diff --git "a/CLIP_TextEncoder.mlmodelc/model.mil" "b/CLIP_TextEncoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/CLIP_TextEncoder.mlmodelc/model.mil" @@ -0,0 +1,1062 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor text) { + tensor var_17_axis_0 = const()[name = tensor("op_17_axis_0"), val = tensor(0)]; + tensor model_token_embedding_weight_to_fp16 = const()[name = tensor("model_token_embedding_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor var_17_cast_fp16 = gather(axis = var_17_axis_0, indices = text, x = model_token_embedding_weight_to_fp16)[name = tensor("op_17_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(50593920)))]; + tensor input_1_cast_fp16 = add(x = var_17_cast_fp16, y = const_0_to_fp16)[name = tensor("input_1_cast_fp16")]; + tensor x_3_axes_0 = const()[name = tensor("x_3_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_0_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_0_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50672832)))]; + tensor model_transformer_resblocks_0_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_0_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50673920)))]; + tensor var_34_to_fp16 = const()[name = tensor("op_34_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = model_transformer_resblocks_0_ln_1_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_0_ln_1_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor model_transformer_resblocks_0_attn_in_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_0_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50675008)))]; + tensor model_transformer_resblocks_0_attn_in_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_0_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52247936)))]; + tensor query_3_cast_fp16 = transpose(perm = query_3_perm_0, x = x_3_cast_fp16)[name = tensor("transpose_85")]; + tensor linear_0_cast_fp16 = linear(bias = model_transformer_resblocks_0_attn_in_proj_bias_to_fp16, weight = model_transformer_resblocks_0_attn_in_proj_weight_to_fp16, x = query_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor concat_0 = const()[name = tensor("concat_0"), val = tensor([77, 1, 3, 512])]; + tensor var_104_cast_fp16 = reshape(shape = concat_0, x = linear_0_cast_fp16)[name = tensor("op_104_cast_fp16")]; + tensor var_105_axes_0 = const()[name = tensor("op_105_axes_0"), val = tensor([0])]; + tensor var_105_cast_fp16 = expand_dims(axes = var_105_axes_0, x = var_104_cast_fp16)[name = tensor("op_105_cast_fp16")]; + tensor var_106_perm_0 = const()[name = tensor("op_106_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_107_axes_0 = const()[name = tensor("op_107_axes_0"), val = tensor([-2])]; + tensor var_106_cast_fp16 = transpose(perm = var_106_perm_0, x = var_105_cast_fp16)[name = tensor("transpose_84")]; + tensor var_107_cast_fp16 = squeeze(axes = var_107_axes_0, x = var_106_cast_fp16)[name = tensor("op_107_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, 77, 1, 512])]; + 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_107_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, 77, 1, 512])]; + 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_107_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, 77, 1, 512])]; + 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_107_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_116 = const()[name = tensor("op_116"), val = tensor([77, 8, 64])]; + tensor var_117_cast_fp16 = reshape(shape = var_116, x = q_1_cast_fp16)[name = tensor("op_117_cast_fp16")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([1, 0, 2])]; + tensor var_123 = const()[name = tensor("op_123"), val = tensor([77, 8, 64])]; + tensor var_124_cast_fp16 = reshape(shape = var_123, x = k_1_cast_fp16)[name = tensor("op_124_cast_fp16")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([1, 0, 2])]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([77, 8, 64])]; + tensor var_131_cast_fp16 = reshape(shape = var_130, x = v_1_cast_fp16)[name = tensor("op_131_cast_fp16")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([1, 0, 2])]; + tensor var_135 = const()[name = tensor("op_135"), val = tensor([1, 8, 77, 64])]; + tensor q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = var_117_cast_fp16)[name = tensor("transpose_83")]; + tensor q_5_cast_fp16 = reshape(shape = var_135, x = q_3_cast_fp16)[name = tensor("q_5_cast_fp16")]; + tensor var_137 = const()[name = tensor("op_137"), val = tensor([1, 8, 77, 64])]; + tensor k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = var_124_cast_fp16)[name = tensor("transpose_82")]; + tensor k_5_cast_fp16 = reshape(shape = var_137, x = k_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor var_139 = const()[name = tensor("op_139"), val = tensor([1, 8, 77, 64])]; + tensor v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_131_cast_fp16)[name = tensor("transpose_81")]; + tensor v_5_cast_fp16 = reshape(shape = var_139, 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 attn_mask_7_to_fp16 = const()[name = tensor("attn_mask_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52251072)))]; + tensor add_0_cast_fp16 = add(x = matmul_0_cast_fp16, y = attn_mask_7_to_fp16)[name = tensor("add_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 = add_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_142 = const()[name = tensor("op_142"), val = tensor([2, 0, 1, 3])]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor([77, 512])]; + tensor var_143_cast_fp16 = transpose(perm = var_142, x = attn_output_1_cast_fp16)[name = tensor("transpose_80")]; + tensor attn_output_3_cast_fp16 = reshape(shape = var_147, x = var_143_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor model_transformer_resblocks_0_attn_out_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_0_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52263040)))]; + tensor model_transformer_resblocks_0_attn_out_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_0_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52787392)))]; + tensor linear_1_cast_fp16 = linear(bias = model_transformer_resblocks_0_attn_out_proj_bias_to_fp16, weight = model_transformer_resblocks_0_attn_out_proj_weight_to_fp16, x = attn_output_3_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_151 = const()[name = tensor("op_151"), val = tensor([77, 1, 512])]; + tensor attn_output_7_cast_fp16 = reshape(shape = var_151, x = linear_1_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_153_perm_0 = const()[name = tensor("op_153_perm_0"), val = tensor([1, 0, 2])]; + tensor var_153_cast_fp16 = transpose(perm = var_153_perm_0, x = attn_output_7_cast_fp16)[name = tensor("transpose_79")]; + tensor input_3_cast_fp16 = add(x = input_1_cast_fp16, y = var_153_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_0_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_0_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52788480)))]; + tensor model_transformer_resblocks_0_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_0_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52789568)))]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = model_transformer_resblocks_0_ln_2_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_0_ln_2_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor model_transformer_resblocks_0_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_0_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52790656)))]; + tensor model_transformer_resblocks_0_mlp_c_fc_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_0_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54887872)))]; + tensor linear_2_cast_fp16 = linear(bias = model_transformer_resblocks_0_mlp_c_fc_bias_to_fp16, weight = model_transformer_resblocks_0_mlp_c_fc_weight_to_fp16, x = x_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_9_mode_0 = const()[name = tensor("input_9_mode_0"), val = tensor("EXACT")]; + tensor input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor model_transformer_resblocks_0_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_0_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54892032)))]; + tensor model_transformer_resblocks_0_mlp_c_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_0_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56989248)))]; + tensor linear_3_cast_fp16 = linear(bias = model_transformer_resblocks_0_mlp_c_proj_bias_to_fp16, weight = model_transformer_resblocks_0_mlp_c_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_11_cast_fp16 = add(x = input_3_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_1_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_1_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56990336)))]; + tensor model_transformer_resblocks_1_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_1_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56991424)))]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = model_transformer_resblocks_1_ln_1_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_1_ln_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor query_7_perm_0 = const()[name = tensor("query_7_perm_0"), val = tensor([1, 0, 2])]; + tensor model_transformer_resblocks_1_attn_in_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_1_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56992512)))]; + tensor model_transformer_resblocks_1_attn_in_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_1_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58565440)))]; + tensor query_7_cast_fp16 = transpose(perm = query_7_perm_0, x = x_7_cast_fp16)[name = tensor("transpose_78")]; + tensor linear_4_cast_fp16 = linear(bias = model_transformer_resblocks_1_attn_in_proj_bias_to_fp16, weight = model_transformer_resblocks_1_attn_in_proj_weight_to_fp16, x = query_7_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([77, 1, 3, 512])]; + tensor var_206_cast_fp16 = reshape(shape = concat_1, x = linear_4_cast_fp16)[name = tensor("op_206_cast_fp16")]; + tensor var_207_axes_0 = const()[name = tensor("op_207_axes_0"), val = tensor([0])]; + tensor var_207_cast_fp16 = expand_dims(axes = var_207_axes_0, x = var_206_cast_fp16)[name = tensor("op_207_cast_fp16")]; + tensor var_208_perm_0 = const()[name = tensor("op_208_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_209_axes_0 = const()[name = tensor("op_209_axes_0"), val = tensor([-2])]; + tensor var_208_cast_fp16 = transpose(perm = var_208_perm_0, x = var_207_cast_fp16)[name = tensor("transpose_77")]; + tensor var_209_cast_fp16 = squeeze(axes = var_209_axes_0, x = var_208_cast_fp16)[name = tensor("op_209_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, 77, 1, 512])]; + 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_209_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, 77, 1, 512])]; + 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_209_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, 77, 1, 512])]; + 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_209_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_218 = const()[name = tensor("op_218"), val = tensor([77, 8, 64])]; + tensor var_219_cast_fp16 = reshape(shape = var_218, x = q_7_cast_fp16)[name = tensor("op_219_cast_fp16")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([1, 0, 2])]; + tensor var_225 = const()[name = tensor("op_225"), val = tensor([77, 8, 64])]; + tensor var_226_cast_fp16 = reshape(shape = var_225, x = k_7_cast_fp16)[name = tensor("op_226_cast_fp16")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([1, 0, 2])]; + tensor var_232 = const()[name = tensor("op_232"), val = tensor([77, 8, 64])]; + tensor var_233_cast_fp16 = reshape(shape = var_232, x = v_7_cast_fp16)[name = tensor("op_233_cast_fp16")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([1, 0, 2])]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 8, 77, 64])]; + tensor q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_219_cast_fp16)[name = tensor("transpose_76")]; + tensor q_11_cast_fp16 = reshape(shape = var_237, x = q_9_cast_fp16)[name = tensor("q_11_cast_fp16")]; + tensor var_239 = const()[name = tensor("op_239"), val = tensor([1, 8, 77, 64])]; + tensor k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_226_cast_fp16)[name = tensor("transpose_75")]; + tensor k_11_cast_fp16 = reshape(shape = var_239, x = k_9_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor var_241 = const()[name = tensor("op_241"), val = tensor([1, 8, 77, 64])]; + tensor v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_233_cast_fp16)[name = tensor("transpose_74")]; + tensor v_11_cast_fp16 = reshape(shape = var_241, 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 add_1_cast_fp16 = add(x = matmul_1_cast_fp16, y = attn_mask_7_to_fp16)[name = tensor("add_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 = add_1_cast_fp16)[name = tensor("softmax_1_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1_cast_fp16, y = v_11_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([2, 0, 1, 3])]; + tensor var_249 = const()[name = tensor("op_249"), val = tensor([77, 512])]; + tensor var_245_cast_fp16 = transpose(perm = var_244, x = attn_output_9_cast_fp16)[name = tensor("transpose_73")]; + tensor attn_output_11_cast_fp16 = reshape(shape = var_249, x = var_245_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor model_transformer_resblocks_1_attn_out_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_1_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58568576)))]; + tensor model_transformer_resblocks_1_attn_out_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_1_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59092928)))]; + tensor linear_5_cast_fp16 = linear(bias = model_transformer_resblocks_1_attn_out_proj_bias_to_fp16, weight = model_transformer_resblocks_1_attn_out_proj_weight_to_fp16, x = attn_output_11_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([77, 1, 512])]; + tensor attn_output_15_cast_fp16 = reshape(shape = var_253, x = linear_5_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_255_perm_0 = const()[name = tensor("op_255_perm_0"), val = tensor([1, 0, 2])]; + tensor var_255_cast_fp16 = transpose(perm = var_255_perm_0, x = attn_output_15_cast_fp16)[name = tensor("transpose_72")]; + tensor input_13_cast_fp16 = add(x = input_11_cast_fp16, y = var_255_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor x_9_axes_0 = const()[name = tensor("x_9_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_1_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_1_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59094016)))]; + tensor model_transformer_resblocks_1_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_1_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59095104)))]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = model_transformer_resblocks_1_ln_2_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_1_ln_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor model_transformer_resblocks_1_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_1_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59096192)))]; + tensor model_transformer_resblocks_1_mlp_c_fc_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_1_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61193408)))]; + tensor linear_6_cast_fp16 = linear(bias = model_transformer_resblocks_1_mlp_c_fc_bias_to_fp16, weight = model_transformer_resblocks_1_mlp_c_fc_weight_to_fp16, x = x_9_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_19_mode_0 = const()[name = tensor("input_19_mode_0"), val = tensor("EXACT")]; + tensor input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_6_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor model_transformer_resblocks_1_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_1_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61197568)))]; + tensor model_transformer_resblocks_1_mlp_c_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_1_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63294784)))]; + tensor linear_7_cast_fp16 = linear(bias = model_transformer_resblocks_1_mlp_c_proj_bias_to_fp16, weight = model_transformer_resblocks_1_mlp_c_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = add(x = input_13_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_2_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_2_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63295872)))]; + tensor model_transformer_resblocks_2_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_2_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63296960)))]; + tensor x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = model_transformer_resblocks_2_ln_1_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_2_ln_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor query_11_perm_0 = const()[name = tensor("query_11_perm_0"), val = tensor([1, 0, 2])]; + tensor model_transformer_resblocks_2_attn_in_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_2_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63298048)))]; + tensor model_transformer_resblocks_2_attn_in_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_2_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64870976)))]; + tensor query_11_cast_fp16 = transpose(perm = query_11_perm_0, x = x_11_cast_fp16)[name = tensor("transpose_71")]; + tensor linear_8_cast_fp16 = linear(bias = model_transformer_resblocks_2_attn_in_proj_bias_to_fp16, weight = model_transformer_resblocks_2_attn_in_proj_weight_to_fp16, x = query_11_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([77, 1, 3, 512])]; + tensor var_308_cast_fp16 = reshape(shape = concat_2, x = linear_8_cast_fp16)[name = tensor("op_308_cast_fp16")]; + tensor var_309_axes_0 = const()[name = tensor("op_309_axes_0"), val = tensor([0])]; + tensor var_309_cast_fp16 = expand_dims(axes = var_309_axes_0, x = var_308_cast_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_310_perm_0 = const()[name = tensor("op_310_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_311_axes_0 = const()[name = tensor("op_311_axes_0"), val = tensor([-2])]; + tensor var_310_cast_fp16 = transpose(perm = var_310_perm_0, x = var_309_cast_fp16)[name = tensor("transpose_70")]; + tensor var_311_cast_fp16 = squeeze(axes = var_311_axes_0, x = var_310_cast_fp16)[name = tensor("op_311_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, 77, 1, 512])]; + 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_311_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, 77, 1, 512])]; + 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_311_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, 77, 1, 512])]; + 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_311_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([77, 8, 64])]; + tensor var_321_cast_fp16 = reshape(shape = var_320, x = q_13_cast_fp16)[name = tensor("op_321_cast_fp16")]; + tensor q_15_perm_0 = const()[name = tensor("q_15_perm_0"), val = tensor([1, 0, 2])]; + tensor var_327 = const()[name = tensor("op_327"), val = tensor([77, 8, 64])]; + tensor var_328_cast_fp16 = reshape(shape = var_327, x = k_13_cast_fp16)[name = tensor("op_328_cast_fp16")]; + tensor k_15_perm_0 = const()[name = tensor("k_15_perm_0"), val = tensor([1, 0, 2])]; + tensor var_334 = const()[name = tensor("op_334"), val = tensor([77, 8, 64])]; + tensor var_335_cast_fp16 = reshape(shape = var_334, x = v_13_cast_fp16)[name = tensor("op_335_cast_fp16")]; + tensor v_15_perm_0 = const()[name = tensor("v_15_perm_0"), val = tensor([1, 0, 2])]; + tensor var_339 = const()[name = tensor("op_339"), val = tensor([1, 8, 77, 64])]; + tensor q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = var_321_cast_fp16)[name = tensor("transpose_69")]; + tensor q_17_cast_fp16 = reshape(shape = var_339, x = q_15_cast_fp16)[name = tensor("q_17_cast_fp16")]; + tensor var_341 = const()[name = tensor("op_341"), val = tensor([1, 8, 77, 64])]; + tensor k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = var_328_cast_fp16)[name = tensor("transpose_68")]; + tensor k_17_cast_fp16 = reshape(shape = var_341, x = k_15_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor var_343 = const()[name = tensor("op_343"), val = tensor([1, 8, 77, 64])]; + tensor v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_335_cast_fp16)[name = tensor("transpose_67")]; + tensor v_17_cast_fp16 = reshape(shape = var_343, 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 add_2_cast_fp16 = add(x = matmul_2_cast_fp16, y = attn_mask_7_to_fp16)[name = tensor("add_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 = add_2_cast_fp16)[name = tensor("softmax_2_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = softmax_2_cast_fp16, y = v_17_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_346 = const()[name = tensor("op_346"), val = tensor([2, 0, 1, 3])]; + tensor var_351 = const()[name = tensor("op_351"), val = tensor([77, 512])]; + tensor var_347_cast_fp16 = transpose(perm = var_346, x = attn_output_17_cast_fp16)[name = tensor("transpose_66")]; + tensor attn_output_19_cast_fp16 = reshape(shape = var_351, x = var_347_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor model_transformer_resblocks_2_attn_out_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_2_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64874112)))]; + tensor model_transformer_resblocks_2_attn_out_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_2_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65398464)))]; + tensor linear_9_cast_fp16 = linear(bias = model_transformer_resblocks_2_attn_out_proj_bias_to_fp16, weight = model_transformer_resblocks_2_attn_out_proj_weight_to_fp16, x = attn_output_19_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_355 = const()[name = tensor("op_355"), val = tensor([77, 1, 512])]; + tensor attn_output_23_cast_fp16 = reshape(shape = var_355, x = linear_9_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_357_perm_0 = const()[name = tensor("op_357_perm_0"), val = tensor([1, 0, 2])]; + tensor var_357_cast_fp16 = transpose(perm = var_357_perm_0, x = attn_output_23_cast_fp16)[name = tensor("transpose_65")]; + tensor input_23_cast_fp16 = add(x = input_21_cast_fp16, y = var_357_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_2_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_2_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65399552)))]; + tensor model_transformer_resblocks_2_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_2_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65400640)))]; + tensor x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = model_transformer_resblocks_2_ln_2_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_2_ln_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor model_transformer_resblocks_2_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_2_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65401728)))]; + tensor model_transformer_resblocks_2_mlp_c_fc_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_2_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67498944)))]; + tensor linear_10_cast_fp16 = linear(bias = model_transformer_resblocks_2_mlp_c_fc_bias_to_fp16, weight = model_transformer_resblocks_2_mlp_c_fc_weight_to_fp16, x = x_13_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_29_mode_0 = const()[name = tensor("input_29_mode_0"), val = tensor("EXACT")]; + tensor input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = linear_10_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor model_transformer_resblocks_2_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_2_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67503104)))]; + tensor model_transformer_resblocks_2_mlp_c_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_2_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69600320)))]; + tensor linear_11_cast_fp16 = linear(bias = model_transformer_resblocks_2_mlp_c_proj_bias_to_fp16, weight = model_transformer_resblocks_2_mlp_c_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_31_cast_fp16 = add(x = input_23_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor x_15_axes_0 = const()[name = tensor("x_15_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_3_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_3_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69601408)))]; + tensor model_transformer_resblocks_3_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_3_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69602496)))]; + tensor x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = model_transformer_resblocks_3_ln_1_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_3_ln_1_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor query_15_perm_0 = const()[name = tensor("query_15_perm_0"), val = tensor([1, 0, 2])]; + tensor model_transformer_resblocks_3_attn_in_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_3_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69603584)))]; + tensor model_transformer_resblocks_3_attn_in_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_3_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71176512)))]; + tensor query_15_cast_fp16 = transpose(perm = query_15_perm_0, x = x_15_cast_fp16)[name = tensor("transpose_64")]; + tensor linear_12_cast_fp16 = linear(bias = model_transformer_resblocks_3_attn_in_proj_bias_to_fp16, weight = model_transformer_resblocks_3_attn_in_proj_weight_to_fp16, x = query_15_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([77, 1, 3, 512])]; + tensor var_410_cast_fp16 = reshape(shape = concat_3, x = linear_12_cast_fp16)[name = tensor("op_410_cast_fp16")]; + tensor var_411_axes_0 = const()[name = tensor("op_411_axes_0"), val = tensor([0])]; + tensor var_411_cast_fp16 = expand_dims(axes = var_411_axes_0, x = var_410_cast_fp16)[name = tensor("op_411_cast_fp16")]; + tensor var_412_perm_0 = const()[name = tensor("op_412_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_413_axes_0 = const()[name = tensor("op_413_axes_0"), val = tensor([-2])]; + tensor var_412_cast_fp16 = transpose(perm = var_412_perm_0, x = var_411_cast_fp16)[name = tensor("transpose_63")]; + tensor var_413_cast_fp16 = squeeze(axes = var_413_axes_0, x = var_412_cast_fp16)[name = tensor("op_413_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, 77, 1, 512])]; + 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_413_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, 77, 1, 512])]; + 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_413_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, 77, 1, 512])]; + 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_413_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_422 = const()[name = tensor("op_422"), val = tensor([77, 8, 64])]; + tensor var_423_cast_fp16 = reshape(shape = var_422, x = q_19_cast_fp16)[name = tensor("op_423_cast_fp16")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([77, 8, 64])]; + tensor var_430_cast_fp16 = reshape(shape = var_429, x = k_19_cast_fp16)[name = tensor("op_430_cast_fp16")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_436 = const()[name = tensor("op_436"), val = tensor([77, 8, 64])]; + tensor var_437_cast_fp16 = reshape(shape = var_436, x = v_19_cast_fp16)[name = tensor("op_437_cast_fp16")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_441 = const()[name = tensor("op_441"), val = tensor([1, 8, 77, 64])]; + tensor q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_423_cast_fp16)[name = tensor("transpose_62")]; + tensor q_23_cast_fp16 = reshape(shape = var_441, x = q_21_cast_fp16)[name = tensor("q_23_cast_fp16")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 8, 77, 64])]; + tensor k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_430_cast_fp16)[name = tensor("transpose_61")]; + tensor k_23_cast_fp16 = reshape(shape = var_443, x = k_21_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 8, 77, 64])]; + tensor v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_437_cast_fp16)[name = tensor("transpose_60")]; + tensor v_23_cast_fp16 = reshape(shape = var_445, 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 add_3_cast_fp16 = add(x = matmul_3_cast_fp16, y = attn_mask_7_to_fp16)[name = tensor("add_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 = add_3_cast_fp16)[name = tensor("softmax_3_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_3_cast_fp16, y = v_23_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_448 = const()[name = tensor("op_448"), val = tensor([2, 0, 1, 3])]; + tensor var_453 = const()[name = tensor("op_453"), val = tensor([77, 512])]; + tensor var_449_cast_fp16 = transpose(perm = var_448, x = attn_output_25_cast_fp16)[name = tensor("transpose_59")]; + tensor attn_output_27_cast_fp16 = reshape(shape = var_453, x = var_449_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor model_transformer_resblocks_3_attn_out_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_3_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71179648)))]; + tensor model_transformer_resblocks_3_attn_out_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_3_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71704000)))]; + tensor linear_13_cast_fp16 = linear(bias = model_transformer_resblocks_3_attn_out_proj_bias_to_fp16, weight = model_transformer_resblocks_3_attn_out_proj_weight_to_fp16, x = attn_output_27_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_457 = const()[name = tensor("op_457"), val = tensor([77, 1, 512])]; + tensor attn_output_31_cast_fp16 = reshape(shape = var_457, x = linear_13_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_459_perm_0 = const()[name = tensor("op_459_perm_0"), val = tensor([1, 0, 2])]; + tensor var_459_cast_fp16 = transpose(perm = var_459_perm_0, x = attn_output_31_cast_fp16)[name = tensor("transpose_58")]; + tensor input_33_cast_fp16 = add(x = input_31_cast_fp16, y = var_459_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_3_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_3_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71705088)))]; + tensor model_transformer_resblocks_3_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_3_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71706176)))]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = model_transformer_resblocks_3_ln_2_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_3_ln_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor model_transformer_resblocks_3_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_3_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71707264)))]; + tensor model_transformer_resblocks_3_mlp_c_fc_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_3_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73804480)))]; + tensor linear_14_cast_fp16 = linear(bias = model_transformer_resblocks_3_mlp_c_fc_bias_to_fp16, weight = model_transformer_resblocks_3_mlp_c_fc_weight_to_fp16, x = x_17_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_39_mode_0 = const()[name = tensor("input_39_mode_0"), val = tensor("EXACT")]; + tensor input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = linear_14_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor model_transformer_resblocks_3_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_3_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73808640)))]; + tensor model_transformer_resblocks_3_mlp_c_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_3_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75905856)))]; + tensor linear_15_cast_fp16 = linear(bias = model_transformer_resblocks_3_mlp_c_proj_bias_to_fp16, weight = model_transformer_resblocks_3_mlp_c_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = input_33_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_4_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_4_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75906944)))]; + tensor model_transformer_resblocks_4_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_4_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75908032)))]; + tensor x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = model_transformer_resblocks_4_ln_1_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_4_ln_1_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor query_19_perm_0 = const()[name = tensor("query_19_perm_0"), val = tensor([1, 0, 2])]; + tensor model_transformer_resblocks_4_attn_in_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_4_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75909120)))]; + tensor model_transformer_resblocks_4_attn_in_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_4_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77482048)))]; + tensor query_19_cast_fp16 = transpose(perm = query_19_perm_0, x = x_19_cast_fp16)[name = tensor("transpose_57")]; + tensor linear_16_cast_fp16 = linear(bias = model_transformer_resblocks_4_attn_in_proj_bias_to_fp16, weight = model_transformer_resblocks_4_attn_in_proj_weight_to_fp16, x = query_19_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([77, 1, 3, 512])]; + tensor var_512_cast_fp16 = reshape(shape = concat_4, x = linear_16_cast_fp16)[name = tensor("op_512_cast_fp16")]; + tensor var_513_axes_0 = const()[name = tensor("op_513_axes_0"), val = tensor([0])]; + tensor var_513_cast_fp16 = expand_dims(axes = var_513_axes_0, x = var_512_cast_fp16)[name = tensor("op_513_cast_fp16")]; + tensor var_514_perm_0 = const()[name = tensor("op_514_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_515_axes_0 = const()[name = tensor("op_515_axes_0"), val = tensor([-2])]; + tensor var_514_cast_fp16 = transpose(perm = var_514_perm_0, x = var_513_cast_fp16)[name = tensor("transpose_56")]; + tensor var_515_cast_fp16 = squeeze(axes = var_515_axes_0, x = var_514_cast_fp16)[name = tensor("op_515_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, 77, 1, 512])]; + 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_515_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, 77, 1, 512])]; + 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_515_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, 77, 1, 512])]; + 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_515_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_524 = const()[name = tensor("op_524"), val = tensor([77, 8, 64])]; + tensor var_525_cast_fp16 = reshape(shape = var_524, x = q_25_cast_fp16)[name = tensor("op_525_cast_fp16")]; + tensor q_27_perm_0 = const()[name = tensor("q_27_perm_0"), val = tensor([1, 0, 2])]; + tensor var_531 = const()[name = tensor("op_531"), val = tensor([77, 8, 64])]; + tensor var_532_cast_fp16 = reshape(shape = var_531, x = k_25_cast_fp16)[name = tensor("op_532_cast_fp16")]; + tensor k_27_perm_0 = const()[name = tensor("k_27_perm_0"), val = tensor([1, 0, 2])]; + tensor var_538 = const()[name = tensor("op_538"), val = tensor([77, 8, 64])]; + tensor var_539_cast_fp16 = reshape(shape = var_538, x = v_25_cast_fp16)[name = tensor("op_539_cast_fp16")]; + tensor v_27_perm_0 = const()[name = tensor("v_27_perm_0"), val = tensor([1, 0, 2])]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 8, 77, 64])]; + tensor q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = var_525_cast_fp16)[name = tensor("transpose_55")]; + tensor q_29_cast_fp16 = reshape(shape = var_543, x = q_27_cast_fp16)[name = tensor("q_29_cast_fp16")]; + tensor var_545 = const()[name = tensor("op_545"), val = tensor([1, 8, 77, 64])]; + tensor k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = var_532_cast_fp16)[name = tensor("transpose_54")]; + tensor k_29_cast_fp16 = reshape(shape = var_545, x = k_27_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor var_547 = const()[name = tensor("op_547"), val = tensor([1, 8, 77, 64])]; + tensor v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_539_cast_fp16)[name = tensor("transpose_53")]; + tensor v_29_cast_fp16 = reshape(shape = var_547, 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 add_4_cast_fp16 = add(x = matmul_4_cast_fp16, y = attn_mask_7_to_fp16)[name = tensor("add_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 = add_4_cast_fp16)[name = tensor("softmax_4_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_4_cast_fp16, y = v_29_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_550 = const()[name = tensor("op_550"), val = tensor([2, 0, 1, 3])]; + tensor var_555 = const()[name = tensor("op_555"), val = tensor([77, 512])]; + tensor var_551_cast_fp16 = transpose(perm = var_550, x = attn_output_33_cast_fp16)[name = tensor("transpose_52")]; + tensor attn_output_35_cast_fp16 = reshape(shape = var_555, x = var_551_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor model_transformer_resblocks_4_attn_out_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_4_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77485184)))]; + tensor model_transformer_resblocks_4_attn_out_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_4_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78009536)))]; + tensor linear_17_cast_fp16 = linear(bias = model_transformer_resblocks_4_attn_out_proj_bias_to_fp16, weight = model_transformer_resblocks_4_attn_out_proj_weight_to_fp16, x = attn_output_35_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_559 = const()[name = tensor("op_559"), val = tensor([77, 1, 512])]; + tensor attn_output_39_cast_fp16 = reshape(shape = var_559, x = linear_17_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_561_perm_0 = const()[name = tensor("op_561_perm_0"), val = tensor([1, 0, 2])]; + tensor var_561_cast_fp16 = transpose(perm = var_561_perm_0, x = attn_output_39_cast_fp16)[name = tensor("transpose_51")]; + tensor input_43_cast_fp16 = add(x = input_41_cast_fp16, y = var_561_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor x_21_axes_0 = const()[name = tensor("x_21_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_4_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_4_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78010624)))]; + tensor model_transformer_resblocks_4_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_4_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78011712)))]; + tensor x_21_cast_fp16 = layer_norm(axes = x_21_axes_0, beta = model_transformer_resblocks_4_ln_2_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_4_ln_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor model_transformer_resblocks_4_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_4_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78012800)))]; + tensor model_transformer_resblocks_4_mlp_c_fc_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_4_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80110016)))]; + tensor linear_18_cast_fp16 = linear(bias = model_transformer_resblocks_4_mlp_c_fc_bias_to_fp16, weight = model_transformer_resblocks_4_mlp_c_fc_weight_to_fp16, x = x_21_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_18_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor model_transformer_resblocks_4_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_4_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80114176)))]; + tensor model_transformer_resblocks_4_mlp_c_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_4_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82211392)))]; + tensor linear_19_cast_fp16 = linear(bias = model_transformer_resblocks_4_mlp_c_proj_bias_to_fp16, weight = model_transformer_resblocks_4_mlp_c_proj_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_43_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_5_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_5_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82212480)))]; + tensor model_transformer_resblocks_5_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_5_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82213568)))]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = model_transformer_resblocks_5_ln_1_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_5_ln_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor query_23_perm_0 = const()[name = tensor("query_23_perm_0"), val = tensor([1, 0, 2])]; + tensor model_transformer_resblocks_5_attn_in_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_5_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82214656)))]; + tensor model_transformer_resblocks_5_attn_in_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_5_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83787584)))]; + tensor query_23_cast_fp16 = transpose(perm = query_23_perm_0, x = x_23_cast_fp16)[name = tensor("transpose_50")]; + tensor linear_20_cast_fp16 = linear(bias = model_transformer_resblocks_5_attn_in_proj_bias_to_fp16, weight = model_transformer_resblocks_5_attn_in_proj_weight_to_fp16, x = query_23_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([77, 1, 3, 512])]; + tensor var_614_cast_fp16 = reshape(shape = concat_5, x = linear_20_cast_fp16)[name = tensor("op_614_cast_fp16")]; + tensor var_615_axes_0 = const()[name = tensor("op_615_axes_0"), val = tensor([0])]; + tensor var_615_cast_fp16 = expand_dims(axes = var_615_axes_0, x = var_614_cast_fp16)[name = tensor("op_615_cast_fp16")]; + tensor var_616_perm_0 = const()[name = tensor("op_616_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_617_axes_0 = const()[name = tensor("op_617_axes_0"), val = tensor([-2])]; + tensor var_616_cast_fp16 = transpose(perm = var_616_perm_0, x = var_615_cast_fp16)[name = tensor("transpose_49")]; + tensor var_617_cast_fp16 = squeeze(axes = var_617_axes_0, x = var_616_cast_fp16)[name = tensor("op_617_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, 77, 1, 512])]; + 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_617_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, 77, 1, 512])]; + 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_617_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, 77, 1, 512])]; + 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_617_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_626 = const()[name = tensor("op_626"), val = tensor([77, 8, 64])]; + tensor var_627_cast_fp16 = reshape(shape = var_626, x = q_31_cast_fp16)[name = tensor("op_627_cast_fp16")]; + tensor q_33_perm_0 = const()[name = tensor("q_33_perm_0"), val = tensor([1, 0, 2])]; + tensor var_633 = const()[name = tensor("op_633"), val = tensor([77, 8, 64])]; + tensor var_634_cast_fp16 = reshape(shape = var_633, x = k_31_cast_fp16)[name = tensor("op_634_cast_fp16")]; + tensor k_33_perm_0 = const()[name = tensor("k_33_perm_0"), val = tensor([1, 0, 2])]; + tensor var_640 = const()[name = tensor("op_640"), val = tensor([77, 8, 64])]; + tensor var_641_cast_fp16 = reshape(shape = var_640, x = v_31_cast_fp16)[name = tensor("op_641_cast_fp16")]; + tensor v_33_perm_0 = const()[name = tensor("v_33_perm_0"), val = tensor([1, 0, 2])]; + tensor var_645 = const()[name = tensor("op_645"), val = tensor([1, 8, 77, 64])]; + tensor q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_627_cast_fp16)[name = tensor("transpose_48")]; + tensor q_35_cast_fp16 = reshape(shape = var_645, x = q_33_cast_fp16)[name = tensor("q_35_cast_fp16")]; + tensor var_647 = const()[name = tensor("op_647"), val = tensor([1, 8, 77, 64])]; + tensor k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_634_cast_fp16)[name = tensor("transpose_47")]; + tensor k_35_cast_fp16 = reshape(shape = var_647, x = k_33_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor var_649 = const()[name = tensor("op_649"), val = tensor([1, 8, 77, 64])]; + tensor v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_641_cast_fp16)[name = tensor("transpose_46")]; + tensor v_35_cast_fp16 = reshape(shape = var_649, 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 add_5_cast_fp16 = add(x = matmul_5_cast_fp16, y = attn_mask_7_to_fp16)[name = tensor("add_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 = add_5_cast_fp16)[name = tensor("softmax_5_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_5_cast_fp16, y = v_35_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_652 = const()[name = tensor("op_652"), val = tensor([2, 0, 1, 3])]; + tensor var_657 = const()[name = tensor("op_657"), val = tensor([77, 512])]; + tensor var_653_cast_fp16 = transpose(perm = var_652, x = attn_output_41_cast_fp16)[name = tensor("transpose_45")]; + tensor attn_output_43_cast_fp16 = reshape(shape = var_657, x = var_653_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor model_transformer_resblocks_5_attn_out_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_5_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83790720)))]; + tensor model_transformer_resblocks_5_attn_out_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_5_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84315072)))]; + tensor linear_21_cast_fp16 = linear(bias = model_transformer_resblocks_5_attn_out_proj_bias_to_fp16, weight = model_transformer_resblocks_5_attn_out_proj_weight_to_fp16, x = attn_output_43_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor([77, 1, 512])]; + tensor attn_output_47_cast_fp16 = reshape(shape = var_661, x = linear_21_cast_fp16)[name = tensor("attn_output_47_cast_fp16")]; + tensor var_663_perm_0 = const()[name = tensor("op_663_perm_0"), val = tensor([1, 0, 2])]; + tensor var_663_cast_fp16 = transpose(perm = var_663_perm_0, x = attn_output_47_cast_fp16)[name = tensor("transpose_44")]; + tensor input_53_cast_fp16 = add(x = input_51_cast_fp16, y = var_663_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_5_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_5_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84316160)))]; + tensor model_transformer_resblocks_5_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_5_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84317248)))]; + tensor x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = model_transformer_resblocks_5_ln_2_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_5_ln_2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor model_transformer_resblocks_5_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_5_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84318336)))]; + tensor model_transformer_resblocks_5_mlp_c_fc_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_5_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86415552)))]; + tensor linear_22_cast_fp16 = linear(bias = model_transformer_resblocks_5_mlp_c_fc_bias_to_fp16, weight = model_transformer_resblocks_5_mlp_c_fc_weight_to_fp16, x = x_25_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_22_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor model_transformer_resblocks_5_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_5_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86419712)))]; + tensor model_transformer_resblocks_5_mlp_c_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_5_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88516928)))]; + tensor linear_23_cast_fp16 = linear(bias = model_transformer_resblocks_5_mlp_c_proj_bias_to_fp16, weight = model_transformer_resblocks_5_mlp_c_proj_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_53_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor x_27_axes_0 = const()[name = tensor("x_27_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_6_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_6_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88518016)))]; + tensor model_transformer_resblocks_6_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_6_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88519104)))]; + tensor x_27_cast_fp16 = layer_norm(axes = x_27_axes_0, beta = model_transformer_resblocks_6_ln_1_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_6_ln_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor query_27_perm_0 = const()[name = tensor("query_27_perm_0"), val = tensor([1, 0, 2])]; + tensor model_transformer_resblocks_6_attn_in_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_6_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88520192)))]; + tensor model_transformer_resblocks_6_attn_in_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_6_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90093120)))]; + tensor query_27_cast_fp16 = transpose(perm = query_27_perm_0, x = x_27_cast_fp16)[name = tensor("transpose_43")]; + tensor linear_24_cast_fp16 = linear(bias = model_transformer_resblocks_6_attn_in_proj_bias_to_fp16, weight = model_transformer_resblocks_6_attn_in_proj_weight_to_fp16, x = query_27_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([77, 1, 3, 512])]; + tensor var_716_cast_fp16 = reshape(shape = concat_6, x = linear_24_cast_fp16)[name = tensor("op_716_cast_fp16")]; + tensor var_717_axes_0 = const()[name = tensor("op_717_axes_0"), val = tensor([0])]; + tensor var_717_cast_fp16 = expand_dims(axes = var_717_axes_0, x = var_716_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_718_perm_0 = const()[name = tensor("op_718_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_719_axes_0 = const()[name = tensor("op_719_axes_0"), val = tensor([-2])]; + tensor var_718_cast_fp16 = transpose(perm = var_718_perm_0, x = var_717_cast_fp16)[name = tensor("transpose_42")]; + tensor var_719_cast_fp16 = squeeze(axes = var_719_axes_0, x = var_718_cast_fp16)[name = tensor("op_719_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, 77, 1, 512])]; + 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_719_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, 