diff --git "a/CoreMLModels/TextEncoder_mobileCLIP_s2.mlmodelc/model.mil" "b/CoreMLModels/TextEncoder_mobileCLIP_s2.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/CoreMLModels/TextEncoder_mobileCLIP_s2.mlmodelc/model.mil" @@ -0,0 +1,726 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "1.13.1"}, {"coremltools-version", "7.0"}})] +{ + func main(tensor input_tokens) { + tensor var_19 = const()[name = tensor("op_19"), val = tensor(-1)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(false)]; + tensor token_emb_1_axis_0 = const()[name = tensor("token_emb_1_axis_0"), val = tensor(0)]; + tensor token_emb_1_batch_dims_0 = const()[name = tensor("token_emb_1_batch_dims_0"), val = tensor(0)]; + tensor original_model_text_encoder_embedding_layer_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_embedding_layer_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor token_emb_1_cast = gather(axis = token_emb_1_axis_0, batch_dims = token_emb_1_batch_dims_0, indices = input_tokens, x = original_model_text_encoder_embedding_layer_weight_to_fp16)[name = tensor("token_emb_1_cast")]; + tensor var_56_to_fp16 = const()[name = tensor("op_56_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50593920)))]; + tensor input_1_cast = add(x = token_emb_1_cast, y = var_56_to_fp16)[name = tensor("input_1_cast")]; + tensor var_68_axes_0 = const()[name = tensor("op_68_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_0_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_0_pre_norm_mha_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50672832)))]; + tensor original_model_text_encoder_transformer_0_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_0_pre_norm_mha_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50673920)))]; + tensor var_6_to_fp16 = const()[name = tensor("op_6_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_68_cast = layer_norm(axes = var_68_axes_0, beta = original_model_text_encoder_transformer_0_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_0_pre_norm_mha_0_weight_to_fp16, x = input_1_cast)[name = tensor("op_68_cast")]; + tensor original_model_text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50675008)))]; + tensor original_model_text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52247936)))]; + tensor linear_0_cast = linear(bias = original_model_text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_68_cast)[name = tensor("linear_0_cast")]; + tensor var_80 = const()[name = tensor("op_80"), val = tensor([1, 77, 3, 8, -1])]; + tensor qkv_1_cast = reshape(shape = var_80, x = linear_0_cast)[name = tensor("qkv_1_cast")]; + tensor var_82_perm_0 = const()[name = tensor("op_82_perm_0"), val = tensor([0, 3, 2, 1, 4])]; + tensor query_1_begin_0 = const()[name = tensor("query_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor query_1_end_0 = const()[name = tensor("query_1_end_0"), val = tensor([1, 8, 1, 77, 64])]; + tensor query_1_end_mask_0 = const()[name = tensor("query_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor query_1_squeeze_mask_0 = const()[name = tensor("query_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor transpose_36 = transpose(perm = var_82_perm_0, x = qkv_1_cast)[name = tensor("transpose_36")]; + tensor query_1_cast = slice_by_index(begin = query_1_begin_0, end = query_1_end_0, end_mask = query_1_end_mask_0, squeeze_mask = query_1_squeeze_mask_0, x = transpose_36)[name = tensor("query_1_cast")]; + tensor key_1_begin_0 = const()[name = tensor("key_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor key_1_end_0 = const()[name = tensor("key_1_end_0"), val = tensor([1, 8, 2, 77, 64])]; + tensor key_1_end_mask_0 = const()[name = tensor("key_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor key_1_squeeze_mask_0 = const()[name = tensor("key_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor key_1_cast = slice_by_index(begin = key_1_begin_0, end = key_1_end_0, end_mask = key_1_end_mask_0, squeeze_mask = key_1_squeeze_mask_0, x = transpose_36)[name = tensor("key_1_cast")]; + tensor value_1_begin_0 = const()[name = tensor("value_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor value_1_end_0 = const()[name = tensor("value_1_end_0"), val = tensor([1, 8, 3, 77, 64])]; + tensor value_1_end_mask_0 = const()[name = tensor("value_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor value_1_squeeze_mask_0 = const()[name = tensor("value_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor value_1_cast = slice_by_index(begin = value_1_begin_0, end = value_1_end_0, end_mask = value_1_end_mask_0, squeeze_mask = value_1_squeeze_mask_0, x = transpose_36)[name = tensor("value_1_cast")]; + tensor var_93_to_fp16 = const()[name = tensor("op_93_to_fp16"), val = tensor(0x1p-3)]; + tensor query_3_cast = mul(x = query_1_cast, y = var_93_to_fp16)[name = tensor("query_3_cast")]; + tensor key_3_perm_0 = const()[name = tensor("key_3_perm_0"), val = tensor([0, 1, -1, -2])]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor transpose_35 = transpose(perm = key_3_perm_0, x = key_1_cast)[name = tensor("transpose_35")]; + tensor attn_1_cast = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = query_3_cast, y = transpose_35)[name = tensor("attn_1_cast")]; + tensor attn_as_float_1_cast = softmax(axis = var_19, x = attn_1_cast)[name = tensor("attn_as_float_1_cast")]; + tensor out_1_transpose_x_0 = const()[name = tensor("out_1_transpose_x_0"), val = tensor(false)]; + tensor out_1_transpose_y_0 = const()[name = tensor("out_1_transpose_y_0"), val = tensor(false)]; + tensor out_1_cast = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = attn_as_float_1_cast, y = value_1_cast)[name = tensor("out_1_cast")]; + tensor var_102_perm_0 = const()[name = tensor("op_102_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_103 = const()[name = tensor("op_103"), val = tensor([1, 77, -1])]; + tensor transpose_34 = transpose(perm = var_102_perm_0, x = out_1_cast)[name = tensor("transpose_34")]; + tensor input_9_cast = reshape(shape = var_103, x = transpose_34)[name = tensor("input_9_cast")]; + tensor original_model_text_encoder_transformer_0_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_0_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52251072)))]; + tensor original_model_text_encoder_transformer_0_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_0_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52775424)))]; + tensor linear_1_cast = linear(bias = original_model_text_encoder_transformer_0_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_0_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_9_cast)[name = tensor("linear_1_cast")]; + tensor x_5_cast = add(x = linear_1_cast, y = input_1_cast)[name = tensor("x_5_cast")]; + tensor var_117_axes_0 = const()[name = tensor("op_117_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_0_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_0_pre_norm_ffn_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52776512)))]; + tensor original_model_text_encoder_transformer_0_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_0_pre_norm_ffn_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52777600)))]; + tensor var_117_cast = layer_norm(axes = var_117_axes_0, beta = original_model_text_encoder_transformer_0_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_0_pre_norm_ffn_0_weight_to_fp16, x = x_5_cast)[name = tensor("op_117_cast")]; + tensor original_model_text_encoder_transformer_0_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_0_pre_norm_ffn_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(52778688)))]; + tensor original_model_text_encoder_transformer_0_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_0_pre_norm_ffn_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54875904)))]; + tensor linear_2_cast = linear(bias = original_model_text_encoder_transformer_0_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_0_pre_norm_ffn_1_weight_to_fp16, x = var_117_cast)[name = tensor("linear_2_cast")]; + tensor input_19_mode_0 = const()[name = tensor("input_19_mode_0"), val = tensor("EXACT")]; + tensor input_19_cast = gelu(mode = input_19_mode_0, x = linear_2_cast)[name = tensor("input_19_cast")]; + tensor original_model_text_encoder_transformer_0_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_0_pre_norm_ffn_4_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54880064)))]; + tensor original_model_text_encoder_transformer_0_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_0_pre_norm_ffn_4_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56977280)))]; + tensor linear_3_cast = linear(bias = original_model_text_encoder_transformer_0_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_0_pre_norm_ffn_4_weight_to_fp16, x = input_19_cast)[name = tensor("linear_3_cast")]; + tensor x_7_cast = add(x = x_5_cast, y = linear_3_cast)[name = tensor("x_7_cast")]; + tensor var_144_axes_0 = const()[name = tensor("op_144_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_1_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_1_pre_norm_mha_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56978368)))]; + tensor original_model_text_encoder_transformer_1_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_1_pre_norm_mha_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56979456)))]; + tensor var_144_cast = layer_norm(axes = var_144_axes_0, beta = original_model_text_encoder_transformer_1_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_1_pre_norm_mha_0_weight_to_fp16, x = x_7_cast)[name = tensor("op_144_cast")]; + tensor original_model_text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56980544)))]; + tensor original_model_text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58553472)))]; + tensor linear_4_cast = linear(bias = original_model_text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_144_cast)[name = tensor("linear_4_cast")]; + tensor var_156 = const()[name = tensor("op_156"), val = tensor([1, 77, 3, 8, -1])]; + tensor qkv_5_cast = reshape(shape = var_156, x = linear_4_cast)[name = tensor("qkv_5_cast")]; + tensor var_158_perm_0 = const()[name = tensor("op_158_perm_0"), val = tensor([0, 3, 2, 1, 4])]; + tensor query_5_begin_0 = const()[name = tensor("query_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor query_5_end_0 = const()[name = tensor("query_5_end_0"), val = tensor([1, 8, 1, 77, 64])]; + tensor query_5_end_mask_0 = const()[name = tensor("query_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor query_5_squeeze_mask_0 = const()[name = tensor("query_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor transpose_33 = transpose(perm = var_158_perm_0, x = qkv_5_cast)[name = tensor("transpose_33")]; + tensor query_5_cast = slice_by_index(begin = query_5_begin_0, end = query_5_end_0, end_mask = query_5_end_mask_0, squeeze_mask = query_5_squeeze_mask_0, x = transpose_33)[name = tensor("query_5_cast")]; + tensor key_5_begin_0 = const()[name = tensor("key_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor key_5_end_0 = const()[name = tensor("key_5_end_0"), val = tensor([1, 8, 2, 77, 64])]; + tensor key_5_end_mask_0 = const()[name = tensor("key_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor key_5_squeeze_mask_0 = const()[name = tensor("key_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor key_5_cast = slice_by_index(begin = key_5_begin_0, end = key_5_end_0, end_mask = key_5_end_mask_0, squeeze_mask = key_5_squeeze_mask_0, x = transpose_33)[name = tensor("key_5_cast")]; + tensor value_3_begin_0 = const()[name = tensor("value_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor value_3_end_0 = const()[name = tensor("value_3_end_0"), val = tensor([1, 8, 3, 77, 64])]; + tensor value_3_end_mask_0 = const()[name = tensor("value_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor value_3_squeeze_mask_0 = const()[name = tensor("value_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor value_3_cast = slice_by_index(begin = value_3_begin_0, end = value_3_end_0, end_mask = value_3_end_mask_0, squeeze_mask = value_3_squeeze_mask_0, x = transpose_33)[name = tensor("value_3_cast")]; + tensor var_169_to_fp16 = const()[name = tensor("op_169_to_fp16"), val = tensor(0x1p-3)]; + tensor query_7_cast = mul(x = query_5_cast, y = var_169_to_fp16)[name = tensor("query_7_cast")]; + tensor key_7_perm_0 = const()[name = tensor("key_7_perm_0"), val = tensor([0, 1, -1, -2])]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor transpose_32 = transpose(perm = key_7_perm_0, x = key_5_cast)[name = tensor("transpose_32")]; + tensor attn_5_cast = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = query_7_cast, y = transpose_32)[name = tensor("attn_5_cast")]; + tensor attn_as_float_3_cast = softmax(axis = var_19, x = attn_5_cast)[name = tensor("attn_as_float_3_cast")]; + tensor out_3_transpose_x_0 = const()[name = tensor("out_3_transpose_x_0"), val = tensor(false)]; + tensor out_3_transpose_y_0 = const()[name = tensor("out_3_transpose_y_0"), val = tensor(false)]; + tensor out_3_cast = matmul(transpose_x = out_3_transpose_x_0, transpose_y = out_3_transpose_y_0, x = attn_as_float_3_cast, y = value_3_cast)[name = tensor("out_3_cast")]; + tensor var_178_perm_0 = const()[name = tensor("op_178_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_179 = const()[name = tensor("op_179"), val = tensor([1, 77, -1])]; + tensor transpose_31 = transpose(perm = var_178_perm_0, x = out_3_cast)[name = tensor("transpose_31")]; + tensor input_31_cast = reshape(shape = var_179, x = transpose_31)[name = tensor("input_31_cast")]; + tensor original_model_text_encoder_transformer_1_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_1_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58556608)))]; + tensor original_model_text_encoder_transformer_1_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_1_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59080960)))]; + tensor linear_5_cast = linear(bias = original_model_text_encoder_transformer_1_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_1_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_31_cast)[name = tensor("linear_5_cast")]; + tensor x_11_cast = add(x = linear_5_cast, y = x_7_cast)[name = tensor("x_11_cast")]; + tensor var_193_axes_0 = const()[name = tensor("op_193_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_1_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_1_pre_norm_ffn_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59082048)))]; + tensor original_model_text_encoder_transformer_1_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_1_pre_norm_ffn_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59083136)))]; + tensor var_193_cast = layer_norm(axes = var_193_axes_0, beta = original_model_text_encoder_transformer_1_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_1_pre_norm_ffn_0_weight_to_fp16, x = x_11_cast)[name = tensor("op_193_cast")]; + tensor original_model_text_encoder_transformer_1_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_1_pre_norm_ffn_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59084224)))]; + tensor original_model_text_encoder_transformer_1_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_1_pre_norm_ffn_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61181440)))]; + tensor linear_6_cast = linear(bias = original_model_text_encoder_transformer_1_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_1_pre_norm_ffn_1_weight_to_fp16, x = var_193_cast)[name = tensor("linear_6_cast")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast = gelu(mode = input_41_mode_0, x = linear_6_cast)[name = tensor("input_41_cast")]; + tensor original_model_text_encoder_transformer_1_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_1_pre_norm_ffn_4_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(61185600)))]; + tensor original_model_text_encoder_transformer_1_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_1_pre_norm_ffn_4_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63282816)))]; + tensor linear_7_cast = linear(bias = original_model_text_encoder_transformer_1_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_1_pre_norm_ffn_4_weight_to_fp16, x = input_41_cast)[name = tensor("linear_7_cast")]; + tensor x_13_cast = add(x = x_11_cast, y = linear_7_cast)[name = tensor("x_13_cast")]; + tensor var_220_axes_0 = const()[name = tensor("op_220_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_2_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_2_pre_norm_mha_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63283904)))]; + tensor original_model_text_encoder_transformer_2_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_2_pre_norm_mha_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63284992)))]; + tensor var_220_cast = layer_norm(axes = var_220_axes_0, beta = original_model_text_encoder_transformer_2_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_2_pre_norm_mha_0_weight_to_fp16, x = x_13_cast)[name = tensor("op_220_cast")]; + tensor original_model_text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63286080)))]; + tensor original_model_text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64859008)))]; + tensor linear_8_cast = linear(bias = original_model_text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_220_cast)[name = tensor("linear_8_cast")]; + tensor var_232 = const()[name = tensor("op_232"), val = tensor([1, 77, 3, 8, -1])]; + tensor qkv_9_cast = reshape(shape = var_232, x = linear_8_cast)[name = tensor("qkv_9_cast")]; + tensor var_234_perm_0 = const()[name = tensor("op_234_perm_0"), val = tensor([0, 3, 2, 1, 4])]; + tensor query_9_begin_0 = const()[name = tensor("query_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor query_9_end_0 = const()[name = tensor("query_9_end_0"), val = tensor([1, 8, 1, 77, 64])]; + tensor query_9_end_mask_0 = const()[name = tensor("query_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor query_9_squeeze_mask_0 = const()[name = tensor("query_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor transpose_30 = transpose(perm = var_234_perm_0, x = qkv_9_cast)[name = tensor("transpose_30")]; + tensor query_9_cast = slice_by_index(begin = query_9_begin_0, end = query_9_end_0, end_mask = query_9_end_mask_0, squeeze_mask = query_9_squeeze_mask_0, x = transpose_30)[name = tensor("query_9_cast")]; + tensor key_9_begin_0 = const()[name = tensor("key_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor key_9_end_0 = const()[name = tensor("key_9_end_0"), val = tensor([1, 8, 2, 77, 64])]; + tensor key_9_end_mask_0 = const()[name = tensor("key_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor key_9_squeeze_mask_0 = const()[name = tensor("key_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor key_9_cast = slice_by_index(begin = key_9_begin_0, end = key_9_end_0, end_mask = key_9_end_mask_0, squeeze_mask = key_9_squeeze_mask_0, x = transpose_30)[name = tensor("key_9_cast")]; + tensor value_5_begin_0 = const()[name = tensor("value_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor value_5_end_0 = const()[name = tensor("value_5_end_0"), val = tensor([1, 8, 3, 77, 64])]; + tensor value_5_end_mask_0 = const()[name = tensor("value_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor value_5_squeeze_mask_0 = const()[name = tensor("value_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor value_5_cast = slice_by_index(begin = value_5_begin_0, end = value_5_end_0, end_mask = value_5_end_mask_0, squeeze_mask = value_5_squeeze_mask_0, x = transpose_30)[name = tensor("value_5_cast")]; + tensor var_245_to_fp16 = const()[name = tensor("op_245_to_fp16"), val = tensor(0x1p-3)]; + tensor query_11_cast = mul(x = query_9_cast, y = var_245_to_fp16)[name = tensor("query_11_cast")]; + tensor key_11_perm_0 = const()[name = tensor("key_11_perm_0"), val = tensor([0, 1, -1, -2])]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor transpose_29 = transpose(perm = key_11_perm_0, x = key_9_cast)[name = tensor("transpose_29")]; + tensor attn_9_cast = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = query_11_cast, y = transpose_29)[name = tensor("attn_9_cast")]; + tensor attn_as_float_5_cast = softmax(axis = var_19, x = attn_9_cast)[name = tensor("attn_as_float_5_cast")]; + tensor out_5_transpose_x_0 = const()[name = tensor("out_5_transpose_x_0"), val = tensor(false)]; + tensor out_5_transpose_y_0 = const()[name = tensor("out_5_transpose_y_0"), val = tensor(false)]; + tensor out_5_cast = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = attn_as_float_5_cast, y = value_5_cast)[name = tensor("out_5_cast")]; + tensor var_254_perm_0 = const()[name = tensor("op_254_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_255 = const()[name = tensor("op_255"), val = tensor([1, 77, -1])]; + tensor transpose_28 = transpose(perm = var_254_perm_0, x = out_5_cast)[name = tensor("transpose_28")]; + tensor input_53_cast = reshape(shape = var_255, x = transpose_28)[name = tensor("input_53_cast")]; + tensor original_model_text_encoder_transformer_2_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_2_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64862144)))]; + tensor original_model_text_encoder_transformer_2_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_2_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65386496)))]; + tensor linear_9_cast = linear(bias = original_model_text_encoder_transformer_2_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_2_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_53_cast)[name = tensor("linear_9_cast")]; + tensor x_17_cast = add(x = linear_9_cast, y = x_13_cast)[name = tensor("x_17_cast")]; + tensor var_269_axes_0 = const()[name = tensor("op_269_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_2_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_2_pre_norm_ffn_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65387584)))]; + tensor original_model_text_encoder_transformer_2_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_2_pre_norm_ffn_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65388672)))]; + tensor var_269_cast = layer_norm(axes = var_269_axes_0, beta = original_model_text_encoder_transformer_2_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_2_pre_norm_ffn_0_weight_to_fp16, x = x_17_cast)[name = tensor("op_269_cast")]; + tensor original_model_text_encoder_transformer_2_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_2_pre_norm_ffn_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65389760)))]; + tensor original_model_text_encoder_transformer_2_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_2_pre_norm_ffn_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67486976)))]; + tensor linear_10_cast = linear(bias = original_model_text_encoder_transformer_2_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_2_pre_norm_ffn_1_weight_to_fp16, x = var_269_cast)[name = tensor("linear_10_cast")]; + tensor input_63_mode_0 = const()[name = tensor("input_63_mode_0"), val = tensor("EXACT")]; + tensor input_63_cast = gelu(mode = input_63_mode_0, x = linear_10_cast)[name = tensor("input_63_cast")]; + tensor original_model_text_encoder_transformer_2_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_2_pre_norm_ffn_4_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67491136)))]; + tensor original_model_text_encoder_transformer_2_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_2_pre_norm_ffn_4_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69588352)))]; + tensor