1311, 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_1463 = const()[name = tensor("op_1463"), val = tensor([1])]; + tensor var_1464_cast_fp16 = reduce_mean(axes = var_1463, keep_dims = var_1311, x = zero_mean_sq_23_cast_fp16)[name = tensor("op_1464_cast_fp16")]; + tensor var_1465_to_fp16 = const()[name = tensor("op_1465_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1466_cast_fp16 = add(x = var_1464_cast_fp16, y = var_1465_to_fp16)[name = tensor("op_1466_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_1466_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(71297856)))]; + 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(71300480)))]; + 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_1480 = const()[name = tensor("op_1480"), val = tensor([1, 1])]; + tensor var_1482 = const()[name = tensor("op_1482"), 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_5_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71303104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(74579968))), 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(74580096)))]; + tensor pretrained_out_69_cast_fp16 = conv(bias = layers_5_fc1_pretrained_bias_to_fp16, dilations = var_1482, groups = var_1310, pad = pretrained_out_69_pad_0, pad_type = pretrained_out_69_pad_type_0, strides = var_1480, weight = layers_5_fc1_pretrained_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = tensor("pretrained_out_69_cast_fp16")]; + tensor var_1486 = const()[name = tensor("op_1486"), val = tensor([1, 1])]; + tensor var_1488 = const()[name = tensor("op_1488"), 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_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(74590400)))]; + tensor input_113_cast_fp16 = conv(dilations = var_1488, groups = var_1310, pad = input_113_pad_0, pad_type = input_113_pad_type_0, strides = var_1486, weight = layers_5_fc1_loraA_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor var_1492 = const()[name = tensor("op_1492"), val = tensor([1, 1])]; + tensor var_1494 = const()[name = tensor("op_1494"), 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(74631424)))]; + tensor lora_out_139_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_1494, groups = var_1310, pad = lora_out_137_pad_0, pad_type = lora_out_137_pad_type_0, strides = var_1492, weight = lora_out_139_weight_0_to_fp16, x = input_113_cast_fp16)[name = tensor("lora_out_139_cast_fp16")]; + tensor input_115_cast_fp16 = add(x = pretrained_out_69_cast_fp16, y = lora_out_139_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_1506 = const()[name = tensor("op_1506"), val = tensor([1, 1])]; + tensor var_1508 = const()[name = tensor("op_1508"), 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_5_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(74795328))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78072192))), 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(78072320)))]; + tensor pretrained_out_71_cast_fp16 = conv(bias = layers_5_fc2_pretrained_bias_to_fp16, dilations = var_1508, groups = var_1310, pad = pretrained_out_71_pad_0, pad_type = pretrained_out_71_pad_type_0, strides = var_1506, weight = layers_5_fc2_pretrained_weight_to_fp16_palettized, x = input_117_cast_fp16)[name = tensor("pretrained_out_71_cast_fp16")]; + tensor var_1512 = const()[name = tensor("op_1512"), val = tensor([1, 1])]; + tensor var_1514 = const()[name = tensor("op_1514"), 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_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(78074944)))]; + tensor input_119_cast_fp16 = conv(dilations = var_1514, groups = var_1310, pad = input_119_pad_0, pad_type = input_119_pad_type_0, strides = var_1512, weight = layers_5_fc2_loraA_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor var_1518 = const()[name = tensor("op_1518"), val = tensor([1, 1])]; + tensor var_1520 = const()[name = tensor("op_1520"), 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(78238848)))]; + tensor lora_out_143_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1520, groups = var_1310, pad = lora_out_141_pad_0, pad_type = lora_out_141_pad_type_0, strides = var_1518, weight = lora_out_143_weight_0_to_fp16, x = input_119_cast_fp16)[name = tensor("lora_out_143_cast_fp16")]; + tensor hidden_states_15_cast_fp16 = add(x = pretrained_out_71_cast_fp16, y = lora_out_143_cast_fp16)[name = tensor("hidden_states_15_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_1534 = const()[name = tensor("op_1534"), val = tensor(3)]; + tensor var_1536 = const()[name = tensor("op_1536"), val = tensor(1)]; + tensor var_1537 = const()[name = tensor("op_1537"), val = tensor(true)]; + tensor var_1547 = const()[name = tensor("op_1547"), val = tensor([1])]; + tensor channels_mean_25_cast_fp16 = reduce_mean(axes = var_1547, keep_dims = var_1537, 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_1551 = const()[name = tensor("op_1551"), val = tensor([1])]; + tensor var_1552_cast_fp16 = reduce_mean(axes = var_1551, keep_dims = var_1537, x = zero_mean_sq_25_cast_fp16)[name = tensor("op_1552_cast_fp16")]; + tensor var_1553_to_fp16 = const()[name = tensor("op_1553_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1554_cast_fp16 = add(x = var_1552_cast_fp16, y = var_1553_to_fp16)[name = tensor("op_1554_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_1554_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_25_gamma_0_to_fp16 = const()[name = tensor("obj_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78279872)))]; + tensor obj_25_beta_0_to_fp16 = const()[name = tensor("obj_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78282496)))]; + tensor obj_25_epsilon_0_to_fp16 = const()[name = tensor("obj_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_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_25_cast_fp16")]; + tensor var_1572 = const()[name = tensor("op_1572"), val = tensor([1, 1])]; + tensor var_1574 = const()[name = tensor("op_1574"), 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_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(78285120))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79104384))), 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(79104512)))]; + tensor pretrained_out_73_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1574, groups = var_1536, pad = pretrained_out_73_pad_0, pad_type = pretrained_out_73_pad_type_0, strides = var_1572, weight = layers_6_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor("pretrained_out_73_cast_fp16")]; + tensor var_1578 = const()[name = tensor("op_1578"), val = tensor([1, 1])]; + tensor var_1580 = const()[name = tensor("op_1580"), 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_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(79107136)))]; + tensor input_121_cast_fp16 = conv(dilations = var_1580, groups = var_1536, pad = input_121_pad_0, pad_type = input_121_pad_type_0, strides = var_1578, weight = layers_6_self_attn_q_proj_loraA_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor var_1584 = const()[name = tensor("op_1584"), val = tensor([1, 1])]; + tensor var_1586 = const()[name = tensor("op_1586"), 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(79148160)))]; + tensor lora_out_147_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1586, groups = var_1536, pad = lora_out_145_pad_0, pad_type = lora_out_145_pad_type_0, strides = var_1584, weight = lora_out_147_weight_0_to_fp16, x = input_121_cast_fp16)[name = tensor("lora_out_147_cast_fp16")]; + tensor query_13_cast_fp16 = add(x = pretrained_out_73_cast_fp16, y = lora_out_147_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor var_1596 = const()[name = tensor("op_1596"), val = tensor([1, 1])]; + tensor var_1598 = const()[name = tensor("op_1598"), 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_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(79189184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80008448))), name = tensor("layers_6_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_75_cast_fp16 = conv(dilations = var_1598, groups = var_1536, pad = pretrained_out_75_pad_0, pad_type = pretrained_out_75_pad_type_0, strides = var_1596, weight = layers_6_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor("pretrained_out_75_cast_fp16")]; + tensor var_1602 = const()[name = tensor("op_1602"), val = tensor([1, 1])]; + tensor var_1604 = const()[name = tensor("op_1604"), 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_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(80008576)))]; + tensor input_123_cast_fp16 = conv(dilations = var_1604, groups = var_1536, pad = input_123_pad_0, pad_type = input_123_pad_type_0, strides = var_1602, weight = layers_6_self_attn_k_proj_loraA_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor var_1608 = const()[name = tensor("op_1608"), val = tensor([1, 1])]; + tensor var_1610 = const()[name = tensor("op_1610"), 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(80049600)))]; + tensor lora_out_151_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1610, groups = var_1536, pad = lora_out_149_pad_0, pad_type = lora_out_149_pad_type_0, strides = var_1608, weight = lora_out_151_weight_0_to_fp16, x = input_123_cast_fp16)[name = tensor("lora_out_151_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = pretrained_out_75_cast_fp16, y = lora_out_151_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_1621 = const()[name = tensor("op_1621"), val = tensor([1, 1])]; + tensor var_1623 = const()[name = tensor("op_1623"), 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_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(80090624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80909888))), 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(80910016)))]; + tensor pretrained_out_77_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1623, groups = var_1536, pad = pretrained_out_77_pad_0, pad_type = pretrained_out_77_pad_type_0, strides = var_1621, weight = layers_6_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_25_cast_fp16)[name = tensor("pretrained_out_77_cast_fp16")]; + tensor var_1627 = const()[name = tensor("op_1627"), val = tensor([1, 1])]; + tensor var_1629 = const()[name = tensor("op_1629"), 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_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(80912640)))]; + tensor input_125_cast_fp16 = conv(dilations = var_1629, groups = var_1536, pad = input_125_pad_0, pad_type = input_125_pad_type_0, strides = var_1627, weight = layers_6_self_attn_v_proj_loraA_weight_to_fp16, x = obj_25_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor var_1633 = const()[name = tensor("op_1633"), val = tensor([1, 1])]; + tensor var_1635 = const()[name = tensor("op_1635"), 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(80953664)))]; + tensor lora_out_155_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1635, groups = var_1536, pad = lora_out_153_pad_0, pad_type = lora_out_153_pad_type_0, strides = var_1633, weight = lora_out_155_weight_0_to_fp16, x = input_125_cast_fp16)[name = tensor("lora_out_155_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = pretrained_out_77_cast_fp16, y = lora_out_155_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_1642 = const()[name = tensor("op_1642"), val = tensor([1, 20, 64, -1])]; + tensor var_1643_cast_fp16 = reshape(shape = var_1642, x = query_13_cast_fp16)[name = tensor("op_1643_cast_fp16")]; + tensor var_1644_to_fp16 = const()[name = tensor("op_1644_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1645_cast_fp16 = mul(x = var_1643_cast_fp16, y = var_1644_to_fp16)[name = tensor("op_1645_cast_fp16")]; + tensor var_1646 = const()[name = tensor("op_1646"), val = tensor([1, 20, 64, -1])]; + tensor var_1647_cast_fp16 = reshape(shape = var_1646, x = key_13_cast_fp16)[name = tensor("op_1647_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_1645_cast_fp16, y = var_1647_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor var_1650_cast_fp16 = softmax(axis = var_1534, x = mh_w_13_cast_fp16)[name = tensor("op_1650_cast_fp16")]; + tensor var_1651 = const()[name = tensor("op_1651"), val = tensor([1, 20, 64, -1])]; + tensor var_1652_cast_fp16 = reshape(shape = var_1651, x = value_13_cast_fp16)[name = tensor("op_1652_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_1652_cast_fp16, y = var_1650_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1655 = const()[name = tensor("op_1655"), val = tensor([1, 1280, 1, -1])]; + tensor input_127_cast_fp16 = reshape(shape = var_1655, x = attn_13_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 1])]; + tensor var_1664 = const()[name = tensor("op_1664"), 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_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(80994688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81813952))), 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(81814080)))]; + tensor pretrained_out_79_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1664, groups = var_1536, pad = pretrained_out_79_pad_0, pad_type = pretrained_out_79_pad_type_0, strides = var_1662, weight = layers_6_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_127_cast_fp16)[name = tensor("pretrained_out_79_cast_fp16")]; + tensor var_1668 = const()[name = tensor("op_1668"), val = tensor([1, 1])]; + tensor var_1670 = const()[name = tensor("op_1670"), 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_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(81816704)))]; + tensor input_129_cast_fp16 = conv(dilations = var_1670, groups = var_1536, pad = input_129_pad_0, pad_type = input_129_pad_type_0, strides = var_1668, weight = layers_6_self_attn_o_proj_loraA_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor var_1674 = const()[name = tensor("op_1674"), val = tensor([1, 1])]; + tensor var_1676 = const()[name = tensor("op_1676"), 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(81857728)))]; + tensor lora_out_159_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1676, groups = var_1536, pad = lora_out_157_pad_0, pad_type = lora_out_157_pad_type_0, strides = var_1674, weight = lora_out_159_weight_0_to_fp16, x = input_129_cast_fp16)[name = tensor("lora_out_159_cast_fp16")]; + tensor obj_27_cast_fp16 = add(x = pretrained_out_79_cast_fp16, y = lora_out_159_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_27_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor var_1685 = const()[name = tensor("op_1685"), val = tensor([1])]; + tensor channels_mean_27_cast_fp16 = reduce_mean(axes = var_1685, keep_dims = var_1537, 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_1689 = const()[name = tensor("op_1689"), val = tensor([1])]; + tensor var_1690_cast_fp16 = reduce_mean(axes = var_1689, keep_dims = var_1537, x = zero_mean_sq_27_cast_fp16)[name = tensor("op_1690_cast_fp16")]; + tensor var_1691_to_fp16 = const()[name = tensor("op_1691_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1692_cast_fp16 = add(x = var_1690_cast_fp16, y = var_1691_to_fp16)[name = tensor("op_1692_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_1692_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 input_131_gamma_0_to_fp16 = const()[name = tensor("input_131_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81898752)))]; + tensor input_131_beta_0_to_fp16 = const()[name = tensor("input_131_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81901376)))]; + tensor input_131_epsilon_0_to_fp16 = const()[name = tensor("input_131_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_131_cast_fp16 = batch_norm(beta = input_131_beta_0_to_fp16, epsilon = input_131_epsilon_0_to_fp16, gamma = input_131_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("input_131_cast_fp16")]; + tensor var_1706 = const()[name = tensor("op_1706"), val = tensor([1, 1])]; + tensor var_1708 = const()[name = tensor("op_1708"), 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_6_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81904000))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85180864))), 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(85180992)))]; + tensor pretrained_out_81_cast_fp16 = conv(bias = layers_6_fc1_pretrained_bias_to_fp16, dilations = var_1708, groups = var_1536, pad = pretrained_out_81_pad_0, pad_type = pretrained_out_81_pad_type_0, strides = var_1706, weight = layers_6_fc1_pretrained_weight_to_fp16_palettized, x = input_131_cast_fp16)[name = tensor("pretrained_out_81_cast_fp16")]; + tensor var_1712 = const()[name = tensor("op_1712"), val = tensor([1, 1])]; + tensor var_1714 = const()[name = tensor("op_1714"), 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_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(85191296)))]; + tensor input_133_cast_fp16 = conv(dilations = var_1714, groups = var_1536, pad = input_133_pad_0, pad_type = input_133_pad_type_0, strides = var_1712, weight = layers_6_fc1_loraA_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor var_1718 = const()[name = tensor("op_1718"), val = tensor([1, 1])]; + tensor var_1720 = const()[name = tensor("op_1720"), 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(85232320)))]; + tensor lora_out_163_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_1720, groups = var_1536, pad = lora_out_161_pad_0, pad_type = lora_out_161_pad_type_0, strides = var_1718, weight = lora_out_163_weight_0_to_fp16, x = input_133_cast_fp16)[name = tensor("lora_out_163_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = pretrained_out_81_cast_fp16, y = lora_out_163_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_mode_0 = const()[name = tensor("input_137_mode_0"), val = tensor("EXACT")]; + tensor input_137_cast_fp16 = gelu(mode = input_137_mode_0, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor var_1732 = const()[name = tensor("op_1732"), val = tensor([1, 1])]; + tensor var_1734 = const()[name = tensor("op_1734"), 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_6_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85396224))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88673088))), 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(88673216)))]; + tensor pretrained_out_83_cast_fp16 = conv(bias = layers_6_fc2_pretrained_bias_to_fp16, dilations = var_1734, groups = var_1536, pad = pretrained_out_83_pad_0, pad_type = pretrained_out_83_pad_type_0, strides = var_1732, weight = layers_6_fc2_pretrained_weight_to_fp16_palettized, x = input_137_cast_fp16)[name = tensor("pretrained_out_83_cast_fp16")]; + tensor var_1738 = const()[name = tensor("op_1738"), val = tensor([1, 1])]; + tensor var_1740 = const()[name = tensor("op_1740"), 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_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(88675840)))]; + tensor input_139_cast_fp16 = conv(dilations = var_1740, groups = var_1536, pad = input_139_pad_0, pad_type = input_139_pad_type_0, strides = var_1738, weight = layers_6_fc2_loraA_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor var_1744 = const()[name = tensor("op_1744"), val = tensor([1, 1])]; + tensor var_1746 = const()[name = tensor("op_1746"), 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(88839744)))]; + tensor lora_out_167_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1746, groups = var_1536, pad = lora_out_165_pad_0, pad_type = lora_out_165_pad_type_0, strides = var_1744, weight = lora_out_167_weight_0_to_fp16, x = input_139_cast_fp16)[name = tensor("lora_out_167_cast_fp16")]; + tensor hidden_states_17_cast_fp16 = add(x = pretrained_out_83_cast_fp16, y = lora_out_167_cast_fp16)[name = tensor("hidden_states_17_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor var_1760 = const()[name = tensor("op_1760"), val = tensor(3)]; + tensor var_1762 = const()[name = tensor("op_1762"), val = tensor(1)]; + tensor var_1763 = const()[name = tensor("op_1763"), val = tensor(true)]; + tensor var_1773 = const()[name = tensor("op_1773"), val = tensor([1])]; + tensor channels_mean_29_cast_fp16 = reduce_mean(axes = var_1773, keep_dims = var_1763, 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_1777 = const()[name = tensor("op_1777"), val = tensor([1])]; + tensor var_1778_cast_fp16 = reduce_mean(axes = var_1777, keep_dims = var_1763, x = zero_mean_sq_29_cast_fp16)[name = tensor("op_1778_cast_fp16")]; + tensor var_1779_to_fp16 = const()[name = tensor("op_1779_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1780_cast_fp16 = add(x = var_1778_cast_fp16, y = var_1779_to_fp16)[name = tensor("op_1780_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_1780_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 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(88880768)))]; + 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(88883392)))]; + 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_29_cast_fp16)[name = tensor("obj_29_cast_fp16")]; + tensor var_1798 = const()[name = tensor("op_1798"), val = tensor([1, 1])]; + tensor var_1800 = const()[name = tensor("op_1800"), 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_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(88886016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89705280))), 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(89705408)))]; + tensor pretrained_out_85_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_1800, groups = var_1762, pad = pretrained_out_85_pad_0, pad_type = pretrained_out_85_pad_type_0, strides = var_1798, weight = layers_7_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor("pretrained_out_85_cast_fp16")]; + tensor var_1804 = const()[name = tensor("op_1804"), val = tensor([1, 1])]; + tensor var_1806 = const()[name = tensor("op_1806"), val = tensor([1, 1])]; + tensor input_141_pad_type_0 = const()[name = tensor("input_141_pad_type_0"), val = tensor("custom")]; + tensor input_141_pad_0 = const()[name = tensor("input_141_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(89708032)))]; + tensor input_141_cast_fp16 = conv(dilations = var_1806, groups = var_1762, pad = input_141_pad_0, pad_type = input_141_pad_type_0, strides = var_1804, weight = layers_7_self_attn_q_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor var_1810 = const()[name = tensor("op_1810"), val = tensor([1, 1])]; + tensor var_1812 = const()[name = tensor("op_1812"), 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(89749056)))]; + tensor lora_out_171_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1812, groups = var_1762, pad = lora_out_169_pad_0, pad_type = lora_out_169_pad_type_0, strides = var_1810, weight = lora_out_171_weight_0_to_fp16, x = input_141_cast_fp16)[name = tensor("lora_out_171_cast_fp16")]; + tensor query_15_cast_fp16 = add(x = pretrained_out_85_cast_fp16, y = lora_out_171_cast_fp16)[name = tensor("query_15_cast_fp16")]; + tensor var_1822 = const()[name = tensor("op_1822"), val = tensor([1, 1])]; + tensor var_1824 = const()[name = tensor("op_1824"), 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_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(89790080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90609344))), name = tensor("layers_7_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_87_cast_fp16 = conv(dilations = var_1824, groups = var_1762, pad = pretrained_out_87_pad_0, pad_type = pretrained_out_87_pad_type_0, strides = var_1822, weight = layers_7_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor("pretrained_out_87_cast_fp16")]; + tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([1, 1])]; + tensor var_1830 = const()[name = tensor("op_1830"), 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_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(90609472)))]; + tensor input_143_cast_fp16 = conv(dilations = var_1830, groups = var_1762, pad = input_143_pad_0, pad_type = input_143_pad_type_0, strides = var_1828, weight = layers_7_self_attn_k_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor var_1834 = const()[name = tensor("op_1834"), val = tensor([1, 1])]; + tensor var_1836 = const()[name = tensor("op_1836"), 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(90650496)))]; + tensor lora_out_175_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1836, groups = var_1762, pad = lora_out_173_pad_0, pad_type = lora_out_173_pad_type_0, strides = var_1834, weight = lora_out_175_weight_0_to_fp16, x = input_143_cast_fp16)[name = tensor("lora_out_175_cast_fp16")]; + tensor key_15_cast_fp16 = add(x = pretrained_out_87_cast_fp16, y = lora_out_175_cast_fp16)[name = tensor("key_15_cast_fp16")]; + tensor var_1847 = const()[name = tensor("op_1847"), val = tensor([1, 1])]; + tensor var_1849 = const()[name = tensor("op_1849"), 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_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(90691520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91510784))), 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(91510912)))]; + tensor pretrained_out_89_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_1849, groups = var_1762, pad = pretrained_out_89_pad_0, pad_type = pretrained_out_89_pad_type_0, strides = var_1847, weight = layers_7_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor("pretrained_out_89_cast_fp16")]; + tensor var_1853 = const()[name = tensor("op_1853"), val = tensor([1, 1])]; + tensor var_1855 = const()[name = tensor("op_1855"), val = tensor([1, 1])]; + tensor input_145_pad_type_0 = const()[name = tensor("input_145_pad_type_0"), val = tensor("custom")]; + tensor input_145_pad_0 = const()[name = tensor("input_145_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(91513536)))]; + tensor input_145_cast_fp16 = conv(dilations = var_1855, groups = var_1762, pad = input_145_pad_0, pad_type = input_145_pad_type_0, strides = var_1853, weight = layers_7_self_attn_v_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor var_1859 = const()[name = tensor("op_1859"), val = tensor([1, 1])]; + tensor var_1861 = const()[name = tensor("op_1861"), 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(91554560)))]; + tensor lora_out_179_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1861, groups = var_1762, pad = lora_out_177_pad_0, pad_type = lora_out_177_pad_type_0, strides = var_1859, weight = lora_out_179_weight_0_to_fp16, x = input_145_cast_fp16)[name = tensor("lora_out_179_cast_fp16")]; + tensor value_15_cast_fp16 = add(x = pretrained_out_89_cast_fp16, y = lora_out_179_cast_fp16)[name = tensor("value_15_cast_fp16")]; + tensor var_1868 = const()[name = tensor("op_1868"), val = tensor([1, 20, 64, -1])]; + tensor var_1869_cast_fp16 = reshape(shape = var_1868, x = query_15_cast_fp16)[name = tensor("op_1869_cast_fp16")]; + tensor var_1870_to_fp16 = const()[name = tensor("op_1870_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1871_cast_fp16 = mul(x = var_1869_cast_fp16, y = var_1870_to_fp16)[name = tensor("op_1871_cast_fp16")]; + tensor var_1872 = const()[name = tensor("op_1872"), val = tensor([1, 20, 64, -1])]; + tensor var_1873_cast_fp16 = reshape(shape = var_1872, x = key_15_cast_fp16)[name = tensor("op_1873_cast_fp16")]; + tensor mh_w_15_transpose_x_0 = const()[name = tensor("mh_w_15_transpose_x_0"), val = tensor(true)]; + tensor mh_w_15_transpose_y_0 = const()[name = tensor("mh_w_15_transpose_y_0"), val = tensor(false)]; + tensor mh_w_15_cast_fp16 = matmul(transpose_x = mh_w_15_transpose_x_0, transpose_y = mh_w_15_transpose_y_0, x = var_1871_cast_fp16, y = var_1873_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_1876_cast_fp16 = softmax(axis = var_1760, x = mh_w_15_cast_fp16)[name = tensor("op_1876_cast_fp16")]; + tensor var_1877 = const()[name = tensor("op_1877"), val = tensor([1, 20, 64, -1])]; + tensor var_1878_cast_fp16 = reshape(shape = var_1877, x = value_15_cast_fp16)[name = tensor("op_1878_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_1878_cast_fp16, y = var_1876_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_1881 = const()[name = tensor("op_1881"), val = tensor([1, 1280, 1, -1])]; + tensor input_147_cast_fp16 = reshape(shape = var_1881, x = attn_15_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor var_1888 = const()[name = tensor("op_1888"), val = tensor([1, 1])]; + tensor var_1890 = const()[name = tensor("op_1890"), 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_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(91595584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92414848))), 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(92414976)))]; + tensor pretrained_out_91_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_1890, groups = var_1762, pad = pretrained_out_91_pad_0, pad_type = pretrained_out_91_pad_type_0, strides = var_1888, weight = layers_7_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = tensor("pretrained_out_91_cast_fp16")]; + tensor var_1894 = const()[name = tensor("op_1894"), val = tensor([1, 1])]; + tensor var_1896 = const()[name = tensor("op_1896"), 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_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(92417600)))]; + tensor input_149_cast_fp16 = conv(dilations = var_1896, groups = var_1762, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = var_1894, weight = layers_7_self_attn_o_proj_loraA_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor var_1900 = const()[name = tensor("op_1900"), val = tensor([1, 1])]; + tensor var_1902 = const()[name = tensor("op_1902"), 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(92458624)))]; + tensor lora_out_183_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1902, groups = var_1762, pad = lora_out_181_pad_0, pad_type = lora_out_181_pad_type_0, strides = var_1900, weight = lora_out_183_weight_0_to_fp16, x = input_149_cast_fp16)[name = tensor("lora_out_183_cast_fp16")]; + tensor obj_31_cast_fp16 = add(x = pretrained_out_91_cast_fp16, y = lora_out_183_cast_fp16)[name = tensor("obj_31_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = obj_31_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor var_1911 = const()[name = tensor("op_1911"), val = tensor([1])]; + tensor channels_mean_31_cast_fp16 = reduce_mean(axes = var_1911, keep_dims = var_1763, 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_1915 = const()[name = tensor("op_1915"), val = tensor([1])]; + tensor var_1916_cast_fp16 = reduce_mean(axes = var_1915, keep_dims = var_1763, x = zero_mean_sq_31_cast_fp16)[name = tensor("op_1916_cast_fp16")]; + tensor var_1917_to_fp16 = const()[name = tensor("op_1917_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1918_cast_fp16 = add(x = var_1916_cast_fp16, y = var_1917_to_fp16)[name = tensor("op_1918_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_1918_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 input_151_gamma_0_to_fp16 = const()[name = tensor("input_151_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92499648)))]; + tensor input_151_beta_0_to_fp16 = const()[name = tensor("input_151_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92502272)))]; + tensor input_151_epsilon_0_to_fp16 = const()[name = tensor("input_151_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_151_cast_fp16 = batch_norm(beta = input_151_beta_0_to_fp16, epsilon = input_151_epsilon_0_to_fp16, gamma = input_151_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("input_151_cast_fp16")]; + tensor var_1932 = const()[name = tensor("op_1932"), val = tensor([1, 1])]; + tensor var_1934 = const()[name = tensor("op_1934"), 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_7_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92504896))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95781760))), 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(95781888)))]; + tensor pretrained_out_93_cast_fp16 = conv(bias = layers_7_fc1_pretrained_bias_to_fp16, dilations = var_1934, groups = var_1762, pad = pretrained_out_93_pad_0, pad_type = pretrained_out_93_pad_type_0, strides = var_1932, weight = layers_7_fc1_pretrained_weight_to_fp16_palettized, x = input_151_cast_fp16)[name = tensor("pretrained_out_93_cast_fp16")]; + tensor var_1938 = const()[name = tensor("op_1938"), val = tensor([1, 1])]; + tensor var_1940 = const()[name = tensor("op_1940"), 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_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(95792192)))]; + tensor input_153_cast_fp16 = conv(dilations = var_1940, groups = var_1762, pad = input_153_pad_0, pad_type = input_153_pad_type_0, strides = var_1938, weight = layers_7_fc1_loraA_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor var_1944 = const()[name = tensor("op_1944"), val = tensor([1, 1])]; + tensor var_1946 = const()[name = tensor("op_1946"), 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(95833216)))]; + tensor lora_out_187_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_1946, groups = var_1762, pad = lora_out_185_pad_0, pad_type = lora_out_185_pad_type_0, strides = var_1944, weight = lora_out_187_weight_0_to_fp16, x = input_153_cast_fp16)[name = tensor("lora_out_187_cast_fp16")]; + tensor input_155_cast_fp16 = add(x = pretrained_out_93_cast_fp16, y = lora_out_187_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor input_157_mode_0 = const()[name = tensor("input_157_mode_0"), val = tensor("EXACT")]; + tensor input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = input_155_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor var_1958 = const()[name = tensor("op_1958"), val = tensor([1, 1])]; + tensor var_1960 = const()[name = tensor("op_1960"), 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_7_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(95997120))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99273984))), 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(99274112)))]; + tensor pretrained_out_95_cast_fp16 = conv(bias = layers_7_fc2_pretrained_bias_to_fp16, dilations = var_1960, groups = var_1762, pad = pretrained_out_95_pad_0, pad_type = pretrained_out_95_pad_type_0, strides = var_1958, weight = layers_7_fc2_pretrained_weight_to_fp16_palettized, x = input_157_cast_fp16)[name = tensor("pretrained_out_95_cast_fp16")]; + tensor var_1964 = const()[name = tensor("op_1964"), val = tensor([1, 1])]; + tensor var_1966 = const()[name = tensor("op_1966"), 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_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(99276736)))]; + tensor input_159_cast_fp16 = conv(dilations = var_1966, groups = var_1762, pad = input_159_pad_0, pad_type = input_159_pad_type_0, strides = var_1964, weight = layers_7_fc2_loraA_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor var_1970 = const()[name = tensor("op_1970"), val = tensor([1, 1])]; + tensor var_1972 = const()[name = tensor("op_1972"), 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(99440640)))]; + tensor lora_out_191_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_1972, groups = var_1762, pad = lora_out_189_pad_0, pad_type = lora_out_189_pad_type_0, strides = var_1970, weight = lora_out_191_weight_0_to_fp16, x = input_159_cast_fp16)[name = tensor("lora_out_191_cast_fp16")]; + tensor hidden_states_19_cast_fp16 = add(x = pretrained_out_95_cast_fp16, y = lora_out_191_cast_fp16)[name = tensor("hidden_states_19_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor var_1986 = const()[name = tensor("op_1986"), val = tensor(3)]; + tensor var_1988 = const()[name = tensor("op_1988"), val = tensor(1)]; + tensor var_1989 = const()[name = tensor("op_1989"), val = tensor(true)]; + tensor var_1999 = const()[name = tensor("op_1999"), val = tensor([1])]; + tensor channels_mean_33_cast_fp16 = reduce_mean(axes = var_1999, keep_dims = var_1989, 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_2003 = const()[name = tensor("op_2003"), val = tensor([1])]; + tensor var_2004_cast_fp16 = reduce_mean(axes = var_2003, keep_dims = var_1989, x = zero_mean_sq_33_cast_fp16)[name = tensor("op_2004_cast_fp16")]; + tensor var_2005_to_fp16 = const()[name = tensor("op_2005_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2006_cast_fp16 = add(x = var_2004_cast_fp16, y = var_2005_to_fp16)[name = tensor("op_2006_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_2006_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_33_gamma_0_to_fp16 = const()[name = tensor("obj_33_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99481664)))]; + tensor obj_33_beta_0_to_fp16 = const()[name = tensor("obj_33_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99484288)))]; + tensor obj_33_epsilon_0_to_fp16 = const()[name = tensor("obj_33_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_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_33_cast_fp16")]; + tensor var_2024 = const()[name = tensor("op_2024"), val = tensor([1, 1])]; + tensor var_2026 = const()[name = tensor("op_2026"), 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_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(99486912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100306176))), 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(100306304)))]; + tensor pretrained_out_97_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2026, groups = var_1988, pad = pretrained_out_97_pad_0, pad_type = pretrained_out_97_pad_type_0, strides = var_2024, weight = layers_8_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = tensor("pretrained_out_97_cast_fp16")]; + tensor var_2030 = const()[name = tensor("op_2030"), val = tensor([1, 1])]; + tensor var_2032 = const()[name = tensor("op_2032"), 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_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(100308928)))]; + tensor input_161_cast_fp16 = conv(dilations = var_2032, groups = var_1988, pad = input_161_pad_0, pad_type = input_161_pad_type_0, strides = var_2030, weight = layers_8_self_attn_q_proj_loraA_weight_to_fp16, x = obj_33_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor var_2036 = const()[name = tensor("op_2036"), val = tensor([1, 1])]; + tensor var_2038 = const()[name = tensor("op_2038"), 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(100349952)))]; + tensor lora_out_195_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2038, groups = var_1988, pad = lora_out_193_pad_0, pad_type = lora_out_193_pad_type_0, strides = var_2036, weight = lora_out_195_weight_0_to_fp16, x = input_161_cast_fp16)[name = tensor("lora_out_195_cast_fp16")]; + tensor query_17_cast_fp16 = add(x = pretrained_out_97_cast_fp16, y = lora_out_195_cast_fp16)[name = tensor("query_17_cast_fp16")]; + tensor var_2048 = const()[name = tensor("op_2048"), val = tensor([1, 1])]; + tensor var_2050 = const()[name = tensor("op_2050"), 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_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(100390976))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101210240))), name = tensor("layers_8_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_99_cast_fp16 = conv(dilations = var_2050, groups = var_1988, pad = pretrained_out_99_pad_0, pad_type = pretrained_out_99_pad_type_0, strides = var_2048, weight = layers_8_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = tensor("pretrained_out_99_cast_fp16")]; + tensor var_2054 = const()[name = tensor("op_2054"), val = tensor([1, 1])]; + tensor var_2056 = const()[name = tensor("op_2056"), 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_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(101210368)))]; + tensor input_163_cast_fp16 = conv(dilations = var_2056, groups = var_1988, pad = input_163_pad_0, pad_type = input_163_pad_type_0, strides = var_2054, weight = layers_8_self_attn_k_proj_loraA_weight_to_fp16, x = obj_33_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor var_2060 = const()[name = tensor("op_2060"), val = tensor([1, 1])]; + tensor var_2062 = const()[name = tensor("op_2062"), 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(101251392)))]; + tensor lora_out_199_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2062, groups = var_1988, pad = lora_out_197_pad_0, pad_type = lora_out_197_pad_type_0, strides = var_2060, weight = lora_out_199_weight_0_to_fp16, x = input_163_cast_fp16)[name = tensor("lora_out_199_cast_fp16")]; + tensor key_17_cast_fp16 = add(x = pretrained_out_99_cast_fp16, y = lora_out_199_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 1])]; + tensor var_2075 = const()[name = tensor("op_2075"), 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_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(101292416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102111680))), 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(102111808)))]; + tensor pretrained_out_101_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2075, groups = var_1988, pad = pretrained_out_101_pad_0, pad_type = pretrained_out_101_pad_type_0, strides = var_2073, weight = layers_8_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_33_cast_fp16)[name = tensor("pretrained_out_101_cast_fp16")]; + tensor var_2079 = const()[name = tensor("op_2079"), val = tensor([1, 1])]; + tensor var_2081 = const()[name = tensor("op_2081"), 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_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(102114432)))]; + tensor input_165_cast_fp16 = conv(dilations = var_2081, groups = var_1988, pad = input_165_pad_0, pad_type = input_165_pad_type_0, strides = var_2079, weight = layers_8_self_attn_v_proj_loraA_weight_to_fp16, x = obj_33_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor var_2085 = const()[name = tensor("op_2085"), val = tensor([1, 1])]; + tensor var_2087 = const()[name = tensor("op_2087"), 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(102155456)))]; + tensor lora_out_203_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2087, groups = var_1988, pad = lora_out_201_pad_0, pad_type = lora_out_201_pad_type_0, strides = var_2085, weight = lora_out_203_weight_0_to_fp16, x = input_165_cast_fp16)[name = tensor("lora_out_203_cast_fp16")]; + tensor value_17_cast_fp16 = add(x = pretrained_out_101_cast_fp16, y = lora_out_203_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_2094 = const()[name = tensor("op_2094"), val = tensor([1, 20, 64, -1])]; + tensor var_2095_cast_fp16 = reshape(shape = var_2094, x = query_17_cast_fp16)[name = tensor("op_2095_cast_fp16")]; + tensor var_2096_to_fp16 = const()[name = tensor("op_2096_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2097_cast_fp16 = mul(x = var_2095_cast_fp16, y = var_2096_to_fp16)[name = tensor("op_2097_cast_fp16")]; + tensor var_2098 = const()[name = tensor("op_2098"), val = tensor([1, 20, 64, -1])]; + tensor var_2099_cast_fp16 = reshape(shape = var_2098, x = key_17_cast_fp16)[name = tensor("op_2099_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_2097_cast_fp16, y = var_2099_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor var_2102_cast_fp16 = softmax(axis = var_1986, x = mh_w_17_cast_fp16)[name = tensor("op_2102_cast_fp16")]; + tensor var_2103 = const()[name = tensor("op_2103"), val = tensor([1, 20, 64, -1])]; + tensor var_2104_cast_fp16 = reshape(shape = var_2103, x = value_17_cast_fp16)[name = tensor("op_2104_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_2104_cast_fp16, y = var_2102_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2107 = const()[name = tensor("op_2107"), val = tensor([1, 1280, 1, -1])]; + tensor input_167_cast_fp16 = reshape(shape = var_2107, x = attn_17_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor var_2114 = const()[name = tensor("op_2114"), val = tensor([1, 1])]; + tensor var_2116 = const()[name = tensor("op_2116"), 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_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(102196480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103015744))), 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(103015872)))]; + tensor pretrained_out_103_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2116, groups = var_1988, pad = pretrained_out_103_pad_0, pad_type = pretrained_out_103_pad_type_0, strides = var_2114, weight = layers_8_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = tensor("pretrained_out_103_cast_fp16")]; + tensor var_2120 = const()[name = tensor("op_2120"), val = tensor([1, 1])]; + tensor var_2122 = const()[name = tensor("op_2122"), 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_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(103018496)))]; + tensor input_169_cast_fp16 = conv(dilations = var_2122, groups = var_1988, pad = input_169_pad_0, pad_type = input_169_pad_type_0, strides = var_2120, weight = layers_8_self_attn_o_proj_loraA_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor var_2126 = const()[name = tensor("op_2126"), val = tensor([1, 1])]; + tensor var_2128 = const()[name = tensor("op_2128"), 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(103059520)))]; + tensor lora_out_207_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2128, groups = var_1988, pad = lora_out_205_pad_0, pad_type = lora_out_205_pad_type_0, strides = var_2126, weight = lora_out_207_weight_0_to_fp16, x = input_169_cast_fp16)[name = tensor("lora_out_207_cast_fp16")]; + tensor obj_35_cast_fp16 = add(x = pretrained_out_103_cast_fp16, y = lora_out_207_cast_fp16)[name = tensor("obj_35_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor var_2137 = const()[name = tensor("op_2137"), val = tensor([1])]; + tensor channels_mean_35_cast_fp16 = reduce_mean(axes = var_2137, keep_dims = var_1989, 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_2141 = const()[name = tensor("op_2141"), val = tensor([1])]; + tensor var_2142_cast_fp16 = reduce_mean(axes = var_2141, keep_dims = var_1989, x = zero_mean_sq_35_cast_fp16)[name = tensor("op_2142_cast_fp16")]; + tensor var_2143_to_fp16 = const()[name = tensor("op_2143_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2144_cast_fp16 = add(x = var_2142_cast_fp16, y = var_2143_to_fp16)[name = tensor("op_2144_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_2144_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(103100544)))]; + 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(103103168)))]; + 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_2158 = const()[name = tensor("op_2158"), val = tensor([1, 1])]; + tensor var_2160 = const()[name = tensor("op_2160"), 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_8_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103105792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106382656))), 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(106382784)))]; + tensor pretrained_out_105_cast_fp16 = conv(bias = layers_8_fc1_pretrained_bias_to_fp16, dilations = var_2160, groups = var_1988, pad = pretrained_out_105_pad_0, pad_type = pretrained_out_105_pad_type_0, strides = var_2158, weight = layers_8_fc1_pretrained_weight_to_fp16_palettized, x = input_171_cast_fp16)[name = tensor("pretrained_out_105_cast_fp16")]; + tensor var_2164 = const()[name = tensor("op_2164"), val = tensor([1, 1])]; + tensor var_2166 = const()[name = tensor("op_2166"), 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_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(106393088)))]; + tensor input_173_cast_fp16 = conv(dilations = var_2166, groups = var_1988, pad = input_173_pad_0, pad_type = input_173_pad_type_0, strides = var_2164, weight = layers_8_fc1_loraA_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor var_2170 = const()[name = tensor("op_2170"), val = tensor([1, 1])]; + tensor var_2172 = const()[name = tensor("op_2172"), 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(106434112)))]; + tensor lora_out_211_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_2172, groups = var_1988, pad = lora_out_209_pad_0, pad_type = lora_out_209_pad_type_0, strides = var_2170, weight = lora_out_211_weight_0_to_fp16, x = input_173_cast_fp16)[name = tensor("lora_out_211_cast_fp16")]; + tensor input_175_cast_fp16 = add(x = pretrained_out_105_cast_fp16, y = lora_out_211_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_2184 = const()[name = tensor("op_2184"), val = tensor([1, 1])]; + tensor var_2186 = const()[name = tensor("op_2186"), 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_8_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106598016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109874880))), 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(109875008)))]; + tensor pretrained_out_107_cast_fp16 = conv(bias = layers_8_fc2_pretrained_bias_to_fp16, dilations = var_2186, groups = var_1988, pad = pretrained_out_107_pad_0, pad_type = pretrained_out_107_pad_type_0, strides = var_2184, weight = layers_8_fc2_pretrained_weight_to_fp16_palettized, x = input_177_cast_fp16)[name = tensor("pretrained_out_107_cast_fp16")]; + tensor var_2190 = const()[name = tensor("op_2190"), val = tensor([1, 1])]; + tensor var_2192 = const()[name = tensor("op_2192"), 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_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(109877632)))]; + tensor input_179_cast_fp16 = conv(dilations = var_2192, groups = var_1988, pad = input_179_pad_0, pad_type = input_179_pad_type_0, strides = var_2190, weight = layers_8_fc2_loraA_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor var_2196 = const()[name = tensor("op_2196"), val = tensor([1, 1])]; + tensor var_2198 = const()[name = tensor("op_2198"), 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(110041536)))]; + tensor lora_out_215_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2198, groups = var_1988, pad = lora_out_213_pad_0, pad_type = lora_out_213_pad_type_0, strides = var_2196, weight = lora_out_215_weight_0_to_fp16, x = input_179_cast_fp16)[name = tensor("lora_out_215_cast_fp16")]; + tensor hidden_states_21_cast_fp16 = add(x = pretrained_out_107_cast_fp16, y = lora_out_215_cast_fp16)[name = tensor("hidden_states_21_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_2212 = const()[name = tensor("op_2212"), val = tensor(3)]; + tensor var_2214 = const()[name = tensor("op_2214"), val = tensor(1)]; + tensor var_2215 = const()[name = tensor("op_2215"), val = tensor(true)]; + tensor var_2225 = const()[name = tensor("op_2225"), val = tensor([1])]; + tensor channels_mean_37_cast_fp16 = reduce_mean(axes = var_2225, keep_dims = var_2215, 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_2229 = const()[name = tensor("op_2229"), val = tensor([1])]; + tensor var_2230_cast_fp16 = reduce_mean(axes = var_2229, keep_dims = var_2215, x = zero_mean_sq_37_cast_fp16)[name = tensor("op_2230_cast_fp16")]; + tensor var_2231_to_fp16 = const()[name = tensor("op_2231_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2232_cast_fp16 = add(x = var_2230_cast_fp16, y = var_2231_to_fp16)[name = tensor("op_2232_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_2232_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_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(110082560)))]; + 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(110085184)))]; + 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_37_cast_fp16)[name = tensor("obj_37_cast_fp16")]; + tensor var_2250 = const()[name = tensor("op_2250"), val = tensor([1, 1])]; + tensor var_2252 = const()[name = tensor("op_2252"), 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_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(110087808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110907072))), 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(110907200)))]; + tensor pretrained_out_109_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2252, groups = var_2214, pad = pretrained_out_109_pad_0, pad_type = pretrained_out_109_pad_type_0, strides = var_2250, weight = layers_9_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor("pretrained_out_109_cast_fp16")]; + tensor var_2256 = const()[name = tensor("op_2256"), val = tensor([1, 1])]; + tensor var_2258 = const()[name = tensor("op_2258"), 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_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(110909824)))]; + tensor input_181_cast_fp16 = conv(dilations = var_2258, groups = var_2214, pad = input_181_pad_0, pad_type = input_181_pad_type_0, strides = var_2256, weight = layers_9_self_attn_q_proj_loraA_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor var_2262 = const()[name = tensor("op_2262"), val = tensor([1, 1])]; + tensor var_2264 = const()[name = tensor("op_2264"), 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(110950848)))]; + tensor lora_out_219_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2264, groups = var_2214, pad = lora_out_217_pad_0, pad_type = lora_out_217_pad_type_0, strides = var_2262, weight = lora_out_219_weight_0_to_fp16, x = input_181_cast_fp16)[name = tensor("lora_out_219_cast_fp16")]; + tensor query_19_cast_fp16 = add(x = pretrained_out_109_cast_fp16, y = lora_out_219_cast_fp16)[name = tensor("query_19_cast_fp16")]; + tensor var_2274 = const()[name = tensor("op_2274"), val = tensor([1, 1])]; + tensor var_2276 = const()[name = tensor("op_2276"), 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_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(110991872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(111811136))), name = tensor("layers_9_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_111_cast_fp16 = conv(dilations = var_2276, groups = var_2214, pad = pretrained_out_111_pad_0, pad_type = pretrained_out_111_pad_type_0, strides = var_2274, weight = layers_9_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor("pretrained_out_111_cast_fp16")]; + tensor var_2280 = const()[name = tensor("op_2280"), val = tensor([1, 1])]; + tensor var_2282 = const()[name = tensor("op_2282"), 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_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(111811264)))]; + tensor input_183_cast_fp16 = conv(dilations = var_2282, groups = var_2214, pad = input_183_pad_0, pad_type = input_183_pad_type_0, strides = var_2280, weight = layers_9_self_attn_k_proj_loraA_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor var_2286 = const()[name = tensor("op_2286"), val = tensor([1, 1])]; + tensor var_2288 = const()[name = tensor("op_2288"), 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(111852288)))]; + tensor lora_out_223_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2288, groups = var_2214, pad = lora_out_221_pad_0, pad_type = lora_out_221_pad_type_0, strides = var_2286, weight = lora_out_223_weight_0_to_fp16, x = input_183_cast_fp16)[name = tensor("lora_out_223_cast_fp16")]; + tensor key_19_cast_fp16 = add(x = pretrained_out_111_cast_fp16, y = lora_out_223_cast_fp16)[name = tensor("key_19_cast_fp16")]; + tensor var_2299 = const()[name = tensor("op_2299"), val = tensor([1, 1])]; + tensor var_2301 = const()[name = tensor("op_2301"), 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_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(111893312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(112712576))), 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(112712704)))]; + tensor pretrained_out_113_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2301, groups = var_2214, pad = pretrained_out_113_pad_0, pad_type = pretrained_out_113_pad_type_0, strides = var_2299, weight = layers_9_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor("pretrained_out_113_cast_fp16")]; + tensor var_2305 = const()[name = tensor("op_2305"), val = tensor([1, 1])]; + tensor var_2307 = const()[name = tensor("op_2307"), 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_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(112715328)))]; + tensor input_185_cast_fp16 = conv(dilations = var_2307, groups = var_2214, pad = input_185_pad_0, pad_type = input_185_pad_type_0, strides = var_2305, weight = layers_9_self_attn_v_proj_loraA_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor var_2311 = const()[name = tensor("op_2311"), val = tensor([1, 1])]; + tensor var_2313 = const()[name = tensor("op_2313"), 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(112756352)))]; + tensor lora_out_227_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2313, groups = var_2214, pad = lora_out_225_pad_0, pad_type = lora_out_225_pad_type_0, strides = var_2311, weight = lora_out_227_weight_0_to_fp16, x = input_185_cast_fp16)[name = tensor("lora_out_227_cast_fp16")]; + tensor value_19_cast_fp16 = add(x = pretrained_out_113_cast_fp16, y = lora_out_227_cast_fp16)[name = tensor("value_19_cast_fp16")]; + tensor var_2320 = const()[name = tensor("op_2320"), val = tensor([1, 20, 64, -1])]; + tensor var_2321_cast_fp16 = reshape(shape = var_2320, x = query_19_cast_fp16)[name = tensor("op_2321_cast_fp16")]; + tensor var_2322_to_fp16 = const()[name = tensor("op_2322_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2323_cast_fp16 = mul(x = var_2321_cast_fp16, y = var_2322_to_fp16)[name = tensor("op_2323_cast_fp16")]; + tensor var_2324 = const()[name = tensor("op_2324"), val = tensor([1, 20, 64, -1])]; + tensor var_2325_cast_fp16 = reshape(shape = var_2324, x = key_19_cast_fp16)[name = tensor("op_2325_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_2323_cast_fp16, y = var_2325_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor var_2328_cast_fp16 = softmax(axis = var_2212, x = mh_w_19_cast_fp16)[name = tensor("op_2328_cast_fp16")]; + tensor var_2329 = const()[name = tensor("op_2329"), val = tensor([1, 20, 64, -1])]; + tensor var_2330_cast_fp16 = reshape(shape = var_2329, x = value_19_cast_fp16)[name = tensor("op_2330_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_2330_cast_fp16, y = var_2328_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2333 = const()[name = tensor("op_2333"), val = tensor([1, 1280, 1, -1])]; + tensor input_187_cast_fp16 = reshape(shape = var_2333, x = attn_19_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor var_2340 = const()[name = tensor("op_2340"), val = tensor([1, 1])]; + tensor var_2342 = const()[name = tensor("op_2342"), 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_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(112797376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113616640))), 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(113616768)))]; + tensor pretrained_out_115_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2342, groups = var_2214, pad = pretrained_out_115_pad_0, pad_type = pretrained_out_115_pad_type_0, strides = var_2340, weight = layers_9_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = tensor("pretrained_out_115_cast_fp16")]; + tensor var_2346 = const()[name = tensor("op_2346"), val = tensor([1, 1])]; + tensor var_2348 = const()[name = tensor("op_2348"), 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_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(113619392)))]; + tensor input_189_cast_fp16 = conv(dilations = var_2348, groups = var_2214, pad = input_189_pad_0, pad_type = input_189_pad_type_0, strides = var_2346, weight = layers_9_self_attn_o_proj_loraA_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor var_2352 = const()[name = tensor("op_2352"), val = tensor([1, 1])]; + tensor var_2354 = const()[name = tensor("op_2354"), 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(113660416)))]; + tensor lora_out_231_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2354, groups = var_2214, pad = lora_out_229_pad_0, pad_type = lora_out_229_pad_type_0, strides = var_2352, weight = lora_out_231_weight_0_to_fp16, x = input_189_cast_fp16)[name = tensor("lora_out_231_cast_fp16")]; + tensor obj_39_cast_fp16 = add(x = pretrained_out_115_cast_fp16, y = lora_out_231_cast_fp16)[name = tensor("obj_39_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor var_2363 = const()[name = tensor("op_2363"), val = tensor([1])]; + tensor channels_mean_39_cast_fp16 = reduce_mean(axes = var_2363, keep_dims = var_2215, 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_2367 = const()[name = tensor("op_2367"), val = tensor([1])]; + tensor var_2368_cast_fp16 = reduce_mean(axes = var_2367, keep_dims = var_2215, x = zero_mean_sq_39_cast_fp16)[name = tensor("op_2368_cast_fp16")]; + tensor var_2369_to_fp16 = const()[name = tensor("op_2369_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2370_cast_fp16 = add(x = var_2368_cast_fp16, y = var_2369_to_fp16)[name = tensor("op_2370_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_2370_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 input_191_gamma_0_to_fp16 = const()[name = tensor("input_191_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113701440)))]; + tensor input_191_beta_0_to_fp16 = const()[name = tensor("input_191_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113704064)))]; + tensor input_191_epsilon_0_to_fp16 = const()[name = tensor("input_191_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_191_cast_fp16 = batch_norm(beta = input_191_beta_0_to_fp16, epsilon = input_191_epsilon_0_to_fp16, gamma = input_191_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("input_191_cast_fp16")]; + tensor var_2384 = const()[name = tensor("op_2384"), val = tensor([1, 1])]; + tensor var_2386 = const()[name = tensor("op_2386"), 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_9_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113706688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116983552))), 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(116983680)))]; + tensor pretrained_out_117_cast_fp16 = conv(bias = layers_9_fc1_pretrained_bias_to_fp16, dilations = var_2386, groups = var_2214, pad = pretrained_out_117_pad_0, pad_type = pretrained_out_117_pad_type_0, strides = var_2384, weight = layers_9_fc1_pretrained_weight_to_fp16_palettized, x = input_191_cast_fp16)[name = tensor("pretrained_out_117_cast_fp16")]; + tensor var_2390 = const()[name = tensor("op_2390"), val = tensor([1, 1])]; + tensor var_2392 = const()[name = tensor("op_2392"), 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_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(116993984)))]; + tensor input_193_cast_fp16 = conv(dilations = var_2392, groups = var_2214, pad = input_193_pad_0, pad_type = input_193_pad_type_0, strides = var_2390, weight = layers_9_fc1_loraA_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor var_2396 = const()[name = tensor("op_2396"), val = tensor([1, 1])]; + tensor var_2398 = const()[name = tensor("op_2398"), 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(117035008)))]; + tensor lora_out_235_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_2398, groups = var_2214, pad = lora_out_233_pad_0, pad_type = lora_out_233_pad_type_0, strides = var_2396, weight = lora_out_235_weight_0_to_fp16, x = input_193_cast_fp16)[name = tensor("lora_out_235_cast_fp16")]; + tensor input_195_cast_fp16 = add(x = pretrained_out_117_cast_fp16, y = lora_out_235_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_mode_0 = const()[name = tensor("input_197_mode_0"), val = tensor("EXACT")]; + tensor input_197_cast_fp16 = gelu(mode = input_197_mode_0, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor var_2410 = const()[name = tensor("op_2410"), val = tensor([1, 1])]; + tensor var_2412 = const()[name = tensor("op_2412"), 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_9_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117198912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120475776))), 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(120475904)))]; + tensor pretrained_out_119_cast_fp16 = conv(bias = layers_9_fc2_pretrained_bias_to_fp16, dilations = var_2412, groups = var_2214, pad = pretrained_out_119_pad_0, pad_type = pretrained_out_119_pad_type_0, strides = var_2410, weight = layers_9_fc2_pretrained_weight_to_fp16_palettized, x = input_197_cast_fp16)[name = tensor("pretrained_out_119_cast_fp16")]; + tensor var_2416 = const()[name = tensor("op_2416"), val = tensor([1, 1])]; + tensor var_2418 = const()[name = tensor("op_2418"), 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_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(120478528)))]; + tensor input_199_cast_fp16 = conv(dilations = var_2418, groups = var_2214, pad = input_199_pad_0, pad_type = input_199_pad_type_0, strides = var_2416, weight = layers_9_fc2_loraA_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor var_2422 = const()[name = tensor("op_2422"), val = tensor([1, 1])]; + tensor var_2424 = const()[name = tensor("op_2424"), 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(120642432)))]; + tensor lora_out_239_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2424, groups = var_2214, pad = lora_out_237_pad_0, pad_type = lora_out_237_pad_type_0, strides = var_2422, weight = lora_out_239_weight_0_to_fp16, x = input_199_cast_fp16)[name = tensor("lora_out_239_cast_fp16")]; + tensor hidden_states_23_cast_fp16 = add(x = pretrained_out_119_cast_fp16, y = lora_out_239_cast_fp16)[name = tensor("hidden_states_23_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_2438 = const()[name = tensor("op_2438"), val = tensor(3)]; + tensor var_2440 = const()[name = tensor("op_2440"), val = tensor(1)]; + tensor var_2441 = const()[name = tensor("op_2441"), val = tensor(true)]; + tensor var_2451 = const()[name = tensor("op_2451"), val = tensor([1])]; + tensor channels_mean_41_cast_fp16 = reduce_mean(axes = var_2451, keep_dims = var_2441, 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_2455 = const()[name = tensor("op_2455"), val = tensor([1])]; + tensor var_2456_cast_fp16 = reduce_mean(axes = var_2455, keep_dims = var_2441, x = zero_mean_sq_41_cast_fp16)[name = tensor("op_2456_cast_fp16")]; + tensor var_2457_to_fp16 = const()[name = tensor("op_2457_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2458_cast_fp16 = add(x = var_2456_cast_fp16, y = var_2457_to_fp16)[name = tensor("op_2458_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_2458_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 obj_41_gamma_0_to_fp16 = const()[name = tensor("obj_41_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120683456)))]; + tensor obj_41_beta_0_to_fp16 = const()[name = tensor("obj_41_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120686080)))]; + tensor obj_41_epsilon_0_to_fp16 = const()[name = tensor("obj_41_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_41_cast_fp16 = batch_norm(beta = obj_41_beta_0_to_fp16, epsilon = obj_41_epsilon_0_to_fp16, gamma = obj_41_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("obj_41_cast_fp16")]; + tensor var_2476 = const()[name = tensor("op_2476"), val = tensor([1, 1])]; + tensor var_2478 = const()[name = tensor("op_2478"), 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_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(120688704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121507968))), 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(121508096)))]; + tensor pretrained_out_121_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2478, groups = var_2440, pad = pretrained_out_121_pad_0, pad_type = pretrained_out_121_pad_type_0, strides = var_2476, weight = layers_10_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_41_cast_fp16)[name = tensor("pretrained_out_121_cast_fp16")]; + tensor var_2482 = const()[name = tensor("op_2482"), val = tensor([1, 1])]; + tensor var_2484 = const()[name = tensor("op_2484"), val = tensor([1, 1])]; + tensor input_201_pad_type_0 = const()[name = tensor("input_201_pad_type_0"), val = tensor("custom")]; + tensor input_201_pad_0 = const()[name = tensor("input_201_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(121510720)))]; + tensor input_201_cast_fp16 = conv(dilations = var_2484, groups = var_2440, pad = input_201_pad_0, pad_type = input_201_pad_type_0, strides = var_2482, weight = layers_10_self_attn_q_proj_loraA_weight_to_fp16, x = obj_41_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor var_2488 = const()[name = tensor("op_2488"), val = tensor([1, 1])]; + tensor var_2490 = const()[name = tensor("op_2490"), 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(121551744)))]; + tensor lora_out_243_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2490, groups = var_2440, pad = lora_out_241_pad_0, pad_type = lora_out_241_pad_type_0, strides = var_2488, weight = lora_out_243_weight_0_to_fp16, x = input_201_cast_fp16)[name = tensor("lora_out_243_cast_fp16")]; + tensor query_21_cast_fp16 = add(x = pretrained_out_121_cast_fp16, y = lora_out_243_cast_fp16)[name = tensor("query_21_cast_fp16")]; + tensor var_2500 = const()[name = tensor("op_2500"), val = tensor([1, 1])]; + tensor var_2502 = const()[name = tensor("op_2502"), 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_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(121592768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122412032))), name = tensor("layers_10_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_123_cast_fp16 = conv(dilations = var_2502, groups = var_2440, pad = pretrained_out_123_pad_0, pad_type = pretrained_out_123_pad_type_0, strides = var_2500, weight = layers_10_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_41_cast_fp16)[name = tensor("pretrained_out_123_cast_fp16")]; + tensor var_2506 = const()[name = tensor("op_2506"), val = tensor([1, 1])]; + tensor var_2508 = const()[name = tensor("op_2508"), 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_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(122412160)))]; + tensor input_203_cast_fp16 = conv(dilations = var_2508, groups = var_2440, pad = input_203_pad_0, pad_type = input_203_pad_type_0, strides = var_2506, weight = layers_10_self_attn_k_proj_loraA_weight_to_fp16, x = obj_41_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor var_2512 = const()[name = tensor("op_2512"), val = tensor([1, 1])]; + tensor var_2514 = const()[name = tensor("op_2514"), 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(122453184)))]; + tensor lora_out_247_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2514, groups = var_2440, pad = lora_out_245_pad_0, pad_type = lora_out_245_pad_type_0, strides = var_2512, weight = lora_out_247_weight_0_to_fp16, x = input_203_cast_fp16)[name = tensor("lora_out_247_cast_fp16")]; + tensor key_21_cast_fp16 = add(x = pretrained_out_123_cast_fp16, y = lora_out_247_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor var_2525 = const()[name = tensor("op_2525"), val = tensor([1, 1])]; + tensor var_2527 = const()[name = tensor("op_2527"), 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_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(122494208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123313472))), 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(123313600)))]; + tensor pretrained_out_125_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2527, groups = var_2440, pad = pretrained_out_125_pad_0, pad_type = pretrained_out_125_pad_type_0, strides = var_2525, weight = layers_10_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_41_cast_fp16)[name = tensor("pretrained_out_125_cast_fp16")]; + tensor var_2531 = const()[name = tensor("op_2531"), val = tensor([1, 1])]; + tensor var_2533 = const()[name = tensor("op_2533"), val = tensor([1, 1])]; + tensor input_205_pad_type_0 = const()[name = tensor("input_205_pad_type_0"), val = tensor("custom")]; + tensor input_205_pad_0 = const()[name = tensor("input_205_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(123316224)))]; + tensor input_205_cast_fp16 = conv(dilations = var_2533, groups = var_2440, pad = input_205_pad_0, pad_type = input_205_pad_type_0, strides = var_2531, weight = layers_10_self_attn_v_proj_loraA_weight_to_fp16, x = obj_41_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor var_2537 = const()[name = tensor("op_2537"), val = tensor([1, 1])]; + tensor var_2539 = const()[name = tensor("op_2539"), 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(123357248)))]; + tensor lora_out_251_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2539, groups = var_2440, pad = lora_out_249_pad_0, pad_type = lora_out_249_pad_type_0, strides = var_2537, weight = lora_out_251_weight_0_to_fp16, x = input_205_cast_fp16)[name = tensor("lora_out_251_cast_fp16")]; + tensor value_21_cast_fp16 = add(x = pretrained_out_125_cast_fp16, y = lora_out_251_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_2546 = const()[name = tensor("op_2546"), val = tensor([1, 20, 64, -1])]; + tensor var_2547_cast_fp16 = reshape(shape = var_2546, x = query_21_cast_fp16)[name = tensor("op_2547_cast_fp16")]; + tensor var_2548_to_fp16 = const()[name = tensor("op_2548_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2549_cast_fp16 = mul(x = var_2547_cast_fp16, y = var_2548_to_fp16)[name = tensor("op_2549_cast_fp16")]; + tensor var_2550 = const()[name = tensor("op_2550"), val = tensor([1, 20, 64, -1])]; + tensor var_2551_cast_fp16 = reshape(shape = var_2550, x = key_21_cast_fp16)[name = tensor("op_2551_cast_fp16")]; + tensor mh_w_21_transpose_x_0 = const()[name = tensor("mh_w_21_transpose_x_0"), val = tensor(true)]; + tensor mh_w_21_transpose_y_0 = const()[name = tensor("mh_w_21_transpose_y_0"), val = tensor(false)]; + tensor mh_w_21_cast_fp16 = matmul(transpose_x = mh_w_21_transpose_x_0, transpose_y = mh_w_21_transpose_y_0, x = var_2549_cast_fp16, y = var_2551_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_2554_cast_fp16 = softmax(axis = var_2438, x = mh_w_21_cast_fp16)[name = tensor("op_2554_cast_fp16")]; + tensor var_2555 = const()[name = tensor("op_2555"), val = tensor([1, 20, 64, -1])]; + tensor var_2556_cast_fp16 = reshape(shape = var_2555, x = value_21_cast_fp16)[name = tensor("op_2556_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_2556_cast_fp16, y = var_2554_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_2559 = const()[name = tensor("op_2559"), val = tensor([1, 1280, 1, -1])]; + tensor input_207_cast_fp16 = reshape(shape = var_2559, x = attn_21_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor var_2566 = const()[name = tensor("op_2566"), val = tensor([1, 1])]; + tensor var_2568 = const()[name = tensor("op_2568"), 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_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(123398272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124217536))), 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(124217664)))]; + tensor pretrained_out_127_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2568, groups = var_2440, pad = pretrained_out_127_pad_0, pad_type = pretrained_out_127_pad_type_0, strides = var_2566, weight = layers_10_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_207_cast_fp16)[name = tensor("pretrained_out_127_cast_fp16")]; + tensor var_2572 = const()[name = tensor("op_2572"), val = tensor([1, 1])]; + tensor var_2574 = const()[name = tensor("op_2574"), 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_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(124220288)))]; + tensor input_209_cast_fp16 = conv(dilations = var_2574, groups = var_2440, pad = input_209_pad_0, pad_type = input_209_pad_type_0, strides = var_2572, weight = layers_10_self_attn_o_proj_loraA_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor var_2578 = const()[name = tensor("op_2578"), val = tensor([1, 1])]; + tensor var_2580 = const()[name = tensor("op_2580"), 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(124261312)))]; + tensor lora_out_255_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2580, groups = var_2440, pad = lora_out_253_pad_0, pad_type = lora_out_253_pad_type_0, strides = var_2578, weight = lora_out_255_weight_0_to_fp16, x = input_209_cast_fp16)[name = tensor("lora_out_255_cast_fp16")]; + tensor obj_43_cast_fp16 = add(x = pretrained_out_127_cast_fp16, y = lora_out_255_cast_fp16)[name = tensor("obj_43_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = obj_43_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor var_2589 = const()[name = tensor("op_2589"), val = tensor([1])]; + tensor channels_mean_43_cast_fp16 = reduce_mean(axes = var_2589, keep_dims = var_2441, 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_2593 = const()[name = tensor("op_2593"), val = tensor([1])]; + tensor var_2594_cast_fp16 = reduce_mean(axes = var_2593, keep_dims = var_2441, x = zero_mean_sq_43_cast_fp16)[name = tensor("op_2594_cast_fp16")]; + tensor var_2595_to_fp16 = const()[name = tensor("op_2595_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2596_cast_fp16 = add(x = var_2594_cast_fp16, y = var_2595_to_fp16)[name = tensor("op_2596_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_2596_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 input_211_gamma_0_to_fp16 = const()[name = tensor("input_211_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124302336)))]; + tensor input_211_beta_0_to_fp16 = const()[name = tensor("input_211_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124304960)))]; + tensor input_211_epsilon_0_to_fp16 = const()[name = tensor("input_211_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_211_cast_fp16 = batch_norm(beta = input_211_beta_0_to_fp16, epsilon = input_211_epsilon_0_to_fp16, gamma = input_211_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("input_211_cast_fp16")]; + tensor var_2610 = const()[name = tensor("op_2610"), val = tensor([1, 1])]; + tensor var_2612 = const()[name = tensor("op_2612"), 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_10_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124307584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127584448))), 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(127584576)))]; + tensor pretrained_out_129_cast_fp16 = conv(bias = layers_10_fc1_pretrained_bias_to_fp16, dilations = var_2612, groups = var_2440, pad = pretrained_out_129_pad_0, pad_type = pretrained_out_129_pad_type_0, strides = var_2610, weight = layers_10_fc1_pretrained_weight_to_fp16_palettized, x = input_211_cast_fp16)[name = tensor("pretrained_out_129_cast_fp16")]; + tensor var_2616 = const()[name = tensor("op_2616"), val = tensor([1, 1])]; + tensor var_2618 = const()[name = tensor("op_2618"), 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_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(127594880)))]; + tensor input_213_cast_fp16 = conv(dilations = var_2618, groups = var_2440, pad = input_213_pad_0, pad_type = input_213_pad_type_0, strides = var_2616, weight = layers_10_fc1_loraA_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor var_2622 = const()[name = tensor("op_2622"), val = tensor([1, 1])]; + tensor var_2624 = const()[name = tensor("op_2624"), 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(127635904)))]; + tensor lora_out_259_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_2624, groups = var_2440, pad = lora_out_257_pad_0, pad_type = lora_out_257_pad_type_0, strides = var_2622, weight = lora_out_259_weight_0_to_fp16, x = input_213_cast_fp16)[name = tensor("lora_out_259_cast_fp16")]; + tensor input_215_cast_fp16 = add(x = pretrained_out_129_cast_fp16, y = lora_out_259_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor input_217_mode_0 = const()[name = tensor("input_217_mode_0"), val = tensor("EXACT")]; + tensor input_217_cast_fp16 = gelu(mode = input_217_mode_0, x = input_215_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor var_2636 = const()[name = tensor("op_2636"), val = tensor([1, 1])]; + tensor var_2638 = const()[name = tensor("op_2638"), 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_10_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127799808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131076672))), 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(131076800)))]; + tensor pretrained_out_131_cast_fp16 = conv(bias = layers_10_fc2_pretrained_bias_to_fp16, dilations = var_2638, groups = var_2440, pad = pretrained_out_131_pad_0, pad_type = pretrained_out_131_pad_type_0, strides = var_2636, weight = layers_10_fc2_pretrained_weight_to_fp16_palettized, x = input_217_cast_fp16)[name = tensor("pretrained_out_131_cast_fp16")]; + tensor var_2642 = const()[name = tensor("op_2642"), val = tensor([1, 1])]; + tensor var_2644 = const()[name = tensor("op_2644"), 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_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(131079424)))]; + tensor input_219_cast_fp16 = conv(dilations = var_2644, groups = var_2440, pad = input_219_pad_0, pad_type = input_219_pad_type_0, strides = var_2642, weight = layers_10_fc2_loraA_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor var_2648 = const()[name = tensor("op_2648"), val = tensor([1, 1])]; + tensor var_2650 = const()[name = tensor("op_2650"), 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(131243328)))]; + tensor lora_out_263_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2650, groups = var_2440, pad = lora_out_261_pad_0, pad_type = lora_out_261_pad_type_0, strides = var_2648, weight = lora_out_263_weight_0_to_fp16, x = input_219_cast_fp16)[name = tensor("lora_out_263_cast_fp16")]; + tensor hidden_states_25_cast_fp16 = add(x = pretrained_out_131_cast_fp16, y = lora_out_263_cast_fp16)[name = tensor("hidden_states_25_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor var_2664 = const()[name = tensor("op_2664"), val = tensor(3)]; + tensor var_2666 = const()[name = tensor("op_2666"), val = tensor(1)]; + tensor var_2667 = const()[name = tensor("op_2667"), val = tensor(true)]; + tensor var_2677 = const()[name = tensor("op_2677"), val = tensor([1])]; + tensor channels_mean_45_cast_fp16 = reduce_mean(axes = var_2677, keep_dims = var_2667, 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_2681 = const()[name = tensor("op_2681"), val = tensor([1])]; + tensor var_2682_cast_fp16 = reduce_mean(axes = var_2681, keep_dims = var_2667, x = zero_mean_sq_45_cast_fp16)[name = tensor("op_2682_cast_fp16")]; + tensor var_2683_to_fp16 = const()[name = tensor("op_2683_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2684_cast_fp16 = add(x = var_2682_cast_fp16, y = var_2683_to_fp16)[name = tensor("op_2684_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_2684_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_45_gamma_0_to_fp16 = const()[name = tensor("obj_45_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131284352)))]; + tensor obj_45_beta_0_to_fp16 = const()[name = tensor("obj_45_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131286976)))]; + tensor obj_45_epsilon_0_to_fp16 = const()[name = tensor("obj_45_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_45_cast_fp16 = batch_norm(beta = obj_45_beta_0_to_fp16, epsilon = obj_45_epsilon_0_to_fp16, gamma = obj_45_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_45_cast_fp16")]; + tensor var_2702 = const()[name = tensor("op_2702"), val = tensor([1, 1])]; + tensor var_2704 = const()[name = tensor("op_2704"), 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_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(131289600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132108864))), 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(132108992)))]; + tensor pretrained_out_133_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2704, groups = var_2666, pad = pretrained_out_133_pad_0, pad_type = pretrained_out_133_pad_type_0, strides = var_2702, weight = layers_11_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = tensor("pretrained_out_133_cast_fp16")]; + tensor var_2708 = const()[name = tensor("op_2708"), val = tensor([1, 1])]; + tensor var_2710 = const()[name = tensor("op_2710"), 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_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(132111616)))]; + tensor input_221_cast_fp16 = conv(dilations = var_2710, groups = var_2666, pad = input_221_pad_0, pad_type = input_221_pad_type_0, strides = var_2708, weight = layers_11_self_attn_q_proj_loraA_weight_to_fp16, x = obj_45_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor var_2714 = const()[name = tensor("op_2714"), val = tensor([1, 1])]; + tensor var_2716 = const()[name = tensor("op_2716"), 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(132152640)))]; + tensor lora_out_267_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2716, groups = var_2666, pad = lora_out_265_pad_0, pad_type = lora_out_265_pad_type_0, strides = var_2714, weight = lora_out_267_weight_0_to_fp16, x = input_221_cast_fp16)[name = tensor("lora_out_267_cast_fp16")]; + tensor query_23_cast_fp16 = add(x = pretrained_out_133_cast_fp16, y = lora_out_267_cast_fp16)[name = tensor("query_23_cast_fp16")]; + tensor var_2726 = const()[name = tensor("op_2726"), val = tensor([1, 1])]; + tensor var_2728 = const()[name = tensor("op_2728"), 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_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(132193664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133012928))), name = tensor("layers_11_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_135_cast_fp16 = conv(dilations = var_2728, groups = var_2666, pad = pretrained_out_135_pad_0, pad_type = pretrained_out_135_pad_type_0, strides = var_2726, weight = layers_11_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = tensor("pretrained_out_135_cast_fp16")]; + tensor var_2732 = const()[name = tensor("op_2732"), val = tensor([1, 1])]; + tensor var_2734 = const()[name = tensor("op_2734"), 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_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(133013056)))]; + tensor input_223_cast_fp16 = conv(dilations = var_2734, groups = var_2666, pad = input_223_pad_0, pad_type = input_223_pad_type_0, strides = var_2732, weight = layers_11_self_attn_k_proj_loraA_weight_to_fp16, x = obj_45_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor var_2738 = const()[name = tensor("op_2738"), val = tensor([1, 1])]; + tensor var_2740 = const()[name = tensor("op_2740"), 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(133054080)))]; + tensor lora_out_271_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2740, groups = var_2666, pad = lora_out_269_pad_0, pad_type = lora_out_269_pad_type_0, strides = var_2738, weight = lora_out_271_weight_0_to_fp16, x = input_223_cast_fp16)[name = tensor("lora_out_271_cast_fp16")]; + tensor key_23_cast_fp16 = add(x = pretrained_out_135_cast_fp16, y = lora_out_271_cast_fp16)[name = tensor("key_23_cast_fp16")]; + tensor var_2751 = const()[name = tensor("op_2751"), val = tensor([1, 1])]; + tensor var_2753 = const()[name = tensor("op_2753"), 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_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(133095104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133914368))), 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(133914496)))]; + tensor pretrained_out_137_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2753, groups = var_2666, pad = pretrained_out_137_pad_0, pad_type = pretrained_out_137_pad_type_0, strides = var_2751, weight = layers_11_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_45_cast_fp16)[name = tensor("pretrained_out_137_cast_fp16")]; + tensor var_2757 = const()[name = tensor("op_2757"), val = tensor([1, 1])]; + tensor var_2759 = const()[name = tensor("op_2759"), 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_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(133917120)))]; + tensor input_225_cast_fp16 = conv(dilations = var_2759, groups = var_2666, pad = input_225_pad_0, pad_type = input_225_pad_type_0, strides = var_2757, weight = layers_11_self_attn_v_proj_loraA_weight_to_fp16, x = obj_45_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor var_2763 = const()[name = tensor("op_2763"), val = tensor([1, 1])]; + tensor var_2765 = const()[name = tensor("op_2765"), 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(133958144)))]; + tensor lora_out_275_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2765, groups = var_2666, pad = lora_out_273_pad_0, pad_type = lora_out_273_pad_type_0, strides = var_2763, weight = lora_out_275_weight_0_to_fp16, x = input_225_cast_fp16)[name = tensor("lora_out_275_cast_fp16")]; + tensor value_23_cast_fp16 = add(x = pretrained_out_137_cast_fp16, y = lora_out_275_cast_fp16)[name = tensor("value_23_cast_fp16")]; + tensor var_2772 = const()[name = tensor("op_2772"), val = tensor([1, 20, 64, -1])]; + tensor var_2773_cast_fp16 = reshape(shape = var_2772, x = query_23_cast_fp16)[name = tensor("op_2773_cast_fp16")]; + tensor var_2774_to_fp16 = const()[name = tensor("op_2774_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2775_cast_fp16 = mul(x = var_2773_cast_fp16, y = var_2774_to_fp16)[name = tensor("op_2775_cast_fp16")]; + tensor var_2776 = const()[name = tensor("op_2776"), val = tensor([1, 20, 64, -1])]; + tensor var_2777_cast_fp16 = reshape(shape = var_2776, x = key_23_cast_fp16)[name = tensor("op_2777_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_2775_cast_fp16, y = var_2777_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor var_2780_cast_fp16 = softmax(axis = var_2664, x = mh_w_23_cast_fp16)[name = tensor("op_2780_cast_fp16")]; + tensor var_2781 = const()[name = tensor("op_2781"), val = tensor([1, 20, 64, -1])]; + tensor var_2782_cast_fp16 = reshape(shape = var_2781, x = value_23_cast_fp16)[name = tensor("op_2782_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_2782_cast_fp16, y = var_2780_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_2785 = const()[name = tensor("op_2785"), val = tensor([1, 1280, 1, -1])]; + tensor input_227_cast_fp16 = reshape(shape = var_2785, x = attn_23_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor var_2792 = const()[name = tensor("op_2792"), val = tensor([1, 1])]; + tensor var_2794 = const()[name = tensor("op_2794"), 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_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(133999168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134818432))), 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(134818560)))]; + tensor pretrained_out_139_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_2794, groups = var_2666, pad = pretrained_out_139_pad_0, pad_type = pretrained_out_139_pad_type_0, strides = var_2792, weight = layers_11_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_227_cast_fp16)[name = tensor("pretrained_out_139_cast_fp16")]; + tensor var_2798 = const()[name = tensor("op_2798"), val = tensor([1, 1])]; + tensor var_2800 = const()[name = tensor("op_2800"), 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_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(134821184)))]; + tensor input_229_cast_fp16 = conv(dilations = var_2800, groups = var_2666, pad = input_229_pad_0, pad_type = input_229_pad_type_0, strides = var_2798, weight = layers_11_self_attn_o_proj_loraA_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor var_2804 = const()[name = tensor("op_2804"), val = tensor([1, 1])]; + tensor var_2806 = const()[name = tensor("op_2806"), 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(134862208)))]; + tensor lora_out_279_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2806, groups = var_2666, pad = lora_out_277_pad_0, pad_type = lora_out_277_pad_type_0, strides = var_2804, weight = lora_out_279_weight_0_to_fp16, x = input_229_cast_fp16)[name = tensor("lora_out_279_cast_fp16")]; + tensor obj_47_cast_fp16 = add(x = pretrained_out_139_cast_fp16, y = lora_out_279_cast_fp16)[name = tensor("obj_47_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_47_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor var_2815 = const()[name = tensor("op_2815"), val = tensor([1])]; + tensor channels_mean_47_cast_fp16 = reduce_mean(axes = var_2815, keep_dims = var_2667, 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_2819 = const()[name = tensor("op_2819"), val = tensor([1])]; + tensor var_2820_cast_fp16 = reduce_mean(axes = var_2819, keep_dims = var_2667, x = zero_mean_sq_47_cast_fp16)[name = tensor("op_2820_cast_fp16")]; + tensor var_2821_to_fp16 = const()[name = tensor("op_2821_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2822_cast_fp16 = add(x = var_2820_cast_fp16, y = var_2821_to_fp16)[name = tensor("op_2822_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_2822_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(134903232)))]; + 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(134905856)))]; + 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_2836 = const()[name = tensor("op_2836"), val = tensor([1, 1])]; + tensor var_2838 = const()[name = tensor("op_2838"), 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_11_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134908480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138185344))), 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(138185472)))]; + tensor pretrained_out_141_cast_fp16 = conv(bias = layers_11_fc1_pretrained_bias_to_fp16, dilations = var_2838, groups = var_2666, pad = pretrained_out_141_pad_0, pad_type = pretrained_out_141_pad_type_0, strides = var_2836, weight = layers_11_fc1_pretrained_weight_to_fp16_palettized, x = input_231_cast_fp16)[name = tensor("pretrained_out_141_cast_fp16")]; + tensor var_2842 = const()[name = tensor("op_2842"), val = tensor([1, 1])]; + tensor var_2844 = const()[name = tensor("op_2844"), 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_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(138195776)))]; + tensor input_233_cast_fp16 = conv(dilations = var_2844, groups = var_2666, pad = input_233_pad_0, pad_type = input_233_pad_type_0, strides = var_2842, weight = layers_11_fc1_loraA_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor var_2848 = const()[name = tensor("op_2848"), val = tensor([1, 1])]; + tensor var_2850 = const()[name = tensor("op_2850"), 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(138236800)))]; + tensor lora_out_283_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_2850, groups = var_2666, pad = lora_out_281_pad_0, pad_type = lora_out_281_pad_type_0, strides = var_2848, weight = lora_out_283_weight_0_to_fp16, x = input_233_cast_fp16)[name = tensor("lora_out_283_cast_fp16")]; + tensor input_235_cast_fp16 = add(x = pretrained_out_141_cast_fp16, y = lora_out_283_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_2862 = const()[name = tensor("op_2862"), val = tensor([1, 1])]; + tensor var_2864 = const()[name = tensor("op_2864"), 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_11_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138400704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141677568))), 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(141677696)))]; + tensor pretrained_out_143_cast_fp16 = conv(bias = layers_11_fc2_pretrained_bias_to_fp16, dilations = var_2864, groups = var_2666, pad = pretrained_out_143_pad_0, pad_type = pretrained_out_143_pad_type_0, strides = var_2862, weight = layers_11_fc2_pretrained_weight_to_fp16_palettized, x = input_237_cast_fp16)[name = tensor("pretrained_out_143_cast_fp16")]; + tensor var_2868 = const()[name = tensor("op_2868"), val = tensor([1, 1])]; + tensor var_2870 = const()[name = tensor("op_2870"), 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_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(141680320)))]; + tensor input_239_cast_fp16 = conv(dilations = var_2870, groups = var_2666, pad = input_239_pad_0, pad_type = input_239_pad_type_0, strides = var_2868, weight = layers_11_fc2_loraA_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor var_2874 = const()[name = tensor("op_2874"), val = tensor([1, 1])]; + tensor var_2876 = const()[name = tensor("op_2876"), 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(141844224)))]; + tensor lora_out_287_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2876, groups = var_2666, pad = lora_out_285_pad_0, pad_type = lora_out_285_pad_type_0, strides = var_2874, weight = lora_out_287_weight_0_to_fp16, x = input_239_cast_fp16)[name = tensor("lora_out_287_cast_fp16")]; + tensor hidden_states_27_cast_fp16 = add(x = pretrained_out_143_cast_fp16, y = lora_out_287_cast_fp16)[name = tensor("hidden_states_27_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_2890 = const()[name = tensor("op_2890"), val = tensor(3)]; + tensor var_2892 = const()[name = tensor("op_2892"), val = tensor(1)]; + tensor var_2893 = const()[name = tensor("op_2893"), val = tensor(true)]; + tensor var_2903 = const()[name = tensor("op_2903"), val = tensor([1])]; + tensor channels_mean_49_cast_fp16 = reduce_mean(axes = var_2903, keep_dims = var_2893, 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_2907 = const()[name = tensor("op_2907"), val = tensor([1])]; + tensor var_2908_cast_fp16 = reduce_mean(axes = var_2907, keep_dims = var_2893, x = zero_mean_sq_49_cast_fp16)[name = tensor("op_2908_cast_fp16")]; + tensor var_2909_to_fp16 = const()[name = tensor("op_2909_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_2910_cast_fp16 = add(x = var_2908_cast_fp16, y = var_2909_to_fp16)[name = tensor("op_2910_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_2910_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_49_gamma_0_to_fp16 = const()[name = tensor("obj_49_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141885248)))]; + tensor obj_49_beta_0_to_fp16 = const()[name = tensor("obj_49_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141887872)))]; + tensor obj_49_epsilon_0_to_fp16 = const()[name = tensor("obj_49_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_49_cast_fp16 = batch_norm(beta = obj_49_beta_0_to_fp16, epsilon = obj_49_epsilon_0_to_fp16, gamma = obj_49_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_49_cast_fp16")]; + tensor var_2928 = const()[name = tensor("op_2928"), val = tensor([1, 1])]; + tensor var_2930 = const()[name = tensor("op_2930"), 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_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(141890496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142709760))), 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(142709888)))]; + tensor pretrained_out_145_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_2930, groups = var_2892, pad = pretrained_out_145_pad_0, pad_type = pretrained_out_145_pad_type_0, strides = var_2928, weight = layers_12_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = tensor("pretrained_out_145_cast_fp16")]; + tensor var_2934 = const()[name = tensor("op_2934"), val = tensor([1, 1])]; + tensor var_2936 = const()[name = tensor("op_2936"), 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_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(142712512)))]; + tensor input_241_cast_fp16 = conv(dilations = var_2936, groups = var_2892, pad = input_241_pad_0, pad_type = input_241_pad_type_0, strides = var_2934, weight = layers_12_self_attn_q_proj_loraA_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor var_2940 = const()[name = tensor("op_2940"), val = tensor([1, 1])]; + tensor var_2942 = const()[name = tensor("op_2942"), 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(142753536)))]; + tensor lora_out_291_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2942, groups = var_2892, pad = lora_out_289_pad_0, pad_type = lora_out_289_pad_type_0, strides = var_2940, weight = lora_out_291_weight_0_to_fp16, x = input_241_cast_fp16)[name = tensor("lora_out_291_cast_fp16")]; + tensor query_25_cast_fp16 = add(x = pretrained_out_145_cast_fp16, y = lora_out_291_cast_fp16)[name = tensor("query_25_cast_fp16")]; + tensor var_2952 = const()[name = tensor("op_2952"), val = tensor([1, 1])]; + tensor var_2954 = const()[name = tensor("op_2954"), 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_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(142794560))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143613824))), name = tensor("layers_12_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_147_cast_fp16 = conv(dilations = var_2954, groups = var_2892, pad = pretrained_out_147_pad_0, pad_type = pretrained_out_147_pad_type_0, strides = var_2952, weight = layers_12_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = tensor("pretrained_out_147_cast_fp16")]; + tensor var_2958 = const()[name = tensor("op_2958"), val = tensor([1, 1])]; + tensor var_2960 = const()[name = tensor("op_2960"), 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_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(143613952)))]; + tensor input_243_cast_fp16 = conv(dilations = var_2960, groups = var_2892, pad = input_243_pad_0, pad_type = input_243_pad_type_0, strides = var_2958, weight = layers_12_self_attn_k_proj_loraA_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor var_2964 = const()[name = tensor("op_2964"), val = tensor([1, 1])]; + tensor var_2966 = const()[name = tensor("op_2966"), 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(143654976)))]; + tensor lora_out_295_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2966, groups = var_2892, pad = lora_out_293_pad_0, pad_type = lora_out_293_pad_type_0, strides = var_2964, weight = lora_out_295_weight_0_to_fp16, x = input_243_cast_fp16)[name = tensor("lora_out_295_cast_fp16")]; + tensor key_25_cast_fp16 = add(x = pretrained_out_147_cast_fp16, y = lora_out_295_cast_fp16)[name = tensor("key_25_cast_fp16")]; + tensor var_2977 = const()[name = tensor("op_2977"), val = tensor([1, 1])]; + tensor var_2979 = const()[name = tensor("op_2979"), 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_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(143696000))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144515264))), 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(144515392)))]; + tensor pretrained_out_149_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_2979, groups = var_2892, pad = pretrained_out_149_pad_0, pad_type = pretrained_out_149_pad_type_0, strides = var_2977, weight = layers_12_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_49_cast_fp16)[name = tensor("pretrained_out_149_cast_fp16")]; + tensor var_2983 = const()[name = tensor("op_2983"), val = tensor([1, 1])]; + tensor var_2985 = const()[name = tensor("op_2985"), 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_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(144518016)))]; + tensor input_245_cast_fp16 = conv(dilations = var_2985, groups = var_2892, pad = input_245_pad_0, pad_type = input_245_pad_type_0, strides = var_2983, weight = layers_12_self_attn_v_proj_loraA_weight_to_fp16, x = obj_49_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor var_2989 = const()[name = tensor("op_2989"), val = tensor([1, 1])]; + tensor var_2991 = const()[name = tensor("op_2991"), 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(144559040)))]; + tensor lora_out_299_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_2991, groups = var_2892, pad = lora_out_297_pad_0, pad_type = lora_out_297_pad_type_0, strides = var_2989, weight = lora_out_299_weight_0_to_fp16, x = input_245_cast_fp16)[name = tensor("lora_out_299_cast_fp16")]; + tensor value_25_cast_fp16 = add(x = pretrained_out_149_cast_fp16, y = lora_out_299_cast_fp16)[name = tensor("value_25_cast_fp16")]; + tensor var_2998 = const()[name = tensor("op_2998"), val = tensor([1, 20, 64, -1])]; + tensor var_2999_cast_fp16 = reshape(shape = var_2998, x = query_25_cast_fp16)[name = tensor("op_2999_cast_fp16")]; + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3001_cast_fp16 = mul(x = var_2999_cast_fp16, y = var_3000_to_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor var_3002 = const()[name = tensor("op_3002"), val = tensor([1, 20, 64, -1])]; + tensor var_3003_cast_fp16 = reshape(shape = var_3002, x = key_25_cast_fp16)[name = tensor("op_3003_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_3001_cast_fp16, y = var_3003_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor var_3006_cast_fp16 = softmax(axis = var_2890, x = mh_w_25_cast_fp16)[name = tensor("op_3006_cast_fp16")]; + tensor var_3007 = const()[name = tensor("op_3007"), val = tensor([1, 20, 64, -1])]; + tensor var_3008_cast_fp16 = reshape(shape = var_3007, x = value_25_cast_fp16)[name = tensor("op_3008_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_3008_cast_fp16, y = var_3006_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3011 = const()[name = tensor("op_3011"), val = tensor([1, 1280, 1, -1])]; + tensor input_247_cast_fp16 = reshape(shape = var_3011, x = attn_25_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor var_3018 = const()[name = tensor("op_3018"), val = tensor([1, 1])]; + tensor var_3020 = const()[name = tensor("op_3020"), 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_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(144600064))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145419328))), 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(145419456)))]; + tensor pretrained_out_151_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3020, groups = var_2892, pad = pretrained_out_151_pad_0, pad_type = pretrained_out_151_pad_type_0, strides = var_3018, weight = layers_12_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = tensor("pretrained_out_151_cast_fp16")]; + tensor var_3024 = const()[name = tensor("op_3024"), val = tensor([1, 1])]; + tensor var_3026 = const()[name = tensor("op_3026"), 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_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(145422080)))]; + tensor input_249_cast_fp16 = conv(dilations = var_3026, groups = var_2892, pad = input_249_pad_0, pad_type = input_249_pad_type_0, strides = var_3024, weight = layers_12_self_attn_o_proj_loraA_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor var_3030 = const()[name = tensor("op_3030"), val = tensor([1, 1])]; + tensor var_3032 = const()[name = tensor("op_3032"), 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(145463104)))]; + tensor lora_out_303_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3032, groups = var_2892, pad = lora_out_301_pad_0, pad_type = lora_out_301_pad_type_0, strides = var_3030, weight = lora_out_303_weight_0_to_fp16, x = input_249_cast_fp16)[name = tensor("lora_out_303_cast_fp16")]; + tensor obj_51_cast_fp16 = add(x = pretrained_out_151_cast_fp16, y = lora_out_303_cast_fp16)[name = tensor("obj_51_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_51_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor var_3041 = const()[name = tensor("op_3041"), val = tensor([1])]; + tensor channels_mean_51_cast_fp16 = reduce_mean(axes = var_3041, keep_dims = var_2893, 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_3045 = const()[name = tensor("op_3045"), val = tensor([1])]; + tensor var_3046_cast_fp16 = reduce_mean(axes = var_3045, keep_dims = var_2893, x = zero_mean_sq_51_cast_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_to_fp16 = const()[name = tensor("op_3047_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3048_cast_fp16 = add(x = var_3046_cast_fp16, y = var_3047_to_fp16)[name = tensor("op_3048_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_3048_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 input_251_gamma_0_to_fp16 = const()[name = tensor("input_251_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145504128)))]; + tensor input_251_beta_0_to_fp16 = const()[name = tensor("input_251_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145506752)))]; + tensor input_251_epsilon_0_to_fp16 = const()[name = tensor("input_251_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = batch_norm(beta = input_251_beta_0_to_fp16, epsilon = input_251_epsilon_0_to_fp16, gamma = input_251_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("input_251_cast_fp16")]; + tensor var_3062 = const()[name = tensor("op_3062"), val = tensor([1, 1])]; + tensor var_3064 = const()[name = tensor("op_3064"), 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_12_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145509376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148786240))), 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(148786368)))]; + tensor pretrained_out_153_cast_fp16 = conv(bias = layers_12_fc1_pretrained_bias_to_fp16, dilations = var_3064, groups = var_2892, pad = pretrained_out_153_pad_0, pad_type = pretrained_out_153_pad_type_0, strides = var_3062, weight = layers_12_fc1_pretrained_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = tensor("pretrained_out_153_cast_fp16")]; + tensor var_3068 = const()[name = tensor("op_3068"), val = tensor([1, 1])]; + tensor var_3070 = const()[name = tensor("op_3070"), 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_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(148796672)))]; + tensor input_253_cast_fp16 = conv(dilations = var_3070, groups = var_2892, pad = input_253_pad_0, pad_type = input_253_pad_type_0, strides = var_3068, weight = layers_12_fc1_loraA_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor var_3074 = const()[name = tensor("op_3074"), val = tensor([1, 1])]; + tensor var_3076 = const()[name = tensor("op_3076"), 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(148837696)))]; + tensor lora_out_307_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_3076, groups = var_2892, pad = lora_out_305_pad_0, pad_type = lora_out_305_pad_type_0, strides = var_3074, weight = lora_out_307_weight_0_to_fp16, x = input_253_cast_fp16)[name = tensor("lora_out_307_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = pretrained_out_153_cast_fp16, y = lora_out_307_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor input_257_mode_0 = const()[name = tensor("input_257_mode_0"), val = tensor("EXACT")]; + tensor input_257_cast_fp16 = gelu(mode = input_257_mode_0, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor var_3088 = const()[name = tensor("op_3088"), val = tensor([1, 1])]; + tensor var_3090 = const()[name = tensor("op_3090"), 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_12_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149001600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152278464))), 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(152278592)))]; + tensor pretrained_out_155_cast_fp16 = conv(bias = layers_12_fc2_pretrained_bias_to_fp16, dilations = var_3090, groups = var_2892, pad = pretrained_out_155_pad_0, pad_type = pretrained_out_155_pad_type_0, strides = var_3088, weight = layers_12_fc2_pretrained_weight_to_fp16_palettized, x = input_257_cast_fp16)[name = tensor("pretrained_out_155_cast_fp16")]; + tensor var_3094 = const()[name = tensor("op_3094"), val = tensor([1, 1])]; + tensor var_3096 = const()[name = tensor("op_3096"), 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_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(152281216)))]; + tensor input_259_cast_fp16 = conv(dilations = var_3096, groups = var_2892, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = var_3094, weight = layers_12_fc2_loraA_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor var_3100 = const()[name = tensor("op_3100"), val = tensor([1, 1])]; + tensor var_3102 = const()[name = tensor("op_3102"), 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(152445120)))]; + tensor lora_out_311_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3102, groups = var_2892, pad = lora_out_309_pad_0, pad_type = lora_out_309_pad_type_0, strides = var_3100, weight = lora_out_311_weight_0_to_fp16, x = input_259_cast_fp16)[name = tensor("lora_out_311_cast_fp16")]; + tensor hidden_states_29_cast_fp16 = add(x = pretrained_out_155_cast_fp16, y = lora_out_311_cast_fp16)[name = tensor("hidden_states_29_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor var_3116 = const()[name = tensor("op_3116"), val = tensor(3)]; + tensor var_3118 = const()[name = tensor("op_3118"), val = tensor(1)]; + tensor var_3119 = const()[name = tensor("op_3119"), val = tensor(true)]; + tensor var_3129 = const()[name = tensor("op_3129"), val = tensor([1])]; + tensor channels_mean_53_cast_fp16 = reduce_mean(axes = var_3129, keep_dims = var_3119, 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_3133 = const()[name = tensor("op_3133"), val = tensor([1])]; + tensor var_3134_cast_fp16 = reduce_mean(axes = var_3133, keep_dims = var_3119, x = zero_mean_sq_53_cast_fp16)[name = tensor("op_3134_cast_fp16")]; + tensor var_3135_to_fp16 = const()[name = tensor("op_3135_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3136_cast_fp16 = add(x = var_3134_cast_fp16, y = var_3135_to_fp16)[name = tensor("op_3136_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_3136_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 obj_53_gamma_0_to_fp16 = const()[name = tensor("obj_53_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152486144)))]; + tensor obj_53_beta_0_to_fp16 = const()[name = tensor("obj_53_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152488768)))]; + tensor obj_53_epsilon_0_to_fp16 = const()[name = tensor("obj_53_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_53_cast_fp16 = batch_norm(beta = obj_53_beta_0_to_fp16, epsilon = obj_53_epsilon_0_to_fp16, gamma = obj_53_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("obj_53_cast_fp16")]; + tensor var_3154 = const()[name = tensor("op_3154"), val = tensor([1, 1])]; + tensor var_3156 = const()[name = tensor("op_3156"), 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_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(152491392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153310656))), 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(153310784)))]; + tensor pretrained_out_157_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3156, groups = var_3118, pad = pretrained_out_157_pad_0, pad_type = pretrained_out_157_pad_type_0, strides = var_3154, weight = layers_13_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_53_cast_fp16)[name = tensor("pretrained_out_157_cast_fp16")]; + tensor var_3160 = const()[name = tensor("op_3160"), val = tensor([1, 1])]; + tensor var_3162 = const()[name = tensor("op_3162"), val = tensor([1, 1])]; + tensor input_261_pad_type_0 = const()[name = tensor("input_261_pad_type_0"), val = tensor("custom")]; + tensor input_261_pad_0 = const()[name = tensor("input_261_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(153313408)))]; + tensor input_261_cast_fp16 = conv(dilations = var_3162, groups = var_3118, pad = input_261_pad_0, pad_type = input_261_pad_type_0, strides = var_3160, weight = layers_13_self_attn_q_proj_loraA_weight_to_fp16, x = obj_53_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor var_3166 = const()[name = tensor("op_3166"), val = tensor([1, 1])]; + tensor var_3168 = const()[name = tensor("op_3168"), 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(153354432)))]; + tensor lora_out_315_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3168, groups = var_3118, pad = lora_out_313_pad_0, pad_type = lora_out_313_pad_type_0, strides = var_3166, weight = lora_out_315_weight_0_to_fp16, x = input_261_cast_fp16)[name = tensor("lora_out_315_cast_fp16")]; + tensor query_27_cast_fp16 = add(x = pretrained_out_157_cast_fp16, y = lora_out_315_cast_fp16)[name = tensor("query_27_cast_fp16")]; + tensor var_3178 = const()[name = tensor("op_3178"), val = tensor([1, 1])]; + tensor var_3180 = const()[name = tensor("op_3180"), 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_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(153395456))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154214720))), name = tensor("layers_13_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_159_cast_fp16 = conv(dilations = var_3180, groups = var_3118, pad = pretrained_out_159_pad_0, pad_type = pretrained_out_159_pad_type_0, strides = var_3178, weight = layers_13_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_53_cast_fp16)[name = tensor("pretrained_out_159_cast_fp16")]; + tensor var_3184 = const()[name = tensor("op_3184"), val = tensor([1, 1])]; + tensor var_3186 = const()[name = tensor("op_3186"), 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_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(154214848)))]; + tensor input_263_cast_fp16 = conv(dilations = var_3186, groups = var_3118, pad = input_263_pad_0, pad_type = input_263_pad_type_0, strides = var_3184, weight = layers_13_self_attn_k_proj_loraA_weight_to_fp16, x = obj_53_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor var_3190 = const()[name = tensor("op_3190"), val = tensor([1, 1])]; + tensor var_3192 = const()[name = tensor("op_3192"), 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(154255872)))]; + tensor lora_out_319_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3192, groups = var_3118, pad = lora_out_317_pad_0, pad_type = lora_out_317_pad_type_0, strides = var_3190, weight = lora_out_319_weight_0_to_fp16, x = input_263_cast_fp16)[name = tensor("lora_out_319_cast_fp16")]; + tensor key_27_cast_fp16 = add(x = pretrained_out_159_cast_fp16, y = lora_out_319_cast_fp16)[name = tensor("key_27_cast_fp16")]; + tensor var_3203 = const()[name = tensor("op_3203"), val = tensor([1, 1])]; + tensor var_3205 = const()[name = tensor("op_3205"), 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_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(154296896))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155116160))), 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(155116288)))]; + tensor pretrained_out_161_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3205, groups = var_3118, pad = pretrained_out_161_pad_0, pad_type = pretrained_out_161_pad_type_0, strides = var_3203, weight = layers_13_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_53_cast_fp16)[name = tensor("pretrained_out_161_cast_fp16")]; + tensor var_3209 = const()[name = tensor("op_3209"), val = tensor([1, 1])]; + tensor var_3211 = const()[name = tensor("op_3211"), val = tensor([1, 1])]; + tensor input_265_pad_type_0 = const()[name = tensor("input_265_pad_type_0"), val = tensor("custom")]; + tensor input_265_pad_0 = const()[name = tensor("input_265_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(155118912)))]; + tensor input_265_cast_fp16 = conv(dilations = var_3211, groups = var_3118, pad = input_265_pad_0, pad_type = input_265_pad_type_0, strides = var_3209, weight = layers_13_self_attn_v_proj_loraA_weight_to_fp16, x = obj_53_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor var_3215 = const()[name = tensor("op_3215"), val = tensor([1, 1])]; + tensor var_3217 = const()[name = tensor("op_3217"), 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(155159936)))]; + tensor lora_out_323_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3217, groups = var_3118, pad = lora_out_321_pad_0, pad_type = lora_out_321_pad_type_0, strides = var_3215, weight = lora_out_323_weight_0_to_fp16, x = input_265_cast_fp16)[name = tensor("lora_out_323_cast_fp16")]; + tensor value_27_cast_fp16 = add(x = pretrained_out_161_cast_fp16, y = lora_out_323_cast_fp16)[name = tensor("value_27_cast_fp16")]; + tensor var_3224 = const()[name = tensor("op_3224"), val = tensor([1, 20, 64, -1])]; + tensor var_3225_cast_fp16 = reshape(shape = var_3224, x = query_27_cast_fp16)[name = tensor("op_3225_cast_fp16")]; + tensor var_3226_to_fp16 = const()[name = tensor("op_3226_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3227_cast_fp16 = mul(x = var_3225_cast_fp16, y = var_3226_to_fp16)[name = tensor("op_3227_cast_fp16")]; + tensor var_3228 = const()[name = tensor("op_3228"), val = tensor([1, 20, 64, -1])]; + tensor var_3229_cast_fp16 = reshape(shape = var_3228, x = key_27_cast_fp16)[name = tensor("op_3229_cast_fp16")]; + tensor mh_w_27_transpose_x_0 = const()[name = tensor("mh_w_27_transpose_x_0"), val = tensor(true)]; + tensor mh_w_27_transpose_y_0 = const()[name = tensor("mh_w_27_transpose_y_0"), val = tensor(false)]; + tensor mh_w_27_cast_fp16 = matmul(transpose_x = mh_w_27_transpose_x_0, transpose_y = mh_w_27_transpose_y_0, x = var_3227_cast_fp16, y = var_3229_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_3232_cast_fp16 = softmax(axis = var_3116, x = mh_w_27_cast_fp16)[name = tensor("op_3232_cast_fp16")]; + tensor var_3233 = const()[name = tensor("op_3233"), val = tensor([1, 20, 64, -1])]; + tensor var_3234_cast_fp16 = reshape(shape = var_3233, x = value_27_cast_fp16)[name = tensor("op_3234_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_3234_cast_fp16, y = var_3232_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3237 = const()[name = tensor("op_3237"), val = tensor([1, 1280, 1, -1])]; + tensor input_267_cast_fp16 = reshape(shape = var_3237, x = attn_27_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor var_3244 = const()[name = tensor("op_3244"), val = tensor([1, 1])]; + tensor var_3246 = const()[name = tensor("op_3246"), 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_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(155200960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156020224))), 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(156020352)))]; + tensor pretrained_out_163_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3246, groups = var_3118, pad = pretrained_out_163_pad_0, pad_type = pretrained_out_163_pad_type_0, strides = var_3244, weight = layers_13_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_267_cast_fp16)[name = tensor("pretrained_out_163_cast_fp16")]; + tensor var_3250 = const()[name = tensor("op_3250"), val = tensor([1, 1])]; + tensor var_3252 = const()[name = tensor("op_3252"), 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_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(156022976)))]; + tensor input_269_cast_fp16 = conv(dilations = var_3252, groups = var_3118, pad = input_269_pad_0, pad_type = input_269_pad_type_0, strides = var_3250, weight = layers_13_self_attn_o_proj_loraA_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor var_3256 = const()[name = tensor("op_3256"), val = tensor([1, 1])]; + tensor var_3258 = const()[name = tensor("op_3258"), 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(156064000)))]; + tensor lora_out_327_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3258, groups = var_3118, pad = lora_out_325_pad_0, pad_type = lora_out_325_pad_type_0, strides = var_3256, weight = lora_out_327_weight_0_to_fp16, x = input_269_cast_fp16)[name = tensor("lora_out_327_cast_fp16")]; + tensor obj_55_cast_fp16 = add(x = pretrained_out_163_cast_fp16, y = lora_out_327_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = obj_55_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor var_3267 = const()[name = tensor("op_3267"), val = tensor([1])]; + tensor channels_mean_55_cast_fp16 = reduce_mean(axes = var_3267, keep_dims = var_3119, 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_3271 = const()[name = tensor("op_3271"), val = tensor([1])]; + tensor var_3272_cast_fp16 = reduce_mean(axes = var_3271, keep_dims = var_3119, x = zero_mean_sq_55_cast_fp16)[name = tensor("op_3272_cast_fp16")]; + tensor var_3273_to_fp16 = const()[name = tensor("op_3273_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3274_cast_fp16 = add(x = var_3272_cast_fp16, y = var_3273_to_fp16)[name = tensor("op_3274_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_3274_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 input_271_gamma_0_to_fp16 = const()[name = tensor("input_271_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156105024)))]; + tensor input_271_beta_0_to_fp16 = const()[name = tensor("input_271_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156107648)))]; + tensor input_271_epsilon_0_to_fp16 = const()[name = tensor("input_271_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_271_cast_fp16 = batch_norm(beta = input_271_beta_0_to_fp16, epsilon = input_271_epsilon_0_to_fp16, gamma = input_271_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("input_271_cast_fp16")]; + tensor var_3288 = const()[name = tensor("op_3288"), val = tensor([1, 1])]; + tensor var_3290 = const()[name = tensor("op_3290"), 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_13_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156110272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159387136))), 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(159387264)))]; + tensor pretrained_out_165_cast_fp16 = conv(bias = layers_13_fc1_pretrained_bias_to_fp16, dilations = var_3290, groups = var_3118, pad = pretrained_out_165_pad_0, pad_type = pretrained_out_165_pad_type_0, strides = var_3288, weight = layers_13_fc1_pretrained_weight_to_fp16_palettized, x = input_271_cast_fp16)[name = tensor("pretrained_out_165_cast_fp16")]; + tensor var_3294 = const()[name = tensor("op_3294"), val = tensor([1, 1])]; + tensor var_3296 = const()[name = tensor("op_3296"), 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_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(159397568)))]; + tensor input_273_cast_fp16 = conv(dilations = var_3296, groups = var_3118, pad = input_273_pad_0, pad_type = input_273_pad_type_0, strides = var_3294, weight = layers_13_fc1_loraA_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor var_3300 = const()[name = tensor("op_3300"), val = tensor([1, 1])]; + tensor var_3302 = const()[name = tensor("op_3302"), 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(159438592)))]; + tensor lora_out_331_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_3302, groups = var_3118, pad = lora_out_329_pad_0, pad_type = lora_out_329_pad_type_0, strides = var_3300, weight = lora_out_331_weight_0_to_fp16, x = input_273_cast_fp16)[name = tensor("lora_out_331_cast_fp16")]; + tensor input_275_cast_fp16 = add(x = pretrained_out_165_cast_fp16, y = lora_out_331_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor input_277_mode_0 = const()[name = tensor("input_277_mode_0"), val = tensor("EXACT")]; + tensor input_277_cast_fp16 = gelu(mode = input_277_mode_0, x = input_275_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor var_3314 = const()[name = tensor("op_3314"), val = tensor([1, 1])]; + tensor var_3316 = const()[name = tensor("op_3316"), 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_13_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159602496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162879360))), 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(162879488)))]; + tensor pretrained_out_167_cast_fp16 = conv(bias = layers_13_fc2_pretrained_bias_to_fp16, dilations = var_3316, groups = var_3118, pad = pretrained_out_167_pad_0, pad_type = pretrained_out_167_pad_type_0, strides = var_3314, weight = layers_13_fc2_pretrained_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = tensor("pretrained_out_167_cast_fp16")]; + tensor var_3320 = const()[name = tensor("op_3320"), val = tensor([1, 1])]; + tensor var_3322 = const()[name = tensor("op_3322"), 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_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(162882112)))]; + tensor input_279_cast_fp16 = conv(dilations = var_3322, groups = var_3118, pad = input_279_pad_0, pad_type = input_279_pad_type_0, strides = var_3320, weight = layers_13_fc2_loraA_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor var_3326 = const()[name = tensor("op_3326"), val = tensor([1, 1])]; + tensor var_3328 = const()[name = tensor("op_3328"), 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(163046016)))]; + tensor lora_out_335_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3328, groups = var_3118, pad = lora_out_333_pad_0, pad_type = lora_out_333_pad_type_0, strides = var_3326, weight = lora_out_335_weight_0_to_fp16, x = input_279_cast_fp16)[name = tensor("lora_out_335_cast_fp16")]; + tensor hidden_states_31_cast_fp16 = add(x = pretrained_out_167_cast_fp16, y = lora_out_335_cast_fp16)[name = tensor("hidden_states_31_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor var_3342 = const()[name = tensor("op_3342"), val = tensor(3)]; + tensor var_3344 = const()[name = tensor("op_3344"), val = tensor(1)]; + tensor var_3345 = const()[name = tensor("op_3345"), val = tensor(true)]; + tensor var_3355 = const()[name = tensor("op_3355"), val = tensor([1])]; + tensor channels_mean_57_cast_fp16 = reduce_mean(axes = var_3355, keep_dims = var_3345, 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_3359 = const()[name = tensor("op_3359"), val = tensor([1])]; + tensor var_3360_cast_fp16 = reduce_mean(axes = var_3359, keep_dims = var_3345, x = zero_mean_sq_57_cast_fp16)[name = tensor("op_3360_cast_fp16")]; + tensor var_3361_to_fp16 = const()[name = tensor("op_3361_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3362_cast_fp16 = add(x = var_3360_cast_fp16, y = var_3361_to_fp16)[name = tensor("op_3362_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_3362_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_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(163087040)))]; + 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(163089664)))]; + 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_57_cast_fp16)[name = tensor("obj_57_cast_fp16")]; + tensor var_3380 = const()[name = tensor("op_3380"), val = tensor([1, 1])]; + tensor var_3382 = const()[name = tensor("op_3382"), 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_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(163092288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163911552))), 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(163911680)))]; + tensor pretrained_out_169_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3382, groups = var_3344, pad = pretrained_out_169_pad_0, pad_type = pretrained_out_169_pad_type_0, strides = var_3380, weight = layers_14_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = tensor("pretrained_out_169_cast_fp16")]; + tensor var_3386 = const()[name = tensor("op_3386"), val = tensor([1, 1])]; + tensor var_3388 = const()[name = tensor("op_3388"), 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_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(163914304)))]; + tensor input_281_cast_fp16 = conv(dilations = var_3388, groups = var_3344, pad = input_281_pad_0, pad_type = input_281_pad_type_0, strides = var_3386, weight = layers_14_self_attn_q_proj_loraA_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor var_3392 = const()[name = tensor("op_3392"), val = tensor([1, 1])]; + tensor var_3394 = const()[name = tensor("op_3394"), 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(163955328)))]; + tensor lora_out_339_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3394, groups = var_3344, pad = lora_out_337_pad_0, pad_type = lora_out_337_pad_type_0, strides = var_3392, weight = lora_out_339_weight_0_to_fp16, x = input_281_cast_fp16)[name = tensor("lora_out_339_cast_fp16")]; + tensor query_29_cast_fp16 = add(x = pretrained_out_169_cast_fp16, y = lora_out_339_cast_fp16)[name = tensor("query_29_cast_fp16")]; + tensor var_3404 = const()[name = tensor("op_3404"), val = tensor([1, 1])]; + tensor var_3406 = const()[name = tensor("op_3406"), 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_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(163996352))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164815616))), name = tensor("layers_14_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_171_cast_fp16 = conv(dilations = var_3406, groups = var_3344, pad = pretrained_out_171_pad_0, pad_type = pretrained_out_171_pad_type_0, strides = var_3404, weight = layers_14_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = tensor("pretrained_out_171_cast_fp16")]; + tensor var_3410 = const()[name = tensor("op_3410"), val = tensor([1, 1])]; + tensor var_3412 = const()[name = tensor("op_3412"), 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_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(164815744)))]; + tensor input_283_cast_fp16 = conv(dilations = var_3412, groups = var_3344, pad = input_283_pad_0, pad_type = input_283_pad_type_0, strides = var_3410, weight = layers_14_self_attn_k_proj_loraA_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor var_3416 = const()[name = tensor("op_3416"), val = tensor([1, 1])]; + tensor var_3418 = const()[name = tensor("op_3418"), 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(164856768)))]; + tensor lora_out_343_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3418, groups = var_3344, pad = lora_out_341_pad_0, pad_type = lora_out_341_pad_type_0, strides = var_3416, weight = lora_out_343_weight_0_to_fp16, x = input_283_cast_fp16)[name = tensor("lora_out_343_cast_fp16")]; + tensor key_29_cast_fp16 = add(x = pretrained_out_171_cast_fp16, y = lora_out_343_cast_fp16)[name = tensor("key_29_cast_fp16")]; + tensor var_3429 = const()[name = tensor("op_3429"), val = tensor([1, 1])]; + tensor var_3431 = const()[name = tensor("op_3431"), 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_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(164897792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165717056))), 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(165717184)))]; + tensor pretrained_out_173_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3431, groups = var_3344, pad = pretrained_out_173_pad_0, pad_type = pretrained_out_173_pad_type_0, strides = var_3429, weight = layers_14_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_57_cast_fp16)[name = tensor("pretrained_out_173_cast_fp16")]; + tensor var_3435 = const()[name = tensor("op_3435"), val = tensor([1, 1])]; + tensor var_3437 = const()[name = tensor("op_3437"), 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_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(165719808)))]; + tensor input_285_cast_fp16 = conv(dilations = var_3437, groups = var_3344, pad = input_285_pad_0, pad_type = input_285_pad_type_0, strides = var_3435, weight = layers_14_self_attn_v_proj_loraA_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor var_3441 = const()[name = tensor("op_3441"), val = tensor([1, 1])]; + tensor var_3443 = const()[name = tensor("op_3443"), 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(165760832)))]; + tensor lora_out_347_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3443, groups = var_3344, pad = lora_out_345_pad_0, pad_type = lora_out_345_pad_type_0, strides = var_3441, weight = lora_out_347_weight_0_to_fp16, x = input_285_cast_fp16)[name = tensor("lora_out_347_cast_fp16")]; + tensor value_29_cast_fp16 = add(x = pretrained_out_173_cast_fp16, y = lora_out_347_cast_fp16)[name = tensor("value_29_cast_fp16")]; + tensor var_3450 = const()[name = tensor("op_3450"), val = tensor([1, 20, 64, -1])]; + tensor var_3451_cast_fp16 = reshape(shape = var_3450, x = query_29_cast_fp16)[name = tensor("op_3451_cast_fp16")]; + tensor var_3452_to_fp16 = const()[name = tensor("op_3452_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3453_cast_fp16 = mul(x = var_3451_cast_fp16, y = var_3452_to_fp16)[name = tensor("op_3453_cast_fp16")]; + tensor var_3454 = const()[name = tensor("op_3454"), val = tensor([1, 20, 64, -1])]; + tensor var_3455_cast_fp16 = reshape(shape = var_3454, x = key_29_cast_fp16)[name = tensor("op_3455_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_3453_cast_fp16, y = var_3455_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor var_3458_cast_fp16 = softmax(axis = var_3342, x = mh_w_29_cast_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 = value_29_cast_fp16)[name = tensor("op_3460_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_3460_cast_fp16, y = var_3458_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_3463 = const()[name = tensor("op_3463"), val = tensor([1, 1280, 1, -1])]; + tensor input_287_cast_fp16 = reshape(shape = var_3463, x = attn_29_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor var_3470 = const()[name = tensor("op_3470"), val = tensor([1, 1])]; + tensor var_3472 = const()[name = tensor("op_3472"), 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_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(165801856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166621120))), 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(166621248)))]; + tensor pretrained_out_175_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3472, groups = var_3344, pad = pretrained_out_175_pad_0, pad_type = pretrained_out_175_pad_type_0, strides = var_3470, weight = layers_14_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = tensor("pretrained_out_175_cast_fp16")]; + tensor var_3476 = const()[name = tensor("op_3476"), val = tensor([1, 1])]; + tensor var_3478 = const()[name = tensor("op_3478"), 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_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(166623872)))]; + tensor input_289_cast_fp16 = conv(dilations = var_3478, groups = var_3344, pad = input_289_pad_0, pad_type = input_289_pad_type_0, strides = var_3476, weight = layers_14_self_attn_o_proj_loraA_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor var_3482 = const()[name = tensor("op_3482"), val = tensor([1, 1])]; + tensor var_3484 = const()[name = tensor("op_3484"), 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(166664896)))]; + tensor lora_out_351_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3484, groups = var_3344, pad = lora_out_349_pad_0, pad_type = lora_out_349_pad_type_0, strides = var_3482, weight = lora_out_351_weight_0_to_fp16, x = input_289_cast_fp16)[name = tensor("lora_out_351_cast_fp16")]; + tensor obj_59_cast_fp16 = add(x = pretrained_out_175_cast_fp16, y = lora_out_351_cast_fp16)[name = tensor("obj_59_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_59_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor var_3493 = const()[name = tensor("op_3493"), val = tensor([1])]; + tensor channels_mean_59_cast_fp16 = reduce_mean(axes = var_3493, keep_dims = var_3345, 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_3497 = const()[name = tensor("op_3497"), val = tensor([1])]; + tensor var_3498_cast_fp16 = reduce_mean(axes = var_3497, keep_dims = var_3345, x = zero_mean_sq_59_cast_fp16)[name = tensor("op_3498_cast_fp16")]; + tensor var_3499_to_fp16 = const()[name = tensor("op_3499_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3500_cast_fp16 = add(x = var_3498_cast_fp16, y = var_3499_to_fp16)[name = tensor("op_3500_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_3500_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(166705920)))]; + 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(166708544)))]; + 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_3514 = const()[name = tensor("op_3514"), val = tensor([1, 1])]; + tensor var_3516 = const()[name = tensor("op_3516"), 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_14_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166711168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169988032))), 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(169988160)))]; + tensor pretrained_out_177_cast_fp16 = conv(bias = layers_14_fc1_pretrained_bias_to_fp16, dilations = var_3516, groups = var_3344, pad = pretrained_out_177_pad_0, pad_type = pretrained_out_177_pad_type_0, strides = var_3514, weight = layers_14_fc1_pretrained_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = tensor("pretrained_out_177_cast_fp16")]; + tensor var_3520 = const()[name = tensor("op_3520"), val = tensor([1, 1])]; + tensor var_3522 = const()[name = tensor("op_3522"), 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_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(169998464)))]; + tensor input_293_cast_fp16 = conv(dilations = var_3522, groups = var_3344, pad = input_293_pad_0, pad_type = input_293_pad_type_0, strides = var_3520, weight = layers_14_fc1_loraA_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor var_3526 = const()[name = tensor("op_3526"), val = tensor([1, 1])]; + tensor var_3528 = const()[name = tensor("op_3528"), 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(170039488)))]; + tensor lora_out_355_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_3528, groups = var_3344, pad = lora_out_353_pad_0, pad_type = lora_out_353_pad_type_0, strides = var_3526, weight = lora_out_355_weight_0_to_fp16, x = input_293_cast_fp16)[name = tensor("lora_out_355_cast_fp16")]; + tensor input_295_cast_fp16 = add(x = pretrained_out_177_cast_fp16, y = lora_out_355_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_3540 = const()[name = tensor("op_3540"), val = tensor([1, 1])]; + tensor var_3542 = const()[name = tensor("op_3542"), 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_14_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170203392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173480256))), 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(173480384)))]; + tensor pretrained_out_179_cast_fp16 = conv(bias = layers_14_fc2_pretrained_bias_to_fp16, dilations = var_3542, groups = var_3344, pad = pretrained_out_179_pad_0, pad_type = pretrained_out_179_pad_type_0, strides = var_3540, weight = layers_14_fc2_pretrained_weight_to_fp16_palettized, x = input_297_cast_fp16)[name = tensor("pretrained_out_179_cast_fp16")]; + tensor var_3546 = const()[name = tensor("op_3546"), val = tensor([1, 1])]; + tensor var_3548 = const()[name = tensor("op_3548"), 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_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(173483008)))]; + tensor input_299_cast_fp16 = conv(dilations = var_3548, groups = var_3344, pad = input_299_pad_0, pad_type = input_299_pad_type_0, strides = var_3546, weight = layers_14_fc2_loraA_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("input_299_cast_fp16")]; + tensor var_3552 = const()[name = tensor("op_3552"), val = tensor([1, 1])]; + tensor var_3554 = const()[name = tensor("op_3554"), 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(173646912)))]; + tensor lora_out_359_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3554, groups = var_3344, pad = lora_out_357_pad_0, pad_type = lora_out_357_pad_type_0, strides = var_3552, weight = lora_out_359_weight_0_to_fp16, x = input_299_cast_fp16)[name = tensor("lora_out_359_cast_fp16")]; + tensor hidden_states_33_cast_fp16 = add(x = pretrained_out_179_cast_fp16, y = lora_out_359_cast_fp16)[name = tensor("hidden_states_33_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_3568 = const()[name = tensor("op_3568"), val = tensor(3)]; + tensor var_3570 = const()[name = tensor("op_3570"), val = tensor(1)]; + tensor var_3571 = const()[name = tensor("op_3571"), val = tensor(true)]; + tensor var_3581 = const()[name = tensor("op_3581"), val = tensor([1])]; + tensor channels_mean_61_cast_fp16 = reduce_mean(axes = var_3581, keep_dims = var_3571, 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_3585 = const()[name = tensor("op_3585"), val = tensor([1])]; + tensor var_3586_cast_fp16 = reduce_mean(axes = var_3585, keep_dims = var_3571, x = zero_mean_sq_61_cast_fp16)[name = tensor("op_3586_cast_fp16")]; + tensor var_3587_to_fp16 = const()[name = tensor("op_3587_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3588_cast_fp16 = add(x = var_3586_cast_fp16, y = var_3587_to_fp16)[name = tensor("op_3588_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_3588_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_61_gamma_0_to_fp16 = const()[name = tensor("obj_61_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173687936)))]; + tensor obj_61_beta_0_to_fp16 = const()[name = tensor("obj_61_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173690560)))]; + tensor obj_61_epsilon_0_to_fp16 = const()[name = tensor("obj_61_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_61_cast_fp16 = batch_norm(beta = obj_61_beta_0_to_fp16, epsilon = obj_61_epsilon_0_to_fp16, gamma = obj_61_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_61_cast_fp16")]; + tensor var_3606 = const()[name = tensor("op_3606"), val = tensor([1, 1])]; + tensor var_3608 = const()[name = tensor("op_3608"), 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_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(173693184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174512448))), 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(174512576)))]; + tensor pretrained_out_181_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3608, groups = var_3570, pad = pretrained_out_181_pad_0, pad_type = pretrained_out_181_pad_type_0, strides = var_3606, weight = layers_15_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = tensor("pretrained_out_181_cast_fp16")]; + tensor var_3612 = const()[name = tensor("op_3612"), val = tensor([1, 1])]; + tensor var_3614 = const()[name = tensor("op_3614"), 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_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(174515200)))]; + tensor input_301_cast_fp16 = conv(dilations = var_3614, groups = var_3570, pad = input_301_pad_0, pad_type = input_301_pad_type_0, strides = var_3612, weight = layers_15_self_attn_q_proj_loraA_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor var_3618 = const()[name = tensor("op_3618"), val = tensor([1, 1])]; + tensor var_3620 = const()[name = tensor("op_3620"), 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(174556224)))]; + tensor lora_out_363_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3620, groups = var_3570, pad = lora_out_361_pad_0, pad_type = lora_out_361_pad_type_0, strides = var_3618, weight = lora_out_363_weight_0_to_fp16, x = input_301_cast_fp16)[name = tensor("lora_out_363_cast_fp16")]; + tensor query_31_cast_fp16 = add(x = pretrained_out_181_cast_fp16, y = lora_out_363_cast_fp16)[name = tensor("query_31_cast_fp16")]; + tensor var_3630 = const()[name = tensor("op_3630"), val = tensor([1, 1])]; + tensor var_3632 = const()[name = tensor("op_3632"), 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_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(174597248))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175416512))), name = tensor("layers_15_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_183_cast_fp16 = conv(dilations = var_3632, groups = var_3570, pad = pretrained_out_183_pad_0, pad_type = pretrained_out_183_pad_type_0, strides = var_3630, weight = layers_15_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = tensor("pretrained_out_183_cast_fp16")]; + tensor var_3636 = const()[name = tensor("op_3636"), val = tensor([1, 1])]; + tensor var_3638 = const()[name = tensor("op_3638"), 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_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(175416640)))]; + tensor input_303_cast_fp16 = conv(dilations = var_3638, groups = var_3570, pad = input_303_pad_0, pad_type = input_303_pad_type_0, strides = var_3636, weight = layers_15_self_attn_k_proj_loraA_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor("input_303_cast_fp16")]; + tensor var_3642 = const()[name = tensor("op_3642"), val = tensor([1, 1])]; + tensor var_3644 = const()[name = tensor("op_3644"), 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(175457664)))]; + tensor lora_out_367_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3644, groups = var_3570, pad = lora_out_365_pad_0, pad_type = lora_out_365_pad_type_0, strides = var_3642, weight = lora_out_367_weight_0_to_fp16, x = input_303_cast_fp16)[name = tensor("lora_out_367_cast_fp16")]; + tensor key_31_cast_fp16 = add(x = pretrained_out_183_cast_fp16, y = lora_out_367_cast_fp16)[name = tensor("key_31_cast_fp16")]; + tensor var_3655 = const()[name = tensor("op_3655"), val = tensor([1, 1])]; + tensor var_3657 = const()[name = tensor("op_3657"), 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_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(175498688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176317952))), 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(176318080)))]; + tensor pretrained_out_185_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3657, groups = var_3570, pad = pretrained_out_185_pad_0, pad_type = pretrained_out_185_pad_type_0, strides = var_3655, weight = layers_15_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_61_cast_fp16)[name = tensor("pretrained_out_185_cast_fp16")]; + tensor var_3661 = const()[name = tensor("op_3661"), val = tensor([1, 1])]; + tensor var_3663 = const()[name = tensor("op_3663"), 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_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(176320704)))]; + tensor input_305_cast_fp16 = conv(dilations = var_3663, groups = var_3570, pad = input_305_pad_0, pad_type = input_305_pad_type_0, strides = var_3661, weight = layers_15_self_attn_v_proj_loraA_weight_to_fp16, x = obj_61_cast_fp16)[name = tensor("input_305_cast_fp16")]; + tensor var_3667 = const()[name = tensor("op_3667"), val = tensor([1, 1])]; + tensor var_3669 = const()[name = tensor("op_3669"), 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(176361728)))]; + tensor lora_out_371_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3669, groups = var_3570, pad = lora_out_369_pad_0, pad_type = lora_out_369_pad_type_0, strides = var_3667, weight = lora_out_371_weight_0_to_fp16, x = input_305_cast_fp16)[name = tensor("lora_out_371_cast_fp16")]; + tensor value_31_cast_fp16 = add(x = pretrained_out_185_cast_fp16, y = lora_out_371_cast_fp16)[name = tensor("value_31_cast_fp16")]; + tensor var_3676 = const()[name = tensor("op_3676"), val = tensor([1, 20, 64, -1])]; + tensor var_3677_cast_fp16 = reshape(shape = var_3676, x = query_31_cast_fp16)[name = tensor("op_3677_cast_fp16")]; + tensor var_3678_to_fp16 = const()[name = tensor("op_3678_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3679_cast_fp16 = mul(x = var_3677_cast_fp16, y = var_3678_to_fp16)[name = tensor("op_3679_cast_fp16")]; + tensor var_3680 = const()[name = tensor("op_3680"), val = tensor([1, 20, 64, -1])]; + tensor var_3681_cast_fp16 = reshape(shape = var_3680, x = key_31_cast_fp16)[name = tensor("op_3681_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_3679_cast_fp16, y = var_3681_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor var_3684_cast_fp16 = softmax(axis = var_3568, x = mh_w_31_cast_fp16)[name = tensor("op_3684_cast_fp16")]; + tensor var_3685 = const()[name = tensor("op_3685"), val = tensor([1, 20, 64, -1])]; + tensor var_3686_cast_fp16 = reshape(shape = var_3685, x = value_31_cast_fp16)[name = tensor("op_3686_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_3686_cast_fp16, y = var_3684_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_3689 = const()[name = tensor("op_3689"), val = tensor([1, 1280, 1, -1])]; + tensor input_307_cast_fp16 = reshape(shape = var_3689, x = attn_31_cast_fp16)[name = tensor("input_307_cast_fp16")]; + tensor var_3696 = const()[name = tensor("op_3696"), val = tensor([1, 1])]; + tensor var_3698 = const()[name = tensor("op_3698"), 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_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(176402752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177222016))), 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(177222144)))]; + tensor pretrained_out_187_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3698, groups = var_3570, pad = pretrained_out_187_pad_0, pad_type = pretrained_out_187_pad_type_0, strides = var_3696, weight = layers_15_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_307_cast_fp16)[name = tensor("pretrained_out_187_cast_fp16")]; + tensor var_3702 = const()[name = tensor("op_3702"), val = tensor([1, 1])]; + tensor var_3704 = const()[name = tensor("op_3704"), 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_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(177224768)))]; + tensor input_309_cast_fp16 = conv(dilations = var_3704, groups = var_3570, pad = input_309_pad_0, pad_type = input_309_pad_type_0, strides = var_3702, weight = layers_15_self_attn_o_proj_loraA_weight_to_fp16, x = input_307_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor var_3708 = const()[name = tensor("op_3708"), val = tensor([1, 1])]; + tensor var_3710 = const()[name = tensor("op_3710"), 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(177265792)))]; + tensor lora_out_375_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3710, groups = var_3570, pad = lora_out_373_pad_0, pad_type = lora_out_373_pad_type_0, strides = var_3708, weight = lora_out_375_weight_0_to_fp16, x = input_309_cast_fp16)[name = tensor("lora_out_375_cast_fp16")]; + tensor obj_63_cast_fp16 = add(x = pretrained_out_187_cast_fp16, y = lora_out_375_cast_fp16)[name = tensor("obj_63_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_63_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor var_3719 = const()[name = tensor("op_3719"), val = tensor([1])]; + tensor channels_mean_63_cast_fp16 = reduce_mean(axes = var_3719, keep_dims = var_3571, 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_3723 = const()[name = tensor("op_3723"), val = tensor([1])]; + tensor var_3724_cast_fp16 = reduce_mean(axes = var_3723, keep_dims = var_3571, x = zero_mean_sq_63_cast_fp16)[name = tensor("op_3724_cast_fp16")]; + tensor var_3725_to_fp16 = const()[name = tensor("op_3725_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3726_cast_fp16 = add(x = var_3724_cast_fp16, y = var_3725_to_fp16)[name = tensor("op_3726_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_3726_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 input_311_gamma_0_to_fp16 = const()[name = tensor("input_311_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177306816)))]; + tensor input_311_beta_0_to_fp16 = const()[name = tensor("input_311_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177309440)))]; + tensor input_311_epsilon_0_to_fp16 = const()[name = tensor("input_311_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_311_cast_fp16 = batch_norm(beta = input_311_beta_0_to_fp16, epsilon = input_311_epsilon_0_to_fp16, gamma = input_311_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("input_311_cast_fp16")]; + tensor var_3740 = const()[name = tensor("op_3740"), val = tensor([1, 1])]; + tensor var_3742 = const()[name = tensor("op_3742"), 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_15_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177312064))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180588928))), 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(180589056)))]; + tensor pretrained_out_189_cast_fp16 = conv(bias = layers_15_fc1_pretrained_bias_to_fp16, dilations = var_3742, groups = var_3570, pad = pretrained_out_189_pad_0, pad_type = pretrained_out_189_pad_type_0, strides = var_3740, weight = layers_15_fc1_pretrained_weight_to_fp16_palettized, x = input_311_cast_fp16)[name = tensor("pretrained_out_189_cast_fp16")]; + tensor var_3746 = const()[name = tensor("op_3746"), val = tensor([1, 1])]; + tensor var_3748 = const()[name = tensor("op_3748"), 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_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(180599360)))]; + tensor input_313_cast_fp16 = conv(dilations = var_3748, groups = var_3570, pad = input_313_pad_0, pad_type = input_313_pad_type_0, strides = var_3746, weight = layers_15_fc1_loraA_weight_to_fp16, x = input_311_cast_fp16)[name = tensor("input_313_cast_fp16")]; + tensor var_3752 = const()[name = tensor("op_3752"), val = tensor([1, 1])]; + tensor var_3754 = const()[name = tensor("op_3754"), 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(180640384)))]; + tensor lora_out_379_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_3754, groups = var_3570, pad = lora_out_377_pad_0, pad_type = lora_out_377_pad_type_0, strides = var_3752, weight = lora_out_379_weight_0_to_fp16, x = input_313_cast_fp16)[name = tensor("lora_out_379_cast_fp16")]; + tensor input_315_cast_fp16 = add(x = pretrained_out_189_cast_fp16, y = lora_out_379_cast_fp16)[name = tensor("input_315_cast_fp16")]; + tensor input_317_mode_0 = const()[name = tensor("input_317_mode_0"), val = tensor("EXACT")]; + tensor input_317_cast_fp16 = gelu(mode = input_317_mode_0, x = input_315_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor var_3766 = const()[name = tensor("op_3766"), val = tensor([1, 1])]; + tensor var_3768 = const()[name = tensor("op_3768"), 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_15_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180804288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184081152))), 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(184081280)))]; + tensor pretrained_out_191_cast_fp16 = conv(bias = layers_15_fc2_pretrained_bias_to_fp16, dilations = var_3768, groups = var_3570, pad = pretrained_out_191_pad_0, pad_type = pretrained_out_191_pad_type_0, strides = var_3766, weight = layers_15_fc2_pretrained_weight_to_fp16_palettized, x = input_317_cast_fp16)[name = tensor("pretrained_out_191_cast_fp16")]; + tensor var_3772 = const()[name = tensor("op_3772"), val = tensor([1, 1])]; + tensor var_3774 = const()[name = tensor("op_3774"), 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_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(184083904)))]; + tensor input_319_cast_fp16 = conv(dilations = var_3774, groups = var_3570, pad = input_319_pad_0, pad_type = input_319_pad_type_0, strides = var_3772, weight = layers_15_fc2_loraA_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("input_319_cast_fp16")]; + tensor var_3778 = const()[name = tensor("op_3778"), val = tensor([1, 1])]; + tensor var_3780 = const()[name = tensor("op_3780"), 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(184247808)))]; + tensor lora_out_383_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3780, groups = var_3570, pad = lora_out_381_pad_0, pad_type = lora_out_381_pad_type_0, strides = var_3778, weight = lora_out_383_weight_0_to_fp16, x = input_319_cast_fp16)[name = tensor("lora_out_383_cast_fp16")]; + tensor hidden_states_35_cast_fp16 = add(x = pretrained_out_191_cast_fp16, y = lora_out_383_cast_fp16)[name = tensor("hidden_states_35_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor var_3794 = const()[name = tensor("op_3794"), val = tensor(3)]; + tensor var_3796 = const()[name = tensor("op_3796"), val = tensor(1)]; + tensor var_3797 = const()[name = tensor("op_3797"), val = tensor(true)]; + tensor var_3807 = const()[name = tensor("op_3807"), val = tensor([1])]; + tensor channels_mean_65_cast_fp16 = reduce_mean(axes = var_3807, keep_dims = var_3797, 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_3811 = const()[name = tensor("op_3811"), val = tensor([1])]; + tensor var_3812_cast_fp16 = reduce_mean(axes = var_3811, keep_dims = var_3797, x = zero_mean_sq_65_cast_fp16)[name = tensor("op_3812_cast_fp16")]; + tensor var_3813_to_fp16 = const()[name = tensor("op_3813_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3814_cast_fp16 = add(x = var_3812_cast_fp16, y = var_3813_to_fp16)[name = tensor("op_3814_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_3814_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 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(184288832)))]; + 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(184291456)))]; + 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_65_cast_fp16)[name = tensor("obj_65_cast_fp16")]; + tensor var_3832 = const()[name = tensor("op_3832"), val = tensor([1, 1])]; + tensor var_3834 = const()[name = tensor("op_3834"), 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_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(184294080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185113344))), 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(185113472)))]; + tensor pretrained_out_193_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_3834, groups = var_3796, pad = pretrained_out_193_pad_0, pad_type = pretrained_out_193_pad_type_0, strides = var_3832, weight = layers_16_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_65_cast_fp16)[name = tensor("pretrained_out_193_cast_fp16")]; + tensor var_3838 = const()[name = tensor("op_3838"), val = tensor([1, 1])]; + tensor var_3840 = const()[name = tensor("op_3840"), val = tensor([1, 1])]; + tensor input_321_pad_type_0 = const()[name = tensor("input_321_pad_type_0"), val = tensor("custom")]; + tensor input_321_pad_0 = const()[name = tensor("input_321_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(185116096)))]; + tensor input_321_cast_fp16 = conv(dilations = var_3840, groups = var_3796, pad = input_321_pad_0, pad_type = input_321_pad_type_0, strides = var_3838, weight = layers_16_self_attn_q_proj_loraA_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("input_321_cast_fp16")]; + tensor var_3844 = const()[name = tensor("op_3844"), val = tensor([1, 1])]; + tensor var_3846 = const()[name = tensor("op_3846"), 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(185157120)))]; + tensor lora_out_387_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3846, groups = var_3796, pad = lora_out_385_pad_0, pad_type = lora_out_385_pad_type_0, strides = var_3844, weight = lora_out_387_weight_0_to_fp16, x = input_321_cast_fp16)[name = tensor("lora_out_387_cast_fp16")]; + tensor query_33_cast_fp16 = add(x = pretrained_out_193_cast_fp16, y = lora_out_387_cast_fp16)[name = tensor("query_33_cast_fp16")]; + tensor var_3856 = const()[name = tensor("op_3856"), val = tensor([1, 1])]; + tensor var_3858 = const()[name = tensor("op_3858"), 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_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(185198144))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186017408))), name = tensor("layers_16_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_195_cast_fp16 = conv(dilations = var_3858, groups = var_3796, pad = pretrained_out_195_pad_0, pad_type = pretrained_out_195_pad_type_0, strides = var_3856, weight = layers_16_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_65_cast_fp16)[name = tensor("pretrained_out_195_cast_fp16")]; + tensor var_3862 = const()[name = tensor("op_3862"), val = tensor([1, 1])]; + tensor var_3864 = const()[name = tensor("op_3864"), 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_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(186017536)))]; + tensor input_323_cast_fp16 = conv(dilations = var_3864, groups = var_3796, pad = input_323_pad_0, pad_type = input_323_pad_type_0, strides = var_3862, weight = layers_16_self_attn_k_proj_loraA_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("input_323_cast_fp16")]; + tensor var_3868 = const()[name = tensor("op_3868"), val = tensor([1, 1])]; + tensor var_3870 = const()[name = tensor("op_3870"), 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(186058560)))]; + tensor lora_out_391_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3870, groups = var_3796, pad = lora_out_389_pad_0, pad_type = lora_out_389_pad_type_0, strides = var_3868, weight = lora_out_391_weight_0_to_fp16, x = input_323_cast_fp16)[name = tensor("lora_out_391_cast_fp16")]; + tensor key_33_cast_fp16 = add(x = pretrained_out_195_cast_fp16, y = lora_out_391_cast_fp16)[name = tensor("key_33_cast_fp16")]; + tensor var_3881 = const()[name = tensor("op_3881"), val = tensor([1, 1])]; + tensor var_3883 = const()[name = tensor("op_3883"), 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_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(186099584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186918848))), 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(186918976)))]; + tensor pretrained_out_197_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_3883, groups = var_3796, pad = pretrained_out_197_pad_0, pad_type = pretrained_out_197_pad_type_0, strides = var_3881, weight = layers_16_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_65_cast_fp16)[name = tensor("pretrained_out_197_cast_fp16")]; + tensor var_3887 = const()[name = tensor("op_3887"), val = tensor([1, 1])]; + tensor var_3889 = const()[name = tensor("op_3889"), val = tensor([1, 1])]; + tensor input_325_pad_type_0 = const()[name = tensor("input_325_pad_type_0"), val = tensor("custom")]; + tensor input_325_pad_0 = const()[name = tensor("input_325_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(186921600)))]; + tensor input_325_cast_fp16 = conv(dilations = var_3889, groups = var_3796, pad = input_325_pad_0, pad_type = input_325_pad_type_0, strides = var_3887, weight = layers_16_self_attn_v_proj_loraA_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor var_3893 = const()[name = tensor("op_3893"), val = tensor([1, 1])]; + tensor var_3895 = const()[name = tensor("op_3895"), 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(186962624)))]; + tensor lora_out_395_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3895, groups = var_3796, pad = lora_out_393_pad_0, pad_type = lora_out_393_pad_type_0, strides = var_3893, weight = lora_out_395_weight_0_to_fp16, x = input_325_cast_fp16)[name = tensor("lora_out_395_cast_fp16")]; + tensor value_33_cast_fp16 = add(x = pretrained_out_197_cast_fp16, y = lora_out_395_cast_fp16)[name = tensor("value_33_cast_fp16")]; + tensor var_3902 = const()[name = tensor("op_3902"), val = tensor([1, 20, 64, -1])]; + tensor var_3903_cast_fp16 = reshape(shape = var_3902, x = query_33_cast_fp16)[name = tensor("op_3903_cast_fp16")]; + tensor var_3904_to_fp16 = const()[name = tensor("op_3904_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3905_cast_fp16 = mul(x = var_3903_cast_fp16, y = var_3904_to_fp16)[name = tensor("op_3905_cast_fp16")]; + tensor var_3906 = const()[name = tensor("op_3906"), val = tensor([1, 20, 64, -1])]; + tensor var_3907_cast_fp16 = reshape(shape = var_3906, x = key_33_cast_fp16)[name = tensor("op_3907_cast_fp16")]; + tensor mh_w_33_transpose_x_0 = const()[name = tensor("mh_w_33_transpose_x_0"), val = tensor(true)]; + tensor mh_w_33_transpose_y_0 = const()[name = tensor("mh_w_33_transpose_y_0"), val = tensor(false)]; + tensor mh_w_33_cast_fp16 = matmul(transpose_x = mh_w_33_transpose_x_0, transpose_y = mh_w_33_transpose_y_0, x = var_3905_cast_fp16, y = var_3907_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor var_3910_cast_fp16 = softmax(axis = var_3794, x = mh_w_33_cast_fp16)[name = tensor("op_3910_cast_fp16")]; + tensor var_3911 = const()[name = tensor("op_3911"), val = tensor([1, 20, 64, -1])]; + tensor var_3912_cast_fp16 = reshape(shape = var_3911, x = value_33_cast_fp16)[name = tensor("op_3912_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_3912_cast_fp16, y = var_3910_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_3915 = const()[name = tensor("op_3915"), val = tensor([1, 1280, 1, -1])]; + tensor input_327_cast_fp16 = reshape(shape = var_3915, x = attn_33_cast_fp16)[name = tensor("input_327_cast_fp16")]; + tensor var_3922 = const()[name = tensor("op_3922"), val = tensor([1, 1])]; + tensor var_3924 = const()[name = tensor("op_3924"), 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_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(187003648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187822912))), 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(187823040)))]; + tensor pretrained_out_199_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_3924, groups = var_3796, pad = pretrained_out_199_pad_0, pad_type = pretrained_out_199_pad_type_0, strides = var_3922, weight = layers_16_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_327_cast_fp16)[name = tensor("pretrained_out_199_cast_fp16")]; + tensor var_3928 = const()[name = tensor("op_3928"), val = tensor([1, 1])]; + tensor var_3930 = const()[name = tensor("op_3930"), 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_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(187825664)))]; + tensor input_329_cast_fp16 = conv(dilations = var_3930, groups = var_3796, pad = input_329_pad_0, pad_type = input_329_pad_type_0, strides = var_3928, weight = layers_16_self_attn_o_proj_loraA_weight_to_fp16, x = input_327_cast_fp16)[name = tensor("input_329_cast_fp16")]; + tensor var_3934 = const()[name = tensor("op_3934"), val = tensor([1, 1])]; + tensor var_3936 = const()[name = tensor("op_3936"), 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(187866688)))]; + tensor lora_out_399_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_3936, groups = var_3796, pad = lora_out_397_pad_0, pad_type = lora_out_397_pad_type_0, strides = var_3934, weight = lora_out_399_weight_0_to_fp16, x = input_329_cast_fp16)[name = tensor("lora_out_399_cast_fp16")]; + tensor obj_67_cast_fp16 = add(x = pretrained_out_199_cast_fp16, y = lora_out_399_cast_fp16)[name = tensor("obj_67_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = obj_67_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor var_3945 = const()[name = tensor("op_3945"), val = tensor([1])]; + tensor channels_mean_67_cast_fp16 = reduce_mean(axes = var_3945, keep_dims = var_3797, 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_3949 = const()[name = tensor("op_3949"), val = tensor([1])]; + tensor var_3950_cast_fp16 = reduce_mean(axes = var_3949, keep_dims = var_3797, x = zero_mean_sq_67_cast_fp16)[name = tensor("op_3950_cast_fp16")]; + tensor var_3951_to_fp16 = const()[name = tensor("op_3951_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3952_cast_fp16 = add(x = var_3950_cast_fp16, y = var_3951_to_fp16)[name = tensor("op_3952_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_3952_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 input_331_gamma_0_to_fp16 = const()[name = tensor("input_331_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187907712)))]; + tensor input_331_beta_0_to_fp16 = const()[name = tensor("input_331_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187910336)))]; + tensor input_331_epsilon_0_to_fp16 = const()[name = tensor("input_331_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_331_cast_fp16 = batch_norm(beta = input_331_beta_0_to_fp16, epsilon = input_331_epsilon_0_to_fp16, gamma = input_331_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("input_331_cast_fp16")]; + tensor var_3966 = const()[name = tensor("op_3966"), val = tensor([1, 1])]; + tensor var_3968 = const()[name = tensor("op_3968"), 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_16_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187912960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191189824))), 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(191189952)))]; + tensor pretrained_out_201_cast_fp16 = conv(bias = layers_16_fc1_pretrained_bias_to_fp16, dilations = var_3968, groups = var_3796, pad = pretrained_out_201_pad_0, pad_type = pretrained_out_201_pad_type_0, strides = var_3966, weight = layers_16_fc1_pretrained_weight_to_fp16_palettized, x = input_331_cast_fp16)[name = tensor("pretrained_out_201_cast_fp16")]; + tensor var_3972 = const()[name = tensor("op_3972"), val = tensor([1, 1])]; + tensor var_3974 = const()[name = tensor("op_3974"), 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_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(191200256)))]; + tensor input_333_cast_fp16 = conv(dilations = var_3974, groups = var_3796, pad = input_333_pad_0, pad_type = input_333_pad_type_0, strides = var_3972, weight = layers_16_fc1_loraA_weight_to_fp16, x = input_331_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor var_3978 = const()[name = tensor("op_3978"), val = tensor([1, 1])]; + tensor var_3980 = const()[name = tensor("op_3980"), 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(191241280)))]; + tensor lora_out_403_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_3980, groups = var_3796, pad = lora_out_401_pad_0, pad_type = lora_out_401_pad_type_0, strides = var_3978, weight = lora_out_403_weight_0_to_fp16, x = input_333_cast_fp16)[name = tensor("lora_out_403_cast_fp16")]; + tensor input_335_cast_fp16 = add(x = pretrained_out_201_cast_fp16, y = lora_out_403_cast_fp16)[name = tensor("input_335_cast_fp16")]; + tensor input_337_mode_0 = const()[name = tensor("input_337_mode_0"), val = tensor("EXACT")]; + tensor input_337_cast_fp16 = gelu(mode = input_337_mode_0, x = input_335_cast_fp16)[name = tensor("input_337_cast_fp16")]; + tensor var_3992 = const()[name = tensor("op_3992"), val = tensor([1, 1])]; + tensor var_3994 = const()[name = tensor("op_3994"), 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_16_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191405184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194682048))), 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(194682176)))]; + tensor pretrained_out_203_cast_fp16 = conv(bias = layers_16_fc2_pretrained_bias_to_fp16, dilations = var_3994, groups = var_3796, pad = pretrained_out_203_pad_0, pad_type = pretrained_out_203_pad_type_0, strides = var_3992, weight = layers_16_fc2_pretrained_weight_to_fp16_palettized, x = input_337_cast_fp16)[name = tensor("pretrained_out_203_cast_fp16")]; + tensor var_3998 = const()[name = tensor("op_3998"), val = tensor([1, 1])]; + tensor var_4000 = const()[name = tensor("op_4000"), 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_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(194684800)))]; + tensor input_339_cast_fp16 = conv(dilations = var_4000, groups = var_3796, pad = input_339_pad_0, pad_type = input_339_pad_type_0, strides = var_3998, weight = layers_16_fc2_loraA_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor var_4004 = const()[name = tensor("op_4004"), val = tensor([1, 1])]; + tensor var_4006 = const()[name = tensor("op_4006"), 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(194848704)))]; + tensor lora_out_407_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4006, groups = var_3796, pad = lora_out_405_pad_0, pad_type = lora_out_405_pad_type_0, strides = var_4004, weight = lora_out_407_weight_0_to_fp16, x = input_339_cast_fp16)[name = tensor("lora_out_407_cast_fp16")]; + tensor hidden_states_37_cast_fp16 = add(x = pretrained_out_203_cast_fp16, y = lora_out_407_cast_fp16)[name = tensor("hidden_states_37_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor var_4020 = const()[name = tensor("op_4020"), val = tensor(3)]; + tensor var_4022 = const()[name = tensor("op_4022"), val = tensor(1)]; + tensor var_4023 = const()[name = tensor("op_4023"), val = tensor(true)]; + tensor var_4033 = const()[name = tensor("op_4033"), val = tensor([1])]; + tensor channels_mean_69_cast_fp16 = reduce_mean(axes = var_4033, keep_dims = var_4023, 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_4037 = const()[name = tensor("op_4037"), val = tensor([1])]; + tensor var_4038_cast_fp16 = reduce_mean(axes = var_4037, keep_dims = var_4023, x = zero_mean_sq_69_cast_fp16)[name = tensor("op_4038_cast_fp16")]; + tensor var_4039_to_fp16 = const()[name = tensor("op_4039_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4040_cast_fp16 = add(x = var_4038_cast_fp16, y = var_4039_to_fp16)[name = tensor("op_4040_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_4040_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_69_gamma_0_to_fp16 = const()[name = tensor("obj_69_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194889728)))]; + tensor obj_69_beta_0_to_fp16 = const()[name = tensor("obj_69_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194892352)))]; + tensor obj_69_epsilon_0_to_fp16 = const()[name = tensor("obj_69_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_69_cast_fp16 = batch_norm(beta = obj_69_beta_0_to_fp16, epsilon = obj_69_epsilon_0_to_fp16, gamma = obj_69_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_69_cast_fp16")]; + tensor var_4058 = const()[name = tensor("op_4058"), val = tensor([1, 1])]; + tensor var_4060 = const()[name = tensor("op_4060"), 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_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(194894976))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195714240))), 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(195714368)))]; + tensor pretrained_out_205_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4060, groups = var_4022, pad = pretrained_out_205_pad_0, pad_type = pretrained_out_205_pad_type_0, strides = var_4058, weight = layers_17_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_69_cast_fp16)[name = tensor("pretrained_out_205_cast_fp16")]; + tensor var_4064 = const()[name = tensor("op_4064"), val = tensor([1, 1])]; + tensor var_4066 = const()[name = tensor("op_4066"), 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_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(195716992)))]; + tensor input_341_cast_fp16 = conv(dilations = var_4066, groups = var_4022, pad = input_341_pad_0, pad_type = input_341_pad_type_0, strides = var_4064, weight = layers_17_self_attn_q_proj_loraA_weight_to_fp16, x = obj_69_cast_fp16)[name = tensor("input_341_cast_fp16")]; + tensor var_4070 = const()[name = tensor("op_4070"), val = tensor([1, 1])]; + tensor var_4072 = const()[name = tensor("op_4072"), 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(195758016)))]; + tensor lora_out_411_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4072, groups = var_4022, pad = lora_out_409_pad_0, pad_type = lora_out_409_pad_type_0, strides = var_4070, weight = lora_out_411_weight_0_to_fp16, x = input_341_cast_fp16)[name = tensor("lora_out_411_cast_fp16")]; + tensor query_35_cast_fp16 = add(x = pretrained_out_205_cast_fp16, y = lora_out_411_cast_fp16)[name = tensor("query_35_cast_fp16")]; + tensor var_4082 = const()[name = tensor("op_4082"), val = tensor([1, 1])]; + tensor var_4084 = const()[name = tensor("op_4084"), 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_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(195799040))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(196618304))), name = tensor("layers_17_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_207_cast_fp16 = conv(dilations = var_4084, groups = var_4022, pad = pretrained_out_207_pad_0, pad_type = pretrained_out_207_pad_type_0, strides = var_4082, weight = layers_17_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_69_cast_fp16)[name = tensor("pretrained_out_207_cast_fp16")]; + tensor var_4088 = const()[name = tensor("op_4088"), val = tensor([1, 1])]; + tensor var_4090 = const()[name = tensor("op_4090"), 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_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(196618432)))]; + tensor input_343_cast_fp16 = conv(dilations = var_4090, groups = var_4022, pad = input_343_pad_0, pad_type = input_343_pad_type_0, strides = var_4088, weight = layers_17_self_attn_k_proj_loraA_weight_to_fp16, x = obj_69_cast_fp16)[name = tensor("input_343_cast_fp16")]; + tensor var_4094 = const()[name = tensor("op_4094"), val = tensor([1, 1])]; + tensor var_4096 = const()[name = tensor("op_4096"), 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(196659456)))]; + tensor lora_out_415_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4096, groups = var_4022, pad = lora_out_413_pad_0, pad_type = lora_out_413_pad_type_0, strides = var_4094, weight = lora_out_415_weight_0_to_fp16, x = input_343_cast_fp16)[name = tensor("lora_out_415_cast_fp16")]; + tensor key_35_cast_fp16 = add(x = pretrained_out_207_cast_fp16, y = lora_out_415_cast_fp16)[name = tensor("key_35_cast_fp16")]; + tensor var_4107 = const()[name = tensor("op_4107"), val = tensor([1, 1])]; + tensor var_4109 = const()[name = tensor("op_4109"), 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_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(196700480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197519744))), 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(197519872)))]; + tensor pretrained_out_209_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_4109, groups = var_4022, pad = pretrained_out_209_pad_0, pad_type = pretrained_out_209_pad_type_0, strides = var_4107, weight = layers_17_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_69_cast_fp16)[name = tensor("pretrained_out_209_cast_fp16")]; + tensor var_4113 = const()[name = tensor("op_4113"), val = tensor([1, 1])]; + tensor var_4115 = const()[name = tensor("op_4115"), 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_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(197522496)))]; + tensor input_345_cast_fp16 = conv(dilations = var_4115, groups = var_4022, pad = input_345_pad_0, pad_type = input_345_pad_type_0, strides = var_4113, weight = layers_17_self_attn_v_proj_loraA_weight_to_fp16, x = obj_69_cast_fp16)[name = tensor("input_345_cast_fp16")]; + tensor var_4119 = const()[name = tensor("op_4119"), val = tensor([1, 1])]; + tensor var_4121 = const()[name = tensor("op_4121"), 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(197563520)))]; + tensor lora_out_419_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4121, groups = var_4022, pad = lora_out_417_pad_0, pad_type = lora_out_417_pad_type_0, strides = var_4119, weight = lora_out_419_weight_0_to_fp16, x = input_345_cast_fp16)[name = tensor("lora_out_419_cast_fp16")]; + tensor value_35_cast_fp16 = add(x = pretrained_out_209_cast_fp16, y = lora_out_419_cast_fp16)[name = tensor("value_35_cast_fp16")]; + tensor var_4128 = const()[name = tensor("op_4128"), val = tensor([1, 20, 64, -1])]; + tensor var_4129_cast_fp16 = reshape(shape = var_4128, x = query_35_cast_fp16)[name = tensor("op_4129_cast_fp16")]; + tensor var_4130_to_fp16 = const()[name = tensor("op_4130_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4131_cast_fp16 = mul(x = var_4129_cast_fp16, y = var_4130_to_fp16)[name = tensor("op_4131_cast_fp16")]; + tensor var_4132 = const()[name = tensor("op_4132"), val = tensor([1, 20, 64, -1])]; + tensor var_4133_cast_fp16 = reshape(shape = var_4132, x = key_35_cast_fp16)[name = tensor("op_4133_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_4131_cast_fp16, y = var_4133_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; + tensor var_4136_cast_fp16 = softmax(axis = var_4020, x = mh_w_35_cast_fp16)[name = tensor("op_4136_cast_fp16")]; + tensor var_4137 = const()[name = tensor("op_4137"), val = tensor([1, 20, 64, -1])]; + tensor var_4138_cast_fp16 = reshape(shape = var_4137, x = value_35_cast_fp16)[name = tensor("op_4138_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_4138_cast_fp16, y = var_4136_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_4141 = const()[name = tensor("op_4141"), val = tensor([1, 1280, 1, -1])]; + tensor input_347_cast_fp16 = reshape(shape = var_4141, x = attn_35_cast_fp16)[name = tensor("input_347_cast_fp16")]; + tensor var_4148 = const()[name = tensor("op_4148"), val = tensor([1, 1])]; + tensor var_4150 = const()[name = tensor("op_4150"), 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_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(197604544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198423808))), 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(198423936)))]; + tensor pretrained_out_211_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4150, groups = var_4022, pad = pretrained_out_211_pad_0, pad_type = pretrained_out_211_pad_type_0, strides = var_4148, weight = layers_17_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_347_cast_fp16)[name = tensor("pretrained_out_211_cast_fp16")]; + tensor var_4154 = const()[name = tensor("op_4154"), val = tensor([1, 1])]; + tensor var_4156 = const()[name = tensor("op_4156"), 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_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(198426560)))]; + tensor input_349_cast_fp16 = conv(dilations = var_4156, groups = var_4022, pad = input_349_pad_0, pad_type = input_349_pad_type_0, strides = var_4154, weight = layers_17_self_attn_o_proj_loraA_weight_to_fp16, x = input_347_cast_fp16)[name = tensor("input_349_cast_fp16")]; + tensor var_4160 = const()[name = tensor("op_4160"), val = tensor([1, 1])]; + tensor var_4162 = const()[name = tensor("op_4162"), 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(198467584)))]; + tensor lora_out_423_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4162, groups = var_4022, pad = lora_out_421_pad_0, pad_type = lora_out_421_pad_type_0, strides = var_4160, weight = lora_out_423_weight_0_to_fp16, x = input_349_cast_fp16)[name = tensor("lora_out_423_cast_fp16")]; + tensor obj_71_cast_fp16 = add(x = pretrained_out_211_cast_fp16, y = lora_out_423_cast_fp16)[name = tensor("obj_71_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_71_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor var_4171 = const()[name = tensor("op_4171"), val = tensor([1])]; + tensor channels_mean_71_cast_fp16 = reduce_mean(axes = var_4171, keep_dims = var_4023, 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_4175 = const()[name = tensor("op_4175"), val = tensor([1])]; + tensor var_4176_cast_fp16 = reduce_mean(axes = var_4175, keep_dims = var_4023, x = zero_mean_sq_71_cast_fp16)[name = tensor("op_4176_cast_fp16")]; + tensor var_4177_to_fp16 = const()[name = tensor("op_4177_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4178_cast_fp16 = add(x = var_4176_cast_fp16, y = var_4177_to_fp16)[name = tensor("op_4178_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_4178_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(198508608)))]; + 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(198511232)))]; + 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_4192 = const()[name = tensor("op_4192"), val = tensor([1, 1])]; + tensor var_4194 = const()[name = tensor("op_4194"), 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_17_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198513856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201790720))), 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(201790848)))]; + tensor pretrained_out_213_cast_fp16 = conv(bias = layers_17_fc1_pretrained_bias_to_fp16, dilations = var_4194, groups = var_4022, pad = pretrained_out_213_pad_0, pad_type = pretrained_out_213_pad_type_0, strides = var_4192, weight = layers_17_fc1_pretrained_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = tensor("pretrained_out_213_cast_fp16")]; + tensor var_4198 = const()[name = tensor("op_4198"), val = tensor([1, 1])]; + tensor var_4200 = const()[name = tensor("op_4200"), 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_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(201801152)))]; + tensor input_353_cast_fp16 = conv(dilations = var_4200, groups = var_4022, pad = input_353_pad_0, pad_type = input_353_pad_type_0, strides = var_4198, weight = layers_17_fc1_loraA_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor var_4204 = const()[name = tensor("op_4204"), val = tensor([1, 1])]; + tensor var_4206 = const()[name = tensor("op_4206"), 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(201842176)))]; + tensor lora_out_427_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_4206, groups = var_4022, pad = lora_out_425_pad_0, pad_type = lora_out_425_pad_type_0, strides = var_4204, weight = lora_out_427_weight_0_to_fp16, x = input_353_cast_fp16)[name = tensor("lora_out_427_cast_fp16")]; + tensor input_355_cast_fp16 = add(x = pretrained_out_213_cast_fp16, y = lora_out_427_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_4218 = const()[name = tensor("op_4218"), val = tensor([1, 1])]; + tensor var_4220 = const()[name = tensor("op_4220"), 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_17_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202006080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205282944))), 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(205283072)))]; + tensor pretrained_out_215_cast_fp16 = conv(bias = layers_17_fc2_pretrained_bias_to_fp16, dilations = var_4220, groups = var_4022, pad = pretrained_out_215_pad_0, pad_type = pretrained_out_215_pad_type_0, strides = var_4218, weight = layers_17_fc2_pretrained_weight_to_fp16_palettized, x = input_357_cast_fp16)[name = tensor("pretrained_out_215_cast_fp16")]; + tensor var_4224 = const()[name = tensor("op_4224"), val = tensor([1, 1])]; + tensor var_4226 = const()[name = tensor("op_4226"), 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_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(205285696)))]; + tensor input_359_cast_fp16 = conv(dilations = var_4226, groups = var_4022, pad = input_359_pad_0, pad_type = input_359_pad_type_0, strides = var_4224, weight = layers_17_fc2_loraA_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("input_359_cast_fp16")]; + tensor var_4230 = const()[name = tensor("op_4230"), val = tensor([1, 1])]; + tensor var_4232 = const()[name = tensor("op_4232"), 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(205449600)))]; + tensor lora_out_431_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4232, groups = var_4022, pad = lora_out_429_pad_0, pad_type = lora_out_429_pad_type_0, strides = var_4230, weight = lora_out_431_weight_0_to_fp16, x = input_359_cast_fp16)[name = tensor("lora_out_431_cast_fp16")]; + tensor hidden_states_39_cast_fp16 = add(x = pretrained_out_215_cast_fp16, y = lora_out_431_cast_fp16)[name = tensor("hidden_states_39_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_4246 = const()[name = tensor("op_4246"), val = tensor(3)]; + tensor var_4248 = const()[name = tensor("op_4248"), val = tensor(1)]; + tensor var_4249 = const()[name = tensor("op_4249"), val = tensor(true)]; + tensor var_4259 = const()[name = tensor("op_4259"), val = tensor([1])]; + tensor channels_mean_73_cast_fp16 = reduce_mean(axes = var_4259, keep_dims = var_4249, 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_4263 = const()[name = tensor("op_4263"), val = tensor([1])]; + tensor var_4264_cast_fp16 = reduce_mean(axes = var_4263, keep_dims = var_4249, x = zero_mean_sq_73_cast_fp16)[name = tensor("op_4264_cast_fp16")]; + tensor var_4265_to_fp16 = const()[name = tensor("op_4265_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4266_cast_fp16 = add(x = var_4264_cast_fp16, y = var_4265_to_fp16)[name = tensor("op_4266_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_4266_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_73_gamma_0_to_fp16 = const()[name = tensor("obj_73_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205490624)))]; + tensor obj_73_beta_0_to_fp16 = const()[name = tensor("obj_73_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205493248)))]; + tensor obj_73_epsilon_0_to_fp16 = const()[name = tensor("obj_73_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_73_cast_fp16 = batch_norm(beta = obj_73_beta_0_to_fp16, epsilon = obj_73_epsilon_0_to_fp16, gamma = obj_73_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_73_cast_fp16")]; + tensor var_4284 = const()[name = tensor("op_4284"), val = tensor([1, 1])]; + tensor var_4286 = const()[name = tensor("op_4286"), 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_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(205495872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(206315136))), 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(206315264)))]; + tensor pretrained_out_217_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4286, groups = var_4248, pad = pretrained_out_217_pad_0, pad_type = pretrained_out_217_pad_type_0, strides = var_4284, weight = layers_18_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = tensor("pretrained_out_217_cast_fp16")]; + tensor var_4290 = const()[name = tensor("op_4290"), val = tensor([1, 1])]; + tensor var_4292 = const()[name = tensor("op_4292"), 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_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(206317888)))]; + tensor input_361_cast_fp16 = conv(dilations = var_4292, groups = var_4248, pad = input_361_pad_0, pad_type = input_361_pad_type_0, strides = var_4290, weight = layers_18_self_attn_q_proj_loraA_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor var_4296 = const()[name = tensor("op_4296"), val = tensor([1, 1])]; + tensor var_4298 = const()[name = tensor("op_4298"), 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(206358912)))]; + tensor lora_out_435_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4298, groups = var_4248, pad = lora_out_433_pad_0, pad_type = lora_out_433_pad_type_0, strides = var_4296, weight = lora_out_435_weight_0_to_fp16, x = input_361_cast_fp16)[name = tensor("lora_out_435_cast_fp16")]; + tensor query_37_cast_fp16 = add(x = pretrained_out_217_cast_fp16, y = lora_out_435_cast_fp16)[name = tensor("query_37_cast_fp16")]; + tensor var_4308 = const()[name = tensor("op_4308"), val = tensor([1, 1])]; + tensor var_4310 = const()[name = tensor("op_4310"), 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_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(206399936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207219200))), name = tensor("layers_18_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_219_cast_fp16 = conv(dilations = var_4310, groups = var_4248, pad = pretrained_out_219_pad_0, pad_type = pretrained_out_219_pad_type_0, strides = var_4308, weight = layers_18_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = tensor("pretrained_out_219_cast_fp16")]; + tensor var_4314 = const()[name = tensor("op_4314"), val = tensor([1, 1])]; + tensor var_4316 = const()[name = tensor("op_4316"), 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_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(207219328)))]; + tensor input_363_cast_fp16 = conv(dilations = var_4316, groups = var_4248, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = var_4314, weight = layers_18_self_attn_k_proj_loraA_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor("input_363_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 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(207260352)))]; + tensor lora_out_439_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4322, groups = var_4248, pad = lora_out_437_pad_0, pad_type = lora_out_437_pad_type_0, strides = var_4320, weight = lora_out_439_weight_0_to_fp16, x = input_363_cast_fp16)[name = tensor("lora_out_439_cast_fp16")]; + tensor key_37_cast_fp16 = add(x = pretrained_out_219_cast_fp16, y = lora_out_439_cast_fp16)[name = tensor("key_37_cast_fp16")]; + tensor var_4333 = const()[name = tensor("op_4333"), val = tensor([1, 1])]; + tensor var_4335 = const()[name = tensor("op_4335"), 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_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(207301376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208120640))), 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(208120768)))]; + tensor pretrained_out_221_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_4335, groups = var_4248, pad = pretrained_out_221_pad_0, pad_type = pretrained_out_221_pad_type_0, strides = var_4333, weight = layers_18_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_73_cast_fp16)[name = tensor("pretrained_out_221_cast_fp16")]; + tensor var_4339 = const()[name = tensor("op_4339"), val = tensor([1, 1])]; + tensor var_4341 = const()[name = tensor("op_4341"), 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_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(208123392)))]; + tensor input_365_cast_fp16 = conv(dilations = var_4341, groups = var_4248, pad = input_365_pad_0, pad_type = input_365_pad_type_0, strides = var_4339, weight = layers_18_self_attn_v_proj_loraA_weight_to_fp16, x = obj_73_cast_fp16)[name = tensor("input_365_cast_fp16")]; + tensor var_4345 = const()[name = tensor("op_4345"), val = tensor([1, 1])]; + tensor var_4347 = const()[name = tensor("op_4347"), 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(208164416)))]; + tensor lora_out_443_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4347, groups = var_4248, pad = lora_out_441_pad_0, pad_type = lora_out_441_pad_type_0, strides = var_4345, weight = lora_out_443_weight_0_to_fp16, x = input_365_cast_fp16)[name = tensor("lora_out_443_cast_fp16")]; + tensor value_37_cast_fp16 = add(x = pretrained_out_221_cast_fp16, y = lora_out_443_cast_fp16)[name = tensor("value_37_cast_fp16")]; + tensor var_4354 = const()[name = tensor("op_4354"), val = tensor([1, 20, 64, -1])]; + tensor var_4355_cast_fp16 = reshape(shape = var_4354, x = query_37_cast_fp16)[name = tensor("op_4355_cast_fp16")]; + tensor var_4356_to_fp16 = const()[name = tensor("op_4356_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4357_cast_fp16 = mul(x = var_4355_cast_fp16, y = var_4356_to_fp16)[name = tensor("op_4357_cast_fp16")]; + tensor var_4358 = const()[name = tensor("op_4358"), val = tensor([1, 20, 64, -1])]; + tensor var_4359_cast_fp16 = reshape(shape = var_4358, x = key_37_cast_fp16)[name = tensor("op_4359_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_4357_cast_fp16, y = var_4359_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; + tensor var_4362_cast_fp16 = softmax(axis = var_4246, x = mh_w_37_cast_fp16)[name = tensor("op_4362_cast_fp16")]; + tensor var_4363 = const()[name = tensor("op_4363"), val = tensor([1, 20, 64, -1])]; + tensor var_4364_cast_fp16 = reshape(shape = var_4363, x = value_37_cast_fp16)[name = tensor("op_4364_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_4364_cast_fp16, y = var_4362_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_4367 = const()[name = tensor("op_4367"), val = tensor([1, 1280, 1, -1])]; + tensor input_367_cast_fp16 = reshape(shape = var_4367, x = attn_37_cast_fp16)[name = tensor("input_367_cast_fp16")]; + tensor var_4374 = const()[name = tensor("op_4374"), val = tensor([1, 1])]; + tensor var_4376 = const()[name = tensor("op_4376"), 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_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(208205440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209024704))), 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(209024832)))]; + tensor pretrained_out_223_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4376, groups = var_4248, pad = pretrained_out_223_pad_0, pad_type = pretrained_out_223_pad_type_0, strides = var_4374, weight = layers_18_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_367_cast_fp16)[name = tensor("pretrained_out_223_cast_fp16")]; + tensor var_4380 = const()[name = tensor("op_4380"), val = tensor([1, 1])]; + tensor var_4382 = const()[name = tensor("op_4382"), 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_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(209027456)))]; + tensor input_369_cast_fp16 = conv(dilations = var_4382, groups = var_4248, pad = input_369_pad_0, pad_type = input_369_pad_type_0, strides = var_4380, weight = layers_18_self_attn_o_proj_loraA_weight_to_fp16, x = input_367_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor var_4386 = const()[name = tensor("op_4386"), val = tensor([1, 1])]; + tensor var_4388 = const()[name = tensor("op_4388"), 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(209068480)))]; + tensor lora_out_447_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4388, groups = var_4248, pad = lora_out_445_pad_0, pad_type = lora_out_445_pad_type_0, strides = var_4386, weight = lora_out_447_weight_0_to_fp16, x = input_369_cast_fp16)[name = tensor("lora_out_447_cast_fp16")]; + tensor obj_75_cast_fp16 = add(x = pretrained_out_223_cast_fp16, y = lora_out_447_cast_fp16)[name = tensor("obj_75_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_75_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor var_4397 = const()[name = tensor("op_4397"), val = tensor([1])]; + tensor channels_mean_75_cast_fp16 = reduce_mean(axes = var_4397, keep_dims = var_4249, 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_4401 = const()[name = tensor("op_4401"), val = tensor([1])]; + tensor var_4402_cast_fp16 = reduce_mean(axes = var_4401, keep_dims = var_4249, x = zero_mean_sq_75_cast_fp16)[name = tensor("op_4402_cast_fp16")]; + tensor var_4403_to_fp16 = const()[name = tensor("op_4403_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4404_cast_fp16 = add(x = var_4402_cast_fp16, y = var_4403_to_fp16)[name = tensor("op_4404_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_4404_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 input_371_gamma_0_to_fp16 = const()[name = tensor("input_371_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209109504)))]; + tensor input_371_beta_0_to_fp16 = const()[name = tensor("input_371_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209112128)))]; + tensor input_371_epsilon_0_to_fp16 = const()[name = tensor("input_371_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_371_cast_fp16 = batch_norm(beta = input_371_beta_0_to_fp16, epsilon = input_371_epsilon_0_to_fp16, gamma = input_371_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("input_371_cast_fp16")]; + tensor var_4418 = const()[name = tensor("op_4418"), val = tensor([1, 1])]; + tensor var_4420 = const()[name = tensor("op_4420"), 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_18_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209114752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212391616))), 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(212391744)))]; + tensor pretrained_out_225_cast_fp16 = conv(bias = layers_18_fc1_pretrained_bias_to_fp16, dilations = var_4420, groups = var_4248, pad = pretrained_out_225_pad_0, pad_type = pretrained_out_225_pad_type_0, strides = var_4418, weight = layers_18_fc1_pretrained_weight_to_fp16_palettized, x = input_371_cast_fp16)[name = tensor("pretrained_out_225_cast_fp16")]; + tensor var_4424 = const()[name = tensor("op_4424"), val = tensor([1, 1])]; + tensor var_4426 = const()[name = tensor("op_4426"), 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_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(212402048)))]; + tensor input_373_cast_fp16 = conv(dilations = var_4426, groups = var_4248, pad = input_373_pad_0, pad_type = input_373_pad_type_0, strides = var_4424, weight = layers_18_fc1_loraA_weight_to_fp16, x = input_371_cast_fp16)[name = tensor("input_373_cast_fp16")]; + tensor var_4430 = const()[name = tensor("op_4430"), val = tensor([1, 1])]; + tensor var_4432 = const()[name = tensor("op_4432"), 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(212443072)))]; + tensor lora_out_451_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_4432, groups = var_4248, pad = lora_out_449_pad_0, pad_type = lora_out_449_pad_type_0, strides = var_4430, weight = lora_out_451_weight_0_to_fp16, x = input_373_cast_fp16)[name = tensor("lora_out_451_cast_fp16")]; + tensor input_375_cast_fp16 = add(x = pretrained_out_225_cast_fp16, y = lora_out_451_cast_fp16)[name = tensor("input_375_cast_fp16")]; + tensor input_377_mode_0 = const()[name = tensor("input_377_mode_0"), val = tensor("EXACT")]; + tensor input_377_cast_fp16 = gelu(mode = input_377_mode_0, x = input_375_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor var_4444 = const()[name = tensor("op_4444"), val = tensor([1, 1])]; + tensor var_4446 = const()[name = tensor("op_4446"), 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_18_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212606976))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(215883840))), 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(215883968)))]; + tensor pretrained_out_227_cast_fp16 = conv(bias = layers_18_fc2_pretrained_bias_to_fp16, dilations = var_4446, groups = var_4248, pad = pretrained_out_227_pad_0, pad_type = pretrained_out_227_pad_type_0, strides = var_4444, weight = layers_18_fc2_pretrained_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = tensor("pretrained_out_227_cast_fp16")]; + tensor var_4450 = const()[name = tensor("op_4450"), val = tensor([1, 1])]; + tensor var_4452 = const()[name = tensor("op_4452"), 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_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(215886592)))]; + tensor input_379_cast_fp16 = conv(dilations = var_4452, groups = var_4248, pad = input_379_pad_0, pad_type = input_379_pad_type_0, strides = var_4450, weight = layers_18_fc2_loraA_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("input_379_cast_fp16")]; + tensor var_4456 = const()[name = tensor("op_4456"), val = tensor([1, 1])]; + tensor var_4458 = const()[name = tensor("op_4458"), 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(216050496)))]; + tensor lora_out_455_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4458, groups = var_4248, pad = lora_out_453_pad_0, pad_type = lora_out_453_pad_type_0, strides = var_4456, weight = lora_out_455_weight_0_to_fp16, x = input_379_cast_fp16)[name = tensor("lora_out_455_cast_fp16")]; + tensor hidden_states_41_cast_fp16 = add(x = pretrained_out_227_cast_fp16, y = lora_out_455_cast_fp16)[name = tensor("hidden_states_41_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor var_4472 = const()[name = tensor("op_4472"), val = tensor(3)]; + tensor var_4474 = const()[name = tensor("op_4474"), val = tensor(1)]; + tensor var_4475 = const()[name = tensor("op_4475"), val = tensor(true)]; + tensor var_4485 = const()[name = tensor("op_4485"), val = tensor([1])]; + tensor channels_mean_77_cast_fp16 = reduce_mean(axes = var_4485, keep_dims = var_4475, 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_4489 = const()[name = tensor("op_4489"), val = tensor([1])]; + tensor var_4490_cast_fp16 = reduce_mean(axes = var_4489, keep_dims = var_4475, x = zero_mean_sq_77_cast_fp16)[name = tensor("op_4490_cast_fp16")]; + tensor var_4491_to_fp16 = const()[name = tensor("op_4491_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4492_cast_fp16 = add(x = var_4490_cast_fp16, y = var_4491_to_fp16)[name = tensor("op_4492_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_4492_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 obj_77_gamma_0_to_fp16 = const()[name = tensor("obj_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216091520)))]; + tensor obj_77_beta_0_to_fp16 = const()[name = tensor("obj_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216094144)))]; + tensor obj_77_epsilon_0_to_fp16 = const()[name = tensor("obj_77_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_77_cast_fp16 = batch_norm(beta = obj_77_beta_0_to_fp16, epsilon = obj_77_epsilon_0_to_fp16, gamma = obj_77_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("obj_77_cast_fp16")]; + tensor var_4510 = const()[name = tensor("op_4510"), val = tensor([1, 1])]; + tensor var_4512 = const()[name = tensor("op_4512"), 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_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(216096768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216916032))), 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(216916160)))]; + tensor pretrained_out_229_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4512, groups = var_4474, pad = pretrained_out_229_pad_0, pad_type = pretrained_out_229_pad_type_0, strides = var_4510, weight = layers_19_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_77_cast_fp16)[name = tensor("pretrained_out_229_cast_fp16")]; + tensor var_4516 = const()[name = tensor("op_4516"), val = tensor([1, 1])]; + tensor var_4518 = const()[name = tensor("op_4518"), val = tensor([1, 1])]; + tensor input_381_pad_type_0 = const()[name = tensor("input_381_pad_type_0"), val = tensor("custom")]; + tensor input_381_pad_0 = const()[name = tensor("input_381_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(216918784)))]; + tensor input_381_cast_fp16 = conv(dilations = var_4518, groups = var_4474, pad = input_381_pad_0, pad_type = input_381_pad_type_0, strides = var_4516, weight = layers_19_self_attn_q_proj_loraA_weight_to_fp16, x = obj_77_cast_fp16)[name = tensor("input_381_cast_fp16")]; + tensor var_4522 = const()[name = tensor("op_4522"), val = tensor([1, 1])]; + tensor var_4524 = const()[name = tensor("op_4524"), 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(216959808)))]; + tensor lora_out_459_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4524, groups = var_4474, pad = lora_out_457_pad_0, pad_type = lora_out_457_pad_type_0, strides = var_4522, weight = lora_out_459_weight_0_to_fp16, x = input_381_cast_fp16)[name = tensor("lora_out_459_cast_fp16")]; + tensor query_39_cast_fp16 = add(x = pretrained_out_229_cast_fp16, y = lora_out_459_cast_fp16)[name = tensor("query_39_cast_fp16")]; + tensor var_4534 = const()[name = tensor("op_4534"), val = tensor([1, 1])]; + tensor var_4536 = const()[name = tensor("op_4536"), 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_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(217000832))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217820096))), name = tensor("layers_19_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_231_cast_fp16 = conv(dilations = var_4536, groups = var_4474, pad = pretrained_out_231_pad_0, pad_type = pretrained_out_231_pad_type_0, strides = var_4534, weight = layers_19_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_77_cast_fp16)[name = tensor("pretrained_out_231_cast_fp16")]; + tensor var_4540 = const()[name = tensor("op_4540"), val = tensor([1, 1])]; + tensor var_4542 = const()[name = tensor("op_4542"), 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_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(217820224)))]; + tensor input_383_cast_fp16 = conv(dilations = var_4542, groups = var_4474, pad = input_383_pad_0, pad_type = input_383_pad_type_0, strides = var_4540, weight = layers_19_self_attn_k_proj_loraA_weight_to_fp16, x = obj_77_cast_fp16)[name = tensor("input_383_cast_fp16")]; + tensor var_4546 = const()[name = tensor("op_4546"), val = tensor([1, 1])]; + tensor var_4548 = const()[name = tensor("op_4548"), 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(217861248)))]; + tensor lora_out_463_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4548, groups = var_4474, pad = lora_out_461_pad_0, pad_type = lora_out_461_pad_type_0, strides = var_4546, weight = lora_out_463_weight_0_to_fp16, x = input_383_cast_fp16)[name = tensor("lora_out_463_cast_fp16")]; + tensor key_39_cast_fp16 = add(x = pretrained_out_231_cast_fp16, y = lora_out_463_cast_fp16)[name = tensor("key_39_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_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_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(217902272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218721536))), 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(218721664)))]; + tensor pretrained_out_233_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_4561, groups = var_4474, pad = pretrained_out_233_pad_0, pad_type = pretrained_out_233_pad_type_0, strides = var_4559, weight = layers_19_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_77_cast_fp16)[name = tensor("pretrained_out_233_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_385_pad_type_0 = const()[name = tensor("input_385_pad_type_0"), val = tensor("custom")]; + tensor input_385_pad_0 = const()[name = tensor("input_385_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(218724288)))]; + tensor input_385_cast_fp16 = conv(dilations = var_4567, groups = var_4474, pad = input_385_pad_0, pad_type = input_385_pad_type_0, strides = var_4565, weight = layers_19_self_attn_v_proj_loraA_weight_to_fp16, x = obj_77_cast_fp16)[name = tensor("input_385_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_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(218765312)))]; + tensor lora_out_467_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4573, groups = var_4474, pad = lora_out_465_pad_0, pad_type = lora_out_465_pad_type_0, strides = var_4571, weight = lora_out_467_weight_0_to_fp16, x = input_385_cast_fp16)[name = tensor("lora_out_467_cast_fp16")]; + tensor value_39_cast_fp16 = add(x = pretrained_out_233_cast_fp16, y = lora_out_467_cast_fp16)[name = tensor("value_39_cast_fp16")]; + tensor var_4580 = const()[name = tensor("op_4580"), val = tensor([1, 20, 64, -1])]; + tensor var_4581_cast_fp16 = reshape(shape = var_4580, x = query_39_cast_fp16)[name = tensor("op_4581_cast_fp16")]; + tensor var_4582_to_fp16 = const()[name = tensor("op_4582_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4583_cast_fp16 = mul(x = var_4581_cast_fp16, y = var_4582_to_fp16)[name = tensor("op_4583_cast_fp16")]; + tensor var_4584 = const()[name = tensor("op_4584"), val = tensor([1, 20, 64, -1])]; + tensor var_4585_cast_fp16 = reshape(shape = var_4584, x = key_39_cast_fp16)[name = tensor("op_4585_cast_fp16")]; + tensor mh_w_39_transpose_x_0 = const()[name = tensor("mh_w_39_transpose_x_0"), val = tensor(true)]; + tensor mh_w_39_transpose_y_0 = const()[name = tensor("mh_w_39_transpose_y_0"), val = tensor(false)]; + tensor mh_w_39_cast_fp16 = matmul(transpose_x = mh_w_39_transpose_x_0, transpose_y = mh_w_39_transpose_y_0, x = var_4583_cast_fp16, y = var_4585_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; + tensor var_4588_cast_fp16 = softmax(axis = var_4472, x = mh_w_39_cast_fp16)[name = tensor("op_4588_cast_fp16")]; + tensor var_4589 = const()[name = tensor("op_4589"), val = tensor([1, 20, 64, -1])]; + tensor var_4590_cast_fp16 = reshape(shape = var_4589, x = value_39_cast_fp16)[name = tensor("op_4590_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_4590_cast_fp16, y = var_4588_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_4593 = const()[name = tensor("op_4593"), val = tensor([1, 1280, 1, -1])]; + tensor input_387_cast_fp16 = reshape(shape = var_4593, x = attn_39_cast_fp16)[name = tensor("input_387_cast_fp16")]; + tensor var_4600 = const()[name = tensor("op_4600"), val = tensor([1, 1])]; + tensor var_4602 = const()[name = tensor("op_4602"), 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_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(218806336))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219625600))), 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(219625728)))]; + tensor pretrained_out_235_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4602, groups = var_4474, pad = pretrained_out_235_pad_0, pad_type = pretrained_out_235_pad_type_0, strides = var_4600, weight = layers_19_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_387_cast_fp16)[name = tensor("pretrained_out_235_cast_fp16")]; + tensor var_4606 = const()[name = tensor("op_4606"), val = tensor([1, 1])]; + tensor var_4608 = const()[name = tensor("op_4608"), 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_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(219628352)))]; + tensor input_389_cast_fp16 = conv(dilations = var_4608, groups = var_4474, pad = input_389_pad_0, pad_type = input_389_pad_type_0, strides = var_4606, weight = layers_19_self_attn_o_proj_loraA_weight_to_fp16, x = input_387_cast_fp16)[name = tensor("input_389_cast_fp16")]; + tensor var_4612 = const()[name = tensor("op_4612"), val = tensor([1, 1])]; + tensor var_4614 = const()[name = tensor("op_4614"), 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(219669376)))]; + tensor lora_out_471_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4614, groups = var_4474, pad = lora_out_469_pad_0, pad_type = lora_out_469_pad_type_0, strides = var_4612, weight = lora_out_471_weight_0_to_fp16, x = input_389_cast_fp16)[name = tensor("lora_out_471_cast_fp16")]; + tensor obj_79_cast_fp16 = add(x = pretrained_out_235_cast_fp16, y = lora_out_471_cast_fp16)[name = tensor("obj_79_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = obj_79_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor var_4623 = const()[name = tensor("op_4623"), val = tensor([1])]; + tensor channels_mean_79_cast_fp16 = reduce_mean(axes = var_4623, keep_dims = var_4475, 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_4627 = const()[name = tensor("op_4627"), val = tensor([1])]; + tensor var_4628_cast_fp16 = reduce_mean(axes = var_4627, keep_dims = var_4475, x = zero_mean_sq_79_cast_fp16)[name = tensor("op_4628_cast_fp16")]; + tensor var_4629_to_fp16 = const()[name = tensor("op_4629_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4630_cast_fp16 = add(x = var_4628_cast_fp16, y = var_4629_to_fp16)[name = tensor("op_4630_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_4630_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 input_391_gamma_0_to_fp16 = const()[name = tensor("input_391_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219710400)))]; + tensor input_391_beta_0_to_fp16 = const()[name = tensor("input_391_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219713024)))]; + tensor input_391_epsilon_0_to_fp16 = const()[name = tensor("input_391_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_391_cast_fp16 = batch_norm(beta = input_391_beta_0_to_fp16, epsilon = input_391_epsilon_0_to_fp16, gamma = input_391_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("input_391_cast_fp16")]; + tensor var_4644 = const()[name = tensor("op_4644"), val = tensor([1, 1])]; + tensor var_4646 = const()[name = tensor("op_4646"), 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_19_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219715648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222992512))), 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(222992640)))]; + tensor pretrained_out_237_cast_fp16 = conv(bias = layers_19_fc1_pretrained_bias_to_fp16, dilations = var_4646, groups = var_4474, pad = pretrained_out_237_pad_0, pad_type = pretrained_out_237_pad_type_0, strides = var_4644, weight = layers_19_fc1_pretrained_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = tensor("pretrained_out_237_cast_fp16")]; + tensor var_4650 = const()[name = tensor("op_4650"), val = tensor([1, 1])]; + tensor var_4652 = const()[name = tensor("op_4652"), 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_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(223002944)))]; + tensor input_393_cast_fp16 = conv(dilations = var_4652, groups = var_4474, pad = input_393_pad_0, pad_type = input_393_pad_type_0, strides = var_4650, weight = layers_19_fc1_loraA_weight_to_fp16, x = input_391_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor var_4656 = const()[name = tensor("op_4656"), val = tensor([1, 1])]; + tensor var_4658 = const()[name = tensor("op_4658"), 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(223043968)))]; + tensor lora_out_475_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_4658, groups = var_4474, pad = lora_out_473_pad_0, pad_type = lora_out_473_pad_type_0, strides = var_4656, weight = lora_out_475_weight_0_to_fp16, x = input_393_cast_fp16)[name = tensor("lora_out_475_cast_fp16")]; + tensor input_395_cast_fp16 = add(x = pretrained_out_237_cast_fp16, y = lora_out_475_cast_fp16)[name = tensor("input_395_cast_fp16")]; + tensor input_397_mode_0 = const()[name = tensor("input_397_mode_0"), val = tensor("EXACT")]; + tensor input_397_cast_fp16 = gelu(mode = input_397_mode_0, x = input_395_cast_fp16)[name = tensor("input_397_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([1, 1])]; + tensor var_4672 = const()[name = tensor("op_4672"), 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_19_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223207872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226484736))), 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(226484864)))]; + tensor pretrained_out_239_cast_fp16 = conv(bias = layers_19_fc2_pretrained_bias_to_fp16, dilations = var_4672, groups = var_4474, pad = pretrained_out_239_pad_0, pad_type = pretrained_out_239_pad_type_0, strides = var_4670, weight = layers_19_fc2_pretrained_weight_to_fp16_palettized, x = input_397_cast_fp16)[name = tensor("pretrained_out_239_cast_fp16")]; + tensor var_4676 = const()[name = tensor("op_4676"), val = tensor([1, 1])]; + tensor var_4678 = const()[name = tensor("op_4678"), 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_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(226487488)))]; + tensor input_399_cast_fp16 = conv(dilations = var_4678, groups = var_4474, pad = input_399_pad_0, pad_type = input_399_pad_type_0, strides = var_4676, weight = layers_19_fc2_loraA_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("input_399_cast_fp16")]; + tensor var_4682 = const()[name = tensor("op_4682"), val = tensor([1, 1])]; + tensor var_4684 = const()[name = tensor("op_4684"), 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(226651392)))]; + tensor lora_out_479_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4684, groups = var_4474, pad = lora_out_477_pad_0, pad_type = lora_out_477_pad_type_0, strides = var_4682, weight = lora_out_479_weight_0_to_fp16, x = input_399_cast_fp16)[name = tensor("lora_out_479_cast_fp16")]; + tensor hidden_states_43_cast_fp16 = add(x = pretrained_out_239_cast_fp16, y = lora_out_479_cast_fp16)[name = tensor("hidden_states_43_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor var_4698 = const()[name = tensor("op_4698"), val = tensor(3)]; + tensor var_4700 = const()[name = tensor("op_4700"), val = tensor(1)]; + tensor var_4701 = const()[name = tensor("op_4701"), val = tensor(true)]; + tensor var_4711 = const()[name = tensor("op_4711"), val = tensor([1])]; + tensor channels_mean_81_cast_fp16 = reduce_mean(axes = var_4711, keep_dims = var_4701, 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_4715 = const()[name = tensor("op_4715"), val = tensor([1])]; + tensor var_4716_cast_fp16 = reduce_mean(axes = var_4715, keep_dims = var_4701, x = zero_mean_sq_81_cast_fp16)[name = tensor("op_4716_cast_fp16")]; + tensor var_4717_to_fp16 = const()[name = tensor("op_4717_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4718_cast_fp16 = add(x = var_4716_cast_fp16, y = var_4717_to_fp16)[name = tensor("op_4718_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_4718_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_81_gamma_0_to_fp16 = const()[name = tensor("obj_81_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226692416)))]; + tensor obj_81_beta_0_to_fp16 = const()[name = tensor("obj_81_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226695040)))]; + tensor obj_81_epsilon_0_to_fp16 = const()[name = tensor("obj_81_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_81_cast_fp16 = batch_norm(beta = obj_81_beta_0_to_fp16, epsilon = obj_81_epsilon_0_to_fp16, gamma = obj_81_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_81_cast_fp16")]; + tensor var_4736 = const()[name = tensor("op_4736"), val = tensor([1, 1])]; + tensor var_4738 = const()[name = tensor("op_4738"), 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_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(226697664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227516928))), 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(227517056)))]; + tensor pretrained_out_241_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4738, groups = var_4700, pad = pretrained_out_241_pad_0, pad_type = pretrained_out_241_pad_type_0, strides = var_4736, weight = layers_20_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_81_cast_fp16)[name = tensor("pretrained_out_241_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 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_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(227519680)))]; + tensor input_401_cast_fp16 = conv(dilations = var_4744, groups = var_4700, pad = input_401_pad_0, pad_type = input_401_pad_type_0, strides = var_4742, weight = layers_20_self_attn_q_proj_loraA_weight_to_fp16, x = obj_81_cast_fp16)[name = tensor("input_401_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 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(227560704)))]; + tensor lora_out_483_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4750, groups = var_4700, pad = lora_out_481_pad_0, pad_type = lora_out_481_pad_type_0, strides = var_4748, weight = lora_out_483_weight_0_to_fp16, x = input_401_cast_fp16)[name = tensor("lora_out_483_cast_fp16")]; + tensor query_41_cast_fp16 = add(x = pretrained_out_241_cast_fp16, y = lora_out_483_cast_fp16)[name = tensor("query_41_cast_fp16")]; + tensor var_4760 = const()[name = tensor("op_4760"), val = tensor([1, 1])]; + tensor var_4762 = const()[name = tensor("op_4762"), 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_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(227601728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(228420992))), name = tensor("layers_20_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_243_cast_fp16 = conv(dilations = var_4762, groups = var_4700, pad = pretrained_out_243_pad_0, pad_type = pretrained_out_243_pad_type_0, strides = var_4760, weight = layers_20_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_81_cast_fp16)[name = tensor("pretrained_out_243_cast_fp16")]; + tensor var_4766 = const()[name = tensor("op_4766"), val = tensor([1, 1])]; + tensor var_4768 = const()[name = tensor("op_4768"), 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_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(228421120)))]; + tensor input_403_cast_fp16 = conv(dilations = var_4768, groups = var_4700, pad = input_403_pad_0, pad_type = input_403_pad_type_0, strides = var_4766, weight = layers_20_self_attn_k_proj_loraA_weight_to_fp16, x = obj_81_cast_fp16)[name = tensor("input_403_cast_fp16")]; + tensor var_4772 = const()[name = tensor("op_4772"), val = tensor([1, 1])]; + tensor var_4774 = const()[name = tensor("op_4774"), 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(228462144)))]; + tensor lora_out_487_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4774, groups = var_4700, pad = lora_out_485_pad_0, pad_type = lora_out_485_pad_type_0, strides = var_4772, weight = lora_out_487_weight_0_to_fp16, x = input_403_cast_fp16)[name = tensor("lora_out_487_cast_fp16")]; + tensor key_41_cast_fp16 = add(x = pretrained_out_243_cast_fp16, y = lora_out_487_cast_fp16)[name = tensor("key_41_cast_fp16")]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1])]; + tensor var_4787 = const()[name = tensor("op_4787"), 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_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(228503168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229322432))), 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(229322560)))]; + tensor pretrained_out_245_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_4787, groups = var_4700, pad = pretrained_out_245_pad_0, pad_type = pretrained_out_245_pad_type_0, strides = var_4785, weight = layers_20_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_81_cast_fp16)[name = tensor("pretrained_out_245_cast_fp16")]; + tensor var_4791 = const()[name = tensor("op_4791"), val = tensor([1, 1])]; + tensor var_4793 = const()[name = tensor("op_4793"), 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_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(229325184)))]; + tensor input_405_cast_fp16 = conv(dilations = var_4793, groups = var_4700, pad = input_405_pad_0, pad_type = input_405_pad_type_0, strides = var_4791, weight = layers_20_self_attn_v_proj_loraA_weight_to_fp16, x = obj_81_cast_fp16)[name = tensor("input_405_cast_fp16")]; + tensor var_4797 = const()[name = tensor("op_4797"), val = tensor([1, 1])]; + tensor var_4799 = const()[name = tensor("op_4799"), 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(229366208)))]; + tensor lora_out_491_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4799, groups = var_4700, pad = lora_out_489_pad_0, pad_type = lora_out_489_pad_type_0, strides = var_4797, weight = lora_out_491_weight_0_to_fp16, x = input_405_cast_fp16)[name = tensor("lora_out_491_cast_fp16")]; + tensor value_41_cast_fp16 = add(x = pretrained_out_245_cast_fp16, y = lora_out_491_cast_fp16)[name = tensor("value_41_cast_fp16")]; + tensor var_4806 = const()[name = tensor("op_4806"), val = tensor([1, 20, 64, -1])]; + tensor var_4807_cast_fp16 = reshape(shape = var_4806, x = query_41_cast_fp16)[name = tensor("op_4807_cast_fp16")]; + tensor var_4808_to_fp16 = const()[name = tensor("op_4808_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4809_cast_fp16 = mul(x = var_4807_cast_fp16, y = var_4808_to_fp16)[name = tensor("op_4809_cast_fp16")]; + tensor var_4810 = const()[name = tensor("op_4810"), val = tensor([1, 20, 64, -1])]; + tensor var_4811_cast_fp16 = reshape(shape = var_4810, x = key_41_cast_fp16)[name = tensor("op_4811_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_4809_cast_fp16, y = var_4811_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; + tensor var_4814_cast_fp16 = softmax(axis = var_4698, x = mh_w_41_cast_fp16)[name = tensor("op_4814_cast_fp16")]; + tensor var_4815 = const()[name = tensor("op_4815"), val = tensor([1, 20, 64, -1])]; + tensor var_4816_cast_fp16 = reshape(shape = var_4815, x = value_41_cast_fp16)[name = tensor("op_4816_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_4816_cast_fp16, y = var_4814_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_4819 = const()[name = tensor("op_4819"), val = tensor([1, 1280, 1, -1])]; + tensor input_407_cast_fp16 = reshape(shape = var_4819, x = attn_41_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor var_4826 = const()[name = tensor("op_4826"), val = tensor([1, 1])]; + tensor var_4828 = const()[name = tensor("op_4828"), 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_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(229407232))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230226496))), 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(230226624)))]; + tensor pretrained_out_247_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_4828, groups = var_4700, pad = pretrained_out_247_pad_0, pad_type = pretrained_out_247_pad_type_0, strides = var_4826, weight = layers_20_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = tensor("pretrained_out_247_cast_fp16")]; + tensor var_4832 = const()[name = tensor("op_4832"), val = tensor([1, 1])]; + tensor var_4834 = const()[name = tensor("op_4834"), 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_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(230229248)))]; + tensor input_409_cast_fp16 = conv(dilations = var_4834, groups = var_4700, pad = input_409_pad_0, pad_type = input_409_pad_type_0, strides = var_4832, weight = layers_20_self_attn_o_proj_loraA_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("input_409_cast_fp16")]; + tensor var_4838 = const()[name = tensor("op_4838"), val = tensor([1, 1])]; + tensor var_4840 = const()[name = tensor("op_4840"), 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(230270272)))]; + tensor lora_out_495_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4840, groups = var_4700, pad = lora_out_493_pad_0, pad_type = lora_out_493_pad_type_0, strides = var_4838, weight = lora_out_495_weight_0_to_fp16, x = input_409_cast_fp16)[name = tensor("lora_out_495_cast_fp16")]; + tensor obj_83_cast_fp16 = add(x = pretrained_out_247_cast_fp16, y = lora_out_495_cast_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_83_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor var_4849 = const()[name = tensor("op_4849"), val = tensor([1])]; + tensor channels_mean_83_cast_fp16 = reduce_mean(axes = var_4849, keep_dims = var_4701, 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_4853 = const()[name = tensor("op_4853"), val = tensor([1])]; + tensor var_4854_cast_fp16 = reduce_mean(axes = var_4853, keep_dims = var_4701, x = zero_mean_sq_83_cast_fp16)[name = tensor("op_4854_cast_fp16")]; + tensor var_4855_to_fp16 = const()[name = tensor("op_4855_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4856_cast_fp16 = add(x = var_4854_cast_fp16, y = var_4855_to_fp16)[name = tensor("op_4856_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_4856_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(230311296)))]; + 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(230313920)))]; + 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_4870 = const()[name = tensor("op_4870"), val = tensor([1, 1])]; + tensor var_4872 = const()[name = tensor("op_4872"), 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_20_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230316544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233593408))), 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(233593536)))]; + tensor pretrained_out_249_cast_fp16 = conv(bias = layers_20_fc1_pretrained_bias_to_fp16, dilations = var_4872, groups = var_4700, pad = pretrained_out_249_pad_0, pad_type = pretrained_out_249_pad_type_0, strides = var_4870, weight = layers_20_fc1_pretrained_weight_to_fp16_palettized, x = input_411_cast_fp16)[name = tensor("pretrained_out_249_cast_fp16")]; + tensor var_4876 = const()[name = tensor("op_4876"), val = tensor([1, 1])]; + tensor var_4878 = const()[name = tensor("op_4878"), 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_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(233603840)))]; + tensor input_413_cast_fp16 = conv(dilations = var_4878, groups = var_4700, pad = input_413_pad_0, pad_type = input_413_pad_type_0, strides = var_4876, weight = layers_20_fc1_loraA_weight_to_fp16, x = input_411_cast_fp16)[name = tensor("input_413_cast_fp16")]; + tensor var_4882 = const()[name = tensor("op_4882"), val = tensor([1, 1])]; + tensor var_4884 = const()[name = tensor("op_4884"), 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(233644864)))]; + tensor lora_out_499_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_4884, groups = var_4700, pad = lora_out_497_pad_0, pad_type = lora_out_497_pad_type_0, strides = var_4882, weight = lora_out_499_weight_0_to_fp16, x = input_413_cast_fp16)[name = tensor("lora_out_499_cast_fp16")]; + tensor input_415_cast_fp16 = add(x = pretrained_out_249_cast_fp16, y = lora_out_499_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_4896 = const()[name = tensor("op_4896"), val = tensor([1, 1])]; + tensor var_4898 = const()[name = tensor("op_4898"), 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_20_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233808768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237085632))), 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(237085760)))]; + tensor pretrained_out_251_cast_fp16 = conv(bias = layers_20_fc2_pretrained_bias_to_fp16, dilations = var_4898, groups = var_4700, pad = pretrained_out_251_pad_0, pad_type = pretrained_out_251_pad_type_0, strides = var_4896, weight = layers_20_fc2_pretrained_weight_to_fp16_palettized, x = input_417_cast_fp16)[name = tensor("pretrained_out_251_cast_fp16")]; + tensor var_4902 = const()[name = tensor("op_4902"), val = tensor([1, 1])]; + tensor var_4904 = const()[name = tensor("op_4904"), 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_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(237088384)))]; + tensor input_419_cast_fp16 = conv(dilations = var_4904, groups = var_4700, pad = input_419_pad_0, pad_type = input_419_pad_type_0, strides = var_4902, weight = layers_20_fc2_loraA_weight_to_fp16, x = input_417_cast_fp16)[name = tensor("input_419_cast_fp16")]; + tensor var_4908 = const()[name = tensor("op_4908"), val = tensor([1, 1])]; + tensor var_4910 = const()[name = tensor("op_4910"), 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(237252288)))]; + tensor lora_out_503_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4910, groups = var_4700, pad = lora_out_501_pad_0, pad_type = lora_out_501_pad_type_0, strides = var_4908, weight = lora_out_503_weight_0_to_fp16, x = input_419_cast_fp16)[name = tensor("lora_out_503_cast_fp16")]; + tensor hidden_states_45_cast_fp16 = add(x = pretrained_out_251_cast_fp16, y = lora_out_503_cast_fp16)[name = tensor("hidden_states_45_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_4924 = const()[name = tensor("op_4924"), val = tensor(3)]; + tensor var_4926 = const()[name = tensor("op_4926"), val = tensor(1)]; + tensor var_4927 = const()[name = tensor("op_4927"), val = tensor(true)]; + tensor var_4937 = const()[name = tensor("op_4937"), val = tensor([1])]; + tensor channels_mean_85_cast_fp16 = reduce_mean(axes = var_4937, keep_dims = var_4927, 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_4941 = const()[name = tensor("op_4941"), val = tensor([1])]; + tensor var_4942_cast_fp16 = reduce_mean(axes = var_4941, keep_dims = var_4927, x = zero_mean_sq_85_cast_fp16)[name = tensor("op_4942_cast_fp16")]; + tensor var_4943_to_fp16 = const()[name = tensor("op_4943_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_4944_cast_fp16 = add(x = var_4942_cast_fp16, y = var_4943_to_fp16)[name = tensor("op_4944_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_4944_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_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(237293312)))]; + 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(237295936)))]; + 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_85_cast_fp16)[name = tensor("obj_85_cast_fp16")]; + tensor var_4962 = const()[name = tensor("op_4962"), val = tensor([1, 1])]; + tensor var_4964 = const()[name = tensor("op_4964"), 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_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(237298560))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238117824))), 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(238117952)))]; + tensor pretrained_out_253_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_4964, groups = var_4926, pad = pretrained_out_253_pad_0, pad_type = pretrained_out_253_pad_type_0, strides = var_4962, weight = layers_21_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor("pretrained_out_253_cast_fp16")]; + tensor var_4968 = const()[name = tensor("op_4968"), val = tensor([1, 1])]; + tensor var_4970 = const()[name = tensor("op_4970"), 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_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(238120576)))]; + tensor input_421_cast_fp16 = conv(dilations = var_4970, groups = var_4926, pad = input_421_pad_0, pad_type = input_421_pad_type_0, strides = var_4968, weight = layers_21_self_attn_q_proj_loraA_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor var_4974 = const()[name = tensor("op_4974"), val = tensor([1, 1])]; + tensor var_4976 = const()[name = tensor("op_4976"), 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(238161600)))]; + tensor lora_out_507_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_4976, groups = var_4926, pad = lora_out_505_pad_0, pad_type = lora_out_505_pad_type_0, strides = var_4974, weight = lora_out_507_weight_0_to_fp16, x = input_421_cast_fp16)[name = tensor("lora_out_507_cast_fp16")]; + tensor query_43_cast_fp16 = add(x = pretrained_out_253_cast_fp16, y = lora_out_507_cast_fp16)[name = tensor("query_43_cast_fp16")]; + tensor var_4986 = const()[name = tensor("op_4986"), val = tensor([1, 1])]; + tensor var_4988 = const()[name = tensor("op_4988"), 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_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(238202624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239021888))), name = tensor("layers_21_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_255_cast_fp16 = conv(dilations = var_4988, groups = var_4926, pad = pretrained_out_255_pad_0, pad_type = pretrained_out_255_pad_type_0, strides = var_4986, weight = layers_21_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor("pretrained_out_255_cast_fp16")]; + tensor var_4992 = const()[name = tensor("op_4992"), val = tensor([1, 1])]; + tensor var_4994 = const()[name = tensor("op_4994"), 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_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(239022016)))]; + tensor input_423_cast_fp16 = conv(dilations = var_4994, groups = var_4926, pad = input_423_pad_0, pad_type = input_423_pad_type_0, strides = var_4992, weight = layers_21_self_attn_k_proj_loraA_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor var_4998 = const()[name = tensor("op_4998"), val = tensor([1, 1])]; + tensor var_5000 = const()[name = tensor("op_5000"), 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(239063040)))]; + tensor lora_out_511_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5000, groups = var_4926, pad = lora_out_509_pad_0, pad_type = lora_out_509_pad_type_0, strides = var_4998, weight = lora_out_511_weight_0_to_fp16, x = input_423_cast_fp16)[name = tensor("lora_out_511_cast_fp16")]; + tensor key_43_cast_fp16 = add(x = pretrained_out_255_cast_fp16, y = lora_out_511_cast_fp16)[name = tensor("key_43_cast_fp16")]; + tensor var_5011 = const()[name = tensor("op_5011"), val = tensor([1, 1])]; + tensor var_5013 = const()[name = tensor("op_5013"), 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_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(239104064))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239923328))), 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(239923456)))]; + tensor pretrained_out_257_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5013, groups = var_4926, pad = pretrained_out_257_pad_0, pad_type = pretrained_out_257_pad_type_0, strides = var_5011, weight = layers_21_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_85_cast_fp16)[name = tensor("pretrained_out_257_cast_fp16")]; + tensor var_5017 = const()[name = tensor("op_5017"), val = tensor([1, 1])]; + tensor var_5019 = const()[name = tensor("op_5019"), 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_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(239926080)))]; + tensor input_425_cast_fp16 = conv(dilations = var_5019, groups = var_4926, pad = input_425_pad_0, pad_type = input_425_pad_type_0, strides = var_5017, weight = layers_21_self_attn_v_proj_loraA_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("input_425_cast_fp16")]; + tensor var_5023 = const()[name = tensor("op_5023"), val = tensor([1, 1])]; + tensor var_5025 = const()[name = tensor("op_5025"), 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(239967104)))]; + tensor lora_out_515_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5025, groups = var_4926, pad = lora_out_513_pad_0, pad_type = lora_out_513_pad_type_0, strides = var_5023, weight = lora_out_515_weight_0_to_fp16, x = input_425_cast_fp16)[name = tensor("lora_out_515_cast_fp16")]; + tensor value_43_cast_fp16 = add(x = pretrained_out_257_cast_fp16, y = lora_out_515_cast_fp16)[name = tensor("value_43_cast_fp16")]; + tensor var_5032 = const()[name = tensor("op_5032"), val = tensor([1, 20, 64, -1])]; + tensor var_5033_cast_fp16 = reshape(shape = var_5032, x = query_43_cast_fp16)[name = tensor("op_5033_cast_fp16")]; + tensor var_5034_to_fp16 = const()[name = tensor("op_5034_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5035_cast_fp16 = mul(x = var_5033_cast_fp16, y = var_5034_to_fp16)[name = tensor("op_5035_cast_fp16")]; + tensor var_5036 = const()[name = tensor("op_5036"), val = tensor([1, 20, 64, -1])]; + tensor var_5037_cast_fp16 = reshape(shape = var_5036, x = key_43_cast_fp16)[name = tensor("op_5037_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_5035_cast_fp16, y = var_5037_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; + tensor var_5040_cast_fp16 = softmax(axis = var_4924, x = mh_w_43_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor var_5041 = const()[name = tensor("op_5041"), val = tensor([1, 20, 64, -1])]; + tensor var_5042_cast_fp16 = reshape(shape = var_5041, x = value_43_cast_fp16)[name = tensor("op_5042_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_5042_cast_fp16, y = var_5040_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_5045 = const()[name = tensor("op_5045"), val = tensor([1, 1280, 1, -1])]; + tensor input_427_cast_fp16 = reshape(shape = var_5045, x = attn_43_cast_fp16)[name = tensor("input_427_cast_fp16")]; + tensor var_5052 = const()[name = tensor("op_5052"), val = tensor([1, 1])]; + tensor var_5054 = const()[name = tensor("op_5054"), 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_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(240008128))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240827392))), 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(240827520)))]; + tensor pretrained_out_259_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5054, groups = var_4926, pad = pretrained_out_259_pad_0, pad_type = pretrained_out_259_pad_type_0, strides = var_5052, weight = layers_21_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_427_cast_fp16)[name = tensor("pretrained_out_259_cast_fp16")]; + tensor var_5058 = const()[name = tensor("op_5058"), val = tensor([1, 1])]; + tensor var_5060 = const()[name = tensor("op_5060"), 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_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(240830144)))]; + tensor input_429_cast_fp16 = conv(dilations = var_5060, groups = var_4926, pad = input_429_pad_0, pad_type = input_429_pad_type_0, strides = var_5058, weight = layers_21_self_attn_o_proj_loraA_weight_to_fp16, x = input_427_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor var_5064 = const()[name = tensor("op_5064"), val = tensor([1, 1])]; + tensor var_5066 = const()[name = tensor("op_5066"), 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(240871168)))]; + tensor lora_out_519_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5066, groups = var_4926, pad = lora_out_517_pad_0, pad_type = lora_out_517_pad_type_0, strides = var_5064, weight = lora_out_519_weight_0_to_fp16, x = input_429_cast_fp16)[name = tensor("lora_out_519_cast_fp16")]; + tensor obj_87_cast_fp16 = add(x = pretrained_out_259_cast_fp16, y = lora_out_519_cast_fp16)[name = tensor("obj_87_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_87_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor var_5075 = const()[name = tensor("op_5075"), val = tensor([1])]; + tensor channels_mean_87_cast_fp16 = reduce_mean(axes = var_5075, keep_dims = var_4927, 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_5079 = const()[name = tensor("op_5079"), val = tensor([1])]; + tensor var_5080_cast_fp16 = reduce_mean(axes = var_5079, keep_dims = var_4927, x = zero_mean_sq_87_cast_fp16)[name = tensor("op_5080_cast_fp16")]; + tensor var_5081_to_fp16 = const()[name = tensor("op_5081_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5082_cast_fp16 = add(x = var_5080_cast_fp16, y = var_5081_to_fp16)[name = tensor("op_5082_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_5082_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 input_431_gamma_0_to_fp16 = const()[name = tensor("input_431_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240912192)))]; + tensor input_431_beta_0_to_fp16 = const()[name = tensor("input_431_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240914816)))]; + tensor input_431_epsilon_0_to_fp16 = const()[name = tensor("input_431_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_431_cast_fp16 = batch_norm(beta = input_431_beta_0_to_fp16, epsilon = input_431_epsilon_0_to_fp16, gamma = input_431_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("input_431_cast_fp16")]; + tensor var_5096 = const()[name = tensor("op_5096"), val = tensor([1, 1])]; + tensor var_5098 = const()[name = tensor("op_5098"), 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_21_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240917440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244194304))), 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(244194432)))]; + tensor pretrained_out_261_cast_fp16 = conv(bias = layers_21_fc1_pretrained_bias_to_fp16, dilations = var_5098, groups = var_4926, pad = pretrained_out_261_pad_0, pad_type = pretrained_out_261_pad_type_0, strides = var_5096, weight = layers_21_fc1_pretrained_weight_to_fp16_palettized, x = input_431_cast_fp16)[name = tensor("pretrained_out_261_cast_fp16")]; + tensor var_5102 = const()[name = tensor("op_5102"), val = tensor([1, 1])]; + tensor var_5104 = const()[name = tensor("op_5104"), 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_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(244204736)))]; + tensor input_433_cast_fp16 = conv(dilations = var_5104, groups = var_4926, pad = input_433_pad_0, pad_type = input_433_pad_type_0, strides = var_5102, weight = layers_21_fc1_loraA_weight_to_fp16, x = input_431_cast_fp16)[name = tensor("input_433_cast_fp16")]; + tensor var_5108 = const()[name = tensor("op_5108"), val = tensor([1, 1])]; + tensor var_5110 = const()[name = tensor("op_5110"), 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(244245760)))]; + tensor lora_out_523_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_5110, groups = var_4926, pad = lora_out_521_pad_0, pad_type = lora_out_521_pad_type_0, strides = var_5108, weight = lora_out_523_weight_0_to_fp16, x = input_433_cast_fp16)[name = tensor("lora_out_523_cast_fp16")]; + tensor input_435_cast_fp16 = add(x = pretrained_out_261_cast_fp16, y = lora_out_523_cast_fp16)[name = tensor("input_435_cast_fp16")]; + tensor input_437_mode_0 = const()[name = tensor("input_437_mode_0"), val = tensor("EXACT")]; + tensor input_437_cast_fp16 = gelu(mode = input_437_mode_0, x = input_435_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor var_5122 = const()[name = tensor("op_5122"), val = tensor([1, 1])]; + tensor var_5124 = const()[name = tensor("op_5124"), 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_21_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244409664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247686528))), 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(247686656)))]; + tensor pretrained_out_263_cast_fp16 = conv(bias = layers_21_fc2_pretrained_bias_to_fp16, dilations = var_5124, groups = var_4926, pad = pretrained_out_263_pad_0, pad_type = pretrained_out_263_pad_type_0, strides = var_5122, weight = layers_21_fc2_pretrained_weight_to_fp16_palettized, x = input_437_cast_fp16)[name = tensor("pretrained_out_263_cast_fp16")]; + tensor var_5128 = const()[name = tensor("op_5128"), val = tensor([1, 1])]; + tensor var_5130 = const()[name = tensor("op_5130"), 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_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(247689280)))]; + tensor input_439_cast_fp16 = conv(dilations = var_5130, groups = var_4926, pad = input_439_pad_0, pad_type = input_439_pad_type_0, strides = var_5128, weight = layers_21_fc2_loraA_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("input_439_cast_fp16")]; + tensor var_5134 = const()[name = tensor("op_5134"), val = tensor([1, 1])]; + tensor var_5136 = const()[name = tensor("op_5136"), 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(247853184)))]; + tensor lora_out_527_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5136, groups = var_4926, pad = lora_out_525_pad_0, pad_type = lora_out_525_pad_type_0, strides = var_5134, weight = lora_out_527_weight_0_to_fp16, x = input_439_cast_fp16)[name = tensor("lora_out_527_cast_fp16")]; + tensor hidden_states_47_cast_fp16 = add(x = pretrained_out_263_cast_fp16, y = lora_out_527_cast_fp16)[name = tensor("hidden_states_47_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor var_5150 = const()[name = tensor("op_5150"), val = tensor(3)]; + tensor var_5152 = const()[name = tensor("op_5152"), val = tensor(1)]; + tensor var_5153 = const()[name = tensor("op_5153"), val = tensor(true)]; + tensor var_5163 = const()[name = tensor("op_5163"), val = tensor([1])]; + tensor channels_mean_89_cast_fp16 = reduce_mean(axes = var_5163, keep_dims = var_5153, 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_5167 = const()[name = tensor("op_5167"), val = tensor([1])]; + tensor var_5168_cast_fp16 = reduce_mean(axes = var_5167, keep_dims = var_5153, x = zero_mean_sq_89_cast_fp16)[name = tensor("op_5168_cast_fp16")]; + tensor var_5169_to_fp16 = const()[name = tensor("op_5169_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5170_cast_fp16 = add(x = var_5168_cast_fp16, y = var_5169_to_fp16)[name = tensor("op_5170_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_5170_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 obj_89_gamma_0_to_fp16 = const()[name = tensor("obj_89_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247894208)))]; + tensor obj_89_beta_0_to_fp16 = const()[name = tensor("obj_89_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247896832)))]; + tensor obj_89_epsilon_0_to_fp16 = const()[name = tensor("obj_89_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_89_cast_fp16 = batch_norm(beta = obj_89_beta_0_to_fp16, epsilon = obj_89_epsilon_0_to_fp16, gamma = obj_89_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("obj_89_cast_fp16")]; + tensor var_5188 = const()[name = tensor("op_5188"), val = tensor([1, 1])]; + tensor var_5190 = const()[name = tensor("op_5190"), 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_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(247899456))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(248718720))), 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(248718848)))]; + tensor pretrained_out_265_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5190, groups = var_5152, pad = pretrained_out_265_pad_0, pad_type = pretrained_out_265_pad_type_0, strides = var_5188, weight = layers_22_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_89_cast_fp16)[name = tensor("pretrained_out_265_cast_fp16")]; + tensor var_5194 = const()[name = tensor("op_5194"), val = tensor([1, 1])]; + tensor var_5196 = const()[name = tensor("op_5196"), val = tensor([1, 1])]; + tensor input_441_pad_type_0 = const()[name = tensor("input_441_pad_type_0"), val = tensor("custom")]; + tensor input_441_pad_0 = const()[name = tensor("input_441_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(248721472)))]; + tensor input_441_cast_fp16 = conv(dilations = var_5196, groups = var_5152, pad = input_441_pad_0, pad_type = input_441_pad_type_0, strides = var_5194, weight = layers_22_self_attn_q_proj_loraA_weight_to_fp16, x = obj_89_cast_fp16)[name = tensor("input_441_cast_fp16")]; + tensor var_5200 = const()[name = tensor("op_5200"), val = tensor([1, 1])]; + tensor var_5202 = const()[name = tensor("op_5202"), 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(248762496)))]; + tensor lora_out_531_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5202, groups = var_5152, pad = lora_out_529_pad_0, pad_type = lora_out_529_pad_type_0, strides = var_5200, weight = lora_out_531_weight_0_to_fp16, x = input_441_cast_fp16)[name = tensor("lora_out_531_cast_fp16")]; + tensor query_45_cast_fp16 = add(x = pretrained_out_265_cast_fp16, y = lora_out_531_cast_fp16)[name = tensor("query_45_cast_fp16")]; + tensor var_5212 = const()[name = tensor("op_5212"), val = tensor([1, 1])]; + tensor var_5214 = const()[name = tensor("op_5214"), 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_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(248803520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249622784))), name = tensor("layers_22_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_267_cast_fp16 = conv(dilations = var_5214, groups = var_5152, pad = pretrained_out_267_pad_0, pad_type = pretrained_out_267_pad_type_0, strides = var_5212, weight = layers_22_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_89_cast_fp16)[name = tensor("pretrained_out_267_cast_fp16")]; + tensor var_5218 = const()[name = tensor("op_5218"), val = tensor([1, 1])]; + tensor var_5220 = const()[name = tensor("op_5220"), 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_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(249622912)))]; + tensor input_443_cast_fp16 = conv(dilations = var_5220, groups = var_5152, pad = input_443_pad_0, pad_type = input_443_pad_type_0, strides = var_5218, weight = layers_22_self_attn_k_proj_loraA_weight_to_fp16, x = obj_89_cast_fp16)[name = tensor("input_443_cast_fp16")]; + tensor var_5224 = const()[name = tensor("op_5224"), val = tensor([1, 1])]; + tensor var_5226 = const()[name = tensor("op_5226"), 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(249663936)))]; + tensor