diff --git "a/whisper-medium/TextDecoder.mlmodelc/model.mil" "b/whisper-medium/TextDecoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/whisper-medium/TextDecoder.mlmodelc/model.mil" @@ -0,0 +1,3570 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3402.3.2"}, {"coremlc-version", "3402.4.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.2"}})] +{ + func main(tensor cache_length, tensor decoder_key_padding_mask, tensor encoder_output_embeds, tensor input_ids, tensor key_cache, tensor kv_cache_update_mask, tensor value_cache) { + tensor var_64_axis_0 = const()[name = tensor("op_64_axis_0"), val = tensor(0)]; + tensor var_64_batch_dims_0 = const()[name = tensor("op_64_batch_dims_0"), val = tensor(0)]; + tensor embed_tokens_weight_to_fp16 = const()[name = tensor("embed_tokens_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor var_64_cast_fp16 = gather(axis = var_64_axis_0, batch_dims = var_64_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_64_cast_fp16")]; + tensor var_68_axis_0 = const()[name = tensor("op_68_axis_0"), val = tensor(0)]; + tensor var_68_batch_dims_0 = const()[name = tensor("op_68_batch_dims_0"), val = tensor(0)]; + tensor embed_positions_weight_to_fp16 = const()[name = tensor("embed_positions_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106219648)))]; + tensor var_68_cast_fp16 = gather(axis = var_68_axis_0, batch_dims = var_68_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor("op_68_cast_fp16")]; + tensor hidden_states_1_cast_fp16 = add(x = var_64_cast_fp16, y = var_68_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; + tensor var_82_axes_0 = const()[name = tensor("op_82_axes_0"), val = tensor([2])]; + tensor var_82_cast_fp16 = expand_dims(axes = var_82_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_82_cast_fp16")]; + tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([3])]; + tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_82_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024])]; + tensor var_87_axis_0 = const()[name = tensor("op_87_axis_0"), val = tensor(1)]; + tensor var_87_cast_fp16_0, tensor var_87_cast_fp16_1, tensor var_87_cast_fp16_2, tensor var_87_cast_fp16_3, tensor var_87_cast_fp16_4, tensor var_87_cast_fp16_5, tensor var_87_cast_fp16_6, tensor var_87_cast_fp16_7, tensor var_87_cast_fp16_8, tensor var_87_cast_fp16_9, tensor var_87_cast_fp16_10, tensor var_87_cast_fp16_11, tensor var_87_cast_fp16_12, tensor var_87_cast_fp16_13, tensor var_87_cast_fp16_14, tensor var_87_cast_fp16_15, tensor var_87_cast_fp16_16, tensor var_87_cast_fp16_17, tensor var_87_cast_fp16_18, tensor var_87_cast_fp16_19, tensor var_87_cast_fp16_20, tensor var_87_cast_fp16_21, tensor var_87_cast_fp16_22, tensor var_87_cast_fp16_23 = split(axis = var_87_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_87_cast_fp16")]; + tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024])]; + tensor var_114_axis_0 = const()[name = tensor("op_114_axis_0"), val = tensor(1)]; + tensor var_114_cast_fp16_0, tensor var_114_cast_fp16_1, tensor var_114_cast_fp16_2, tensor var_114_cast_fp16_3, tensor var_114_cast_fp16_4, tensor var_114_cast_fp16_5, tensor var_114_cast_fp16_6, tensor var_114_cast_fp16_7, tensor var_114_cast_fp16_8, tensor var_114_cast_fp16_9, tensor var_114_cast_fp16_10, tensor var_114_cast_fp16_11, tensor var_114_cast_fp16_12, tensor var_114_cast_fp16_13, tensor var_114_cast_fp16_14, tensor var_114_cast_fp16_15, tensor var_114_cast_fp16_16, tensor var_114_cast_fp16_17, tensor var_114_cast_fp16_18, tensor var_114_cast_fp16_19, tensor var_114_cast_fp16_20, tensor var_114_cast_fp16_21, tensor var_114_cast_fp16_22, tensor var_114_cast_fp16_23 = split(axis = var_114_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_114_cast_fp16")]; + tensor var_144 = const()[name = tensor("op_144"), val = tensor(3)]; + tensor out_1_axes_0 = const()[name = tensor("out_1_axes_0"), val = tensor([1])]; + tensor var_169_to_fp16 = const()[name = tensor("op_169_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_169_to_fp16, x = inputs_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; + tensor obj_1_mean_0_to_fp16 = const()[name = tensor("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107137216)))]; + tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107139328)))]; + tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107141440)))]; + tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107143552)))]; + tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("obj_1_cast_fp16")]; + tensor query_1_pad_type_0 = const()[name = tensor("query_1_pad_type_0"), val = tensor("valid")]; + tensor query_1_strides_0 = const()[name = tensor("query_1_strides_0"), val = tensor([1, 1])]; + tensor query_1_pad_0 = const()[name = tensor("query_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_1_dilations_0 = const()[name = tensor("query_1_dilations_0"), val = tensor([1, 1])]; + tensor query_1_groups_0 = const()[name = tensor("query_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107145664)))]; + tensor layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109242880)))]; + tensor query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; + tensor current_key_1_pad_type_0 = const()[name = tensor("current_key_1_pad_type_0"), val = tensor("valid")]; + tensor current_key_1_strides_0 = const()[name = tensor("current_key_1_strides_0"), val = tensor([1, 1])]; + tensor current_key_1_pad_0 = const()[name = tensor("current_key_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_1_dilations_0 = const()[name = tensor("current_key_1_dilations_0"), val = tensor([1, 1])]; + tensor current_key_1_groups_0 = const()[name = tensor("current_key_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109244992)))]; + tensor current_key_1_cast_fp16 = conv(dilations = current_key_1_dilations_0, groups = current_key_1_groups_0, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = current_key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; + tensor current_value_1_pad_type_0 = const()[name = tensor("current_value_1_pad_type_0"), val = tensor("valid")]; + tensor current_value_1_strides_0 = const()[name = tensor("current_value_1_strides_0"), val = tensor([1, 1])]; + tensor current_value_1_pad_0 = const()[name = tensor("current_value_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_1_dilations_0 = const()[name = tensor("current_value_1_dilations_0"), val = tensor([1, 1])]; + tensor current_value_1_groups_0 = const()[name = tensor("current_value_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(111342208)))]; + tensor layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113439424)))]; + tensor current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = current_value_1_dilations_0, groups = current_value_1_groups_0, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = current_value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; + tensor var_204_axes_0 = const()[name = tensor("op_204_axes_0"), val = tensor([1])]; + tensor var_204_cast_fp16 = expand_dims(axes = var_204_axes_0, x = kv_cache_update_mask)[name = tensor("op_204_cast_fp16")]; + tensor var_205_axes_0 = const()[name = tensor("op_205_axes_0"), val = tensor([2])]; + tensor var_205_cast_fp16 = expand_dims(axes = var_205_axes_0, x = var_204_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor var_207_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_207_cast_fp16")]; + tensor key_1_cast_fp16 = add(x = var_87_cast_fp16_0, y = var_207_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor var_209_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_209_cast_fp16")]; + tensor value_1_cast_fp16 = add(x = var_114_cast_fp16_0, y = var_209_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_1_cast_fp16 = reshape(shape = var_212, x = query_1_cast_fp16)[name = tensor("mh_q_1_cast_fp16")]; + tensor var_214_to_fp16 = const()[name = tensor("op_214_to_fp16"), val = tensor(0x1p-3)]; + tensor var_215_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_214_to_fp16)[name = tensor("op_215_cast_fp16")]; + tensor var_216 = const()[name = tensor("op_216"), val = tensor([1, 16, 64, -1])]; + tensor var_217_cast_fp16 = reshape(shape = var_216, x = key_1_cast_fp16)[name = tensor("op_217_cast_fp16")]; + tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; + tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; + tensor mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_215_cast_fp16, y = var_217_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor var_221_axes_0 = const()[name = tensor("op_221_axes_0"), val = tensor([1])]; + tensor var_221_cast_fp16 = expand_dims(axes = var_221_axes_0, x = decoder_key_padding_mask)[name = tensor("op_221_cast_fp16")]; + tensor var_222_axes_0 = const()[name = tensor("op_222_axes_0"), val = tensor([2])]; + tensor var_222_cast_fp16 = expand_dims(axes = var_222_axes_0, x = var_221_cast_fp16)[name = tensor("op_222_cast_fp16")]; + tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_225_cast_fp16 = softmax(axis = var_144, x = mh_w_3_cast_fp16)[name = tensor("op_225_cast_fp16")]; + tensor var_226 = const()[name = tensor("op_226"), val = tensor([1, 16, 64, -1])]; + tensor var_227_cast_fp16 = reshape(shape = var_226, x = value_1_cast_fp16)[name = tensor("op_227_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(true)]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_227_cast_fp16, y = var_225_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_230 = const()[name = tensor("op_230"), val = tensor([1, 1024, 1, -1])]; + tensor input_1_cast_fp16 = reshape(shape = var_230, x = attn_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor obj_7_pad_type_0 = const()[name = tensor("obj_7_pad_type_0"), val = tensor("valid")]; + tensor obj_7_strides_0 = const()[name = tensor("obj_7_strides_0"), val = tensor([1, 1])]; + tensor obj_7_pad_0 = const()[name = tensor("obj_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_7_dilations_0 = const()[name = tensor("obj_7_dilations_0"), val = tensor([1, 1])]; + tensor obj_7_groups_0 = const()[name = tensor("obj_7_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113441536)))]; + tensor layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115538752)))]; + tensor obj_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_7_dilations_0, groups = obj_7_groups_0, pad = obj_7_pad_0, pad_type = obj_7_pad_type_0, strides = obj_7_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("obj_7_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([1])]; + tensor var_252_to_fp16 = const()[name = tensor("op_252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_252_to_fp16, x = inputs_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; + tensor obj_9_gamma_0_to_fp16 = const()[name = tensor("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115540864)))]; + tensor obj_9_beta_0_to_fp16 = const()[name = tensor("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115542976)))]; + tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor("obj_9_cast_fp16")]; + tensor query_3_pad_type_0 = const()[name = tensor("query_3_pad_type_0"), val = tensor("valid")]; + tensor query_3_strides_0 = const()[name = tensor("query_3_strides_0"), val = tensor([1, 1])]; + tensor query_3_pad_0 = const()[name = tensor("query_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_3_dilations_0 = const()[name = tensor("query_3_dilations_0"), val = tensor([1, 1])]; + tensor query_3_groups_0 = const()[name = tensor("query_3_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(115545088)))]; + tensor layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117642304)))]; + tensor query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("query_3_cast_fp16")]; + tensor key_3_pad_type_0 = const()[name = tensor("key_3_pad_type_0"), val = tensor("valid")]; + tensor key_3_strides_0 = const()[name = tensor("key_3_strides_0"), val = tensor([1, 1])]; + tensor key_3_pad_0 = const()[name = tensor("key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_3_dilations_0 = const()[name = tensor("key_3_dilations_0"), val = tensor([1, 1])]; + tensor key_3_groups_0 = const()[name = tensor("key_3_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117644416)))]; + tensor key_3_cast_fp16 = conv(dilations = key_3_dilations_0, groups = key_3_groups_0, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = key_3_strides_0, weight = layers_0_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_3_cast_fp16")]; + tensor value_3_pad_type_0 = const()[name = tensor("value_3_pad_type_0"), val = tensor("valid")]; + tensor value_3_strides_0 = const()[name = tensor("value_3_strides_0"), val = tensor([1, 1])]; + tensor value_3_pad_0 = const()[name = tensor("value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_3_dilations_0 = const()[name = tensor("value_3_dilations_0"), val = tensor([1, 1])]; + tensor value_3_groups_0 = const()[name = tensor("value_3_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(119741632)))]; + tensor layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121838848)))]; + tensor value_3_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_bias_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_0_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_3_cast_fp16")]; + tensor var_287 = const()[name = tensor("op_287"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_3_cast_fp16 = reshape(shape = var_287, x = query_3_cast_fp16)[name = tensor("mh_q_3_cast_fp16")]; + tensor var_289_to_fp16 = const()[name = tensor("op_289_to_fp16"), val = tensor(0x1p-3)]; + tensor var_290_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_289_to_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_291 = const()[name = tensor("op_291"), val = tensor([1, 16, 64, -1])]; + tensor var_292_cast_fp16 = reshape(shape = var_291, x = key_3_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor mh_w_5_transpose_x_0 = const()[name = tensor("mh_w_5_transpose_x_0"), val = tensor(true)]; + tensor mh_w_5_transpose_y_0 = const()[name = tensor("mh_w_5_transpose_y_0"), val = tensor(false)]; + tensor mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_290_cast_fp16, y = var_292_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor obj_13_cast_fp16 = softmax(axis = var_144, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor var_296 = const()[name = tensor("op_296"), val = tensor([1, 16, 64, -1])]; + tensor var_297_cast_fp16 = reshape(shape = var_296, x = value_3_cast_fp16)[name = tensor("op_297_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(true)]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_297_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_300 = const()[name = tensor("op_300"), val = tensor([1, 1024, 1, -1])]; + tensor input_3_cast_fp16 = reshape(shape = var_300, x = attn_3_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor obj_11_pad_type_0 = const()[name = tensor("obj_11_pad_type_0"), val = tensor("valid")]; + tensor obj_11_strides_0 = const()[name = tensor("obj_11_strides_0"), val = tensor([1, 1])]; + tensor obj_11_pad_0 = const()[name = tensor("obj_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_11_dilations_0 = const()[name = tensor("obj_11_dilations_0"), val = tensor([1, 1])]; + tensor obj_11_groups_0 = const()[name = tensor("obj_11_groups_0"), val = tensor(1)]; + tensor layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121840960)))]; + tensor layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123938176)))]; + tensor obj_11_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("obj_11_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor out_5_axes_0 = const()[name = tensor("out_5_axes_0"), val = tensor([1])]; + tensor var_318_to_fp16 = const()[name = tensor("op_318_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_318_to_fp16, x = inputs_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; + tensor input_5_gamma_0_to_fp16 = const()[name = tensor("input_5_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123940288)))]; + tensor input_5_beta_0_to_fp16 = const()[name = tensor("input_5_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123942400)))]; + tensor input_5_epsilon_0_to_fp16 = const()[name = tensor("input_5_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor input_7_pad_type_0 = const()[name = tensor("input_7_pad_type_0"), val = tensor("valid")]; + tensor input_7_strides_0 = const()[name = tensor("input_7_strides_0"), val = tensor([1, 1])]; + tensor input_7_pad_0 = const()[name = tensor("input_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_7_dilations_0 = const()[name = tensor("input_7_dilations_0"), val = tensor([1, 1])]; + tensor input_7_groups_0 = const()[name = tensor("input_7_groups_0"), val = tensor(1)]; + tensor layers_0_fc1_weight_to_fp16 = const()[name = tensor("layers_0_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123944512)))]; + tensor layers_0_fc1_bias_to_fp16 = const()[name = tensor("layers_0_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132333184)))]; + tensor input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_mode_0 = const()[name = tensor("input_9_mode_0"), val = tensor("EXACT")]; + tensor input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor hidden_states_3_pad_type_0 = const()[name = tensor("hidden_states_3_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_3_strides_0 = const()[name = tensor("hidden_states_3_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_3_pad_0 = const()[name = tensor("hidden_states_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_3_dilations_0 = const()[name = tensor("hidden_states_3_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_3_groups_0 = const()[name = tensor("hidden_states_3_groups_0"), val = tensor(1)]; + tensor layers_0_fc2_weight_to_fp16 = const()[name = tensor("layers_0_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132341440)))]; + tensor layers_0_fc2_bias_to_fp16 = const()[name = tensor("layers_0_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140730112)))]; + tensor hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor var_353 = const()[name = tensor("op_353"), val = tensor(3)]; + tensor out_7_axes_0 = const()[name = tensor("out_7_axes_0"), val = tensor([1])]; + tensor var_378_to_fp16 = const()[name = tensor("op_378_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_378_to_fp16, x = inputs_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor obj_15_gamma_0_to_fp16 = const()[name = tensor("obj_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140732224)))]; + tensor obj_15_beta_0_to_fp16 = const()[name = tensor("obj_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140734336)))]; + tensor obj_15_epsilon_0_to_fp16 = const()[name = tensor("obj_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("obj_15_cast_fp16")]; + tensor query_5_pad_type_0 = const()[name = tensor("query_5_pad_type_0"), val = tensor("valid")]; + tensor query_5_strides_0 = const()[name = tensor("query_5_strides_0"), val = tensor([1, 1])]; + tensor query_5_pad_0 = const()[name = tensor("query_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_5_dilations_0 = const()[name = tensor("query_5_dilations_0"), val = tensor([1, 1])]; + tensor query_5_groups_0 = const()[name = tensor("query_5_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140736448)))]; + tensor layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142833664)))]; + tensor query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor current_key_3_pad_type_0 = const()[name = tensor("current_key_3_pad_type_0"), val = tensor("valid")]; + tensor current_key_3_strides_0 = const()[name = tensor("current_key_3_strides_0"), val = tensor([1, 1])]; + tensor current_key_3_pad_0 = const()[name = tensor("current_key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_3_dilations_0 = const()[name = tensor("current_key_3_dilations_0"), val = tensor([1, 1])]; + tensor current_key_3_groups_0 = const()[name = tensor("current_key_3_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142835776)))]; + tensor current_key_3_cast_fp16 = conv(dilations = current_key_3_dilations_0, groups = current_key_3_groups_0, pad = current_key_3_pad_0, pad_type = current_key_3_pad_type_0, strides = current_key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_key_3_cast_fp16")]; + tensor current_value_3_pad_type_0 = const()[name = tensor("current_value_3_pad_type_0"), val = tensor("valid")]; + tensor current_value_3_strides_0 = const()[name = tensor("current_value_3_strides_0"), val = tensor([1, 1])]; + tensor current_value_3_pad_0 = const()[name = tensor("current_value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_3_dilations_0 = const()[name = tensor("current_value_3_dilations_0"), val = tensor([1, 1])]; + tensor current_value_3_groups_0 = const()[name = tensor("current_value_3_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144932992)))]; + tensor layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147030208)))]; + tensor current_value_3_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = current_value_3_dilations_0, groups = current_value_3_groups_0, pad = current_value_3_pad_0, pad_type = current_value_3_pad_type_0, strides = current_value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; + tensor var_416_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_416_cast_fp16")]; + tensor key_5_cast_fp16 = add(x = var_87_cast_fp16_1, y = var_416_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor var_418_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_418_cast_fp16")]; + tensor value_5_cast_fp16 = add(x = var_114_cast_fp16_1, y = var_418_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_421 = const()[name = tensor("op_421"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_5_cast_fp16 = reshape(shape = var_421, x = query_5_cast_fp16)[name = tensor("mh_q_5_cast_fp16")]; + tensor