linear_11_cast = linear(bias = original_model_text_encoder_transformer_2_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_2_pre_norm_ffn_4_weight_to_fp16, x = input_63_cast)[name = tensor("linear_11_cast")]; + tensor x_19_cast = add(x = x_17_cast, y = linear_11_cast)[name = tensor("x_19_cast")]; + tensor var_296_axes_0 = const()[name = tensor("op_296_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_3_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_3_pre_norm_mha_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69589440)))]; + tensor original_model_text_encoder_transformer_3_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_3_pre_norm_mha_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69590528)))]; + tensor var_296_cast = layer_norm(axes = var_296_axes_0, beta = original_model_text_encoder_transformer_3_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_3_pre_norm_mha_0_weight_to_fp16, x = x_19_cast)[name = tensor("op_296_cast")]; + tensor original_model_text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69591616)))]; + tensor original_model_text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71164544)))]; + tensor linear_12_cast = linear(bias = original_model_text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_296_cast)[name = tensor("linear_12_cast")]; + tensor var_308 = const()[name = tensor("op_308"), val = tensor([1, 77, 3, 8, -1])]; + tensor qkv_13_cast = reshape(shape = var_308, x = linear_12_cast)[name = tensor("qkv_13_cast")]; + tensor var_310_perm_0 = const()[name = tensor("op_310_perm_0"), val = tensor([0, 3, 2, 1, 4])]; + tensor query_13_begin_0 = const()[name = tensor("query_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor query_13_end_0 = const()[name = tensor("query_13_end_0"), val = tensor([1, 8, 1, 77, 64])]; + tensor query_13_end_mask_0 = const()[name = tensor("query_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor query_13_squeeze_mask_0 = const()[name = tensor("query_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor transpose_27 = transpose(perm = var_310_perm_0, x = qkv_13_cast)[name = tensor("transpose_27")]; + tensor query_13_cast = slice_by_index(begin = query_13_begin_0, end = query_13_end_0, end_mask = query_13_end_mask_0, squeeze_mask = query_13_squeeze_mask_0, x = transpose_27)[name = tensor("query_13_cast")]; + tensor key_13_begin_0 = const()[name = tensor("key_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor key_13_end_0 = const()[name = tensor("key_13_end_0"), val = tensor([1, 8, 2, 77, 64])]; + tensor key_13_end_mask_0 = const()[name = tensor("key_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor key_13_squeeze_mask_0 = const()[name = tensor("key_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor key_13_cast = slice_by_index(begin = key_13_begin_0, end = key_13_end_0, end_mask = key_13_end_mask_0, squeeze_mask = key_13_squeeze_mask_0, x = transpose_27)[name = tensor("key_13_cast")]; + tensor value_7_begin_0 = const()[name = tensor("value_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor value_7_end_0 = const()[name = tensor("value_7_end_0"), val = tensor([1, 8, 3, 77, 64])]; + tensor value_7_end_mask_0 = const()[name = tensor("value_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor value_7_squeeze_mask_0 = const()[name = tensor("value_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor value_7_cast = slice_by_index(begin = value_7_begin_0, end = value_7_end_0, end_mask = value_7_end_mask_0, squeeze_mask = value_7_squeeze_mask_0, x = transpose_27)[name = tensor("value_7_cast")]; + tensor var_321_to_fp16 = const()[name = tensor("op_321_to_fp16"), val = tensor(0x1p-3)]; + tensor query_15_cast = mul(x = query_13_cast, y = var_321_to_fp16)[name = tensor("query_15_cast")]; + tensor key_15_perm_0 = const()[name = tensor("key_15_perm_0"), val = tensor([0, 1, -1, -2])]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(false)]; + tensor transpose_26 = transpose(perm = key_15_perm_0, x = key_13_cast)[name = tensor("transpose_26")]; + tensor attn_13_cast = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = query_15_cast, y = transpose_26)[name = tensor("attn_13_cast")]; + tensor attn_as_float_7_cast = softmax(axis = var_19, x = attn_13_cast)[name = tensor("attn_as_float_7_cast")]; + tensor out_7_transpose_x_0 = const()[name = tensor("out_7_transpose_x_0"), val = tensor(false)]; + tensor out_7_transpose_y_0 = const()[name = tensor("out_7_transpose_y_0"), val = tensor(false)]; + tensor out_7_cast = matmul(transpose_x = out_7_transpose_x_0, transpose_y = out_7_transpose_y_0, x = attn_as_float_7_cast, y = value_7_cast)[name = tensor("out_7_cast")]; + tensor var_330_perm_0 = const()[name = tensor("op_330_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_331 = const()[name = tensor("op_331"), val = tensor([1, 77, -1])]; + tensor transpose_25 = transpose(perm = var_330_perm_0, x = out_7_cast)[name = tensor("transpose_25")]; + tensor input_75_cast = reshape(shape = var_331, x = transpose_25)[name = tensor("input_75_cast")]; + tensor original_model_text_encoder_transformer_3_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_3_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71167680)))]; + tensor original_model_text_encoder_transformer_3_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_3_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71692032)))]; + tensor linear_13_cast = linear(bias = original_model_text_encoder_transformer_3_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_3_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_75_cast)[name = tensor("linear_13_cast")]; + tensor x_23_cast = add(x = linear_13_cast, y = x_19_cast)[name = tensor("x_23_cast")]; + tensor var_345_axes_0 = const()[name = tensor("op_345_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_3_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_3_pre_norm_ffn_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71693120)))]; + tensor original_model_text_encoder_transformer_3_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_3_pre_norm_ffn_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71694208)))]; + tensor var_345_cast = layer_norm(axes = var_345_axes_0, beta = original_model_text_encoder_transformer_3_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_3_pre_norm_ffn_0_weight_to_fp16, x = x_23_cast)[name = tensor("op_345_cast")]; + tensor original_model_text_encoder_transformer_3_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_3_pre_norm_ffn_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71695296)))]; + tensor original_model_text_encoder_transformer_3_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_3_pre_norm_ffn_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73792512)))]; + tensor linear_14_cast = linear(bias = original_model_text_encoder_transformer_3_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_3_pre_norm_ffn_1_weight_to_fp16, x = var_345_cast)[name = tensor("linear_14_cast")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("EXACT")]; + tensor input_85_cast = gelu(mode = input_85_mode_0, x = linear_14_cast)[name = tensor("input_85_cast")]; + tensor original_model_text_encoder_transformer_3_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_3_pre_norm_ffn_4_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(73796672)))]; + tensor original_model_text_encoder_transformer_3_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_3_pre_norm_ffn_4_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75893888)))]; + tensor linear_15_cast = linear(bias = original_model_text_encoder_transformer_3_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_3_pre_norm_ffn_4_weight_to_fp16, x = input_85_cast)[name = tensor("linear_15_cast")]; + tensor x_25_cast = add(x = x_23_cast, y = linear_15_cast)[name = tensor("x_25_cast")]; + tensor var_372_axes_0 = const()[name = tensor("op_372_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_4_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_4_pre_norm_mha_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75894976)))]; + tensor original_model_text_encoder_transformer_4_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_4_pre_norm_mha_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75896064)))]; + tensor var_372_cast = layer_norm(axes = var_372_axes_0, beta = original_model_text_encoder_transformer_4_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_4_pre_norm_mha_0_weight_to_fp16, x = x_25_cast)[name = tensor("op_372_cast")]; + tensor original_model_text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75897152)))]; + tensor original_model_text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77470080)))]; + tensor linear_16_cast = linear(bias = original_model_text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_372_cast)[name = tensor("linear_16_cast")]; + tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 77, 3, 8, -1])]; + tensor qkv_17_cast = reshape(shape = var_384, x = linear_16_cast)[name = tensor("qkv_17_cast")]; + tensor var_386_perm_0 = const()[name = tensor("op_386_perm_0"), val = tensor([0, 3, 2, 1, 4])]; + tensor query_17_begin_0 = const()[name = tensor("query_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor query_17_end_0 = const()[name = tensor("query_17_end_0"), val = tensor([1, 8, 1, 77, 64])]; + tensor query_17_end_mask_0 = const()[name = tensor("query_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor query_17_squeeze_mask_0 = const()[name = tensor("query_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor transpose_24 = transpose(perm = var_386_perm_0, x = qkv_17_cast)[name = tensor("transpose_24")]; + tensor query_17_cast = slice_by_index(begin = query_17_begin_0, end = query_17_end_0, end_mask = query_17_end_mask_0, squeeze_mask = query_17_squeeze_mask_0, x = transpose_24)[name = tensor("query_17_cast")]; + tensor key_17_begin_0 = const()[name = tensor("key_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor key_17_end_0 = const()[name = tensor("key_17_end_0"), val = tensor([1, 8, 2, 77, 64])]; + tensor key_17_end_mask_0 = const()[name = tensor("key_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor key_17_squeeze_mask_0 = const()[name = tensor("key_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor key_17_cast = slice_by_index(begin = key_17_begin_0, end = key_17_end_0, end_mask = key_17_end_mask_0, squeeze_mask = key_17_squeeze_mask_0, x = transpose_24)[name = tensor("key_17_cast")]; + tensor value_9_begin_0 = const()[name = tensor("value_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor value_9_end_0 = const()[name = tensor("value_9_end_0"), val = tensor([1, 8, 3, 77, 64])]; + tensor value_9_end_mask_0 = const()[name = tensor("value_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor value_9_squeeze_mask_0 = const()[name = tensor("value_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor value_9_cast = slice_by_index(begin = value_9_begin_0, end = value_9_end_0, end_mask = value_9_end_mask_0, squeeze_mask = value_9_squeeze_mask_0, x = transpose_24)[name = tensor("value_9_cast")]; + tensor var_397_to_fp16 = const()[name = tensor("op_397_to_fp16"), val = tensor(0x1p-3)]; + tensor query_19_cast = mul(x = query_17_cast, y = var_397_to_fp16)[name = tensor("query_19_cast")]; + tensor key_19_perm_0 = const()[name = tensor("key_19_perm_0"), val = tensor([0, 1, -1, -2])]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(false)]; + tensor transpose_23 = transpose(perm = key_19_perm_0, x = key_17_cast)[name = tensor("transpose_23")]; + tensor attn_17_cast = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = query_19_cast, y = transpose_23)[name = tensor("attn_17_cast")]; + tensor