lora_out_535_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5226, groups = var_5152, pad = lora_out_533_pad_0, pad_type = lora_out_533_pad_type_0, strides = var_5224, weight = lora_out_535_weight_0_to_fp16, x = input_443_cast_fp16)[name = tensor("lora_out_535_cast_fp16")]; + tensor key_45_cast_fp16 = add(x = pretrained_out_267_cast_fp16, y = lora_out_535_cast_fp16)[name = tensor("key_45_cast_fp16")]; + tensor var_5237 = const()[name = tensor("op_5237"), val = tensor([1, 1])]; + tensor var_5239 = const()[name = tensor("op_5239"), 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_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(249704960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(250524224))), 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(250524352)))]; + tensor pretrained_out_269_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5239, groups = var_5152, pad = pretrained_out_269_pad_0, pad_type = pretrained_out_269_pad_type_0, strides = var_5237, weight = layers_22_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_89_cast_fp16)[name = tensor("pretrained_out_269_cast_fp16")]; + tensor var_5243 = const()[name = tensor("op_5243"), val = tensor([1, 1])]; + tensor var_5245 = const()[name = tensor("op_5245"), val = tensor([1, 1])]; + tensor input_445_pad_type_0 = const()[name = tensor("input_445_pad_type_0"), val = tensor("custom")]; + tensor input_445_pad_0 = const()[name = tensor("input_445_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(250526976)))]; + tensor input_445_cast_fp16 = conv(dilations = var_5245, groups = var_5152, pad = input_445_pad_0, pad_type = input_445_pad_type_0, strides = var_5243, weight = layers_22_self_attn_v_proj_loraA_weight_to_fp16, x = obj_89_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor var_5249 = const()[name = tensor("op_5249"), val = tensor([1, 1])]; + tensor var_5251 = const()[name = tensor("op_5251"), 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(250568000)))]; + tensor lora_out_539_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5251, groups = var_5152, pad = lora_out_537_pad_0, pad_type = lora_out_537_pad_type_0, strides = var_5249, weight = lora_out_539_weight_0_to_fp16, x = input_445_cast_fp16)[name = tensor("lora_out_539_cast_fp16")]; + tensor value_45_cast_fp16 = add(x = pretrained_out_269_cast_fp16, y = lora_out_539_cast_fp16)[name = tensor("value_45_cast_fp16")]; + tensor var_5258 = const()[name = tensor("op_5258"), val = tensor([1, 20, 64, -1])]; + tensor var_5259_cast_fp16 = reshape(shape = var_5258, x = query_45_cast_fp16)[name = tensor("op_5259_cast_fp16")]; + tensor var_5260_to_fp16 = const()[name = tensor("op_5260_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5261_cast_fp16 = mul(x = var_5259_cast_fp16, y = var_5260_to_fp16)[name = tensor("op_5261_cast_fp16")]; + tensor var_5262 = const()[name = tensor("op_5262"), val = tensor([1, 20, 64, -1])]; + tensor var_5263_cast_fp16 = reshape(shape = var_5262, x = key_45_cast_fp16)[name = tensor("op_5263_cast_fp16")]; + tensor mh_w_45_transpose_x_0 = const()[name = tensor("mh_w_45_transpose_x_0"), val = tensor(true)]; + tensor mh_w_45_transpose_y_0 = const()[name = tensor("mh_w_45_transpose_y_0"), val = tensor(false)]; + tensor mh_w_45_cast_fp16 = matmul(transpose_x = mh_w_45_transpose_x_0, transpose_y = mh_w_45_transpose_y_0, x = var_5261_cast_fp16, y = var_5263_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; + tensor var_5266_cast_fp16 = softmax(axis = var_5150, x = mh_w_45_cast_fp16)[name = tensor("op_5266_cast_fp16")]; + tensor var_5267 = const()[name = tensor("op_5267"), val = tensor([1, 20, 64, -1])]; + tensor var_5268_cast_fp16 = reshape(shape = var_5267, x = value_45_cast_fp16)[name = tensor("op_5268_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_5268_cast_fp16, y = var_5266_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_5271 = const()[name = tensor("op_5271"), val = tensor([1, 1280, 1, -1])]; + tensor input_447_cast_fp16 = reshape(shape = var_5271, x = attn_45_cast_fp16)[name = tensor("input_447_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 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_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(250609024))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251428288))), 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(251428416)))]; + tensor pretrained_out_271_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5280, groups = var_5152, pad = pretrained_out_271_pad_0, pad_type = pretrained_out_271_pad_type_0, strides = var_5278, weight = layers_22_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = tensor("pretrained_out_271_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 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_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(251431040)))]; + tensor input_449_cast_fp16 = conv(dilations = var_5286, groups = var_5152, pad = input_449_pad_0, pad_type = input_449_pad_type_0, strides = var_5284, weight = layers_22_self_attn_o_proj_loraA_weight_to_fp16, x = input_447_cast_fp16)[name = tensor("input_449_cast_fp16")]; + tensor var_5290 = const()[name = tensor("op_5290"), val = tensor([1, 1])]; + tensor var_5292 = const()[name = tensor("op_5292"), 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(251472064)))]; + tensor lora_out_543_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5292, groups = var_5152, pad = lora_out_541_pad_0, pad_type = lora_out_541_pad_type_0, strides = var_5290, weight = lora_out_543_weight_0_to_fp16, x = input_449_cast_fp16)[name = tensor("lora_out_543_cast_fp16")]; + tensor obj_91_cast_fp16 = add(x = pretrained_out_271_cast_fp16, y = lora_out_543_cast_fp16)[name = tensor("obj_91_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = obj_91_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor var_5301 = const()[name = tensor("op_5301"), val = tensor([1])]; + tensor channels_mean_91_cast_fp16 = reduce_mean(axes = var_5301, keep_dims = var_5153, 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_5305 = const()[name = tensor("op_5305"), val = tensor([1])]; + tensor var_5306_cast_fp16 = reduce_mean(axes = var_5305, keep_dims = var_5153, x = zero_mean_sq_91_cast_fp16)[name = tensor("op_5306_cast_fp16")]; + tensor var_5307_to_fp16 = const()[name = tensor("op_5307_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5308_cast_fp16 = add(x = var_5306_cast_fp16, y = var_5307_to_fp16)[name = tensor("op_5308_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_5308_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 input_451_gamma_0_to_fp16 = const()[name = tensor("input_451_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251513088)))]; + tensor input_451_beta_0_to_fp16 = const()[name = tensor("input_451_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251515712)))]; + tensor input_451_epsilon_0_to_fp16 = const()[name = tensor("input_451_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_451_cast_fp16 = batch_norm(beta = input_451_beta_0_to_fp16, epsilon = input_451_epsilon_0_to_fp16, gamma = input_451_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("input_451_cast_fp16")]; + tensor var_5322 = const()[name = tensor("op_5322"), val = tensor([1, 1])]; + tensor var_5324 = const()[name = tensor("op_5324"), 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_22_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251518336))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254795200))), 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(254795328)))]; + tensor pretrained_out_273_cast_fp16 = conv(bias = layers_22_fc1_pretrained_bias_to_fp16, dilations = var_5324, groups = var_5152, pad = pretrained_out_273_pad_0, pad_type = pretrained_out_273_pad_type_0, strides = var_5322, weight = layers_22_fc1_pretrained_weight_to_fp16_palettized, x = input_451_cast_fp16)[name = tensor("pretrained_out_273_cast_fp16")]; + tensor var_5328 = const()[name = tensor("op_5328"), val = tensor([1, 1])]; + tensor var_5330 = const()[name = tensor("op_5330"), 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_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(254805632)))]; + tensor input_453_cast_fp16 = conv(dilations = var_5330, groups = var_5152, pad = input_453_pad_0, pad_type = input_453_pad_type_0, strides = var_5328, weight = layers_22_fc1_loraA_weight_to_fp16, x = input_451_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor var_5334 = const()[name = tensor("op_5334"), val = tensor([1, 1])]; + tensor var_5336 = const()[name = tensor("op_5336"), 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(254846656)))]; + tensor lora_out_547_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_5336, groups = var_5152, pad = lora_out_545_pad_0, pad_type = lora_out_545_pad_type_0, strides = var_5334, weight = lora_out_547_weight_0_to_fp16, x = input_453_cast_fp16)[name = tensor("lora_out_547_cast_fp16")]; + tensor input_455_cast_fp16 = add(x = pretrained_out_273_cast_fp16, y = lora_out_547_cast_fp16)[name = tensor("input_455_cast_fp16")]; + tensor input_457_mode_0 = const()[name = tensor("input_457_mode_0"), val = tensor("EXACT")]; + tensor input_457_cast_fp16 = gelu(mode = input_457_mode_0, x = input_455_cast_fp16)[name = tensor("input_457_cast_fp16")]; + tensor var_5348 = const()[name = tensor("op_5348"), val = tensor([1, 1])]; + tensor var_5350 = const()[name = tensor("op_5350"), 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_22_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255010560))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258287424))), 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(258287552)))]; + tensor pretrained_out_275_cast_fp16 = conv(bias = layers_22_fc2_pretrained_bias_to_fp16, dilations = var_5350, groups = var_5152, pad = pretrained_out_275_pad_0, pad_type = pretrained_out_275_pad_type_0, strides = var_5348, weight = layers_22_fc2_pretrained_weight_to_fp16_palettized, x = input_457_cast_fp16)[name = tensor("pretrained_out_275_cast_fp16")]; + tensor var_5354 = const()[name = tensor("op_5354"), val = tensor([1, 1])]; + tensor var_5356 = const()[name = tensor("op_5356"), 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_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(258290176)))]; + tensor input_459_cast_fp16 = conv(dilations = var_5356, groups = var_5152, pad = input_459_pad_0, pad_type = input_459_pad_type_0, strides = var_5354, weight = layers_22_fc2_loraA_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("input_459_cast_fp16")]; + tensor var_5360 = const()[name = tensor("op_5360"), val = tensor([1, 1])]; + tensor var_5362 = const()[name = tensor("op_5362"), 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(258454080)))]; + tensor lora_out_551_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5362, groups = var_5152, pad = lora_out_549_pad_0, pad_type = lora_out_549_pad_type_0, strides = var_5360, weight = lora_out_551_weight_0_to_fp16, x = input_459_cast_fp16)[name = tensor("lora_out_551_cast_fp16")]; + tensor hidden_states_49_cast_fp16 = add(x = pretrained_out_275_cast_fp16, y = lora_out_551_cast_fp16)[name = tensor("hidden_states_49_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor var_5376 = const()[name = tensor("op_5376"), val = tensor(3)]; + tensor var_5378 = const()[name = tensor("op_5378"), val = tensor(1)]; + tensor var_5379 = const()[name = tensor("op_5379"), val = tensor(true)]; + tensor var_5389 = const()[name = tensor("op_5389"), val = tensor([1])]; + tensor channels_mean_93_cast_fp16 = reduce_mean(axes = var_5389, keep_dims = var_5379, 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_5393 = const()[name = tensor("op_5393"), val = tensor([1])]; + tensor var_5394_cast_fp16 = reduce_mean(axes = var_5393, keep_dims = var_5379, x = zero_mean_sq_93_cast_fp16)[name = tensor("op_5394_cast_fp16")]; + tensor var_5395_to_fp16 = const()[name = tensor("op_5395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5396_cast_fp16 = add(x = var_5394_cast_fp16, y = var_5395_to_fp16)[name = tensor("op_5396_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_5396_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_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(258495104)))]; + 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(258497728)))]; + 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_93_cast_fp16)[name = tensor("obj_93_cast_fp16")]; + tensor var_5414 = const()[name = tensor("op_5414"), val = tensor([1, 1])]; + tensor var_5416 = const()[name = tensor("op_5416"), 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_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(258500352))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259319616))), 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(259319744)))]; + tensor pretrained_out_277_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5416, groups = var_5378, pad = pretrained_out_277_pad_0, pad_type = pretrained_out_277_pad_type_0, strides = var_5414, weight = layers_23_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_93_cast_fp16)[name = tensor("pretrained_out_277_cast_fp16")]; + tensor var_5420 = const()[name = tensor("op_5420"), val = tensor([1, 1])]; + tensor var_5422 = const()[name = tensor("op_5422"), 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_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(259322368)))]; + tensor input_461_cast_fp16 = conv(dilations = var_5422, groups = var_5378, pad = input_461_pad_0, pad_type = input_461_pad_type_0, strides = var_5420, weight = layers_23_self_attn_q_proj_loraA_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor var_5426 = const()[name = tensor("op_5426"), val = tensor([1, 1])]; + tensor var_5428 = const()[name = tensor("op_5428"), 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(259363392)))]; + tensor lora_out_555_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5428, groups = var_5378, pad = lora_out_553_pad_0, pad_type = lora_out_553_pad_type_0, strides = var_5426, weight = lora_out_555_weight_0_to_fp16, x = input_461_cast_fp16)[name = tensor("lora_out_555_cast_fp16")]; + tensor query_47_cast_fp16 = add(x = pretrained_out_277_cast_fp16, y = lora_out_555_cast_fp16)[name = tensor("query_47_cast_fp16")]; + tensor var_5438 = const()[name = tensor("op_5438"), val = tensor([1, 1])]; + tensor var_5440 = const()[name = tensor("op_5440"), 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_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(259404416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260223680))), name = tensor("layers_23_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_279_cast_fp16 = conv(dilations = var_5440, groups = var_5378, pad = pretrained_out_279_pad_0, pad_type = pretrained_out_279_pad_type_0, strides = var_5438, weight = layers_23_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_93_cast_fp16)[name = tensor("pretrained_out_279_cast_fp16")]; + tensor var_5444 = const()[name = tensor("op_5444"), val = tensor([1, 1])]; + tensor var_5446 = const()[name = tensor("op_5446"), 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_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(260223808)))]; + tensor input_463_cast_fp16 = conv(dilations = var_5446, groups = var_5378, pad = input_463_pad_0, pad_type = input_463_pad_type_0, strides = var_5444, weight = layers_23_self_attn_k_proj_loraA_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("input_463_cast_fp16")]; + tensor var_5450 = const()[name = tensor("op_5450"), val = tensor([1, 1])]; + tensor var_5452 = const()[name = tensor("op_5452"), 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(260264832)))]; + tensor lora_out_559_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5452, groups = var_5378, pad = lora_out_557_pad_0, pad_type = lora_out_557_pad_type_0, strides = var_5450, weight = lora_out_559_weight_0_to_fp16, x = input_463_cast_fp16)[name = tensor("lora_out_559_cast_fp16")]; + tensor key_47_cast_fp16 = add(x = pretrained_out_279_cast_fp16, y = lora_out_559_cast_fp16)[name = tensor("key_47_cast_fp16")]; + tensor var_5463 = const()[name = tensor("op_5463"), val = tensor([1, 1])]; + tensor var_5465 = const()[name = tensor("op_5465"), 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_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(260305856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261125120))), 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(261125248)))]; + tensor pretrained_out_281_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5465, groups = var_5378, pad = pretrained_out_281_pad_0, pad_type = pretrained_out_281_pad_type_0, strides = var_5463, weight = layers_23_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_93_cast_fp16)[name = tensor("pretrained_out_281_cast_fp16")]; + tensor var_5469 = const()[name = tensor("op_5469"), val = tensor([1, 1])]; + tensor var_5471 = const()[name = tensor("op_5471"), 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_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(261127872)))]; + tensor input_465_cast_fp16 = conv(dilations = var_5471, groups = var_5378, pad = input_465_pad_0, pad_type = input_465_pad_type_0, strides = var_5469, weight = layers_23_self_attn_v_proj_loraA_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("input_465_cast_fp16")]; + tensor var_5475 = const()[name = tensor("op_5475"), val = tensor([1, 1])]; + tensor var_5477 = const()[name = tensor("op_5477"), 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(261168896)))]; + tensor lora_out_563_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5477, groups = var_5378, pad = lora_out_561_pad_0, pad_type = lora_out_561_pad_type_0, strides = var_5475, weight = lora_out_563_weight_0_to_fp16, x = input_465_cast_fp16)[name = tensor("lora_out_563_cast_fp16")]; + tensor value_47_cast_fp16 = add(x = pretrained_out_281_cast_fp16, y = lora_out_563_cast_fp16)[name = tensor("value_47_cast_fp16")]; + tensor var_5484 = const()[name = tensor("op_5484"), val = tensor([1, 20, 64, -1])]; + tensor var_5485_cast_fp16 = reshape(shape = var_5484, x = query_47_cast_fp16)[name = tensor("op_5485_cast_fp16")]; + tensor var_5486_to_fp16 = const()[name = tensor("op_5486_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5487_cast_fp16 = mul(x = var_5485_cast_fp16, y = var_5486_to_fp16)[name = tensor("op_5487_cast_fp16")]; + tensor var_5488 = const()[name = tensor("op_5488"), val = tensor([1, 20, 64, -1])]; + tensor var_5489_cast_fp16 = reshape(shape = var_5488, x = key_47_cast_fp16)[name = tensor("op_5489_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_5487_cast_fp16, y = var_5489_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; + tensor var_5492_cast_fp16 = softmax(axis = var_5376, x = mh_w_47_cast_fp16)[name = tensor("op_5492_cast_fp16")]; + tensor var_5493 = const()[name = tensor("op_5493"), val = tensor([1, 20, 64, -1])]; + tensor var_5494_cast_fp16 = reshape(shape = var_5493, x = value_47_cast_fp16)[name = tensor("op_5494_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_5494_cast_fp16, y = var_5492_cast_fp16)[name = tensor("attn_47_cast_fp16")]; + tensor var_5497 = const()[name = tensor("op_5497"), val = tensor([1, 1280, 1, -1])]; + tensor input_467_cast_fp16 = reshape(shape = var_5497, x = attn_47_cast_fp16)[name = tensor("input_467_cast_fp16")]; + tensor var_5504 = const()[name = tensor("op_5504"), val = tensor([1, 1])]; + tensor var_5506 = const()[name = tensor("op_5506"), 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_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(261209920))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262029184))), 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(262029312)))]; + tensor pretrained_out_283_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5506, groups = var_5378, pad = pretrained_out_283_pad_0, pad_type = pretrained_out_283_pad_type_0, strides = var_5504, weight = layers_23_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_467_cast_fp16)[name = tensor("pretrained_out_283_cast_fp16")]; + tensor var_5510 = const()[name = tensor("op_5510"), val = tensor([1, 1])]; + tensor var_5512 = const()[name = tensor("op_5512"), 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_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(262031936)))]; + tensor input_469_cast_fp16 = conv(dilations = var_5512, groups = var_5378, pad = input_469_pad_0, pad_type = input_469_pad_type_0, strides = var_5510, weight = layers_23_self_attn_o_proj_loraA_weight_to_fp16, x = input_467_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor var_5516 = const()[name = tensor("op_5516"), val = tensor([1, 1])]; + tensor var_5518 = const()[name = tensor("op_5518"), 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(262072960)))]; + tensor lora_out_567_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5518, groups = var_5378, pad = lora_out_565_pad_0, pad_type = lora_out_565_pad_type_0, strides = var_5516, weight = lora_out_567_weight_0_to_fp16, x = input_469_cast_fp16)[name = tensor("lora_out_567_cast_fp16")]; + tensor obj_95_cast_fp16 = add(x = pretrained_out_283_cast_fp16, y = lora_out_567_cast_fp16)[name = tensor("obj_95_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_95_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor var_5527 = const()[name = tensor("op_5527"), val = tensor([1])]; + tensor channels_mean_95_cast_fp16 = reduce_mean(axes = var_5527, keep_dims = var_5379, 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_5531 = const()[name = tensor("op_5531"), val = tensor([1])]; + tensor var_5532_cast_fp16 = reduce_mean(axes = var_5531, keep_dims = var_5379, x = zero_mean_sq_95_cast_fp16)[name = tensor("op_5532_cast_fp16")]; + tensor var_5533_to_fp16 = const()[name = tensor("op_5533_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5534_cast_fp16 = add(x = var_5532_cast_fp16, y = var_5533_to_fp16)[name = tensor("op_5534_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_5534_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(262113984)))]; + 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(262116608)))]; + 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_5548 = const()[name = tensor("op_5548"), val = tensor([1, 1])]; + tensor var_5550 = const()[name = tensor("op_5550"), 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_23_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262119232))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265396096))), 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(265396224)))]; + tensor pretrained_out_285_cast_fp16 = conv(bias = layers_23_fc1_pretrained_bias_to_fp16, dilations = var_5550, groups = var_5378, pad = pretrained_out_285_pad_0, pad_type = pretrained_out_285_pad_type_0, strides = var_5548, weight = layers_23_fc1_pretrained_weight_to_fp16_palettized, x = input_471_cast_fp16)[name = tensor("pretrained_out_285_cast_fp16")]; + tensor var_5554 = const()[name = tensor("op_5554"), val = tensor([1, 1])]; + tensor var_5556 = const()[name = tensor("op_5556"), 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_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(265406528)))]; + tensor input_473_cast_fp16 = conv(dilations = var_5556, groups = var_5378, pad = input_473_pad_0, pad_type = input_473_pad_type_0, strides = var_5554, weight = layers_23_fc1_loraA_weight_to_fp16, x = input_471_cast_fp16)[name = tensor("input_473_cast_fp16")]; + tensor var_5560 = const()[name = tensor("op_5560"), val = tensor([1, 1])]; + tensor var_5562 = const()[name = tensor("op_5562"), 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(265447552)))]; + tensor lora_out_571_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_5562, groups = var_5378, pad = lora_out_569_pad_0, pad_type = lora_out_569_pad_type_0, strides = var_5560, weight = lora_out_571_weight_0_to_fp16, x = input_473_cast_fp16)[name = tensor("lora_out_571_cast_fp16")]; + tensor input_475_cast_fp16 = add(x = pretrained_out_285_cast_fp16, y = lora_out_571_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_5574 = const()[name = tensor("op_5574"), val = tensor([1, 1])]; + tensor var_5576 = const()[name = tensor("op_5576"), 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_23_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265611456))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268888320))), 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(268888448)))]; + tensor pretrained_out_287_cast_fp16 = conv(bias = layers_23_fc2_pretrained_bias_to_fp16, dilations = var_5576, groups = var_5378, pad = pretrained_out_287_pad_0, pad_type = pretrained_out_287_pad_type_0, strides = var_5574, weight = layers_23_fc2_pretrained_weight_to_fp16_palettized, x = input_477_cast_fp16)[name = tensor("pretrained_out_287_cast_fp16")]; + tensor var_5580 = const()[name = tensor("op_5580"), val = tensor([1, 1])]; + tensor var_5582 = const()[name = tensor("op_5582"), 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_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(268891072)))]; + tensor input_479_cast_fp16 = conv(dilations = var_5582, groups = var_5378, pad = input_479_pad_0, pad_type = input_479_pad_type_0, strides = var_5580, weight = layers_23_fc2_loraA_weight_to_fp16, x = input_477_cast_fp16)[name = tensor("input_479_cast_fp16")]; + tensor var_5586 = const()[name = tensor("op_5586"), val = tensor([1, 1])]; + tensor var_5588 = const()[name = tensor("op_5588"), 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(269054976)))]; + tensor lora_out_575_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5588, groups = var_5378, pad = lora_out_573_pad_0, pad_type = lora_out_573_pad_type_0, strides = var_5586, weight = lora_out_575_weight_0_to_fp16, x = input_479_cast_fp16)[name = tensor("lora_out_575_cast_fp16")]; + tensor hidden_states_51_cast_fp16 = add(x = pretrained_out_287_cast_fp16, y = lora_out_575_cast_fp16)[name = tensor("hidden_states_51_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_51_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; + tensor var_5602 = const()[name = tensor("op_5602"), val = tensor(3)]; + tensor var_5604 = const()[name = tensor("op_5604"), val = tensor(1)]; + tensor var_5605 = const()[name = tensor("op_5605"), val = tensor(true)]; + tensor var_5615 = const()[name = tensor("op_5615"), val = tensor([1])]; + tensor channels_mean_97_cast_fp16 = reduce_mean(axes = var_5615, keep_dims = var_5605, 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_5619 = const()[name = tensor("op_5619"), val = tensor([1])]; + tensor var_5620_cast_fp16 = reduce_mean(axes = var_5619, keep_dims = var_5605, x = zero_mean_sq_97_cast_fp16)[name = tensor("op_5620_cast_fp16")]; + tensor var_5621_to_fp16 = const()[name = tensor("op_5621_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5622_cast_fp16 = add(x = var_5620_cast_fp16, y = var_5621_to_fp16)[name = tensor("op_5622_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_5622_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_97_gamma_0_to_fp16 = const()[name = tensor("obj_97_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269096000)))]; + tensor obj_97_beta_0_to_fp16 = const()[name = tensor("obj_97_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269098624)))]; + tensor obj_97_epsilon_0_to_fp16 = const()[name = tensor("obj_97_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_97_cast_fp16 = batch_norm(beta = obj_97_beta_0_to_fp16, epsilon = obj_97_epsilon_0_to_fp16, gamma = obj_97_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_97_cast_fp16")]; + tensor var_5640 = const()[name = tensor("op_5640"), val = tensor([1, 1])]; + tensor var_5642 = const()[name = tensor("op_5642"), 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_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(269101248))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269920512))), 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(269920640)))]; + tensor pretrained_out_289_cast_fp16 = conv(bias = layers_24_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5642, groups = var_5604, pad = pretrained_out_289_pad_0, pad_type = pretrained_out_289_pad_type_0, strides = var_5640, weight = layers_24_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = tensor("pretrained_out_289_cast_fp16")]; + tensor var_5646 = const()[name = tensor("op_5646"), val = tensor([1, 1])]; + tensor var_5648 = const()[name = tensor("op_5648"), 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_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(269923264)))]; + tensor input_481_cast_fp16 = conv(dilations = var_5648, groups = var_5604, pad = input_481_pad_0, pad_type = input_481_pad_type_0, strides = var_5646, weight = layers_24_self_attn_q_proj_loraA_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor("input_481_cast_fp16")]; + tensor var_5652 = const()[name = tensor("op_5652"), val = tensor([1, 1])]; + tensor var_5654 = const()[name = tensor("op_5654"), 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(269964288)))]; + tensor lora_out_579_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5654, groups = var_5604, pad = lora_out_577_pad_0, pad_type = lora_out_577_pad_type_0, strides = var_5652, weight = lora_out_579_weight_0_to_fp16, x = input_481_cast_fp16)[name = tensor("lora_out_579_cast_fp16")]; + tensor query_49_cast_fp16 = add(x = pretrained_out_289_cast_fp16, y = lora_out_579_cast_fp16)[name = tensor("query_49_cast_fp16")]; + tensor var_5664 = const()[name = tensor("op_5664"), val = tensor([1, 1])]; + tensor var_5666 = const()[name = tensor("op_5666"), 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_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(270005312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(270824576))), name = tensor("layers_24_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_291_cast_fp16 = conv(dilations = var_5666, groups = var_5604, pad = pretrained_out_291_pad_0, pad_type = pretrained_out_291_pad_type_0, strides = var_5664, weight = layers_24_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = tensor("pretrained_out_291_cast_fp16")]; + tensor var_5670 = const()[name = tensor("op_5670"), val = tensor([1, 1])]; + tensor var_5672 = const()[name = tensor("op_5672"), 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_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(270824704)))]; + tensor input_483_cast_fp16 = conv(dilations = var_5672, groups = var_5604, pad = input_483_pad_0, pad_type = input_483_pad_type_0, strides = var_5670, weight = layers_24_self_attn_k_proj_loraA_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor("input_483_cast_fp16")]; + tensor var_5676 = const()[name = tensor("op_5676"), val = tensor([1, 1])]; + tensor var_5678 = const()[name = tensor("op_5678"), 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(270865728)))]; + tensor lora_out_583_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5678, groups = var_5604, pad = lora_out_581_pad_0, pad_type = lora_out_581_pad_type_0, strides = var_5676, weight = lora_out_583_weight_0_to_fp16, x = input_483_cast_fp16)[name = tensor("lora_out_583_cast_fp16")]; + tensor key_49_cast_fp16 = add(x = pretrained_out_291_cast_fp16, y = lora_out_583_cast_fp16)[name = tensor("key_49_cast_fp16")]; + tensor var_5689 = const()[name = tensor("op_5689"), val = tensor([1, 1])]; + tensor var_5691 = const()[name = tensor("op_5691"), 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_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(270906752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271726016))), 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(271726144)))]; + tensor pretrained_out_293_cast_fp16 = conv(bias = layers_24_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5691, groups = var_5604, pad = pretrained_out_293_pad_0, pad_type = pretrained_out_293_pad_type_0, strides = var_5689, weight = layers_24_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_97_cast_fp16)[name = tensor("pretrained_out_293_cast_fp16")]; + tensor var_5695 = const()[name = tensor("op_5695"), val = tensor([1, 1])]; + tensor var_5697 = const()[name = tensor("op_5697"), 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_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(271728768)))]; + tensor input_485_cast_fp16 = conv(dilations = var_5697, groups = var_5604, pad = input_485_pad_0, pad_type = input_485_pad_type_0, strides = var_5695, weight = layers_24_self_attn_v_proj_loraA_weight_to_fp16, x = obj_97_cast_fp16)[name = tensor("input_485_cast_fp16")]; + tensor var_5701 = const()[name = tensor("op_5701"), val = tensor([1, 1])]; + tensor var_5703 = const()[name = tensor("op_5703"), 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(271769792)))]; + tensor lora_out_587_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5703, groups = var_5604, pad = lora_out_585_pad_0, pad_type = lora_out_585_pad_type_0, strides = var_5701, weight = lora_out_587_weight_0_to_fp16, x = input_485_cast_fp16)[name = tensor("lora_out_587_cast_fp16")]; + tensor value_49_cast_fp16 = add(x = pretrained_out_293_cast_fp16, y = lora_out_587_cast_fp16)[name = tensor("value_49_cast_fp16")]; + tensor var_5710 = const()[name = tensor("op_5710"), val = tensor([1, 20, 64, -1])]; + tensor var_5711_cast_fp16 = reshape(shape = var_5710, x = query_49_cast_fp16)[name = tensor("op_5711_cast_fp16")]; + tensor var_5712_to_fp16 = const()[name = tensor("op_5712_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5713_cast_fp16 = mul(x = var_5711_cast_fp16, y = var_5712_to_fp16)[name = tensor("op_5713_cast_fp16")]; + tensor var_5714 = const()[name = tensor("op_5714"), val = tensor([1, 20, 64, -1])]; + tensor var_5715_cast_fp16 = reshape(shape = var_5714, x = key_49_cast_fp16)[name = tensor("op_5715_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_5713_cast_fp16, y = var_5715_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; + tensor var_5718_cast_fp16 = softmax(axis = var_5602, x = mh_w_49_cast_fp16)[name = tensor("op_5718_cast_fp16")]; + tensor var_5719 = const()[name = tensor("op_5719"), val = tensor([1, 20, 64, -1])]; + tensor var_5720_cast_fp16 = reshape(shape = var_5719, x = value_49_cast_fp16)[name = tensor("op_5720_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_5720_cast_fp16, y = var_5718_cast_fp16)[name = tensor("attn_49_cast_fp16")]; + tensor var_5723 = const()[name = tensor("op_5723"), val = tensor([1, 1280, 1, -1])]; + tensor input_487_cast_fp16 = reshape(shape = var_5723, x = attn_49_cast_fp16)[name = tensor("input_487_cast_fp16")]; + tensor var_5730 = const()[name = tensor("op_5730"), val = tensor([1, 1])]; + tensor var_5732 = const()[name = tensor("op_5732"), 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_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(271810816))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272630080))), 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(272630208)))]; + tensor pretrained_out_295_cast_fp16 = conv(bias = layers_24_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5732, groups = var_5604, pad = pretrained_out_295_pad_0, pad_type = pretrained_out_295_pad_type_0, strides = var_5730, weight = layers_24_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_487_cast_fp16)[name = tensor("pretrained_out_295_cast_fp16")]; + tensor var_5736 = const()[name = tensor("op_5736"), val = tensor([1, 1])]; + tensor var_5738 = const()[name = tensor("op_5738"), 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_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(272632832)))]; + tensor input_489_cast_fp16 = conv(dilations = var_5738, groups = var_5604, pad = input_489_pad_0, pad_type = input_489_pad_type_0, strides = var_5736, weight = layers_24_self_attn_o_proj_loraA_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor var_5742 = const()[name = tensor("op_5742"), val = tensor([1, 1])]; + tensor var_5744 = const()[name = tensor("op_5744"), 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(272673856)))]; + tensor lora_out_591_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5744, groups = var_5604, pad = lora_out_589_pad_0, pad_type = lora_out_589_pad_type_0, strides = var_5742, weight = lora_out_591_weight_0_to_fp16, x = input_489_cast_fp16)[name = tensor("lora_out_591_cast_fp16")]; + tensor obj_99_cast_fp16 = add(x = pretrained_out_295_cast_fp16, y = lora_out_591_cast_fp16)[name = tensor("obj_99_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_99_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; + tensor var_5753 = const()[name = tensor("op_5753"), val = tensor([1])]; + tensor channels_mean_99_cast_fp16 = reduce_mean(axes = var_5753, keep_dims = var_5605, 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_5757 = const()[name = tensor("op_5757"), val = tensor([1])]; + tensor var_5758_cast_fp16 = reduce_mean(axes = var_5757, keep_dims = var_5605, x = zero_mean_sq_99_cast_fp16)[name = tensor("op_5758_cast_fp16")]; + tensor var_5759_to_fp16 = const()[name = tensor("op_5759_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5760_cast_fp16 = add(x = var_5758_cast_fp16, y = var_5759_to_fp16)[name = tensor("op_5760_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_5760_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 input_491_gamma_0_to_fp16 = const()[name = tensor("input_491_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272714880)))]; + tensor input_491_beta_0_to_fp16 = const()[name = tensor("input_491_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272717504)))]; + tensor input_491_epsilon_0_to_fp16 = const()[name = tensor("input_491_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_491_cast_fp16 = batch_norm(beta = input_491_beta_0_to_fp16, epsilon = input_491_epsilon_0_to_fp16, gamma = input_491_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("input_491_cast_fp16")]; + tensor var_5774 = const()[name = tensor("op_5774"), val = tensor([1, 1])]; + tensor var_5776 = const()[name = tensor("op_5776"), 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_24_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(272720128))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275996992))), 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(275997120)))]; + tensor pretrained_out_297_cast_fp16 = conv(bias = layers_24_fc1_pretrained_bias_to_fp16, dilations = var_5776, groups = var_5604, pad = pretrained_out_297_pad_0, pad_type = pretrained_out_297_pad_type_0, strides = var_5774, weight = layers_24_fc1_pretrained_weight_to_fp16_palettized, x = input_491_cast_fp16)[name = tensor("pretrained_out_297_cast_fp16")]; + tensor var_5780 = const()[name = tensor("op_5780"), val = tensor([1, 1])]; + tensor var_5782 = const()[name = tensor("op_5782"), 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_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(276007424)))]; + tensor input_493_cast_fp16 = conv(dilations = var_5782, groups = var_5604, pad = input_493_pad_0, pad_type = input_493_pad_type_0, strides = var_5780, weight = layers_24_fc1_loraA_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("input_493_cast_fp16")]; + tensor var_5786 = const()[name = tensor("op_5786"), val = tensor([1, 1])]; + tensor var_5788 = const()[name = tensor("op_5788"), 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(276048448)))]; + tensor lora_out_595_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_5788, groups = var_5604, pad = lora_out_593_pad_0, pad_type = lora_out_593_pad_type_0, strides = var_5786, weight = lora_out_595_weight_0_to_fp16, x = input_493_cast_fp16)[name = tensor("lora_out_595_cast_fp16")]; + tensor input_495_cast_fp16 = add(x = pretrained_out_297_cast_fp16, y = lora_out_595_cast_fp16)[name = tensor("input_495_cast_fp16")]; + tensor input_497_mode_0 = const()[name = tensor("input_497_mode_0"), val = tensor("EXACT")]; + tensor input_497_cast_fp16 = gelu(mode = input_497_mode_0, x = input_495_cast_fp16)[name = tensor("input_497_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 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_24_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276212352))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279489216))), 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(279489344)))]; + tensor pretrained_out_299_cast_fp16 = conv(bias = layers_24_fc2_pretrained_bias_to_fp16, dilations = var_5802, groups = var_5604, pad = pretrained_out_299_pad_0, pad_type = pretrained_out_299_pad_type_0, strides = var_5800, weight = layers_24_fc2_pretrained_weight_to_fp16_palettized, x = input_497_cast_fp16)[name = tensor("pretrained_out_299_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 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_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(279491968)))]; + tensor input_499_cast_fp16 = conv(dilations = var_5808, groups = var_5604, pad = input_499_pad_0, pad_type = input_499_pad_type_0, strides = var_5806, weight = layers_24_fc2_loraA_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("input_499_cast_fp16")]; + tensor var_5812 = const()[name = tensor("op_5812"), val = tensor([1, 1])]; + tensor var_5814 = const()[name = tensor("op_5814"), 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(279655872)))]; + tensor lora_out_599_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5814, groups = var_5604, pad = lora_out_597_pad_0, pad_type = lora_out_597_pad_type_0, strides = var_5812, weight = lora_out_599_weight_0_to_fp16, x = input_499_cast_fp16)[name = tensor("lora_out_599_cast_fp16")]; + tensor hidden_states_53_cast_fp16 = add(x = pretrained_out_299_cast_fp16, y = lora_out_599_cast_fp16)[name = tensor("hidden_states_53_cast_fp16")]; + tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = hidden_states_53_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor var_5828 = const()[name = tensor("op_5828"), val = tensor(3)]; + tensor var_5830 = const()[name = tensor("op_5830"), val = tensor(1)]; + tensor var_5831 = const()[name = tensor("op_5831"), val = tensor(true)]; + tensor var_5841 = const()[name = tensor("op_5841"), val = tensor([1])]; + tensor channels_mean_101_cast_fp16 = reduce_mean(axes = var_5841, keep_dims = var_5831, 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_5845 = const()[name = tensor("op_5845"), val = tensor([1])]; + tensor var_5846_cast_fp16 = reduce_mean(axes = var_5845, keep_dims = var_5831, x = zero_mean_sq_101_cast_fp16)[name = tensor("op_5846_cast_fp16")]; + tensor var_5847_to_fp16 = const()[name = tensor("op_5847_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5848_cast_fp16 = add(x = var_5846_cast_fp16, y = var_5847_to_fp16)[name = tensor("op_5848_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_5848_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 obj_101_gamma_0_to_fp16 = const()[name = tensor("obj_101_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279696896)))]; + tensor obj_101_beta_0_to_fp16 = const()[name = tensor("obj_101_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279699520)))]; + tensor obj_101_epsilon_0_to_fp16 = const()[name = tensor("obj_101_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_101_cast_fp16 = batch_norm(beta = obj_101_beta_0_to_fp16, epsilon = obj_101_epsilon_0_to_fp16, gamma = obj_101_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("obj_101_cast_fp16")]; + tensor var_5866 = const()[name = tensor("op_5866"), val = tensor([1, 1])]; + tensor var_5868 = const()[name = tensor("op_5868"), 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_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(279702144))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280521408))), 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(280521536)))]; + tensor pretrained_out_301_cast_fp16 = conv(bias = layers_25_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_5868, groups = var_5830, pad = pretrained_out_301_pad_0, pad_type = pretrained_out_301_pad_type_0, strides = var_5866, weight = layers_25_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_101_cast_fp16)[name = tensor("pretrained_out_301_cast_fp16")]; + tensor var_5872 = const()[name = tensor("op_5872"), val = tensor([1, 1])]; + tensor var_5874 = const()[name = tensor("op_5874"), val = tensor([1, 1])]; + tensor input_501_pad_type_0 = const()[name = tensor("input_501_pad_type_0"), val = tensor("custom")]; + tensor input_501_pad_0 = const()[name = tensor("input_501_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(280524160)))]; + tensor input_501_cast_fp16 = conv(dilations = var_5874, groups = var_5830, pad = input_501_pad_0, pad_type = input_501_pad_type_0, strides = var_5872, weight = layers_25_self_attn_q_proj_loraA_weight_to_fp16, x = obj_101_cast_fp16)[name = tensor("input_501_cast_fp16")]; + tensor var_5878 = const()[name = tensor("op_5878"), val = tensor([1, 1])]; + tensor var_5880 = const()[name = tensor("op_5880"), 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(280565184)))]; + tensor lora_out_603_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5880, groups = var_5830, pad = lora_out_601_pad_0, pad_type = lora_out_601_pad_type_0, strides = var_5878, weight = lora_out_603_weight_0_to_fp16, x = input_501_cast_fp16)[name = tensor("lora_out_603_cast_fp16")]; + tensor query_51_cast_fp16 = add(x = pretrained_out_301_cast_fp16, y = lora_out_603_cast_fp16)[name = tensor("query_51_cast_fp16")]; + tensor var_5890 = const()[name = tensor("op_5890"), val = tensor([1, 1])]; + tensor var_5892 = const()[name = tensor("op_5892"), 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_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(280606208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281425472))), name = tensor("layers_25_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_303_cast_fp16 = conv(dilations = var_5892, groups = var_5830, pad = pretrained_out_303_pad_0, pad_type = pretrained_out_303_pad_type_0, strides = var_5890, weight = layers_25_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_101_cast_fp16)[name = tensor("pretrained_out_303_cast_fp16")]; + tensor var_5896 = const()[name = tensor("op_5896"), val = tensor([1, 1])]; + tensor var_5898 = const()[name = tensor("op_5898"), 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_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(281425600)))]; + tensor input_503_cast_fp16 = conv(dilations = var_5898, groups = var_5830, pad = input_503_pad_0, pad_type = input_503_pad_type_0, strides = var_5896, weight = layers_25_self_attn_k_proj_loraA_weight_to_fp16, x = obj_101_cast_fp16)[name = tensor("input_503_cast_fp16")]; + tensor var_5902 = const()[name = tensor("op_5902"), val = tensor([1, 1])]; + tensor var_5904 = const()[name = tensor("op_5904"), 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(281466624)))]; + tensor lora_out_607_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5904, groups = var_5830, pad = lora_out_605_pad_0, pad_type = lora_out_605_pad_type_0, strides = var_5902, weight = lora_out_607_weight_0_to_fp16, x = input_503_cast_fp16)[name = tensor("lora_out_607_cast_fp16")]; + tensor key_51_cast_fp16 = add(x = pretrained_out_303_cast_fp16, y = lora_out_607_cast_fp16)[name = tensor("key_51_cast_fp16")]; + tensor var_5915 = const()[name = tensor("op_5915"), val = tensor([1, 1])]; + tensor var_5917 = const()[name = tensor("op_5917"), 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_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(281507648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282326912))), 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(282327040)))]; + tensor pretrained_out_305_cast_fp16 = conv(bias = layers_25_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_5917, groups = var_5830, pad = pretrained_out_305_pad_0, pad_type = pretrained_out_305_pad_type_0, strides = var_5915, weight = layers_25_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_101_cast_fp16)[name = tensor("pretrained_out_305_cast_fp16")]; + tensor var_5921 = const()[name = tensor("op_5921"), val = tensor([1, 1])]; + tensor var_5923 = const()[name = tensor("op_5923"), val = tensor([1, 1])]; + tensor input_505_pad_type_0 = const()[name = tensor("input_505_pad_type_0"), val = tensor("custom")]; + tensor input_505_pad_0 = const()[name = tensor("input_505_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(282329664)))]; + tensor input_505_cast_fp16 = conv(dilations = var_5923, groups = var_5830, pad = input_505_pad_0, pad_type = input_505_pad_type_0, strides = var_5921, weight = layers_25_self_attn_v_proj_loraA_weight_to_fp16, x = obj_101_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor var_5927 = const()[name = tensor("op_5927"), val = tensor([1, 1])]; + tensor var_5929 = const()[name = tensor("op_5929"), 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(282370688)))]; + tensor lora_out_611_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5929, groups = var_5830, pad = lora_out_609_pad_0, pad_type = lora_out_609_pad_type_0, strides = var_5927, weight = lora_out_611_weight_0_to_fp16, x = input_505_cast_fp16)[name = tensor("lora_out_611_cast_fp16")]; + tensor value_51_cast_fp16 = add(x = pretrained_out_305_cast_fp16, y = lora_out_611_cast_fp16)[name = tensor("value_51_cast_fp16")]; + tensor var_5936 = const()[name = tensor("op_5936"), val = tensor([1, 20, 64, -1])]; + tensor var_5937_cast_fp16 = reshape(shape = var_5936, x = query_51_cast_fp16)[name = tensor("op_5937_cast_fp16")]; + tensor var_5938_to_fp16 = const()[name = tensor("op_5938_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5939_cast_fp16 = mul(x = var_5937_cast_fp16, y = var_5938_to_fp16)[name = tensor("op_5939_cast_fp16")]; + tensor var_5940 = const()[name = tensor("op_5940"), val = tensor([1, 20, 64, -1])]; + tensor var_5941_cast_fp16 = reshape(shape = var_5940, x = key_51_cast_fp16)[name = tensor("op_5941_cast_fp16")]; + tensor mh_w_51_transpose_x_0 = const()[name = tensor("mh_w_51_transpose_x_0"), val = tensor(true)]; + tensor mh_w_51_transpose_y_0 = const()[name = tensor("mh_w_51_transpose_y_0"), val = tensor(false)]; + tensor mh_w_51_cast_fp16 = matmul(transpose_x = mh_w_51_transpose_x_0, transpose_y = mh_w_51_transpose_y_0, x = var_5939_cast_fp16, y = var_5941_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; + tensor var_5944_cast_fp16 = softmax(axis = var_5828, x = mh_w_51_cast_fp16)[name = tensor("op_5944_cast_fp16")]; + tensor var_5945 = const()[name = tensor("op_5945"), val = tensor([1, 20, 64, -1])]; + tensor var_5946_cast_fp16 = reshape(shape = var_5945, x = value_51_cast_fp16)[name = tensor("op_5946_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_5946_cast_fp16, y = var_5944_cast_fp16)[name = tensor("attn_51_cast_fp16")]; + tensor var_5949 = const()[name = tensor("op_5949"), val = tensor([1, 1280, 1, -1])]; + tensor input_507_cast_fp16 = reshape(shape = var_5949, x = attn_51_cast_fp16)[name = tensor("input_507_cast_fp16")]; + tensor var_5956 = const()[name = tensor("op_5956"), val = tensor([1, 1])]; + tensor var_5958 = const()[name = tensor("op_5958"), 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_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(282411712))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283230976))), 