diff --git "a/CoreMLModels/ImageEncoder_float32_batch_16.mlmodelc/model.mil" "b/CoreMLModels/ImageEncoder_float32_batch_16.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/CoreMLModels/ImageEncoder_float32_batch_16.mlmodelc/model.mil" @@ -0,0 +1,1072 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3404.23.1"}, {"coremltools-component-torch", "2.6.0+cu124"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] +{ + func main(tensor colorImage) [FlexibleShapeInformation = tuple, dict, tensor>>, tuple, dict, list, ?>>>>((("DefaultShapes", {{"colorImage", [1, 3, 224, 224]}}), ("RangeDims", {{"colorImage", [[1, 1], [3, 3], [224, 224], [224, 224]]}})))] { + tensor colorImage__scaled___y_0 = const()[name = tensor("colorImage__scaled___y_0"), val = tensor(0x1.de77c4p-7)]; + tensor colorImage__scaled__ = mul(x = colorImage, y = colorImage__scaled___y_0)[name = tensor("colorImage__scaled__")]; + tensor colorImage__biased___y_0 = const()[name = tensor("colorImage__biased___y_0"), val = tensor([[[[-0x1.cad1b8p+0]], [[-0x1.c0897p+0]], [[-0x1.7aefaep+0]]]])]; + tensor colorImage__biased__ = add(x = colorImage__scaled__, y = colorImage__biased___y_0)[name = tensor("colorImage__biased__")]; + tensor x_3_pad_type_0 = const()[name = tensor("x_3_pad_type_0"), val = tensor("valid")]; + tensor x_3_strides_0 = const()[name = tensor("x_3_strides_0"), val = tensor([32, 32])]; + tensor x_3_pad_0 = const()[name = tensor("x_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_3_dilations_0 = const()[name = tensor("x_3_dilations_0"), val = tensor([1, 1])]; + tensor x_3_groups_0 = const()[name = tensor("x_3_groups_0"), val = tensor(1)]; + tensor colorImage_to_fp16_dtype_0 = const()[name = tensor("colorImage_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor conv1_weight_to_fp16 = const()[name = tensor("conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor colorImage_to_fp16 = cast(dtype = colorImage_to_fp16_dtype_0, x = colorImage__biased__)[name = tensor("cast_186")]; + tensor x_3_cast_fp16 = conv(dilations = x_3_dilations_0, groups = x_3_groups_0, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = x_3_strides_0, weight = conv1_weight_to_fp16, x = colorImage_to_fp16)[name = tensor("x_3_cast_fp16")]; + tensor var_22_shape_cast_fp16 = shape(x = x_3_cast_fp16)[name = tensor("op_22_shape_cast_fp16")]; + tensor gather_0_indices_0 = const()[name = tensor("gather_0_indices_0"), val = tensor(0)]; + tensor gather_0_axis_0 = const()[name = tensor("gather_0_axis_0"), val = tensor(0)]; + tensor gather_0_batch_dims_0 = const()[name = tensor("gather_0_batch_dims_0"), val = tensor(0)]; + tensor gather_0 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = gather_0_indices_0, x = var_22_shape_cast_fp16)[name = tensor("gather_0")]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(768)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(-1)]; + tensor concat_0_axis_0 = const()[name = tensor("concat_0_axis_0"), val = tensor(0)]; + tensor concat_0_interleave_0 = const()[name = tensor("concat_0_interleave_0"), val = tensor(false)]; + tensor concat_0 = concat(axis = concat_0_axis_0, interleave = concat_0_interleave_0, values = (gather_0, var_28, var_29))[name = tensor("concat_0")]; + tensor x_5_cast_fp16 = reshape(shape = concat_0, x = x_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor var_35 = const()[name = tensor("op_35"), val = tensor([0, 2, 1])]; + tensor x_7_cast_fp16 = transpose(perm = var_35, x = x_5_cast_fp16)[name = tensor("transpose_63")]; + tensor var_43_shape_cast_fp16 = shape(x = x_7_cast_fp16)[name = tensor("op_43_shape_cast_fp16")]; + tensor gather_2_indices_0 = const()[name = tensor("gather_2_indices_0"), val = tensor(0)]; + tensor gather_2_axis_0 = const()[name = tensor("gather_2_axis_0"), val = tensor(0)]; + tensor gather_2_batch_dims_0 = const()[name = tensor("gather_2_batch_dims_0"), val = tensor(0)]; + tensor gather_2 = gather(axis = gather_2_axis_0, batch_dims = gather_2_batch_dims_0, indices = gather_2_indices_0, x = var_43_shape_cast_fp16)[name = tensor("gather_2")]; + tensor var_49 = const()[name = tensor("op_49"), val = tensor(768)]; + tensor var_50 = const()[name = tensor("op_50"), val = tensor(1)]; + tensor concat_1_axis_0 = const()[name = tensor("concat_1_axis_0"), val = tensor(0)]; + tensor concat_1_interleave_0 = const()[name = tensor("concat_1_interleave_0"), val = tensor(false)]; + tensor concat_1 = concat(axis = concat_1_axis_0, interleave = concat_1_interleave_0, values = (gather_2, var_50, var_49))[name = tensor("concat_1")]; + tensor var_56_value_0_to_fp16 = const()[name = tensor("op_56_value_0_to_fp16"), val = tensor(0x0p+0)]; + tensor var_56_cast_fp16 = fill(shape = concat_1, value = var_56_value_0_to_fp16)[name = tensor("op_56_cast_fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4718720)))]; + tensor var_58_cast_fp16 = add(x = const_0_to_fp16, y = var_56_cast_fp16)[name = tensor("op_58_cast_fp16")]; + tensor var_60 = const()[name = tensor("op_60"), val = tensor(1)]; + tensor x_9_interleave_0 = const()[name = tensor("x_9_interleave_0"), val = tensor(false)]; + tensor x_9_cast_fp16 = concat(axis = var_60, interleave = x_9_interleave_0, values = (var_58_cast_fp16, x_7_cast_fp16))[name = tensor("x_9_cast_fp16")]; + tensor const_1_to_fp16 = const()[name = tensor("const_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4720320)))]; + tensor x_11_cast_fp16 = add(x = x_9_cast_fp16, y = const_1_to_fp16)[name = tensor("x_11_cast_fp16")]; + tensor ret_1_axes_0 = const()[name = tensor("ret_1_axes_0"), val = tensor([-1])]; + tensor ln_pre_weight_to_fp16 = const()[name = tensor("ln_pre_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4797184)))]; + tensor ln_pre_bias_to_fp16 = const()[name = tensor("ln_pre_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4798784)))]; + tensor var_70_to_fp16 = const()[name = tensor("op_70_to_fp16"), val = tensor(0x1.5p-17)]; + tensor ret_1_cast_fp16 = layer_norm(axes = ret_1_axes_0, beta = ln_pre_bias_to_fp16, epsilon = var_70_to_fp16, gamma = ln_pre_weight_to_fp16, x = x_11_cast_fp16)[name = tensor("ret_1_cast_fp16")]; + tensor var_84 = const()[name = tensor("op_84"), val = tensor([1, 0, 2])]; + tensor ret_3_axes_0 = const()[name = tensor("ret_3_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_0_ln_1_weight_to_fp16 = const()[name = tensor("transformer_resblocks_0_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4800384)))]; + tensor transformer_resblocks_0_ln_1_bias_to_fp16 = const()[name = tensor("transformer_resblocks_0_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4801984)))]; + tensor var_100_to_fp16 = const()[name = tensor("op_100_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_15_cast_fp16 = transpose(perm = var_84, x = ret_1_cast_fp16)[name = tensor("transpose_62")]; + tensor ret_3_cast_fp16 = layer_norm(axes = ret_3_axes_0, beta = transformer_resblocks_0_ln_1_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_0_ln_1_weight_to_fp16, x = x_15_cast_fp16)[name = tensor("ret_3_cast_fp16")]; + tensor transformer_resblocks_0_attn_in_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_0_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4803584)))]; + tensor transformer_resblocks_0_attn_in_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_0_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8342592)))]; + tensor linear_0_cast_fp16 = linear(bias = transformer_resblocks_0_attn_in_proj_bias_to_fp16, weight = transformer_resblocks_0_attn_in_proj_weight_to_fp16, x = ret_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor concat_2x = const()[name = tensor("concat_2x"), val = tensor([50, -1, 3, 768])]; + tensor var_150_cast_fp16 = reshape(shape = concat_2x, x = linear_0_cast_fp16)[name = tensor("op_150_cast_fp16")]; + tensor var_151_axes_0 = const()[name = tensor("op_151_axes_0"), val = tensor([0])]; + tensor var_151_cast_fp16 = expand_dims(axes = var_151_axes_0, x = var_150_cast_fp16)[name = tensor("op_151_cast_fp16")]; + tensor var_152_perm_0 = const()[name = tensor("op_152_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_153_axes_0 = const()[name = tensor("op_153_axes_0"), val = tensor([-2])]; + tensor var_152_cast_fp16 = transpose(perm = var_152_perm_0, x = var_151_cast_fp16)[name = tensor("transpose_61")]; + tensor var_153_cast_fp16 = squeeze(axes = var_153_axes_0, x = var_152_cast_fp16)[name = tensor("op_153_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 50, 0, 768])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = var_153_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([2, 50, 0, 768])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = var_153_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([3, 50, 0, 768])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = var_153_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor concat_3x = const()[name = tensor("concat_3x"), val = tensor([50, -1, 64])]; + tensor var_162_cast_fp16 = reshape(shape = concat_3x, x = q_1_cast_fp16)[name = tensor("op_162_cast_fp16")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_4x = const()[name = tensor("concat_4x"), val = tensor([50, -1, 64])]; + tensor var_169_cast_fp16 = reshape(shape = concat_4x, x = k_1_cast_fp16)[name = tensor("op_169_cast_fp16")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_5x = const()[name = tensor("concat_5x"), val = tensor([50, -1, 64])]; + tensor var_176_cast_fp16 = reshape(shape = concat_5x, x = v_1_cast_fp16)[name = tensor("op_176_cast_fp16")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_6x = const()[name = tensor("concat_6x"), val = tensor([-1, 12, 50, 64])]; + tensor q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_162_cast_fp16)[name = tensor("transpose_60")]; + tensor q_5_cast_fp16 = reshape(shape = concat_6x, x = q_3_cast_fp16)[name = tensor("q_5_cast_fp16")]; + tensor concat_7x = const()[name = tensor("concat_7x"), val = tensor([-1, 12, 50, 64])]; + tensor k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_169_cast_fp16)[name = tensor("transpose_59")]; + tensor k_5_cast_fp16 = reshape(shape = concat_7x, x = k_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor concat_8x = const()[name = tensor("concat_8x"), val = tensor([-1, 12, 50, 64])]; + tensor v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_176_cast_fp16)[name = tensor("transpose_58")]; + tensor v_5_cast_fp16 = reshape(shape = concat_8x, x = v_3_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor mul_1_y_0_to_fp16 = const()[name = tensor("mul_1_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_1_cast_fp16 = mul(x = q_5_cast_fp16, y = mul_1_y_0_to_fp16)[name = tensor("mul_1_cast_fp16")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1_cast_fp16, y = k_5_cast_fp16)[name = tensor("matmul_0_cast_fp16")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = matmul_0_cast_fp16)[name = tensor("softmax_0_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = v_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_186 = const()[name = tensor("op_186"), val = tensor([2, 0, 1, 3])]; + tensor concat_9x = const()[name = tensor("concat_9x"), val = tensor([-1, 768])]; + tensor var_187_cast_fp16 = transpose(perm = var_186, x = attn_output_1_cast_fp16)[name = tensor("transpose_57")]; + tensor attn_output_3_cast_fp16 = reshape(shape = concat_9x, x = var_187_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor transformer_resblocks_0_attn_out_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_0_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8347264)))]; + tensor transformer_resblocks_0_attn_out_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_0_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9526976)))]; + tensor linear_1_cast_fp16 = linear(bias = transformer_resblocks_0_attn_out_proj_bias_to_fp16, weight = transformer_resblocks_0_attn_out_proj_weight_to_fp16, x = attn_output_3_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor concat_10x = const()[name = tensor("concat_10x"), val = tensor([50, -1, 768])]; + tensor var_196_cast_fp16 = reshape(shape = concat_10x, x = linear_1_cast_fp16)[name = tensor("op_196_cast_fp16")]; + tensor x_17_cast_fp16 = add(x = x_15_cast_fp16, y = var_196_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor ret_5_axes_0 = const()[name = tensor("ret_5_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_0_ln_2_weight_to_fp16 = const()[name = tensor("transformer_resblocks_0_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9528576)))]; + tensor transformer_resblocks_0_ln_2_bias_to_fp16 = const()[name = tensor("transformer_resblocks_0_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9530176)))]; + tensor ret_5_cast_fp16 = layer_norm(axes = ret_5_axes_0, beta = transformer_resblocks_0_ln_2_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_0_ln_2_weight_to_fp16, x = x_17_cast_fp16)[name = tensor("ret_5_cast_fp16")]; + tensor transformer_resblocks_0_mlp_c_fc_weight_to_fp16 = const()[name = tensor("transformer_resblocks_0_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9531776)))]; + tensor transformer_resblocks_0_mlp_c_fc_bias_to_fp16 = const()[name = tensor("transformer_resblocks_0_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14250432)))]; + tensor linear_2_cast_fp16 = linear(bias = transformer_resblocks_0_mlp_c_fc_bias_to_fp16, weight = transformer_resblocks_0_mlp_c_fc_weight_to_fp16, x = ret_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_212_to_fp16 = const()[name = tensor("op_212_to_fp16"), val = tensor(0x1.b3cp+0)]; + tensor var_213_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_212_to_fp16)[name = tensor("op_213_cast_fp16")]; + tensor var_214_cast_fp16 = sigmoid(x = var_213_cast_fp16)[name = tensor("op_214_cast_fp16")]; + tensor input_9_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_214_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor transformer_resblocks_0_mlp_c_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_0_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14256640)))]; + tensor transformer_resblocks_0_mlp_c_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_0_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18975296)))]; + tensor linear_3_cast_fp16 = linear(bias = transformer_resblocks_0_mlp_c_proj_bias_to_fp16, weight = transformer_resblocks_0_mlp_c_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor x_21_cast_fp16 = add(x = x_17_cast_fp16, y = linear_3_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor ret_7_axes_0 = const()[name = tensor("ret_7_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_1_ln_1_weight_to_fp16 = const()[name = tensor("transformer_resblocks_1_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18976896)))]; + tensor transformer_resblocks_1_ln_1_bias_to_fp16 = const()[name = tensor("transformer_resblocks_1_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18978496)))]; + tensor ret_7_cast_fp16 = layer_norm(axes = ret_7_axes_0, beta = transformer_resblocks_1_ln_1_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_1_ln_1_weight_to_fp16, x = x_21_cast_fp16)[name = tensor("ret_7_cast_fp16")]; + tensor transformer_resblocks_1_attn_in_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_1_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18980096)))]; + tensor transformer_resblocks_1_attn_in_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_1_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22519104)))]; + tensor linear_4_cast_fp16 = linear(bias = transformer_resblocks_1_attn_in_proj_bias_to_fp16, weight = transformer_resblocks_1_attn_in_proj_weight_to_fp16, x = ret_7_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor concat_11x = const()[name = tensor("concat_11x"), val = tensor([50, -1, 3, 768])]; + tensor var_255_cast_fp16 = reshape(shape = concat_11x, x = linear_4_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor var_256_axes_0 = const()[name = tensor("op_256_axes_0"), val = tensor([0])]; + tensor var_256_cast_fp16 = expand_dims(axes = var_256_axes_0, x = var_255_cast_fp16)[name = tensor("op_256_cast_fp16")]; + tensor var_257_perm_0 = const()[name = tensor("op_257_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_258_axes_0 = const()[name = tensor("op_258_axes_0"), val = tensor([-2])]; + tensor var_257_cast_fp16 = transpose(perm = var_257_perm_0, x = var_256_cast_fp16)[name = tensor("transpose_56")]; + tensor var_258_cast_fp16 = squeeze(axes = var_258_axes_0, x = var_257_cast_fp16)[name = tensor("op_258_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 50, 0, 768])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = var_258_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_7_begin_0 = const()[name = tensor("k_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_7_end_0 = const()[name = tensor("k_7_end_0"), val = tensor([2, 50, 0, 768])]; + tensor k_7_end_mask_0 = const()[name = tensor("k_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_7_squeeze_mask_0 = const()[name = tensor("k_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_7_cast_fp16 = slice_by_index(begin = k_7_begin_0, end = k_7_end_0, end_mask = k_7_end_mask_0, squeeze_mask = k_7_squeeze_mask_0, x = var_258_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([3, 50, 0, 768])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = var_258_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor concat_12x = const()[name = tensor("concat_12x"), val = tensor([50, -1, 64])]; + tensor var_267_cast_fp16 = reshape(shape = concat_12x, x = q_7_cast_fp16)[name = tensor("op_267_cast_fp16")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_13x = const()[name = tensor("concat_13x"), val = tensor([50, -1, 64])]; + tensor var_274_cast_fp16 = reshape(shape = concat_13x, x = k_7_cast_fp16)[name = tensor("op_274_cast_fp16")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_14x = const()[name = tensor("concat_14x"), val = tensor([50, -1, 64])]; + tensor var_281_cast_fp16 = reshape(shape = concat_14x, x = v_7_cast_fp16)[name = tensor("op_281_cast_fp16")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_15x = const()[name = tensor("concat_15x"), val = tensor([-1, 12, 50, 64])]; + tensor q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_267_cast_fp16)[name = tensor("transpose_55")]; + tensor q_11_cast_fp16 = reshape(shape = concat_15x, x = q_9_cast_fp16)[name = tensor("q_11_cast_fp16")]; + tensor concat_16x = const()[name = tensor("concat_16x"), val = tensor([-1, 12, 50, 64])]; + tensor k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_274_cast_fp16)[name = tensor("transpose_54")]; + tensor k_11_cast_fp16 = reshape(shape = concat_16x, x = k_9_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor concat_17x = const()[name = tensor("concat_17x"), val = tensor([-1, 12, 50, 64])]; + tensor v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_281_cast_fp16)[name = tensor("transpose_53")]; + tensor v_11_cast_fp16 = reshape(shape = concat_17x, x = v_9_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor mul_3_y_0_to_fp16 = const()[name = tensor("mul_3_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_3_cast_fp16 = mul(x = q_11_cast_fp16, y = mul_3_y_0_to_fp16)[name = tensor("mul_3_cast_fp16")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3_cast_fp16, y = k_11_cast_fp16)[name = tensor("matmul_1_cast_fp16")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = matmul_1_cast_fp16)[name = tensor("softmax_1_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = softmax_1_cast_fp16, y = v_11_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_291 = const()[name = tensor("op_291"), val = tensor([2, 0, 1, 3])]; + tensor concat_18x = const()[name = tensor("concat_18x"), val = tensor([-1, 768])]; + tensor var_292_cast_fp16 = transpose(perm = var_291, x = attn_output_7_cast_fp16)[name = tensor("transpose_52")]; + tensor attn_output_9_cast_fp16 = reshape(shape = concat_18x, x = var_292_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor transformer_resblocks_1_attn_out_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_1_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22523776)))]; + tensor transformer_resblocks_1_attn_out_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_1_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23703488)))]; + tensor linear_5_cast_fp16 = linear(bias = transformer_resblocks_1_attn_out_proj_bias_to_fp16, weight = transformer_resblocks_1_attn_out_proj_weight_to_fp16, x = attn_output_9_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor concat_19x = const()[name = tensor("concat_19x"), val = tensor([50, -1, 768])]; + tensor var_301_cast_fp16 = reshape(shape = concat_19x, x = linear_5_cast_fp16)[name = tensor("op_301_cast_fp16")]; + tensor x_23_cast_fp16 = add(x = x_21_cast_fp16, y = var_301_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor ret_9_axes_0 = const()[name = tensor("ret_9_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_1_ln_2_weight_to_fp16 = const()[name = tensor("transformer_resblocks_1_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23705088)))]; + tensor transformer_resblocks_1_ln_2_bias_to_fp16 = const()[name = tensor("transformer_resblocks_1_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23706688)))]; + tensor ret_9_cast_fp16 = layer_norm(axes = ret_9_axes_0, beta = transformer_resblocks_1_ln_2_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_1_ln_2_weight_to_fp16, x = x_23_cast_fp16)[name = tensor("ret_9_cast_fp16")]; + tensor transformer_resblocks_1_mlp_c_fc_weight_to_fp16 = const()[name = tensor("transformer_resblocks_1_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23708288)))]; + tensor transformer_resblocks_1_mlp_c_fc_bias_to_fp16 = const()[name = tensor("transformer_resblocks_1_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28426944)))]; + tensor linear_6_cast_fp16 = linear(bias = transformer_resblocks_1_mlp_c_fc_bias_to_fp16, weight = transformer_resblocks_1_mlp_c_fc_weight_to_fp16, x = ret_9_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_317_to_fp16 = const()[name = tensor("op_317_to_fp16"), val = tensor(0x1.b3cp+0)]; + tensor var_318_cast_fp16 = mul(x = linear_6_cast_fp16, y = var_317_to_fp16)[name = tensor("op_318_cast_fp16")]; + tensor var_319_cast_fp16 = sigmoid(x = var_318_cast_fp16)[name = tensor("op_319_cast_fp16")]; + tensor input_17_cast_fp16 = mul(x = linear_6_cast_fp16, y = var_319_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor transformer_resblocks_1_mlp_c_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_1_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28433152)))]; + tensor transformer_resblocks_1_mlp_c_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_1_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33151808)))]; + tensor linear_7_cast_fp16 = linear(bias = transformer_resblocks_1_mlp_c_proj_bias_to_fp16, weight = transformer_resblocks_1_mlp_c_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor x_27_cast_fp16 = add(x = x_23_cast_fp16, y = linear_7_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor ret_11_axes_0 = const()[name = tensor("ret_11_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_2_ln_1_weight_to_fp16 = const()[name = tensor("transformer_resblocks_2_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33153408)))]; + tensor transformer_resblocks_2_ln_1_bias_to_fp16 = const()[name = tensor("transformer_resblocks_2_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33155008)))]; + tensor ret_11_cast_fp16 = layer_norm(axes = ret_11_axes_0, beta = transformer_resblocks_2_ln_1_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_2_ln_1_weight_to_fp16, x = x_27_cast_fp16)[name = tensor("ret_11_cast_fp16")]; + tensor transformer_resblocks_2_attn_in_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_2_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33156608)))]; + tensor transformer_resblocks_2_attn_in_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_2_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36695616)))]; + tensor linear_8_cast_fp16 = linear(bias = transformer_resblocks_2_attn_in_proj_bias_to_fp16, weight = transformer_resblocks_2_attn_in_proj_weight_to_fp16, x = ret_11_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor concat_20x = const()[name = tensor("concat_20x"), val = tensor([50, -1, 3, 768])]; + tensor var_360_cast_fp16 = reshape(shape = concat_20x, x = linear_8_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor var_361_axes_0 = const()[name = tensor("op_361_axes_0"), val = tensor([0])]; + tensor var_361_cast_fp16 = expand_dims(axes = var_361_axes_0, x = var_360_cast_fp16)[name = tensor("op_361_cast_fp16")]; + tensor var_362_perm_0 = const()[name = tensor("op_362_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_363_axes_0 = const()[name = tensor("op_363_axes_0"), val = tensor([-2])]; + tensor var_362_cast_fp16 = transpose(perm = var_362_perm_0, x = var_361_cast_fp16)[name = tensor("transpose_51")]; + tensor var_363_cast_fp16 = squeeze(axes = var_363_axes_0, x = var_362_cast_fp16)[name = tensor("op_363_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 50, 0, 768])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = var_363_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([2, 50, 0, 768])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = var_363_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([3, 50, 0, 768])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = var_363_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor concat_21x = const()[name = tensor("concat_21x"), val = tensor([50, -1, 64])]; + tensor var_372_cast_fp16 = reshape(shape = concat_21x, x = q_13_cast_fp16)[name = tensor("op_372_cast_fp16")]; + tensor q_15_perm_0 = const()[name = tensor("q_15_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_22x = const()[name = tensor("concat_22x"), val = tensor([50, -1, 64])]; + tensor var_379_cast_fp16 = reshape(shape = concat_22x, x = k_13_cast_fp16)[name = tensor("op_379_cast_fp16")]; + tensor k_15_perm_0 = const()[name = tensor("k_15_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_23x = const()[name = tensor("concat_23x"), val = tensor([50, -1, 64])]; + tensor var_386_cast_fp16 = reshape(shape = concat_23x, x = v_13_cast_fp16)[name = tensor("op_386_cast_fp16")]; + tensor v_15_perm_0 = const()[name = tensor("v_15_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_24x = const()[name = tensor("concat_24x"), val = tensor([-1, 12, 50, 64])]; + tensor q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_372_cast_fp16)[name = tensor("transpose_50")]; + tensor q_17_cast_fp16 = reshape(shape = concat_24x, x = q_15_cast_fp16)[name = tensor("q_17_cast_fp16")]; + tensor concat_25x = const()[name = tensor("concat_25x"), val = tensor([-1, 12, 50, 64])]; + tensor k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_379_cast_fp16)[name = tensor("transpose_49")]; + tensor k_17_cast_fp16 = reshape(shape = concat_25x, x = k_15_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor concat_26x = const()[name = tensor("concat_26x"), val = tensor([-1, 12, 50, 64])]; + tensor v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_386_cast_fp16)[name = tensor("transpose_48")]; + tensor v_17_cast_fp16 = reshape(shape = concat_26x, x = v_15_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor mul_5_y_0_to_fp16 = const()[name = tensor("mul_5_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_5_cast_fp16 = mul(x = q_17_cast_fp16, y = mul_5_y_0_to_fp16)[name = tensor("mul_5_cast_fp16")]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(true)]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5_cast_fp16, y = k_17_cast_fp16)[name = tensor("matmul_2_cast_fp16")]; + tensor softmax_2_axis_0 = const()[name = tensor("softmax_2_axis_0"), val = tensor(-1)]; + tensor softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = matmul_2_cast_fp16)[name = tensor("softmax_2_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_2_cast_fp16, y = v_17_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_396 = const()[name = tensor("op_396"), val = tensor([2, 0, 1, 3])]; + tensor concat_27x = const()[name = tensor("concat_27x"), val = tensor([-1, 768])]; + tensor var_397_cast_fp16 = transpose(perm = var_396, x = attn_output_13_cast_fp16)[name = tensor("transpose_47")]; + tensor attn_output_15_cast_fp16 = reshape(shape = concat_27x, x = var_397_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor transformer_resblocks_2_attn_out_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_2_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36700288)))]; + tensor transformer_resblocks_2_attn_out_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_2_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37880000)))]; + tensor linear_9_cast_fp16 = linear(bias = transformer_resblocks_2_attn_out_proj_bias_to_fp16, weight = transformer_resblocks_2_attn_out_proj_weight_to_fp16, x = attn_output_15_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor concat_28x = const()[name = tensor("concat_28x"), val = tensor([50, -1, 768])]; + tensor var_406_cast_fp16 = reshape(shape = concat_28x, x = linear_9_cast_fp16)[name = tensor("op_406_cast_fp16")]; + tensor x_29_cast_fp16 = add(x = x_27_cast_fp16, y = var_406_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor ret_13_axes_0 = const()[name = tensor("ret_13_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_2_ln_2_weight_to_fp16 = const()[name = tensor("transformer_resblocks_2_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37881600)))]; + tensor transformer_resblocks_2_ln_2_bias_to_fp16 = const()[name = tensor("transformer_resblocks_2_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37883200)))]; + tensor ret_13_cast_fp16 = layer_norm(axes = ret_13_axes_0, beta = transformer_resblocks_2_ln_2_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_2_ln_2_weight_to_fp16, x = x_29_cast_fp16)[name = tensor("ret_13_cast_fp16")]; + tensor transformer_resblocks_2_mlp_c_fc_weight_to_fp16 = const()[name = tensor("transformer_resblocks_2_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37884800)))]; + tensor transformer_resblocks_2_mlp_c_fc_bias_to_fp16 = const()[name = tensor("transformer_resblocks_2_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42603456)))]; + tensor linear_10_cast_fp16 = linear(bias = transformer_resblocks_2_mlp_c_fc_bias_to_fp16, weight = transformer_resblocks_2_mlp_c_fc_weight_to_fp16, x = ret_13_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_422_to_fp16 = const()[name = tensor("op_422_to_fp16"), val = tensor(0x1.b3cp+0)]; + tensor var_423_cast_fp16 = mul(x = linear_10_cast_fp16, y = var_422_to_fp16)[name = tensor("op_423_cast_fp16")]; + tensor var_424_cast_fp16 = sigmoid(x = var_423_cast_fp16)[name = tensor("op_424_cast_fp16")]; + tensor input_25_cast_fp16 = mul(x = linear_10_cast_fp16, y = var_424_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor transformer_resblocks_2_mlp_c_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_2_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42609664)))]; + tensor transformer_resblocks_2_mlp_c_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_2_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47328320)))]; + tensor linear_11_cast_fp16 = linear(bias = transformer_resblocks_2_mlp_c_proj_bias_to_fp16, weight = transformer_resblocks_2_mlp_c_proj_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor x_33_cast_fp16 = add(x = x_29_cast_fp16, y = linear_11_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor ret_15_axes_0 = const()[name = tensor("ret_15_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_3_ln_1_weight_to_fp16 = const()[name = tensor("transformer_resblocks_3_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47329920)))]; + tensor transformer_resblocks_3_ln_1_bias_to_fp16 = const()[name = tensor("transformer_resblocks_3_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47331520)))]; + tensor ret_15_cast_fp16 = layer_norm(axes = ret_15_axes_0, beta = transformer_resblocks_3_ln_1_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_3_ln_1_weight_to_fp16, x = x_33_cast_fp16)[name = tensor("ret_15_cast_fp16")]; + tensor transformer_resblocks_3_attn_in_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_3_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47333120)))]; + tensor transformer_resblocks_3_attn_in_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_3_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50872128)))]; + tensor linear_12_cast_fp16 = linear(bias = transformer_resblocks_3_attn_in_proj_bias_to_fp16, weight = transformer_resblocks_3_attn_in_proj_weight_to_fp16, x = ret_15_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor concat_29x = const()[name = tensor("concat_29x"), val = tensor([50, -1, 3, 768])]; + tensor var_465_cast_fp16 = reshape(shape = concat_29x, x = linear_12_cast_fp16)[name = tensor("op_465_cast_fp16")]; + tensor var_466_axes_0 = const()[name = tensor("op_466_axes_0"), val = tensor([0])]; + tensor var_466_cast_fp16 = expand_dims(axes = var_466_axes_0, x = var_465_cast_fp16)[name = tensor("op_466_cast_fp16")]; + tensor var_467_perm_0 = const()[name = tensor("op_467_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_468_axes_0 = const()[name = tensor("op_468_axes_0"), val = tensor([-2])]; + tensor var_467_cast_fp16 = transpose(perm = var_467_perm_0, x = var_466_cast_fp16)[name = tensor("transpose_46")]; + tensor var_468_cast_fp16 = squeeze(axes = var_468_axes_0, x = var_467_cast_fp16)[name = tensor("op_468_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 50, 0, 768])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_468_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 50, 0, 768])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19_cast_fp16 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_468_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 50, 0, 768])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_468_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor concat_30x = const()[name = tensor("concat_30x"), val = tensor([50, -1, 64])]; + tensor var_477_cast_fp16 = reshape(shape = concat_30x, x = q_19_cast_fp16)[name = tensor("op_477_cast_fp16")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_31x = const()[name = tensor("concat_31x"), val = tensor([50, -1, 64])]; + tensor var_484_cast_fp16 = reshape(shape = concat_31x, x = k_19_cast_fp16)[name = tensor("op_484_cast_fp16")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_32x = const()[name = tensor("concat_32x"), val = tensor([50, -1, 64])]; + tensor var_491_cast_fp16 = reshape(shape = concat_32x, x = v_19_cast_fp16)[name = tensor("op_491_cast_fp16")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_33x = const()[name = tensor("concat_33x"), val = tensor([-1, 12, 50, 64])]; + tensor q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_477_cast_fp16)[name = tensor("transpose_45")]; + tensor q_23_cast_fp16 = reshape(shape = concat_33x, x = q_21_cast_fp16)[name = tensor("q_23_cast_fp16")]; + tensor concat_34x = const()[name = tensor("concat_34x"), val = tensor([-1, 12, 50, 64])]; + tensor k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_484_cast_fp16)[name = tensor("transpose_44")]; + tensor k_23_cast_fp16 = reshape(shape = concat_34x, x = k_21_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor concat_35x = const()[name = tensor("concat_35x"), val = tensor([-1, 12, 50, 64])]; + tensor v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_491_cast_fp16)[name = tensor("transpose_43")]; + tensor v_23_cast_fp16 = reshape(shape = concat_35x, x = v_21_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor mul_7_y_0_to_fp16 = const()[name = tensor("mul_7_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_7_cast_fp16 = mul(x = q_23_cast_fp16, y = mul_7_y_0_to_fp16)[name = tensor("mul_7_cast_fp16")]; + tensor matmul_3_transpose_y_0 = const()[name = tensor("matmul_3_transpose_y_0"), val = tensor(true)]; + tensor matmul_3_transpose_x_0 = const()[name = tensor("matmul_3_transpose_x_0"), val = tensor(false)]; + tensor matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7_cast_fp16, y = k_23_cast_fp16)[name = tensor("matmul_3_cast_fp16")]; + tensor softmax_3_axis_0 = const()[name = tensor("softmax_3_axis_0"), val = tensor(-1)]; + tensor softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = matmul_3_cast_fp16)[name = tensor("softmax_3_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = softmax_3_cast_fp16, y = v_23_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_501 = const()[name = tensor("op_501"), val = tensor([2, 0, 1, 3])]; + tensor concat_36x = const()[name = tensor("concat_36x"), val = tensor([-1, 768])]; + tensor var_502_cast_fp16 = transpose(perm = var_501, x = attn_output_19_cast_fp16)[name = tensor("transpose_42")]; + tensor attn_output_21_cast_fp16 = reshape(shape = concat_36x, x = var_502_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor transformer_resblocks_3_attn_out_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_3_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50876800)))]; + tensor transformer_resblocks_3_attn_out_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_3_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52056512)))]; + tensor linear_13_cast_fp16 = linear(bias = transformer_resblocks_3_attn_out_proj_bias_to_fp16, weight = transformer_resblocks_3_attn_out_proj_weight_to_fp16, x = attn_output_21_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor concat_37x = const()[name = tensor("concat_37x"), val = tensor([50, -1, 768])]; + tensor var_511_cast_fp16 = reshape(shape = concat_37x, x = linear_13_cast_fp16)[name = tensor("op_511_cast_fp16")]; + tensor x_35_cast_fp16 = add(x = x_33_cast_fp16, y = var_511_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor ret_17_axes_0 = const()[name = tensor("ret_17_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_3_ln_2_weight_to_fp16 = const()[name = tensor("transformer_resblocks_3_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52058112)))]; + tensor transformer_resblocks_3_ln_2_bias_to_fp16 = const()[name = tensor("transformer_resblocks_3_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52059712)))]; + tensor ret_17_cast_fp16 = layer_norm(axes = ret_17_axes_0, beta = transformer_resblocks_3_ln_2_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_3_ln_2_weight_to_fp16, x = x_35_cast_fp16)[name = tensor("ret_17_cast_fp16")]; + tensor transformer_resblocks_3_mlp_c_fc_weight_to_fp16 = const()[name = tensor("transformer_resblocks_3_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52061312)))]; + tensor transformer_resblocks_3_mlp_c_fc_bias_to_fp16 = const()[name = tensor("transformer_resblocks_3_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56779968)))]; + tensor linear_14_cast_fp16 = linear(bias = transformer_resblocks_3_mlp_c_fc_bias_to_fp16, weight = transformer_resblocks_3_mlp_c_fc_weight_to_fp16, x = ret_17_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_527_to_fp16 = const()[name = tensor("op_527_to_fp16"), val = tensor(0x1.b3cp+0)]; + tensor var_528_cast_fp16 = mul(x = linear_14_cast_fp16, y = var_527_to_fp16)[name = tensor("op_528_cast_fp16")]; + tensor var_529_cast_fp16 = sigmoid(x = var_528_cast_fp16)[name = tensor("op_529_cast_fp16")]; + tensor input_33_cast_fp16 = mul(x = linear_14_cast_fp16, y = var_529_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor transformer_resblocks_3_mlp_c_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_3_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56786176)))]; + tensor transformer_resblocks_3_mlp_c_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_3_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61504832)))]; + tensor linear_15_cast_fp16 = linear(bias = transformer_resblocks_3_mlp_c_proj_bias_to_fp16, weight = transformer_resblocks_3_mlp_c_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor x_39_cast_fp16 = add(x = x_35_cast_fp16, y = linear_15_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor ret_19_axes_0 = const()[name = tensor("ret_19_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_4_ln_1_weight_to_fp16 = const()[name = tensor("transformer_resblocks_4_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61506432)))]; + tensor transformer_resblocks_4_ln_1_bias_to_fp16 = const()[name = tensor("transformer_resblocks_4_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61508032)))]; + tensor ret_19_cast_fp16 = layer_norm(axes = ret_19_axes_0, beta = transformer_resblocks_4_ln_1_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_4_ln_1_weight_to_fp16, x = x_39_cast_fp16)[name = tensor("ret_19_cast_fp16")]; + tensor transformer_resblocks_4_attn_in_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_4_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61509632)))]; + tensor transformer_resblocks_4_attn_in_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_4_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65048640)))]; + tensor linear_16_cast_fp16 = linear(bias = transformer_resblocks_4_attn_in_proj_bias_to_fp16, weight = transformer_resblocks_4_attn_in_proj_weight_to_fp16, x = ret_19_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor concat_38x = const()[name = tensor("concat_38x"), val = tensor([50, -1, 3, 768])]; + tensor var_570_cast_fp16 = reshape(shape = concat_38x, x = linear_16_cast_fp16)[name = tensor("op_570_cast_fp16")]; + tensor var_571_axes_0 = const()[name = tensor("op_571_axes_0"), val = tensor([0])]; + tensor var_571_cast_fp16 = expand_dims(axes = var_571_axes_0, x = var_570_cast_fp16)[name = tensor("op_571_cast_fp16")]; + tensor var_572_perm_0 = const()[name = tensor("op_572_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_573_axes_0 = const()[name = tensor("op_573_axes_0"), val = tensor([-2])]; + tensor var_572_cast_fp16 = transpose(perm = var_572_perm_0, x = var_571_cast_fp16)[name = tensor("transpose_41")]; + tensor var_573_cast_fp16 = squeeze(axes = var_573_axes_0, x = var_572_cast_fp16)[name = tensor("op_573_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 50, 0, 768])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = var_573_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([2, 50, 0, 768])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = var_573_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([3, 50, 0, 768])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = var_573_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor concat_39x = const()[name = tensor("concat_39x"), val = tensor([50, -1, 64])]; + tensor var_582_cast_fp16 = reshape(shape = concat_39x, x = q_25_cast_fp16)[name = tensor("op_582_cast_fp16")]; + tensor q_27_perm_0 = const()[name = tensor("q_27_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_40x = const()[name = tensor("concat_40x"), val = tensor([50, -1, 64])]; + tensor var_589_cast_fp16 = reshape(shape = concat_40x, x = k_25_cast_fp16)[name = tensor("op_589_cast_fp16")]; + tensor k_27_perm_0 = const()[name = tensor("k_27_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_41x = const()[name = tensor("concat_41x"), val = tensor([50, -1, 64])]; + tensor var_596_cast_fp16 = reshape(shape = concat_41x, x = v_25_cast_fp16)[name = tensor("op_596_cast_fp16")]; + tensor v_27_perm_0 = const()[name = tensor("v_27_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_42x = const()[name = tensor("concat_42x"), val = tensor([-1, 12, 50, 64])]; + tensor q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_582_cast_fp16)[name = tensor("transpose_40")]; + tensor q_29_cast_fp16 = reshape(shape = concat_42x, x = q_27_cast_fp16)[name = tensor("q_29_cast_fp16")]; + tensor concat_43x = const()[name = tensor("concat_43x"), val = tensor([-1, 12, 50, 64])]; + tensor