attn_as_float_9_cast = softmax(axis = var_19, x = attn_17_cast)[name = tensor("attn_as_float_9_cast")]; + tensor out_9_transpose_x_0 = const()[name = tensor("out_9_transpose_x_0"), val = tensor(false)]; + tensor out_9_transpose_y_0 = const()[name = tensor("out_9_transpose_y_0"), val = tensor(false)]; + tensor out_9_cast = matmul(transpose_x = out_9_transpose_x_0, transpose_y = out_9_transpose_y_0, x = attn_as_float_9_cast, y = value_9_cast)[name = tensor("out_9_cast")]; + tensor var_406_perm_0 = const()[name = tensor("op_406_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_407 = const()[name = tensor("op_407"), val = tensor([1, 77, -1])]; + tensor transpose_22 = transpose(perm = var_406_perm_0, x = out_9_cast)[name = tensor("transpose_22")]; + tensor input_97_cast = reshape(shape = var_407, x = transpose_22)[name = tensor("input_97_cast")]; + tensor original_model_text_encoder_transformer_4_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_4_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77473216)))]; + tensor original_model_text_encoder_transformer_4_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_4_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77997568)))]; + tensor linear_17_cast = linear(bias = original_model_text_encoder_transformer_4_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_4_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_97_cast)[name = tensor("linear_17_cast")]; + tensor x_29_cast = add(x = linear_17_cast, y = x_25_cast)[name = tensor("x_29_cast")]; + tensor var_421_axes_0 = const()[name = tensor("op_421_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_4_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_4_pre_norm_ffn_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77998656)))]; + tensor original_model_text_encoder_transformer_4_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_4_pre_norm_ffn_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(77999744)))]; + tensor var_421_cast = layer_norm(axes = var_421_axes_0, beta = original_model_text_encoder_transformer_4_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_4_pre_norm_ffn_0_weight_to_fp16, x = x_29_cast)[name = tensor("op_421_cast")]; + tensor original_model_text_encoder_transformer_4_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_4_pre_norm_ffn_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78000832)))]; + tensor original_model_text_encoder_transformer_4_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_4_pre_norm_ffn_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80098048)))]; + tensor linear_18_cast = linear(bias = original_model_text_encoder_transformer_4_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_4_pre_norm_ffn_1_weight_to_fp16, x = var_421_cast)[name = tensor("linear_18_cast")]; + tensor input_107_mode_0 = const()[name = tensor("input_107_mode_0"), val = tensor("EXACT")]; + tensor input_107_cast = gelu(mode = input_107_mode_0, x = linear_18_cast)[name = tensor("input_107_cast")]; + tensor original_model_text_encoder_transformer_4_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_4_pre_norm_ffn_4_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80102208)))]; + tensor original_model_text_encoder_transformer_4_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_4_pre_norm_ffn_4_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82199424)))]; + tensor linear_19_cast = linear(bias = original_model_text_encoder_transformer_4_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_4_pre_norm_ffn_4_weight_to_fp16, x = input_107_cast)[name = tensor("linear_19_cast")]; + tensor x_31_cast = add(x = x_29_cast, y = linear_19_cast)[name = tensor("x_31_cast")]; + tensor var_448_axes_0 = const()[name = tensor("op_448_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_5_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_5_pre_norm_mha_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82200512)))]; + tensor original_model_text_encoder_transformer_5_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_5_pre_norm_mha_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82201600)))]; + tensor var_448_cast = layer_norm(axes = var_448_axes_0, beta = original_model_text_encoder_transformer_5_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_5_pre_norm_mha_0_weight_to_fp16, x = x_31_cast)[name = tensor("op_448_cast")]; + tensor original_model_text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82202688)))]; + tensor original_model_text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83775616)))]; + tensor linear_20_cast = linear(bias = original_model_text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_448_cast)[name = tensor("linear_20_cast")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 77, 3, 8, -1])]; + tensor qkv_21_cast = reshape(shape = var_460, x = linear_20_cast)[name = tensor("qkv_21_cast")]; + tensor var_462_perm_0 = const()[name = tensor("op_462_perm_0"), val = tensor([0, 3, 2, 1, 4])]; + tensor query_21_begin_0 = const()[name = tensor("query_21_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor query_21_end_0 = const()[name = tensor("query_21_end_0"), val = tensor([1, 8, 1, 77, 64])]; + tensor query_21_end_mask_0 = const()[name = tensor("query_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor query_21_squeeze_mask_0 = const()[name = tensor("query_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor transpose_21 = transpose(perm = var_462_perm_0, x = qkv_21_cast)[name = tensor("transpose_21")]; + tensor query_21_cast = slice_by_index(begin = query_21_begin_0, end = query_21_end_0, end_mask = query_21_end_mask_0, squeeze_mask = query_21_squeeze_mask_0, x = transpose_21)[name = tensor("query_21_cast")]; + tensor key_21_begin_0 = const()[name = tensor("key_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor key_21_end_0 = const()[name = tensor("key_21_end_0"), val = tensor([1, 8, 2, 77, 64])]; + tensor key_21_end_mask_0 = const()[name = tensor("key_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor key_21_squeeze_mask_0 = const()[name = tensor("key_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor key_21_cast = slice_by_index(begin = key_21_begin_0, end = key_21_end_0, end_mask = key_21_end_mask_0, squeeze_mask = key_21_squeeze_mask_0, x = transpose_21)[name = tensor("key_21_cast")]; + tensor value_11_begin_0 = const()[name = tensor("value_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor value_11_end_0 = const()[name = tensor("value_11_end_0"), val = tensor([1, 8, 3, 77, 64])]; + tensor value_11_end_mask_0 = const()[name = tensor("value_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor value_11_squeeze_mask_0 = const()[name = tensor("value_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor value_11_cast = slice_by_index(begin = value_11_begin_0, end = value_11_end_0, end_mask = value_11_end_mask_0, squeeze_mask = value_11_squeeze_mask_0, x = transpose_21)[name = tensor("value_11_cast")]; + tensor var_473_to_fp16 = const()[name = tensor("op_473_to_fp16"), val = tensor(0x1p-3)]; + tensor query_23_cast = mul(x = query_21_cast, y = var_473_to_fp16)[name = tensor("query_23_cast")]; + tensor key_23_perm_0 = const()[name = tensor("key_23_perm_0"), val = tensor([0, 1, -1, -2])]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(false)]; + tensor transpose_20 = transpose(perm = key_23_perm_0, x = key_21_cast)[name = tensor("transpose_20")]; + tensor attn_21_cast = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = query_23_cast, y = transpose_20)[name = tensor("attn_21_cast")]; + tensor attn_as_float_11_cast = softmax(axis = var_19, x = attn_21_cast)[name = tensor("attn_as_float_11_cast")]; + tensor out_11_transpose_x_0 = const()[name = tensor("out_11_transpose_x_0"), val = tensor(false)]; + tensor out_11_transpose_y_0 = const()[name = tensor("out_11_transpose_y_0"), val = tensor(false)]; + tensor out_11_cast = matmul(transpose_x = out_11_transpose_x_0, transpose_y = out_11_transpose_y_0, x = attn_as_float_11_cast, y = value_11_cast)[name = tensor("out_11_cast")]; + tensor var_482_perm_0 = const()[name = tensor("op_482_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_483 = const()[name = tensor("op_483"), val = tensor([1, 77, -1])]; + tensor transpose_19 = transpose(perm = var_482_perm_0, x = out_11_cast)[name = tensor("transpose_19")]; + tensor input_119_cast = reshape(shape = var_483, x = transpose_19)[name = tensor("input_119_cast")]; + tensor original_model_text_encoder_transformer_5_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_5_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83778752)))]; + tensor original_model_text_encoder_transformer_5_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_5_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84303104)))]; + tensor linear_21_cast = linear(bias = original_model_text_encoder_transformer_5_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_5_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_119_cast)[name = tensor("linear_21_cast")]; + tensor x_35_cast = add(x = linear_21_cast, y = x_31_cast)[name = tensor("x_35_cast")]; + tensor var_497_axes_0 = const()[name = tensor("op_497_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_5_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_5_pre_norm_ffn_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84304192)))]; + tensor original_model_text_encoder_transformer_5_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_5_pre_norm_ffn_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84305280)))]; + tensor var_497_cast = layer_norm(axes = var_497_axes_0, beta = original_model_text_encoder_transformer_5_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_5_pre_norm_ffn_0_weight_to_fp16, x = x_35_cast)[name = tensor("op_497_cast")]; + tensor original_model_text_encoder_transformer_5_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_5_pre_norm_ffn_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84306368)))]; + tensor original_model_text_encoder_transformer_5_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_5_pre_norm_ffn_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86403584)))]; + tensor linear_22_cast = linear(bias = original_model_text_encoder_transformer_5_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_5_pre_norm_ffn_1_weight_to_fp16, x = var_497_cast)[name = tensor("linear_22_cast")]; + tensor input_129_mode_0 = const()[name = tensor("input_129_mode_0"), val = tensor("EXACT")]; + tensor input_129_cast = gelu(mode = input_129_mode_0, x = linear_22_cast)[name = tensor("input_129_cast")]; + tensor original_model_text_encoder_transformer_5_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_5_pre_norm_ffn_4_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86407744)))]; + tensor original_model_text_encoder_transformer_5_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_5_pre_norm_ffn_4_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88504960)))]; + tensor linear_23_cast = linear(bias = original_model_text_encoder_transformer_5_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_5_pre_norm_ffn_4_weight_to_fp16, x = input_129_cast)[name = tensor("linear_23_cast")]; + tensor x_37_cast = add(x = x_35_cast, y = linear_23_cast)[name = tensor("x_37_cast")]; + tensor var_524_axes_0 = const()[name = tensor("op_524_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_6_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_6_pre_norm_mha_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88506048)))]; + tensor original_model_text_encoder_transformer_6_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_6_pre_norm_mha_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88507136)))]; + tensor