var_423_to_fp16 = const()[name = tensor("op_423_to_fp16"), val = tensor(0x1p-3)]; + tensor var_424_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_423_to_fp16)[name = tensor("op_424_cast_fp16")]; + tensor var_425 = const()[name = tensor("op_425"), val = tensor([1, 16, 64, -1])]; + tensor var_426_cast_fp16 = reshape(shape = var_425, x = key_5_cast_fp16)[name = tensor("op_426_cast_fp16")]; + tensor mh_w_7_transpose_x_0 = const()[name = tensor("mh_w_7_transpose_x_0"), val = tensor(true)]; + tensor mh_w_7_transpose_y_0 = const()[name = tensor("mh_w_7_transpose_y_0"), val = tensor(false)]; + tensor mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_424_cast_fp16, y = var_426_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor var_434_cast_fp16 = softmax(axis = var_353, x = mh_w_9_cast_fp16)[name = tensor("op_434_cast_fp16")]; + tensor var_435 = const()[name = tensor("op_435"), val = tensor([1, 16, 64, -1])]; + tensor var_436_cast_fp16 = reshape(shape = var_435, x = value_5_cast_fp16)[name = tensor("op_436_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(true)]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_436_cast_fp16, y = var_434_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_439 = const()[name = tensor("op_439"), val = tensor([1, 1024, 1, -1])]; + tensor input_11_cast_fp16 = reshape(shape = var_439, x = attn_5_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor obj_21_pad_type_0 = const()[name = tensor("obj_21_pad_type_0"), val = tensor("valid")]; + tensor obj_21_strides_0 = const()[name = tensor("obj_21_strides_0"), val = tensor([1, 1])]; + tensor obj_21_pad_0 = const()[name = tensor("obj_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_21_dilations_0 = const()[name = tensor("obj_21_dilations_0"), val = tensor([1, 1])]; + tensor obj_21_groups_0 = const()[name = tensor("obj_21_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147032320)))]; + tensor layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149129536)))]; + tensor obj_21_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("obj_21_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([1])]; + tensor var_461_to_fp16 = const()[name = tensor("op_461_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_461_to_fp16, x = inputs_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; + tensor obj_23_gamma_0_to_fp16 = const()[name = tensor("obj_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149131648)))]; + tensor obj_23_beta_0_to_fp16 = const()[name = tensor("obj_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149133760)))]; + tensor obj_23_epsilon_0_to_fp16 = const()[name = tensor("obj_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_23_cast_fp16")]; + tensor query_7_pad_type_0 = const()[name = tensor("query_7_pad_type_0"), val = tensor("valid")]; + tensor query_7_strides_0 = const()[name = tensor("query_7_strides_0"), val = tensor([1, 1])]; + tensor query_7_pad_0 = const()[name = tensor("query_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_7_dilations_0 = const()[name = tensor("query_7_dilations_0"), val = tensor([1, 1])]; + tensor query_7_groups_0 = const()[name = tensor("query_7_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149135872)))]; + tensor layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151233088)))]; + tensor query_7_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = query_7_dilations_0, groups = query_7_groups_0, pad = query_7_pad_0, pad_type = query_7_pad_type_0, strides = query_7_strides_0, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor("query_7_cast_fp16")]; + tensor key_7_pad_type_0 = const()[name = tensor("key_7_pad_type_0"), val = tensor("valid")]; + tensor key_7_strides_0 = const()[name = tensor("key_7_strides_0"), val = tensor([1, 1])]; + tensor key_7_pad_0 = const()[name = tensor("key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_7_dilations_0 = const()[name = tensor("key_7_dilations_0"), val = tensor([1, 1])]; + tensor key_7_groups_0 = const()[name = tensor("key_7_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151235200)))]; + tensor key_7_cast_fp16 = conv(dilations = key_7_dilations_0, groups = key_7_groups_0, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = key_7_strides_0, weight = layers_1_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_7_cast_fp16")]; + tensor value_7_pad_type_0 = const()[name = tensor("value_7_pad_type_0"), val = tensor("valid")]; + tensor value_7_strides_0 = const()[name = tensor("value_7_strides_0"), val = tensor([1, 1])]; + tensor value_7_pad_0 = const()[name = tensor("value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_7_dilations_0 = const()[name = tensor("value_7_dilations_0"), val = tensor([1, 1])]; + tensor value_7_groups_0 = const()[name = tensor("value_7_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153332416)))]; + tensor layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155429632)))]; + tensor value_7_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_bias_to_fp16, dilations = value_7_dilations_0, groups = value_7_groups_0, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = value_7_strides_0, weight = layers_1_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_7_cast_fp16")]; + tensor var_496 = const()[name = tensor("op_496"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_7_cast_fp16 = reshape(shape = var_496, x = query_7_cast_fp16)[name = tensor("mh_q_7_cast_fp16")]; + tensor var_498_to_fp16 = const()[name = tensor("op_498_to_fp16"), val = tensor(0x1p-3)]; + tensor var_499_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_498_to_fp16)[name = tensor("op_499_cast_fp16")]; + tensor var_500 = const()[name = tensor("op_500"), val = tensor([1, 16, 64, -1])]; + tensor var_501_cast_fp16 = reshape(shape = var_500, x = key_7_cast_fp16)[name = tensor("op_501_cast_fp16")]; + tensor mh_w_11_transpose_x_0 = const()[name = tensor("mh_w_11_transpose_x_0"), val = tensor(true)]; + tensor mh_w_11_transpose_y_0 = const()[name = tensor("mh_w_11_transpose_y_0"), val = tensor(false)]; + tensor mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_499_cast_fp16, y = var_501_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor obj_27_cast_fp16 = softmax(axis = var_353, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor var_505 = const()[name = tensor("op_505"), val = tensor([1, 16, 64, -1])]; + tensor var_506_cast_fp16 = reshape(shape = var_505, x = value_7_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(true)]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_506_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_509 = const()[name = tensor("op_509"), val = tensor([1, 1024, 1, -1])]; + tensor input_13_cast_fp16 = reshape(shape = var_509, x = attn_7_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor obj_25_pad_type_0 = const()[name = tensor("obj_25_pad_type_0"), val = tensor("valid")]; + tensor obj_25_strides_0 = const()[name = tensor("obj_25_strides_0"), val = tensor([1, 1])]; + tensor obj_25_pad_0 = const()[name = tensor("obj_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_25_dilations_0 = const()[name = tensor("obj_25_dilations_0"), val = tensor([1, 1])]; + tensor obj_25_groups_0 = const()[name = tensor("obj_25_groups_0"), val = tensor(1)]; + tensor layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155431744)))]; + tensor layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157528960)))]; + tensor obj_25_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = obj_25_dilations_0, groups = obj_25_groups_0, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = obj_25_strides_0, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("obj_25_cast_fp16")]; + tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor out_11_axes_0 = const()[name = tensor("out_11_axes_0"), val = tensor([1])]; + tensor var_527_to_fp16 = const()[name = tensor("op_527_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_527_to_fp16, x = inputs_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; + tensor input_15_gamma_0_to_fp16 = const()[name = tensor("input_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157531072)))]; + tensor input_15_beta_0_to_fp16 = const()[name = tensor("input_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157533184)))]; + tensor input_15_epsilon_0_to_fp16 = const()[name = tensor("input_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_pad_type_0 = const()[name = tensor("input_17_pad_type_0"), val = tensor("valid")]; + tensor input_17_strides_0 = const()[name = tensor("input_17_strides_0"), val = tensor([1, 1])]; + tensor input_17_pad_0 = const()[name = tensor("input_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_17_dilations_0 = const()[name = tensor("input_17_dilations_0"), val = tensor([1, 1])]; + tensor input_17_groups_0 = const()[name = tensor("input_17_groups_0"), val = tensor(1)]; + tensor layers_1_fc1_weight_to_fp16 = const()[name = tensor("layers_1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157535296)))]; + tensor layers_1_fc1_bias_to_fp16 = const()[name = tensor("layers_1_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165923968)))]; + tensor input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_mode_0 = const()[name = tensor("input_19_mode_0"), val = tensor("EXACT")]; + tensor input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor hidden_states_5_pad_type_0 = const()[name = tensor("hidden_states_5_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_5_strides_0 = const()[name = tensor("hidden_states_5_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_5_pad_0 = const()[name = tensor("hidden_states_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_5_dilations_0 = const()[name = tensor("hidden_states_5_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_5_groups_0 = const()[name = tensor("hidden_states_5_groups_0"), val = tensor(1)]; + tensor layers_1_fc2_weight_to_fp16 = const()[name = tensor("layers_1_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165932224)))]; + tensor layers_1_fc2_bias_to_fp16 = const()[name = tensor("layers_1_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174320896)))]; + tensor hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("hidden_states_5_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor var_562 = const()[name = tensor("op_562"), val = tensor(3)]; + tensor out_13_axes_0 = const()[name = tensor("out_13_axes_0"), val = tensor([1])]; + tensor var_587_to_fp16 = const()[name = tensor("op_587_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_587_to_fp16, x = inputs_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; + tensor obj_29_gamma_0_to_fp16 = const()[name = tensor("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174323008)))]; + tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174325120)))]; + tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_29_cast_fp16")]; + tensor query_9_pad_type_0 = const()[name = tensor("query_9_pad_type_0"), val = tensor("valid")]; + tensor query_9_strides_0 = const()[name = tensor("query_9_strides_0"), val = tensor([1, 1])]; + tensor query_9_pad_0 = const()[name = tensor("query_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_9_dilations_0 = const()[name = tensor("query_9_dilations_0"), val = tensor([1, 1])]; + tensor query_9_groups_0 = const()[name = tensor("query_9_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(174327232)))]; + tensor layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176424448)))]; + tensor query_9_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_bias_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor current_key_5_pad_type_0 = const()[name = tensor("current_key_5_pad_type_0"), val = tensor("valid")]; + tensor current_key_5_strides_0 = const()[name = tensor("current_key_5_strides_0"), val = tensor([1, 1])]; + tensor current_key_5_pad_0 = const()[name = tensor("current_key_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_5_dilations_0 = const()[name = tensor("current_key_5_dilations_0"), val = tensor([1, 1])]; + tensor current_key_5_groups_0 = const()[name = tensor("current_key_5_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176426560)))]; + tensor current_key_5_cast_fp16 = conv(dilations = current_key_5_dilations_0, groups = current_key_5_groups_0, pad = current_key_5_pad_0, pad_type = current_key_5_pad_type_0, strides = current_key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; + tensor current_value_5_pad_type_0 = const()[name = tensor("current_value_5_pad_type_0"), val = tensor("valid")]; + tensor current_value_5_strides_0 = const()[name = tensor("current_value_5_strides_0"), val = tensor([1, 1])]; + tensor current_value_5_pad_0 = const()[name = tensor("current_value_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_5_dilations_0 = const()[name = tensor("current_value_5_dilations_0"), val = tensor([1, 1])]; + tensor current_value_5_groups_0 = const()[name = tensor("current_value_5_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178523776)))]; + tensor layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180620992)))]; + tensor current_value_5_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_bias_to_fp16, dilations = current_value_5_dilations_0, groups = current_value_5_groups_0, pad = current_value_5_pad_0, pad_type = current_value_5_pad_type_0, strides = current_value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; + tensor var_625_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor key_9_cast_fp16 = add(x = var_87_cast_fp16_2, y = var_625_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor var_627_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_627_cast_fp16")]; + tensor value_9_cast_fp16 = add(x = var_114_cast_fp16_2, y = var_627_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_630 = const()[name = tensor("op_630"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_9_cast_fp16 = reshape(shape = var_630, x = query_9_cast_fp16)[name = tensor("mh_q_9_cast_fp16")]; + tensor var_632_to_fp16 = const()[name = tensor("op_632_to_fp16"), val = tensor(0x1p-3)]; + tensor var_633_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_632_to_fp16)[name = tensor("op_633_cast_fp16")]; + tensor var_634 = const()[name = tensor("op_634"), val = tensor([1, 16, 64, -1])]; + tensor var_635_cast_fp16 = reshape(shape = var_634, x = key_9_cast_fp16)[name = tensor("op_635_cast_fp16")]; + tensor mh_w_13_transpose_x_0 = const()[name = tensor("mh_w_13_transpose_x_0"), val = tensor(true)]; + tensor mh_w_13_transpose_y_0 = const()[name = tensor("mh_w_13_transpose_y_0"), val = tensor(false)]; + tensor mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_633_cast_fp16, y = var_635_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_643_cast_fp16 = softmax(axis = var_562, x = mh_w_15_cast_fp16)[name = tensor("op_643_cast_fp16")]; + tensor var_644 = const()[name = tensor("op_644"), val = tensor([1, 16, 64, -1])]; + tensor var_645_cast_fp16 = reshape(shape = var_644, x = value_9_cast_fp16)[name = tensor("op_645_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(true)]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_645_cast_fp16, y = var_643_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_648 = const()[name = tensor("op_648"), val = tensor([1, 1024, 1, -1])]; + tensor input_21_cast_fp16 = reshape(shape = var_648, x = attn_9_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor obj_35_pad_type_0 = const()[name = tensor("obj_35_pad_type_0"), val = tensor("valid")]; + tensor obj_35_strides_0 = const()[name = tensor("obj_35_strides_0"), val = tensor([1, 1])]; + tensor obj_35_pad_0 = const()[name = tensor("obj_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_35_dilations_0 = const()[name = tensor("obj_35_dilations_0"), val = tensor([1, 1])]; + tensor obj_35_groups_0 = const()[name = tensor("obj_35_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180623104)))]; + tensor layers_2_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182720320)))]; + tensor obj_35_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_bias_to_fp16, dilations = obj_35_dilations_0, groups = obj_35_groups_0, pad = obj_35_pad_0, pad_type = obj_35_pad_type_0, strides = obj_35_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("obj_35_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([1])]; + tensor var_670_to_fp16 = const()[name = tensor("op_670_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_670_to_fp16, x = inputs_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; + tensor obj_37_gamma_0_to_fp16 = const()[name = tensor("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182722432)))]; + tensor obj_37_beta_0_to_fp16 = const()[name = tensor("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182724544)))]; + tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("obj_37_cast_fp16")]; + tensor query_11_pad_type_0 = const()[name = tensor("query_11_pad_type_0"), val = tensor("valid")]; + tensor query_11_strides_0 = const()[name = tensor("query_11_strides_0"), val = tensor([1, 1])]; + tensor query_11_pad_0 = const()[name = tensor("query_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_11_dilations_0 = const()[name = tensor("query_11_dilations_0"), val = tensor([1, 1])]; + tensor query_11_groups_0 = const()[name = tensor("query_11_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182726656)))]; + tensor layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184823872)))]; + tensor query_11_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_bias_to_fp16, dilations = query_11_dilations_0, groups = query_11_groups_0, pad = query_11_pad_0, pad_type = query_11_pad_type_0, strides = query_11_strides_0, weight = layers_2_encoder_attn_q_proj_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("query_11_cast_fp16")]; + tensor key_11_pad_type_0 = const()[name = tensor("key_11_pad_type_0"), val = tensor("valid")]; + tensor key_11_strides_0 = const()[name = tensor("key_11_strides_0"), val = tensor([1, 1])]; + tensor key_11_pad_0 = const()[name = tensor("key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_11_dilations_0 = const()[name = tensor("key_11_dilations_0"), val = tensor([1, 1])]; + tensor key_11_groups_0 = const()[name = tensor("key_11_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184825984)))]; + tensor key_11_cast_fp16 = conv(dilations = key_11_dilations_0, groups = key_11_groups_0, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = key_11_strides_0, weight = layers_2_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_11_cast_fp16")]; + tensor value_11_pad_type_0 = const()[name = tensor("value_11_pad_type_0"), val = tensor("valid")]; + tensor value_11_strides_0 = const()[name = tensor("value_11_strides_0"), val = tensor([1, 1])]; + tensor value_11_pad_0 = const()[name = tensor("value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_11_dilations_0 = const()[name = tensor("value_11_dilations_0"), val = tensor([1, 1])]; + tensor value_11_groups_0 = const()[name = tensor("value_11_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186923200)))]; + tensor layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189020416)))]; + tensor value_11_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_bias_to_fp16, dilations = value_11_dilations_0, groups = value_11_groups_0, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = value_11_strides_0, weight = layers_2_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_11_cast_fp16")]; + tensor var_705 = const()[name = tensor("op_705"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_11_cast_fp16 = reshape(shape = var_705, x = query_11_cast_fp16)[name = tensor("mh_q_11_cast_fp16")]; + tensor var_707_to_fp16 = const()[name = tensor("op_707_to_fp16"), val = tensor(0x1p-3)]; + tensor var_708_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_707_to_fp16)[name = tensor("op_708_cast_fp16")]; + tensor var_709 = const()[name = tensor("op_709"), val = tensor([1, 16, 64, -1])]; + tensor var_710_cast_fp16 = reshape(shape = var_709, x = key_11_cast_fp16)[name = tensor("op_710_cast_fp16")]; + tensor mh_w_17_transpose_x_0 = const()[name = tensor("mh_w_17_transpose_x_0"), val = tensor(true)]; + tensor mh_w_17_transpose_y_0 = const()[name = tensor("mh_w_17_transpose_y_0"), val = tensor(false)]; + tensor mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_708_cast_fp16, y = var_710_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor obj_41_cast_fp16 = softmax(axis = var_562, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor var_714 = const()[name = tensor("op_714"), val = tensor([1, 16, 64, -1])]; + tensor var_715_cast_fp16 = reshape(shape = var_714, x = value_11_cast_fp16)[name = tensor("op_715_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(true)]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_715_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_718 = const()[name = tensor("op_718"), val = tensor([1, 1024, 1, -1])]; + tensor input_23_cast_fp16 = reshape(shape = var_718, x = attn_11_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor obj_39_pad_type_0 = const()[name = tensor("obj_39_pad_type_0"), val = tensor("valid")]; + tensor obj_39_strides_0 = const()[name = tensor("obj_39_strides_0"), val = tensor([1, 1])]; + tensor obj_39_pad_0 = const()[name = tensor("obj_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_39_dilations_0 = const()[name = tensor("obj_39_dilations_0"), val = tensor([1, 1])]; + tensor obj_39_groups_0 = const()[name = tensor("obj_39_groups_0"), val = tensor(1)]; + tensor layers_2_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189022528)))]; + tensor layers_2_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191119744)))]; + tensor obj_39_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_2_encoder_attn_o_proj_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("obj_39_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor out_17_axes_0 = const()[name = tensor("out_17_axes_0"), val = tensor([1])]; + tensor var_736_to_fp16 = const()[name = tensor("op_736_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_736_to_fp16, x = inputs_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; + tensor input_25_gamma_0_to_fp16 = const()[name = tensor("input_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191121856)))]; + tensor input_25_beta_0_to_fp16 = const()[name = tensor("input_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191123968)))]; + tensor input_25_epsilon_0_to_fp16 = const()[name = tensor("input_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_pad_type_0 = const()[name = tensor("input_27_pad_type_0"), val = tensor("valid")]; + tensor input_27_strides_0 = const()[name = tensor("input_27_strides_0"), val = tensor([1, 1])]; + tensor input_27_pad_0 = const()[name = tensor("input_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_27_dilations_0 = const()[name = tensor("input_27_dilations_0"), val = tensor([1, 1])]; + tensor input_27_groups_0 = const()[name = tensor("input_27_groups_0"), val = tensor(1)]; + tensor layers_2_fc1_weight_to_fp16 = const()[name = tensor("layers_2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(191126080)))]; + tensor layers_2_fc1_bias_to_fp16 = const()[name = tensor("layers_2_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199514752)))]; + tensor input_27_cast_fp16 = conv(bias = layers_2_fc1_bias_to_fp16, dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = layers_2_fc1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_mode_0 = const()[name = tensor("input_29_mode_0"), val = tensor("EXACT")]; + tensor input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor hidden_states_7_pad_type_0 = const()[name = tensor("hidden_states_7_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_7_strides_0 = const()[name = tensor("hidden_states_7_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_7_pad_0 = const()[name = tensor("hidden_states_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_7_dilations_0 = const()[name = tensor("hidden_states_7_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_7_groups_0 = const()[name = tensor("hidden_states_7_groups_0"), val = tensor(1)]; + tensor layers_2_fc2_weight_to_fp16 = const()[name = tensor("layers_2_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199523008)))]; + tensor layers_2_fc2_bias_to_fp16 = const()[name = tensor("layers_2_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207911680)))]; + tensor hidden_states_7_cast_fp16 = conv(bias = layers_2_fc2_bias_to_fp16, dilations = hidden_states_7_dilations_0, groups = hidden_states_7_groups_0, pad = hidden_states_7_pad_0, pad_type = hidden_states_7_pad_type_0, strides = hidden_states_7_strides_0, weight = layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor(3)]; + tensor out_19_axes_0 = const()[name = tensor("out_19_axes_0"), val = tensor([1])]; + tensor var_796_to_fp16 = const()[name = tensor("op_796_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_796_to_fp16, x = inputs_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; + tensor obj_43_gamma_0_to_fp16 = const()[name = tensor("obj_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207913792)))]; + tensor obj_43_beta_0_to_fp16 = const()[name = tensor("obj_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207915904)))]; + tensor obj_43_epsilon_0_to_fp16 = const()[name = tensor("obj_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("obj_43_cast_fp16")]; + tensor query_13_pad_type_0 = const()[name = tensor("query_13_pad_type_0"), val = tensor("valid")]; + tensor query_13_strides_0 = const()[name = tensor("query_13_strides_0"), val = tensor([1, 1])]; + tensor query_13_pad_0 = const()[name = tensor("query_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_13_dilations_0 = const()[name = tensor("query_13_dilations_0"), val = tensor([1, 1])]; + tensor query_13_groups_0 = const()[name = tensor("query_13_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207918016)))]; + tensor layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210015232)))]; + tensor query_13_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_bias_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor current_key_7_pad_type_0 = const()[name = tensor("current_key_7_pad_type_0"), val = tensor("valid")]; + tensor current_key_7_strides_0 = const()[name = tensor("current_key_7_strides_0"), val = tensor([1, 1])]; + tensor current_key_7_pad_0 = const()[name = tensor("current_key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_7_dilations_0 = const()[name = tensor("current_key_7_dilations_0"), val = tensor([1, 1])]; + tensor current_key_7_groups_0 = const()[name = tensor("current_key_7_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210017344)))]; + tensor current_key_7_cast_fp16 = conv(dilations = current_key_7_dilations_0, groups = current_key_7_groups_0, pad = current_key_7_pad_0, pad_type = current_key_7_pad_type_0, strides = current_key_7_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_key_7_cast_fp16")]; + tensor current_value_7_pad_type_0 = const()[name = tensor("current_value_7_pad_type_0"), val = tensor("valid")]; + tensor current_value_7_strides_0 = const()[name = tensor("current_value_7_strides_0"), val = tensor([1, 1])]; + tensor current_value_7_pad_0 = const()[name = tensor("current_value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_7_dilations_0 = const()[name = tensor("current_value_7_dilations_0"), val = tensor([1, 1])]; + tensor current_value_7_groups_0 = const()[name = tensor("current_value_7_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(212114560)))]; + tensor layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214211776)))]; + tensor current_value_7_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_bias_to_fp16, dilations = current_value_7_dilations_0, groups = current_value_7_groups_0, pad = current_value_7_pad_0, pad_type = current_value_7_pad_type_0, strides = current_value_7_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_value_7_cast_fp16")]; + tensor var_834_cast_fp16 = mul(x = current_key_7_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_834_cast_fp16")]; + tensor key_13_cast_fp16 = add(x = var_87_cast_fp16_3, y = var_834_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor var_836_cast_fp16 = mul(x = current_value_7_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_836_cast_fp16")]; + tensor value_13_cast_fp16 = add(x = var_114_cast_fp16_3, y = var_836_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_839 = const()[name = tensor("op_839"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_13_cast_fp16 = reshape(shape = var_839, x = query_13_cast_fp16)[name = tensor("mh_q_13_cast_fp16")]; + tensor var_841_to_fp16 = const()[name = tensor("op_841_to_fp16"), val = tensor(0x1p-3)]; + tensor var_842_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_841_to_fp16)[name = tensor("op_842_cast_fp16")]; + tensor var_843 = const()[name = tensor("op_843"), val = tensor([1, 16, 64, -1])]; + tensor var_844_cast_fp16 = reshape(shape = var_843, x = key_13_cast_fp16)[name = tensor("op_844_cast_fp16")]; + tensor