77, 1, 512])]; + 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_719_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, 77, 1, 512])]; + 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_719_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_728 = const()[name = tensor("op_728"), val = tensor([77, 8, 64])]; + tensor var_729_cast_fp16 = reshape(shape = var_728, x = q_37_cast_fp16)[name = tensor("op_729_cast_fp16")]; + tensor q_39_perm_0 = const()[name = tensor("q_39_perm_0"), val = tensor([1, 0, 2])]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor([77, 8, 64])]; + tensor var_736_cast_fp16 = reshape(shape = var_735, x = k_37_cast_fp16)[name = tensor("op_736_cast_fp16")]; + tensor k_39_perm_0 = const()[name = tensor("k_39_perm_0"), val = tensor([1, 0, 2])]; + tensor var_742 = const()[name = tensor("op_742"), val = tensor([77, 8, 64])]; + tensor var_743_cast_fp16 = reshape(shape = var_742, x = v_37_cast_fp16)[name = tensor("op_743_cast_fp16")]; + tensor v_39_perm_0 = const()[name = tensor("v_39_perm_0"), val = tensor([1, 0, 2])]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 8, 77, 64])]; + tensor q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = var_729_cast_fp16)[name = tensor("transpose_41")]; + tensor q_41_cast_fp16 = reshape(shape = var_747, x = q_39_cast_fp16)[name = tensor("q_41_cast_fp16")]; + tensor var_749 = const()[name = tensor("op_749"), val = tensor([1, 8, 77, 64])]; + tensor k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = var_736_cast_fp16)[name = tensor("transpose_40")]; + tensor k_41_cast_fp16 = reshape(shape = var_749, x = k_39_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor var_751 = const()[name = tensor("op_751"), val = tensor([1, 8, 77, 64])]; + tensor v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_743_cast_fp16)[name = tensor("transpose_39")]; + tensor v_41_cast_fp16 = reshape(shape = var_751, 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 add_6_cast_fp16 = add(x = matmul_6_cast_fp16, y = attn_mask_7_to_fp16)[name = tensor("add_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 = add_6_cast_fp16)[name = tensor("softmax_6_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_6_cast_fp16, y = v_41_cast_fp16)[name = tensor("attn_output_49_cast_fp16")]; + tensor var_754 = const()[name = tensor("op_754"), val = tensor([2, 0, 1, 3])]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor([77, 512])]; + tensor var_755_cast_fp16 = transpose(perm = var_754, x = attn_output_49_cast_fp16)[name = tensor("transpose_38")]; + tensor attn_output_51_cast_fp16 = reshape(shape = var_759, x = var_755_cast_fp16)[name = tensor("attn_output_51_cast_fp16")]; + tensor model_transformer_resblocks_6_attn_out_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_6_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90096256)))]; + tensor model_transformer_resblocks_6_attn_out_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_6_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90620608)))]; + tensor linear_25_cast_fp16 = linear(bias = model_transformer_resblocks_6_attn_out_proj_bias_to_fp16, weight = model_transformer_resblocks_6_attn_out_proj_weight_to_fp16, x = attn_output_51_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor([77, 1, 512])]; + tensor attn_output_55_cast_fp16 = reshape(shape = var_763, x = linear_25_cast_fp16)[name = tensor("attn_output_55_cast_fp16")]; + tensor var_765_perm_0 = const()[name = tensor("op_765_perm_0"), val = tensor([1, 0, 2])]; + tensor var_765_cast_fp16 = transpose(perm = var_765_perm_0, x = attn_output_55_cast_fp16)[name = tensor("transpose_37")]; + tensor input_63_cast_fp16 = add(x = input_61_cast_fp16, y = var_765_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor x_29_axes_0 = const()[name = tensor("x_29_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_6_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_6_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90621696)))]; + tensor model_transformer_resblocks_6_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_6_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90622784)))]; + tensor x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, beta = model_transformer_resblocks_6_ln_2_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_6_ln_2_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor model_transformer_resblocks_6_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_6_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90623872)))]; + tensor model_transformer_resblocks_6_mlp_c_fc_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_6_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92721088)))]; + tensor linear_26_cast_fp16 = linear(bias = model_transformer_resblocks_6_mlp_c_fc_bias_to_fp16, weight = model_transformer_resblocks_6_mlp_c_fc_weight_to_fp16, x = x_29_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_69_mode_0 = const()[name = tensor("input_69_mode_0"), val = tensor("EXACT")]; + tensor input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = linear_26_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor model_transformer_resblocks_6_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_6_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92725248)))]; + tensor model_transformer_resblocks_6_mlp_c_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_6_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94822464)))]; + tensor linear_27_cast_fp16 = linear(bias = model_transformer_resblocks_6_mlp_c_proj_bias_to_fp16, weight = model_transformer_resblocks_6_mlp_c_proj_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_71_cast_fp16 = add(x = input_63_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor x_31_axes_0 = const()[name = tensor("x_31_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_7_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_7_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94823552)))]; + tensor model_transformer_resblocks_7_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_7_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94824640)))]; + tensor x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = model_transformer_resblocks_7_ln_1_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_7_ln_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor query_31_perm_0 = const()[name = tensor("query_31_perm_0"), val = tensor([1, 0, 2])]; + tensor model_transformer_resblocks_7_attn_in_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_7_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94825728)))]; + tensor model_transformer_resblocks_7_attn_in_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_7_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96398656)))]; + tensor query_31_cast_fp16 = transpose(perm = query_31_perm_0, x = x_31_cast_fp16)[name = tensor("transpose_36")]; + tensor linear_28_cast_fp16 = linear(bias = model_transformer_resblocks_7_attn_in_proj_bias_to_fp16, weight = model_transformer_resblocks_7_attn_in_proj_weight_to_fp16, x = query_31_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([77, 1, 3, 512])]; + tensor var_818_cast_fp16 = reshape(shape = concat_7, x = linear_28_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor var_819_axes_0 = const()[name = tensor("op_819_axes_0"), val = tensor([0])]; + tensor var_819_cast_fp16 = expand_dims(axes = var_819_axes_0, x = var_818_cast_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_820_perm_0 = const()[name = tensor("op_820_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_821_axes_0 = const()[name = tensor("op_821_axes_0"), val = tensor([-2])]; + tensor var_820_cast_fp16 = transpose(perm = var_820_perm_0, x = var_819_cast_fp16)[name = tensor("transpose_35")]; + tensor var_821_cast_fp16 = squeeze(axes = var_821_axes_0, x = var_820_cast_fp16)[name = tensor("op_821_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, 77, 1, 512])]; + 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_821_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, 77, 1, 512])]; + 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_821_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, 77, 1, 512])]; + 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_821_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_830 = const()[name = tensor("op_830"), val = tensor([77, 8, 64])]; + tensor var_831_cast_fp16 = reshape(shape = var_830, x = q_43_cast_fp16)[name = tensor("op_831_cast_fp16")]; + tensor q_45_perm_0 = const()[name = tensor("q_45_perm_0"), val = tensor([1, 0, 2])]; + tensor var_837 = const()[name = tensor("op_837"), val = tensor([77, 8, 64])]; + tensor var_838_cast_fp16 = reshape(shape = var_837, x = k_43_cast_fp16)[name = tensor("op_838_cast_fp16")]; + tensor k_45_perm_0 = const()[name = tensor("k_45_perm_0"), val = tensor([1, 0, 2])]; + tensor var_844 = const()[name = tensor("op_844"), val = tensor([77, 8, 64])]; + tensor var_845_cast_fp16 = reshape(shape = var_844, x = v_43_cast_fp16)[name = tensor("op_845_cast_fp16")]; + tensor v_45_perm_0 = const()[name = tensor("v_45_perm_0"), val = tensor([1, 0, 2])]; + tensor var_849 = const()[name = tensor("op_849"), val = tensor([1, 8, 77, 64])]; + tensor q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_831_cast_fp16)[name = tensor("transpose_34")]; + tensor q_47_cast_fp16 = reshape(shape = var_849, x = q_45_cast_fp16)[name = tensor("q_47_cast_fp16")]; + tensor var_851 = const()[name = tensor("op_851"), val = tensor([1, 8, 77, 64])]; + tensor k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_838_cast_fp16)[name = tensor("transpose_33")]; + tensor k_47_cast_fp16 = reshape(shape = var_851, x = k_45_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 8, 77, 64])]; + tensor v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_845_cast_fp16)[name = tensor("transpose_32")]; + tensor v_47_cast_fp16 = reshape(shape = var_853, 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 add_7_cast_fp16 = add(x = matmul_7_cast_fp16, y = attn_mask_7_to_fp16)[name = tensor("add_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 = add_7_cast_fp16)[name = tensor("softmax_7_cast_fp16")]; + tensor attn_output_57_transpose_x_0 = const()[name = tensor("attn_output_57_transpose_x_0"), val = tensor(false)]; + tensor attn_output_57_transpose_y_0 = const()[name = tensor("attn_output_57_transpose_y_0"), val = tensor(false)]; + tensor attn_output_57_cast_fp16 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = softmax_7_cast_fp16, y = v_47_cast_fp16)[name = tensor("attn_output_57_cast_fp16")]; + tensor var_856 = const()[name = tensor("op_856"), val = tensor([2, 0, 1, 3])]; + tensor var_861 = const()[name = tensor("op_861"), val = tensor([77, 512])]; + tensor var_857_cast_fp16 = transpose(perm = var_856, x = attn_output_57_cast_fp16)[name = tensor("transpose_31")]; + tensor attn_output_59_cast_fp16 = reshape(shape = var_861, x = var_857_cast_fp16)[name = tensor("attn_output_59_cast_fp16")]; + tensor model_transformer_resblocks_7_attn_out_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_7_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96401792)))]; + tensor model_transformer_resblocks_7_attn_out_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_7_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96926144)))]; + tensor linear_29_cast_fp16 = linear(bias = model_transformer_resblocks_7_attn_out_proj_bias_to_fp16, weight = model_transformer_resblocks_7_attn_out_proj_weight_to_fp16, x = attn_output_59_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_865 = const()[name = tensor("op_865"), val = tensor([77, 1, 512])]; + tensor attn_output_63_cast_fp16 = reshape(shape = var_865, x = linear_29_cast_fp16)[name = tensor("attn_output_63_cast_fp16")]; + tensor var_867_perm_0 = const()[name = tensor("op_867_perm_0"), val = tensor([1, 0, 2])]; + tensor var_867_cast_fp16 = transpose(perm = var_867_perm_0, x = attn_output_63_cast_fp16)[name = tensor("transpose_30")]; + tensor input_73_cast_fp16 = add(x = input_71_cast_fp16, y = var_867_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor x_33_axes_0 = const()[name = tensor("x_33_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_7_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_7_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96927232)))]; + tensor