k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_589_cast_fp16)[name = tensor("transpose_39")]; + tensor k_29_cast_fp16 = reshape(shape = concat_43x, x = k_27_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor concat_44x = const()[name = tensor("concat_44x"), val = tensor([-1, 12, 50, 64])]; + tensor v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_596_cast_fp16)[name = tensor("transpose_38")]; + tensor v_29_cast_fp16 = reshape(shape = concat_44x, x = v_27_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor mul_9_y_0_to_fp16 = const()[name = tensor("mul_9_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_9_cast_fp16 = mul(x = q_29_cast_fp16, y = mul_9_y_0_to_fp16)[name = tensor("mul_9_cast_fp16")]; + tensor matmul_4_transpose_y_0 = const()[name = tensor("matmul_4_transpose_y_0"), val = tensor(true)]; + tensor matmul_4_transpose_x_0 = const()[name = tensor("matmul_4_transpose_x_0"), val = tensor(false)]; + tensor matmul_4_cast_fp16 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9_cast_fp16, y = k_29_cast_fp16)[name = tensor("matmul_4_cast_fp16")]; + tensor softmax_4_axis_0 = const()[name = tensor("softmax_4_axis_0"), val = tensor(-1)]; + tensor softmax_4_cast_fp16 = softmax(axis = softmax_4_axis_0, x = matmul_4_cast_fp16)[name = tensor("softmax_4_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_4_cast_fp16, y = v_29_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_606 = const()[name = tensor("op_606"), val = tensor([2, 0, 1, 3])]; + tensor concat_45x = const()[name = tensor("concat_45x"), val = tensor([-1, 768])]; + tensor var_607_cast_fp16 = transpose(perm = var_606, x = attn_output_25_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_output_27_cast_fp16 = reshape(shape = concat_45x, x = var_607_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor transformer_resblocks_4_attn_out_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_4_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65053312)))]; + tensor transformer_resblocks_4_attn_out_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_4_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66233024)))]; + tensor linear_17_cast_fp16 = linear(bias = transformer_resblocks_4_attn_out_proj_bias_to_fp16, weight = transformer_resblocks_4_attn_out_proj_weight_to_fp16, x = attn_output_27_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor concat_46x = const()[name = tensor("concat_46x"), val = tensor([50, -1, 768])]; + tensor var_616_cast_fp16 = reshape(shape = concat_46x, x = linear_17_cast_fp16)[name = tensor("op_616_cast_fp16")]; + tensor x_41_cast_fp16 = add(x = x_39_cast_fp16, y = var_616_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor ret_21_axes_0 = const()[name = tensor("ret_21_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_4_ln_2_weight_to_fp16 = const()[name = tensor("transformer_resblocks_4_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66234624)))]; + tensor transformer_resblocks_4_ln_2_bias_to_fp16 = const()[name = tensor("transformer_resblocks_4_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66236224)))]; + tensor ret_21_cast_fp16 = layer_norm(axes = ret_21_axes_0, beta = transformer_resblocks_4_ln_2_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_4_ln_2_weight_to_fp16, x = x_41_cast_fp16)[name = tensor("ret_21_cast_fp16")]; + tensor transformer_resblocks_4_mlp_c_fc_weight_to_fp16 = const()[name = tensor("transformer_resblocks_4_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66237824)))]; + tensor transformer_resblocks_4_mlp_c_fc_bias_to_fp16 = const()[name = tensor("transformer_resblocks_4_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70956480)))]; + tensor linear_18_cast_fp16 = linear(bias = transformer_resblocks_4_mlp_c_fc_bias_to_fp16, weight = transformer_resblocks_4_mlp_c_fc_weight_to_fp16, x = ret_21_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_632_to_fp16 = const()[name = tensor("op_632_to_fp16"), val = tensor(0x1.b3cp+0)]; + tensor var_633_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_632_to_fp16)[name = tensor("op_633_cast_fp16")]; + tensor var_634_cast_fp16 = sigmoid(x = var_633_cast_fp16)[name = tensor("op_634_cast_fp16")]; + tensor input_41_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_634_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor transformer_resblocks_4_mlp_c_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_4_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70962688)))]; + tensor transformer_resblocks_4_mlp_c_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_4_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75681344)))]; + tensor linear_19_cast_fp16 = linear(bias = transformer_resblocks_4_mlp_c_proj_bias_to_fp16, weight = transformer_resblocks_4_mlp_c_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor x_45_cast_fp16 = add(x = x_41_cast_fp16, y = linear_19_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor ret_23_axes_0 = const()[name = tensor("ret_23_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_5_ln_1_weight_to_fp16 = const()[name = tensor("transformer_resblocks_5_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75682944)))]; + tensor transformer_resblocks_5_ln_1_bias_to_fp16 = const()[name = tensor("transformer_resblocks_5_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75684544)))]; + tensor ret_23_cast_fp16 = layer_norm(axes = ret_23_axes_0, beta = transformer_resblocks_5_ln_1_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_5_ln_1_weight_to_fp16, x = x_45_cast_fp16)[name = tensor("ret_23_cast_fp16")]; + tensor transformer_resblocks_5_attn_in_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_5_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75686144)))]; + tensor transformer_resblocks_5_attn_in_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_5_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79225152)))]; + tensor linear_20_cast_fp16 = linear(bias = transformer_resblocks_5_attn_in_proj_bias_to_fp16, weight = transformer_resblocks_5_attn_in_proj_weight_to_fp16, x = ret_23_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor concat_47x = const()[name = tensor("concat_47x"), val = tensor([50, -1, 3, 768])]; + tensor var_675_cast_fp16 = reshape(shape = concat_47x, x = linear_20_cast_fp16)[name = tensor("op_675_cast_fp16")]; + tensor var_676_axes_0 = const()[name = tensor("op_676_axes_0"), val = tensor([0])]; + tensor var_676_cast_fp16 = expand_dims(axes = var_676_axes_0, x = var_675_cast_fp16)[name = tensor("op_676_cast_fp16")]; + tensor var_677_perm_0 = const()[name = tensor("op_677_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_678_axes_0 = const()[name = tensor("op_678_axes_0"), val = tensor([-2])]; + tensor var_677_cast_fp16 = transpose(perm = var_677_perm_0, x = var_676_cast_fp16)[name = tensor("transpose_36")]; + tensor var_678_cast_fp16 = squeeze(axes = var_678_axes_0, x = var_677_cast_fp16)[name = tensor("op_678_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 50, 0, 768])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = var_678_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_31_begin_0 = const()[name = tensor("k_31_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_31_end_0 = const()[name = tensor("k_31_end_0"), val = tensor([2, 50, 0, 768])]; + tensor k_31_end_mask_0 = const()[name = tensor("k_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_31_squeeze_mask_0 = const()[name = tensor("k_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_31_cast_fp16 = slice_by_index(begin = k_31_begin_0, end = k_31_end_0, end_mask = k_31_end_mask_0, squeeze_mask = k_31_squeeze_mask_0, x = var_678_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([3, 50, 0, 768])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = var_678_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor concat_48x = const()[name = tensor("concat_48x"), val = tensor([50, -1, 64])]; + tensor var_687_cast_fp16 = reshape(shape = concat_48x, x = q_31_cast_fp16)[name = tensor("op_687_cast_fp16")]; + tensor q_33_perm_0 = const()[name = tensor("q_33_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_49x = const()[name = tensor("concat_49x"), val = tensor([50, -1, 64])]; + tensor var_694_cast_fp16 = reshape(shape = concat_49x, x = k_31_cast_fp16)[name = tensor("op_694_cast_fp16")]; + tensor k_33_perm_0 = const()[name = tensor("k_33_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_50x = const()[name = tensor("concat_50x"), val = tensor([50, -1, 64])]; + tensor var_701_cast_fp16 = reshape(shape = concat_50x, x = v_31_cast_fp16)[name = tensor("op_701_cast_fp16")]; + tensor v_33_perm_0 = const()[name = tensor("v_33_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_51x = const()[name = tensor("concat_51x"), val = tensor([-1, 12, 50, 64])]; + tensor q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_687_cast_fp16)[name = tensor("transpose_35")]; + tensor q_35_cast_fp16 = reshape(shape = concat_51x, x = q_33_cast_fp16)[name = tensor("q_35_cast_fp16")]; + tensor concat_52x = const()[name = tensor("concat_52x"), val = tensor([-1, 12, 50, 64])]; + tensor k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_694_cast_fp16)[name = tensor("transpose_34")]; + tensor k_35_cast_fp16 = reshape(shape = concat_52x, x = k_33_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor concat_53x = const()[name = tensor("concat_53x"), val = tensor([-1, 12, 50, 64])]; + tensor v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_701_cast_fp16)[name = tensor("transpose_33")]; + tensor v_35_cast_fp16 = reshape(shape = concat_53x, x = v_33_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor mul_11_y_0_to_fp16 = const()[name = tensor("mul_11_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_11_cast_fp16 = mul(x = q_35_cast_fp16, y = mul_11_y_0_to_fp16)[name = tensor("mul_11_cast_fp16")]; + tensor matmul_5_transpose_y_0 = const()[name = tensor("matmul_5_transpose_y_0"), val = tensor(true)]; + tensor matmul_5_transpose_x_0 = const()[name = tensor("matmul_5_transpose_x_0"), val = tensor(false)]; + tensor matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11_cast_fp16, y = k_35_cast_fp16)[name = tensor("matmul_5_cast_fp16")]; + tensor softmax_5_axis_0 = const()[name = tensor("softmax_5_axis_0"), val = tensor(-1)]; + tensor softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = matmul_5_cast_fp16)[name = tensor("softmax_5_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = softmax_5_cast_fp16, y = v_35_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_711 = const()[name = tensor("op_711"), val = tensor([2, 0, 1, 3])]; + tensor concat_54x = const()[name = tensor("concat_54x"), val = tensor([-1, 768])]; + tensor var_712_cast_fp16 = transpose(perm = var_711, x = attn_output_31_cast_fp16)[name = tensor("transpose_32")]; + tensor attn_output_33_cast_fp16 = reshape(shape = concat_54x, x = var_712_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor transformer_resblocks_5_attn_out_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_5_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79229824)))]; + tensor transformer_resblocks_5_attn_out_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_5_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80409536)))]; + tensor linear_21_cast_fp16 = linear(bias = transformer_resblocks_5_attn_out_proj_bias_to_fp16, weight = transformer_resblocks_5_attn_out_proj_weight_to_fp16, x = attn_output_33_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor concat_55x = const()[name = tensor("concat_55x"), val = tensor([50, -1, 768])]; + tensor var_721_cast_fp16 = reshape(shape = concat_55x, x = linear_21_cast_fp16)[name = tensor("op_721_cast_fp16")]; + tensor x_47_cast_fp16 = add(x = x_45_cast_fp16, y = var_721_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor ret_25_axes_0 = const()[name = tensor("ret_25_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_5_ln_2_weight_to_fp16 = const()[name = tensor("transformer_resblocks_5_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80411136)))]; + tensor transformer_resblocks_5_ln_2_bias_to_fp16 = const()[name = tensor("transformer_resblocks_5_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80412736)))]; + tensor ret_25_cast_fp16 = layer_norm(axes = ret_25_axes_0, beta = transformer_resblocks_5_ln_2_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_5_ln_2_weight_to_fp16, x = x_47_cast_fp16)[name = tensor("ret_25_cast_fp16")]; + tensor transformer_resblocks_5_mlp_c_fc_weight_to_fp16 = const()[name = tensor("transformer_resblocks_5_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80414336)))]; + tensor transformer_resblocks_5_mlp_c_fc_bias_to_fp16 = const()[name = tensor("transformer_resblocks_5_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85132992)))]; + tensor linear_22_cast_fp16 = linear(bias = transformer_resblocks_5_mlp_c_fc_bias_to_fp16, weight = transformer_resblocks_5_mlp_c_fc_weight_to_fp16, x = ret_25_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_737_to_fp16 = const()[name = tensor("op_737_to_fp16"), val = tensor(0x1.b3cp+0)]; + tensor var_738_cast_fp16 = mul(x = linear_22_cast_fp16, y = var_737_to_fp16)[name = tensor("op_738_cast_fp16")]; + tensor var_739_cast_fp16 = sigmoid(x = var_738_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor input_49_cast_fp16 = mul(x = linear_22_cast_fp16, y = var_739_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor transformer_resblocks_5_mlp_c_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_5_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85139200)))]; + tensor transformer_resblocks_5_mlp_c_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_5_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89857856)))]; + tensor linear_23_cast_fp16 = linear(bias = transformer_resblocks_5_mlp_c_proj_bias_to_fp16, weight = transformer_resblocks_5_mlp_c_proj_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor x_51_cast_fp16 = add(x = x_47_cast_fp16, y = linear_23_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor ret_27_axes_0 = const()[name = tensor("ret_27_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_6_ln_1_weight_to_fp16 = const()[name = tensor("transformer_resblocks_6_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89859456)))]; + tensor transformer_resblocks_6_ln_1_bias_to_fp16 = const()[name = tensor("transformer_resblocks_6_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89861056)))]; + tensor ret_27_cast_fp16 = layer_norm(axes = ret_27_axes_0, beta = transformer_resblocks_6_ln_1_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_6_ln_1_weight_to_fp16, x = x_51_cast_fp16)[name = tensor("ret_27_cast_fp16")]; + tensor transformer_resblocks_6_attn_in_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_6_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89862656)))]; + tensor transformer_resblocks_6_attn_in_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_6_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93401664)))]; + tensor linear_24_cast_fp16 = linear(bias = transformer_resblocks_6_attn_in_proj_bias_to_fp16, weight = transformer_resblocks_6_attn_in_proj_weight_to_fp16, x = ret_27_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor concat_56x = const()[name = tensor("concat_56x"), val = tensor([50, -1, 3, 768])]; + tensor var_780_cast_fp16 = reshape(shape = concat_56x, x = linear_24_cast_fp16)[name = tensor("op_780_cast_fp16")]; + tensor var_781_axes_0 = const()[name = tensor("op_781_axes_0"), val = tensor([0])]; + tensor var_781_cast_fp16 = expand_dims(axes = var_781_axes_0, x = var_780_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor var_782_perm_0 = const()[name = tensor("op_782_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_783_axes_0 = const()[name = tensor("op_783_axes_0"), val = tensor([-2])]; + tensor var_782_cast_fp16 = transpose(perm = var_782_perm_0, x = var_781_cast_fp16)[name = tensor("transpose_31")]; + tensor var_783_cast_fp16 = squeeze(axes = var_783_axes_0, x = var_782_cast_fp16)[name = tensor("op_783_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 50, 0, 768])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = var_783_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([2, 50, 0, 768])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = var_783_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([3, 50, 0, 768])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = var_783_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor concat_57x = const()[name = tensor("concat_57x"), val = tensor([50, -1, 64])]; + tensor var_792_cast_fp16 = reshape(shape = concat_57x, x = q_37_cast_fp16)[name = tensor("op_792_cast_fp16")]; + tensor q_39_perm_0 = const()[name = tensor("q_39_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_58x = const()[name = tensor("concat_58x"), val = tensor([50, -1, 64])]; + tensor var_799_cast_fp16 = reshape(shape = concat_58x, x = k_37_cast_fp16)[name = tensor("op_799_cast_fp16")]; + tensor k_39_perm_0 = const()[name = tensor("k_39_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_59x = const()[name = tensor("concat_59x"), val = tensor([50, -1, 64])]; + tensor var_806_cast_fp16 = reshape(shape = concat_59x, x = v_37_cast_fp16)[name = tensor("op_806_cast_fp16")]; + tensor v_39_perm_0 = const()[name = tensor("v_39_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_60x = const()[name = tensor("concat_60x"), val = tensor([-1, 12, 50, 64])]; + tensor q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_792_cast_fp16)[name = tensor("transpose_30")]; + tensor q_41_cast_fp16 = reshape(shape = concat_60x, x = q_39_cast_fp16)[name = tensor("q_41_cast_fp16")]; + tensor concat_61x = const()[name = tensor("concat_61x"), val = tensor([-1, 12, 50, 64])]; + tensor k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_799_cast_fp16)[name = tensor("transpose_29")]; + tensor k_41_cast_fp16 = reshape(shape = concat_61x, x = k_39_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor concat_62x = const()[name = tensor("concat_62x"), val = tensor([-1, 12, 50, 64])]; + tensor v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_806_cast_fp16)[name = tensor("transpose_28")]; + tensor v_41_cast_fp16 = reshape(shape = concat_62x, x = v_39_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor mul_13_y_0_to_fp16 = const()[name = tensor("mul_13_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_13_cast_fp16 = mul(x = q_41_cast_fp16, y = mul_13_y_0_to_fp16)[name = tensor("mul_13_cast_fp16")]; + tensor matmul_6_transpose_y_0 = const()[name = tensor("matmul_6_transpose_y_0"), val = tensor(true)]; + tensor matmul_6_transpose_x_0 = const()[name = tensor("matmul_6_transpose_x_0"), val = tensor(false)]; + tensor matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13_cast_fp16, y = k_41_cast_fp16)[name = tensor("matmul_6_cast_fp16")]; + tensor softmax_6_axis_0 = const()[name = tensor("softmax_6_axis_0"), val = tensor(-1)]; + tensor softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = matmul_6_cast_fp16)[name = tensor("softmax_6_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_6_cast_fp16, y = v_41_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_816 = const()[name = tensor("op_816"), val = tensor([2, 0, 1, 3])]; + tensor concat_63x = const()[name = tensor("concat_63x"), val = tensor([-1, 768])]; + tensor var_817_cast_fp16 = transpose(perm = var_816, x = attn_output_37_cast_fp16)[name = tensor("transpose_27")]; + tensor attn_output_39_cast_fp16 = reshape(shape = concat_63x, x = var_817_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor transformer_resblocks_6_attn_out_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_6_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93406336)))]; + tensor transformer_resblocks_6_attn_out_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_6_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94586048)))]; + tensor linear_25_cast_fp16 = linear(bias = transformer_resblocks_6_attn_out_proj_bias_to_fp16, weight = transformer_resblocks_6_attn_out_proj_weight_to_fp16, x = attn_output_39_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor concat_64x = const()[name = tensor("concat_64x"), val = tensor([50, -1, 768])]; + tensor var_826_cast_fp16 = reshape(shape = concat_64x, x = linear_25_cast_fp16)[name = tensor("op_826_cast_fp16")]; + tensor x_53_cast_fp16 = add(x = x_51_cast_fp16, y = var_826_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor ret_29_axes_0 = const()[name = tensor("ret_29_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_6_ln_2_weight_to_fp16 = const()[name = tensor("transformer_resblocks_6_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94587648)))]; + tensor transformer_resblocks_6_ln_2_bias_to_fp16 = const()[name = tensor("transformer_resblocks_6_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94589248)))]; + tensor ret_29_cast_fp16 = layer_norm(axes = ret_29_axes_0, beta = transformer_resblocks_6_ln_2_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_6_ln_2_weight_to_fp16, x = x_53_cast_fp16)[name = tensor("ret_29_cast_fp16")]; + tensor transformer_resblocks_6_mlp_c_fc_weight_to_fp16 = const()[name = tensor("transformer_resblocks_6_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94590848)))]; + tensor transformer_resblocks_6_mlp_c_fc_bias_to_fp16 = const()[name = tensor("transformer_resblocks_6_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99309504)))]; + tensor linear_26_cast_fp16 = linear(bias = transformer_resblocks_6_mlp_c_fc_bias_to_fp16, weight = transformer_resblocks_6_mlp_c_fc_weight_to_fp16, x = ret_29_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_842_to_fp16 = const()[name = tensor("op_842_to_fp16"), val = tensor(0x1.b3cp+0)]; + tensor var_843_cast_fp16 = mul(x = linear_26_cast_fp16, y = var_842_to_fp16)[name = tensor("op_843_cast_fp16")]; + tensor var_844_cast_fp16 = sigmoid(x = var_843_cast_fp16)[name = tensor("op_844_cast_fp16")]; + tensor input_57_cast_fp16 = mul(x = linear_26_cast_fp16, y = var_844_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor transformer_resblocks_6_mlp_c_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_6_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99315712)))]; + tensor transformer_resblocks_6_mlp_c_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_6_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104034368)))]; + tensor linear_27_cast_fp16 = linear(bias = transformer_resblocks_6_mlp_c_proj_bias_to_fp16, weight = transformer_resblocks_6_mlp_c_proj_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor x_57_cast_fp16 = add(x = x_53_cast_fp16, y = linear_27_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor ret_31_axes_0 = const()[name = tensor("ret_31_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_7_ln_1_weight_to_fp16 = const()[name = tensor("transformer_resblocks_7_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104035968)))]; + tensor transformer_resblocks_7_ln_1_bias_to_fp16 = const()[name = tensor("transformer_resblocks_7_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104037568)))]; + tensor ret_31_cast_fp16 = layer_norm(axes = ret_31_axes_0, beta = transformer_resblocks_7_ln_1_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_7_ln_1_weight_to_fp16, x = x_57_cast_fp16)[name = tensor("ret_31_cast_fp16")]; + tensor transformer_resblocks_7_attn_in_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_7_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104039168)))]; + tensor transformer_resblocks_7_attn_in_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_7_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107578176)))]; + tensor linear_28_cast_fp16 = linear(bias = transformer_resblocks_7_attn_in_proj_bias_to_fp16, weight = transformer_resblocks_7_attn_in_proj_weight_to_fp16, x = ret_31_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor concat_65x = const()[name = tensor("concat_65x"), val = tensor([50, -1, 3, 768])]; + tensor var_885_cast_fp16 = reshape(shape = concat_65x, x = linear_28_cast_fp16)[name = tensor("op_885_cast_fp16")]; + tensor var_886_axes_0 = const()[name = tensor("op_886_axes_0"), val = tensor([0])]; + tensor var_886_cast_fp16 = expand_dims(axes = var_886_axes_0, x = var_885_cast_fp16)[name = tensor("op_886_cast_fp16")]; + tensor var_887_perm_0 = const()[name = tensor("op_887_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_888_axes_0 = const()[name = tensor("op_888_axes_0"), val = tensor([-2])]; + tensor var_887_cast_fp16 = transpose(perm = var_887_perm_0, x = var_886_cast_fp16)[name = tensor("transpose_26")]; + tensor var_888_cast_fp16 = squeeze(axes = var_888_axes_0, x = var_887_cast_fp16)[name = tensor("op_888_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 50, 0, 768])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = var_888_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_43_begin_0 = const()[name = tensor("k_43_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_43_end_0 = const()[name = tensor("k_43_end_0"), val = tensor([2, 50, 0, 768])]; + tensor k_43_end_mask_0 = const()[name = tensor("k_43_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_43_squeeze_mask_0 = const()[name = tensor("k_43_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_43_cast_fp16 = slice_by_index(begin = k_43_begin_0, end = k_43_end_0, end_mask = k_43_end_mask_0, squeeze_mask = k_43_squeeze_mask_0, x = var_888_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([3, 50, 0, 768])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = var_888_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor concat_66x = const()[name = tensor("concat_66x"), val = tensor([50, -1, 64])]; + tensor var_897_cast_fp16 = reshape(shape = concat_66x, x = q_43_cast_fp16)[name = tensor("op_897_cast_fp16")]; + tensor q_45_perm_0 = const()[name = tensor("q_45_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_67x = const()[name = tensor("concat_67x"), val = tensor([50, -1, 64])]; + tensor var_904_cast_fp16 = reshape(shape = concat_67x, x = k_43_cast_fp16)[name = tensor("op_904_cast_fp16")]; + tensor k_45_perm_0 = const()[name = tensor("k_45_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_68x = const()[name = tensor("concat_68x"), val = tensor([50, -1, 64])]; + tensor var_911_cast_fp16 = reshape(shape = concat_68x, x = v_43_cast_fp16)[name = tensor("op_911_cast_fp16")]; + tensor v_45_perm_0 = const()[name = tensor("v_45_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_69x = const()[name = tensor("concat_69x"), val = tensor([-1, 12, 50, 64])]; + tensor q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_897_cast_fp16)[name = tensor("transpose_25")]; + tensor q_47_cast_fp16 = reshape(shape = concat_69x, x = q_45_cast_fp16)[name = tensor("q_47_cast_fp16")]; + tensor concat_70x = const()[name = tensor("concat_70x"), val = tensor([-1, 12, 50, 64])]; + tensor k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_904_cast_fp16)[name = tensor("transpose_24")]; + tensor k_47_cast_fp16 = reshape(shape = concat_70x, x = k_45_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor concat_71x = const()[name = tensor("concat_71x"), val = tensor([-1, 12, 50, 64])]; + tensor v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_911_cast_fp16)[name = tensor("transpose_23")]; + tensor v_47_cast_fp16 = reshape(shape = concat_71x, x = v_45_cast_fp16)[name = tensor("v_47_cast_fp16")]; + tensor mul_15_y_0_to_fp16 = const()[name = tensor("mul_15_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_15_cast_fp16 = mul(x = q_47_cast_fp16, y = mul_15_y_0_to_fp16)[name = tensor("mul_15_cast_fp16")]; + tensor matmul_7_transpose_y_0 = const()[name = tensor("matmul_7_transpose_y_0"), val = tensor(true)]; + tensor matmul_7_transpose_x_0 = const()[name = tensor("matmul_7_transpose_x_0"), val = tensor(false)]; + tensor matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15_cast_fp16, y = k_47_cast_fp16)[name = tensor("matmul_7_cast_fp16")]; + tensor softmax_7_axis_0 = const()[name = tensor("softmax_7_axis_0"), val = tensor(-1)]; + tensor softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = matmul_7_cast_fp16)[name = tensor("softmax_7_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = softmax_7_cast_fp16, y = v_47_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor([2, 0, 1, 3])]; + tensor concat_72x = const()[name = tensor("concat_72x"), val = tensor([-1, 768])]; + tensor var_922_cast_fp16 = transpose(perm = var_921, x = attn_output_43_cast_fp16)[name = tensor("transpose_22")]; + tensor attn_output_45_cast_fp16 = reshape(shape = concat_72x, x = var_922_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor transformer_resblocks_7_attn_out_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_7_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107582848)))]; + tensor transformer_resblocks_7_attn_out_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_7_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108762560)))]; + tensor linear_29_cast_fp16 = linear(bias = transformer_resblocks_7_attn_out_proj_bias_to_fp16, weight = transformer_resblocks_7_attn_out_proj_weight_to_fp16, x = attn_output_45_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor concat_73x = const()[name = tensor("concat_73x"), val = tensor([50, -1, 768])]; + tensor var_931_cast_fp16 = reshape(shape = concat_73x, x = linear_29_cast_fp16)[name = tensor("op_931_cast_fp16")]; + tensor x_59_cast_fp16 = add(x = x_57_cast_fp16, y = var_931_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor ret_33_axes_0 = const()[name = tensor("ret_33_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_7_ln_2_weight_to_fp16 = const()[name = tensor("transformer_resblocks_7_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108764160)))]; + tensor transformer_resblocks_7_ln_2_bias_to_fp16 = const()[name = tensor("transformer_resblocks_7_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108765760)))]; + tensor ret_33_cast_fp16 = layer_norm(axes = ret_33_axes_0, beta = transformer_resblocks_7_ln_2_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_7_ln_2_weight_to_fp16, x = x_59_cast_fp16)[name = tensor("ret_33_cast_fp16")]; + tensor transformer_resblocks_7_mlp_c_fc_weight_to_fp16 = const()[name = tensor("transformer_resblocks_7_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108767360)))]; + tensor transformer_resblocks_7_mlp_c_fc_bias_to_fp16 = const()[name = tensor("transformer_resblocks_7_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113486016)))]; + tensor linear_30_cast_fp16 = linear(bias = transformer_resblocks_7_mlp_c_fc_bias_to_fp16, weight = transformer_resblocks_7_mlp_c_fc_weight_to_fp16, x = ret_33_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_947_to_fp16 = const()[name = tensor("op_947_to_fp16"), val = tensor(0x1.b3cp+0)]; + tensor var_948_cast_fp16 = mul(x = linear_30_cast_fp16, y = var_947_to_fp16)[name = tensor("op_948_cast_fp16")]; + tensor var_949_cast_fp16 = sigmoid(x = var_948_cast_fp16)[name = tensor("op_949_cast_fp16")]; + tensor input_65_cast_fp16 = mul(x = linear_30_cast_fp16, y = var_949_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor transformer_resblocks_7_mlp_c_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_7_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113492224)))]; + tensor transformer_resblocks_7_mlp_c_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_7_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118210880)))]; + tensor linear_31_cast_fp16 = linear(bias = transformer_resblocks_7_mlp_c_proj_bias_to_fp16, weight = transformer_resblocks_7_mlp_c_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor x_63_cast_fp16 = add(x = x_59_cast_fp16, y = linear_31_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor ret_35_axes_0 = const()[name = tensor("ret_35_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_8_ln_1_weight_to_fp16 = const()[name = tensor("transformer_resblocks_8_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118212480)))]; + tensor transformer_resblocks_8_ln_1_bias_to_fp16 = const()[name = tensor("transformer_resblocks_8_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118214080)))]; + tensor ret_35_cast_fp16 = layer_norm(axes = ret_35_axes_0, beta = transformer_resblocks_8_ln_1_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_8_ln_1_weight_to_fp16, x = x_63_cast_fp16)[name = tensor("ret_35_cast_fp16")]; + tensor transformer_resblocks_8_attn_in_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_8_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118215680)))]; + tensor transformer_resblocks_8_attn_in_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_8_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121754688)))]; + tensor linear_32_cast_fp16 = linear(bias = transformer_resblocks_8_attn_in_proj_bias_to_fp16, weight = transformer_resblocks_8_attn_in_proj_weight_to_fp16, x = ret_35_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor concat_74x = const()[name = tensor("concat_74x"), val = tensor([50, -1, 3, 768])]; + tensor var_990_cast_fp16 = reshape(shape = concat_74x, x = linear_32_cast_fp16)[name = tensor("op_990_cast_fp16")]; + tensor var_991_axes_0 = const()[name = tensor("op_991_axes_0"), val = tensor([0])]; + tensor var_991_cast_fp16 = expand_dims(axes = var_991_axes_0, x = var_990_cast_fp16)[name = tensor("op_991_cast_fp16")]; + tensor var_992_perm_0 = const()[name = tensor("op_992_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_993_axes_0 = const()[name = tensor("op_993_axes_0"), val = tensor([-2])]; + tensor var_992_cast_fp16 = transpose(perm = var_992_perm_0, x = var_991_cast_fp16)[name = tensor("transpose_21")]; + tensor var_993_cast_fp16 = squeeze(axes = var_993_axes_0, x = var_992_cast_fp16)[name = tensor("op_993_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 50, 0, 768])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = var_993_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([2, 50, 0, 768])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = var_993_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_49_begin_0 = const()[name = tensor("v_49_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_49_end_0 = const()[name = tensor("v_49_end_0"), val = tensor([3, 50, 0, 768])]; + tensor v_49_end_mask_0 = const()[name = tensor("v_49_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_49_squeeze_mask_0 = const()[name = tensor("v_49_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_49_cast_fp16 = slice_by_index(begin = v_49_begin_0, end = v_49_end_0, end_mask = v_49_end_mask_0, squeeze_mask = v_49_squeeze_mask_0, x = var_993_cast_fp16)[name = tensor("v_49_cast_fp16")]; + tensor concat_75x = const()[name = tensor("concat_75x"), val = tensor([50, -1, 64])]; + tensor var_1002_cast_fp16 = reshape(shape = concat_75x, x = q_49_cast_fp16)[name = tensor("op_1002_cast_fp16")]; + tensor q_51_perm_0 = const()[name = tensor("q_51_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_76x = const()[name = tensor("concat_76x"), val = tensor([50, -1, 64])]; + tensor var_1009_cast_fp16 = reshape(shape = concat_76x, x = k_49_cast_fp16)[name = tensor("op_1009_cast_fp16")]; + tensor k_51_perm_0 = const()[name = tensor("k_51_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_77x = const()[name = tensor("concat_77x"), val = tensor([50, -1, 64])]; + tensor var_1016_cast_fp16 = reshape(shape = concat_77x, x = v_49_cast_fp16)[name = tensor("op_1016_cast_fp16")]; + tensor v_51_perm_0 = const()[name = tensor("v_51_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_78x = const()[name = tensor("concat_78x"), val = tensor([-1, 12, 50, 64])]; + tensor q_51_cast_fp16 = transpose(perm = q_51_perm_0, x = var_1002_cast_fp16)[name = tensor("transpose_20")]; + tensor q_53_cast_fp16 = reshape(shape = concat_78x, x = q_51_cast_fp16)[name = tensor("q_53_cast_fp16")]; + tensor concat_79x = const()[name = tensor("concat_79x"), val = tensor([-1, 12, 50, 64])]; + tensor k_51_cast_fp16 = transpose(perm = k_51_perm_0, x = var_1009_cast_fp16)[name = tensor("transpose_19")]; + tensor k_53_cast_fp16 = reshape(shape = concat_79x, x = k_51_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor concat_80x = const()[name = tensor("concat_80x"), val = tensor([-1, 12, 50, 64])]; + tensor v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = var_1016_cast_fp16)[name = tensor("transpose_18")]; + tensor v_53_cast_fp16 = reshape(shape = concat_80x, x = v_51_cast_fp16)[name = tensor("v_53_cast_fp16")]; + tensor mul_17_y_0_to_fp16 = const()[name = tensor("mul_17_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_17_cast_fp16 = mul(x = q_53_cast_fp16, y = mul_17_y_0_to_fp16)[name = tensor("mul_17_cast_fp16")]; + tensor matmul_8_transpose_y_0 = const()[name = tensor("matmul_8_transpose_y_0"), val = tensor(true)]; + tensor matmul_8_transpose_x_0 = const()[name = tensor("matmul_8_transpose_x_0"), val = tensor(false)]; + tensor matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17_cast_fp16, y = k_53_cast_fp16)[name = tensor("matmul_8_cast_fp16")]; + tensor softmax_8_axis_0 = const()[name = tensor("softmax_8_axis_0"), val = tensor(-1)]; + tensor softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = matmul_8_cast_fp16)[name = tensor("softmax_8_cast_fp16")]; + tensor attn_output_49_transpose_x_0 = const()[name = tensor("attn_output_49_transpose_x_0"), val = tensor(false)]; + tensor attn_output_49_transpose_y_0 = const()[name = tensor("attn_output_49_transpose_y_0"), val = tensor(false)]; + tensor attn_output_49_cast_fp16 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = softmax_8_cast_fp16, y = v_53_cast_fp16)[name = tensor("attn_output_49_cast_fp16")]; + tensor var_1026 = const()[name = tensor("op_1026"), val = tensor([2, 0, 1, 3])]; + tensor concat_81x = const()[name = tensor("concat_81x"), val = tensor([-1, 768])]; + tensor var_1027_cast_fp16 = transpose(perm = var_1026, x = attn_output_49_cast_fp16)[name = tensor("transpose_17")]; + tensor attn_output_51_cast_fp16 = reshape(shape = concat_81x, x = var_1027_cast_fp16)[name = tensor("attn_output_51_cast_fp16")]; + tensor transformer_resblocks_8_attn_out_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_8_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121759360)))]; + tensor transformer_resblocks_8_attn_out_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_8_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122939072)))]; + tensor linear_33_cast_fp16 = linear(bias = transformer_resblocks_8_attn_out_proj_bias_to_fp16, weight = transformer_resblocks_8_attn_out_proj_weight_to_fp16, x = attn_output_51_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor concat_82x = const()[name = tensor("concat_82x"), val = tensor([50, -1, 768])]; + tensor var_1036_cast_fp16 = reshape(shape = concat_82x, x = linear_33_cast_fp16)[name = tensor("op_1036_cast_fp16")]; + tensor x_65_cast_fp16 = add(x = x_63_cast_fp16, y = var_1036_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor ret_37_axes_0 = const()[name = tensor("ret_37_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_8_ln_2_weight_to_fp16 = const()[name = tensor("transformer_resblocks_8_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122940672)))]; + tensor transformer_resblocks_8_ln_2_bias_to_fp16 = const()[name = tensor("transformer_resblocks_8_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122942272)))]; + tensor ret_37_cast_fp16 = layer_norm(axes = ret_37_axes_0, beta = transformer_resblocks_8_ln_2_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_8_ln_2_weight_to_fp16, x = x_65_cast_fp16)[name = tensor("ret_37_cast_fp16")]; + tensor transformer_resblocks_8_mlp_c_fc_weight_to_fp16 = const()[name = tensor("transformer_resblocks_8_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122943872)))]; + tensor transformer_resblocks_8_mlp_c_fc_bias_to_fp16 = const()[name = tensor("transformer_resblocks_8_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127662528)))]; + tensor linear_34_cast_fp16 = linear(bias = transformer_resblocks_8_mlp_c_fc_bias_to_fp16, weight = transformer_resblocks_8_mlp_c_fc_weight_to_fp16, x = ret_37_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_1052_to_fp16 = const()[name = tensor("op_1052_to_fp16"), val = tensor(0x1.b3cp+0)]; + tensor var_1053_cast_fp16 = mul(x = linear_34_cast_fp16, y = var_1052_to_fp16)[name = tensor("op_1053_cast_fp16")]; + tensor var_1054_cast_fp16 = sigmoid(x = var_1053_cast_fp16)[name = tensor("op_1054_cast_fp16")]; + tensor input_73_cast_fp16 = mul(x = linear_34_cast_fp16, y = var_1054_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor transformer_resblocks_8_mlp_c_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_8_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127668736)))]; + tensor transformer_resblocks_8_mlp_c_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_8_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132387392)))]; + tensor linear_35_cast_fp16 = linear(bias = transformer_resblocks_8_mlp_c_proj_bias_to_fp16, weight = transformer_resblocks_8_mlp_c_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor x_69_cast_fp16 = add(x = x_65_cast_fp16, y = linear_35_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor ret_39_axes_0 = const()[name = tensor("ret_39_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_9_ln_1_weight_to_fp16 = const()[name = tensor("transformer_resblocks_9_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132388992)))]; + tensor transformer_resblocks_9_ln_1_bias_to_fp16 = const()[name = tensor("transformer_resblocks_9_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132390592)))]; + tensor ret_39_cast_fp16 = layer_norm(axes = ret_39_axes_0, beta = transformer_resblocks_9_ln_1_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_9_ln_1_weight_to_fp16, x = x_69_cast_fp16)[name = tensor("ret_39_cast_fp16")]; + tensor transformer_resblocks_9_attn_in_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_9_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132392192)))]; + tensor transformer_resblocks_9_attn_in_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_9_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135931200)))]; + tensor linear_36_cast_fp16 = linear(bias = transformer_resblocks_9_attn_in_proj_bias_to_fp16, weight = transformer_resblocks_9_attn_in_proj_weight_to_fp16, x = ret_39_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor concat_83x = const()[name = tensor("concat_83x"), val = tensor([50, -1, 3, 768])]; + tensor var_1095_cast_fp16 = reshape(shape = concat_83x, x = linear_36_cast_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_axes_0 = const()[name = tensor("op_1096_axes_0"), val = tensor([0])]; + tensor var_1096_cast_fp16 = expand_dims(axes = var_1096_axes_0, x = var_1095_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1097_perm_0 = const()[name = tensor("op_1097_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1098_axes_0 = const()[name = tensor("op_1098_axes_0"), val = tensor([-2])]; + tensor var_1097_cast_fp16 = transpose(perm = var_1097_perm_0, x = var_1096_cast_fp16)[name = tensor("transpose_16")]; + tensor var_1098_cast_fp16 = squeeze(axes = var_1098_axes_0, x = var_1097_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 50, 0, 768])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = var_1098_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_55_begin_0 = const()[name = tensor("k_55_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_55_end_0 = const()[name = tensor("k_55_end_0"), val = tensor([2, 50, 0, 768])]; + tensor k_55_end_mask_0 = const()[name = tensor("k_55_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_55_squeeze_mask_0 = const()[name = tensor("k_55_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_55_cast_fp16 = slice_by_index(begin = k_55_begin_0, end = k_55_end_0, end_mask = k_55_end_mask_0, squeeze_mask = k_55_squeeze_mask_0, x = var_1098_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor v_55_begin_0 = const()[name = tensor("v_55_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_55_end_0 = const()[name = tensor("v_55_end_0"), val = tensor([3, 50, 0, 768])]; + tensor v_55_end_mask_0 = const()[name = tensor("v_55_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_55_squeeze_mask_0 = const()[name = tensor("v_55_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_55_cast_fp16 = slice_by_index(begin = v_55_begin_0, end = v_55_end_0, end_mask = v_55_end_mask_0, squeeze_mask = v_55_squeeze_mask_0, x = var_1098_cast_fp16)[name = tensor("v_55_cast_fp16")]; + tensor concat_84x = const()[name = tensor("concat_84x"), val = tensor([50, -1, 64])]; + tensor var_1107_cast_fp16 = reshape(shape = concat_84x, x = q_55_cast_fp16)[name = tensor("op_1107_cast_fp16")]; + tensor q_57_perm_0 = const()[name = tensor("q_57_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_85x = const()[name = tensor("concat_85x"), val = tensor([50, -1, 64])]; + tensor var_1114_cast_fp16 = reshape(shape = concat_85x, x = k_55_cast_fp16)[name = tensor("op_1114_cast_fp16")]; + tensor k_57_perm_0 = const()[name = tensor("k_57_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_86x = const()[name = tensor("concat_86x"), val = tensor([50, -1, 64])]; + tensor var_1121_cast_fp16 = reshape(shape = concat_86x, x = v_55_cast_fp16)[name = tensor("op_1121_cast_fp16")]; + tensor v_57_perm_0 = const()[name = tensor("v_57_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_87x = const()[name = tensor("concat_87x"), val = tensor([-1, 12, 50, 64])]; + tensor q_57_cast_fp16 = transpose(perm = q_57_perm_0, x = var_1107_cast_fp16)[name = tensor("transpose_15")]; + tensor q_59_cast_fp16 = reshape(shape = concat_87x, x = q_57_cast_fp16)[name = tensor("q_59_cast_fp16")]; + tensor concat_88x = const()[name = tensor("concat_88x"), val = tensor([-1, 12, 50, 64])]; + tensor k_57_cast_fp16 = transpose(perm = k_57_perm_0, x = var_1114_cast_fp16)[name = tensor("transpose_14")]; + tensor k_59_cast_fp16 = reshape(shape = concat_88x, x = k_57_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor concat_89x = const()[name = tensor("concat_89x"), val = tensor([-1, 12, 50, 64])]; + tensor v_57_cast_fp16 = transpose(perm = v_57_perm_0, x = var_1121_cast_fp16)[name = tensor("transpose_13")]; + tensor v_59_cast_fp16 = reshape(shape = concat_89x, x = v_57_cast_fp16)[name = tensor("v_59_cast_fp16")]; + tensor mul_19_y_0_to_fp16 = const()[name = tensor("mul_19_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_19_cast_fp16 = mul(x = q_59_cast_fp16, y = mul_19_y_0_to_fp16)[name = tensor("mul_19_cast_fp16")]; + tensor matmul_9_transpose_y_0 = const()[name = tensor("matmul_9_transpose_y_0"), val = tensor(true)]; + tensor matmul_9_transpose_x_0 = const()[name = tensor("matmul_9_transpose_x_0"), val = tensor(false)]; + tensor matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19_cast_fp16, y = k_59_cast_fp16)[name = tensor("matmul_9_cast_fp16")]; + tensor softmax_9_axis_0 = const()[name = tensor("softmax_9_axis_0"), val = tensor(-1)]; + tensor softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = matmul_9_cast_fp16)[name = tensor("softmax_9_cast_fp16")]; + tensor attn_output_55_transpose_x_0 = const()[name = tensor("attn_output_55_transpose_x_0"), val = tensor(false)]; + tensor attn_output_55_transpose_y_0 = const()[name = tensor("attn_output_55_transpose_y_0"), val = tensor(false)]; + tensor attn_output_55_cast_fp16 = matmul(transpose_x = attn_output_55_transpose_x_0, transpose_y = attn_output_55_transpose_y_0, x = softmax_9_cast_fp16, y = v_59_cast_fp16)[name = tensor("attn_output_55_cast_fp16")]; + tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([2, 0, 1, 3])]; + tensor concat_90x = const()[name = tensor("concat_90x"), val = tensor([-1, 768])]; + tensor var_1132_cast_fp16 = transpose(perm = var_1131, x = attn_output_55_cast_fp16)[name = tensor("transpose_12")]; + tensor attn_output_57_cast_fp16 = reshape(shape = concat_90x, x = var_1132_cast_fp16)[name = tensor("attn_output_57_cast_fp16")]; + tensor transformer_resblocks_9_attn_out_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_9_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135935872)))]; + tensor transformer_resblocks_9_attn_out_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_9_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137115584)))]; + tensor linear_37_cast_fp16 = linear(bias = transformer_resblocks_9_attn_out_proj_bias_to_fp16, weight = transformer_resblocks_9_attn_out_proj_weight_to_fp16, x = attn_output_57_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor concat_91x = const()[name = tensor("concat_91x"), val = tensor([50, -1, 768])]; + tensor var_1141_cast_fp16 = reshape(shape = concat_91x, x = linear_37_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor x_71_cast_fp16 = add(x = x_69_cast_fp16, y = var_1141_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor ret_41_axes_0 = const()[name = tensor("ret_41_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_9_ln_2_weight_to_fp16 = const()[name = tensor("transformer_resblocks_9_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137117184)))]; + tensor transformer_resblocks_9_ln_2_bias_to_fp16 = const()[name = tensor("transformer_resblocks_9_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137118784)))]; + tensor ret_41_cast_fp16 = layer_norm(axes = ret_41_axes_0, beta = transformer_resblocks_9_ln_2_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_9_ln_2_weight_to_fp16, x = x_71_cast_fp16)[name = tensor("ret_41_cast_fp16")]; + tensor transformer_resblocks_9_mlp_c_fc_weight_to_fp16 = const()[name = tensor("transformer_resblocks_9_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137120384)))]; + tensor transformer_resblocks_9_mlp_c_fc_bias_to_fp16 = const()[name = tensor("transformer_resblocks_9_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141839040)))]; + tensor linear_38_cast_fp16 = linear(bias = transformer_resblocks_9_mlp_c_fc_bias_to_fp16, weight = transformer_resblocks_9_mlp_c_fc_weight_to_fp16, x = ret_41_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_1157_to_fp16 = const()[name = tensor("op_1157_to_fp16"), val = tensor(0x1.b3cp+0)]; + tensor var_1158_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1157_to_fp16)[name = tensor("op_1158_cast_fp16")]; + tensor var_1159_cast_fp16 = sigmoid(x = var_1158_cast_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor input_81_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1159_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor transformer_resblocks_9_mlp_c_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_9_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141845248)))]; + tensor transformer_resblocks_9_mlp_c_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_9_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146563904)))]; + tensor linear_39_cast_fp16 = linear(bias = transformer_resblocks_9_mlp_c_proj_bias_to_fp16, weight = transformer_resblocks_9_mlp_c_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor x_75_cast_fp16 = add(x = x_71_cast_fp16, y = linear_39_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor ret_43_axes_0 = const()[name = tensor("ret_43_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_10_ln_1_weight_to_fp16 = const()[name = tensor("transformer_resblocks_10_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146565504)))]; + tensor transformer_resblocks_10_ln_1_bias_to_fp16 = const()[name = tensor("transformer_resblocks_10_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146567104)))]; + tensor ret_43_cast_fp16 = layer_norm(axes = ret_43_axes_0, beta = transformer_resblocks_10_ln_1_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_10_ln_1_weight_to_fp16, x = x_75_cast_fp16)[name = tensor("ret_43_cast_fp16")]; + tensor transformer_resblocks_10_attn_in_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_10_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146568704)))]; + tensor transformer_resblocks_10_attn_in_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_10_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150107712)))]; + tensor linear_40_cast_fp16 = linear(bias = transformer_resblocks_10_attn_in_proj_bias_to_fp16, weight = transformer_resblocks_10_attn_in_proj_weight_to_fp16, x = ret_43_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor concat_92x = const()[name = tensor("concat_92x"), val = tensor([50, -1, 3, 768])]; + tensor var_1200_cast_fp16 = reshape(shape = concat_92x, x = linear_40_cast_fp16)[name = tensor("op_1200_cast_fp16")]; + tensor var_1201_axes_0 = const()[name = tensor("op_1201_axes_0"), val = tensor([0])]; + tensor var_1201_cast_fp16 = expand_dims(axes = var_1201_axes_0, x = var_1200_cast_fp16)[name = tensor("op_1201_cast_fp16")]; + tensor var_1202_perm_0 = const()[name = tensor("op_1202_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1203_axes_0 = const()[name = tensor("op_1203_axes_0"), val = tensor([-2])]; + tensor var_1202_cast_fp16 = transpose(perm = var_1202_perm_0, x = var_1201_cast_fp16)[name = tensor("transpose_11")]; + tensor var_1203_cast_fp16 = squeeze(axes = var_1203_axes_0, x = var_1202_cast_fp16)[name = tensor("op_1203_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 50, 0, 768])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = var_1203_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([2, 50, 0, 768])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = var_1203_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_61_begin_0 = const()[name = tensor("v_61_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_61_end_0 = const()[name = tensor("v_61_end_0"), val = tensor([3, 50, 0, 768])]; + tensor v_61_end_mask_0 = const()[name = tensor("v_61_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_61_squeeze_mask_0 = const()[name = tensor("v_61_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_61_cast_fp16 = slice_by_index(begin = v_61_begin_0, end = v_61_end_0, end_mask = v_61_end_mask_0, squeeze_mask = v_61_squeeze_mask_0, x = var_1203_cast_fp16)[name = tensor("v_61_cast_fp16")]; + tensor concat_93x = const()[name = tensor("concat_93x"), val = tensor([50, -1, 64])]; + tensor var_1212_cast_fp16 = reshape(shape = concat_93x, x = q_61_cast_fp16)[name = tensor("op_1212_cast_fp16")]; + tensor q_63_perm_0 = const()[name = tensor("q_63_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_94x = const()[name = tensor("concat_94x"), val = tensor([50, -1, 64])]; + tensor var_1219_cast_fp16 = reshape(shape = concat_94x, x = k_61_cast_fp16)[name = tensor("op_1219_cast_fp16")]; + tensor k_63_perm_0 = const()[name = tensor("k_63_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_95x = const()[name = tensor("concat_95x"), val = tensor([50, -1, 64])]; + tensor var_1226_cast_fp16 = reshape(shape = concat_95x, x = v_61_cast_fp16)[name = tensor("op_1226_cast_fp16")]; + tensor v_63_perm_0 = const()[name = tensor("v_63_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_96x = const()[name = tensor("concat_96x"), val = tensor([-1, 12, 50, 64])]; + tensor q_63_cast_fp16 = transpose(perm = q_63_perm_0, x = var_1212_cast_fp16)[name = tensor("transpose_10")]; + tensor q_65_cast_fp16 = reshape(shape = concat_96x, x = q_63_cast_fp16)[name = tensor("q_65_cast_fp16")]; + tensor concat_97x = const()[name = tensor("concat_97x"), val = tensor([-1, 12, 50, 64])]; + tensor k_63_cast_fp16 = transpose(perm = k_63_perm_0, x = var_1219_cast_fp16)[name = tensor("transpose_9")]; + tensor k_65_cast_fp16 = reshape(shape = concat_97x, x = k_63_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor concat_98x = const()[name = tensor("concat_98x"), val = tensor([-1, 12, 50, 64])]; + tensor v_63_cast_fp16 = transpose(perm = v_63_perm_0, x = var_1226_cast_fp16)[name = tensor("transpose_8")]; + tensor v_65_cast_fp16 = reshape(shape = concat_98x, x = v_63_cast_fp16)[name = tensor("v_65_cast_fp16")]; + tensor mul_21_y_0_to_fp16 = const()[name = tensor("mul_21_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_21_cast_fp16 = mul(x = q_65_cast_fp16, y = mul_21_y_0_to_fp16)[name = tensor("mul_21_cast_fp16")]; + tensor matmul_10_transpose_y_0 = const()[name = tensor("matmul_10_transpose_y_0"), val = tensor(true)]; + tensor matmul_10_transpose_x_0 = const()[name = tensor("matmul_10_transpose_x_0"), val = tensor(false)]; + tensor matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21_cast_fp16, y = k_65_cast_fp16)[name = tensor("matmul_10_cast_fp16")]; + tensor softmax_10_axis_0 = const()[name = tensor("softmax_10_axis_0"), val = tensor(-1)]; + tensor softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = matmul_10_cast_fp16)[name = tensor("softmax_10_cast_fp16")]; + tensor attn_output_61_transpose_x_0 = const()[name = tensor("attn_output_61_transpose_x_0"), val = tensor(false)]; + tensor attn_output_61_transpose_y_0 = const()[name = tensor("attn_output_61_transpose_y_0"), val = tensor(false)]; + tensor attn_output_61_cast_fp16 = matmul(transpose_x = attn_output_61_transpose_x_0, transpose_y = attn_output_61_transpose_y_0, x = softmax_10_cast_fp16, y = v_65_cast_fp16)[name = tensor("attn_output_61_cast_fp16")]; + tensor var_1236 = const()[name = tensor("op_1236"), val = tensor([2, 0, 1, 3])]; + tensor concat_99x = const()[name = tensor("concat_99x"), val = tensor([-1, 768])]; + tensor var_1237_cast_fp16 = transpose(perm = var_1236, x = attn_output_61_cast_fp16)[name = tensor("transpose_7")]; + tensor attn_output_63_cast_fp16 = reshape(shape = concat_99x, x = var_1237_cast_fp16)[name = tensor("attn_output_63_cast_fp16")]; + tensor transformer_resblocks_10_attn_out_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_10_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150112384)))]; + tensor transformer_resblocks_10_attn_out_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_10_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151292096)))]; + tensor linear_41_cast_fp16 = linear(bias = transformer_resblocks_10_attn_out_proj_bias_to_fp16, weight = transformer_resblocks_10_attn_out_proj_weight_to_fp16, x = attn_output_63_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor concat_100x = const()[name = tensor("concat_100x"), val = tensor([50, -1, 768])]; + tensor var_1246_cast_fp16 = reshape(shape = concat_100x, x = linear_41_cast_fp16)[name = tensor("op_1246_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = x_75_cast_fp16, y = var_1246_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor ret_45_axes_0 = const()[name = tensor("ret_45_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_10_ln_2_weight_to_fp16 = const()[name = tensor("transformer_resblocks_10_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151293696)))]; + tensor transformer_resblocks_10_ln_2_bias_to_fp16 = const()[name = tensor("transformer_resblocks_10_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151295296)))]; + tensor ret_45_cast_fp16 = layer_norm(axes = ret_45_axes_0, beta = transformer_resblocks_10_ln_2_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_10_ln_2_weight_to_fp16, x = x_77_cast_fp16)[name = tensor("ret_45_cast_fp16")]; + tensor transformer_resblocks_10_mlp_c_fc_weight_to_fp16 = const()[name = tensor("transformer_resblocks_10_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151296896)))]; + tensor transformer_resblocks_10_mlp_c_fc_bias_to_fp16 = const()[name = tensor("transformer_resblocks_10_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156015552)))]; + tensor linear_42_cast_fp16 = linear(bias = transformer_resblocks_10_mlp_c_fc_bias_to_fp16, weight = transformer_resblocks_10_mlp_c_fc_weight_to_fp16, x = ret_45_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_1262_to_fp16 = const()[name = tensor("op_1262_to_fp16"), val = tensor(0x1.b3cp+0)]; + tensor var_1263_cast_fp16 = mul(x = linear_42_cast_fp16, y = var_1262_to_fp16)[name = tensor("op_1263_cast_fp16")]; + tensor var_1264_cast_fp16 = sigmoid(x = var_1263_cast_fp16)[name = tensor("op_1264_cast_fp16")]; + tensor input_89_cast_fp16 = mul(x = linear_42_cast_fp16, y = var_1264_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor transformer_resblocks_10_mlp_c_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_10_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(156021760)))]; + tensor transformer_resblocks_10_mlp_c_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_10_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160740416)))]; + tensor linear_43_cast_fp16 = linear(bias = transformer_resblocks_10_mlp_c_proj_bias_to_fp16, weight = transformer_resblocks_10_mlp_c_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor x_81_cast_fp16 = add(x = x_77_cast_fp16, y = linear_43_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor ret_47_axes_0 = const()[name = tensor("ret_47_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_11_ln_1_weight_to_fp16 = const()[name = tensor("transformer_resblocks_11_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160742016)))]; + tensor transformer_resblocks_11_ln_1_bias_to_fp16 = const()[name = tensor("transformer_resblocks_11_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160743616)))]; + tensor ret_47_cast_fp16 = layer_norm(axes = ret_47_axes_0, beta = transformer_resblocks_11_ln_1_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_11_ln_1_weight_to_fp16, x = x_81_cast_fp16)[name = tensor("ret_47_cast_fp16")]; + tensor transformer_resblocks_11_attn_in_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_11_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160745216)))]; + tensor transformer_resblocks_11_attn_in_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_11_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164284224)))]; + tensor linear_44_cast_fp16 = linear(bias = transformer_resblocks_11_attn_in_proj_bias_to_fp16, weight = transformer_resblocks_11_attn_in_proj_weight_to_fp16, x = ret_47_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor concat_101x = const()[name = tensor("concat_101x"), val = tensor([50, -1, 3, 768])]; + tensor var_1305_cast_fp16 = reshape(shape = concat_101x, x = linear_44_cast_fp16)[name = tensor("op_1305_cast_fp16")]; + tensor var_1306_axes_0 = const()[name = tensor("op_1306_axes_0"), val = tensor([0])]; + tensor var_1306_cast_fp16 = expand_dims(axes = var_1306_axes_0, x = var_1305_cast_fp16)[name = tensor("op_1306_cast_fp16")]; + tensor var_1307_perm_0 = const()[name = tensor("op_1307_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1308_axes_0 = const()[name = tensor("op_1308_axes_0"), val = tensor([-2])]; + tensor var_1307_cast_fp16 = transpose(perm = var_1307_perm_0, x = var_1306_cast_fp16)[name = tensor("transpose_6")]; + tensor var_1308_cast_fp16 = squeeze(axes = var_1308_axes_0, x = var_1307_cast_fp16)[name = tensor("op_1308_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 50, 0, 768])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = var_1308_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_67_begin_0 = const()[name = tensor("k_67_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_67_end_0 = const()[name = tensor("k_67_end_0"), val = tensor([2, 50, 0, 768])]; + tensor k_67_end_mask_0 = const()[name = tensor("k_67_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_67_squeeze_mask_0 = const()[name = tensor("k_67_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_67_cast_fp16 = slice_by_index(begin = k_67_begin_0, end = k_67_end_0, end_mask = k_67_end_mask_0, squeeze_mask = k_67_squeeze_mask_0, x = var_1308_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor v_67_begin_0 = const()[name = tensor("v_67_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_67_end_0 = const()[name = tensor("v_67_end_0"), val = tensor([3, 50, 0, 768])]; + tensor v_67_end_mask_0 = const()[name = tensor("v_67_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_67_squeeze_mask_0 = const()[name = tensor("v_67_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_67_cast_fp16 = slice_by_index(begin = v_67_begin_0, end = v_67_end_0, end_mask = v_67_end_mask_0, squeeze_mask = v_67_squeeze_mask_0, x = var_1308_cast_fp16)[name = tensor("v_67_cast_fp16")]; + tensor concat_102x = const()[name = tensor("concat_102x"), val = tensor([50, -1, 64])]; + tensor var_1317_cast_fp16 = reshape(shape = concat_102x, x = q_67_cast_fp16)[name = tensor("op_1317_cast_fp16")]; + tensor q_69_perm_0 = const()[name = tensor("q_69_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_103x = const()[name = tensor("concat_103x"), val = tensor([50, -1, 64])]; + tensor var_1324_cast_fp16 = reshape(shape = concat_103x, x = k_67_cast_fp16)[name = tensor("op_1324_cast_fp16")]; + tensor k_69_perm_0 = const()[name = tensor("k_69_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_104x = const()[name = tensor("concat_104x"), val = tensor([50, -1, 64])]; + tensor var_1331_cast_fp16 = reshape(shape = concat_104x, x = v_67_cast_fp16)[name = tensor("op_1331_cast_fp16")]; + tensor v_69_perm_0 = const()[name = tensor("v_69_perm_0"), val = tensor([1, 0, 2])]; + tensor concat_105x = const()[name = tensor("concat_105x"), val = tensor([-1, 12, 50, 64])]; + tensor q_69_cast_fp16 = transpose(perm = q_69_perm_0, x = var_1317_cast_fp16)[name = tensor("transpose_5")]; + tensor q_cast_fp16 = reshape(shape = concat_105x, x = q_69_cast_fp16)[name = tensor("q_cast_fp16")]; + tensor concat_106x = const()[name = tensor("concat_106x"), val = tensor([-1, 12, 50, 64])]; + tensor k_69_cast_fp16 = transpose(perm = k_69_perm_0, x = var_1324_cast_fp16)[name = tensor("transpose_4")]; + tensor k_cast_fp16 = reshape(shape = concat_106x, x = k_69_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor concat_107x = const()[name = tensor("concat_107x"), val = tensor([-1, 12, 50, 64])]; + tensor v_69_cast_fp16 = transpose(perm = v_69_perm_0, x = var_1331_cast_fp16)[name = tensor("transpose_3")]; + tensor v_cast_fp16 = reshape(shape = concat_107x, x = v_69_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor mul_23_y_0_to_fp16 = const()[name = tensor("mul_23_y_0_to_fp16"), val = tensor(0x1p-3)]; + tensor mul_23_cast_fp16 = mul(x = q_cast_fp16, y = mul_23_y_0_to_fp16)[name = tensor("mul_23_cast_fp16")]; + tensor matmul_11_transpose_y_0 = const()[name = tensor("matmul_11_transpose_y_0"), val = tensor(true)]; + tensor matmul_11_transpose_x_0 = const()[name = tensor("matmul_11_transpose_x_0"), val = tensor(false)]; + tensor matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23_cast_fp16, y = k_cast_fp16)[name = tensor("matmul_11_cast_fp16")]; + tensor softmax_11_axis_0 = const()[name = tensor("softmax_11_axis_0"), val = tensor(-1)]; + tensor softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = matmul_11_cast_fp16)[name = tensor("softmax_11_cast_fp16")]; + tensor attn_output_67_transpose_x_0 = const()[name = tensor("attn_output_67_transpose_x_0"), val = tensor(false)]; + tensor attn_output_67_transpose_y_0 = const()[name = tensor("attn_output_67_transpose_y_0"), val = tensor(false)]; + tensor attn_output_67_cast_fp16 = matmul(transpose_x = attn_output_67_transpose_x_0, transpose_y = attn_output_67_transpose_y_0, x = softmax_11_cast_fp16, y = v_cast_fp16)[name = tensor("attn_output_67_cast_fp16")]; + tensor var_1341 = const()[name = tensor("op_1341"), val = tensor([2, 0, 1, 3])]; + tensor concat_108x = const()[name = tensor("concat_108x"), val = tensor([-1, 768])]; + tensor var_1342_cast_fp16 = transpose(perm = var_1341, x = attn_output_67_cast_fp16)[name = tensor("transpose_2")]; + tensor attn_output_69_cast_fp16 = reshape(shape = concat_108x, x = var_1342_cast_fp16)[name = tensor("attn_output_69_cast_fp16")]; + tensor transformer_resblocks_11_attn_out_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_11_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164288896)))]; + tensor transformer_resblocks_11_attn_out_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_11_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165468608)))]; + tensor linear_45_cast_fp16 = linear(bias = transformer_resblocks_11_attn_out_proj_bias_to_fp16, weight = transformer_resblocks_11_attn_out_proj_weight_to_fp16, x = attn_output_69_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor concat_109x = const()[name = tensor("concat_109x"), val = tensor([50, -1, 768])]; + tensor var_1351_cast_fp16 = reshape(shape = concat_109x, x = linear_45_cast_fp16)[name = tensor("op_1351_cast_fp16")]; + tensor x_83_cast_fp16 = add(x = x_81_cast_fp16, y = var_1351_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor ret_49_axes_0 = const()[name = tensor("ret_49_axes_0"), val = tensor([-1])]; + tensor transformer_resblocks_11_ln_2_weight_to_fp16 = const()[name = tensor("transformer_resblocks_11_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165470208)))]; + tensor transformer_resblocks_11_ln_2_bias_to_fp16 = const()[name = tensor("transformer_resblocks_11_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165471808)))]; + tensor ret_49_cast_fp16 = layer_norm(axes = ret_49_axes_0, beta = transformer_resblocks_11_ln_2_bias_to_fp16, epsilon = var_100_to_fp16, gamma = transformer_resblocks_11_ln_2_weight_to_fp16, x = x_83_cast_fp16)[name = tensor("ret_49_cast_fp16")]; + tensor transformer_resblocks_11_mlp_c_fc_weight_to_fp16 = const()[name = tensor("transformer_resblocks_11_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165473408)))]; + tensor transformer_resblocks_11_mlp_c_fc_bias_to_fp16 = const()[name = tensor("transformer_resblocks_11_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170192064)))]; + tensor linear_46_cast_fp16 = linear(bias = transformer_resblocks_11_mlp_c_fc_bias_to_fp16, weight = transformer_resblocks_11_mlp_c_fc_weight_to_fp16, x = ret_49_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_1367_to_fp16 = const()[name = tensor("op_1367_to_fp16"), val = tensor(0x1.b3cp+0)]; + tensor var_1368_cast_fp16 = mul(x = linear_46_cast_fp16, y = var_1367_to_fp16)[name = tensor("op_1368_cast_fp16")]; + tensor var_1369_cast_fp16 = sigmoid(x = var_1368_cast_fp16)[name = tensor("op_1369_cast_fp16")]; + tensor input_97_cast_fp16 = mul(x = linear_46_cast_fp16, y = var_1369_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor transformer_resblocks_11_mlp_c_proj_weight_to_fp16 = const()[name = tensor("transformer_resblocks_11_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170198272)))]; + tensor transformer_resblocks_11_mlp_c_proj_bias_to_fp16 = const()[name = tensor("transformer_resblocks_11_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174916928)))]; + tensor linear_47_cast_fp16 = linear(bias = transformer_resblocks_11_mlp_c_proj_bias_to_fp16, weight = transformer_resblocks_11_mlp_c_proj_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor x_87_cast_fp16 = add(x = x_83_cast_fp16, y = linear_47_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor var_1378 = const()[name = tensor("op_1378"), val = tensor([1, 0, 2])]; + tensor var_1387_begin_0 = const()[name = tensor("op_1387_begin_0"), val = tensor([0, 0, 0])]; + tensor var_1387_end_0 = const()[name = tensor("op_1387_end_0"), val = tensor([0, 1, 768])]; + tensor var_1387_end_mask_0 = const()[name = tensor("op_1387_end_mask_0"), val = tensor([true, false, true])]; + tensor var_1387_squeeze_mask_0 = const()[name = tensor("op_1387_squeeze_mask_0"), val = tensor([false, true, false])]; + tensor x_89_cast_fp16 = transpose(perm = var_1378, x = x_87_cast_fp16)[name = tensor("transpose_1")]; + tensor var_1387_cast_fp16 = slice_by_index(begin = var_1387_begin_0, end = var_1387_end_0, end_mask = var_1387_end_mask_0, squeeze_mask = var_1387_squeeze_mask_0, x = x_89_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor ret_axes_0 = const()[name = tensor("ret_axes_0"), val = tensor([-1])]; + tensor ln_post_weight_to_fp16 = const()[name = tensor("ln_post_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174918528)))]; + tensor ln_post_bias_to_fp16 = const()[name = tensor("ln_post_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174920128)))]; + tensor var_1394_to_fp16 = const()[name = tensor("op_1394_to_fp16"), val = tensor(0x1.5p-17)]; + tensor ret_cast_fp16 = layer_norm(axes = ret_axes_0, beta = ln_post_bias_to_fp16, epsilon = var_1394_to_fp16, gamma = ln_post_weight_to_fp16, x = var_1387_cast_fp16)[name = tensor("ret_cast_fp16")]; + tensor transpose_0_to_fp16 = const()[name = tensor("transpose_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174921728)))]; + tensor var_1405_bias_0_to_fp16 = const()[name = tensor("op_1405_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(175708224)))]; + tensor var_1405_cast_fp16 = linear(bias = var_1405_bias_0_to_fp16, weight = transpose_0_to_fp16, x = ret_cast_fp16)[name = tensor("op_1405_cast_fp16")]; + tensor var_1405_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1405_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor embOutput = cast(dtype = var_1405_cast_fp16_to_fp32_dtype_0, x = var_1405_cast_fp16)[name = tensor("cast_185")]; + } -> (embOutput); +} \ No newline at end of file