program(1.0) [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3304.5.2"}, {"coremlc-version", "3304.6.2"}})] { func main<ios16>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 448]> decoder_key_padding_mask, tensor<fp16, [1, 1280, 1, 1500]> encoder_output_embeds, tensor<int32, [1]> input_ids, tensor<fp16, [1, 5120, 1, 448]> key_cache, tensor<fp16, [1, 448]> kv_cache_update_mask, tensor<fp16, [1, 5120, 1, 448]> value_cache) { tensor<int32, []> var_24_axis_0 = const()[name = tensor<string, []>("op_24_axis_0"), val = tensor<int32, []>(0)]; tensor<int32, []> var_24_batch_dims_0 = const()[name = tensor<string, []>("op_24_batch_dims_0"), val = tensor<int32, []>(0)]; tensor<fp16, [51866, 1280]> embed_tokens_weight_to_fp16 = const()[name = tensor<string, []>("embed_tokens_weight_to_fp16"), val = tensor<fp16, [51866, 1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))]; tensor<fp16, [1, 1280]> var_24_cast_fp16 = gather(axis = var_24_axis_0, batch_dims = var_24_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor<string, []>("op_24_cast_fp16")]; tensor<int32, []> var_28_axis_0 = const()[name = tensor<string, []>("op_28_axis_0"), val = tensor<int32, []>(0)]; tensor<int32, []> var_28_batch_dims_0 = const()[name = tensor<string, []>("op_28_batch_dims_0"), val = tensor<int32, []>(0)]; tensor<fp16, [448, 1280]> embed_positions_weight_to_fp16 = const()[name = tensor<string, []>("embed_positions_weight_to_fp16"), val = tensor<fp16, [448, 1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132777088)))]; tensor<fp16, [1, 1280]> var_28_cast_fp16 = gather(axis = var_28_axis_0, batch_dims = var_28_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor<string, []>("op_28_cast_fp16")]; tensor<fp16, [1, 1280]> hidden_states_1_cast_fp16 = add(x = var_24_cast_fp16, y = var_28_cast_fp16)[name = tensor<string, []>("hidden_states_1_cast_fp16")]; tensor<int32, [1]> var_42_axes_0 = const()[name = tensor<string, []>("op_42_axes_0"), val = tensor<int32, [1]>([2])]; tensor<fp16, [1, 1280, 1]> var_42_cast_fp16 = expand_dims(axes = var_42_axes_0, x = hidden_states_1_cast_fp16)[name = tensor<string, []>("op_42_cast_fp16")]; tensor<int32, [1]> inputs_1_axes_0 = const()[name = tensor<string, []>("inputs_1_axes_0"), val = tensor<int32, [1]>([3])]; tensor<fp16, [1, 1280, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_42_cast_fp16)[name = tensor<string, []>("inputs_1_cast_fp16")]; tensor<int32, [4]> tile_0 = const()[name = tensor<string, []>("tile_0"), val = tensor<int32, [4]>([1280, 1280, 1280, 1280])]; tensor<int32, []> var_47_axis_0 = const()[name = tensor<string, []>("op_47_axis_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1, 1280, 1, 448]> var_47_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_47_cast_fp16_1, tensor<fp16, [1, 1280, 1, 448]> var_47_cast_fp16_2, tensor<fp16, [1, 1280, 1, 448]> var_47_cast_fp16_3 = split(axis = var_47_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor<string, []>("op_47_cast_fp16")]; tensor<int32, [4]> tile_1 = const()[name = tensor<string, []>("tile_1"), val = tensor<int32, [4]>([1280, 1280, 1280, 1280])]; tensor<int32, []> var_54_axis_0 = const()[name = tensor<string, []>("op_54_axis_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1, 1280, 1, 448]> var_54_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_54_cast_fp16_1, tensor<fp16, [1, 1280, 1, 448]> var_54_cast_fp16_2, tensor<fp16, [1, 1280, 1, 448]> var_54_cast_fp16_3 = split(axis = var_54_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor<string, []>("op_54_cast_fp16")]; tensor<int32, []> var_64 = const()[name = tensor<string, []>("op_64"), val = tensor<int32, []>(3)]; tensor<int32, [1]> out_1_axes_0 = const()[name = tensor<string, []>("out_1_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_90_to_fp16 = const()[name = tensor<string, []>("op_90_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_90_to_fp16, x = inputs_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")]; tensor<fp16, [1280]> obj_1_mean_0_to_fp16 = const()[name = tensor<string, []>("obj_1_mean_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133924032)))]; tensor<fp16, [1280]> obj_1_variance_0_to_fp16 = const()[name = tensor<string, []>("obj_1_variance_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133926656)))]; tensor<fp16, [1280]> obj_1_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_1_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133929280)))]; tensor<fp16, [1280]> obj_1_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_1_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133931904)))]; tensor<fp16, []> obj_1_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_1_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> 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<string, []>("obj_1_cast_fp16")]; tensor<string, []> pretrained_out_1_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_1_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_1_strides_0 = const()[name = tensor<string, []>("pretrained_out_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_1_pad_0 = const()[name = tensor<string, []>("pretrained_out_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_1_dilations_0 = const()[name = tensor<string, []>("pretrained_out_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_1_groups_0 = const()[name = tensor<string, []>("pretrained_out_1_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(133934528))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134753792))), name = tensor<string, []>("layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_0_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134753920)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_1_dilations_0, groups = pretrained_out_1_groups_0, pad = pretrained_out_1_pad_0, pad_type = pretrained_out_1_pad_type_0, strides = pretrained_out_1_strides_0, weight = layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("pretrained_out_1_cast_fp16")]; tensor<string, []> input_1_pad_type_0 = const()[name = tensor<string, []>("input_1_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_1_strides_0 = const()[name = tensor<string, []>("input_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_1_pad_0 = const()[name = tensor<string, []>("input_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_1_dilations_0 = const()[name = tensor<string, []>("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_1_groups_0 = const()[name = tensor<string, []>("input_1_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_0_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134756544)))]; tensor<fp16, [1, 16, 1, 1]> input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_self_attn_q_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")]; tensor<string, []> lora_out_1_pad_type_0 = const()[name = tensor<string, []>("lora_out_1_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_1_strides_0 = const()[name = tensor<string, []>("lora_out_1_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_1_pad_0 = const()[name = tensor<string, []>("lora_out_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_1_dilations_0 = const()[name = tensor<string, []>("lora_out_1_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_1_groups_0 = const()[name = tensor<string, []>("lora_out_1_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_0_self_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134797568)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_1_cast_fp16 = conv(dilations = lora_out_1_dilations_0, groups = lora_out_1_groups_0, pad = lora_out_1_pad_0, pad_type = lora_out_1_pad_type_0, strides = lora_out_1_strides_0, weight = layers_0_self_attn_q_proj_loraB_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("lora_out_1_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> query_1_cast_fp16 = add(x = pretrained_out_1_cast_fp16, y = lora_out_1_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")]; tensor<string, []> pretrained_out_3_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_3_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_3_strides_0 = const()[name = tensor<string, []>("pretrained_out_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_3_pad_0 = const()[name = tensor<string, []>("pretrained_out_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_3_dilations_0 = const()[name = tensor<string, []>("pretrained_out_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_3_groups_0 = const()[name = tensor<string, []>("pretrained_out_3_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134838592))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135657856))), name = tensor<string, []>("layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_3_cast_fp16 = conv(dilations = pretrained_out_3_dilations_0, groups = pretrained_out_3_groups_0, pad = pretrained_out_3_pad_0, pad_type = pretrained_out_3_pad_type_0, strides = pretrained_out_3_strides_0, weight = layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("pretrained_out_3_cast_fp16")]; tensor<string, []> input_3_pad_type_0 = const()[name = tensor<string, []>("input_3_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_3_strides_0 = const()[name = tensor<string, []>("input_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_3_dilations_0 = const()[name = tensor<string, []>("input_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_3_groups_0 = const()[name = tensor<string, []>("input_3_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_0_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135657984)))]; tensor<fp16, [1, 16, 1, 1]> input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_0_self_attn_k_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")]; tensor<string, []> lora_out_3_pad_type_0 = const()[name = tensor<string, []>("lora_out_3_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_3_strides_0 = const()[name = tensor<string, []>("lora_out_3_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_3_pad_0 = const()[name = tensor<string, []>("lora_out_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_3_dilations_0 = const()[name = tensor<string, []>("lora_out_3_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_3_groups_0 = const()[name = tensor<string, []>("lora_out_3_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_0_self_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135699008)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_3_cast_fp16 = conv(dilations = lora_out_3_dilations_0, groups = lora_out_3_groups_0, pad = lora_out_3_pad_0, pad_type = lora_out_3_pad_type_0, strides = lora_out_3_strides_0, weight = layers_0_self_attn_k_proj_loraB_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("lora_out_3_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> current_key_1_cast_fp16 = add(x = pretrained_out_3_cast_fp16, y = lora_out_3_cast_fp16)[name = tensor<string, []>("current_key_1_cast_fp16")]; tensor<string, []> pretrained_out_5_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_5_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_5_strides_0 = const()[name = tensor<string, []>("pretrained_out_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_5_pad_0 = const()[name = tensor<string, []>("pretrained_out_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_5_dilations_0 = const()[name = tensor<string, []>("pretrained_out_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_5_groups_0 = const()[name = tensor<string, []>("pretrained_out_5_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135740032))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136559296))), name = tensor<string, []>("layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_0_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136559424)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_5_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_5_dilations_0, groups = pretrained_out_5_groups_0, pad = pretrained_out_5_pad_0, pad_type = pretrained_out_5_pad_type_0, strides = pretrained_out_5_strides_0, weight = layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor<string, []>("pretrained_out_5_cast_fp16")]; tensor<string, []> input_5_pad_type_0 = const()[name = tensor<string, []>("input_5_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_5_strides_0 = const()[name = tensor<string, []>("input_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_5_dilations_0 = const()[name = tensor<string, []>("input_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_5_groups_0 = const()[name = tensor<string, []>("input_5_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_0_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136562048)))]; tensor<fp16, [1, 16, 1, 1]> input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_0_self_attn_v_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")]; tensor<string, []> lora_out_5_pad_type_0 = const()[name = tensor<string, []>("lora_out_5_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_5_strides_0 = const()[name = tensor<string, []>("lora_out_5_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_5_pad_0 = const()[name = tensor<string, []>("lora_out_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_5_dilations_0 = const()[name = tensor<string, []>("lora_out_5_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_5_groups_0 = const()[name = tensor<string, []>("lora_out_5_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_0_self_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136603072)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_5_cast_fp16 = conv(dilations = lora_out_5_dilations_0, groups = lora_out_5_groups_0, pad = lora_out_5_pad_0, pad_type = lora_out_5_pad_type_0, strides = lora_out_5_strides_0, weight = layers_0_self_attn_v_proj_loraB_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("lora_out_5_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> current_value_1_cast_fp16 = add(x = pretrained_out_5_cast_fp16, y = lora_out_5_cast_fp16)[name = tensor<string, []>("current_value_1_cast_fp16")]; tensor<int32, [1]> var_173_axes_0 = const()[name = tensor<string, []>("op_173_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, [1, 1, 448]> var_173_cast_fp16 = expand_dims(axes = var_173_axes_0, x = kv_cache_update_mask)[name = tensor<string, []>("op_173_cast_fp16")]; tensor<int32, [1]> var_174_axes_0 = const()[name = tensor<string, []>("op_174_axes_0"), val = tensor<int32, [1]>([2])]; tensor<fp16, [1, 1, 1, 448]> var_174_cast_fp16 = expand_dims(axes = var_174_axes_0, x = var_173_cast_fp16)[name = tensor<string, []>("op_174_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_176_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_176_cast_fp16")]; tensor<fp16, []> var_65_to_fp16 = const()[name = tensor<string, []>("op_65_to_fp16"), val = tensor<fp16, []>(0x1p+0)]; tensor<fp16, [1, 1, 1, 448]> var_177_cast_fp16 = sub(x = var_65_to_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_177_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_178_cast_fp16 = mul(x = var_47_cast_fp16_0, y = var_177_cast_fp16)[name = tensor<string, []>("op_178_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> key_1_cast_fp16 = add(x = var_176_cast_fp16, y = var_178_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_180_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_180_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_182_cast_fp16 = mul(x = var_54_cast_fp16_0, y = var_177_cast_fp16)[name = tensor<string, []>("op_182_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> value_1_cast_fp16 = add(x = var_180_cast_fp16, y = var_182_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")]; tensor<int32, [4]> var_185 = const()[name = tensor<string, []>("op_185"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1]> mh_q_1_cast_fp16 = reshape(shape = var_185, x = query_1_cast_fp16)[name = tensor<string, []>("mh_q_1_cast_fp16")]; tensor<fp16, []> var_187_to_fp16 = const()[name = tensor<string, []>("op_187_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; tensor<fp16, [1, 20, 64, 1]> var_188_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_187_to_fp16)[name = tensor<string, []>("op_188_cast_fp16")]; tensor<int32, [4]> var_189 = const()[name = tensor<string, []>("op_189"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 448]> var_190_cast_fp16 = reshape(shape = var_189, x = key_1_cast_fp16)[name = tensor<string, []>("op_190_cast_fp16")]; tensor<bool, []> mh_w_1_transpose_x_0 = const()[name = tensor<string, []>("mh_w_1_transpose_x_0"), val = tensor<bool, []>(true)]; tensor<bool, []> mh_w_1_transpose_y_0 = const()[name = tensor<string, []>("mh_w_1_transpose_y_0"), val = tensor<bool, []>(false)]; tensor<fp16, [1, 20, 1, 448]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_188_cast_fp16, y = var_190_cast_fp16)[name = tensor<string, []>("mh_w_1_cast_fp16")]; tensor<int32, [1]> var_194_axes_0 = const()[name = tensor<string, []>("op_194_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, [1, 1, 448]> var_194_cast_fp16 = expand_dims(axes = var_194_axes_0, x = decoder_key_padding_mask)[name = tensor<string, []>("op_194_cast_fp16")]; tensor<int32, [1]> var_195_axes_0 = const()[name = tensor<string, []>("op_195_axes_0"), val = tensor<int32, [1]>([2])]; tensor<fp16, [1, 1, 1, 448]> var_195_cast_fp16 = expand_dims(axes = var_195_axes_0, x = var_194_cast_fp16)[name = tensor<string, []>("op_195_cast_fp16")]; tensor<fp16, [1, 20, 1, 448]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_195_cast_fp16)[name = tensor<string, []>("mh_w_3_cast_fp16")]; tensor<fp16, [1, 20, 1, 448]> var_198_cast_fp16 = softmax(axis = var_64, x = mh_w_3_cast_fp16)[name = tensor<string, []>("op_198_cast_fp16")]; tensor<int32, [4]> var_199 = const()[name = tensor<string, []>("op_199"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 448]> var_200_cast_fp16 = reshape(shape = var_199, x = value_1_cast_fp16)[name = tensor<string, []>("op_200_cast_fp16")]; tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)]; tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(true)]; tensor<fp16, [1, 20, 64, 1]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_200_cast_fp16, y = var_198_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")]; tensor<int32, [4]> var_203 = const()[name = tensor<string, []>("op_203"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; tensor<fp16, [1, 1280, 1, 1]> input_7_cast_fp16 = reshape(shape = var_203, x = attn_1_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")]; tensor<string, []> pretrained_out_7_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_7_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_7_strides_0 = const()[name = tensor<string, []>("pretrained_out_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_7_pad_0 = const()[name = tensor<string, []>("pretrained_out_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_7_dilations_0 = const()[name = tensor<string, []>("pretrained_out_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_7_groups_0 = const()[name = tensor<string, []>("pretrained_out_7_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136644096))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137463360))), name = tensor<string, []>("layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_0_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137463488)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_7_dilations_0, groups = pretrained_out_7_groups_0, pad = pretrained_out_7_pad_0, pad_type = pretrained_out_7_pad_type_0, strides = pretrained_out_7_strides_0, weight = layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = tensor<string, []>("pretrained_out_7_cast_fp16")]; tensor<string, []> input_9_pad_type_0 = const()[name = tensor<string, []>("input_9_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_9_strides_0 = const()[name = tensor<string, []>("input_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_9_dilations_0 = const()[name = tensor<string, []>("input_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_9_groups_0 = const()[name = tensor<string, []>("input_9_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_0_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137466112)))]; tensor<fp16, [1, 16, 1, 1]> input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_0_self_attn_o_proj_loraA_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")]; tensor<string, []> lora_out_7_pad_type_0 = const()[name = tensor<string, []>("lora_out_7_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_7_strides_0 = const()[name = tensor<string, []>("lora_out_7_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_7_pad_0 = const()[name = tensor<string, []>("lora_out_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_7_dilations_0 = const()[name = tensor<string, []>("lora_out_7_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_7_groups_0 = const()[name = tensor<string, []>("lora_out_7_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_0_self_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137507136)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_7_cast_fp16 = conv(dilations = lora_out_7_dilations_0, groups = lora_out_7_groups_0, pad = lora_out_7_pad_0, pad_type = lora_out_7_pad_type_0, strides = lora_out_7_strides_0, weight = layers_0_self_attn_o_proj_loraB_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("lora_out_7_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> obj_7_cast_fp16 = add(x = pretrained_out_7_cast_fp16, y = lora_out_7_cast_fp16)[name = tensor<string, []>("obj_7_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor<string, []>("inputs_3_cast_fp16")]; tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_241_to_fp16 = const()[name = tensor<string, []>("op_241_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_241_to_fp16, x = inputs_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")]; tensor<fp16, [1280]> obj_9_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_9_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137548160)))]; tensor<fp16, [1280]> obj_9_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_9_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137550784)))]; tensor<fp16, []> obj_9_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_9_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor<string, []>("obj_9_cast_fp16")]; tensor<string, []> pretrained_out_9_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_9_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_9_strides_0 = const()[name = tensor<string, []>("pretrained_out_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_9_pad_0 = const()[name = tensor<string, []>("pretrained_out_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_9_dilations_0 = const()[name = tensor<string, []>("pretrained_out_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_9_groups_0 = const()[name = tensor<string, []>("pretrained_out_9_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137553408))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138372672))), name = tensor<string, []>("layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138372800)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_9_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_9_dilations_0, groups = pretrained_out_9_groups_0, pad = pretrained_out_9_pad_0, pad_type = pretrained_out_9_pad_type_0, strides = pretrained_out_9_strides_0, weight = layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor<string, []>("pretrained_out_9_cast_fp16")]; tensor<string, []> input_11_pad_type_0 = const()[name = tensor<string, []>("input_11_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_11_strides_0 = const()[name = tensor<string, []>("input_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_11_pad_0 = const()[name = tensor<string, []>("input_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_11_dilations_0 = const()[name = tensor<string, []>("input_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_11_groups_0 = const()[name = tensor<string, []>("input_11_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_0_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138375424)))]; tensor<fp16, [1, 16, 1, 1]> input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_0_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")]; tensor<string, []> lora_out_9_pad_type_0 = const()[name = tensor<string, []>("lora_out_9_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_9_strides_0 = const()[name = tensor<string, []>("lora_out_9_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_9_pad_0 = const()[name = tensor<string, []>("lora_out_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_9_dilations_0 = const()[name = tensor<string, []>("lora_out_9_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_9_groups_0 = const()[name = tensor<string, []>("lora_out_9_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_0_encoder_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138416448)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_9_cast_fp16 = conv(dilations = lora_out_9_dilations_0, groups = lora_out_9_groups_0, pad = lora_out_9_pad_0, pad_type = lora_out_9_pad_type_0, strides = lora_out_9_strides_0, weight = layers_0_encoder_attn_q_proj_loraB_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("lora_out_9_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> query_3_cast_fp16 = add(x = pretrained_out_9_cast_fp16, y = lora_out_9_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")]; tensor<string, []> pretrained_out_11_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_11_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_11_strides_0 = const()[name = tensor<string, []>("pretrained_out_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_11_pad_0 = const()[name = tensor<string, []>("pretrained_out_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_11_dilations_0 = const()[name = tensor<string, []>("pretrained_out_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_11_groups_0 = const()[name = tensor<string, []>("pretrained_out_11_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138457472))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139276736))), name = tensor<string, []>("layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_11_cast_fp16 = conv(dilations = pretrained_out_11_dilations_0, groups = pretrained_out_11_groups_0, pad = pretrained_out_11_pad_0, pad_type = pretrained_out_11_pad_type_0, strides = pretrained_out_11_strides_0, weight = layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_11_cast_fp16")]; tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_13_strides_0 = const()[name = tensor<string, []>("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_13_dilations_0 = const()[name = tensor<string, []>("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_13_groups_0 = const()[name = tensor<string, []>("input_13_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_0_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139276864)))]; tensor<fp16, [1, 16, 1, 1500]> input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_0_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_13_cast_fp16")]; tensor<string, []> lora_out_11_pad_type_0 = const()[name = tensor<string, []>("lora_out_11_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_11_strides_0 = const()[name = tensor<string, []>("lora_out_11_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_11_pad_0 = const()[name = tensor<string, []>("lora_out_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_11_dilations_0 = const()[name = tensor<string, []>("lora_out_11_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_11_groups_0 = const()[name = tensor<string, []>("lora_out_11_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_0_encoder_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139317888)))]; tensor<fp16, [1, 1280, 1, 1500]> lora_out_11_cast_fp16 = conv(dilations = lora_out_11_dilations_0, groups = lora_out_11_groups_0, pad = lora_out_11_pad_0, pad_type = lora_out_11_pad_type_0, strides = lora_out_11_strides_0, weight = layers_0_encoder_attn_k_proj_loraB_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("lora_out_11_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1500]> key_3_cast_fp16 = add(x = pretrained_out_11_cast_fp16, y = lora_out_11_cast_fp16)[name = tensor<string, []>("key_3_cast_fp16")]; tensor<string, []> pretrained_out_13_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_13_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_13_strides_0 = const()[name = tensor<string, []>("pretrained_out_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_13_pad_0 = const()[name = tensor<string, []>("pretrained_out_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_13_dilations_0 = const()[name = tensor<string, []>("pretrained_out_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_13_groups_0 = const()[name = tensor<string, []>("pretrained_out_13_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139358912))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140178176))), name = tensor<string, []>("layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140178304)))]; tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_13_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_13_dilations_0, groups = pretrained_out_13_groups_0, pad = pretrained_out_13_pad_0, pad_type = pretrained_out_13_pad_type_0, strides = pretrained_out_13_strides_0, weight = layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_13_cast_fp16")]; tensor<string, []> input_15_pad_type_0 = const()[name = tensor<string, []>("input_15_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_15_strides_0 = const()[name = tensor<string, []>("input_15_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_15_pad_0 = const()[name = tensor<string, []>("input_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_15_dilations_0 = const()[name = tensor<string, []>("input_15_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_15_groups_0 = const()[name = tensor<string, []>("input_15_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_0_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140180928)))]; tensor<fp16, [1, 16, 1, 1500]> input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_0_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_15_cast_fp16")]; tensor<string, []> lora_out_13_pad_type_0 = const()[name = tensor<string, []>("lora_out_13_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_13_strides_0 = const()[name = tensor<string, []>("lora_out_13_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_13_pad_0 = const()[name = tensor<string, []>("lora_out_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_13_dilations_0 = const()[name = tensor<string, []>("lora_out_13_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_13_groups_0 = const()[name = tensor<string, []>("lora_out_13_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_0_encoder_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140221952)))]; tensor<fp16, [1, 1280, 1, 1500]> lora_out_13_cast_fp16 = conv(dilations = lora_out_13_dilations_0, groups = lora_out_13_groups_0, pad = lora_out_13_pad_0, pad_type = lora_out_13_pad_type_0, strides = lora_out_13_strides_0, weight = layers_0_encoder_attn_v_proj_loraB_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("lora_out_13_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1500]> value_3_cast_fp16 = add(x = pretrained_out_13_cast_fp16, y = lora_out_13_cast_fp16)[name = tensor<string, []>("value_3_cast_fp16")]; tensor<int32, [4]> var_324 = const()[name = tensor<string, []>("op_324"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1]> mh_q_3_cast_fp16 = reshape(shape = var_324, x = query_3_cast_fp16)[name = tensor<string, []>("mh_q_3_cast_fp16")]; tensor<fp16, []> var_326_to_fp16 = const()[name = tensor<string, []>("op_326_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; tensor<fp16, [1, 20, 64, 1]> var_327_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_326_to_fp16)[name = tensor<string, []>("op_327_cast_fp16")]; tensor<int32, [4]> var_328 = const()[name = tensor<string, []>("op_328"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1500]> var_329_cast_fp16 = reshape(shape = var_328, x = key_3_cast_fp16)[name = tensor<string, []>("op_329_cast_fp16")]; tensor<bool, []> mh_w_5_transpose_x_0 = const()[name = tensor<string, []>("mh_w_5_transpose_x_0"), val = tensor<bool, []>(true)]; tensor<bool, []> mh_w_5_transpose_y_0 = const()[name = tensor<string, []>("mh_w_5_transpose_y_0"), val = tensor<bool, []>(false)]; tensor<fp16, [1, 20, 1, 1500]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_327_cast_fp16, y = var_329_cast_fp16)[name = tensor<string, []>("mh_w_5_cast_fp16")]; tensor<fp16, [1, 20, 1, 1500]> obj_13_cast_fp16 = softmax(axis = var_64, x = mh_w_5_cast_fp16)[name = tensor<string, []>("obj_13_cast_fp16")]; tensor<int32, [4]> var_333 = const()[name = tensor<string, []>("op_333"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1500]> var_334_cast_fp16 = reshape(shape = var_333, x = value_3_cast_fp16)[name = tensor<string, []>("op_334_cast_fp16")]; tensor<bool, []> attn_3_transpose_x_0 = const()[name = tensor<string, []>("attn_3_transpose_x_0"), val = tensor<bool, []>(false)]; tensor<bool, []> attn_3_transpose_y_0 = const()[name = tensor<string, []>("attn_3_transpose_y_0"), val = tensor<bool, []>(true)]; tensor<fp16, [1, 20, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_334_cast_fp16, y = obj_13_cast_fp16)[name = tensor<string, []>("attn_3_cast_fp16")]; tensor<int32, [4]> var_337 = const()[name = tensor<string, []>("op_337"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; tensor<fp16, [1, 1280, 1, 1]> input_17_cast_fp16 = reshape(shape = var_337, x = attn_3_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")]; tensor<string, []> pretrained_out_15_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_15_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_15_strides_0 = const()[name = tensor<string, []>("pretrained_out_15_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_15_pad_0 = const()[name = tensor<string, []>("pretrained_out_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_15_dilations_0 = const()[name = tensor<string, []>("pretrained_out_15_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_15_groups_0 = const()[name = tensor<string, []>("pretrained_out_15_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140262976))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141082240))), name = tensor<string, []>("layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141082368)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_15_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_15_dilations_0, groups = pretrained_out_15_groups_0, pad = pretrained_out_15_pad_0, pad_type = pretrained_out_15_pad_type_0, strides = pretrained_out_15_strides_0, weight = layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = tensor<string, []>("pretrained_out_15_cast_fp16")]; tensor<string, []> input_19_pad_type_0 = const()[name = tensor<string, []>("input_19_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_19_strides_0 = const()[name = tensor<string, []>("input_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_19_dilations_0 = const()[name = tensor<string, []>("input_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_19_groups_0 = const()[name = tensor<string, []>("input_19_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_0_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141084992)))]; tensor<fp16, [1, 16, 1, 1]> input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_0_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")]; tensor<string, []> lora_out_15_pad_type_0 = const()[name = tensor<string, []>("lora_out_15_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_15_strides_0 = const()[name = tensor<string, []>("lora_out_15_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_15_pad_0 = const()[name = tensor<string, []>("lora_out_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_15_dilations_0 = const()[name = tensor<string, []>("lora_out_15_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_15_groups_0 = const()[name = tensor<string, []>("lora_out_15_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_0_encoder_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_encoder_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141126016)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_15_cast_fp16 = conv(dilations = lora_out_15_dilations_0, groups = lora_out_15_groups_0, pad = lora_out_15_pad_0, pad_type = lora_out_15_pad_type_0, strides = lora_out_15_strides_0, weight = layers_0_encoder_attn_o_proj_loraB_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("lora_out_15_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> obj_11_cast_fp16 = add(x = pretrained_out_15_cast_fp16, y = lora_out_15_cast_fp16)[name = tensor<string, []>("obj_11_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor<string, []>("inputs_5_cast_fp16")]; tensor<int32, [1]> out_5_axes_0 = const()[name = tensor<string, []>("out_5_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_371_to_fp16 = const()[name = tensor<string, []>("op_371_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_371_to_fp16, x = inputs_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")]; tensor<fp16, [1280]> input_21_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_21_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141167040)))]; tensor<fp16, [1280]> input_21_beta_0_to_fp16 = const()[name = tensor<string, []>("input_21_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141169664)))]; tensor<fp16, []> input_21_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_21_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> input_21_cast_fp16 = batch_norm(beta = input_21_beta_0_to_fp16, epsilon = input_21_epsilon_0_to_fp16, gamma = input_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")]; tensor<string, []> pretrained_out_17_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_17_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_17_strides_0 = const()[name = tensor<string, []>("pretrained_out_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_17_pad_0 = const()[name = tensor<string, []>("pretrained_out_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_17_dilations_0 = const()[name = tensor<string, []>("pretrained_out_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_17_groups_0 = const()[name = tensor<string, []>("pretrained_out_17_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141172288))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144449152))), name = tensor<string, []>("layers_0_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])]; tensor<fp16, [5120]> layers_0_fc1_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_pretrained_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144449280)))]; tensor<fp16, [1, 5120, 1, 1]> pretrained_out_17_cast_fp16 = conv(bias = layers_0_fc1_pretrained_bias_to_fp16, dilations = pretrained_out_17_dilations_0, groups = pretrained_out_17_groups_0, pad = pretrained_out_17_pad_0, pad_type = pretrained_out_17_pad_type_0, strides = pretrained_out_17_strides_0, weight = layers_0_fc1_pretrained_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = tensor<string, []>("pretrained_out_17_cast_fp16")]; tensor<string, []> input_23_pad_type_0 = const()[name = tensor<string, []>("input_23_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_23_strides_0 = const()[name = tensor<string, []>("input_23_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_23_pad_0 = const()[name = tensor<string, []>("input_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_23_dilations_0 = const()[name = tensor<string, []>("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_23_groups_0 = const()[name = tensor<string, []>("input_23_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_0_fc1_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144459584)))]; tensor<fp16, [1, 16, 1, 1]> input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_0_fc1_loraA_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")]; tensor<string, []> lora_out_17_pad_type_0 = const()[name = tensor<string, []>("lora_out_17_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_17_strides_0 = const()[name = tensor<string, []>("lora_out_17_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_17_pad_0 = const()[name = tensor<string, []>("lora_out_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_17_dilations_0 = const()[name = tensor<string, []>("lora_out_17_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_17_groups_0 = const()[name = tensor<string, []>("lora_out_17_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [5120, 16, 1, 1]> layers_0_fc1_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc1_loraB_weight_to_fp16"), val = tensor<fp16, [5120, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144500608)))]; tensor<fp16, [1, 5120, 1, 1]> lora_out_17_cast_fp16 = conv(dilations = lora_out_17_dilations_0, groups = lora_out_17_groups_0, pad = lora_out_17_pad_0, pad_type = lora_out_17_pad_type_0, strides = lora_out_17_strides_0, weight = layers_0_fc1_loraB_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("lora_out_17_cast_fp16")]; tensor<fp16, [1, 5120, 1, 1]> input_25_cast_fp16 = add(x = pretrained_out_17_cast_fp16, y = lora_out_17_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")]; tensor<string, []> input_27_mode_0 = const()[name = tensor<string, []>("input_27_mode_0"), val = tensor<string, []>("EXACT")]; tensor<fp16, [1, 5120, 1, 1]> input_27_cast_fp16 = gelu(mode = input_27_mode_0, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")]; tensor<string, []> pretrained_out_19_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_19_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_19_strides_0 = const()[name = tensor<string, []>("pretrained_out_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_19_pad_0 = const()[name = tensor<string, []>("pretrained_out_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_19_dilations_0 = const()[name = tensor<string, []>("pretrained_out_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_19_groups_0 = const()[name = tensor<string, []>("pretrained_out_19_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(144664512))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147941376))), name = tensor<string, []>("layers_0_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])]; tensor<fp16, [1280]> layers_0_fc2_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147941504)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_19_cast_fp16 = conv(bias = layers_0_fc2_pretrained_bias_to_fp16, dilations = pretrained_out_19_dilations_0, groups = pretrained_out_19_groups_0, pad = pretrained_out_19_pad_0, pad_type = pretrained_out_19_pad_type_0, strides = pretrained_out_19_strides_0, weight = layers_0_fc2_pretrained_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = tensor<string, []>("pretrained_out_19_cast_fp16")]; tensor<string, []> input_29_pad_type_0 = const()[name = tensor<string, []>("input_29_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_29_strides_0 = const()[name = tensor<string, []>("input_29_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_29_pad_0 = const()[name = tensor<string, []>("input_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_29_dilations_0 = const()[name = tensor<string, []>("input_29_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_29_groups_0 = const()[name = tensor<string, []>("input_29_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 5120, 1, 1]> layers_0_fc2_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_loraA_weight_to_fp16"), val = tensor<fp16, [16, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(147944128)))]; tensor<fp16, [1, 16, 1, 1]> input_29_cast_fp16 = conv(dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = layers_0_fc2_loraA_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")]; tensor<string, []> lora_out_19_pad_type_0 = const()[name = tensor<string, []>("lora_out_19_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_19_strides_0 = const()[name = tensor<string, []>("lora_out_19_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_19_pad_0 = const()[name = tensor<string, []>("lora_out_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_19_dilations_0 = const()[name = tensor<string, []>("lora_out_19_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_19_groups_0 = const()[name = tensor<string, []>("lora_out_19_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_0_fc2_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_fc2_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148108032)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_19_cast_fp16 = conv(dilations = lora_out_19_dilations_0, groups = lora_out_19_groups_0, pad = lora_out_19_pad_0, pad_type = lora_out_19_pad_type_0, strides = lora_out_19_strides_0, weight = layers_0_fc2_loraB_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("lora_out_19_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> hidden_states_3_cast_fp16 = add(x = pretrained_out_19_cast_fp16, y = lora_out_19_cast_fp16)[name = tensor<string, []>("hidden_states_3_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor<string, []>("inputs_7_cast_fp16")]; tensor<int32, []> var_438 = const()[name = tensor<string, []>("op_438"), val = tensor<int32, []>(3)]; tensor<int32, [1]> out_7_axes_0 = const()[name = tensor<string, []>("out_7_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_464_to_fp16 = const()[name = tensor<string, []>("op_464_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_464_to_fp16, x = inputs_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")]; tensor<fp16, [1280]> obj_15_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_15_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148149056)))]; tensor<fp16, [1280]> obj_15_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_15_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148151680)))]; tensor<fp16, []> obj_15_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_15_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor<string, []>("obj_15_cast_fp16")]; tensor<string, []> pretrained_out_21_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_21_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_21_strides_0 = const()[name = tensor<string, []>("pretrained_out_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_21_pad_0 = const()[name = tensor<string, []>("pretrained_out_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_21_dilations_0 = const()[name = tensor<string, []>("pretrained_out_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_21_groups_0 = const()[name = tensor<string, []>("pretrained_out_21_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148154304))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148973568))), name = tensor<string, []>("layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_1_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148973696)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_21_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_21_dilations_0, groups = pretrained_out_21_groups_0, pad = pretrained_out_21_pad_0, pad_type = pretrained_out_21_pad_type_0, strides = pretrained_out_21_strides_0, weight = layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor<string, []>("pretrained_out_21_cast_fp16")]; tensor<string, []> input_31_pad_type_0 = const()[name = tensor<string, []>("input_31_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_31_strides_0 = const()[name = tensor<string, []>("input_31_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_31_pad_0 = const()[name = tensor<string, []>("input_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_31_dilations_0 = const()[name = tensor<string, []>("input_31_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_31_groups_0 = const()[name = tensor<string, []>("input_31_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_1_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148976320)))]; tensor<fp16, [1, 16, 1, 1]> input_31_cast_fp16 = conv(dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = layers_1_self_attn_q_proj_loraA_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")]; tensor<string, []> lora_out_21_pad_type_0 = const()[name = tensor<string, []>("lora_out_21_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_21_strides_0 = const()[name = tensor<string, []>("lora_out_21_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_21_pad_0 = const()[name = tensor<string, []>("lora_out_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_21_dilations_0 = const()[name = tensor<string, []>("lora_out_21_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_21_groups_0 = const()[name = tensor<string, []>("lora_out_21_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_1_self_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149017344)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_21_cast_fp16 = conv(dilations = lora_out_21_dilations_0, groups = lora_out_21_groups_0, pad = lora_out_21_pad_0, pad_type = lora_out_21_pad_type_0, strides = lora_out_21_strides_0, weight = layers_1_self_attn_q_proj_loraB_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("lora_out_21_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> query_5_cast_fp16 = add(x = pretrained_out_21_cast_fp16, y = lora_out_21_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")]; tensor<string, []> pretrained_out_23_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_23_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_23_strides_0 = const()[name = tensor<string, []>("pretrained_out_23_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_23_pad_0 = const()[name = tensor<string, []>("pretrained_out_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_23_dilations_0 = const()[name = tensor<string, []>("pretrained_out_23_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_23_groups_0 = const()[name = tensor<string, []>("pretrained_out_23_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149058368))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149877632))), name = tensor<string, []>("layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_23_cast_fp16 = conv(dilations = pretrained_out_23_dilations_0, groups = pretrained_out_23_groups_0, pad = pretrained_out_23_pad_0, pad_type = pretrained_out_23_pad_type_0, strides = pretrained_out_23_strides_0, weight = layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor<string, []>("pretrained_out_23_cast_fp16")]; tensor<string, []> input_33_pad_type_0 = const()[name = tensor<string, []>("input_33_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_33_strides_0 = const()[name = tensor<string, []>("input_33_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_33_pad_0 = const()[name = tensor<string, []>("input_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_33_dilations_0 = const()[name = tensor<string, []>("input_33_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_33_groups_0 = const()[name = tensor<string, []>("input_33_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_1_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149877760)))]; tensor<fp16, [1, 16, 1, 1]> input_33_cast_fp16 = conv(dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = layers_1_self_attn_k_proj_loraA_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")]; tensor<string, []> lora_out_23_pad_type_0 = const()[name = tensor<string, []>("lora_out_23_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_23_strides_0 = const()[name = tensor<string, []>("lora_out_23_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_23_pad_0 = const()[name = tensor<string, []>("lora_out_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_23_dilations_0 = const()[name = tensor<string, []>("lora_out_23_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_23_groups_0 = const()[name = tensor<string, []>("lora_out_23_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_1_self_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149918784)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_23_cast_fp16 = conv(dilations = lora_out_23_dilations_0, groups = lora_out_23_groups_0, pad = lora_out_23_pad_0, pad_type = lora_out_23_pad_type_0, strides = lora_out_23_strides_0, weight = layers_1_self_attn_k_proj_loraB_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("lora_out_23_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> current_key_3_cast_fp16 = add(x = pretrained_out_23_cast_fp16, y = lora_out_23_cast_fp16)[name = tensor<string, []>("current_key_3_cast_fp16")]; tensor<string, []> pretrained_out_25_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_25_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_25_strides_0 = const()[name = tensor<string, []>("pretrained_out_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_25_pad_0 = const()[name = tensor<string, []>("pretrained_out_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_25_dilations_0 = const()[name = tensor<string, []>("pretrained_out_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_25_groups_0 = const()[name = tensor<string, []>("pretrained_out_25_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149959808))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150779072))), name = tensor<string, []>("layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_1_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150779200)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_25_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_25_dilations_0, groups = pretrained_out_25_groups_0, pad = pretrained_out_25_pad_0, pad_type = pretrained_out_25_pad_type_0, strides = pretrained_out_25_strides_0, weight = layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor<string, []>("pretrained_out_25_cast_fp16")]; tensor<string, []> input_35_pad_type_0 = const()[name = tensor<string, []>("input_35_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_35_strides_0 = const()[name = tensor<string, []>("input_35_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_35_pad_0 = const()[name = tensor<string, []>("input_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_35_dilations_0 = const()[name = tensor<string, []>("input_35_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_35_groups_0 = const()[name = tensor<string, []>("input_35_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_1_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150781824)))]; tensor<fp16, [1, 16, 1, 1]> input_35_cast_fp16 = conv(dilations = input_35_dilations_0, groups = input_35_groups_0, pad = input_35_pad_0, pad_type = input_35_pad_type_0, strides = input_35_strides_0, weight = layers_1_self_attn_v_proj_loraA_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")]; tensor<string, []> lora_out_25_pad_type_0 = const()[name = tensor<string, []>("lora_out_25_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_25_strides_0 = const()[name = tensor<string, []>("lora_out_25_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_25_pad_0 = const()[name = tensor<string, []>("lora_out_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_25_dilations_0 = const()[name = tensor<string, []>("lora_out_25_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_25_groups_0 = const()[name = tensor<string, []>("lora_out_25_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_1_self_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150822848)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_25_cast_fp16 = conv(dilations = lora_out_25_dilations_0, groups = lora_out_25_groups_0, pad = lora_out_25_pad_0, pad_type = lora_out_25_pad_type_0, strides = lora_out_25_strides_0, weight = layers_1_self_attn_v_proj_loraB_weight_to_fp16, x = input_35_cast_fp16)[name = tensor<string, []>("lora_out_25_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> current_value_3_cast_fp16 = add(x = pretrained_out_25_cast_fp16, y = lora_out_25_cast_fp16)[name = tensor<string, []>("current_value_3_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_550_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_550_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_552_cast_fp16 = mul(x = var_47_cast_fp16_1, y = var_177_cast_fp16)[name = tensor<string, []>("op_552_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> key_5_cast_fp16 = add(x = var_550_cast_fp16, y = var_552_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_554_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_554_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_556_cast_fp16 = mul(x = var_54_cast_fp16_1, y = var_177_cast_fp16)[name = tensor<string, []>("op_556_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> value_5_cast_fp16 = add(x = var_554_cast_fp16, y = var_556_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")]; tensor<int32, [4]> var_559 = const()[name = tensor<string, []>("op_559"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1]> mh_q_5_cast_fp16 = reshape(shape = var_559, x = query_5_cast_fp16)[name = tensor<string, []>("mh_q_5_cast_fp16")]; tensor<fp16, []> var_561_to_fp16 = const()[name = tensor<string, []>("op_561_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; tensor<fp16, [1, 20, 64, 1]> var_562_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_561_to_fp16)[name = tensor<string, []>("op_562_cast_fp16")]; tensor<int32, [4]> var_563 = const()[name = tensor<string, []>("op_563"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 448]> var_564_cast_fp16 = reshape(shape = var_563, x = key_5_cast_fp16)[name = tensor<string, []>("op_564_cast_fp16")]; tensor<bool, []> mh_w_7_transpose_x_0 = const()[name = tensor<string, []>("mh_w_7_transpose_x_0"), val = tensor<bool, []>(true)]; tensor<bool, []> mh_w_7_transpose_y_0 = const()[name = tensor<string, []>("mh_w_7_transpose_y_0"), val = tensor<bool, []>(false)]; tensor<fp16, [1, 20, 1, 448]> mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_562_cast_fp16, y = var_564_cast_fp16)[name = tensor<string, []>("mh_w_7_cast_fp16")]; tensor<fp16, [1, 20, 1, 448]> mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_195_cast_fp16)[name = tensor<string, []>("mh_w_9_cast_fp16")]; tensor<fp16, [1, 20, 1, 448]> var_572_cast_fp16 = softmax(axis = var_438, x = mh_w_9_cast_fp16)[name = tensor<string, []>("op_572_cast_fp16")]; tensor<int32, [4]> var_573 = const()[name = tensor<string, []>("op_573"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 448]> var_574_cast_fp16 = reshape(shape = var_573, x = value_5_cast_fp16)[name = tensor<string, []>("op_574_cast_fp16")]; tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)]; tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(true)]; tensor<fp16, [1, 20, 64, 1]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_574_cast_fp16, y = var_572_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")]; tensor<int32, [4]> var_577 = const()[name = tensor<string, []>("op_577"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; tensor<fp16, [1, 1280, 1, 1]> input_37_cast_fp16 = reshape(shape = var_577, x = attn_5_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")]; tensor<string, []> pretrained_out_27_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_27_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_27_strides_0 = const()[name = tensor<string, []>("pretrained_out_27_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_27_pad_0 = const()[name = tensor<string, []>("pretrained_out_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_27_dilations_0 = const()[name = tensor<string, []>("pretrained_out_27_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_27_groups_0 = const()[name = tensor<string, []>("pretrained_out_27_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150863872))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151683136))), name = tensor<string, []>("layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_1_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151683264)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_27_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_27_dilations_0, groups = pretrained_out_27_groups_0, pad = pretrained_out_27_pad_0, pad_type = pretrained_out_27_pad_type_0, strides = pretrained_out_27_strides_0, weight = layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = tensor<string, []>("pretrained_out_27_cast_fp16")]; tensor<string, []> input_39_pad_type_0 = const()[name = tensor<string, []>("input_39_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_39_strides_0 = const()[name = tensor<string, []>("input_39_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_39_pad_0 = const()[name = tensor<string, []>("input_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_39_dilations_0 = const()[name = tensor<string, []>("input_39_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_39_groups_0 = const()[name = tensor<string, []>("input_39_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_1_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151685888)))]; tensor<fp16, [1, 16, 1, 1]> input_39_cast_fp16 = conv(dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = layers_1_self_attn_o_proj_loraA_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")]; tensor<string, []> lora_out_27_pad_type_0 = const()[name = tensor<string, []>("lora_out_27_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_27_strides_0 = const()[name = tensor<string, []>("lora_out_27_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_27_pad_0 = const()[name = tensor<string, []>("lora_out_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_27_dilations_0 = const()[name = tensor<string, []>("lora_out_27_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_27_groups_0 = const()[name = tensor<string, []>("lora_out_27_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_1_self_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151726912)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_27_cast_fp16 = conv(dilations = lora_out_27_dilations_0, groups = lora_out_27_groups_0, pad = lora_out_27_pad_0, pad_type = lora_out_27_pad_type_0, strides = lora_out_27_strides_0, weight = layers_1_self_attn_o_proj_loraB_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("lora_out_27_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> obj_21_cast_fp16 = add(x = pretrained_out_27_cast_fp16, y = lora_out_27_cast_fp16)[name = tensor<string, []>("obj_21_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor<string, []>("inputs_9_cast_fp16")]; tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_615_to_fp16 = const()[name = tensor<string, []>("op_615_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_615_to_fp16, x = inputs_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")]; tensor<fp16, [1280]> obj_23_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_23_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151767936)))]; tensor<fp16, [1280]> obj_23_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_23_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151770560)))]; tensor<fp16, []> obj_23_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_23_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor<string, []>("obj_23_cast_fp16")]; tensor<string, []> pretrained_out_29_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_29_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_29_strides_0 = const()[name = tensor<string, []>("pretrained_out_29_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_29_pad_0 = const()[name = tensor<string, []>("pretrained_out_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_29_dilations_0 = const()[name = tensor<string, []>("pretrained_out_29_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_29_groups_0 = const()[name = tensor<string, []>("pretrained_out_29_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151773184))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152592448))), name = tensor<string, []>("layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152592576)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_29_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_29_dilations_0, groups = pretrained_out_29_groups_0, pad = pretrained_out_29_pad_0, pad_type = pretrained_out_29_pad_type_0, strides = pretrained_out_29_strides_0, weight = layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_23_cast_fp16)[name = tensor<string, []>("pretrained_out_29_cast_fp16")]; tensor<string, []> input_41_pad_type_0 = const()[name = tensor<string, []>("input_41_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_41_strides_0 = const()[name = tensor<string, []>("input_41_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_41_pad_0 = const()[name = tensor<string, []>("input_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_41_dilations_0 = const()[name = tensor<string, []>("input_41_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_41_groups_0 = const()[name = tensor<string, []>("input_41_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_1_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152595200)))]; tensor<fp16, [1, 16, 1, 1]> input_41_cast_fp16 = conv(dilations = input_41_dilations_0, groups = input_41_groups_0, pad = input_41_pad_0, pad_type = input_41_pad_type_0, strides = input_41_strides_0, weight = layers_1_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")]; tensor<string, []> lora_out_29_pad_type_0 = const()[name = tensor<string, []>("lora_out_29_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_29_strides_0 = const()[name = tensor<string, []>("lora_out_29_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_29_pad_0 = const()[name = tensor<string, []>("lora_out_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_29_dilations_0 = const()[name = tensor<string, []>("lora_out_29_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_29_groups_0 = const()[name = tensor<string, []>("lora_out_29_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_1_encoder_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152636224)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_29_cast_fp16 = conv(dilations = lora_out_29_dilations_0, groups = lora_out_29_groups_0, pad = lora_out_29_pad_0, pad_type = lora_out_29_pad_type_0, strides = lora_out_29_strides_0, weight = layers_1_encoder_attn_q_proj_loraB_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("lora_out_29_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> query_7_cast_fp16 = add(x = pretrained_out_29_cast_fp16, y = lora_out_29_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")]; tensor<string, []> pretrained_out_31_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_31_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_31_strides_0 = const()[name = tensor<string, []>("pretrained_out_31_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_31_pad_0 = const()[name = tensor<string, []>("pretrained_out_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_31_dilations_0 = const()[name = tensor<string, []>("pretrained_out_31_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_31_groups_0 = const()[name = tensor<string, []>("pretrained_out_31_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(152677248))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153496512))), name = tensor<string, []>("layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_31_cast_fp16 = conv(dilations = pretrained_out_31_dilations_0, groups = pretrained_out_31_groups_0, pad = pretrained_out_31_pad_0, pad_type = pretrained_out_31_pad_type_0, strides = pretrained_out_31_strides_0, weight = layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_31_cast_fp16")]; tensor<string, []> input_43_pad_type_0 = const()[name = tensor<string, []>("input_43_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_43_strides_0 = const()[name = tensor<string, []>("input_43_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_43_pad_0 = const()[name = tensor<string, []>("input_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_43_dilations_0 = const()[name = tensor<string, []>("input_43_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_43_groups_0 = const()[name = tensor<string, []>("input_43_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_1_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153496640)))]; tensor<fp16, [1, 16, 1, 1500]> input_43_cast_fp16 = conv(dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = layers_1_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_43_cast_fp16")]; tensor<string, []> lora_out_31_pad_type_0 = const()[name = tensor<string, []>("lora_out_31_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_31_strides_0 = const()[name = tensor<string, []>("lora_out_31_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_31_pad_0 = const()[name = tensor<string, []>("lora_out_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_31_dilations_0 = const()[name = tensor<string, []>("lora_out_31_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_31_groups_0 = const()[name = tensor<string, []>("lora_out_31_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_1_encoder_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153537664)))]; tensor<fp16, [1, 1280, 1, 1500]> lora_out_31_cast_fp16 = conv(dilations = lora_out_31_dilations_0, groups = lora_out_31_groups_0, pad = lora_out_31_pad_0, pad_type = lora_out_31_pad_type_0, strides = lora_out_31_strides_0, weight = layers_1_encoder_attn_k_proj_loraB_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("lora_out_31_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1500]> key_7_cast_fp16 = add(x = pretrained_out_31_cast_fp16, y = lora_out_31_cast_fp16)[name = tensor<string, []>("key_7_cast_fp16")]; tensor<string, []> pretrained_out_33_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_33_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_33_strides_0 = const()[name = tensor<string, []>("pretrained_out_33_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_33_pad_0 = const()[name = tensor<string, []>("pretrained_out_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_33_dilations_0 = const()[name = tensor<string, []>("pretrained_out_33_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_33_groups_0 = const()[name = tensor<string, []>("pretrained_out_33_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(153578688))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154397952))), name = tensor<string, []>("layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154398080)))]; tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_33_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_33_dilations_0, groups = pretrained_out_33_groups_0, pad = pretrained_out_33_pad_0, pad_type = pretrained_out_33_pad_type_0, strides = pretrained_out_33_strides_0, weight = layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_33_cast_fp16")]; tensor<string, []> input_45_pad_type_0 = const()[name = tensor<string, []>("input_45_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_45_strides_0 = const()[name = tensor<string, []>("input_45_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_45_pad_0 = const()[name = tensor<string, []>("input_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_45_dilations_0 = const()[name = tensor<string, []>("input_45_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_45_groups_0 = const()[name = tensor<string, []>("input_45_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_1_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154400704)))]; tensor<fp16, [1, 16, 1, 1500]> input_45_cast_fp16 = conv(dilations = input_45_dilations_0, groups = input_45_groups_0, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = input_45_strides_0, weight = layers_1_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_45_cast_fp16")]; tensor<string, []> lora_out_33_pad_type_0 = const()[name = tensor<string, []>("lora_out_33_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_33_strides_0 = const()[name = tensor<string, []>("lora_out_33_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_33_pad_0 = const()[name = tensor<string, []>("lora_out_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_33_dilations_0 = const()[name = tensor<string, []>("lora_out_33_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_33_groups_0 = const()[name = tensor<string, []>("lora_out_33_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_1_encoder_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154441728)))]; tensor<fp16, [1, 1280, 1, 1500]> lora_out_33_cast_fp16 = conv(dilations = lora_out_33_dilations_0, groups = lora_out_33_groups_0, pad = lora_out_33_pad_0, pad_type = lora_out_33_pad_type_0, strides = lora_out_33_strides_0, weight = layers_1_encoder_attn_v_proj_loraB_weight_to_fp16, x = input_45_cast_fp16)[name = tensor<string, []>("lora_out_33_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1500]> value_7_cast_fp16 = add(x = pretrained_out_33_cast_fp16, y = lora_out_33_cast_fp16)[name = tensor<string, []>("value_7_cast_fp16")]; tensor<int32, [4]> var_698 = const()[name = tensor<string, []>("op_698"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1]> mh_q_7_cast_fp16 = reshape(shape = var_698, x = query_7_cast_fp16)[name = tensor<string, []>("mh_q_7_cast_fp16")]; tensor<fp16, []> var_700_to_fp16 = const()[name = tensor<string, []>("op_700_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; tensor<fp16, [1, 20, 64, 1]> var_701_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_700_to_fp16)[name = tensor<string, []>("op_701_cast_fp16")]; tensor<int32, [4]> var_702 = const()[name = tensor<string, []>("op_702"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1500]> var_703_cast_fp16 = reshape(shape = var_702, x = key_7_cast_fp16)[name = tensor<string, []>("op_703_cast_fp16")]; tensor<bool, []> mh_w_11_transpose_x_0 = const()[name = tensor<string, []>("mh_w_11_transpose_x_0"), val = tensor<bool, []>(true)]; tensor<bool, []> mh_w_11_transpose_y_0 = const()[name = tensor<string, []>("mh_w_11_transpose_y_0"), val = tensor<bool, []>(false)]; tensor<fp16, [1, 20, 1, 1500]> mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_701_cast_fp16, y = var_703_cast_fp16)[name = tensor<string, []>("mh_w_11_cast_fp16")]; tensor<fp16, [1, 20, 1, 1500]> obj_27_cast_fp16 = softmax(axis = var_438, x = mh_w_11_cast_fp16)[name = tensor<string, []>("obj_27_cast_fp16")]; tensor<int32, [4]> var_707 = const()[name = tensor<string, []>("op_707"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1500]> var_708_cast_fp16 = reshape(shape = var_707, x = value_7_cast_fp16)[name = tensor<string, []>("op_708_cast_fp16")]; tensor<bool, []> attn_7_transpose_x_0 = const()[name = tensor<string, []>("attn_7_transpose_x_0"), val = tensor<bool, []>(false)]; tensor<bool, []> attn_7_transpose_y_0 = const()[name = tensor<string, []>("attn_7_transpose_y_0"), val = tensor<bool, []>(true)]; tensor<fp16, [1, 20, 64, 1]> attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_708_cast_fp16, y = obj_27_cast_fp16)[name = tensor<string, []>("attn_7_cast_fp16")]; tensor<int32, [4]> var_711 = const()[name = tensor<string, []>("op_711"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; tensor<fp16, [1, 1280, 1, 1]> input_47_cast_fp16 = reshape(shape = var_711, x = attn_7_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")]; tensor<string, []> pretrained_out_35_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_35_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_35_strides_0 = const()[name = tensor<string, []>("pretrained_out_35_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_35_pad_0 = const()[name = tensor<string, []>("pretrained_out_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_35_dilations_0 = const()[name = tensor<string, []>("pretrained_out_35_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_35_groups_0 = const()[name = tensor<string, []>("pretrained_out_35_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154482752))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155302016))), name = tensor<string, []>("layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155302144)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_35_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_35_dilations_0, groups = pretrained_out_35_groups_0, pad = pretrained_out_35_pad_0, pad_type = pretrained_out_35_pad_type_0, strides = pretrained_out_35_strides_0, weight = layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_47_cast_fp16)[name = tensor<string, []>("pretrained_out_35_cast_fp16")]; tensor<string, []> input_49_pad_type_0 = const()[name = tensor<string, []>("input_49_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_49_strides_0 = const()[name = tensor<string, []>("input_49_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_49_pad_0 = const()[name = tensor<string, []>("input_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_49_dilations_0 = const()[name = tensor<string, []>("input_49_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_49_groups_0 = const()[name = tensor<string, []>("input_49_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_1_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155304768)))]; tensor<fp16, [1, 16, 1, 1]> input_49_cast_fp16 = conv(dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = layers_1_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")]; tensor<string, []> lora_out_35_pad_type_0 = const()[name = tensor<string, []>("lora_out_35_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_35_strides_0 = const()[name = tensor<string, []>("lora_out_35_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_35_pad_0 = const()[name = tensor<string, []>("lora_out_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_35_dilations_0 = const()[name = tensor<string, []>("lora_out_35_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_35_groups_0 = const()[name = tensor<string, []>("lora_out_35_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_1_encoder_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_encoder_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155345792)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_35_cast_fp16 = conv(dilations = lora_out_35_dilations_0, groups = lora_out_35_groups_0, pad = lora_out_35_pad_0, pad_type = lora_out_35_pad_type_0, strides = lora_out_35_strides_0, weight = layers_1_encoder_attn_o_proj_loraB_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("lora_out_35_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> obj_25_cast_fp16 = add(x = pretrained_out_35_cast_fp16, y = lora_out_35_cast_fp16)[name = tensor<string, []>("obj_25_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor<string, []>("inputs_11_cast_fp16")]; tensor<int32, [1]> out_11_axes_0 = const()[name = tensor<string, []>("out_11_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_745_to_fp16 = const()[name = tensor<string, []>("op_745_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_745_to_fp16, x = inputs_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")]; tensor<fp16, [1280]> input_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_51_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155386816)))]; tensor<fp16, [1280]> input_51_beta_0_to_fp16 = const()[name = tensor<string, []>("input_51_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155389440)))]; tensor<fp16, []> input_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> 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<string, []>("input_51_cast_fp16")]; tensor<string, []> pretrained_out_37_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_37_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_37_strides_0 = const()[name = tensor<string, []>("pretrained_out_37_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_37_pad_0 = const()[name = tensor<string, []>("pretrained_out_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_37_dilations_0 = const()[name = tensor<string, []>("pretrained_out_37_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_37_groups_0 = const()[name = tensor<string, []>("pretrained_out_37_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155392064))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158668928))), name = tensor<string, []>("layers_1_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])]; tensor<fp16, [5120]> layers_1_fc1_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_pretrained_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158669056)))]; tensor<fp16, [1, 5120, 1, 1]> pretrained_out_37_cast_fp16 = conv(bias = layers_1_fc1_pretrained_bias_to_fp16, dilations = pretrained_out_37_dilations_0, groups = pretrained_out_37_groups_0, pad = pretrained_out_37_pad_0, pad_type = pretrained_out_37_pad_type_0, strides = pretrained_out_37_strides_0, weight = layers_1_fc1_pretrained_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = tensor<string, []>("pretrained_out_37_cast_fp16")]; tensor<string, []> input_53_pad_type_0 = const()[name = tensor<string, []>("input_53_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_53_strides_0 = const()[name = tensor<string, []>("input_53_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_53_pad_0 = const()[name = tensor<string, []>("input_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_53_dilations_0 = const()[name = tensor<string, []>("input_53_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_53_groups_0 = const()[name = tensor<string, []>("input_53_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_1_fc1_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158679360)))]; tensor<fp16, [1, 16, 1, 1]> input_53_cast_fp16 = conv(dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = layers_1_fc1_loraA_weight_to_fp16, x = input_51_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")]; tensor<string, []> lora_out_37_pad_type_0 = const()[name = tensor<string, []>("lora_out_37_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_37_strides_0 = const()[name = tensor<string, []>("lora_out_37_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_37_pad_0 = const()[name = tensor<string, []>("lora_out_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_37_dilations_0 = const()[name = tensor<string, []>("lora_out_37_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_37_groups_0 = const()[name = tensor<string, []>("lora_out_37_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [5120, 16, 1, 1]> layers_1_fc1_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc1_loraB_weight_to_fp16"), val = tensor<fp16, [5120, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158720384)))]; tensor<fp16, [1, 5120, 1, 1]> lora_out_37_cast_fp16 = conv(dilations = lora_out_37_dilations_0, groups = lora_out_37_groups_0, pad = lora_out_37_pad_0, pad_type = lora_out_37_pad_type_0, strides = lora_out_37_strides_0, weight = layers_1_fc1_loraB_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("lora_out_37_cast_fp16")]; tensor<fp16, [1, 5120, 1, 1]> input_55_cast_fp16 = add(x = pretrained_out_37_cast_fp16, y = lora_out_37_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")]; tensor<string, []> input_57_mode_0 = const()[name = tensor<string, []>("input_57_mode_0"), val = tensor<string, []>("EXACT")]; tensor<fp16, [1, 5120, 1, 1]> input_57_cast_fp16 = gelu(mode = input_57_mode_0, x = input_55_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")]; tensor<string, []> pretrained_out_39_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_39_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_39_strides_0 = const()[name = tensor<string, []>("pretrained_out_39_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_39_pad_0 = const()[name = tensor<string, []>("pretrained_out_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_39_dilations_0 = const()[name = tensor<string, []>("pretrained_out_39_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_39_groups_0 = const()[name = tensor<string, []>("pretrained_out_39_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158884288))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162161152))), name = tensor<string, []>("layers_1_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])]; tensor<fp16, [1280]> layers_1_fc2_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162161280)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_39_cast_fp16 = conv(bias = layers_1_fc2_pretrained_bias_to_fp16, dilations = pretrained_out_39_dilations_0, groups = pretrained_out_39_groups_0, pad = pretrained_out_39_pad_0, pad_type = pretrained_out_39_pad_type_0, strides = pretrained_out_39_strides_0, weight = layers_1_fc2_pretrained_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = tensor<string, []>("pretrained_out_39_cast_fp16")]; tensor<string, []> input_59_pad_type_0 = const()[name = tensor<string, []>("input_59_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_59_strides_0 = const()[name = tensor<string, []>("input_59_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_59_pad_0 = const()[name = tensor<string, []>("input_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_59_dilations_0 = const()[name = tensor<string, []>("input_59_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_59_groups_0 = const()[name = tensor<string, []>("input_59_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 5120, 1, 1]> layers_1_fc2_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_loraA_weight_to_fp16"), val = tensor<fp16, [16, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162163904)))]; tensor<fp16, [1, 16, 1, 1]> input_59_cast_fp16 = conv(dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = layers_1_fc2_loraA_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")]; tensor<string, []> lora_out_39_pad_type_0 = const()[name = tensor<string, []>("lora_out_39_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_39_strides_0 = const()[name = tensor<string, []>("lora_out_39_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_39_pad_0 = const()[name = tensor<string, []>("lora_out_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_39_dilations_0 = const()[name = tensor<string, []>("lora_out_39_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_39_groups_0 = const()[name = tensor<string, []>("lora_out_39_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_1_fc2_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_fc2_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162327808)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_39_cast_fp16 = conv(dilations = lora_out_39_dilations_0, groups = lora_out_39_groups_0, pad = lora_out_39_pad_0, pad_type = lora_out_39_pad_type_0, strides = lora_out_39_strides_0, weight = layers_1_fc2_loraB_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("lora_out_39_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> hidden_states_5_cast_fp16 = add(x = pretrained_out_39_cast_fp16, y = lora_out_39_cast_fp16)[name = tensor<string, []>("hidden_states_5_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor<string, []>("inputs_13_cast_fp16")]; tensor<int32, []> var_812 = const()[name = tensor<string, []>("op_812"), val = tensor<int32, []>(3)]; tensor<int32, [1]> out_13_axes_0 = const()[name = tensor<string, []>("out_13_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_838_to_fp16 = const()[name = tensor<string, []>("op_838_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_838_to_fp16, x = inputs_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")]; tensor<fp16, [1280]> obj_29_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_29_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162368832)))]; tensor<fp16, [1280]> obj_29_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_29_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162371456)))]; tensor<fp16, []> obj_29_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_29_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor<string, []>("obj_29_cast_fp16")]; tensor<string, []> pretrained_out_41_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_41_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_41_strides_0 = const()[name = tensor<string, []>("pretrained_out_41_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_41_pad_0 = const()[name = tensor<string, []>("pretrained_out_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_41_dilations_0 = const()[name = tensor<string, []>("pretrained_out_41_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_41_groups_0 = const()[name = tensor<string, []>("pretrained_out_41_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162374080))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163193344))), name = tensor<string, []>("layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_2_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163193472)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_41_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_41_dilations_0, groups = pretrained_out_41_groups_0, pad = pretrained_out_41_pad_0, pad_type = pretrained_out_41_pad_type_0, strides = pretrained_out_41_strides_0, weight = layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("pretrained_out_41_cast_fp16")]; tensor<string, []> input_61_pad_type_0 = const()[name = tensor<string, []>("input_61_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_61_strides_0 = const()[name = tensor<string, []>("input_61_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_61_pad_0 = const()[name = tensor<string, []>("input_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_61_dilations_0 = const()[name = tensor<string, []>("input_61_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_61_groups_0 = const()[name = tensor<string, []>("input_61_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_2_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163196096)))]; tensor<fp16, [1, 16, 1, 1]> input_61_cast_fp16 = conv(dilations = input_61_dilations_0, groups = input_61_groups_0, pad = input_61_pad_0, pad_type = input_61_pad_type_0, strides = input_61_strides_0, weight = layers_2_self_attn_q_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")]; tensor<string, []> lora_out_41_pad_type_0 = const()[name = tensor<string, []>("lora_out_41_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_41_strides_0 = const()[name = tensor<string, []>("lora_out_41_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_41_pad_0 = const()[name = tensor<string, []>("lora_out_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_41_dilations_0 = const()[name = tensor<string, []>("lora_out_41_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_41_groups_0 = const()[name = tensor<string, []>("lora_out_41_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_2_self_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163237120)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_41_cast_fp16 = conv(dilations = lora_out_41_dilations_0, groups = lora_out_41_groups_0, pad = lora_out_41_pad_0, pad_type = lora_out_41_pad_type_0, strides = lora_out_41_strides_0, weight = layers_2_self_attn_q_proj_loraB_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("lora_out_41_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> query_9_cast_fp16 = add(x = pretrained_out_41_cast_fp16, y = lora_out_41_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")]; tensor<string, []> pretrained_out_43_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_43_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_43_strides_0 = const()[name = tensor<string, []>("pretrained_out_43_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_43_pad_0 = const()[name = tensor<string, []>("pretrained_out_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_43_dilations_0 = const()[name = tensor<string, []>("pretrained_out_43_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_43_groups_0 = const()[name = tensor<string, []>("pretrained_out_43_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163278144))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164097408))), name = tensor<string, []>("layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_43_cast_fp16 = conv(dilations = pretrained_out_43_dilations_0, groups = pretrained_out_43_groups_0, pad = pretrained_out_43_pad_0, pad_type = pretrained_out_43_pad_type_0, strides = pretrained_out_43_strides_0, weight = layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("pretrained_out_43_cast_fp16")]; tensor<string, []> input_63_pad_type_0 = const()[name = tensor<string, []>("input_63_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_63_strides_0 = const()[name = tensor<string, []>("input_63_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_63_pad_0 = const()[name = tensor<string, []>("input_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_63_dilations_0 = const()[name = tensor<string, []>("input_63_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_63_groups_0 = const()[name = tensor<string, []>("input_63_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_2_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164097536)))]; tensor<fp16, [1, 16, 1, 1]> input_63_cast_fp16 = conv(dilations = input_63_dilations_0, groups = input_63_groups_0, pad = input_63_pad_0, pad_type = input_63_pad_type_0, strides = input_63_strides_0, weight = layers_2_self_attn_k_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")]; tensor<string, []> lora_out_43_pad_type_0 = const()[name = tensor<string, []>("lora_out_43_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_43_strides_0 = const()[name = tensor<string, []>("lora_out_43_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_43_pad_0 = const()[name = tensor<string, []>("lora_out_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_43_dilations_0 = const()[name = tensor<string, []>("lora_out_43_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_43_groups_0 = const()[name = tensor<string, []>("lora_out_43_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_2_self_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164138560)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_43_cast_fp16 = conv(dilations = lora_out_43_dilations_0, groups = lora_out_43_groups_0, pad = lora_out_43_pad_0, pad_type = lora_out_43_pad_type_0, strides = lora_out_43_strides_0, weight = layers_2_self_attn_k_proj_loraB_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("lora_out_43_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> current_key_5_cast_fp16 = add(x = pretrained_out_43_cast_fp16, y = lora_out_43_cast_fp16)[name = tensor<string, []>("current_key_5_cast_fp16")]; tensor<string, []> pretrained_out_45_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_45_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_45_strides_0 = const()[name = tensor<string, []>("pretrained_out_45_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_45_pad_0 = const()[name = tensor<string, []>("pretrained_out_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_45_dilations_0 = const()[name = tensor<string, []>("pretrained_out_45_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_45_groups_0 = const()[name = tensor<string, []>("pretrained_out_45_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164179584))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164998848))), name = tensor<string, []>("layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_2_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164998976)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_45_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_45_dilations_0, groups = pretrained_out_45_groups_0, pad = pretrained_out_45_pad_0, pad_type = pretrained_out_45_pad_type_0, strides = pretrained_out_45_strides_0, weight = layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor<string, []>("pretrained_out_45_cast_fp16")]; tensor<string, []> input_65_pad_type_0 = const()[name = tensor<string, []>("input_65_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_65_strides_0 = const()[name = tensor<string, []>("input_65_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_65_pad_0 = const()[name = tensor<string, []>("input_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_65_dilations_0 = const()[name = tensor<string, []>("input_65_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_65_groups_0 = const()[name = tensor<string, []>("input_65_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_2_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165001600)))]; tensor<fp16, [1, 16, 1, 1]> input_65_cast_fp16 = conv(dilations = input_65_dilations_0, groups = input_65_groups_0, pad = input_65_pad_0, pad_type = input_65_pad_type_0, strides = input_65_strides_0, weight = layers_2_self_attn_v_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")]; tensor<string, []> lora_out_45_pad_type_0 = const()[name = tensor<string, []>("lora_out_45_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_45_strides_0 = const()[name = tensor<string, []>("lora_out_45_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_45_pad_0 = const()[name = tensor<string, []>("lora_out_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_45_dilations_0 = const()[name = tensor<string, []>("lora_out_45_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_45_groups_0 = const()[name = tensor<string, []>("lora_out_45_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_2_self_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165042624)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_45_cast_fp16 = conv(dilations = lora_out_45_dilations_0, groups = lora_out_45_groups_0, pad = lora_out_45_pad_0, pad_type = lora_out_45_pad_type_0, strides = lora_out_45_strides_0, weight = layers_2_self_attn_v_proj_loraB_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("lora_out_45_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> current_value_5_cast_fp16 = add(x = pretrained_out_45_cast_fp16, y = lora_out_45_cast_fp16)[name = tensor<string, []>("current_value_5_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_924_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_924_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_926_cast_fp16 = mul(x = var_47_cast_fp16_2, y = var_177_cast_fp16)[name = tensor<string, []>("op_926_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> key_9_cast_fp16 = add(x = var_924_cast_fp16, y = var_926_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_928_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_928_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_930_cast_fp16 = mul(x = var_54_cast_fp16_2, y = var_177_cast_fp16)[name = tensor<string, []>("op_930_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> value_9_cast_fp16 = add(x = var_928_cast_fp16, y = var_930_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")]; tensor<int32, [4]> var_933 = const()[name = tensor<string, []>("op_933"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1]> mh_q_9_cast_fp16 = reshape(shape = var_933, x = query_9_cast_fp16)[name = tensor<string, []>("mh_q_9_cast_fp16")]; tensor<fp16, []> var_935_to_fp16 = const()[name = tensor<string, []>("op_935_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; tensor<fp16, [1, 20, 64, 1]> var_936_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_935_to_fp16)[name = tensor<string, []>("op_936_cast_fp16")]; tensor<int32, [4]> var_937 = const()[name = tensor<string, []>("op_937"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 448]> var_938_cast_fp16 = reshape(shape = var_937, x = key_9_cast_fp16)[name = tensor<string, []>("op_938_cast_fp16")]; tensor<bool, []> mh_w_13_transpose_x_0 = const()[name = tensor<string, []>("mh_w_13_transpose_x_0"), val = tensor<bool, []>(true)]; tensor<bool, []> mh_w_13_transpose_y_0 = const()[name = tensor<string, []>("mh_w_13_transpose_y_0"), val = tensor<bool, []>(false)]; tensor<fp16, [1, 20, 1, 448]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_936_cast_fp16, y = var_938_cast_fp16)[name = tensor<string, []>("mh_w_13_cast_fp16")]; tensor<fp16, [1, 20, 1, 448]> mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_195_cast_fp16)[name = tensor<string, []>("mh_w_15_cast_fp16")]; tensor<fp16, [1, 20, 1, 448]> var_946_cast_fp16 = softmax(axis = var_812, x = mh_w_15_cast_fp16)[name = tensor<string, []>("op_946_cast_fp16")]; tensor<int32, [4]> var_947 = const()[name = tensor<string, []>("op_947"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 448]> var_948_cast_fp16 = reshape(shape = var_947, x = value_9_cast_fp16)[name = tensor<string, []>("op_948_cast_fp16")]; tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)]; tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(true)]; tensor<fp16, [1, 20, 64, 1]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_948_cast_fp16, y = var_946_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")]; tensor<int32, [4]> var_951 = const()[name = tensor<string, []>("op_951"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; tensor<fp16, [1, 1280, 1, 1]> input_67_cast_fp16 = reshape(shape = var_951, x = attn_9_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")]; tensor<string, []> pretrained_out_47_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_47_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_47_strides_0 = const()[name = tensor<string, []>("pretrained_out_47_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_47_pad_0 = const()[name = tensor<string, []>("pretrained_out_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_47_dilations_0 = const()[name = tensor<string, []>("pretrained_out_47_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_47_groups_0 = const()[name = tensor<string, []>("pretrained_out_47_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165083648))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165902912))), name = tensor<string, []>("layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_2_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165903040)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_47_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_47_dilations_0, groups = pretrained_out_47_groups_0, pad = pretrained_out_47_pad_0, pad_type = pretrained_out_47_pad_type_0, strides = pretrained_out_47_strides_0, weight = layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = tensor<string, []>("pretrained_out_47_cast_fp16")]; tensor<string, []> input_69_pad_type_0 = const()[name = tensor<string, []>("input_69_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_69_strides_0 = const()[name = tensor<string, []>("input_69_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_69_pad_0 = const()[name = tensor<string, []>("input_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_69_dilations_0 = const()[name = tensor<string, []>("input_69_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_69_groups_0 = const()[name = tensor<string, []>("input_69_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_2_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165905664)))]; tensor<fp16, [1, 16, 1, 1]> input_69_cast_fp16 = conv(dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = layers_2_self_attn_o_proj_loraA_weight_to_fp16, x = input_67_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")]; tensor<string, []> lora_out_47_pad_type_0 = const()[name = tensor<string, []>("lora_out_47_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_47_strides_0 = const()[name = tensor<string, []>("lora_out_47_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_47_pad_0 = const()[name = tensor<string, []>("lora_out_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_47_dilations_0 = const()[name = tensor<string, []>("lora_out_47_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_47_groups_0 = const()[name = tensor<string, []>("lora_out_47_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_2_self_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165946688)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_47_cast_fp16 = conv(dilations = lora_out_47_dilations_0, groups = lora_out_47_groups_0, pad = lora_out_47_pad_0, pad_type = lora_out_47_pad_type_0, strides = lora_out_47_strides_0, weight = layers_2_self_attn_o_proj_loraB_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("lora_out_47_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> obj_35_cast_fp16 = add(x = pretrained_out_47_cast_fp16, y = lora_out_47_cast_fp16)[name = tensor<string, []>("obj_35_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor<string, []>("inputs_15_cast_fp16")]; tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_989_to_fp16 = const()[name = tensor<string, []>("op_989_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_989_to_fp16, x = inputs_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")]; tensor<fp16, [1280]> obj_37_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_37_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165987712)))]; tensor<fp16, [1280]> obj_37_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_37_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165990336)))]; tensor<fp16, []> obj_37_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_37_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor<string, []>("obj_37_cast_fp16")]; tensor<string, []> pretrained_out_49_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_49_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_49_strides_0 = const()[name = tensor<string, []>("pretrained_out_49_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_49_pad_0 = const()[name = tensor<string, []>("pretrained_out_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_49_dilations_0 = const()[name = tensor<string, []>("pretrained_out_49_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_49_groups_0 = const()[name = tensor<string, []>("pretrained_out_49_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165992960))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166812224))), name = tensor<string, []>("layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166812352)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_49_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_49_dilations_0, groups = pretrained_out_49_groups_0, pad = pretrained_out_49_pad_0, pad_type = pretrained_out_49_pad_type_0, strides = pretrained_out_49_strides_0, weight = layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor<string, []>("pretrained_out_49_cast_fp16")]; tensor<string, []> input_71_pad_type_0 = const()[name = tensor<string, []>("input_71_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_71_strides_0 = const()[name = tensor<string, []>("input_71_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_71_pad_0 = const()[name = tensor<string, []>("input_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_71_dilations_0 = const()[name = tensor<string, []>("input_71_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_71_groups_0 = const()[name = tensor<string, []>("input_71_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_2_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166814976)))]; tensor<fp16, [1, 16, 1, 1]> input_71_cast_fp16 = conv(dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = layers_2_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")]; tensor<string, []> lora_out_49_pad_type_0 = const()[name = tensor<string, []>("lora_out_49_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_49_strides_0 = const()[name = tensor<string, []>("lora_out_49_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_49_pad_0 = const()[name = tensor<string, []>("lora_out_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_49_dilations_0 = const()[name = tensor<string, []>("lora_out_49_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_49_groups_0 = const()[name = tensor<string, []>("lora_out_49_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_2_encoder_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166856000)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_49_cast_fp16 = conv(dilations = lora_out_49_dilations_0, groups = lora_out_49_groups_0, pad = lora_out_49_pad_0, pad_type = lora_out_49_pad_type_0, strides = lora_out_49_strides_0, weight = layers_2_encoder_attn_q_proj_loraB_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("lora_out_49_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> query_11_cast_fp16 = add(x = pretrained_out_49_cast_fp16, y = lora_out_49_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")]; tensor<string, []> pretrained_out_51_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_51_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_51_strides_0 = const()[name = tensor<string, []>("pretrained_out_51_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_51_pad_0 = const()[name = tensor<string, []>("pretrained_out_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_51_dilations_0 = const()[name = tensor<string, []>("pretrained_out_51_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_51_groups_0 = const()[name = tensor<string, []>("pretrained_out_51_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166897024))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167716288))), name = tensor<string, []>("layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_51_cast_fp16 = conv(dilations = pretrained_out_51_dilations_0, groups = pretrained_out_51_groups_0, pad = pretrained_out_51_pad_0, pad_type = pretrained_out_51_pad_type_0, strides = pretrained_out_51_strides_0, weight = layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_51_cast_fp16")]; tensor<string, []> input_73_pad_type_0 = const()[name = tensor<string, []>("input_73_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_73_strides_0 = const()[name = tensor<string, []>("input_73_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_73_pad_0 = const()[name = tensor<string, []>("input_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_73_dilations_0 = const()[name = tensor<string, []>("input_73_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_73_groups_0 = const()[name = tensor<string, []>("input_73_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_2_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167716416)))]; tensor<fp16, [1, 16, 1, 1500]> input_73_cast_fp16 = conv(dilations = input_73_dilations_0, groups = input_73_groups_0, pad = input_73_pad_0, pad_type = input_73_pad_type_0, strides = input_73_strides_0, weight = layers_2_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_73_cast_fp16")]; tensor<string, []> lora_out_51_pad_type_0 = const()[name = tensor<string, []>("lora_out_51_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_51_strides_0 = const()[name = tensor<string, []>("lora_out_51_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_51_pad_0 = const()[name = tensor<string, []>("lora_out_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_51_dilations_0 = const()[name = tensor<string, []>("lora_out_51_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_51_groups_0 = const()[name = tensor<string, []>("lora_out_51_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_2_encoder_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167757440)))]; tensor<fp16, [1, 1280, 1, 1500]> lora_out_51_cast_fp16 = conv(dilations = lora_out_51_dilations_0, groups = lora_out_51_groups_0, pad = lora_out_51_pad_0, pad_type = lora_out_51_pad_type_0, strides = lora_out_51_strides_0, weight = layers_2_encoder_attn_k_proj_loraB_weight_to_fp16, x = input_73_cast_fp16)[name = tensor<string, []>("lora_out_51_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1500]> key_11_cast_fp16 = add(x = pretrained_out_51_cast_fp16, y = lora_out_51_cast_fp16)[name = tensor<string, []>("key_11_cast_fp16")]; tensor<string, []> pretrained_out_53_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_53_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_53_strides_0 = const()[name = tensor<string, []>("pretrained_out_53_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_53_pad_0 = const()[name = tensor<string, []>("pretrained_out_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_53_dilations_0 = const()[name = tensor<string, []>("pretrained_out_53_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_53_groups_0 = const()[name = tensor<string, []>("pretrained_out_53_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167798464))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168617728))), name = tensor<string, []>("layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168617856)))]; tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_53_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_53_dilations_0, groups = pretrained_out_53_groups_0, pad = pretrained_out_53_pad_0, pad_type = pretrained_out_53_pad_type_0, strides = pretrained_out_53_strides_0, weight = layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_53_cast_fp16")]; tensor<string, []> input_75_pad_type_0 = const()[name = tensor<string, []>("input_75_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_75_strides_0 = const()[name = tensor<string, []>("input_75_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_75_pad_0 = const()[name = tensor<string, []>("input_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_75_dilations_0 = const()[name = tensor<string, []>("input_75_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_75_groups_0 = const()[name = tensor<string, []>("input_75_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_2_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168620480)))]; tensor<fp16, [1, 16, 1, 1500]> input_75_cast_fp16 = conv(dilations = input_75_dilations_0, groups = input_75_groups_0, pad = input_75_pad_0, pad_type = input_75_pad_type_0, strides = input_75_strides_0, weight = layers_2_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_75_cast_fp16")]; tensor<string, []> lora_out_53_pad_type_0 = const()[name = tensor<string, []>("lora_out_53_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_53_strides_0 = const()[name = tensor<string, []>("lora_out_53_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_53_pad_0 = const()[name = tensor<string, []>("lora_out_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_53_dilations_0 = const()[name = tensor<string, []>("lora_out_53_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_53_groups_0 = const()[name = tensor<string, []>("lora_out_53_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_2_encoder_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168661504)))]; tensor<fp16, [1, 1280, 1, 1500]> lora_out_53_cast_fp16 = conv(dilations = lora_out_53_dilations_0, groups = lora_out_53_groups_0, pad = lora_out_53_pad_0, pad_type = lora_out_53_pad_type_0, strides = lora_out_53_strides_0, weight = layers_2_encoder_attn_v_proj_loraB_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("lora_out_53_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1500]> value_11_cast_fp16 = add(x = pretrained_out_53_cast_fp16, y = lora_out_53_cast_fp16)[name = tensor<string, []>("value_11_cast_fp16")]; tensor<int32, [4]> var_1072 = const()[name = tensor<string, []>("op_1072"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1]> mh_q_11_cast_fp16 = reshape(shape = var_1072, x = query_11_cast_fp16)[name = tensor<string, []>("mh_q_11_cast_fp16")]; tensor<fp16, []> var_1074_to_fp16 = const()[name = tensor<string, []>("op_1074_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; tensor<fp16, [1, 20, 64, 1]> var_1075_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_1074_to_fp16)[name = tensor<string, []>("op_1075_cast_fp16")]; tensor<int32, [4]> var_1076 = const()[name = tensor<string, []>("op_1076"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1500]> var_1077_cast_fp16 = reshape(shape = var_1076, x = key_11_cast_fp16)[name = tensor<string, []>("op_1077_cast_fp16")]; tensor<bool, []> mh_w_17_transpose_x_0 = const()[name = tensor<string, []>("mh_w_17_transpose_x_0"), val = tensor<bool, []>(true)]; tensor<bool, []> mh_w_17_transpose_y_0 = const()[name = tensor<string, []>("mh_w_17_transpose_y_0"), val = tensor<bool, []>(false)]; tensor<fp16, [1, 20, 1, 1500]> mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_1075_cast_fp16, y = var_1077_cast_fp16)[name = tensor<string, []>("mh_w_17_cast_fp16")]; tensor<fp16, [1, 20, 1, 1500]> obj_41_cast_fp16 = softmax(axis = var_812, x = mh_w_17_cast_fp16)[name = tensor<string, []>("obj_41_cast_fp16")]; tensor<int32, [4]> var_1081 = const()[name = tensor<string, []>("op_1081"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1500]> var_1082_cast_fp16 = reshape(shape = var_1081, x = value_11_cast_fp16)[name = tensor<string, []>("op_1082_cast_fp16")]; tensor<bool, []> attn_11_transpose_x_0 = const()[name = tensor<string, []>("attn_11_transpose_x_0"), val = tensor<bool, []>(false)]; tensor<bool, []> attn_11_transpose_y_0 = const()[name = tensor<string, []>("attn_11_transpose_y_0"), val = tensor<bool, []>(true)]; tensor<fp16, [1, 20, 64, 1]> attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_1082_cast_fp16, y = obj_41_cast_fp16)[name = tensor<string, []>("attn_11_cast_fp16")]; tensor<int32, [4]> var_1085 = const()[name = tensor<string, []>("op_1085"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; tensor<fp16, [1, 1280, 1, 1]> input_77_cast_fp16 = reshape(shape = var_1085, x = attn_11_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")]; tensor<string, []> pretrained_out_55_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_55_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_55_strides_0 = const()[name = tensor<string, []>("pretrained_out_55_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_55_pad_0 = const()[name = tensor<string, []>("pretrained_out_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_55_dilations_0 = const()[name = tensor<string, []>("pretrained_out_55_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_55_groups_0 = const()[name = tensor<string, []>("pretrained_out_55_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(168702528))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169521792))), name = tensor<string, []>("layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169521920)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_55_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_55_dilations_0, groups = pretrained_out_55_groups_0, pad = pretrained_out_55_pad_0, pad_type = pretrained_out_55_pad_type_0, strides = pretrained_out_55_strides_0, weight = layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_77_cast_fp16)[name = tensor<string, []>("pretrained_out_55_cast_fp16")]; tensor<string, []> input_79_pad_type_0 = const()[name = tensor<string, []>("input_79_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_79_strides_0 = const()[name = tensor<string, []>("input_79_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_79_pad_0 = const()[name = tensor<string, []>("input_79_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_79_dilations_0 = const()[name = tensor<string, []>("input_79_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_79_groups_0 = const()[name = tensor<string, []>("input_79_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_2_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169524544)))]; tensor<fp16, [1, 16, 1, 1]> input_79_cast_fp16 = conv(dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = layers_2_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")]; tensor<string, []> lora_out_55_pad_type_0 = const()[name = tensor<string, []>("lora_out_55_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_55_strides_0 = const()[name = tensor<string, []>("lora_out_55_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_55_pad_0 = const()[name = tensor<string, []>("lora_out_55_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_55_dilations_0 = const()[name = tensor<string, []>("lora_out_55_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_55_groups_0 = const()[name = tensor<string, []>("lora_out_55_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_2_encoder_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_encoder_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169565568)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_55_cast_fp16 = conv(dilations = lora_out_55_dilations_0, groups = lora_out_55_groups_0, pad = lora_out_55_pad_0, pad_type = lora_out_55_pad_type_0, strides = lora_out_55_strides_0, weight = layers_2_encoder_attn_o_proj_loraB_weight_to_fp16, x = input_79_cast_fp16)[name = tensor<string, []>("lora_out_55_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> obj_39_cast_fp16 = add(x = pretrained_out_55_cast_fp16, y = lora_out_55_cast_fp16)[name = tensor<string, []>("obj_39_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor<string, []>("inputs_17_cast_fp16")]; tensor<int32, [1]> out_17_axes_0 = const()[name = tensor<string, []>("out_17_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_1122_to_fp16 = const()[name = tensor<string, []>("op_1122_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1122_to_fp16, x = inputs_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")]; tensor<fp16, [1280]> input_81_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_81_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169606592)))]; tensor<fp16, [1280]> input_81_beta_0_to_fp16 = const()[name = tensor<string, []>("input_81_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169609216)))]; tensor<fp16, []> input_81_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_81_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> input_81_cast_fp16 = batch_norm(beta = input_81_beta_0_to_fp16, epsilon = input_81_epsilon_0_to_fp16, gamma = input_81_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")]; tensor<string, []> pretrained_out_57_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_57_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_57_strides_0 = const()[name = tensor<string, []>("pretrained_out_57_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_57_pad_0 = const()[name = tensor<string, []>("pretrained_out_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_57_dilations_0 = const()[name = tensor<string, []>("pretrained_out_57_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_57_groups_0 = const()[name = tensor<string, []>("pretrained_out_57_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [5120, 1280, 1, 1]> layers_2_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(169611840))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172888704))), name = tensor<string, []>("layers_2_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])]; tensor<fp16, [5120]> layers_2_fc1_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_pretrained_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172888832)))]; tensor<fp16, [1, 5120, 1, 1]> pretrained_out_57_cast_fp16 = conv(bias = layers_2_fc1_pretrained_bias_to_fp16, dilations = pretrained_out_57_dilations_0, groups = pretrained_out_57_groups_0, pad = pretrained_out_57_pad_0, pad_type = pretrained_out_57_pad_type_0, strides = pretrained_out_57_strides_0, weight = layers_2_fc1_pretrained_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = tensor<string, []>("pretrained_out_57_cast_fp16")]; tensor<string, []> input_83_pad_type_0 = const()[name = tensor<string, []>("input_83_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_83_strides_0 = const()[name = tensor<string, []>("input_83_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_83_pad_0 = const()[name = tensor<string, []>("input_83_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_83_dilations_0 = const()[name = tensor<string, []>("input_83_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_83_groups_0 = const()[name = tensor<string, []>("input_83_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_2_fc1_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172899136)))]; tensor<fp16, [1, 16, 1, 1]> input_83_cast_fp16 = conv(dilations = input_83_dilations_0, groups = input_83_groups_0, pad = input_83_pad_0, pad_type = input_83_pad_type_0, strides = input_83_strides_0, weight = layers_2_fc1_loraA_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")]; tensor<string, []> lora_out_57_pad_type_0 = const()[name = tensor<string, []>("lora_out_57_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_57_strides_0 = const()[name = tensor<string, []>("lora_out_57_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_57_pad_0 = const()[name = tensor<string, []>("lora_out_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_57_dilations_0 = const()[name = tensor<string, []>("lora_out_57_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_57_groups_0 = const()[name = tensor<string, []>("lora_out_57_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [5120, 16, 1, 1]> layers_2_fc1_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc1_loraB_weight_to_fp16"), val = tensor<fp16, [5120, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(172940160)))]; tensor<fp16, [1, 5120, 1, 1]> lora_out_57_cast_fp16 = conv(dilations = lora_out_57_dilations_0, groups = lora_out_57_groups_0, pad = lora_out_57_pad_0, pad_type = lora_out_57_pad_type_0, strides = lora_out_57_strides_0, weight = layers_2_fc1_loraB_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("lora_out_57_cast_fp16")]; tensor<fp16, [1, 5120, 1, 1]> input_85_cast_fp16 = add(x = pretrained_out_57_cast_fp16, y = lora_out_57_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")]; tensor<string, []> input_87_mode_0 = const()[name = tensor<string, []>("input_87_mode_0"), val = tensor<string, []>("EXACT")]; tensor<fp16, [1, 5120, 1, 1]> input_87_cast_fp16 = gelu(mode = input_87_mode_0, x = input_85_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")]; tensor<string, []> pretrained_out_59_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_59_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_59_strides_0 = const()[name = tensor<string, []>("pretrained_out_59_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_59_pad_0 = const()[name = tensor<string, []>("pretrained_out_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_59_dilations_0 = const()[name = tensor<string, []>("pretrained_out_59_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_59_groups_0 = const()[name = tensor<string, []>("pretrained_out_59_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 5120, 1, 1]> layers_2_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(173104064))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176380928))), name = tensor<string, []>("layers_2_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])]; tensor<fp16, [1280]> layers_2_fc2_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176381056)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_59_cast_fp16 = conv(bias = layers_2_fc2_pretrained_bias_to_fp16, dilations = pretrained_out_59_dilations_0, groups = pretrained_out_59_groups_0, pad = pretrained_out_59_pad_0, pad_type = pretrained_out_59_pad_type_0, strides = pretrained_out_59_strides_0, weight = layers_2_fc2_pretrained_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = tensor<string, []>("pretrained_out_59_cast_fp16")]; tensor<string, []> input_89_pad_type_0 = const()[name = tensor<string, []>("input_89_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_89_strides_0 = const()[name = tensor<string, []>("input_89_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_89_pad_0 = const()[name = tensor<string, []>("input_89_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_89_dilations_0 = const()[name = tensor<string, []>("input_89_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_89_groups_0 = const()[name = tensor<string, []>("input_89_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 5120, 1, 1]> layers_2_fc2_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_loraA_weight_to_fp16"), val = tensor<fp16, [16, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176383680)))]; tensor<fp16, [1, 16, 1, 1]> input_89_cast_fp16 = conv(dilations = input_89_dilations_0, groups = input_89_groups_0, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = input_89_strides_0, weight = layers_2_fc2_loraA_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("input_89_cast_fp16")]; tensor<string, []> lora_out_59_pad_type_0 = const()[name = tensor<string, []>("lora_out_59_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_59_strides_0 = const()[name = tensor<string, []>("lora_out_59_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_59_pad_0 = const()[name = tensor<string, []>("lora_out_59_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_59_dilations_0 = const()[name = tensor<string, []>("lora_out_59_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_59_groups_0 = const()[name = tensor<string, []>("lora_out_59_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_2_fc2_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_fc2_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176547584)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_59_cast_fp16 = conv(dilations = lora_out_59_dilations_0, groups = lora_out_59_groups_0, pad = lora_out_59_pad_0, pad_type = lora_out_59_pad_type_0, strides = lora_out_59_strides_0, weight = layers_2_fc2_loraB_weight_to_fp16, x = input_89_cast_fp16)[name = tensor<string, []>("lora_out_59_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> hidden_states_7_cast_fp16 = add(x = pretrained_out_59_cast_fp16, y = lora_out_59_cast_fp16)[name = tensor<string, []>("hidden_states_7_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor<string, []>("inputs_19_cast_fp16")]; tensor<int32, []> var_1190 = const()[name = tensor<string, []>("op_1190"), val = tensor<int32, []>(3)]; tensor<int32, [1]> out_19_axes_0 = const()[name = tensor<string, []>("out_19_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_1216_to_fp16 = const()[name = tensor<string, []>("op_1216_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1216_to_fp16, x = inputs_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")]; tensor<fp16, [1280]> obj_43_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_43_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176588608)))]; tensor<fp16, [1280]> obj_43_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_43_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176591232)))]; tensor<fp16, []> obj_43_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_43_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor<string, []>("obj_43_cast_fp16")]; tensor<string, []> pretrained_out_61_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_61_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_61_strides_0 = const()[name = tensor<string, []>("pretrained_out_61_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_61_pad_0 = const()[name = tensor<string, []>("pretrained_out_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_61_dilations_0 = const()[name = tensor<string, []>("pretrained_out_61_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_61_groups_0 = const()[name = tensor<string, []>("pretrained_out_61_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(176593856))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177413120))), name = tensor<string, []>("layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_3_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177413248)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_61_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_61_dilations_0, groups = pretrained_out_61_groups_0, pad = pretrained_out_61_pad_0, pad_type = pretrained_out_61_pad_type_0, strides = pretrained_out_61_strides_0, weight = layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor<string, []>("pretrained_out_61_cast_fp16")]; tensor<string, []> input_91_pad_type_0 = const()[name = tensor<string, []>("input_91_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_91_strides_0 = const()[name = tensor<string, []>("input_91_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_91_pad_0 = const()[name = tensor<string, []>("input_91_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_91_dilations_0 = const()[name = tensor<string, []>("input_91_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_91_groups_0 = const()[name = tensor<string, []>("input_91_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_3_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177415872)))]; tensor<fp16, [1, 16, 1, 1]> input_91_cast_fp16 = conv(dilations = input_91_dilations_0, groups = input_91_groups_0, pad = input_91_pad_0, pad_type = input_91_pad_type_0, strides = input_91_strides_0, weight = layers_3_self_attn_q_proj_loraA_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")]; tensor<string, []> lora_out_61_pad_type_0 = const()[name = tensor<string, []>("lora_out_61_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_61_strides_0 = const()[name = tensor<string, []>("lora_out_61_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_61_pad_0 = const()[name = tensor<string, []>("lora_out_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_61_dilations_0 = const()[name = tensor<string, []>("lora_out_61_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_61_groups_0 = const()[name = tensor<string, []>("lora_out_61_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_3_self_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177456896)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_61_cast_fp16 = conv(dilations = lora_out_61_dilations_0, groups = lora_out_61_groups_0, pad = lora_out_61_pad_0, pad_type = lora_out_61_pad_type_0, strides = lora_out_61_strides_0, weight = layers_3_self_attn_q_proj_loraB_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("lora_out_61_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> query_13_cast_fp16 = add(x = pretrained_out_61_cast_fp16, y = lora_out_61_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")]; tensor<string, []> pretrained_out_63_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_63_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_63_strides_0 = const()[name = tensor<string, []>("pretrained_out_63_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_63_pad_0 = const()[name = tensor<string, []>("pretrained_out_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_63_dilations_0 = const()[name = tensor<string, []>("pretrained_out_63_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_63_groups_0 = const()[name = tensor<string, []>("pretrained_out_63_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(177497920))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178317184))), name = tensor<string, []>("layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_63_cast_fp16 = conv(dilations = pretrained_out_63_dilations_0, groups = pretrained_out_63_groups_0, pad = pretrained_out_63_pad_0, pad_type = pretrained_out_63_pad_type_0, strides = pretrained_out_63_strides_0, weight = layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor<string, []>("pretrained_out_63_cast_fp16")]; tensor<string, []> input_93_pad_type_0 = const()[name = tensor<string, []>("input_93_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_93_strides_0 = const()[name = tensor<string, []>("input_93_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_93_pad_0 = const()[name = tensor<string, []>("input_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_93_dilations_0 = const()[name = tensor<string, []>("input_93_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_93_groups_0 = const()[name = tensor<string, []>("input_93_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_3_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178317312)))]; tensor<fp16, [1, 16, 1, 1]> input_93_cast_fp16 = conv(dilations = input_93_dilations_0, groups = input_93_groups_0, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = input_93_strides_0, weight = layers_3_self_attn_k_proj_loraA_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("input_93_cast_fp16")]; tensor<string, []> lora_out_63_pad_type_0 = const()[name = tensor<string, []>("lora_out_63_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_63_strides_0 = const()[name = tensor<string, []>("lora_out_63_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_63_pad_0 = const()[name = tensor<string, []>("lora_out_63_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_63_dilations_0 = const()[name = tensor<string, []>("lora_out_63_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_63_groups_0 = const()[name = tensor<string, []>("lora_out_63_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_3_self_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178358336)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_63_cast_fp16 = conv(dilations = lora_out_63_dilations_0, groups = lora_out_63_groups_0, pad = lora_out_63_pad_0, pad_type = lora_out_63_pad_type_0, strides = lora_out_63_strides_0, weight = layers_3_self_attn_k_proj_loraB_weight_to_fp16, x = input_93_cast_fp16)[name = tensor<string, []>("lora_out_63_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> current_key_cast_fp16 = add(x = pretrained_out_63_cast_fp16, y = lora_out_63_cast_fp16)[name = tensor<string, []>("current_key_cast_fp16")]; tensor<string, []> pretrained_out_65_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_65_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_65_strides_0 = const()[name = tensor<string, []>("pretrained_out_65_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_65_pad_0 = const()[name = tensor<string, []>("pretrained_out_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_65_dilations_0 = const()[name = tensor<string, []>("pretrained_out_65_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_65_groups_0 = const()[name = tensor<string, []>("pretrained_out_65_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178399360))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179218624))), name = tensor<string, []>("layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_3_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179218752)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_65_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_65_dilations_0, groups = pretrained_out_65_groups_0, pad = pretrained_out_65_pad_0, pad_type = pretrained_out_65_pad_type_0, strides = pretrained_out_65_strides_0, weight = layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor<string, []>("pretrained_out_65_cast_fp16")]; tensor<string, []> input_95_pad_type_0 = const()[name = tensor<string, []>("input_95_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_95_strides_0 = const()[name = tensor<string, []>("input_95_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_95_pad_0 = const()[name = tensor<string, []>("input_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_95_dilations_0 = const()[name = tensor<string, []>("input_95_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_95_groups_0 = const()[name = tensor<string, []>("input_95_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_3_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179221376)))]; tensor<fp16, [1, 16, 1, 1]> input_95_cast_fp16 = conv(dilations = input_95_dilations_0, groups = input_95_groups_0, pad = input_95_pad_0, pad_type = input_95_pad_type_0, strides = input_95_strides_0, weight = layers_3_self_attn_v_proj_loraA_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")]; tensor<string, []> lora_out_65_pad_type_0 = const()[name = tensor<string, []>("lora_out_65_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_65_strides_0 = const()[name = tensor<string, []>("lora_out_65_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_65_pad_0 = const()[name = tensor<string, []>("lora_out_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_65_dilations_0 = const()[name = tensor<string, []>("lora_out_65_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_65_groups_0 = const()[name = tensor<string, []>("lora_out_65_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_3_self_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179262400)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_65_cast_fp16 = conv(dilations = lora_out_65_dilations_0, groups = lora_out_65_groups_0, pad = lora_out_65_pad_0, pad_type = lora_out_65_pad_type_0, strides = lora_out_65_strides_0, weight = layers_3_self_attn_v_proj_loraB_weight_to_fp16, x = input_95_cast_fp16)[name = tensor<string, []>("lora_out_65_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> current_value_cast_fp16 = add(x = pretrained_out_65_cast_fp16, y = lora_out_65_cast_fp16)[name = tensor<string, []>("current_value_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_1302_cast_fp16 = mul(x = current_key_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_1302_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_1304_cast_fp16 = mul(x = var_47_cast_fp16_3, y = var_177_cast_fp16)[name = tensor<string, []>("op_1304_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> key_13_cast_fp16 = add(x = var_1302_cast_fp16, y = var_1304_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_1306_cast_fp16 = mul(x = current_value_cast_fp16, y = var_174_cast_fp16)[name = tensor<string, []>("op_1306_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> var_1308_cast_fp16 = mul(x = var_54_cast_fp16_3, y = var_177_cast_fp16)[name = tensor<string, []>("op_1308_cast_fp16")]; tensor<fp16, [1, 1280, 1, 448]> value_13_cast_fp16 = add(x = var_1306_cast_fp16, y = var_1308_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")]; tensor<int32, [4]> var_1311 = const()[name = tensor<string, []>("op_1311"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1]> mh_q_13_cast_fp16 = reshape(shape = var_1311, x = query_13_cast_fp16)[name = tensor<string, []>("mh_q_13_cast_fp16")]; tensor<fp16, []> var_1313_to_fp16 = const()[name = tensor<string, []>("op_1313_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; tensor<fp16, [1, 20, 64, 1]> var_1314_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_1313_to_fp16)[name = tensor<string, []>("op_1314_cast_fp16")]; tensor<int32, [4]> var_1315 = const()[name = tensor<string, []>("op_1315"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 448]> var_1316_cast_fp16 = reshape(shape = var_1315, x = key_13_cast_fp16)[name = tensor<string, []>("op_1316_cast_fp16")]; tensor<bool, []> mh_w_19_transpose_x_0 = const()[name = tensor<string, []>("mh_w_19_transpose_x_0"), val = tensor<bool, []>(true)]; tensor<bool, []> mh_w_19_transpose_y_0 = const()[name = tensor<string, []>("mh_w_19_transpose_y_0"), val = tensor<bool, []>(false)]; tensor<fp16, [1, 20, 1, 448]> mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_1314_cast_fp16, y = var_1316_cast_fp16)[name = tensor<string, []>("mh_w_19_cast_fp16")]; tensor<fp16, [1, 20, 1, 448]> mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_195_cast_fp16)[name = tensor<string, []>("mh_w_21_cast_fp16")]; tensor<fp16, [1, 20, 1, 448]> var_1324_cast_fp16 = softmax(axis = var_1190, x = mh_w_21_cast_fp16)[name = tensor<string, []>("op_1324_cast_fp16")]; tensor<int32, [4]> var_1325 = const()[name = tensor<string, []>("op_1325"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 448]> var_1326_cast_fp16 = reshape(shape = var_1325, x = value_13_cast_fp16)[name = tensor<string, []>("op_1326_cast_fp16")]; tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)]; tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(true)]; tensor<fp16, [1, 20, 64, 1]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_1326_cast_fp16, y = var_1324_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")]; tensor<int32, [4]> var_1329 = const()[name = tensor<string, []>("op_1329"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; tensor<fp16, [1, 1280, 1, 1]> input_97_cast_fp16 = reshape(shape = var_1329, x = attn_13_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")]; tensor<string, []> pretrained_out_67_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_67_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_67_strides_0 = const()[name = tensor<string, []>("pretrained_out_67_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_67_pad_0 = const()[name = tensor<string, []>("pretrained_out_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_67_dilations_0 = const()[name = tensor<string, []>("pretrained_out_67_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_67_groups_0 = const()[name = tensor<string, []>("pretrained_out_67_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179303424))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180122688))), name = tensor<string, []>("layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_3_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180122816)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_67_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_67_dilations_0, groups = pretrained_out_67_groups_0, pad = pretrained_out_67_pad_0, pad_type = pretrained_out_67_pad_type_0, strides = pretrained_out_67_strides_0, weight = layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_97_cast_fp16)[name = tensor<string, []>("pretrained_out_67_cast_fp16")]; tensor<string, []> input_99_pad_type_0 = const()[name = tensor<string, []>("input_99_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_99_strides_0 = const()[name = tensor<string, []>("input_99_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_99_pad_0 = const()[name = tensor<string, []>("input_99_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_99_dilations_0 = const()[name = tensor<string, []>("input_99_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_99_groups_0 = const()[name = tensor<string, []>("input_99_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_3_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180125440)))]; tensor<fp16, [1, 16, 1, 1]> input_99_cast_fp16 = conv(dilations = input_99_dilations_0, groups = input_99_groups_0, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = input_99_strides_0, weight = layers_3_self_attn_o_proj_loraA_weight_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")]; tensor<string, []> lora_out_67_pad_type_0 = const()[name = tensor<string, []>("lora_out_67_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_67_strides_0 = const()[name = tensor<string, []>("lora_out_67_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_67_pad_0 = const()[name = tensor<string, []>("lora_out_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_67_dilations_0 = const()[name = tensor<string, []>("lora_out_67_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_67_groups_0 = const()[name = tensor<string, []>("lora_out_67_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_3_self_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180166464)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_67_cast_fp16 = conv(dilations = lora_out_67_dilations_0, groups = lora_out_67_groups_0, pad = lora_out_67_pad_0, pad_type = lora_out_67_pad_type_0, strides = lora_out_67_strides_0, weight = layers_3_self_attn_o_proj_loraB_weight_to_fp16, x = input_99_cast_fp16)[name = tensor<string, []>("lora_out_67_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> obj_49_cast_fp16 = add(x = pretrained_out_67_cast_fp16, y = lora_out_67_cast_fp16)[name = tensor<string, []>("obj_49_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor<string, []>("inputs_21_cast_fp16")]; tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_1367_to_fp16 = const()[name = tensor<string, []>("op_1367_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1367_to_fp16, x = inputs_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")]; tensor<fp16, [1280]> obj_51_gamma_0_to_fp16 = const()[name = tensor<string, []>("obj_51_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180207488)))]; tensor<fp16, [1280]> obj_51_beta_0_to_fp16 = const()[name = tensor<string, []>("obj_51_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180210112)))]; tensor<fp16, []> obj_51_epsilon_0_to_fp16 = const()[name = tensor<string, []>("obj_51_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor<string, []>("obj_51_cast_fp16")]; tensor<string, []> pretrained_out_69_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_69_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_69_strides_0 = const()[name = tensor<string, []>("pretrained_out_69_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_69_pad_0 = const()[name = tensor<string, []>("pretrained_out_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_69_dilations_0 = const()[name = tensor<string, []>("pretrained_out_69_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_69_groups_0 = const()[name = tensor<string, []>("pretrained_out_69_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(180212736))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181032000))), name = tensor<string, []>("layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181032128)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_69_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_69_dilations_0, groups = pretrained_out_69_groups_0, pad = pretrained_out_69_pad_0, pad_type = pretrained_out_69_pad_type_0, strides = pretrained_out_69_strides_0, weight = layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_51_cast_fp16)[name = tensor<string, []>("pretrained_out_69_cast_fp16")]; tensor<string, []> input_101_pad_type_0 = const()[name = tensor<string, []>("input_101_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_101_strides_0 = const()[name = tensor<string, []>("input_101_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_101_pad_0 = const()[name = tensor<string, []>("input_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_101_dilations_0 = const()[name = tensor<string, []>("input_101_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_101_groups_0 = const()[name = tensor<string, []>("input_101_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_3_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181034752)))]; tensor<fp16, [1, 16, 1, 1]> input_101_cast_fp16 = conv(dilations = input_101_dilations_0, groups = input_101_groups_0, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = input_101_strides_0, weight = layers_3_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor<string, []>("input_101_cast_fp16")]; tensor<string, []> lora_out_69_pad_type_0 = const()[name = tensor<string, []>("lora_out_69_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_69_strides_0 = const()[name = tensor<string, []>("lora_out_69_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_69_pad_0 = const()[name = tensor<string, []>("lora_out_69_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_69_dilations_0 = const()[name = tensor<string, []>("lora_out_69_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_69_groups_0 = const()[name = tensor<string, []>("lora_out_69_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_3_encoder_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_q_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181075776)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_69_cast_fp16 = conv(dilations = lora_out_69_dilations_0, groups = lora_out_69_groups_0, pad = lora_out_69_pad_0, pad_type = lora_out_69_pad_type_0, strides = lora_out_69_strides_0, weight = layers_3_encoder_attn_q_proj_loraB_weight_to_fp16, x = input_101_cast_fp16)[name = tensor<string, []>("lora_out_69_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> query_cast_fp16 = add(x = pretrained_out_69_cast_fp16, y = lora_out_69_cast_fp16)[name = tensor<string, []>("query_cast_fp16")]; tensor<string, []> pretrained_out_71_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_71_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_71_strides_0 = const()[name = tensor<string, []>("pretrained_out_71_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_71_pad_0 = const()[name = tensor<string, []>("pretrained_out_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_71_dilations_0 = const()[name = tensor<string, []>("pretrained_out_71_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_71_groups_0 = const()[name = tensor<string, []>("pretrained_out_71_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181116800))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181936064))), name = tensor<string, []>("layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_71_cast_fp16 = conv(dilations = pretrained_out_71_dilations_0, groups = pretrained_out_71_groups_0, pad = pretrained_out_71_pad_0, pad_type = pretrained_out_71_pad_type_0, strides = pretrained_out_71_strides_0, weight = layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_71_cast_fp16")]; tensor<string, []> input_103_pad_type_0 = const()[name = tensor<string, []>("input_103_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_103_strides_0 = const()[name = tensor<string, []>("input_103_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_103_pad_0 = const()[name = tensor<string, []>("input_103_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_103_dilations_0 = const()[name = tensor<string, []>("input_103_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_103_groups_0 = const()[name = tensor<string, []>("input_103_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_3_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181936192)))]; tensor<fp16, [1, 16, 1, 1500]> input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = layers_3_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_103_cast_fp16")]; tensor<string, []> lora_out_71_pad_type_0 = const()[name = tensor<string, []>("lora_out_71_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_71_strides_0 = const()[name = tensor<string, []>("lora_out_71_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_71_pad_0 = const()[name = tensor<string, []>("lora_out_71_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_71_dilations_0 = const()[name = tensor<string, []>("lora_out_71_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_71_groups_0 = const()[name = tensor<string, []>("lora_out_71_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_3_encoder_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_k_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181977216)))]; tensor<fp16, [1, 1280, 1, 1500]> lora_out_71_cast_fp16 = conv(dilations = lora_out_71_dilations_0, groups = lora_out_71_groups_0, pad = lora_out_71_pad_0, pad_type = lora_out_71_pad_type_0, strides = lora_out_71_strides_0, weight = layers_3_encoder_attn_k_proj_loraB_weight_to_fp16, x = input_103_cast_fp16)[name = tensor<string, []>("lora_out_71_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1500]> key_cast_fp16 = add(x = pretrained_out_71_cast_fp16, y = lora_out_71_cast_fp16)[name = tensor<string, []>("key_cast_fp16")]; tensor<string, []> pretrained_out_73_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_73_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_73_strides_0 = const()[name = tensor<string, []>("pretrained_out_73_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_73_pad_0 = const()[name = tensor<string, []>("pretrained_out_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_73_dilations_0 = const()[name = tensor<string, []>("pretrained_out_73_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_73_groups_0 = const()[name = tensor<string, []>("pretrained_out_73_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182018240))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182837504))), name = tensor<string, []>("layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182837632)))]; tensor<fp16, [1, 1280, 1, 1500]> pretrained_out_73_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_73_dilations_0, groups = pretrained_out_73_groups_0, pad = pretrained_out_73_pad_0, pad_type = pretrained_out_73_pad_type_0, strides = pretrained_out_73_strides_0, weight = layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor<string, []>("pretrained_out_73_cast_fp16")]; tensor<string, []> input_105_pad_type_0 = const()[name = tensor<string, []>("input_105_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_105_strides_0 = const()[name = tensor<string, []>("input_105_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_105_pad_0 = const()[name = tensor<string, []>("input_105_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_105_dilations_0 = const()[name = tensor<string, []>("input_105_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_105_groups_0 = const()[name = tensor<string, []>("input_105_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_3_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182840256)))]; tensor<fp16, [1, 16, 1, 1500]> input_105_cast_fp16 = conv(dilations = input_105_dilations_0, groups = input_105_groups_0, pad = input_105_pad_0, pad_type = input_105_pad_type_0, strides = input_105_strides_0, weight = layers_3_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor<string, []>("input_105_cast_fp16")]; tensor<string, []> lora_out_73_pad_type_0 = const()[name = tensor<string, []>("lora_out_73_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_73_strides_0 = const()[name = tensor<string, []>("lora_out_73_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_73_pad_0 = const()[name = tensor<string, []>("lora_out_73_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_73_dilations_0 = const()[name = tensor<string, []>("lora_out_73_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_73_groups_0 = const()[name = tensor<string, []>("lora_out_73_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_3_encoder_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_v_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182881280)))]; tensor<fp16, [1, 1280, 1, 1500]> lora_out_73_cast_fp16 = conv(dilations = lora_out_73_dilations_0, groups = lora_out_73_groups_0, pad = lora_out_73_pad_0, pad_type = lora_out_73_pad_type_0, strides = lora_out_73_strides_0, weight = layers_3_encoder_attn_v_proj_loraB_weight_to_fp16, x = input_105_cast_fp16)[name = tensor<string, []>("lora_out_73_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1500]> value_cast_fp16 = add(x = pretrained_out_73_cast_fp16, y = lora_out_73_cast_fp16)[name = tensor<string, []>("value_cast_fp16")]; tensor<int32, [4]> var_1450 = const()[name = tensor<string, []>("op_1450"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1]> mh_q_cast_fp16 = reshape(shape = var_1450, x = query_cast_fp16)[name = tensor<string, []>("mh_q_cast_fp16")]; tensor<fp16, []> var_1452_to_fp16 = const()[name = tensor<string, []>("op_1452_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; tensor<fp16, [1, 20, 64, 1]> var_1453_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_1452_to_fp16)[name = tensor<string, []>("op_1453_cast_fp16")]; tensor<int32, [4]> var_1454 = const()[name = tensor<string, []>("op_1454"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1500]> var_1455_cast_fp16 = reshape(shape = var_1454, x = key_cast_fp16)[name = tensor<string, []>("op_1455_cast_fp16")]; tensor<bool, []> mh_w_transpose_x_0 = const()[name = tensor<string, []>("mh_w_transpose_x_0"), val = tensor<bool, []>(true)]; tensor<bool, []> mh_w_transpose_y_0 = const()[name = tensor<string, []>("mh_w_transpose_y_0"), val = tensor<bool, []>(false)]; tensor<fp16, [1, 20, 1, 1500]> mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_1453_cast_fp16, y = var_1455_cast_fp16)[name = tensor<string, []>("mh_w_cast_fp16")]; tensor<fp16, [1, 20, 1, 1500]> obj_55_cast_fp16 = softmax(axis = var_1190, x = mh_w_cast_fp16)[name = tensor<string, []>("obj_55_cast_fp16")]; tensor<int32, [4]> var_1459 = const()[name = tensor<string, []>("op_1459"), val = tensor<int32, [4]>([1, 20, 64, -1])]; tensor<fp16, [1, 20, 64, 1500]> var_1460_cast_fp16 = reshape(shape = var_1459, x = value_cast_fp16)[name = tensor<string, []>("op_1460_cast_fp16")]; tensor<bool, []> attn_transpose_x_0 = const()[name = tensor<string, []>("attn_transpose_x_0"), val = tensor<bool, []>(false)]; tensor<bool, []> attn_transpose_y_0 = const()[name = tensor<string, []>("attn_transpose_y_0"), val = tensor<bool, []>(true)]; tensor<fp16, [1, 20, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_1460_cast_fp16, y = obj_55_cast_fp16)[name = tensor<string, []>("attn_cast_fp16")]; tensor<int32, [4]> var_1463 = const()[name = tensor<string, []>("op_1463"), val = tensor<int32, [4]>([1, 1280, 1, -1])]; tensor<fp16, [1, 1280, 1, 1]> input_107_cast_fp16 = reshape(shape = var_1463, x = attn_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")]; tensor<string, []> pretrained_out_75_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_75_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_75_strides_0 = const()[name = tensor<string, []>("pretrained_out_75_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_75_pad_0 = const()[name = tensor<string, []>("pretrained_out_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_75_dilations_0 = const()[name = tensor<string, []>("pretrained_out_75_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_75_groups_0 = const()[name = tensor<string, []>("pretrained_out_75_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 1280, 1, 1]> layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [819200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182922304))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183741568))), name = tensor<string, []>("layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 1280, 1, 1])]; tensor<fp16, [1280]> layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183741696)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_75_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_75_dilations_0, groups = pretrained_out_75_groups_0, pad = pretrained_out_75_pad_0, pad_type = pretrained_out_75_pad_type_0, strides = pretrained_out_75_strides_0, weight = layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_107_cast_fp16)[name = tensor<string, []>("pretrained_out_75_cast_fp16")]; tensor<string, []> input_109_pad_type_0 = const()[name = tensor<string, []>("input_109_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_109_strides_0 = const()[name = tensor<string, []>("input_109_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_109_pad_0 = const()[name = tensor<string, []>("input_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_109_dilations_0 = const()[name = tensor<string, []>("input_109_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_109_groups_0 = const()[name = tensor<string, []>("input_109_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_3_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183744320)))]; tensor<fp16, [1, 16, 1, 1]> input_109_cast_fp16 = conv(dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = layers_3_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")]; tensor<string, []> lora_out_75_pad_type_0 = const()[name = tensor<string, []>("lora_out_75_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_75_strides_0 = const()[name = tensor<string, []>("lora_out_75_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_75_pad_0 = const()[name = tensor<string, []>("lora_out_75_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_75_dilations_0 = const()[name = tensor<string, []>("lora_out_75_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_75_groups_0 = const()[name = tensor<string, []>("lora_out_75_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_3_encoder_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_encoder_attn_o_proj_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183785344)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_75_cast_fp16 = conv(dilations = lora_out_75_dilations_0, groups = lora_out_75_groups_0, pad = lora_out_75_pad_0, pad_type = lora_out_75_pad_type_0, strides = lora_out_75_strides_0, weight = layers_3_encoder_attn_o_proj_loraB_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("lora_out_75_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> obj_53_cast_fp16 = add(x = pretrained_out_75_cast_fp16, y = lora_out_75_cast_fp16)[name = tensor<string, []>("obj_53_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor<string, []>("inputs_23_cast_fp16")]; tensor<int32, [1]> out_23_axes_0 = const()[name = tensor<string, []>("out_23_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_1500_to_fp16 = const()[name = tensor<string, []>("op_1500_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_1500_to_fp16, x = inputs_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")]; tensor<fp16, [1280]> input_111_gamma_0_to_fp16 = const()[name = tensor<string, []>("input_111_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183826368)))]; tensor<fp16, [1280]> input_111_beta_0_to_fp16 = const()[name = tensor<string, []>("input_111_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183828992)))]; tensor<fp16, []> input_111_epsilon_0_to_fp16 = const()[name = tensor<string, []>("input_111_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> 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<string, []>("input_111_cast_fp16")]; tensor<string, []> pretrained_out_77_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_77_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_77_strides_0 = const()[name = tensor<string, []>("pretrained_out_77_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_77_pad_0 = const()[name = tensor<string, []>("pretrained_out_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_77_dilations_0 = const()[name = tensor<string, []>("pretrained_out_77_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_77_groups_0 = const()[name = tensor<string, []>("pretrained_out_77_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [5120, 1280, 1, 1]> layers_3_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3276800]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183831616))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187108480))), name = tensor<string, []>("layers_3_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([5120, 1280, 1, 1])]; tensor<fp16, [5120]> layers_3_fc1_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_pretrained_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187108608)))]; tensor<fp16, [1, 5120, 1, 1]> pretrained_out_77_cast_fp16 = conv(bias = layers_3_fc1_pretrained_bias_to_fp16, dilations = pretrained_out_77_dilations_0, groups = pretrained_out_77_groups_0, pad = pretrained_out_77_pad_0, pad_type = pretrained_out_77_pad_type_0, strides = pretrained_out_77_strides_0, weight = layers_3_fc1_pretrained_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = tensor<string, []>("pretrained_out_77_cast_fp16")]; tensor<string, []> input_113_pad_type_0 = const()[name = tensor<string, []>("input_113_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_113_strides_0 = const()[name = tensor<string, []>("input_113_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_113_pad_0 = const()[name = tensor<string, []>("input_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_113_dilations_0 = const()[name = tensor<string, []>("input_113_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_113_groups_0 = const()[name = tensor<string, []>("input_113_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 1280, 1, 1]> layers_3_fc1_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_loraA_weight_to_fp16"), val = tensor<fp16, [16, 1280, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187118912)))]; tensor<fp16, [1, 16, 1, 1]> input_113_cast_fp16 = conv(dilations = input_113_dilations_0, groups = input_113_groups_0, pad = input_113_pad_0, pad_type = input_113_pad_type_0, strides = input_113_strides_0, weight = layers_3_fc1_loraA_weight_to_fp16, x = input_111_cast_fp16)[name = tensor<string, []>("input_113_cast_fp16")]; tensor<string, []> lora_out_77_pad_type_0 = const()[name = tensor<string, []>("lora_out_77_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_77_strides_0 = const()[name = tensor<string, []>("lora_out_77_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_77_pad_0 = const()[name = tensor<string, []>("lora_out_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_77_dilations_0 = const()[name = tensor<string, []>("lora_out_77_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_77_groups_0 = const()[name = tensor<string, []>("lora_out_77_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [5120, 16, 1, 1]> layers_3_fc1_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc1_loraB_weight_to_fp16"), val = tensor<fp16, [5120, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187159936)))]; tensor<fp16, [1, 5120, 1, 1]> lora_out_77_cast_fp16 = conv(dilations = lora_out_77_dilations_0, groups = lora_out_77_groups_0, pad = lora_out_77_pad_0, pad_type = lora_out_77_pad_type_0, strides = lora_out_77_strides_0, weight = layers_3_fc1_loraB_weight_to_fp16, x = input_113_cast_fp16)[name = tensor<string, []>("lora_out_77_cast_fp16")]; tensor<fp16, [1, 5120, 1, 1]> input_115_cast_fp16 = add(x = pretrained_out_77_cast_fp16, y = lora_out_77_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")]; tensor<string, []> input_117_mode_0 = const()[name = tensor<string, []>("input_117_mode_0"), val = tensor<string, []>("EXACT")]; tensor<fp16, [1, 5120, 1, 1]> input_117_cast_fp16 = gelu(mode = input_117_mode_0, x = input_115_cast_fp16)[name = tensor<string, []>("input_117_cast_fp16")]; tensor<string, []> pretrained_out_pad_type_0 = const()[name = tensor<string, []>("pretrained_out_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> pretrained_out_strides_0 = const()[name = tensor<string, []>("pretrained_out_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> pretrained_out_pad_0 = const()[name = tensor<string, []>("pretrained_out_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> pretrained_out_dilations_0 = const()[name = tensor<string, []>("pretrained_out_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> pretrained_out_groups_0 = const()[name = tensor<string, []>("pretrained_out_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 5120, 1, 1]> layers_3_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4915200]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187323840))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192239104))), name = tensor<string, []>("layers_3_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([1280, 5120, 1, 1])]; tensor<fp16, [1280]> layers_3_fc2_pretrained_bias_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_pretrained_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192239296)))]; tensor<fp16, [1, 1280, 1, 1]> pretrained_out_cast_fp16 = conv(bias = layers_3_fc2_pretrained_bias_to_fp16, dilations = pretrained_out_dilations_0, groups = pretrained_out_groups_0, pad = pretrained_out_pad_0, pad_type = pretrained_out_pad_type_0, strides = pretrained_out_strides_0, weight = layers_3_fc2_pretrained_weight_to_fp16_palettized, x = input_117_cast_fp16)[name = tensor<string, []>("pretrained_out_cast_fp16")]; tensor<string, []> input_pad_type_0 = const()[name = tensor<string, []>("input_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> input_strides_0 = const()[name = tensor<string, []>("input_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> input_pad_0 = const()[name = tensor<string, []>("input_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> input_dilations_0 = const()[name = tensor<string, []>("input_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> input_groups_0 = const()[name = tensor<string, []>("input_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [16, 5120, 1, 1]> layers_3_fc2_loraA_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_loraA_weight_to_fp16"), val = tensor<fp16, [16, 5120, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192241920)))]; tensor<fp16, [1, 16, 1, 1]> input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_3_fc2_loraA_weight_to_fp16, x = input_117_cast_fp16)[name = tensor<string, []>("input_cast_fp16")]; tensor<string, []> lora_out_pad_type_0 = const()[name = tensor<string, []>("lora_out_pad_type_0"), val = tensor<string, []>("valid")]; tensor<int32, [2]> lora_out_strides_0 = const()[name = tensor<string, []>("lora_out_strides_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, [4]> lora_out_pad_0 = const()[name = tensor<string, []>("lora_out_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [2]> lora_out_dilations_0 = const()[name = tensor<string, []>("lora_out_dilations_0"), val = tensor<int32, [2]>([1, 1])]; tensor<int32, []> lora_out_groups_0 = const()[name = tensor<string, []>("lora_out_groups_0"), val = tensor<int32, []>(1)]; tensor<fp16, [1280, 16, 1, 1]> layers_3_fc2_loraB_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_fc2_loraB_weight_to_fp16"), val = tensor<fp16, [1280, 16, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192405824)))]; tensor<fp16, [1, 1280, 1, 1]> lora_out_cast_fp16 = conv(dilations = lora_out_dilations_0, groups = lora_out_groups_0, pad = lora_out_pad_0, pad_type = lora_out_pad_type_0, strides = lora_out_strides_0, weight = layers_3_fc2_loraB_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("lora_out_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> hidden_states_9_cast_fp16 = add(x = pretrained_out_cast_fp16, y = lora_out_cast_fp16)[name = tensor<string, []>("hidden_states_9_cast_fp16")]; tensor<fp16, [1, 1280, 1, 1]> inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor<string, []>("inputs_cast_fp16")]; tensor<int32, [1]> out_axes_0 = const()[name = tensor<string, []>("out_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, []> var_1575_to_fp16 = const()[name = tensor<string, []>("op_1575_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_1575_to_fp16, x = inputs_cast_fp16)[name = tensor<string, []>("out_cast_fp16")]; tensor<fp16, [1280]> hidden_states_gamma_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192446848)))]; tensor<fp16, [1280]> hidden_states_beta_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192449472)))]; tensor<fp16, []> hidden_states_epsilon_0_to_fp16 = const()[name = tensor<string, []>("hidden_states_epsilon_0_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; tensor<fp16, [1, 1280, 1, 1]> hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")]; tensor<int32, [1]> var_1586_axes_0 = const()[name = tensor<string, []>("op_1586_axes_0"), val = tensor<int32, [1]>([2])]; tensor<fp16, [1, 1280, 1]> var_1586_cast_fp16 = squeeze(axes = var_1586_axes_0, x = hidden_states_cast_fp16)[name = tensor<string, []>("op_1586_cast_fp16")]; tensor<int32, [3]> var_1589_perm_0 = const()[name = tensor<string, []>("op_1589_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; tensor<fp16, [51866]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [51866]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(192452096)))]; tensor<fp16, [1, 1, 1280]> var_1589_cast_fp16 = transpose(perm = var_1589_perm_0, x = var_1586_cast_fp16)[name = tensor<string, []>("transpose_0")]; tensor<fp16, [1, 1, 51866]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_1589_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")]; tensor<int32, []> var_1593 = const()[name = tensor<string, []>("op_1593"), val = tensor<int32, []>(1)]; tensor<bool, []> obj_59_interleave_0 = const()[name = tensor<string, []>("obj_59_interleave_0"), val = tensor<bool, []>(false)]; tensor<fp16, [1, 5120, 1, 1]> key_cache_updates = concat(axis = var_1593, interleave = obj_59_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_cast_fp16))[name = tensor<string, []>("obj_59_cast_fp16")]; tensor<int32, []> var_1596 = const()[name = tensor<string, []>("op_1596"), val = tensor<int32, []>(1)]; tensor<bool, []> obj_61_interleave_0 = const()[name = tensor<string, []>("obj_61_interleave_0"), val = tensor<bool, []>(false)]; tensor<fp16, [1, 5120, 1, 1]> value_cache_updates = concat(axis = var_1596, interleave = obj_61_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_cast_fp16))[name = tensor<string, []>("obj_61_cast_fp16")]; tensor<int32, [4]> var_1607_begin_0 = const()[name = tensor<string, []>("op_1607_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])]; tensor<int32, [4]> var_1607_end_0 = const()[name = tensor<string, []>("op_1607_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1500])]; tensor<bool, [4]> var_1607_end_mask_0 = const()[name = tensor<string, []>("op_1607_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; tensor<fp16, [1, 1, 1, 1500]> var_1607_cast_fp16 = slice_by_index(begin = var_1607_begin_0, end = var_1607_end_0, end_mask = var_1607_end_mask_0, x = obj_41_cast_fp16)[name = tensor<string, []>("op_1607_cast_fp16")]; tensor<int32, [4]> var_1610_begin_0 = const()[name = tensor<string, []>("op_1610_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [4]> var_1610_end_0 = const()[name = tensor<string, []>("op_1610_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; tensor<bool, [4]> var_1610_end_mask_0 = const()[name = tensor<string, []>("op_1610_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; tensor<bool, [4]> var_1610_squeeze_mask_0 = const()[name = tensor<string, []>("op_1610_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; tensor<fp16, [1, 1, 1500]> var_1610_cast_fp16 = slice_by_index(begin = var_1610_begin_0, end = var_1610_end_0, end_mask = var_1610_end_mask_0, squeeze_mask = var_1610_squeeze_mask_0, x = var_1607_cast_fp16)[name = tensor<string, []>("op_1610_cast_fp16")]; tensor<int32, [4]> var_1625_begin_0 = const()[name = tensor<string, []>("op_1625_begin_0"), val = tensor<int32, [4]>([0, 11, 0, 0])]; tensor<int32, [4]> var_1625_end_0 = const()[name = tensor<string, []>("op_1625_end_0"), val = tensor<int32, [4]>([1, 12, 1, 1500])]; tensor<bool, [4]> var_1625_end_mask_0 = const()[name = tensor<string, []>("op_1625_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; tensor<fp16, [1, 1, 1, 1500]> var_1625_cast_fp16 = slice_by_index(begin = var_1625_begin_0, end = var_1625_end_0, end_mask = var_1625_end_mask_0, x = obj_41_cast_fp16)[name = tensor<string, []>("op_1625_cast_fp16")]; tensor<int32, [4]> var_1628_begin_0 = const()[name = tensor<string, []>("op_1628_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [4]> var_1628_end_0 = const()[name = tensor<string, []>("op_1628_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; tensor<bool, [4]> var_1628_end_mask_0 = const()[name = tensor<string, []>("op_1628_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; tensor<bool, [4]> var_1628_squeeze_mask_0 = const()[name = tensor<string, []>("op_1628_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; tensor<fp16, [1, 1, 1500]> var_1628_cast_fp16 = slice_by_index(begin = var_1628_begin_0, end = var_1628_end_0, end_mask = var_1628_end_mask_0, squeeze_mask = var_1628_squeeze_mask_0, x = var_1625_cast_fp16)[name = tensor<string, []>("op_1628_cast_fp16")]; tensor<int32, [4]> var_1643_begin_0 = const()[name = tensor<string, []>("op_1643_begin_0"), val = tensor<int32, [4]>([0, 3, 0, 0])]; tensor<int32, [4]> var_1643_end_0 = const()[name = tensor<string, []>("op_1643_end_0"), val = tensor<int32, [4]>([1, 4, 1, 1500])]; tensor<bool, [4]> var_1643_end_mask_0 = const()[name = tensor<string, []>("op_1643_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; tensor<fp16, [1, 1, 1, 1500]> var_1643_cast_fp16 = slice_by_index(begin = var_1643_begin_0, end = var_1643_end_0, end_mask = var_1643_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1643_cast_fp16")]; tensor<int32, [4]> var_1646_begin_0 = const()[name = tensor<string, []>("op_1646_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [4]> var_1646_end_0 = const()[name = tensor<string, []>("op_1646_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; tensor<bool, [4]> var_1646_end_mask_0 = const()[name = tensor<string, []>("op_1646_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; tensor<bool, [4]> var_1646_squeeze_mask_0 = const()[name = tensor<string, []>("op_1646_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; tensor<fp16, [1, 1, 1500]> var_1646_cast_fp16 = slice_by_index(begin = var_1646_begin_0, end = var_1646_end_0, end_mask = var_1646_end_mask_0, squeeze_mask = var_1646_squeeze_mask_0, x = var_1643_cast_fp16)[name = tensor<string, []>("op_1646_cast_fp16")]; tensor<int32, [4]> var_1661_begin_0 = const()[name = tensor<string, []>("op_1661_begin_0"), val = tensor<int32, [4]>([0, 6, 0, 0])]; tensor<int32, [4]> var_1661_end_0 = const()[name = tensor<string, []>("op_1661_end_0"), val = tensor<int32, [4]>([1, 7, 1, 1500])]; tensor<bool, [4]> var_1661_end_mask_0 = const()[name = tensor<string, []>("op_1661_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; tensor<fp16, [1, 1, 1, 1500]> var_1661_cast_fp16 = slice_by_index(begin = var_1661_begin_0, end = var_1661_end_0, end_mask = var_1661_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1661_cast_fp16")]; tensor<int32, [4]> var_1664_begin_0 = const()[name = tensor<string, []>("op_1664_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [4]> var_1664_end_0 = const()[name = tensor<string, []>("op_1664_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; tensor<bool, [4]> var_1664_end_mask_0 = const()[name = tensor<string, []>("op_1664_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; tensor<bool, [4]> var_1664_squeeze_mask_0 = const()[name = tensor<string, []>("op_1664_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; tensor<fp16, [1, 1, 1500]> var_1664_cast_fp16 = slice_by_index(begin = var_1664_begin_0, end = var_1664_end_0, end_mask = var_1664_end_mask_0, squeeze_mask = var_1664_squeeze_mask_0, x = var_1661_cast_fp16)[name = tensor<string, []>("op_1664_cast_fp16")]; tensor<int32, [4]> var_1679_begin_0 = const()[name = tensor<string, []>("op_1679_begin_0"), val = tensor<int32, [4]>([0, 11, 0, 0])]; tensor<int32, [4]> var_1679_end_0 = const()[name = tensor<string, []>("op_1679_end_0"), val = tensor<int32, [4]>([1, 12, 1, 1500])]; tensor<bool, [4]> var_1679_end_mask_0 = const()[name = tensor<string, []>("op_1679_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; tensor<fp16, [1, 1, 1, 1500]> var_1679_cast_fp16 = slice_by_index(begin = var_1679_begin_0, end = var_1679_end_0, end_mask = var_1679_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1679_cast_fp16")]; tensor<int32, [4]> var_1682_begin_0 = const()[name = tensor<string, []>("op_1682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [4]> var_1682_end_0 = const()[name = tensor<string, []>("op_1682_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; tensor<bool, [4]> var_1682_end_mask_0 = const()[name = tensor<string, []>("op_1682_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; tensor<bool, [4]> var_1682_squeeze_mask_0 = const()[name = tensor<string, []>("op_1682_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; tensor<fp16, [1, 1, 1500]> var_1682_cast_fp16 = slice_by_index(begin = var_1682_begin_0, end = var_1682_end_0, end_mask = var_1682_end_mask_0, squeeze_mask = var_1682_squeeze_mask_0, x = var_1679_cast_fp16)[name = tensor<string, []>("op_1682_cast_fp16")]; tensor<int32, [4]> var_1697_begin_0 = const()[name = tensor<string, []>("op_1697_begin_0"), val = tensor<int32, [4]>([0, 14, 0, 0])]; tensor<int32, [4]> var_1697_end_0 = const()[name = tensor<string, []>("op_1697_end_0"), val = tensor<int32, [4]>([1, 15, 1, 1500])]; tensor<bool, [4]> var_1697_end_mask_0 = const()[name = tensor<string, []>("op_1697_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])]; tensor<fp16, [1, 1, 1, 1500]> var_1697_cast_fp16 = slice_by_index(begin = var_1697_begin_0, end = var_1697_end_0, end_mask = var_1697_end_mask_0, x = obj_55_cast_fp16)[name = tensor<string, []>("op_1697_cast_fp16")]; tensor<int32, [4]> var_1700_begin_0 = const()[name = tensor<string, []>("op_1700_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; tensor<int32, [4]> var_1700_end_0 = const()[name = tensor<string, []>("op_1700_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1500])]; tensor<bool, [4]> var_1700_end_mask_0 = const()[name = tensor<string, []>("op_1700_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])]; tensor<bool, [4]> var_1700_squeeze_mask_0 = const()[name = tensor<string, []>("op_1700_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])]; tensor<fp16, [1, 1, 1500]> var_1700_cast_fp16 = slice_by_index(begin = var_1700_begin_0, end = var_1700_end_0, end_mask = var_1700_end_mask_0, squeeze_mask = var_1700_squeeze_mask_0, x = var_1697_cast_fp16)[name = tensor<string, []>("op_1700_cast_fp16")]; tensor<int32, []> var_1707 = const()[name = tensor<string, []>("op_1707"), val = tensor<int32, []>(1)]; tensor<bool, []> var_1708_interleave_0 = const()[name = tensor<string, []>("op_1708_interleave_0"), val = tensor<bool, []>(false)]; tensor<fp16, [1, 6, 1500]> var_1708_cast_fp16 = concat(axis = var_1707, interleave = var_1708_interleave_0, values = (var_1610_cast_fp16, var_1628_cast_fp16, var_1646_cast_fp16, var_1664_cast_fp16, var_1682_cast_fp16, var_1700_cast_fp16))[name = tensor<string, []>("op_1708_cast_fp16")]; tensor<bool, []> var_1711 = const()[name = tensor<string, []>("op_1711"), val = tensor<bool, []>(false)]; tensor<int32, [1]> obj_axes_0 = const()[name = tensor<string, []>("obj_axes_0"), val = tensor<int32, [1]>([1])]; tensor<fp16, [1, 1500]> alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = var_1711, x = var_1708_cast_fp16)[name = tensor<string, []>("obj_cast_fp16")]; } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); }