model_transformer_resblocks_7_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_7_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96928320)))]; + tensor x_33_cast_fp16 = layer_norm(axes = x_33_axes_0, beta = model_transformer_resblocks_7_ln_2_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_7_ln_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor model_transformer_resblocks_7_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_7_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96929408)))]; + tensor model_transformer_resblocks_7_mlp_c_fc_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_7_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99026624)))]; + tensor linear_30_cast_fp16 = linear(bias = model_transformer_resblocks_7_mlp_c_fc_bias_to_fp16, weight = model_transformer_resblocks_7_mlp_c_fc_weight_to_fp16, x = x_33_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor input_79_mode_0 = const()[name = tensor("input_79_mode_0"), val = tensor("EXACT")]; + tensor input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = linear_30_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor model_transformer_resblocks_7_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_7_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99030784)))]; + tensor model_transformer_resblocks_7_mlp_c_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_7_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101128000)))]; + tensor linear_31_cast_fp16 = linear(bias = model_transformer_resblocks_7_mlp_c_proj_bias_to_fp16, weight = model_transformer_resblocks_7_mlp_c_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = input_73_cast_fp16, y = linear_31_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor x_35_axes_0 = const()[name = tensor("x_35_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_8_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_8_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101129088)))]; + tensor model_transformer_resblocks_8_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_8_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101130176)))]; + tensor x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = model_transformer_resblocks_8_ln_1_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_8_ln_1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor query_35_perm_0 = const()[name = tensor("query_35_perm_0"), val = tensor([1, 0, 2])]; + tensor model_transformer_resblocks_8_attn_in_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_8_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101131264)))]; + tensor model_transformer_resblocks_8_attn_in_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_8_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102704192)))]; + tensor query_35_cast_fp16 = transpose(perm = query_35_perm_0, x = x_35_cast_fp16)[name = tensor("transpose_29")]; + tensor linear_32_cast_fp16 = linear(bias = model_transformer_resblocks_8_attn_in_proj_bias_to_fp16, weight = model_transformer_resblocks_8_attn_in_proj_weight_to_fp16, x = query_35_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor concat_8 = const()[name = tensor("concat_8"), val = tensor([77, 1, 3, 512])]; + tensor var_920_cast_fp16 = reshape(shape = concat_8, x = linear_32_cast_fp16)[name = tensor("op_920_cast_fp16")]; + tensor var_921_axes_0 = const()[name = tensor("op_921_axes_0"), val = tensor([0])]; + tensor var_921_cast_fp16 = expand_dims(axes = var_921_axes_0, x = var_920_cast_fp16)[name = tensor("op_921_cast_fp16")]; + tensor var_922_perm_0 = const()[name = tensor("op_922_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_923_axes_0 = const()[name = tensor("op_923_axes_0"), val = tensor([-2])]; + tensor var_922_cast_fp16 = transpose(perm = var_922_perm_0, x = var_921_cast_fp16)[name = tensor("transpose_28")]; + tensor var_923_cast_fp16 = squeeze(axes = var_923_axes_0, x = var_922_cast_fp16)[name = tensor("op_923_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, 77, 1, 512])]; + 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_923_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, 77, 1, 512])]; + 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_923_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, 77, 1, 512])]; + 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_923_cast_fp16)[name = tensor("v_49_cast_fp16")]; + tensor var_932 = const()[name = tensor("op_932"), val = tensor([77, 8, 64])]; + tensor var_933_cast_fp16 = reshape(shape = var_932, x = q_49_cast_fp16)[name = tensor("op_933_cast_fp16")]; + tensor q_51_perm_0 = const()[name = tensor("q_51_perm_0"), val = tensor([1, 0, 2])]; + tensor var_939 = const()[name = tensor("op_939"), val = tensor([77, 8, 64])]; + tensor var_940_cast_fp16 = reshape(shape = var_939, x = k_49_cast_fp16)[name = tensor("op_940_cast_fp16")]; + tensor k_51_perm_0 = const()[name = tensor("k_51_perm_0"), val = tensor([1, 0, 2])]; + tensor var_946 = const()[name = tensor("op_946"), val = tensor([77, 8, 64])]; + tensor var_947_cast_fp16 = reshape(shape = var_946, x = v_49_cast_fp16)[name = tensor("op_947_cast_fp16")]; + tensor v_51_perm_0 = const()[name = tensor("v_51_perm_0"), val = tensor([1, 0, 2])]; + tensor var_951 = const()[name = tensor("op_951"), val = tensor([1, 8, 77, 64])]; + tensor q_51_cast_fp16 = transpose(perm = q_51_perm_0, x = var_933_cast_fp16)[name = tensor("transpose_27")]; + tensor q_53_cast_fp16 = reshape(shape = var_951, x = q_51_cast_fp16)[name = tensor("q_53_cast_fp16")]; + tensor var_953 = const()[name = tensor("op_953"), val = tensor([1, 8, 77, 64])]; + tensor k_51_cast_fp16 = transpose(perm = k_51_perm_0, x = var_940_cast_fp16)[name = tensor("transpose_26")]; + tensor k_53_cast_fp16 = reshape(shape = var_953, x = k_51_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor var_955 = const()[name = tensor("op_955"), val = tensor([1, 8, 77, 64])]; + tensor v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = var_947_cast_fp16)[name = tensor("transpose_25")]; + tensor v_53_cast_fp16 = reshape(shape = var_955, 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 add_8_cast_fp16 = add(x = matmul_8_cast_fp16, y = attn_mask_7_to_fp16)[name = tensor("add_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 = add_8_cast_fp16)[name = tensor("softmax_8_cast_fp16")]; + tensor attn_output_65_transpose_x_0 = const()[name = tensor("attn_output_65_transpose_x_0"), val = tensor(false)]; + tensor attn_output_65_transpose_y_0 = const()[name = tensor("attn_output_65_transpose_y_0"), val = tensor(false)]; + tensor attn_output_65_cast_fp16 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = softmax_8_cast_fp16, y = v_53_cast_fp16)[name = tensor("attn_output_65_cast_fp16")]; + tensor var_958 = const()[name = tensor("op_958"), val = tensor([2, 0, 1, 3])]; + tensor var_963 = const()[name = tensor("op_963"), val = tensor([77, 512])]; + tensor var_959_cast_fp16 = transpose(perm = var_958, x = attn_output_65_cast_fp16)[name = tensor("transpose_24")]; + tensor attn_output_67_cast_fp16 = reshape(shape = var_963, x = var_959_cast_fp16)[name = tensor("attn_output_67_cast_fp16")]; + tensor model_transformer_resblocks_8_attn_out_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_8_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102707328)))]; + tensor model_transformer_resblocks_8_attn_out_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_8_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103231680)))]; + tensor linear_33_cast_fp16 = linear(bias = model_transformer_resblocks_8_attn_out_proj_bias_to_fp16, weight = model_transformer_resblocks_8_attn_out_proj_weight_to_fp16, x = attn_output_67_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_967 = const()[name = tensor("op_967"), val = tensor([77, 1, 512])]; + tensor attn_output_71_cast_fp16 = reshape(shape = var_967, x = linear_33_cast_fp16)[name = tensor("attn_output_71_cast_fp16")]; + tensor var_969_perm_0 = const()[name = tensor("op_969_perm_0"), val = tensor([1, 0, 2])]; + tensor var_969_cast_fp16 = transpose(perm = var_969_perm_0, x = attn_output_71_cast_fp16)[name = tensor("transpose_23")]; + tensor input_83_cast_fp16 = add(x = input_81_cast_fp16, y = var_969_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor x_37_axes_0 = const()[name = tensor("x_37_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_8_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_8_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103232768)))]; + tensor model_transformer_resblocks_8_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_8_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103233856)))]; + tensor x_37_cast_fp16 = layer_norm(axes = x_37_axes_0, beta = model_transformer_resblocks_8_ln_2_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_8_ln_2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor model_transformer_resblocks_8_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_8_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103234944)))]; + tensor model_transformer_resblocks_8_mlp_c_fc_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_8_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105332160)))]; + tensor linear_34_cast_fp16 = linear(bias = model_transformer_resblocks_8_mlp_c_fc_bias_to_fp16, weight = model_transformer_resblocks_8_mlp_c_fc_weight_to_fp16, x = x_37_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_89_mode_0 = const()[name = tensor("input_89_mode_0"), val = tensor("EXACT")]; + tensor input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = linear_34_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor model_transformer_resblocks_8_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_8_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105336320)))]; + tensor model_transformer_resblocks_8_mlp_c_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_8_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107433536)))]; + tensor linear_35_cast_fp16 = linear(bias = model_transformer_resblocks_8_mlp_c_proj_bias_to_fp16, weight = model_transformer_resblocks_8_mlp_c_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_91_cast_fp16 = add(x = input_83_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor x_39_axes_0 = const()[name = tensor("x_39_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_9_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_9_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107434624)))]; + tensor model_transformer_resblocks_9_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_9_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107435712)))]; + tensor x_39_cast_fp16 = layer_norm(axes = x_39_axes_0, beta = model_transformer_resblocks_9_ln_1_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_9_ln_1_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor query_39_perm_0 = const()[name = tensor("query_39_perm_0"), val = tensor([1, 0, 2])]; + tensor model_transformer_resblocks_9_attn_in_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_9_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107436800)))]; + tensor model_transformer_resblocks_9_attn_in_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_9_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109009728)))]; + tensor query_39_cast_fp16 = transpose(perm = query_39_perm_0, x = x_39_cast_fp16)[name = tensor("transpose_22")]; + tensor linear_36_cast_fp16 = linear(bias = model_transformer_resblocks_9_attn_in_proj_bias_to_fp16, weight = model_transformer_resblocks_9_attn_in_proj_weight_to_fp16, x = query_39_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([77, 1, 3, 512])]; + tensor var_1022_cast_fp16 = reshape(shape = concat_9, x = linear_36_cast_fp16)[name = tensor("op_1022_cast_fp16")]; + tensor var_1023_axes_0 = const()[name = tensor("op_1023_axes_0"), val = tensor([0])]; + tensor var_1023_cast_fp16 = expand_dims(axes = var_1023_axes_0, x = var_1022_cast_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_perm_0 = const()[name = tensor("op_1024_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1025_axes_0 = const()[name = tensor("op_1025_axes_0"), val = tensor([-2])]; + tensor var_1024_cast_fp16 = transpose(perm = var_1024_perm_0, x = var_1023_cast_fp16)[name = tensor("transpose_21")]; + tensor var_1025_cast_fp16 = squeeze(axes = var_1025_axes_0, x = var_1024_cast_fp16)[name = tensor("op_1025_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, 77, 1, 512])]; + 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_1025_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, 77, 1, 512])]; + 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_1025_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, 77, 1, 512])]; + 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_1025_cast_fp16)[name = tensor("v_55_cast_fp16")]; + tensor var_1034 = const()[name = tensor("op_1034"), val = tensor([77, 8, 64])]; + tensor var_1035_cast_fp16 = reshape(shape = var_1034, x = q_55_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor q_57_perm_0 = const()[name = tensor("q_57_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1041 = const()[name = tensor("op_1041"), val = tensor([77, 8, 64])]; + tensor var_1042_cast_fp16 = reshape(shape = var_1041, x = k_55_cast_fp16)[name = tensor("op_1042_cast_fp16")]; + tensor k_57_perm_0 = const()[name = tensor("k_57_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1048 = const()[name = tensor("op_1048"), val = tensor([77, 8, 64])]; + tensor var_1049_cast_fp16 = reshape(shape = var_1048, x = v_55_cast_fp16)[name = tensor("op_1049_cast_fp16")]; + tensor v_57_perm_0 = const()[name = tensor("v_57_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1053 = const()[name = tensor("op_1053"), val = tensor([1, 8, 77, 64])]; + tensor q_57_cast_fp16 = transpose(perm = q_57_perm_0, x = var_1035_cast_fp16)[name = tensor("transpose_20")]; + tensor q_59_cast_fp16 = reshape(shape = var_1053, x = q_57_cast_fp16)[name = tensor("q_59_cast_fp16")]; + tensor var_1055 = const()[name = tensor("op_1055"), val = tensor([1, 8, 77, 64])]; + tensor k_57_cast_fp16 = transpose(perm = k_57_perm_0, x = var_1042_cast_fp16)[name = tensor("transpose_19")]; + tensor k_59_cast_fp16 = reshape(shape = var_1055, x = k_57_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor var_1057 = const()[name = tensor("op_1057"), val = tensor([1, 8, 77, 64])]; + tensor v_57_cast_fp16 = transpose(perm = v_57_perm_0, x = var_1049_cast_fp16)[name = tensor("transpose_18")]; + tensor v_59_cast_fp16 = reshape(shape = var_1057, 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 add_9_cast_fp16 = add(x = matmul_9_cast_fp16, y = attn_mask_7_to_fp16)[name = tensor("add_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 = add_9_cast_fp16)[name = tensor("softmax_9_cast_fp16")]; + tensor attn_output_73_transpose_x_0 = const()[name = tensor("attn_output_73_transpose_x_0"), val = tensor(false)]; + tensor attn_output_73_transpose_y_0 = const()[name = tensor("attn_output_73_transpose_y_0"), val = tensor(false)]; + tensor attn_output_73_cast_fp16 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = softmax_9_cast_fp16, y = v_59_cast_fp16)[name = tensor("attn_output_73_cast_fp16")]; + tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([2, 0, 1, 3])]; + tensor var_1065 = const()[name = tensor("op_1065"), val = tensor([77, 512])]; + tensor var_1061_cast_fp16 = transpose(perm = var_1060, x = attn_output_73_cast_fp16)[name = tensor("transpose_17")]; + tensor attn_output_75_cast_fp16 = reshape(shape = var_1065, x = var_1061_cast_fp16)[name = tensor("attn_output_75_cast_fp16")]; + tensor model_transformer_resblocks_9_attn_out_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_9_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109012864)))]; + tensor model_transformer_resblocks_9_attn_out_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_9_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109537216)))]; + tensor linear_37_cast_fp16 = linear(bias = model_transformer_resblocks_9_attn_out_proj_bias_to_fp16, weight = model_transformer_resblocks_9_attn_out_proj_weight_to_fp16, x = attn_output_75_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_1069 = const()[name = tensor("op_1069"), val = tensor([77, 1, 512])]; + tensor attn_output_79_cast_fp16 = reshape(shape = var_1069, x = linear_37_cast_fp16)[name = tensor("attn_output_79_cast_fp16")]; + tensor var_1071_perm_0 = const()[name = tensor("op_1071_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1071_cast_fp16 = transpose(perm = var_1071_perm_0, x = attn_output_79_cast_fp16)[name = tensor("transpose_16")]; + tensor input_93_cast_fp16 = add(x = input_91_cast_fp16, y = var_1071_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor x_41_axes_0 = const()[name = tensor("x_41_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_9_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_9_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109538304)))]; + tensor model_transformer_resblocks_9_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_9_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109539392)))]; + tensor x_41_cast_fp16 = layer_norm(axes = x_41_axes_0, beta = model_transformer_resblocks_9_ln_2_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_9_ln_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor model_transformer_resblocks_9_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_9_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109540480)))]; + tensor model_transformer_resblocks_9_mlp_c_fc_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_9_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(111637696)))]; + tensor linear_38_cast_fp16 = linear(bias = model_transformer_resblocks_9_mlp_c_fc_bias_to_fp16, weight = model_transformer_resblocks_9_mlp_c_fc_weight_to_fp16, x = x_41_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor input_99_mode_0 = const()[name = tensor("input_99_mode_0"), val = tensor("EXACT")]; + tensor input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = linear_38_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor model_transformer_resblocks_9_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_9_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(111641856)))]; + tensor model_transformer_resblocks_9_mlp_c_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_9_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113739072)))]; + tensor linear_39_cast_fp16 = linear(bias = model_transformer_resblocks_9_mlp_c_proj_bias_to_fp16, weight = model_transformer_resblocks_9_mlp_c_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_101_cast_fp16 = add(x = input_93_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor x_43_axes_0 = const()[name = tensor("x_43_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_10_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_10_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113740160)))]; + tensor model_transformer_resblocks_10_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_10_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113741248)))]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = model_transformer_resblocks_10_ln_1_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_10_ln_1_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor query_43_perm_0 = const()[name = tensor("query_43_perm_0"), val = tensor([1, 0, 2])]; + tensor model_transformer_resblocks_10_attn_in_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_10_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113742336)))]; + tensor model_transformer_resblocks_10_attn_in_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_10_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115315264)))]; + tensor query_43_cast_fp16 = transpose(perm = query_43_perm_0, x = x_43_cast_fp16)[name = tensor("transpose_15")]; + tensor linear_40_cast_fp16 = linear(bias = model_transformer_resblocks_10_attn_in_proj_bias_to_fp16, weight = model_transformer_resblocks_10_attn_in_proj_weight_to_fp16, x = query_43_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([77, 1, 3, 512])]; + tensor var_1124_cast_fp16 = reshape(shape = concat_10, x = linear_40_cast_fp16)[name = tensor("op_1124_cast_fp16")]; + tensor var_1125_axes_0 = const()[name = tensor("op_1125_axes_0"), val = tensor([0])]; + tensor var_1125_cast_fp16 = expand_dims(axes = var_1125_axes_0, x = var_1124_cast_fp16)[name = tensor("op_1125_cast_fp16")]; + tensor var_1126_perm_0 = const()[name = tensor("op_1126_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([-2])]; + tensor var_1126_cast_fp16 = transpose(perm = var_1126_perm_0, x = var_1125_cast_fp16)[name = tensor("transpose_14")]; + tensor var_1127_cast_fp16 = squeeze(axes = var_1127_axes_0, x = var_1126_cast_fp16)[name = tensor("op_1127_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, 77, 1, 512])]; + 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_1127_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, 77, 1, 512])]; + 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_1127_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, 77, 1, 512])]; + 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_1127_cast_fp16)[name = tensor("v_61_cast_fp16")]; + tensor var_1136 = const()[name = tensor("op_1136"), val = tensor([77, 8, 64])]; + tensor var_1137_cast_fp16 = reshape(shape = var_1136, x = q_61_cast_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor q_63_perm_0 = const()[name = tensor("q_63_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([77, 8, 64])]; + tensor var_1144_cast_fp16 = reshape(shape = var_1143, x = k_61_cast_fp16)[name = tensor("op_1144_cast_fp16")]; + tensor k_63_perm_0 = const()[name = tensor("k_63_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1150 = const()[name = tensor("op_1150"), val = tensor([77, 8, 64])]; + tensor var_1151_cast_fp16 = reshape(shape = var_1150, x = v_61_cast_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor v_63_perm_0 = const()[name = tensor("v_63_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1155 = const()[name = tensor("op_1155"), val = tensor([1, 8, 77, 64])]; + tensor q_63_cast_fp16 = transpose(perm = q_63_perm_0, x = var_1137_cast_fp16)[name = tensor("transpose_13")]; + tensor q_65_cast_fp16 = reshape(shape = var_1155, x = q_63_cast_fp16)[name = tensor("q_65_cast_fp16")]; + tensor var_1157 = const()[name = tensor("op_1157"), val = tensor([1, 8, 77, 64])]; + tensor k_63_cast_fp16 = transpose(perm = k_63_perm_0, x = var_1144_cast_fp16)[name = tensor("transpose_12")]; + tensor k_65_cast_fp16 = reshape(shape = var_1157, x = k_63_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([1, 8, 77, 64])]; + tensor v_63_cast_fp16 = transpose(perm = v_63_perm_0, x = var_1151_cast_fp16)[name = tensor("transpose_11")]; + tensor v_65_cast_fp16 = reshape(shape = var_1159, 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 add_10_cast_fp16 = add(x = matmul_10_cast_fp16, y = attn_mask_7_to_fp16)[name = tensor("add_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 = add_10_cast_fp16)[name = tensor("softmax_10_cast_fp16")]; + tensor attn_output_81_transpose_x_0 = const()[name = tensor("attn_output_81_transpose_x_0"), val = tensor(false)]; + tensor attn_output_81_transpose_y_0 = const()[name = tensor("attn_output_81_transpose_y_0"), val = tensor(false)]; + tensor attn_output_81_cast_fp16 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = softmax_10_cast_fp16, y = v_65_cast_fp16)[name = tensor("attn_output_81_cast_fp16")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([2, 0, 1, 3])]; + tensor var_1167 = const()[name = tensor("op_1167"), val = tensor([77, 512])]; + tensor var_1163_cast_fp16 = transpose(perm = var_1162, x = attn_output_81_cast_fp16)[name = tensor("transpose_10")]; + tensor attn_output_83_cast_fp16 = reshape(shape = var_1167, x = var_1163_cast_fp16)[name = tensor("attn_output_83_cast_fp16")]; + tensor model_transformer_resblocks_10_attn_out_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_10_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115318400)))]; + tensor model_transformer_resblocks_10_attn_out_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_10_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115842752)))]; + tensor linear_41_cast_fp16 = linear(bias = model_transformer_resblocks_10_attn_out_proj_bias_to_fp16, weight = model_transformer_resblocks_10_attn_out_proj_weight_to_fp16, x = attn_output_83_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_1171 = const()[name = tensor("op_1171"), val = tensor([77, 1, 512])]; + tensor attn_output_87_cast_fp16 = reshape(shape = var_1171, x = linear_41_cast_fp16)[name = tensor("attn_output_87_cast_fp16")]; + tensor var_1173_perm_0 = const()[name = tensor("op_1173_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1173_cast_fp16 = transpose(perm = var_1173_perm_0, x = attn_output_87_cast_fp16)[name = tensor("transpose_9")]; + tensor input_103_cast_fp16 = add(x = input_101_cast_fp16, y = var_1173_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor x_45_axes_0 = const()[name = tensor("x_45_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_10_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_10_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115843840)))]; + tensor model_transformer_resblocks_10_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_10_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115844928)))]; + tensor x_45_cast_fp16 = layer_norm(axes = x_45_axes_0, beta = model_transformer_resblocks_10_ln_2_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_10_ln_2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor model_transformer_resblocks_10_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_10_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115846016)))]; + tensor model_transformer_resblocks_10_mlp_c_fc_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_10_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117943232)))]; + tensor linear_42_cast_fp16 = linear(bias = model_transformer_resblocks_10_mlp_c_fc_bias_to_fp16, weight = model_transformer_resblocks_10_mlp_c_fc_weight_to_fp16, x = x_45_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_42_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor model_transformer_resblocks_10_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_10_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117947392)))]; + tensor model_transformer_resblocks_10_mlp_c_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_10_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120044608)))]; + tensor linear_43_cast_fp16 = linear(bias = model_transformer_resblocks_10_mlp_c_proj_bias_to_fp16, weight = model_transformer_resblocks_10_mlp_c_proj_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_111_cast_fp16 = add(x = input_103_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor x_47_axes_0 = const()[name = tensor("x_47_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_11_ln_1_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_11_ln_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120045696)))]; + tensor model_transformer_resblocks_11_ln_1_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_11_ln_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120046784)))]; + tensor x_47_cast_fp16 = layer_norm(axes = x_47_axes_0, beta = model_transformer_resblocks_11_ln_1_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_11_ln_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor model_transformer_resblocks_11_attn_in_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_11_attn_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120047872)))]; + tensor model_transformer_resblocks_11_attn_in_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_11_attn_in_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121620800)))]; + tensor query_cast_fp16 = transpose(perm = query_perm_0, x = x_47_cast_fp16)[name = tensor("transpose_8")]; + tensor linear_44_cast_fp16 = linear(bias = model_transformer_resblocks_11_attn_in_proj_bias_to_fp16, weight = model_transformer_resblocks_11_attn_in_proj_weight_to_fp16, x = query_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([77, 1, 3, 512])]; + tensor var_1226_cast_fp16 = reshape(shape = concat_11, x = linear_44_cast_fp16)[name = tensor("op_1226_cast_fp16")]; + tensor var_1227_axes_0 = const()[name = tensor("op_1227_axes_0"), val = tensor([0])]; + tensor var_1227_cast_fp16 = expand_dims(axes = var_1227_axes_0, x = var_1226_cast_fp16)[name = tensor("op_1227_cast_fp16")]; + tensor var_1228_perm_0 = const()[name = tensor("op_1228_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1229_axes_0 = const()[name = tensor("op_1229_axes_0"), val = tensor([-2])]; + tensor var_1228_cast_fp16 = transpose(perm = var_1228_perm_0, x = var_1227_cast_fp16)[name = tensor("transpose_7")]; + tensor var_1229_cast_fp16 = squeeze(axes = var_1229_axes_0, x = var_1228_cast_fp16)[name = tensor("op_1229_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, 77, 1, 512])]; + 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_1229_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, 77, 1, 512])]; + 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_1229_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, 77, 1, 512])]; + 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_1229_cast_fp16)[name = tensor("v_67_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([77, 8, 64])]; + tensor var_1239_cast_fp16 = reshape(shape = var_1238, x = q_67_cast_fp16)[name = tensor("op_1239_cast_fp16")]; + tensor q_69_perm_0 = const()[name = tensor("q_69_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([77, 8, 64])]; + tensor var_1246_cast_fp16 = reshape(shape = var_1245, x = k_67_cast_fp16)[name = tensor("op_1246_cast_fp16")]; + tensor k_69_perm_0 = const()[name = tensor("k_69_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([77, 8, 64])]; + tensor var_1253_cast_fp16 = reshape(shape = var_1252, x = v_67_cast_fp16)[name = tensor("op_1253_cast_fp16")]; + tensor v_69_perm_0 = const()[name = tensor("v_69_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([1, 8, 77, 64])]; + tensor q_69_cast_fp16 = transpose(perm = q_69_perm_0, x = var_1239_cast_fp16)[name = tensor("transpose_6")]; + tensor q_cast_fp16 = reshape(shape = var_1257, x = q_69_cast_fp16)[name = tensor("q_cast_fp16")]; + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([1, 8, 77, 64])]; + tensor k_69_cast_fp16 = transpose(perm = k_69_perm_0, x = var_1246_cast_fp16)[name = tensor("transpose_5")]; + tensor k_cast_fp16 = reshape(shape = var_1259, x = k_69_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([1, 8, 77, 64])]; + tensor v_69_cast_fp16 = transpose(perm = v_69_perm_0, x = var_1253_cast_fp16)[name = tensor("transpose_4")]; + tensor v_cast_fp16 = reshape(shape = var_1261, 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 add_11_cast_fp16 = add(x = matmul_11_cast_fp16, y = attn_mask_7_to_fp16)[name = tensor("add_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 = add_11_cast_fp16)[name = tensor("softmax_11_cast_fp16")]; + tensor attn_output_89_transpose_x_0 = const()[name = tensor("attn_output_89_transpose_x_0"), val = tensor(false)]; + tensor attn_output_89_transpose_y_0 = const()[name = tensor("attn_output_89_transpose_y_0"), val = tensor(false)]; + tensor attn_output_89_cast_fp16 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = softmax_11_cast_fp16, y = v_cast_fp16)[name = tensor("attn_output_89_cast_fp16")]; + tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([2, 0, 1, 3])]; + tensor var_1269 = const()[name = tensor("op_1269"), val = tensor([77, 512])]; + tensor var_1265_cast_fp16 = transpose(perm = var_1264, x = attn_output_89_cast_fp16)[name = tensor("transpose_3")]; + tensor attn_output_91_cast_fp16 = reshape(shape = var_1269, x = var_1265_cast_fp16)[name = tensor("attn_output_91_cast_fp16")]; + tensor model_transformer_resblocks_11_attn_out_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_11_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121623936)))]; + tensor model_transformer_resblocks_11_attn_out_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_11_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122148288)))]; + tensor linear_45_cast_fp16 = linear(bias = model_transformer_resblocks_11_attn_out_proj_bias_to_fp16, weight = model_transformer_resblocks_11_attn_out_proj_weight_to_fp16, x = attn_output_91_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_1273 = const()[name = tensor("op_1273"), val = tensor([77, 1, 512])]; + tensor attn_output_cast_fp16 = reshape(shape = var_1273, x = linear_45_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_1275_perm_0 = const()[name = tensor("op_1275_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1275_cast_fp16 = transpose(perm = var_1275_perm_0, x = attn_output_cast_fp16)[name = tensor("transpose_2")]; + tensor input_113_cast_fp16 = add(x = input_111_cast_fp16, y = var_1275_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor x_49_axes_0 = const()[name = tensor("x_49_axes_0"), val = tensor([-1])]; + tensor model_transformer_resblocks_11_ln_2_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_11_ln_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122149376)))]; + tensor model_transformer_resblocks_11_ln_2_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_11_ln_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122150464)))]; + tensor x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = model_transformer_resblocks_11_ln_2_bias_to_fp16, epsilon = var_34_to_fp16, gamma = model_transformer_resblocks_11_ln_2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor model_transformer_resblocks_11_mlp_c_fc_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_11_mlp_c_fc_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122151552)))]; + tensor model_transformer_resblocks_11_mlp_c_fc_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_11_mlp_c_fc_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124248768)))]; + tensor linear_46_cast_fp16 = linear(bias = model_transformer_resblocks_11_mlp_c_fc_bias_to_fp16, weight = model_transformer_resblocks_11_mlp_c_fc_weight_to_fp16, x = x_49_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_46_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor model_transformer_resblocks_11_mlp_c_proj_weight_to_fp16 = const()[name = tensor("model_transformer_resblocks_11_mlp_c_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124252928)))]; + tensor model_transformer_resblocks_11_mlp_c_proj_bias_to_fp16 = const()[name = tensor("model_transformer_resblocks_11_mlp_c_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126350144)))]; + tensor linear_47_cast_fp16 = linear(bias = model_transformer_resblocks_11_mlp_c_proj_bias_to_fp16, weight = model_transformer_resblocks_11_mlp_c_proj_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_cast_fp16 = add(x = input_113_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor x_51_axes_0 = const()[name = tensor("x_51_axes_0"), val = tensor([-1])]; + tensor model_ln_final_weight_to_fp16 = const()[name = tensor("model_ln_final_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126351232)))]; + tensor model_ln_final_bias_to_fp16 = const()[name = tensor("model_ln_final_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126352320)))]; + tensor var_1296_to_fp16 = const()[name = tensor("op_1296_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_51_cast_fp16 = layer_norm(axes = x_51_axes_0, beta = model_ln_final_bias_to_fp16, epsilon = var_1296_to_fp16, gamma = model_ln_final_weight_to_fp16, x = input_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor var_1311 = const()[name = tensor("op_1311"), val = tensor([0])]; + tensor var_1314_axis_0 = const()[name = tensor("op_1314_axis_0"), val = tensor(-1)]; + tensor var_1314_keep_dims_0 = const()[name = tensor("op_1314_keep_dims_0"), val = tensor(false)]; + tensor var_1314 = reduce_argmax(axis = var_1314_axis_0, keep_dims = var_1314_keep_dims_0, x = text)[name = tensor("op_1314")]; + tensor stack_0_axis_0 = const()[name = tensor("stack_0_axis_0"), val = tensor(1)]; + tensor stack_0 = stack(axis = stack_0_axis_0, values = (var_1311, var_1314))[name = tensor("stack_0")]; + tensor x_transpose_cast_fp16 = gather_nd(indices = stack_0, x = x_51_cast_fp16)[name = tensor("x_transpose_cast_fp16")]; + tensor transpose_1_to_fp16 = const()[name = tensor("transpose_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126353408)))]; + tensor var_1317_bias_0_to_fp16 = const()[name = tensor("op_1317_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126877760)))]; + tensor var_1317_cast_fp16 = linear(bias = var_1317_bias_0_to_fp16, weight = transpose_1_to_fp16, x = x_transpose_cast_fp16)[name = tensor("op_1317_cast_fp16")]; + tensor var_1317_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1317_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1317 = cast(dtype = var_1317_cast_fp16_to_fp32_dtype_0, x = var_1317_cast_fp16)[name = tensor("cast_133")]; + } -> (var_1317); +} \ No newline at end of file