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(283231104)))]; + tensor pretrained_out_307_cast_fp16 = conv(bias = layers_25_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_5958, groups = var_5830, pad = pretrained_out_307_pad_0, pad_type = pretrained_out_307_pad_type_0, strides = var_5956, weight = layers_25_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = tensor("pretrained_out_307_cast_fp16")]; + tensor var_5962 = const()[name = tensor("op_5962"), val = tensor([1, 1])]; + tensor var_5964 = const()[name = tensor("op_5964"), 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_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(283233728)))]; + tensor input_509_cast_fp16 = conv(dilations = var_5964, groups = var_5830, pad = input_509_pad_0, pad_type = input_509_pad_type_0, strides = var_5962, weight = layers_25_self_attn_o_proj_loraA_weight_to_fp16, x = input_507_cast_fp16)[name = tensor("input_509_cast_fp16")]; + tensor var_5968 = const()[name = tensor("op_5968"), val = tensor([1, 1])]; + tensor var_5970 = const()[name = tensor("op_5970"), 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(283274752)))]; + tensor lora_out_615_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_5970, groups = var_5830, pad = lora_out_613_pad_0, pad_type = lora_out_613_pad_type_0, strides = var_5968, weight = lora_out_615_weight_0_to_fp16, x = input_509_cast_fp16)[name = tensor("lora_out_615_cast_fp16")]; + tensor obj_103_cast_fp16 = add(x = pretrained_out_307_cast_fp16, y = lora_out_615_cast_fp16)[name = tensor("obj_103_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = obj_103_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; + tensor var_5979 = const()[name = tensor("op_5979"), val = tensor([1])]; + tensor channels_mean_103_cast_fp16 = reduce_mean(axes = var_5979, keep_dims = var_5831, 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_5983 = const()[name = tensor("op_5983"), val = tensor([1])]; + tensor var_5984_cast_fp16 = reduce_mean(axes = var_5983, keep_dims = var_5831, x = zero_mean_sq_103_cast_fp16)[name = tensor("op_5984_cast_fp16")]; + tensor var_5985_to_fp16 = const()[name = tensor("op_5985_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_5986_cast_fp16 = add(x = var_5984_cast_fp16, y = var_5985_to_fp16)[name = tensor("op_5986_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_5986_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 input_511_gamma_0_to_fp16 = const()[name = tensor("input_511_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283315776)))]; + tensor input_511_beta_0_to_fp16 = const()[name = tensor("input_511_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283318400)))]; + tensor input_511_epsilon_0_to_fp16 = const()[name = tensor("input_511_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_511_cast_fp16 = batch_norm(beta = input_511_beta_0_to_fp16, epsilon = input_511_epsilon_0_to_fp16, gamma = input_511_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("input_511_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 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_25_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283321024))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286597888))), 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(286598016)))]; + tensor pretrained_out_309_cast_fp16 = conv(bias = layers_25_fc1_pretrained_bias_to_fp16, dilations = var_6002, groups = var_5830, pad = pretrained_out_309_pad_0, pad_type = pretrained_out_309_pad_type_0, strides = var_6000, weight = layers_25_fc1_pretrained_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = tensor("pretrained_out_309_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 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_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(286608320)))]; + tensor input_513_cast_fp16 = conv(dilations = var_6008, groups = var_5830, pad = input_513_pad_0, pad_type = input_513_pad_type_0, strides = var_6006, weight = layers_25_fc1_loraA_weight_to_fp16, x = input_511_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor var_6012 = const()[name = tensor("op_6012"), val = tensor([1, 1])]; + tensor var_6014 = const()[name = tensor("op_6014"), 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(286649344)))]; + tensor lora_out_619_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_6014, groups = var_5830, pad = lora_out_617_pad_0, pad_type = lora_out_617_pad_type_0, strides = var_6012, weight = lora_out_619_weight_0_to_fp16, x = input_513_cast_fp16)[name = tensor("lora_out_619_cast_fp16")]; + tensor input_515_cast_fp16 = add(x = pretrained_out_309_cast_fp16, y = lora_out_619_cast_fp16)[name = tensor("input_515_cast_fp16")]; + tensor input_517_mode_0 = const()[name = tensor("input_517_mode_0"), val = tensor("EXACT")]; + tensor input_517_cast_fp16 = gelu(mode = input_517_mode_0, x = input_515_cast_fp16)[name = tensor("input_517_cast_fp16")]; + tensor var_6026 = const()[name = tensor("op_6026"), val = tensor([1, 1])]; + tensor var_6028 = const()[name = tensor("op_6028"), 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_25_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286813248))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290090112))), 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(290090240)))]; + tensor pretrained_out_311_cast_fp16 = conv(bias = layers_25_fc2_pretrained_bias_to_fp16, dilations = var_6028, groups = var_5830, pad = pretrained_out_311_pad_0, pad_type = pretrained_out_311_pad_type_0, strides = var_6026, weight = layers_25_fc2_pretrained_weight_to_fp16_palettized, x = input_517_cast_fp16)[name = tensor("pretrained_out_311_cast_fp16")]; + tensor var_6032 = const()[name = tensor("op_6032"), val = tensor([1, 1])]; + tensor var_6034 = const()[name = tensor("op_6034"), 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_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(290092864)))]; + tensor input_519_cast_fp16 = conv(dilations = var_6034, groups = var_5830, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = var_6032, weight = layers_25_fc2_loraA_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("input_519_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 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(290256768)))]; + tensor lora_out_623_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6040, groups = var_5830, pad = lora_out_621_pad_0, pad_type = lora_out_621_pad_type_0, strides = var_6038, weight = lora_out_623_weight_0_to_fp16, x = input_519_cast_fp16)[name = tensor("lora_out_623_cast_fp16")]; + tensor hidden_states_55_cast_fp16 = add(x = pretrained_out_311_cast_fp16, y = lora_out_623_cast_fp16)[name = tensor("hidden_states_55_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = hidden_states_55_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; + tensor var_6054 = const()[name = tensor("op_6054"), val = tensor(3)]; + tensor var_6056 = const()[name = tensor("op_6056"), val = tensor(1)]; + tensor var_6057 = const()[name = tensor("op_6057"), val = tensor(true)]; + tensor var_6067 = const()[name = tensor("op_6067"), val = tensor([1])]; + tensor channels_mean_105_cast_fp16 = reduce_mean(axes = var_6067, keep_dims = var_6057, 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_6071 = const()[name = tensor("op_6071"), val = tensor([1])]; + tensor var_6072_cast_fp16 = reduce_mean(axes = var_6071, keep_dims = var_6057, x = zero_mean_sq_105_cast_fp16)[name = tensor("op_6072_cast_fp16")]; + tensor var_6073_to_fp16 = const()[name = tensor("op_6073_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6074_cast_fp16 = add(x = var_6072_cast_fp16, y = var_6073_to_fp16)[name = tensor("op_6074_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_6074_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_105_gamma_0_to_fp16 = const()[name = tensor("obj_105_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290297792)))]; + tensor obj_105_beta_0_to_fp16 = const()[name = tensor("obj_105_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290300416)))]; + tensor obj_105_epsilon_0_to_fp16 = const()[name = tensor("obj_105_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_105_cast_fp16 = batch_norm(beta = obj_105_beta_0_to_fp16, epsilon = obj_105_epsilon_0_to_fp16, gamma = obj_105_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_105_cast_fp16")]; + tensor var_6092 = const()[name = tensor("op_6092"), val = tensor([1, 1])]; + tensor var_6094 = const()[name = tensor("op_6094"), 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_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(290303040))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291122304))), 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(291122432)))]; + tensor pretrained_out_313_cast_fp16 = conv(bias = layers_26_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6094, groups = var_6056, pad = pretrained_out_313_pad_0, pad_type = pretrained_out_313_pad_type_0, strides = var_6092, weight = layers_26_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_105_cast_fp16)[name = tensor("pretrained_out_313_cast_fp16")]; + tensor var_6098 = const()[name = tensor("op_6098"), val = tensor([1, 1])]; + tensor var_6100 = const()[name = tensor("op_6100"), 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_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(291125056)))]; + tensor input_521_cast_fp16 = conv(dilations = var_6100, groups = var_6056, pad = input_521_pad_0, pad_type = input_521_pad_type_0, strides = var_6098, weight = layers_26_self_attn_q_proj_loraA_weight_to_fp16, x = obj_105_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor var_6104 = const()[name = tensor("op_6104"), val = tensor([1, 1])]; + tensor var_6106 = const()[name = tensor("op_6106"), 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(291166080)))]; + tensor lora_out_627_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6106, groups = var_6056, pad = lora_out_625_pad_0, pad_type = lora_out_625_pad_type_0, strides = var_6104, weight = lora_out_627_weight_0_to_fp16, x = input_521_cast_fp16)[name = tensor("lora_out_627_cast_fp16")]; + tensor query_53_cast_fp16 = add(x = pretrained_out_313_cast_fp16, y = lora_out_627_cast_fp16)[name = tensor("query_53_cast_fp16")]; + tensor var_6116 = const()[name = tensor("op_6116"), val = tensor([1, 1])]; + tensor var_6118 = const()[name = tensor("op_6118"), 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_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(291207104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292026368))), name = tensor("layers_26_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_315_cast_fp16 = conv(dilations = var_6118, groups = var_6056, pad = pretrained_out_315_pad_0, pad_type = pretrained_out_315_pad_type_0, strides = var_6116, weight = layers_26_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_105_cast_fp16)[name = tensor("pretrained_out_315_cast_fp16")]; + tensor var_6122 = const()[name = tensor("op_6122"), val = tensor([1, 1])]; + tensor var_6124 = const()[name = tensor("op_6124"), 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_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(292026496)))]; + tensor input_523_cast_fp16 = conv(dilations = var_6124, groups = var_6056, pad = input_523_pad_0, pad_type = input_523_pad_type_0, strides = var_6122, weight = layers_26_self_attn_k_proj_loraA_weight_to_fp16, x = obj_105_cast_fp16)[name = tensor("input_523_cast_fp16")]; + tensor var_6128 = const()[name = tensor("op_6128"), val = tensor([1, 1])]; + tensor var_6130 = const()[name = tensor("op_6130"), 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(292067520)))]; + tensor lora_out_631_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6130, groups = var_6056, pad = lora_out_629_pad_0, pad_type = lora_out_629_pad_type_0, strides = var_6128, weight = lora_out_631_weight_0_to_fp16, x = input_523_cast_fp16)[name = tensor("lora_out_631_cast_fp16")]; + tensor key_53_cast_fp16 = add(x = pretrained_out_315_cast_fp16, y = lora_out_631_cast_fp16)[name = tensor("key_53_cast_fp16")]; + tensor var_6141 = const()[name = tensor("op_6141"), val = tensor([1, 1])]; + tensor var_6143 = const()[name = tensor("op_6143"), 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_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(292108544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292927808))), 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(292927936)))]; + tensor pretrained_out_317_cast_fp16 = conv(bias = layers_26_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6143, groups = var_6056, pad = pretrained_out_317_pad_0, pad_type = pretrained_out_317_pad_type_0, strides = var_6141, weight = layers_26_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_105_cast_fp16)[name = tensor("pretrained_out_317_cast_fp16")]; + tensor var_6147 = const()[name = tensor("op_6147"), val = tensor([1, 1])]; + tensor var_6149 = const()[name = tensor("op_6149"), 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_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(292930560)))]; + tensor input_525_cast_fp16 = conv(dilations = var_6149, groups = var_6056, pad = input_525_pad_0, pad_type = input_525_pad_type_0, strides = var_6147, weight = layers_26_self_attn_v_proj_loraA_weight_to_fp16, x = obj_105_cast_fp16)[name = tensor("input_525_cast_fp16")]; + tensor var_6153 = const()[name = tensor("op_6153"), val = tensor([1, 1])]; + tensor var_6155 = const()[name = tensor("op_6155"), 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(292971584)))]; + tensor lora_out_635_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6155, groups = var_6056, pad = lora_out_633_pad_0, pad_type = lora_out_633_pad_type_0, strides = var_6153, weight = lora_out_635_weight_0_to_fp16, x = input_525_cast_fp16)[name = tensor("lora_out_635_cast_fp16")]; + tensor value_53_cast_fp16 = add(x = pretrained_out_317_cast_fp16, y = lora_out_635_cast_fp16)[name = tensor("value_53_cast_fp16")]; + tensor var_6162 = const()[name = tensor("op_6162"), val = tensor([1, 20, 64, -1])]; + tensor var_6163_cast_fp16 = reshape(shape = var_6162, x = query_53_cast_fp16)[name = tensor("op_6163_cast_fp16")]; + tensor var_6164_to_fp16 = const()[name = tensor("op_6164_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6165_cast_fp16 = mul(x = var_6163_cast_fp16, y = var_6164_to_fp16)[name = tensor("op_6165_cast_fp16")]; + tensor var_6166 = const()[name = tensor("op_6166"), val = tensor([1, 20, 64, -1])]; + tensor var_6167_cast_fp16 = reshape(shape = var_6166, x = key_53_cast_fp16)[name = tensor("op_6167_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_6165_cast_fp16, y = var_6167_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; + tensor var_6170_cast_fp16 = softmax(axis = var_6054, x = mh_w_53_cast_fp16)[name = tensor("op_6170_cast_fp16")]; + tensor var_6171 = const()[name = tensor("op_6171"), val = tensor([1, 20, 64, -1])]; + tensor var_6172_cast_fp16 = reshape(shape = var_6171, x = value_53_cast_fp16)[name = tensor("op_6172_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_6172_cast_fp16, y = var_6170_cast_fp16)[name = tensor("attn_53_cast_fp16")]; + tensor var_6175 = const()[name = tensor("op_6175"), val = tensor([1, 1280, 1, -1])]; + tensor input_527_cast_fp16 = reshape(shape = var_6175, x = attn_53_cast_fp16)[name = tensor("input_527_cast_fp16")]; + tensor var_6182 = const()[name = tensor("op_6182"), val = tensor([1, 1])]; + tensor var_6184 = const()[name = tensor("op_6184"), 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_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(293012608))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293831872))), 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(293832000)))]; + tensor pretrained_out_319_cast_fp16 = conv(bias = layers_26_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_6184, groups = var_6056, pad = pretrained_out_319_pad_0, pad_type = pretrained_out_319_pad_type_0, strides = var_6182, weight = layers_26_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_527_cast_fp16)[name = tensor("pretrained_out_319_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 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_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(293834624)))]; + tensor input_529_cast_fp16 = conv(dilations = var_6190, groups = var_6056, pad = input_529_pad_0, pad_type = input_529_pad_type_0, strides = var_6188, weight = layers_26_self_attn_o_proj_loraA_weight_to_fp16, x = input_527_cast_fp16)[name = tensor("input_529_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 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(293875648)))]; + tensor lora_out_639_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6196, groups = var_6056, pad = lora_out_637_pad_0, pad_type = lora_out_637_pad_type_0, strides = var_6194, weight = lora_out_639_weight_0_to_fp16, x = input_529_cast_fp16)[name = tensor("lora_out_639_cast_fp16")]; + tensor obj_107_cast_fp16 = add(x = pretrained_out_319_cast_fp16, y = lora_out_639_cast_fp16)[name = tensor("obj_107_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_107_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; + tensor var_6205 = const()[name = tensor("op_6205"), val = tensor([1])]; + tensor channels_mean_107_cast_fp16 = reduce_mean(axes = var_6205, keep_dims = var_6057, 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_6209 = const()[name = tensor("op_6209"), val = tensor([1])]; + tensor var_6210_cast_fp16 = reduce_mean(axes = var_6209, keep_dims = var_6057, x = zero_mean_sq_107_cast_fp16)[name = tensor("op_6210_cast_fp16")]; + tensor var_6211_to_fp16 = const()[name = tensor("op_6211_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6212_cast_fp16 = add(x = var_6210_cast_fp16, y = var_6211_to_fp16)[name = tensor("op_6212_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_6212_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(293916672)))]; + 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(293919296)))]; + 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_6226 = const()[name = tensor("op_6226"), val = tensor([1, 1])]; + tensor var_6228 = const()[name = tensor("op_6228"), 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_26_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293921920))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297198784))), 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(297198912)))]; + tensor pretrained_out_321_cast_fp16 = conv(bias = layers_26_fc1_pretrained_bias_to_fp16, dilations = var_6228, groups = var_6056, pad = pretrained_out_321_pad_0, pad_type = pretrained_out_321_pad_type_0, strides = var_6226, weight = layers_26_fc1_pretrained_weight_to_fp16_palettized, x = input_531_cast_fp16)[name = tensor("pretrained_out_321_cast_fp16")]; + tensor var_6232 = const()[name = tensor("op_6232"), val = tensor([1, 1])]; + tensor var_6234 = const()[name = tensor("op_6234"), 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_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(297209216)))]; + tensor input_533_cast_fp16 = conv(dilations = var_6234, groups = var_6056, pad = input_533_pad_0, pad_type = input_533_pad_type_0, strides = var_6232, weight = layers_26_fc1_loraA_weight_to_fp16, x = input_531_cast_fp16)[name = tensor("input_533_cast_fp16")]; + tensor var_6238 = const()[name = tensor("op_6238"), val = tensor([1, 1])]; + tensor var_6240 = const()[name = tensor("op_6240"), 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(297250240)))]; + tensor lora_out_643_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_6240, groups = var_6056, pad = lora_out_641_pad_0, pad_type = lora_out_641_pad_type_0, strides = var_6238, weight = lora_out_643_weight_0_to_fp16, x = input_533_cast_fp16)[name = tensor("lora_out_643_cast_fp16")]; + tensor input_535_cast_fp16 = add(x = pretrained_out_321_cast_fp16, y = lora_out_643_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_6252 = const()[name = tensor("op_6252"), val = tensor([1, 1])]; + tensor var_6254 = const()[name = tensor("op_6254"), 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_26_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297414144))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(300691008))), 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(300691136)))]; + tensor pretrained_out_323_cast_fp16 = conv(bias = layers_26_fc2_pretrained_bias_to_fp16, dilations = var_6254, groups = var_6056, pad = pretrained_out_323_pad_0, pad_type = pretrained_out_323_pad_type_0, strides = var_6252, weight = layers_26_fc2_pretrained_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = tensor("pretrained_out_323_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 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_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(300693760)))]; + tensor input_539_cast_fp16 = conv(dilations = var_6260, groups = var_6056, pad = input_539_pad_0, pad_type = input_539_pad_type_0, strides = var_6258, weight = layers_26_fc2_loraA_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("input_539_cast_fp16")]; + tensor var_6264 = const()[name = tensor("op_6264"), val = tensor([1, 1])]; + tensor var_6266 = const()[name = tensor("op_6266"), 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(300857664)))]; + tensor lora_out_647_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6266, groups = var_6056, pad = lora_out_645_pad_0, pad_type = lora_out_645_pad_type_0, strides = var_6264, weight = lora_out_647_weight_0_to_fp16, x = input_539_cast_fp16)[name = tensor("lora_out_647_cast_fp16")]; + tensor hidden_states_57_cast_fp16 = add(x = pretrained_out_323_cast_fp16, y = lora_out_647_cast_fp16)[name = tensor("hidden_states_57_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_57_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; + tensor var_6280 = const()[name = tensor("op_6280"), val = tensor(3)]; + tensor var_6282 = const()[name = tensor("op_6282"), val = tensor(1)]; + tensor var_6283 = const()[name = tensor("op_6283"), val = tensor(true)]; + tensor var_6293 = const()[name = tensor("op_6293"), val = tensor([1])]; + tensor channels_mean_109_cast_fp16 = reduce_mean(axes = var_6293, keep_dims = var_6283, 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_6297 = const()[name = tensor("op_6297"), val = tensor([1])]; + tensor var_6298_cast_fp16 = reduce_mean(axes = var_6297, keep_dims = var_6283, x = zero_mean_sq_109_cast_fp16)[name = tensor("op_6298_cast_fp16")]; + tensor var_6299_to_fp16 = const()[name = tensor("op_6299_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6300_cast_fp16 = add(x = var_6298_cast_fp16, y = var_6299_to_fp16)[name = tensor("op_6300_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_6300_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_109_gamma_0_to_fp16 = const()[name = tensor("obj_109_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(300898688)))]; + tensor obj_109_beta_0_to_fp16 = const()[name = tensor("obj_109_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(300901312)))]; + tensor obj_109_epsilon_0_to_fp16 = const()[name = tensor("obj_109_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_109_cast_fp16 = batch_norm(beta = obj_109_beta_0_to_fp16, epsilon = obj_109_epsilon_0_to_fp16, gamma = obj_109_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_109_cast_fp16")]; + tensor var_6318 = const()[name = tensor("op_6318"), val = tensor([1, 1])]; + tensor var_6320 = const()[name = tensor("op_6320"), 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_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(300903936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(301723200))), 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(301723328)))]; + tensor pretrained_out_325_cast_fp16 = conv(bias = layers_27_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6320, groups = var_6282, pad = pretrained_out_325_pad_0, pad_type = pretrained_out_325_pad_type_0, strides = var_6318, weight = layers_27_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = tensor("pretrained_out_325_cast_fp16")]; + tensor var_6324 = const()[name = tensor("op_6324"), val = tensor([1, 1])]; + tensor var_6326 = const()[name = tensor("op_6326"), 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_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(301725952)))]; + tensor input_541_cast_fp16 = conv(dilations = var_6326, groups = var_6282, pad = input_541_pad_0, pad_type = input_541_pad_type_0, strides = var_6324, weight = layers_27_self_attn_q_proj_loraA_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor("input_541_cast_fp16")]; + tensor var_6330 = const()[name = tensor("op_6330"), val = tensor([1, 1])]; + tensor var_6332 = const()[name = tensor("op_6332"), 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(301766976)))]; + tensor lora_out_651_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6332, groups = var_6282, pad = lora_out_649_pad_0, pad_type = lora_out_649_pad_type_0, strides = var_6330, weight = lora_out_651_weight_0_to_fp16, x = input_541_cast_fp16)[name = tensor("lora_out_651_cast_fp16")]; + tensor query_55_cast_fp16 = add(x = pretrained_out_325_cast_fp16, y = lora_out_651_cast_fp16)[name = tensor("query_55_cast_fp16")]; + tensor var_6342 = const()[name = tensor("op_6342"), val = tensor([1, 1])]; + tensor var_6344 = const()[name = tensor("op_6344"), 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_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(301808000))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302627264))), name = tensor("layers_27_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_327_cast_fp16 = conv(dilations = var_6344, groups = var_6282, pad = pretrained_out_327_pad_0, pad_type = pretrained_out_327_pad_type_0, strides = var_6342, weight = layers_27_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = tensor("pretrained_out_327_cast_fp16")]; + tensor var_6348 = const()[name = tensor("op_6348"), val = tensor([1, 1])]; + tensor var_6350 = const()[name = tensor("op_6350"), 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_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(302627392)))]; + tensor input_543_cast_fp16 = conv(dilations = var_6350, groups = var_6282, pad = input_543_pad_0, pad_type = input_543_pad_type_0, strides = var_6348, weight = layers_27_self_attn_k_proj_loraA_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor("input_543_cast_fp16")]; + tensor var_6354 = const()[name = tensor("op_6354"), val = tensor([1, 1])]; + tensor var_6356 = const()[name = tensor("op_6356"), 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(302668416)))]; + tensor lora_out_655_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6356, groups = var_6282, pad = lora_out_653_pad_0, pad_type = lora_out_653_pad_type_0, strides = var_6354, weight = lora_out_655_weight_0_to_fp16, x = input_543_cast_fp16)[name = tensor("lora_out_655_cast_fp16")]; + tensor key_55_cast_fp16 = add(x = pretrained_out_327_cast_fp16, y = lora_out_655_cast_fp16)[name = tensor("key_55_cast_fp16")]; + tensor var_6367 = const()[name = tensor("op_6367"), val = tensor([1, 1])]; + tensor var_6369 = const()[name = tensor("op_6369"), 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_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(302709440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303528704))), 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(303528832)))]; + tensor pretrained_out_329_cast_fp16 = conv(bias = layers_27_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6369, groups = var_6282, pad = pretrained_out_329_pad_0, pad_type = pretrained_out_329_pad_type_0, strides = var_6367, weight = layers_27_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_109_cast_fp16)[name = tensor("pretrained_out_329_cast_fp16")]; + tensor var_6373 = const()[name = tensor("op_6373"), val = tensor([1, 1])]; + tensor var_6375 = const()[name = tensor("op_6375"), 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_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(303531456)))]; + tensor input_545_cast_fp16 = conv(dilations = var_6375, groups = var_6282, pad = input_545_pad_0, pad_type = input_545_pad_type_0, strides = var_6373, weight = layers_27_self_attn_v_proj_loraA_weight_to_fp16, x = obj_109_cast_fp16)[name = tensor("input_545_cast_fp16")]; + tensor var_6379 = const()[name = tensor("op_6379"), val = tensor([1, 1])]; + tensor var_6381 = const()[name = tensor("op_6381"), 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(303572480)))]; + tensor lora_out_659_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6381, groups = var_6282, pad = lora_out_657_pad_0, pad_type = lora_out_657_pad_type_0, strides = var_6379, weight = lora_out_659_weight_0_to_fp16, x = input_545_cast_fp16)[name = tensor("lora_out_659_cast_fp16")]; + tensor value_55_cast_fp16 = add(x = pretrained_out_329_cast_fp16, y = lora_out_659_cast_fp16)[name = tensor("value_55_cast_fp16")]; + tensor var_6388 = const()[name = tensor("op_6388"), val = tensor([1, 20, 64, -1])]; + tensor var_6389_cast_fp16 = reshape(shape = var_6388, x = query_55_cast_fp16)[name = tensor("op_6389_cast_fp16")]; + tensor var_6390_to_fp16 = const()[name = tensor("op_6390_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6391_cast_fp16 = mul(x = var_6389_cast_fp16, y = var_6390_to_fp16)[name = tensor("op_6391_cast_fp16")]; + tensor var_6392 = const()[name = tensor("op_6392"), val = tensor([1, 20, 64, -1])]; + tensor var_6393_cast_fp16 = reshape(shape = var_6392, x = key_55_cast_fp16)[name = tensor("op_6393_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_6391_cast_fp16, y = var_6393_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; + tensor var_6396_cast_fp16 = softmax(axis = var_6280, x = mh_w_55_cast_fp16)[name = tensor("op_6396_cast_fp16")]; + tensor var_6397 = const()[name = tensor("op_6397"), val = tensor([1, 20, 64, -1])]; + tensor var_6398_cast_fp16 = reshape(shape = var_6397, x = value_55_cast_fp16)[name = tensor("op_6398_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_6398_cast_fp16, y = var_6396_cast_fp16)[name = tensor("attn_55_cast_fp16")]; + tensor var_6401 = const()[name = tensor("op_6401"), val = tensor([1, 1280, 1, -1])]; + tensor input_547_cast_fp16 = reshape(shape = var_6401, x = attn_55_cast_fp16)[name = tensor("input_547_cast_fp16")]; + tensor var_6408 = const()[name = tensor("op_6408"), val = tensor([1, 1])]; + tensor var_6410 = const()[name = tensor("op_6410"), 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_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(303613504))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(304432768))), 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(304432896)))]; + tensor pretrained_out_331_cast_fp16 = conv(bias = layers_27_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_6410, groups = var_6282, pad = pretrained_out_331_pad_0, pad_type = pretrained_out_331_pad_type_0, strides = var_6408, weight = layers_27_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = tensor("pretrained_out_331_cast_fp16")]; + tensor var_6414 = const()[name = tensor("op_6414"), val = tensor([1, 1])]; + tensor var_6416 = const()[name = tensor("op_6416"), 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_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(304435520)))]; + tensor input_549_cast_fp16 = conv(dilations = var_6416, groups = var_6282, pad = input_549_pad_0, pad_type = input_549_pad_type_0, strides = var_6414, weight = layers_27_self_attn_o_proj_loraA_weight_to_fp16, x = input_547_cast_fp16)[name = tensor("input_549_cast_fp16")]; + tensor var_6420 = const()[name = tensor("op_6420"), val = tensor([1, 1])]; + tensor var_6422 = const()[name = tensor("op_6422"), 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(304476544)))]; + tensor lora_out_663_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6422, groups = var_6282, pad = lora_out_661_pad_0, pad_type = lora_out_661_pad_type_0, strides = var_6420, weight = lora_out_663_weight_0_to_fp16, x = input_549_cast_fp16)[name = tensor("lora_out_663_cast_fp16")]; + tensor obj_111_cast_fp16 = add(x = pretrained_out_331_cast_fp16, y = lora_out_663_cast_fp16)[name = tensor("obj_111_cast_fp16")]; + tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_111_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor var_6431 = const()[name = tensor("op_6431"), val = tensor([1])]; + tensor channels_mean_111_cast_fp16 = reduce_mean(axes = var_6431, keep_dims = var_6283, 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_6435 = const()[name = tensor("op_6435"), val = tensor([1])]; + tensor var_6436_cast_fp16 = reduce_mean(axes = var_6435, keep_dims = var_6283, x = zero_mean_sq_111_cast_fp16)[name = tensor("op_6436_cast_fp16")]; + tensor var_6437_to_fp16 = const()[name = tensor("op_6437_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6438_cast_fp16 = add(x = var_6436_cast_fp16, y = var_6437_to_fp16)[name = tensor("op_6438_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_6438_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 input_551_gamma_0_to_fp16 = const()[name = tensor("input_551_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(304517568)))]; + tensor input_551_beta_0_to_fp16 = const()[name = tensor("input_551_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(304520192)))]; + tensor input_551_epsilon_0_to_fp16 = const()[name = tensor("input_551_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_551_cast_fp16 = batch_norm(beta = input_551_beta_0_to_fp16, epsilon = input_551_epsilon_0_to_fp16, gamma = input_551_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("input_551_cast_fp16")]; + tensor var_6452 = const()[name = tensor("op_6452"), val = tensor([1, 1])]; + tensor var_6454 = const()[name = tensor("op_6454"), 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_27_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(304522816))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(307799680))), 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(307799808)))]; + tensor pretrained_out_333_cast_fp16 = conv(bias = layers_27_fc1_pretrained_bias_to_fp16, dilations = var_6454, groups = var_6282, pad = pretrained_out_333_pad_0, pad_type = pretrained_out_333_pad_type_0, strides = var_6452, weight = layers_27_fc1_pretrained_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = tensor("pretrained_out_333_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 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_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(307810112)))]; + tensor input_553_cast_fp16 = conv(dilations = var_6460, groups = var_6282, pad = input_553_pad_0, pad_type = input_553_pad_type_0, strides = var_6458, weight = layers_27_fc1_loraA_weight_to_fp16, x = input_551_cast_fp16)[name = tensor("input_553_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 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(307851136)))]; + tensor lora_out_667_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_6466, groups = var_6282, pad = lora_out_665_pad_0, pad_type = lora_out_665_pad_type_0, strides = var_6464, weight = lora_out_667_weight_0_to_fp16, x = input_553_cast_fp16)[name = tensor("lora_out_667_cast_fp16")]; + tensor input_555_cast_fp16 = add(x = pretrained_out_333_cast_fp16, y = lora_out_667_cast_fp16)[name = tensor("input_555_cast_fp16")]; + tensor input_557_mode_0 = const()[name = tensor("input_557_mode_0"), val = tensor("EXACT")]; + tensor input_557_cast_fp16 = gelu(mode = input_557_mode_0, x = input_555_cast_fp16)[name = tensor("input_557_cast_fp16")]; + tensor var_6478 = const()[name = tensor("op_6478"), val = tensor([1, 1])]; + tensor var_6480 = const()[name = tensor("op_6480"), 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_27_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308015040))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311291904))), 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(311292032)))]; + tensor pretrained_out_335_cast_fp16 = conv(bias = layers_27_fc2_pretrained_bias_to_fp16, dilations = var_6480, groups = var_6282, pad = pretrained_out_335_pad_0, pad_type = pretrained_out_335_pad_type_0, strides = var_6478, weight = layers_27_fc2_pretrained_weight_to_fp16_palettized, x = input_557_cast_fp16)[name = tensor("pretrained_out_335_cast_fp16")]; + tensor var_6484 = const()[name = tensor("op_6484"), val = tensor([1, 1])]; + tensor var_6486 = const()[name = tensor("op_6486"), 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_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(311294656)))]; + tensor input_559_cast_fp16 = conv(dilations = var_6486, groups = var_6282, pad = input_559_pad_0, pad_type = input_559_pad_type_0, strides = var_6484, weight = layers_27_fc2_loraA_weight_to_fp16, x = input_557_cast_fp16)[name = tensor("input_559_cast_fp16")]; + tensor var_6490 = const()[name = tensor("op_6490"), val = tensor([1, 1])]; + tensor var_6492 = const()[name = tensor("op_6492"), 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(311458560)))]; + tensor lora_out_671_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6492, groups = var_6282, pad = lora_out_669_pad_0, pad_type = lora_out_669_pad_type_0, strides = var_6490, weight = lora_out_671_weight_0_to_fp16, x = input_559_cast_fp16)[name = tensor("lora_out_671_cast_fp16")]; + tensor hidden_states_59_cast_fp16 = add(x = pretrained_out_335_cast_fp16, y = lora_out_671_cast_fp16)[name = tensor("hidden_states_59_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = hidden_states_59_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; + tensor var_6506 = const()[name = tensor("op_6506"), val = tensor(3)]; + tensor var_6508 = const()[name = tensor("op_6508"), val = tensor(1)]; + tensor var_6509 = const()[name = tensor("op_6509"), val = tensor(true)]; + tensor var_6519 = const()[name = tensor("op_6519"), val = tensor([1])]; + tensor channels_mean_113_cast_fp16 = reduce_mean(axes = var_6519, keep_dims = var_6509, 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_6523 = const()[name = tensor("op_6523"), val = tensor([1])]; + tensor var_6524_cast_fp16 = reduce_mean(axes = var_6523, keep_dims = var_6509, x = zero_mean_sq_113_cast_fp16)[name = tensor("op_6524_cast_fp16")]; + tensor var_6525_to_fp16 = const()[name = tensor("op_6525_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6526_cast_fp16 = add(x = var_6524_cast_fp16, y = var_6525_to_fp16)[name = tensor("op_6526_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_6526_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 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(311499584)))]; + 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(311502208)))]; + 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_113_cast_fp16)[name = tensor("obj_113_cast_fp16")]; + tensor var_6544 = const()[name = tensor("op_6544"), val = tensor([1, 1])]; + tensor var_6546 = const()[name = tensor("op_6546"), 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_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(311504832))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312324096))), 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(312324224)))]; + tensor pretrained_out_337_cast_fp16 = conv(bias = layers_28_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6546, groups = var_6508, pad = pretrained_out_337_pad_0, pad_type = pretrained_out_337_pad_type_0, strides = var_6544, weight = layers_28_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_113_cast_fp16)[name = tensor("pretrained_out_337_cast_fp16")]; + tensor var_6550 = const()[name = tensor("op_6550"), val = tensor([1, 1])]; + tensor var_6552 = const()[name = tensor("op_6552"), val = tensor([1, 1])]; + tensor input_561_pad_type_0 = const()[name = tensor("input_561_pad_type_0"), val = tensor("custom")]; + tensor input_561_pad_0 = const()[name = tensor("input_561_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(312326848)))]; + tensor input_561_cast_fp16 = conv(dilations = var_6552, groups = var_6508, pad = input_561_pad_0, pad_type = input_561_pad_type_0, strides = var_6550, weight = layers_28_self_attn_q_proj_loraA_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("input_561_cast_fp16")]; + tensor var_6556 = const()[name = tensor("op_6556"), val = tensor([1, 1])]; + tensor var_6558 = const()[name = tensor("op_6558"), 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(312367872)))]; + tensor lora_out_675_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6558, groups = var_6508, pad = lora_out_673_pad_0, pad_type = lora_out_673_pad_type_0, strides = var_6556, weight = lora_out_675_weight_0_to_fp16, x = input_561_cast_fp16)[name = tensor("lora_out_675_cast_fp16")]; + tensor query_57_cast_fp16 = add(x = pretrained_out_337_cast_fp16, y = lora_out_675_cast_fp16)[name = tensor("query_57_cast_fp16")]; + tensor var_6568 = const()[name = tensor("op_6568"), val = tensor([1, 1])]; + tensor var_6570 = const()[name = tensor("op_6570"), 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_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(312408896))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313228160))), name = tensor("layers_28_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_339_cast_fp16 = conv(dilations = var_6570, groups = var_6508, pad = pretrained_out_339_pad_0, pad_type = pretrained_out_339_pad_type_0, strides = var_6568, weight = layers_28_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_113_cast_fp16)[name = tensor("pretrained_out_339_cast_fp16")]; + tensor var_6574 = const()[name = tensor("op_6574"), val = tensor([1, 1])]; + tensor var_6576 = const()[name = tensor("op_6576"), 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_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(313228288)))]; + tensor input_563_cast_fp16 = conv(dilations = var_6576, groups = var_6508, pad = input_563_pad_0, pad_type = input_563_pad_type_0, strides = var_6574, weight = layers_28_self_attn_k_proj_loraA_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("input_563_cast_fp16")]; + tensor var_6580 = const()[name = tensor("op_6580"), val = tensor([1, 1])]; + tensor var_6582 = const()[name = tensor("op_6582"), 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(313269312)))]; + tensor lora_out_679_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6582, groups = var_6508, pad = lora_out_677_pad_0, pad_type = lora_out_677_pad_type_0, strides = var_6580, weight = lora_out_679_weight_0_to_fp16, x = input_563_cast_fp16)[name = tensor("lora_out_679_cast_fp16")]; + tensor key_57_cast_fp16 = add(x = pretrained_out_339_cast_fp16, y = lora_out_679_cast_fp16)[name = tensor("key_57_cast_fp16")]; + tensor var_6593 = const()[name = tensor("op_6593"), val = tensor([1, 1])]; + tensor var_6595 = const()[name = tensor("op_6595"), 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_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(313310336))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314129600))), 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(314129728)))]; + tensor pretrained_out_341_cast_fp16 = conv(bias = layers_28_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6595, groups = var_6508, pad = pretrained_out_341_pad_0, pad_type = pretrained_out_341_pad_type_0, strides = var_6593, weight = layers_28_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_113_cast_fp16)[name = tensor("pretrained_out_341_cast_fp16")]; + tensor var_6599 = const()[name = tensor("op_6599"), val = tensor([1, 1])]; + tensor var_6601 = const()[name = tensor("op_6601"), val = tensor([1, 1])]; + tensor input_565_pad_type_0 = const()[name = tensor("input_565_pad_type_0"), val = tensor("custom")]; + tensor input_565_pad_0 = const()[name = tensor("input_565_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(314132352)))]; + tensor input_565_cast_fp16 = conv(dilations = var_6601, groups = var_6508, pad = input_565_pad_0, pad_type = input_565_pad_type_0, strides = var_6599, weight = layers_28_self_attn_v_proj_loraA_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("input_565_cast_fp16")]; + tensor var_6605 = const()[name = tensor("op_6605"), val = tensor([1, 1])]; + tensor var_6607 = const()[name = tensor("op_6607"), 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(314173376)))]; + tensor lora_out_683_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6607, groups = var_6508, pad = lora_out_681_pad_0, pad_type = lora_out_681_pad_type_0, strides = var_6605, weight = lora_out_683_weight_0_to_fp16, x = input_565_cast_fp16)[name = tensor("lora_out_683_cast_fp16")]; + tensor value_57_cast_fp16 = add(x = pretrained_out_341_cast_fp16, y = lora_out_683_cast_fp16)[name = tensor("value_57_cast_fp16")]; + tensor var_6614 = const()[name = tensor("op_6614"), val = tensor([1, 20, 64, -1])]; + tensor var_6615_cast_fp16 = reshape(shape = var_6614, x = query_57_cast_fp16)[name = tensor("op_6615_cast_fp16")]; + tensor var_6616_to_fp16 = const()[name = tensor("op_6616_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6617_cast_fp16 = mul(x = var_6615_cast_fp16, y = var_6616_to_fp16)[name = tensor("op_6617_cast_fp16")]; + tensor var_6618 = const()[name = tensor("op_6618"), val = tensor([1, 20, 64, -1])]; + tensor var_6619_cast_fp16 = reshape(shape = var_6618, x = key_57_cast_fp16)[name = tensor("op_6619_cast_fp16")]; + tensor mh_w_57_transpose_x_0 = const()[name = tensor("mh_w_57_transpose_x_0"), val = tensor(true)]; + tensor mh_w_57_transpose_y_0 = const()[name = tensor("mh_w_57_transpose_y_0"), val = tensor(false)]; + tensor mh_w_57_cast_fp16 = matmul(transpose_x = mh_w_57_transpose_x_0, transpose_y = mh_w_57_transpose_y_0, x = var_6617_cast_fp16, y = var_6619_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; + tensor var_6622_cast_fp16 = softmax(axis = var_6506, x = mh_w_57_cast_fp16)[name = tensor("op_6622_cast_fp16")]; + tensor var_6623 = const()[name = tensor("op_6623"), val = tensor([1, 20, 64, -1])]; + tensor var_6624_cast_fp16 = reshape(shape = var_6623, x = value_57_cast_fp16)[name = tensor("op_6624_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_6624_cast_fp16, y = var_6622_cast_fp16)[name = tensor("attn_57_cast_fp16")]; + tensor var_6627 = const()[name = tensor("op_6627"), val = tensor([1, 1280, 1, -1])]; + tensor input_567_cast_fp16 = reshape(shape = var_6627, x = attn_57_cast_fp16)[name = tensor("input_567_cast_fp16")]; + tensor var_6634 = const()[name = tensor("op_6634"), val = tensor([1, 1])]; + tensor var_6636 = const()[name = tensor("op_6636"), 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_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(314214400))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315033664))), 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(315033792)))]; + tensor pretrained_out_343_cast_fp16 = conv(bias = layers_28_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_6636, groups = var_6508, pad = pretrained_out_343_pad_0, pad_type = pretrained_out_343_pad_type_0, strides = var_6634, weight = layers_28_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_567_cast_fp16)[name = tensor("pretrained_out_343_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 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_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(315036416)))]; + tensor input_569_cast_fp16 = conv(dilations = var_6642, groups = var_6508, pad = input_569_pad_0, pad_type = input_569_pad_type_0, strides = var_6640, weight = layers_28_self_attn_o_proj_loraA_weight_to_fp16, x = input_567_cast_fp16)[name = tensor("input_569_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 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(315077440)))]; + tensor lora_out_687_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6648, groups = var_6508, pad = lora_out_685_pad_0, pad_type = lora_out_685_pad_type_0, strides = var_6646, weight = lora_out_687_weight_0_to_fp16, x = input_569_cast_fp16)[name = tensor("lora_out_687_cast_fp16")]; + tensor obj_115_cast_fp16 = add(x = pretrained_out_343_cast_fp16, y = lora_out_687_cast_fp16)[name = tensor("obj_115_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = obj_115_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; + tensor var_6657 = const()[name = tensor("op_6657"), val = tensor([1])]; + tensor channels_mean_115_cast_fp16 = reduce_mean(axes = var_6657, keep_dims = var_6509, 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_6661 = const()[name = tensor("op_6661"), val = tensor([1])]; + tensor var_6662_cast_fp16 = reduce_mean(axes = var_6661, keep_dims = var_6509, x = zero_mean_sq_115_cast_fp16)[name = tensor("op_6662_cast_fp16")]; + tensor var_6663_to_fp16 = const()[name = tensor("op_6663_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6664_cast_fp16 = add(x = var_6662_cast_fp16, y = var_6663_to_fp16)[name = tensor("op_6664_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_6664_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 input_571_gamma_0_to_fp16 = const()[name = tensor("input_571_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315118464)))]; + tensor input_571_beta_0_to_fp16 = const()[name = tensor("input_571_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315121088)))]; + tensor input_571_epsilon_0_to_fp16 = const()[name = tensor("input_571_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_571_cast_fp16 = batch_norm(beta = input_571_beta_0_to_fp16, epsilon = input_571_epsilon_0_to_fp16, gamma = input_571_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("input_571_cast_fp16")]; + tensor var_6678 = const()[name = tensor("op_6678"), val = tensor([1, 1])]; + tensor var_6680 = const()[name = tensor("op_6680"), 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_28_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315123712))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318400576))), 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(318400704)))]; + tensor pretrained_out_345_cast_fp16 = conv(bias = layers_28_fc1_pretrained_bias_to_fp16, dilations = var_6680, groups = var_6508, pad = pretrained_out_345_pad_0, pad_type = pretrained_out_345_pad_type_0, strides = var_6678, weight = layers_28_fc1_pretrained_weight_to_fp16_palettized, x = input_571_cast_fp16)[name = tensor("pretrained_out_345_cast_fp16")]; + tensor var_6684 = const()[name = tensor("op_6684"), val = tensor([1, 1])]; + tensor var_6686 = const()[name = tensor("op_6686"), 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_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(318411008)))]; + tensor input_573_cast_fp16 = conv(dilations = var_6686, groups = var_6508, pad = input_573_pad_0, pad_type = input_573_pad_type_0, strides = var_6684, weight = layers_28_fc1_loraA_weight_to_fp16, x = input_571_cast_fp16)[name = tensor("input_573_cast_fp16")]; + tensor var_6690 = const()[name = tensor("op_6690"), val = tensor([1, 1])]; + tensor var_6692 = const()[name = tensor("op_6692"), 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(318452032)))]; + tensor