var_524_cast = layer_norm(axes = var_524_axes_0, beta = original_model_text_encoder_transformer_6_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_6_pre_norm_mha_0_weight_to_fp16, x = x_37_cast)[name = tensor("op_524_cast")]; + tensor original_model_text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88508224)))]; + tensor original_model_text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90081152)))]; + tensor linear_24_cast = linear(bias = original_model_text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_524_cast)[name = tensor("linear_24_cast")]; + tensor var_536 = const()[name = tensor("op_536"), val = tensor([1, 77, 3, 8, -1])]; + tensor qkv_25_cast = reshape(shape = var_536, x = linear_24_cast)[name = tensor("qkv_25_cast")]; + tensor var_538_perm_0 = const()[name = tensor("op_538_perm_0"), val = tensor([0, 3, 2, 1, 4])]; + tensor query_25_begin_0 = const()[name = tensor("query_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor query_25_end_0 = const()[name = tensor("query_25_end_0"), val = tensor([1, 8, 1, 77, 64])]; + tensor query_25_end_mask_0 = const()[name = tensor("query_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor query_25_squeeze_mask_0 = const()[name = tensor("query_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor transpose_18 = transpose(perm = var_538_perm_0, x = qkv_25_cast)[name = tensor("transpose_18")]; + tensor query_25_cast = slice_by_index(begin = query_25_begin_0, end = query_25_end_0, end_mask = query_25_end_mask_0, squeeze_mask = query_25_squeeze_mask_0, x = transpose_18)[name = tensor("query_25_cast")]; + tensor key_25_begin_0 = const()[name = tensor("key_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor key_25_end_0 = const()[name = tensor("key_25_end_0"), val = tensor([1, 8, 2, 77, 64])]; + tensor key_25_end_mask_0 = const()[name = tensor("key_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor key_25_squeeze_mask_0 = const()[name = tensor("key_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor key_25_cast = slice_by_index(begin = key_25_begin_0, end = key_25_end_0, end_mask = key_25_end_mask_0, squeeze_mask = key_25_squeeze_mask_0, x = transpose_18)[name = tensor("key_25_cast")]; + tensor value_13_begin_0 = const()[name = tensor("value_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor value_13_end_0 = const()[name = tensor("value_13_end_0"), val = tensor([1, 8, 3, 77, 64])]; + tensor value_13_end_mask_0 = const()[name = tensor("value_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor value_13_squeeze_mask_0 = const()[name = tensor("value_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor value_13_cast = slice_by_index(begin = value_13_begin_0, end = value_13_end_0, end_mask = value_13_end_mask_0, squeeze_mask = value_13_squeeze_mask_0, x = transpose_18)[name = tensor("value_13_cast")]; + tensor var_549_to_fp16 = const()[name = tensor("op_549_to_fp16"), val = tensor(0x1p-3)]; + tensor query_27_cast = mul(x = query_25_cast, y = var_549_to_fp16)[name = tensor("query_27_cast")]; + tensor key_27_perm_0 = const()[name = tensor("key_27_perm_0"), val = tensor([0, 1, -1, -2])]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(false)]; + tensor transpose_17 = transpose(perm = key_27_perm_0, x = key_25_cast)[name = tensor("transpose_17")]; + tensor attn_25_cast = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = query_27_cast, y = transpose_17)[name = tensor("attn_25_cast")]; + tensor attn_as_float_13_cast = softmax(axis = var_19, x = attn_25_cast)[name = tensor("attn_as_float_13_cast")]; + tensor out_13_transpose_x_0 = const()[name = tensor("out_13_transpose_x_0"), val = tensor(false)]; + tensor out_13_transpose_y_0 = const()[name = tensor("out_13_transpose_y_0"), val = tensor(false)]; + tensor out_13_cast = matmul(transpose_x = out_13_transpose_x_0, transpose_y = out_13_transpose_y_0, x = attn_as_float_13_cast, y = value_13_cast)[name = tensor("out_13_cast")]; + tensor var_558_perm_0 = const()[name = tensor("op_558_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 77, -1])]; + tensor transpose_16 = transpose(perm = var_558_perm_0, x = out_13_cast)[name = tensor("transpose_16")]; + tensor input_141_cast = reshape(shape = var_559, x = transpose_16)[name = tensor("input_141_cast")]; + tensor original_model_text_encoder_transformer_6_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_6_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90084288)))]; + tensor original_model_text_encoder_transformer_6_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_6_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90608640)))]; + tensor linear_25_cast = linear(bias = original_model_text_encoder_transformer_6_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_6_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_141_cast)[name = tensor("linear_25_cast")]; + tensor x_41_cast = add(x = linear_25_cast, y = x_37_cast)[name = tensor("x_41_cast")]; + tensor var_573_axes_0 = const()[name = tensor("op_573_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_6_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_6_pre_norm_ffn_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90609728)))]; + tensor original_model_text_encoder_transformer_6_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_6_pre_norm_ffn_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90610816)))]; + tensor var_573_cast = layer_norm(axes = var_573_axes_0, beta = original_model_text_encoder_transformer_6_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_6_pre_norm_ffn_0_weight_to_fp16, x = x_41_cast)[name = tensor("op_573_cast")]; + tensor original_model_text_encoder_transformer_6_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_6_pre_norm_ffn_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(90611904)))]; + tensor original_model_text_encoder_transformer_6_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_6_pre_norm_ffn_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92709120)))]; + tensor linear_26_cast = linear(bias = original_model_text_encoder_transformer_6_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_6_pre_norm_ffn_1_weight_to_fp16, x = var_573_cast)[name = tensor("linear_26_cast")]; + tensor input_151_mode_0 = const()[name = tensor("input_151_mode_0"), val = tensor("EXACT")]; + tensor input_151_cast = gelu(mode = input_151_mode_0, x = linear_26_cast)[name = tensor("input_151_cast")]; + tensor original_model_text_encoder_transformer_6_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_6_pre_norm_ffn_4_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92713280)))]; + tensor original_model_text_encoder_transformer_6_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_6_pre_norm_ffn_4_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94810496)))]; + tensor linear_27_cast = linear(bias = original_model_text_encoder_transformer_6_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_6_pre_norm_ffn_4_weight_to_fp16, x = input_151_cast)[name = tensor("linear_27_cast")]; + tensor x_43_cast = add(x = x_41_cast, y = linear_27_cast)[name = tensor("x_43_cast")]; + tensor var_600_axes_0 = const()[name = tensor("op_600_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_7_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_7_pre_norm_mha_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94811584)))]; + tensor original_model_text_encoder_transformer_7_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_7_pre_norm_mha_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94812672)))]; + tensor var_600_cast = layer_norm(axes = var_600_axes_0, beta = original_model_text_encoder_transformer_7_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_7_pre_norm_mha_0_weight_to_fp16, x = x_43_cast)[name = tensor("op_600_cast")]; + tensor original_model_text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94813760)))]; + tensor original_model_text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96386688)))]; + tensor linear_28_cast = linear(bias = original_model_text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_600_cast)[name = tensor("linear_28_cast")]; + tensor var_612 = const()[name = tensor("op_612"), val = tensor([1, 77, 3, 8, -1])]; + tensor qkv_29_cast = reshape(shape = var_612, x = linear_28_cast)[name = tensor("qkv_29_cast")]; + tensor var_614_perm_0 = const()[name = tensor("op_614_perm_0"), val = tensor([0, 3, 2, 1, 4])]; + tensor query_29_begin_0 = const()[name = tensor("query_29_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor query_29_end_0 = const()[name = tensor("query_29_end_0"), val = tensor([1, 8, 1, 77, 64])]; + tensor query_29_end_mask_0 = const()[name = tensor("query_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor query_29_squeeze_mask_0 = const()[name = tensor("query_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor transpose_15 = transpose(perm = var_614_perm_0, x = qkv_29_cast)[name = tensor("transpose_15")]; + tensor query_29_cast = slice_by_index(begin = query_29_begin_0, end = query_29_end_0, end_mask = query_29_end_mask_0, squeeze_mask = query_29_squeeze_mask_0, x = transpose_15)[name = tensor("query_29_cast")]; + tensor key_29_begin_0 = const()[name = tensor("key_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor key_29_end_0 = const()[name = tensor("key_29_end_0"), val = tensor([1, 8, 2, 77, 64])]; + tensor key_29_end_mask_0 = const()[name = tensor("key_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor key_29_squeeze_mask_0 = const()[name = tensor("key_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor key_29_cast = slice_by_index(begin = key_29_begin_0, end = key_29_end_0, end_mask = key_29_end_mask_0, squeeze_mask = key_29_squeeze_mask_0, x = transpose_15)[name = tensor("key_29_cast")]; + tensor value_15_begin_0 = const()[name = tensor("value_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor value_15_end_0 = const()[name = tensor("value_15_end_0"), val = tensor([1, 8, 3, 77, 64])]; + tensor value_15_end_mask_0 = const()[name = tensor("value_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor value_15_squeeze_mask_0 = const()[name = tensor("value_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor value_15_cast = slice_by_index(begin = value_15_begin_0, end = value_15_end_0, end_mask = value_15_end_mask_0, squeeze_mask = value_15_squeeze_mask_0, x = transpose_15)[name = tensor("value_15_cast")]; + tensor var_625_to_fp16 = const()[name = tensor("op_625_to_fp16"), val = tensor(0x1p-3)]; + tensor query_31_cast = mul(x = query_29_cast, y = var_625_to_fp16)[name = tensor("query_31_cast")]; + tensor key_31_perm_0 = const()[name = tensor("key_31_perm_0"), val = tensor([0, 1, -1, -2])]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(false)]; + tensor transpose_14 = transpose(perm = key_31_perm_0, x = key_29_cast)[name = tensor("transpose_14")]; + tensor attn_29_cast = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = query_31_cast, y = transpose_14)[name = tensor("attn_29_cast")]; + tensor attn_as_float_15_cast = softmax(axis = var_19, x = attn_29_cast)[name = tensor("attn_as_float_15_cast")]; + tensor out_15_transpose_x_0 = const()[name = tensor("out_15_transpose_x_0"), val = tensor(false)]; + tensor out_15_transpose_y_0 = const()[name = tensor("out_15_transpose_y_0"), val = tensor(false)]; + tensor out_15_cast = matmul(transpose_x = out_15_transpose_x_0, transpose_y = out_15_transpose_y_0, x = attn_as_float_15_cast, y = value_15_cast)[name = tensor("out_15_cast")]; + tensor var_634_perm_0 = const()[name = tensor("op_634_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor([1, 77, -1])]; + tensor transpose_13 = transpose(perm = var_634_perm_0, x = out_15_cast)[name = tensor("transpose_13")]; + tensor input_163_cast = reshape(shape = var_635, x = transpose_13)[name = tensor("input_163_cast")]; + tensor original_model_text_encoder_transformer_7_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_7_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96389824)))]; + tensor original_model_text_encoder_transformer_7_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_7_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96914176)))]; + tensor linear_29_cast = linear(bias = original_model_text_encoder_transformer_7_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_7_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_163_cast)[name = tensor("linear_29_cast")]; + tensor x_47_cast = add(x = linear_29_cast, y = x_43_cast)[name = tensor("x_47_cast")]; + tensor var_649_axes_0 = const()[name = tensor("op_649_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_7_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_7_pre_norm_ffn_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96915264)))]; + tensor original_model_text_encoder_transformer_7_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_7_pre_norm_ffn_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96916352)))]; + tensor var_649_cast = layer_norm(axes = var_649_axes_0, beta = original_model_text_encoder_transformer_7_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_7_pre_norm_ffn_0_weight_to_fp16, x = x_47_cast)[name = tensor("op_649_cast")]; + tensor original_model_text_encoder_transformer_7_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_7_pre_norm_ffn_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96917440)))]; + tensor original_model_text_encoder_transformer_7_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_7_pre_norm_ffn_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99014656)))]; + tensor linear_30_cast = linear(bias = original_model_text_encoder_transformer_7_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_7_pre_norm_ffn_1_weight_to_fp16, x = var_649_cast)[name = tensor("linear_30_cast")]; + tensor input_173_mode_0 = const()[name = tensor("input_173_mode_0"), val = tensor("EXACT")]; + tensor input_173_cast = gelu(mode = input_173_mode_0, x = linear_30_cast)[name = tensor("input_173_cast")]; + tensor original_model_text_encoder_transformer_7_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_7_pre_norm_ffn_4_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(99018816)))]; + tensor original_model_text_encoder_transformer_7_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_7_pre_norm_ffn_4_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101116032)))]; + tensor linear_31_cast = linear(bias = original_model_text_encoder_transformer_7_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_7_pre_norm_ffn_4_weight_to_fp16, x = input_173_cast)[name = tensor("linear_31_cast")]; + tensor x_49_cast = add(x = x_47_cast, y = linear_31_cast)[name = tensor("x_49_cast")]; + tensor var_676_axes_0 = const()[name = tensor("op_676_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_8_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_8_pre_norm_mha_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101117120)))]; + tensor original_model_text_encoder_transformer_8_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_8_pre_norm_mha_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101118208)))]; + tensor var_676_cast = layer_norm(axes = var_676_axes_0, beta = original_model_text_encoder_transformer_8_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_8_pre_norm_mha_0_weight_to_fp16, x = x_49_cast)[name = tensor("op_676_cast")]; + tensor original_model_text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101119296)))]; + tensor original_model_text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102692224)))]; + tensor linear_32_cast = linear(bias = original_model_text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_676_cast)[name = tensor("linear_32_cast")]; + tensor var_688 = const()[name = tensor("op_688"), val = tensor([1, 77, 3, 8, -1])]; + tensor qkv_33_cast = reshape(shape = var_688, x = linear_32_cast)[name = tensor("qkv_33_cast")]; + tensor var_690_perm_0 = const()[name = tensor("op_690_perm_0"), val = tensor([0, 3, 2, 1, 4])]; + tensor query_33_begin_0 = const()[name = tensor("query_33_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor query_33_end_0 = const()[name = tensor("query_33_end_0"), val = tensor([1, 8, 1, 77, 64])]; + tensor query_33_end_mask_0 = const()[name = tensor("query_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor query_33_squeeze_mask_0 = const()[name = tensor("query_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor transpose_12 = transpose(perm = var_690_perm_0, x = qkv_33_cast)[name = tensor("transpose_12")]; + tensor query_33_cast = slice_by_index(begin = query_33_begin_0, end = query_33_end_0, end_mask = query_33_end_mask_0, squeeze_mask = query_33_squeeze_mask_0, x = transpose_12)[name = tensor("query_33_cast")]; + tensor key_33_begin_0 = const()[name = tensor("key_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor key_33_end_0 = const()[name = tensor("key_33_end_0"), val = tensor([1, 8, 2, 77, 64])]; + tensor key_33_end_mask_0 = const()[name = tensor("key_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor key_33_squeeze_mask_0 = const()[name = tensor("key_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor key_33_cast = slice_by_index(begin = key_33_begin_0, end = key_33_end_0, end_mask = key_33_end_mask_0, squeeze_mask = key_33_squeeze_mask_0, x = transpose_12)[name = tensor("key_33_cast")]; + tensor value_17_begin_0 = const()[name = tensor("value_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor value_17_end_0 = const()[name = tensor("value_17_end_0"), val = tensor([1, 8, 3, 77, 64])]; + tensor value_17_end_mask_0 = const()[name = tensor("value_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor value_17_squeeze_mask_0 = const()[name = tensor("value_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor value_17_cast = slice_by_index(begin = value_17_begin_0, end = value_17_end_0, end_mask = value_17_end_mask_0, squeeze_mask = value_17_squeeze_mask_0, x = transpose_12)[name = tensor("value_17_cast")]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1p-3)]; + tensor query_35_cast = mul(x = query_33_cast, y = var_701_to_fp16)[name = tensor("query_35_cast")]; + tensor key_35_perm_0 = const()[name = tensor("key_35_perm_0"), val = tensor([0, 1, -1, -2])]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(false)]; + tensor transpose_11 = transpose(perm = key_35_perm_0, x = key_33_cast)[name = tensor("transpose_11")]; + tensor attn_33_cast = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = query_35_cast, y = transpose_11)[name = tensor("attn_33_cast")]; + tensor attn_as_float_17_cast = softmax(axis = var_19, x = attn_33_cast)[name = tensor("attn_as_float_17_cast")]; + tensor out_17_transpose_x_0 = const()[name = tensor("out_17_transpose_x_0"), val = tensor(false)]; + tensor out_17_transpose_y_0 = const()[name = tensor("out_17_transpose_y_0"), val = tensor(false)]; + tensor out_17_cast = matmul(transpose_x = out_17_transpose_x_0, transpose_y = out_17_transpose_y_0, x = attn_as_float_17_cast, y = value_17_cast)[name = tensor("out_17_cast")]; + tensor var_710_perm_0 = const()[name = tensor("op_710_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_711 = const()[name = tensor("op_711"), val = tensor([1, 77, -1])]; + tensor transpose_10 = transpose(perm = var_710_perm_0, x = out_17_cast)[name = tensor("transpose_10")]; + tensor input_185_cast = reshape(shape = var_711, x = transpose_10)[name = tensor("input_185_cast")]; + tensor original_model_text_encoder_transformer_8_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_8_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(102695360)))]; + tensor original_model_text_encoder_transformer_8_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_8_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103219712)))]; + tensor linear_33_cast = linear(bias = original_model_text_encoder_transformer_8_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_8_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_185_cast)[name = tensor("linear_33_cast")]; + tensor x_53_cast = add(x = linear_33_cast, y = x_49_cast)[name = tensor("x_53_cast")]; + tensor var_725_axes_0 = const()[name = tensor("op_725_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_8_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_8_pre_norm_ffn_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103220800)))]; + tensor original_model_text_encoder_transformer_8_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_8_pre_norm_ffn_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103221888)))]; + tensor var_725_cast = layer_norm(axes = var_725_axes_0, beta = original_model_text_encoder_transformer_8_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_8_pre_norm_ffn_0_weight_to_fp16, x = x_53_cast)[name = tensor("op_725_cast")]; + tensor original_model_text_encoder_transformer_8_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_8_pre_norm_ffn_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103222976)))]; + tensor original_model_text_encoder_transformer_8_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_8_pre_norm_ffn_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105320192)))]; + tensor linear_34_cast = linear(bias = original_model_text_encoder_transformer_8_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_8_pre_norm_ffn_1_weight_to_fp16, x = var_725_cast)[name = tensor("linear_34_cast")]; + tensor input_195_mode_0 = const()[name = tensor("input_195_mode_0"), val = tensor("EXACT")]; + tensor input_195_cast = gelu(mode = input_195_mode_0, x = linear_34_cast)[name = tensor("input_195_cast")]; + tensor original_model_text_encoder_transformer_8_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_8_pre_norm_ffn_4_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105324352)))]; + tensor original_model_text_encoder_transformer_8_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_8_pre_norm_ffn_4_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107421568)))]; + tensor linear_35_cast = linear(bias = original_model_text_encoder_transformer_8_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_8_pre_norm_ffn_4_weight_to_fp16, x = input_195_cast)[name = tensor("linear_35_cast")]; + tensor x_55_cast = add(x = x_53_cast, y = linear_35_cast)[name = tensor("x_55_cast")]; + tensor var_752_axes_0 = const()[name = tensor("op_752_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_9_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_9_pre_norm_mha_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107422656)))]; + tensor original_model_text_encoder_transformer_9_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_9_pre_norm_mha_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107423744)))]; + tensor var_752_cast = layer_norm(axes = var_752_axes_0, beta = original_model_text_encoder_transformer_9_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_9_pre_norm_mha_0_weight_to_fp16, x = x_55_cast)[name = tensor("op_752_cast")]; + tensor original_model_text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107424832)))]; + tensor original_model_text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108997760)))]; + tensor linear_36_cast = linear(bias = original_model_text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_752_cast)[name = tensor("linear_36_cast")]; + tensor var_764 = const()[name = tensor("op_764"), val = tensor([1, 77, 3, 8, -1])]; + tensor qkv_37_cast = reshape(shape = var_764, x = linear_36_cast)[name = tensor("qkv_37_cast")]; + tensor var_766_perm_0 = const()[name = tensor("op_766_perm_0"), val = tensor([0, 3, 2, 1, 4])]; + tensor query_37_begin_0 = const()[name = tensor("query_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor query_37_end_0 = const()[name = tensor("query_37_end_0"), val = tensor([1, 8, 1, 77, 64])]; + tensor query_37_end_mask_0 = const()[name = tensor("query_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor query_37_squeeze_mask_0 = const()[name = tensor("query_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor transpose_9 = transpose(perm = var_766_perm_0, x = qkv_37_cast)[name = tensor("transpose_9")]; + tensor query_37_cast = slice_by_index(begin = query_37_begin_0, end = query_37_end_0, end_mask = query_37_end_mask_0, squeeze_mask = query_37_squeeze_mask_0, x = transpose_9)[name = tensor("query_37_cast")]; + tensor key_37_begin_0 = const()[name = tensor("key_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor key_37_end_0 = const()[name = tensor("key_37_end_0"), val = tensor([1, 8, 2, 77, 64])]; + tensor key_37_end_mask_0 = const()[name = tensor("key_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor key_37_squeeze_mask_0 = const()[name = tensor("key_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor key_37_cast = slice_by_index(begin = key_37_begin_0, end = key_37_end_0, end_mask = key_37_end_mask_0, squeeze_mask = key_37_squeeze_mask_0, x = transpose_9)[name = tensor("key_37_cast")]; + tensor value_19_begin_0 = const()[name = tensor("value_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor value_19_end_0 = const()[name = tensor("value_19_end_0"), val = tensor([1, 8, 3, 77, 64])]; + tensor value_19_end_mask_0 = const()[name = tensor("value_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor value_19_squeeze_mask_0 = const()[name = tensor("value_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor value_19_cast = slice_by_index(begin = value_19_begin_0, end = value_19_end_0, end_mask = value_19_end_mask_0, squeeze_mask = value_19_squeeze_mask_0, x = transpose_9)[name = tensor("value_19_cast")]; + tensor var_777_to_fp16 = const()[name = tensor("op_777_to_fp16"), val = tensor(0x1p-3)]; + tensor query_39_cast = mul(x = query_37_cast, y = var_777_to_fp16)[name = tensor("query_39_cast")]; + tensor key_39_perm_0 = const()[name = tensor("key_39_perm_0"), val = tensor([0, 1, -1, -2])]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(false)]; + tensor transpose_8 = transpose(perm = key_39_perm_0, x = key_37_cast)[name = tensor("transpose_8")]; + tensor attn_37_cast = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = query_39_cast, y = transpose_8)[name = tensor("attn_37_cast")]; + tensor attn_as_float_19_cast = softmax(axis = var_19, x = attn_37_cast)[name = tensor("attn_as_float_19_cast")]; + tensor out_19_transpose_x_0 = const()[name = tensor("out_19_transpose_x_0"), val = tensor(false)]; + tensor out_19_transpose_y_0 = const()[name = tensor("out_19_transpose_y_0"), val = tensor(false)]; + tensor out_19_cast = matmul(transpose_x = out_19_transpose_x_0, transpose_y = out_19_transpose_y_0, x = attn_as_float_19_cast, y = value_19_cast)[name = tensor("out_19_cast")]; + tensor var_786_perm_0 = const()[name = tensor("op_786_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor([1, 77, -1])]; + tensor transpose_7 = transpose(perm = var_786_perm_0, x = out_19_cast)[name = tensor("transpose_7")]; + tensor input_207_cast = reshape(shape = var_787, x = transpose_7)[name = tensor("input_207_cast")]; + tensor original_model_text_encoder_transformer_9_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_9_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109000896)))]; + tensor original_model_text_encoder_transformer_9_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_9_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109525248)))]; + tensor linear_37_cast = linear(bias = original_model_text_encoder_transformer_9_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_9_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_207_cast)[name = tensor("linear_37_cast")]; + tensor x_59_cast = add(x = linear_37_cast, y = x_55_cast)[name = tensor("x_59_cast")]; + tensor var_801_axes_0 = const()[name = tensor("op_801_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_9_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_9_pre_norm_ffn_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109526336)))]; + tensor original_model_text_encoder_transformer_9_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_9_pre_norm_ffn_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109527424)))]; + tensor var_801_cast = layer_norm(axes = var_801_axes_0, beta = original_model_text_encoder_transformer_9_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_9_pre_norm_ffn_0_weight_to_fp16, x = x_59_cast)[name = tensor("op_801_cast")]; + tensor original_model_text_encoder_transformer_9_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_9_pre_norm_ffn_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109528512)))]; + tensor original_model_text_encoder_transformer_9_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_9_pre_norm_ffn_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(111625728)))]; + tensor linear_38_cast = linear(bias = original_model_text_encoder_transformer_9_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_9_pre_norm_ffn_1_weight_to_fp16, x = var_801_cast)[name = tensor("linear_38_cast")]; + tensor input_217_mode_0 = const()[name = tensor("input_217_mode_0"), val = tensor("EXACT")]; + tensor input_217_cast = gelu(mode = input_217_mode_0, x = linear_38_cast)[name = tensor("input_217_cast")]; + tensor original_model_text_encoder_transformer_9_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_9_pre_norm_ffn_4_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(111629888)))]; + tensor original_model_text_encoder_transformer_9_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_9_pre_norm_ffn_4_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113727104)))]; + tensor linear_39_cast = linear(bias = original_model_text_encoder_transformer_9_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_9_pre_norm_ffn_4_weight_to_fp16, x = input_217_cast)[name = tensor("linear_39_cast")]; + tensor x_61_cast = add(x = x_59_cast, y = linear_39_cast)[name = tensor("x_61_cast")]; + tensor var_828_axes_0 = const()[name = tensor("op_828_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_10_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_10_pre_norm_mha_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113728192)))]; + tensor original_model_text_encoder_transformer_10_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_10_pre_norm_mha_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113729280)))]; + tensor var_828_cast = layer_norm(axes = var_828_axes_0, beta = original_model_text_encoder_transformer_10_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_10_pre_norm_mha_0_weight_to_fp16, x = x_61_cast)[name = tensor("op_828_cast")]; + tensor original_model_text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113730368)))]; + tensor original_model_text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115303296)))]; + tensor linear_40_cast = linear(bias = original_model_text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_828_cast)[name = tensor("linear_40_cast")]; + tensor var_840 = const()[name = tensor("op_840"), val = tensor([1, 77, 3, 8, -1])]; + tensor qkv_41_cast = reshape(shape = var_840, x = linear_40_cast)[name = tensor("qkv_41_cast")]; + tensor var_842_perm_0 = const()[name = tensor("op_842_perm_0"), val = tensor([0, 3, 2, 1, 4])]; + tensor query_41_begin_0 = const()[name = tensor("query_41_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor query_41_end_0 = const()[name = tensor("query_41_end_0"), val = tensor([1, 8, 1, 77, 64])]; + tensor query_41_end_mask_0 = const()[name = tensor("query_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor query_41_squeeze_mask_0 = const()[name = tensor("query_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor transpose_6 = transpose(perm = var_842_perm_0, x = qkv_41_cast)[name = tensor("transpose_6")]; + tensor query_41_cast = slice_by_index(begin = query_41_begin_0, end = query_41_end_0, end_mask = query_41_end_mask_0, squeeze_mask = query_41_squeeze_mask_0, x = transpose_6)[name = tensor("query_41_cast")]; + tensor key_41_begin_0 = const()[name = tensor("key_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor key_41_end_0 = const()[name = tensor("key_41_end_0"), val = tensor([1, 8, 2, 77, 64])]; + tensor key_41_end_mask_0 = const()[name = tensor("key_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor key_41_squeeze_mask_0 = const()[name = tensor("key_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor key_41_cast = slice_by_index(begin = key_41_begin_0, end = key_41_end_0, end_mask = key_41_end_mask_0, squeeze_mask = key_41_squeeze_mask_0, x = transpose_6)[name = tensor("key_41_cast")]; + tensor value_21_begin_0 = const()[name = tensor("value_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor value_21_end_0 = const()[name = tensor("value_21_end_0"), val = tensor([1, 8, 3, 77, 64])]; + tensor value_21_end_mask_0 = const()[name = tensor("value_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor value_21_squeeze_mask_0 = const()[name = tensor("value_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor value_21_cast = slice_by_index(begin = value_21_begin_0, end = value_21_end_0, end_mask = value_21_end_mask_0, squeeze_mask = value_21_squeeze_mask_0, x = transpose_6)[name = tensor("value_21_cast")]; + tensor var_853_to_fp16 = const()[name = tensor("op_853_to_fp16"), val = tensor(0x1p-3)]; + tensor query_43_cast = mul(x = query_41_cast, y = var_853_to_fp16)[name = tensor("query_43_cast")]; + tensor key_43_perm_0 = const()[name = tensor("key_43_perm_0"), val = tensor([0, 1, -1, -2])]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(false)]; + tensor transpose_5 = transpose(perm = key_43_perm_0, x = key_41_cast)[name = tensor("transpose_5")]; + tensor attn_41_cast = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = query_43_cast, y = transpose_5)[name = tensor("attn_41_cast")]; + tensor attn_as_float_21_cast = softmax(axis = var_19, x = attn_41_cast)[name = tensor("attn_as_float_21_cast")]; + tensor out_21_transpose_x_0 = const()[name = tensor("out_21_transpose_x_0"), val = tensor(false)]; + tensor out_21_transpose_y_0 = const()[name = tensor("out_21_transpose_y_0"), val = tensor(false)]; + tensor out_21_cast = matmul(transpose_x = out_21_transpose_x_0, transpose_y = out_21_transpose_y_0, x = attn_as_float_21_cast, y = value_21_cast)[name = tensor("out_21_cast")]; + tensor var_862_perm_0 = const()[name = tensor("op_862_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_863 = const()[name = tensor("op_863"), val = tensor([1, 77, -1])]; + tensor transpose_4 = transpose(perm = var_862_perm_0, x = out_21_cast)[name = tensor("transpose_4")]; + tensor input_229_cast = reshape(shape = var_863, x = transpose_4)[name = tensor("input_229_cast")]; + tensor original_model_text_encoder_transformer_10_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_10_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115306432)))]; + tensor original_model_text_encoder_transformer_10_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_10_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115830784)))]; + tensor linear_41_cast = linear(bias = original_model_text_encoder_transformer_10_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_10_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_229_cast)[name = tensor("linear_41_cast")]; + tensor x_65_cast = add(x = linear_41_cast, y = x_61_cast)[name = tensor("x_65_cast")]; + tensor var_877_axes_0 = const()[name = tensor("op_877_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_10_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_10_pre_norm_ffn_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115831872)))]; + tensor original_model_text_encoder_transformer_10_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_10_pre_norm_ffn_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115832960)))]; + tensor var_877_cast = layer_norm(axes = var_877_axes_0, beta = original_model_text_encoder_transformer_10_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_10_pre_norm_ffn_0_weight_to_fp16, x = x_65_cast)[name = tensor("op_877_cast")]; + tensor original_model_text_encoder_transformer_10_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_10_pre_norm_ffn_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115834048)))]; + tensor original_model_text_encoder_transformer_10_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_10_pre_norm_ffn_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117931264)))]; + tensor linear_42_cast = linear(bias = original_model_text_encoder_transformer_10_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_10_pre_norm_ffn_1_weight_to_fp16, x = var_877_cast)[name = tensor("linear_42_cast")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast = gelu(mode = input_239_mode_0, x = linear_42_cast)[name = tensor("input_239_cast")]; + tensor original_model_text_encoder_transformer_10_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_10_pre_norm_ffn_4_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117935424)))]; + tensor original_model_text_encoder_transformer_10_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_10_pre_norm_ffn_4_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120032640)))]; + tensor linear_43_cast = linear(bias = original_model_text_encoder_transformer_10_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_10_pre_norm_ffn_4_weight_to_fp16, x = input_239_cast)[name = tensor("linear_43_cast")]; + tensor x_67_cast = add(x = x_65_cast, y = linear_43_cast)[name = tensor("x_67_cast")]; + tensor var_904_axes_0 = const()[name = tensor("op_904_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_11_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_11_pre_norm_mha_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120033728)))]; + tensor original_model_text_encoder_transformer_11_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_11_pre_norm_mha_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120034816)))]; + tensor var_904_cast = layer_norm(axes = var_904_axes_0, beta = original_model_text_encoder_transformer_11_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_11_pre_norm_mha_0_weight_to_fp16, x = x_67_cast)[name = tensor("op_904_cast")]; + tensor original_model_text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120035904)))]; + tensor original_model_text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121608832)))]; + tensor linear_44_cast = linear(bias = original_model_text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_904_cast)[name = tensor("linear_44_cast")]; + tensor var_916 = const()[name = tensor("op_916"), val = tensor([1, 77, 3, 8, -1])]; + tensor qkv_45_cast = reshape(shape = var_916, x = linear_44_cast)[name = tensor("qkv_45_cast")]; + tensor var_918_perm_0 = const()[name = tensor("op_918_perm_0"), val = tensor([0, 3, 2, 1, 4])]; + tensor query_45_begin_0 = const()[name = tensor("query_45_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor query_45_end_0 = const()[name = tensor("query_45_end_0"), val = tensor([1, 8, 1, 77, 64])]; + tensor query_45_end_mask_0 = const()[name = tensor("query_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor query_45_squeeze_mask_0 = const()[name = tensor("query_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor transpose_3 = transpose(perm = var_918_perm_0, x = qkv_45_cast)[name = tensor("transpose_3")]; + tensor query_45_cast = slice_by_index(begin = query_45_begin_0, end = query_45_end_0, end_mask = query_45_end_mask_0, squeeze_mask = query_45_squeeze_mask_0, x = transpose_3)[name = tensor("query_45_cast")]; + tensor key_45_begin_0 = const()[name = tensor("key_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor key_45_end_0 = const()[name = tensor("key_45_end_0"), val = tensor([1, 8, 2, 77, 64])]; + tensor key_45_end_mask_0 = const()[name = tensor("key_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor key_45_squeeze_mask_0 = const()[name = tensor("key_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor key_45_cast = slice_by_index(begin = key_45_begin_0, end = key_45_end_0, end_mask = key_45_end_mask_0, squeeze_mask = key_45_squeeze_mask_0, x = transpose_3)[name = tensor("key_45_cast")]; + tensor value_begin_0 = const()[name = tensor("value_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor value_end_0 = const()[name = tensor("value_end_0"), val = tensor([1, 8, 3, 77, 64])]; + tensor value_end_mask_0 = const()[name = tensor("value_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor value_squeeze_mask_0 = const()[name = tensor("value_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor value_cast = slice_by_index(begin = value_begin_0, end = value_end_0, end_mask = value_end_mask_0, squeeze_mask = value_squeeze_mask_0, x = transpose_3)[name = tensor("value_cast")]; + tensor var_929_to_fp16 = const()[name = tensor("op_929_to_fp16"), val = tensor(0x1p-3)]; + tensor query_cast = mul(x = query_45_cast, y = var_929_to_fp16)[name = tensor("query_cast")]; + tensor key_perm_0 = const()[name = tensor("key_perm_0"), val = tensor([0, 1, -1, -2])]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(false)]; + tensor transpose_2 = transpose(perm = key_perm_0, x = key_45_cast)[name = tensor("transpose_2")]; + tensor attn_45_cast = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = query_cast, y = transpose_2)[name = tensor("attn_45_cast")]; + tensor attn_as_float_cast = softmax(axis = var_19, x = attn_45_cast)[name = tensor("attn_as_float_cast")]; + tensor out_transpose_x_0 = const()[name = tensor("out_transpose_x_0"), val = tensor(false)]; + tensor out_transpose_y_0 = const()[name = tensor("out_transpose_y_0"), val = tensor(false)]; + tensor out_cast = matmul(transpose_x = out_transpose_x_0, transpose_y = out_transpose_y_0, x = attn_as_float_cast, y = value_cast)[name = tensor("out_cast")]; + tensor var_938_perm_0 = const()[name = tensor("op_938_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_939 = const()[name = tensor("op_939"), val = tensor([1, 77, -1])]; + tensor transpose_1 = transpose(perm = var_938_perm_0, x = out_cast)[name = tensor("transpose_1")]; + tensor input_251_cast = reshape(shape = var_939, x = transpose_1)[name = tensor("input_251_cast")]; + tensor original_model_text_encoder_transformer_11_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_11_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121611968)))]; + tensor original_model_text_encoder_transformer_11_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_11_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122136320)))]; + tensor linear_45_cast = linear(bias = original_model_text_encoder_transformer_11_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_11_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_251_cast)[name = tensor("linear_45_cast")]; + tensor x_71_cast = add(x = linear_45_cast, y = x_67_cast)[name = tensor("x_71_cast")]; + tensor var_953_axes_0 = const()[name = tensor("op_953_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_transformer_11_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_11_pre_norm_ffn_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122137408)))]; + tensor original_model_text_encoder_transformer_11_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_11_pre_norm_ffn_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122138496)))]; + tensor var_953_cast = layer_norm(axes = var_953_axes_0, beta = original_model_text_encoder_transformer_11_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_11_pre_norm_ffn_0_weight_to_fp16, x = x_71_cast)[name = tensor("op_953_cast")]; + tensor original_model_text_encoder_transformer_11_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_11_pre_norm_ffn_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(122139584)))]; + tensor original_model_text_encoder_transformer_11_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_11_pre_norm_ffn_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124236800)))]; + tensor linear_46_cast = linear(bias = original_model_text_encoder_transformer_11_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_11_pre_norm_ffn_1_weight_to_fp16, x = var_953_cast)[name = tensor("linear_46_cast")]; + tensor input_261_mode_0 = const()[name = tensor("input_261_mode_0"), val = tensor("EXACT")]; + tensor input_261_cast = gelu(mode = input_261_mode_0, x = linear_46_cast)[name = tensor("input_261_cast")]; + tensor original_model_text_encoder_transformer_11_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_11_pre_norm_ffn_4_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124240960)))]; + tensor original_model_text_encoder_transformer_11_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_transformer_11_pre_norm_ffn_4_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126338176)))]; + tensor linear_47_cast = linear(bias = original_model_text_encoder_transformer_11_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_11_pre_norm_ffn_4_weight_to_fp16, x = input_261_cast)[name = tensor("linear_47_cast")]; + tensor x_cast = add(x = x_71_cast, y = linear_47_cast)[name = tensor("x_cast")]; + tensor var_975_axes_0 = const()[name = tensor("op_975_axes_0"), val = tensor([-1])]; + tensor original_model_text_encoder_final_layer_norm_weight_to_fp16 = const()[name = tensor("original_model_text_encoder_final_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126339264)))]; + tensor original_model_text_encoder_final_layer_norm_bias_to_fp16 = const()[name = tensor("original_model_text_encoder_final_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126340352)))]; + tensor var_975_cast = layer_norm(axes = var_975_axes_0, beta = original_model_text_encoder_final_layer_norm_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_final_layer_norm_weight_to_fp16, x = x_cast)[name = tensor("op_975_cast")]; + tensor var_978 = const()[name = tensor("op_978"), val = tensor([0])]; + tensor var_979 = reduce_argmax(axis = var_19, keep_dims = var_20, x = input_tokens)[name = tensor("op_979")]; + 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_978, var_979))[name = tensor("stack_0")]; + tensor token_emb_transpose_batch_dims_0 = const()[name = tensor("token_emb_transpose_batch_dims_0"), val = tensor(0)]; + tensor token_emb_transpose_cast = gather_nd(batch_dims = token_emb_transpose_batch_dims_0, indices = stack_0, x = var_975_cast)[name = tensor("token_emb_transpose_cast")]; + tensor var_982_weight_0_to_fp16 = const()[name = tensor("op_982_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126341440)))]; + tensor var_982_bias_0_to_fp16 = const()[name = tensor("op_982_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126865792)))]; + tensor var_982_cast = linear(bias = var_982_bias_0_to_fp16, weight = var_982_weight_0_to_fp16, x = token_emb_transpose_cast)[name = tensor("op_982_cast")]; + tensor var_982_cast_to_fp32_dtype_0 = const()[name = tensor("op_982_cast_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor text_embeddings = cast(dtype = var_982_cast_to_fp32_dtype_0, x = var_982_cast)[name = tensor("cast_28")]; + } -> (text_embeddings); +} \ No newline at end of file