lora_out_691_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_6692, groups = var_6508, pad = lora_out_689_pad_0, pad_type = lora_out_689_pad_type_0, strides = var_6690, weight = lora_out_691_weight_0_to_fp16, x = input_573_cast_fp16)[name = tensor("lora_out_691_cast_fp16")]; + tensor input_575_cast_fp16 = add(x = pretrained_out_345_cast_fp16, y = lora_out_691_cast_fp16)[name = tensor("input_575_cast_fp16")]; + tensor input_577_mode_0 = const()[name = tensor("input_577_mode_0"), val = tensor("EXACT")]; + tensor input_577_cast_fp16 = gelu(mode = input_577_mode_0, x = input_575_cast_fp16)[name = tensor("input_577_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 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_28_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318615936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321892800))), 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(321892928)))]; + tensor pretrained_out_347_cast_fp16 = conv(bias = layers_28_fc2_pretrained_bias_to_fp16, dilations = var_6706, groups = var_6508, pad = pretrained_out_347_pad_0, pad_type = pretrained_out_347_pad_type_0, strides = var_6704, weight = layers_28_fc2_pretrained_weight_to_fp16_palettized, x = input_577_cast_fp16)[name = tensor("pretrained_out_347_cast_fp16")]; + tensor var_6710 = const()[name = tensor("op_6710"), val = tensor([1, 1])]; + tensor var_6712 = const()[name = tensor("op_6712"), 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_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(321895552)))]; + tensor input_579_cast_fp16 = conv(dilations = var_6712, groups = var_6508, pad = input_579_pad_0, pad_type = input_579_pad_type_0, strides = var_6710, weight = layers_28_fc2_loraA_weight_to_fp16, x = input_577_cast_fp16)[name = tensor("input_579_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 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(322059456)))]; + tensor lora_out_695_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6718, groups = var_6508, pad = lora_out_693_pad_0, pad_type = lora_out_693_pad_type_0, strides = var_6716, weight = lora_out_695_weight_0_to_fp16, x = input_579_cast_fp16)[name = tensor("lora_out_695_cast_fp16")]; + tensor hidden_states_61_cast_fp16 = add(x = pretrained_out_347_cast_fp16, y = lora_out_695_cast_fp16)[name = tensor("hidden_states_61_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = hidden_states_61_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; + tensor var_6732 = const()[name = tensor("op_6732"), val = tensor(3)]; + tensor var_6734 = const()[name = tensor("op_6734"), val = tensor(1)]; + tensor var_6735 = const()[name = tensor("op_6735"), val = tensor(true)]; + tensor var_6745 = const()[name = tensor("op_6745"), val = tensor([1])]; + tensor channels_mean_117_cast_fp16 = reduce_mean(axes = var_6745, keep_dims = var_6735, 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_6749 = const()[name = tensor("op_6749"), val = tensor([1])]; + tensor var_6750_cast_fp16 = reduce_mean(axes = var_6749, keep_dims = var_6735, x = zero_mean_sq_117_cast_fp16)[name = tensor("op_6750_cast_fp16")]; + tensor var_6751_to_fp16 = const()[name = tensor("op_6751_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6752_cast_fp16 = add(x = var_6750_cast_fp16, y = var_6751_to_fp16)[name = tensor("op_6752_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_6752_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_117_gamma_0_to_fp16 = const()[name = tensor("obj_117_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322100480)))]; + tensor obj_117_beta_0_to_fp16 = const()[name = tensor("obj_117_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322103104)))]; + tensor obj_117_epsilon_0_to_fp16 = const()[name = tensor("obj_117_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_117_cast_fp16 = batch_norm(beta = obj_117_beta_0_to_fp16, epsilon = obj_117_epsilon_0_to_fp16, gamma = obj_117_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_117_cast_fp16")]; + tensor var_6770 = const()[name = tensor("op_6770"), val = tensor([1, 1])]; + tensor var_6772 = const()[name = tensor("op_6772"), 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_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(322105728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(322924992))), 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(322925120)))]; + tensor pretrained_out_349_cast_fp16 = conv(bias = layers_29_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6772, groups = var_6734, pad = pretrained_out_349_pad_0, pad_type = pretrained_out_349_pad_type_0, strides = var_6770, weight = layers_29_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_117_cast_fp16)[name = tensor("pretrained_out_349_cast_fp16")]; + tensor var_6776 = const()[name = tensor("op_6776"), val = tensor([1, 1])]; + tensor var_6778 = const()[name = tensor("op_6778"), 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_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(322927744)))]; + tensor input_581_cast_fp16 = conv(dilations = var_6778, groups = var_6734, pad = input_581_pad_0, pad_type = input_581_pad_type_0, strides = var_6776, weight = layers_29_self_attn_q_proj_loraA_weight_to_fp16, x = obj_117_cast_fp16)[name = tensor("input_581_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 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(322968768)))]; + tensor lora_out_699_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6784, groups = var_6734, pad = lora_out_697_pad_0, pad_type = lora_out_697_pad_type_0, strides = var_6782, weight = lora_out_699_weight_0_to_fp16, x = input_581_cast_fp16)[name = tensor("lora_out_699_cast_fp16")]; + tensor query_59_cast_fp16 = add(x = pretrained_out_349_cast_fp16, y = lora_out_699_cast_fp16)[name = tensor("query_59_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 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_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(323009792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323829056))), name = tensor("layers_29_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_351_cast_fp16 = conv(dilations = var_6796, groups = var_6734, pad = pretrained_out_351_pad_0, pad_type = pretrained_out_351_pad_type_0, strides = var_6794, weight = layers_29_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_117_cast_fp16)[name = tensor("pretrained_out_351_cast_fp16")]; + tensor var_6800 = const()[name = tensor("op_6800"), val = tensor([1, 1])]; + tensor var_6802 = const()[name = tensor("op_6802"), 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_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(323829184)))]; + tensor input_583_cast_fp16 = conv(dilations = var_6802, groups = var_6734, pad = input_583_pad_0, pad_type = input_583_pad_type_0, strides = var_6800, weight = layers_29_self_attn_k_proj_loraA_weight_to_fp16, x = obj_117_cast_fp16)[name = tensor("input_583_cast_fp16")]; + tensor var_6806 = const()[name = tensor("op_6806"), val = tensor([1, 1])]; + tensor var_6808 = const()[name = tensor("op_6808"), 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(323870208)))]; + tensor lora_out_703_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6808, groups = var_6734, pad = lora_out_701_pad_0, pad_type = lora_out_701_pad_type_0, strides = var_6806, weight = lora_out_703_weight_0_to_fp16, x = input_583_cast_fp16)[name = tensor("lora_out_703_cast_fp16")]; + tensor key_59_cast_fp16 = add(x = pretrained_out_351_cast_fp16, y = lora_out_703_cast_fp16)[name = tensor("key_59_cast_fp16")]; + tensor var_6819 = const()[name = tensor("op_6819"), val = tensor([1, 1])]; + tensor var_6821 = const()[name = tensor("op_6821"), 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_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(323911232))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(324730496))), 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(324730624)))]; + tensor pretrained_out_353_cast_fp16 = conv(bias = layers_29_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_6821, groups = var_6734, pad = pretrained_out_353_pad_0, pad_type = pretrained_out_353_pad_type_0, strides = var_6819, weight = layers_29_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_117_cast_fp16)[name = tensor("pretrained_out_353_cast_fp16")]; + tensor var_6825 = const()[name = tensor("op_6825"), val = tensor([1, 1])]; + tensor var_6827 = const()[name = tensor("op_6827"), 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_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(324733248)))]; + tensor input_585_cast_fp16 = conv(dilations = var_6827, groups = var_6734, pad = input_585_pad_0, pad_type = input_585_pad_type_0, strides = var_6825, weight = layers_29_self_attn_v_proj_loraA_weight_to_fp16, x = obj_117_cast_fp16)[name = tensor("input_585_cast_fp16")]; + tensor var_6831 = const()[name = tensor("op_6831"), val = tensor([1, 1])]; + tensor var_6833 = const()[name = tensor("op_6833"), 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(324774272)))]; + tensor lora_out_707_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6833, groups = var_6734, pad = lora_out_705_pad_0, pad_type = lora_out_705_pad_type_0, strides = var_6831, weight = lora_out_707_weight_0_to_fp16, x = input_585_cast_fp16)[name = tensor("lora_out_707_cast_fp16")]; + tensor value_59_cast_fp16 = add(x = pretrained_out_353_cast_fp16, y = lora_out_707_cast_fp16)[name = tensor("value_59_cast_fp16")]; + tensor var_6840 = const()[name = tensor("op_6840"), val = tensor([1, 20, 64, -1])]; + tensor var_6841_cast_fp16 = reshape(shape = var_6840, x = query_59_cast_fp16)[name = tensor("op_6841_cast_fp16")]; + tensor var_6842_to_fp16 = const()[name = tensor("op_6842_to_fp16"), val = tensor(0x1p-3)]; + tensor var_6843_cast_fp16 = mul(x = var_6841_cast_fp16, y = var_6842_to_fp16)[name = tensor("op_6843_cast_fp16")]; + tensor var_6844 = const()[name = tensor("op_6844"), val = tensor([1, 20, 64, -1])]; + tensor var_6845_cast_fp16 = reshape(shape = var_6844, x = key_59_cast_fp16)[name = tensor("op_6845_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_6843_cast_fp16, y = var_6845_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; + tensor var_6848_cast_fp16 = softmax(axis = var_6732, x = mh_w_59_cast_fp16)[name = tensor("op_6848_cast_fp16")]; + tensor var_6849 = const()[name = tensor("op_6849"), val = tensor([1, 20, 64, -1])]; + tensor var_6850_cast_fp16 = reshape(shape = var_6849, x = value_59_cast_fp16)[name = tensor("op_6850_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_6850_cast_fp16, y = var_6848_cast_fp16)[name = tensor("attn_59_cast_fp16")]; + tensor var_6853 = const()[name = tensor("op_6853"), val = tensor([1, 1280, 1, -1])]; + tensor input_587_cast_fp16 = reshape(shape = var_6853, x = attn_59_cast_fp16)[name = tensor("input_587_cast_fp16")]; + tensor var_6860 = const()[name = tensor("op_6860"), val = tensor([1, 1])]; + tensor var_6862 = const()[name = tensor("op_6862"), 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_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(324815296))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325634560))), 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(325634688)))]; + tensor pretrained_out_355_cast_fp16 = conv(bias = layers_29_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_6862, groups = var_6734, pad = pretrained_out_355_pad_0, pad_type = pretrained_out_355_pad_type_0, strides = var_6860, weight = layers_29_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_587_cast_fp16)[name = tensor("pretrained_out_355_cast_fp16")]; + tensor var_6866 = const()[name = tensor("op_6866"), val = tensor([1, 1])]; + tensor var_6868 = const()[name = tensor("op_6868"), 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_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(325637312)))]; + tensor input_589_cast_fp16 = conv(dilations = var_6868, groups = var_6734, pad = input_589_pad_0, pad_type = input_589_pad_type_0, strides = var_6866, weight = layers_29_self_attn_o_proj_loraA_weight_to_fp16, x = input_587_cast_fp16)[name = tensor("input_589_cast_fp16")]; + tensor var_6872 = const()[name = tensor("op_6872"), val = tensor([1, 1])]; + tensor var_6874 = const()[name = tensor("op_6874"), 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(325678336)))]; + tensor lora_out_711_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6874, groups = var_6734, pad = lora_out_709_pad_0, pad_type = lora_out_709_pad_type_0, strides = var_6872, weight = lora_out_711_weight_0_to_fp16, x = input_589_cast_fp16)[name = tensor("lora_out_711_cast_fp16")]; + tensor obj_119_cast_fp16 = add(x = pretrained_out_355_cast_fp16, y = lora_out_711_cast_fp16)[name = tensor("obj_119_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_119_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; + tensor var_6883 = const()[name = tensor("op_6883"), val = tensor([1])]; + tensor channels_mean_119_cast_fp16 = reduce_mean(axes = var_6883, keep_dims = var_6735, 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_6887 = const()[name = tensor("op_6887"), val = tensor([1])]; + tensor var_6888_cast_fp16 = reduce_mean(axes = var_6887, keep_dims = var_6735, x = zero_mean_sq_119_cast_fp16)[name = tensor("op_6888_cast_fp16")]; + tensor var_6889_to_fp16 = const()[name = tensor("op_6889_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6890_cast_fp16 = add(x = var_6888_cast_fp16, y = var_6889_to_fp16)[name = tensor("op_6890_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_6890_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(325719360)))]; + 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(325721984)))]; + 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_6904 = const()[name = tensor("op_6904"), val = tensor([1, 1])]; + tensor var_6906 = const()[name = tensor("op_6906"), 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_29_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325724608))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329001472))), 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(329001600)))]; + tensor pretrained_out_357_cast_fp16 = conv(bias = layers_29_fc1_pretrained_bias_to_fp16, dilations = var_6906, groups = var_6734, pad = pretrained_out_357_pad_0, pad_type = pretrained_out_357_pad_type_0, strides = var_6904, weight = layers_29_fc1_pretrained_weight_to_fp16_palettized, x = input_591_cast_fp16)[name = tensor("pretrained_out_357_cast_fp16")]; + tensor var_6910 = const()[name = tensor("op_6910"), val = tensor([1, 1])]; + tensor var_6912 = const()[name = tensor("op_6912"), 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_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(329011904)))]; + tensor input_593_cast_fp16 = conv(dilations = var_6912, groups = var_6734, pad = input_593_pad_0, pad_type = input_593_pad_type_0, strides = var_6910, weight = layers_29_fc1_loraA_weight_to_fp16, x = input_591_cast_fp16)[name = tensor("input_593_cast_fp16")]; + tensor var_6916 = const()[name = tensor("op_6916"), val = tensor([1, 1])]; + tensor var_6918 = const()[name = tensor("op_6918"), 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(329052928)))]; + tensor lora_out_715_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_6918, groups = var_6734, pad = lora_out_713_pad_0, pad_type = lora_out_713_pad_type_0, strides = var_6916, weight = lora_out_715_weight_0_to_fp16, x = input_593_cast_fp16)[name = tensor("lora_out_715_cast_fp16")]; + tensor input_595_cast_fp16 = add(x = pretrained_out_357_cast_fp16, y = lora_out_715_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_6930 = const()[name = tensor("op_6930"), val = tensor([1, 1])]; + tensor var_6932 = const()[name = tensor("op_6932"), 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_29_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(329216832))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332493696))), 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(332493824)))]; + tensor pretrained_out_359_cast_fp16 = conv(bias = layers_29_fc2_pretrained_bias_to_fp16, dilations = var_6932, groups = var_6734, pad = pretrained_out_359_pad_0, pad_type = pretrained_out_359_pad_type_0, strides = var_6930, weight = layers_29_fc2_pretrained_weight_to_fp16_palettized, x = input_597_cast_fp16)[name = tensor("pretrained_out_359_cast_fp16")]; + tensor var_6936 = const()[name = tensor("op_6936"), val = tensor([1, 1])]; + tensor var_6938 = const()[name = tensor("op_6938"), 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_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(332496448)))]; + tensor input_599_cast_fp16 = conv(dilations = var_6938, groups = var_6734, pad = input_599_pad_0, pad_type = input_599_pad_type_0, strides = var_6936, weight = layers_29_fc2_loraA_weight_to_fp16, x = input_597_cast_fp16)[name = tensor("input_599_cast_fp16")]; + tensor var_6942 = const()[name = tensor("op_6942"), val = tensor([1, 1])]; + tensor var_6944 = const()[name = tensor("op_6944"), 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(332660352)))]; + tensor lora_out_719_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_6944, groups = var_6734, pad = lora_out_717_pad_0, pad_type = lora_out_717_pad_type_0, strides = var_6942, weight = lora_out_719_weight_0_to_fp16, x = input_599_cast_fp16)[name = tensor("lora_out_719_cast_fp16")]; + tensor hidden_states_63_cast_fp16 = add(x = pretrained_out_359_cast_fp16, y = lora_out_719_cast_fp16)[name = tensor("hidden_states_63_cast_fp16")]; + tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_63_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_6958 = const()[name = tensor("op_6958"), val = tensor(3)]; + tensor var_6960 = const()[name = tensor("op_6960"), val = tensor(1)]; + tensor var_6961 = const()[name = tensor("op_6961"), val = tensor(true)]; + tensor var_6971 = const()[name = tensor("op_6971"), val = tensor([1])]; + tensor channels_mean_121_cast_fp16 = reduce_mean(axes = var_6971, keep_dims = var_6961, 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_6975 = const()[name = tensor("op_6975"), val = tensor([1])]; + tensor var_6976_cast_fp16 = reduce_mean(axes = var_6975, keep_dims = var_6961, x = zero_mean_sq_121_cast_fp16)[name = tensor("op_6976_cast_fp16")]; + tensor var_6977_to_fp16 = const()[name = tensor("op_6977_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_6978_cast_fp16 = add(x = var_6976_cast_fp16, y = var_6977_to_fp16)[name = tensor("op_6978_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_6978_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_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(332701376)))]; + 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(332704000)))]; + 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_121_cast_fp16)[name = tensor("obj_121_cast_fp16")]; + tensor var_6996 = const()[name = tensor("op_6996"), val = tensor([1, 1])]; + tensor var_6998 = const()[name = tensor("op_6998"), 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_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(332706624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333525888))), 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(333526016)))]; + tensor pretrained_out_361_cast_fp16 = conv(bias = layers_30_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_6998, groups = var_6960, pad = pretrained_out_361_pad_0, pad_type = pretrained_out_361_pad_type_0, strides = var_6996, weight = layers_30_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor("pretrained_out_361_cast_fp16")]; + tensor var_7002 = const()[name = tensor("op_7002"), val = tensor([1, 1])]; + tensor var_7004 = const()[name = tensor("op_7004"), 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_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(333528640)))]; + tensor input_601_cast_fp16 = conv(dilations = var_7004, groups = var_6960, pad = input_601_pad_0, pad_type = input_601_pad_type_0, strides = var_7002, weight = layers_30_self_attn_q_proj_loraA_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("input_601_cast_fp16")]; + tensor var_7008 = const()[name = tensor("op_7008"), val = tensor([1, 1])]; + tensor var_7010 = const()[name = tensor("op_7010"), 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(333569664)))]; + tensor lora_out_723_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7010, groups = var_6960, pad = lora_out_721_pad_0, pad_type = lora_out_721_pad_type_0, strides = var_7008, weight = lora_out_723_weight_0_to_fp16, x = input_601_cast_fp16)[name = tensor("lora_out_723_cast_fp16")]; + tensor query_61_cast_fp16 = add(x = pretrained_out_361_cast_fp16, y = lora_out_723_cast_fp16)[name = tensor("query_61_cast_fp16")]; + tensor var_7020 = const()[name = tensor("op_7020"), val = tensor([1, 1])]; + tensor var_7022 = const()[name = tensor("op_7022"), 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_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(333610688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334429952))), name = tensor("layers_30_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_363_cast_fp16 = conv(dilations = var_7022, groups = var_6960, pad = pretrained_out_363_pad_0, pad_type = pretrained_out_363_pad_type_0, strides = var_7020, weight = layers_30_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor("pretrained_out_363_cast_fp16")]; + tensor var_7026 = const()[name = tensor("op_7026"), val = tensor([1, 1])]; + tensor var_7028 = const()[name = tensor("op_7028"), 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_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(334430080)))]; + tensor input_603_cast_fp16 = conv(dilations = var_7028, groups = var_6960, pad = input_603_pad_0, pad_type = input_603_pad_type_0, strides = var_7026, weight = layers_30_self_attn_k_proj_loraA_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("input_603_cast_fp16")]; + tensor var_7032 = const()[name = tensor("op_7032"), val = tensor([1, 1])]; + tensor var_7034 = const()[name = tensor("op_7034"), 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(334471104)))]; + tensor lora_out_727_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7034, groups = var_6960, pad = lora_out_725_pad_0, pad_type = lora_out_725_pad_type_0, strides = var_7032, weight = lora_out_727_weight_0_to_fp16, x = input_603_cast_fp16)[name = tensor("lora_out_727_cast_fp16")]; + tensor key_61_cast_fp16 = add(x = pretrained_out_363_cast_fp16, y = lora_out_727_cast_fp16)[name = tensor("key_61_cast_fp16")]; + tensor var_7045 = const()[name = tensor("op_7045"), val = tensor([1, 1])]; + tensor var_7047 = const()[name = tensor("op_7047"), 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_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(334512128))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335331392))), 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(335331520)))]; + tensor pretrained_out_365_cast_fp16 = conv(bias = layers_30_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_7047, groups = var_6960, pad = pretrained_out_365_pad_0, pad_type = pretrained_out_365_pad_type_0, strides = var_7045, weight = layers_30_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_121_cast_fp16)[name = tensor("pretrained_out_365_cast_fp16")]; + tensor var_7051 = const()[name = tensor("op_7051"), val = tensor([1, 1])]; + tensor var_7053 = const()[name = tensor("op_7053"), 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_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(335334144)))]; + tensor input_605_cast_fp16 = conv(dilations = var_7053, groups = var_6960, pad = input_605_pad_0, pad_type = input_605_pad_type_0, strides = var_7051, weight = layers_30_self_attn_v_proj_loraA_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("input_605_cast_fp16")]; + tensor var_7057 = const()[name = tensor("op_7057"), val = tensor([1, 1])]; + tensor var_7059 = const()[name = tensor("op_7059"), 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(335375168)))]; + tensor lora_out_731_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7059, groups = var_6960, pad = lora_out_729_pad_0, pad_type = lora_out_729_pad_type_0, strides = var_7057, weight = lora_out_731_weight_0_to_fp16, x = input_605_cast_fp16)[name = tensor("lora_out_731_cast_fp16")]; + tensor value_61_cast_fp16 = add(x = pretrained_out_365_cast_fp16, y = lora_out_731_cast_fp16)[name = tensor("value_61_cast_fp16")]; + tensor var_7066 = const()[name = tensor("op_7066"), val = tensor([1, 20, 64, -1])]; + tensor var_7067_cast_fp16 = reshape(shape = var_7066, x = query_61_cast_fp16)[name = tensor("op_7067_cast_fp16")]; + tensor var_7068_to_fp16 = const()[name = tensor("op_7068_to_fp16"), val = tensor(0x1p-3)]; + tensor var_7069_cast_fp16 = mul(x = var_7067_cast_fp16, y = var_7068_to_fp16)[name = tensor("op_7069_cast_fp16")]; + tensor var_7070 = const()[name = tensor("op_7070"), val = tensor([1, 20, 64, -1])]; + tensor var_7071_cast_fp16 = reshape(shape = var_7070, x = key_61_cast_fp16)[name = tensor("op_7071_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_7069_cast_fp16, y = var_7071_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; + tensor var_7074_cast_fp16 = softmax(axis = var_6958, x = mh_w_61_cast_fp16)[name = tensor("op_7074_cast_fp16")]; + tensor var_7075 = const()[name = tensor("op_7075"), val = tensor([1, 20, 64, -1])]; + tensor var_7076_cast_fp16 = reshape(shape = var_7075, x = value_61_cast_fp16)[name = tensor("op_7076_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_7076_cast_fp16, y = var_7074_cast_fp16)[name = tensor("attn_61_cast_fp16")]; + tensor var_7079 = const()[name = tensor("op_7079"), val = tensor([1, 1280, 1, -1])]; + tensor input_607_cast_fp16 = reshape(shape = var_7079, x = attn_61_cast_fp16)[name = tensor("input_607_cast_fp16")]; + tensor var_7086 = const()[name = tensor("op_7086"), val = tensor([1, 1])]; + tensor var_7088 = const()[name = tensor("op_7088"), 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_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(335416192))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336235456))), 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(336235584)))]; + tensor pretrained_out_367_cast_fp16 = conv(bias = layers_30_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7088, groups = var_6960, pad = pretrained_out_367_pad_0, pad_type = pretrained_out_367_pad_type_0, strides = var_7086, weight = layers_30_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_607_cast_fp16)[name = tensor("pretrained_out_367_cast_fp16")]; + tensor var_7092 = const()[name = tensor("op_7092"), val = tensor([1, 1])]; + tensor var_7094 = const()[name = tensor("op_7094"), 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_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(336238208)))]; + tensor input_609_cast_fp16 = conv(dilations = var_7094, groups = var_6960, pad = input_609_pad_0, pad_type = input_609_pad_type_0, strides = var_7092, weight = layers_30_self_attn_o_proj_loraA_weight_to_fp16, x = input_607_cast_fp16)[name = tensor("input_609_cast_fp16")]; + tensor var_7098 = const()[name = tensor("op_7098"), val = tensor([1, 1])]; + tensor var_7100 = const()[name = tensor("op_7100"), 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(336279232)))]; + tensor lora_out_735_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7100, groups = var_6960, pad = lora_out_733_pad_0, pad_type = lora_out_733_pad_type_0, strides = var_7098, weight = lora_out_735_weight_0_to_fp16, x = input_609_cast_fp16)[name = tensor("lora_out_735_cast_fp16")]; + tensor obj_123_cast_fp16 = add(x = pretrained_out_367_cast_fp16, y = lora_out_735_cast_fp16)[name = tensor("obj_123_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_123_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; + tensor var_7109 = const()[name = tensor("op_7109"), val = tensor([1])]; + tensor channels_mean_123_cast_fp16 = reduce_mean(axes = var_7109, keep_dims = var_6961, 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_7113 = const()[name = tensor("op_7113"), val = tensor([1])]; + tensor var_7114_cast_fp16 = reduce_mean(axes = var_7113, keep_dims = var_6961, x = zero_mean_sq_123_cast_fp16)[name = tensor("op_7114_cast_fp16")]; + tensor var_7115_to_fp16 = const()[name = tensor("op_7115_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7116_cast_fp16 = add(x = var_7114_cast_fp16, y = var_7115_to_fp16)[name = tensor("op_7116_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_7116_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 input_611_gamma_0_to_fp16 = const()[name = tensor("input_611_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336320256)))]; + tensor input_611_beta_0_to_fp16 = const()[name = tensor("input_611_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336322880)))]; + tensor input_611_epsilon_0_to_fp16 = const()[name = tensor("input_611_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_611_cast_fp16 = batch_norm(beta = input_611_beta_0_to_fp16, epsilon = input_611_epsilon_0_to_fp16, gamma = input_611_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("input_611_cast_fp16")]; + tensor var_7130 = const()[name = tensor("op_7130"), val = tensor([1, 1])]; + tensor var_7132 = const()[name = tensor("op_7132"), 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_30_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336325504))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339602368))), 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(339602496)))]; + tensor pretrained_out_369_cast_fp16 = conv(bias = layers_30_fc1_pretrained_bias_to_fp16, dilations = var_7132, groups = var_6960, pad = pretrained_out_369_pad_0, pad_type = pretrained_out_369_pad_type_0, strides = var_7130, weight = layers_30_fc1_pretrained_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = tensor("pretrained_out_369_cast_fp16")]; + tensor var_7136 = const()[name = tensor("op_7136"), val = tensor([1, 1])]; + tensor var_7138 = const()[name = tensor("op_7138"), 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_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(339612800)))]; + tensor input_613_cast_fp16 = conv(dilations = var_7138, groups = var_6960, pad = input_613_pad_0, pad_type = input_613_pad_type_0, strides = var_7136, weight = layers_30_fc1_loraA_weight_to_fp16, x = input_611_cast_fp16)[name = tensor("input_613_cast_fp16")]; + tensor var_7142 = const()[name = tensor("op_7142"), val = tensor([1, 1])]; + tensor var_7144 = const()[name = tensor("op_7144"), 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(339653824)))]; + tensor lora_out_739_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_7144, groups = var_6960, pad = lora_out_737_pad_0, pad_type = lora_out_737_pad_type_0, strides = var_7142, weight = lora_out_739_weight_0_to_fp16, x = input_613_cast_fp16)[name = tensor("lora_out_739_cast_fp16")]; + tensor input_615_cast_fp16 = add(x = pretrained_out_369_cast_fp16, y = lora_out_739_cast_fp16)[name = tensor("input_615_cast_fp16")]; + tensor input_617_mode_0 = const()[name = tensor("input_617_mode_0"), val = tensor("EXACT")]; + tensor input_617_cast_fp16 = gelu(mode = input_617_mode_0, x = input_615_cast_fp16)[name = tensor("input_617_cast_fp16")]; + tensor var_7156 = const()[name = tensor("op_7156"), val = tensor([1, 1])]; + tensor var_7158 = const()[name = tensor("op_7158"), 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_30_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339817728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343094592))), 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(343094720)))]; + tensor pretrained_out_371_cast_fp16 = conv(bias = layers_30_fc2_pretrained_bias_to_fp16, dilations = var_7158, groups = var_6960, pad = pretrained_out_371_pad_0, pad_type = pretrained_out_371_pad_type_0, strides = var_7156, weight = layers_30_fc2_pretrained_weight_to_fp16_palettized, x = input_617_cast_fp16)[name = tensor("pretrained_out_371_cast_fp16")]; + tensor var_7162 = const()[name = tensor("op_7162"), val = tensor([1, 1])]; + tensor var_7164 = const()[name = tensor("op_7164"), 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_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(343097344)))]; + tensor input_619_cast_fp16 = conv(dilations = var_7164, groups = var_6960, pad = input_619_pad_0, pad_type = input_619_pad_type_0, strides = var_7162, weight = layers_30_fc2_loraA_weight_to_fp16, x = input_617_cast_fp16)[name = tensor("input_619_cast_fp16")]; + tensor var_7168 = const()[name = tensor("op_7168"), val = tensor([1, 1])]; + tensor var_7170 = const()[name = tensor("op_7170"), 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(343261248)))]; + tensor lora_out_743_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7170, groups = var_6960, pad = lora_out_741_pad_0, pad_type = lora_out_741_pad_type_0, strides = var_7168, weight = lora_out_743_weight_0_to_fp16, x = input_619_cast_fp16)[name = tensor("lora_out_743_cast_fp16")]; + tensor hidden_states_65_cast_fp16 = add(x = pretrained_out_371_cast_fp16, y = lora_out_743_cast_fp16)[name = tensor("hidden_states_65_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = hidden_states_65_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; + tensor var_7184 = const()[name = tensor("op_7184"), val = tensor(3)]; + tensor var_7186 = const()[name = tensor("op_7186"), val = tensor(1)]; + tensor var_7187 = const()[name = tensor("op_7187"), val = tensor(true)]; + tensor var_7197 = const()[name = tensor("op_7197"), val = tensor([1])]; + tensor channels_mean_125_cast_fp16 = reduce_mean(axes = var_7197, keep_dims = var_7187, 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_7201 = const()[name = tensor("op_7201"), val = tensor([1])]; + tensor var_7202_cast_fp16 = reduce_mean(axes = var_7201, keep_dims = var_7187, x = zero_mean_sq_125_cast_fp16)[name = tensor("op_7202_cast_fp16")]; + tensor var_7203_to_fp16 = const()[name = tensor("op_7203_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7204_cast_fp16 = add(x = var_7202_cast_fp16, y = var_7203_to_fp16)[name = tensor("op_7204_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_7204_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 obj_125_gamma_0_to_fp16 = const()[name = tensor("obj_125_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343302272)))]; + tensor obj_125_beta_0_to_fp16 = const()[name = tensor("obj_125_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343304896)))]; + tensor obj_125_epsilon_0_to_fp16 = const()[name = tensor("obj_125_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_125_cast_fp16 = batch_norm(beta = obj_125_beta_0_to_fp16, epsilon = obj_125_epsilon_0_to_fp16, gamma = obj_125_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("obj_125_cast_fp16")]; + tensor var_7222 = const()[name = tensor("op_7222"), val = tensor([1, 1])]; + tensor var_7224 = const()[name = tensor("op_7224"), 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_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(343307520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344126784))), 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(344126912)))]; + tensor pretrained_out_373_cast_fp16 = conv(bias = layers_31_self_attn_q_proj_pretrained_bias_to_fp16, dilations = var_7224, groups = var_7186, pad = pretrained_out_373_pad_0, pad_type = pretrained_out_373_pad_type_0, strides = var_7222, weight = layers_31_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_125_cast_fp16)[name = tensor("pretrained_out_373_cast_fp16")]; + tensor var_7228 = const()[name = tensor("op_7228"), val = tensor([1, 1])]; + tensor var_7230 = const()[name = tensor("op_7230"), val = tensor([1, 1])]; + tensor input_621_pad_type_0 = const()[name = tensor("input_621_pad_type_0"), val = tensor("custom")]; + tensor input_621_pad_0 = const()[name = tensor("input_621_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(344129536)))]; + tensor input_621_cast_fp16 = conv(dilations = var_7230, groups = var_7186, pad = input_621_pad_0, pad_type = input_621_pad_type_0, strides = var_7228, weight = layers_31_self_attn_q_proj_loraA_weight_to_fp16, x = obj_125_cast_fp16)[name = tensor("input_621_cast_fp16")]; + tensor var_7234 = const()[name = tensor("op_7234"), val = tensor([1, 1])]; + tensor var_7236 = const()[name = tensor("op_7236"), 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(344170560)))]; + tensor lora_out_747_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7236, groups = var_7186, pad = lora_out_745_pad_0, pad_type = lora_out_745_pad_type_0, strides = var_7234, weight = lora_out_747_weight_0_to_fp16, x = input_621_cast_fp16)[name = tensor("lora_out_747_cast_fp16")]; + tensor query_cast_fp16 = add(x = pretrained_out_373_cast_fp16, y = lora_out_747_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor var_7246 = const()[name = tensor("op_7246"), val = tensor([1, 1])]; + tensor var_7248 = const()[name = tensor("op_7248"), 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_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(344211584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345030848))), name = tensor("layers_31_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; + tensor pretrained_out_375_cast_fp16 = conv(dilations = var_7248, groups = var_7186, pad = pretrained_out_375_pad_0, pad_type = pretrained_out_375_pad_type_0, strides = var_7246, weight = layers_31_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_125_cast_fp16)[name = tensor("pretrained_out_375_cast_fp16")]; + tensor var_7252 = const()[name = tensor("op_7252"), val = tensor([1, 1])]; + tensor var_7254 = const()[name = tensor("op_7254"), 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_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(345030976)))]; + tensor input_623_cast_fp16 = conv(dilations = var_7254, groups = var_7186, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = var_7252, weight = layers_31_self_attn_k_proj_loraA_weight_to_fp16, x = obj_125_cast_fp16)[name = tensor("input_623_cast_fp16")]; + tensor var_7258 = const()[name = tensor("op_7258"), val = tensor([1, 1])]; + tensor var_7260 = const()[name = tensor("op_7260"), 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(345072000)))]; + tensor lora_out_751_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7260, groups = var_7186, pad = lora_out_749_pad_0, pad_type = lora_out_749_pad_type_0, strides = var_7258, weight = lora_out_751_weight_0_to_fp16, x = input_623_cast_fp16)[name = tensor("lora_out_751_cast_fp16")]; + tensor key_cast_fp16 = add(x = pretrained_out_375_cast_fp16, y = lora_out_751_cast_fp16)[name = tensor("key_cast_fp16")]; + tensor var_7271 = const()[name = tensor("op_7271"), val = tensor([1, 1])]; + tensor var_7273 = const()[name = tensor("op_7273"), 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_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(345113024))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345932288))), 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(345932416)))]; + tensor pretrained_out_377_cast_fp16 = conv(bias = layers_31_self_attn_v_proj_pretrained_bias_to_fp16, dilations = var_7273, groups = var_7186, pad = pretrained_out_377_pad_0, pad_type = pretrained_out_377_pad_type_0, strides = var_7271, weight = layers_31_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_125_cast_fp16)[name = tensor("pretrained_out_377_cast_fp16")]; + tensor var_7277 = const()[name = tensor("op_7277"), val = tensor([1, 1])]; + tensor var_7279 = const()[name = tensor("op_7279"), val = tensor([1, 1])]; + tensor input_625_pad_type_0 = const()[name = tensor("input_625_pad_type_0"), val = tensor("custom")]; + tensor input_625_pad_0 = const()[name = tensor("input_625_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(345935040)))]; + tensor input_625_cast_fp16 = conv(dilations = var_7279, groups = var_7186, pad = input_625_pad_0, pad_type = input_625_pad_type_0, strides = var_7277, weight = layers_31_self_attn_v_proj_loraA_weight_to_fp16, x = obj_125_cast_fp16)[name = tensor("input_625_cast_fp16")]; + tensor var_7283 = const()[name = tensor("op_7283"), val = tensor([1, 1])]; + tensor var_7285 = const()[name = tensor("op_7285"), 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(345976064)))]; + tensor lora_out_755_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7285, groups = var_7186, pad = lora_out_753_pad_0, pad_type = lora_out_753_pad_type_0, strides = var_7283, weight = lora_out_755_weight_0_to_fp16, x = input_625_cast_fp16)[name = tensor("lora_out_755_cast_fp16")]; + tensor value_cast_fp16 = add(x = pretrained_out_377_cast_fp16, y = lora_out_755_cast_fp16)[name = tensor("value_cast_fp16")]; + tensor var_7292 = const()[name = tensor("op_7292"), val = tensor([1, 20, 64, -1])]; + tensor var_7293_cast_fp16 = reshape(shape = var_7292, x = query_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor var_7294_to_fp16 = const()[name = tensor("op_7294_to_fp16"), val = tensor(0x1p-3)]; + tensor var_7295_cast_fp16 = mul(x = var_7293_cast_fp16, y = var_7294_to_fp16)[name = tensor("op_7295_cast_fp16")]; + tensor var_7296 = const()[name = tensor("op_7296"), val = tensor([1, 20, 64, -1])]; + tensor var_7297_cast_fp16 = reshape(shape = var_7296, x = key_cast_fp16)[name = tensor("op_7297_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_7295_cast_fp16, y = var_7297_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor var_7300_cast_fp16 = softmax(axis = var_7184, x = mh_w_cast_fp16)[name = tensor("op_7300_cast_fp16")]; + tensor var_7301 = const()[name = tensor("op_7301"), val = tensor([1, 20, 64, -1])]; + tensor var_7302_cast_fp16 = reshape(shape = var_7301, x = value_cast_fp16)[name = tensor("op_7302_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_7302_cast_fp16, y = var_7300_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_7305 = const()[name = tensor("op_7305"), val = tensor([1, 1280, 1, -1])]; + tensor input_627_cast_fp16 = reshape(shape = var_7305, x = attn_cast_fp16)[name = tensor("input_627_cast_fp16")]; + tensor var_7312 = const()[name = tensor("op_7312"), val = tensor([1, 1])]; + tensor var_7314 = const()[name = tensor("op_7314"), 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_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(346017088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(346836352))), 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(346836480)))]; + tensor pretrained_out_379_cast_fp16 = conv(bias = layers_31_self_attn_o_proj_pretrained_bias_to_fp16, dilations = var_7314, groups = var_7186, pad = pretrained_out_379_pad_0, pad_type = pretrained_out_379_pad_type_0, strides = var_7312, weight = layers_31_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_627_cast_fp16)[name = tensor("pretrained_out_379_cast_fp16")]; + tensor var_7318 = const()[name = tensor("op_7318"), val = tensor([1, 1])]; + tensor var_7320 = const()[name = tensor("op_7320"), 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_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(346839104)))]; + tensor input_629_cast_fp16 = conv(dilations = var_7320, groups = var_7186, pad = input_629_pad_0, pad_type = input_629_pad_type_0, strides = var_7318, weight = layers_31_self_attn_o_proj_loraA_weight_to_fp16, x = input_627_cast_fp16)[name = tensor("input_629_cast_fp16")]; + tensor var_7324 = const()[name = tensor("op_7324"), val = tensor([1, 1])]; + tensor var_7326 = const()[name = tensor("op_7326"), 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(346880128)))]; + tensor lora_out_759_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7326, groups = var_7186, pad = lora_out_757_pad_0, pad_type = lora_out_757_pad_type_0, strides = var_7324, weight = lora_out_759_weight_0_to_fp16, x = input_629_cast_fp16)[name = tensor("lora_out_759_cast_fp16")]; + tensor obj_cast_fp16 = add(x = pretrained_out_379_cast_fp16, y = lora_out_759_cast_fp16)[name = tensor("obj_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = obj_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; + tensor var_7335 = const()[name = tensor("op_7335"), val = tensor([1])]; + tensor channels_mean_127_cast_fp16 = reduce_mean(axes = var_7335, keep_dims = var_7187, 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_7339 = const()[name = tensor("op_7339"), val = tensor([1])]; + tensor var_7340_cast_fp16 = reduce_mean(axes = var_7339, keep_dims = var_7187, x = zero_mean_sq_127_cast_fp16)[name = tensor("op_7340_cast_fp16")]; + tensor var_7341_to_fp16 = const()[name = tensor("op_7341_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7342_cast_fp16 = add(x = var_7340_cast_fp16, y = var_7341_to_fp16)[name = tensor("op_7342_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_7342_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 input_631_gamma_0_to_fp16 = const()[name = tensor("input_631_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(346921152)))]; + tensor input_631_beta_0_to_fp16 = const()[name = tensor("input_631_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(346923776)))]; + tensor input_631_epsilon_0_to_fp16 = const()[name = tensor("input_631_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_631_cast_fp16 = batch_norm(beta = input_631_beta_0_to_fp16, epsilon = input_631_epsilon_0_to_fp16, gamma = input_631_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("input_631_cast_fp16")]; + tensor var_7356 = const()[name = tensor("op_7356"), val = tensor([1, 1])]; + tensor var_7358 = const()[name = tensor("op_7358"), 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_31_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(346926400))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350203264))), 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(350203392)))]; + tensor pretrained_out_381_cast_fp16 = conv(bias = layers_31_fc1_pretrained_bias_to_fp16, dilations = var_7358, groups = var_7186, pad = pretrained_out_381_pad_0, pad_type = pretrained_out_381_pad_type_0, strides = var_7356, weight = layers_31_fc1_pretrained_weight_to_fp16_palettized, x = input_631_cast_fp16)[name = tensor("pretrained_out_381_cast_fp16")]; + tensor var_7362 = const()[name = tensor("op_7362"), val = tensor([1, 1])]; + tensor var_7364 = const()[name = tensor("op_7364"), 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_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(350213696)))]; + tensor input_633_cast_fp16 = conv(dilations = var_7364, groups = var_7186, pad = input_633_pad_0, pad_type = input_633_pad_type_0, strides = var_7362, weight = layers_31_fc1_loraA_weight_to_fp16, x = input_631_cast_fp16)[name = tensor("input_633_cast_fp16")]; + tensor var_7368 = const()[name = tensor("op_7368"), val = tensor([1, 1])]; + tensor var_7370 = const()[name = tensor("op_7370"), 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(350254720)))]; + tensor lora_out_763_cast_fp16 = conv(bias = lora_out_19_bias_0_to_fp16, dilations = var_7370, groups = var_7186, pad = lora_out_761_pad_0, pad_type = lora_out_761_pad_type_0, strides = var_7368, weight = lora_out_763_weight_0_to_fp16, x = input_633_cast_fp16)[name = tensor("lora_out_763_cast_fp16")]; + tensor input_635_cast_fp16 = add(x = pretrained_out_381_cast_fp16, y = lora_out_763_cast_fp16)[name = tensor("input_635_cast_fp16")]; + tensor input_637_mode_0 = const()[name = tensor("input_637_mode_0"), val = tensor("EXACT")]; + tensor input_637_cast_fp16 = gelu(mode = input_637_mode_0, x = input_635_cast_fp16)[name = tensor("input_637_cast_fp16")]; + tensor var_7382 = const()[name = tensor("op_7382"), val = tensor([1, 1])]; + tensor var_7384 = const()[name = tensor("op_7384"), 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(350418624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353695488))), 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(353695616)))]; + tensor pretrained_out_cast_fp16 = conv(bias = layers_31_fc2_pretrained_bias_to_fp16, dilations = var_7384, groups = var_7186, pad = pretrained_out_pad_0, pad_type = pretrained_out_pad_type_0, strides = var_7382, weight = layers_31_fc2_pretrained_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = tensor("pretrained_out_cast_fp16")]; + tensor var_7388 = const()[name = tensor("op_7388"), val = tensor([1, 1])]; + tensor var_7390 = const()[name = tensor("op_7390"), 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(353698240)))]; + tensor input_cast_fp16 = conv(dilations = var_7390, groups = var_7186, pad = input_pad_0, pad_type = input_pad_type_0, strides = var_7388, weight = layers_31_fc2_loraA_weight_to_fp16, x = input_637_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor var_7394 = const()[name = tensor("op_7394"), val = tensor([1, 1])]; + tensor var_7396 = const()[name = tensor("op_7396"), 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_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(353862144)))]; + tensor lora_out_cast_fp16 = conv(bias = obj_1_mean_0_to_fp16, dilations = var_7396, groups = var_7186, pad = lora_out_765_pad_0, pad_type = lora_out_765_pad_type_0, strides = var_7394, weight = lora_out_weight_0_to_fp16, x = input_cast_fp16)[name = tensor("lora_out_cast_fp16")]; + tensor hidden_states_cast_fp16 = add(x = pretrained_out_cast_fp16, y = lora_out_cast_fp16)[name = tensor("hidden_states_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_127_cast_fp16, y = hidden_states_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor var_7405 = const()[name = tensor("op_7405"), val = tensor(true)]; + tensor var_7409 = const()[name = tensor("op_7409"), val = tensor([1])]; + tensor channels_mean_cast_fp16 = reduce_mean(axes = var_7409, keep_dims = var_7405, 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_7413 = const()[name = tensor("op_7413"), val = tensor([1])]; + tensor var_7414_cast_fp16 = reduce_mean(axes = var_7413, keep_dims = var_7405, x = zero_mean_sq_cast_fp16)[name = tensor("op_7414_cast_fp16")]; + tensor var_7415_to_fp16 = const()[name = tensor("op_7415_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_7416_cast_fp16 = add(x = var_7414_cast_fp16, y = var_7415_to_fp16)[name = tensor("op_7416_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_7416_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 encoder_output_embeds_type_fp32_gamma_0_to_fp16 = const()[name = tensor("encoder_output_embeds_type_fp32_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353903168)))]; + tensor encoder_output_embeds_type_fp32_beta_0_to_fp16 = const()[name = tensor("encoder_output_embeds_type_fp32_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353905792)))]; + tensor encoder_output_embeds_type_fp32_epsilon_0_to_fp16 = const()[name = tensor("encoder_output_embeds_type_fp32_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor encoder_output_embeds = batch_norm(beta = encoder_output_embeds_type_fp32_beta_0_to_fp16, epsilon = encoder_output_embeds_type_fp32_epsilon_0_to_fp16, gamma = encoder_output_embeds_type_fp32_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor("encoder_output_embeds_type_fp32_cast_fp16")]; + } -> (encoder_output_embeds); +} \ No newline at end of file