mh_w_19_transpose_x_0 = const()[name = tensor("mh_w_19_transpose_x_0"), val = tensor(true)]; + tensor mh_w_19_transpose_y_0 = const()[name = tensor("mh_w_19_transpose_y_0"), val = tensor(false)]; + tensor mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_842_cast_fp16, y = var_844_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor var_852_cast_fp16 = softmax(axis = var_771, x = mh_w_21_cast_fp16)[name = tensor("op_852_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 16, 64, -1])]; + tensor var_854_cast_fp16 = reshape(shape = var_853, x = value_13_cast_fp16)[name = tensor("op_854_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(true)]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_854_cast_fp16, y = var_852_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_857 = const()[name = tensor("op_857"), val = tensor([1, 1024, 1, -1])]; + tensor input_31_cast_fp16 = reshape(shape = var_857, x = attn_13_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor obj_49_pad_type_0 = const()[name = tensor("obj_49_pad_type_0"), val = tensor("valid")]; + tensor obj_49_strides_0 = const()[name = tensor("obj_49_strides_0"), val = tensor([1, 1])]; + tensor obj_49_pad_0 = const()[name = tensor("obj_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_49_dilations_0 = const()[name = tensor("obj_49_dilations_0"), val = tensor([1, 1])]; + tensor obj_49_groups_0 = const()[name = tensor("obj_49_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214213888)))]; + tensor layers_3_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216311104)))]; + tensor obj_49_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_bias_to_fp16, dilations = obj_49_dilations_0, groups = obj_49_groups_0, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = obj_49_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("obj_49_cast_fp16")]; + tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([1])]; + tensor var_879_to_fp16 = const()[name = tensor("op_879_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_879_to_fp16, x = inputs_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor obj_51_gamma_0_to_fp16 = const()[name = tensor("obj_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216313216)))]; + tensor obj_51_beta_0_to_fp16 = const()[name = tensor("obj_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216315328)))]; + tensor obj_51_epsilon_0_to_fp16 = const()[name = tensor("obj_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_51_cast_fp16")]; + tensor query_15_pad_type_0 = const()[name = tensor("query_15_pad_type_0"), val = tensor("valid")]; + tensor query_15_strides_0 = const()[name = tensor("query_15_strides_0"), val = tensor([1, 1])]; + tensor query_15_pad_0 = const()[name = tensor("query_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_15_dilations_0 = const()[name = tensor("query_15_dilations_0"), val = tensor([1, 1])]; + tensor query_15_groups_0 = const()[name = tensor("query_15_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(216317440)))]; + tensor layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218414656)))]; + tensor query_15_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_bias_to_fp16, dilations = query_15_dilations_0, groups = query_15_groups_0, pad = query_15_pad_0, pad_type = query_15_pad_type_0, strides = query_15_strides_0, weight = layers_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor("query_15_cast_fp16")]; + tensor key_15_pad_type_0 = const()[name = tensor("key_15_pad_type_0"), val = tensor("valid")]; + tensor key_15_strides_0 = const()[name = tensor("key_15_strides_0"), val = tensor([1, 1])]; + tensor key_15_pad_0 = const()[name = tensor("key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_15_dilations_0 = const()[name = tensor("key_15_dilations_0"), val = tensor([1, 1])]; + tensor key_15_groups_0 = const()[name = tensor("key_15_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218416768)))]; + tensor key_15_cast_fp16 = conv(dilations = key_15_dilations_0, groups = key_15_groups_0, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = key_15_strides_0, weight = layers_3_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_15_cast_fp16")]; + tensor value_15_pad_type_0 = const()[name = tensor("value_15_pad_type_0"), val = tensor("valid")]; + tensor value_15_strides_0 = const()[name = tensor("value_15_strides_0"), val = tensor([1, 1])]; + tensor value_15_pad_0 = const()[name = tensor("value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_15_dilations_0 = const()[name = tensor("value_15_dilations_0"), val = tensor([1, 1])]; + tensor value_15_groups_0 = const()[name = tensor("value_15_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220513984)))]; + tensor layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222611200)))]; + tensor value_15_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_bias_to_fp16, dilations = value_15_dilations_0, groups = value_15_groups_0, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = value_15_strides_0, weight = layers_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_15_cast_fp16")]; + tensor var_914 = const()[name = tensor("op_914"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_15_cast_fp16 = reshape(shape = var_914, x = query_15_cast_fp16)[name = tensor("mh_q_15_cast_fp16")]; + tensor var_916_to_fp16 = const()[name = tensor("op_916_to_fp16"), val = tensor(0x1p-3)]; + tensor var_917_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_916_to_fp16)[name = tensor("op_917_cast_fp16")]; + tensor var_918 = const()[name = tensor("op_918"), val = tensor([1, 16, 64, -1])]; + tensor var_919_cast_fp16 = reshape(shape = var_918, x = key_15_cast_fp16)[name = tensor("op_919_cast_fp16")]; + tensor mh_w_23_transpose_x_0 = const()[name = tensor("mh_w_23_transpose_x_0"), val = tensor(true)]; + tensor mh_w_23_transpose_y_0 = const()[name = tensor("mh_w_23_transpose_y_0"), val = tensor(false)]; + tensor mh_w_23_cast_fp16 = matmul(transpose_x = mh_w_23_transpose_x_0, transpose_y = mh_w_23_transpose_y_0, x = var_917_cast_fp16, y = var_919_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor obj_55_cast_fp16 = softmax(axis = var_771, x = mh_w_23_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor([1, 16, 64, -1])]; + tensor var_924_cast_fp16 = reshape(shape = var_923, x = value_15_cast_fp16)[name = tensor("op_924_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(true)]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_924_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_927 = const()[name = tensor("op_927"), val = tensor([1, 1024, 1, -1])]; + tensor input_33_cast_fp16 = reshape(shape = var_927, x = attn_15_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor obj_53_pad_type_0 = const()[name = tensor("obj_53_pad_type_0"), val = tensor("valid")]; + tensor obj_53_strides_0 = const()[name = tensor("obj_53_strides_0"), val = tensor([1, 1])]; + tensor obj_53_pad_0 = const()[name = tensor("obj_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_53_dilations_0 = const()[name = tensor("obj_53_dilations_0"), val = tensor([1, 1])]; + tensor obj_53_groups_0 = const()[name = tensor("obj_53_groups_0"), val = tensor(1)]; + tensor layers_3_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222613312)))]; + tensor layers_3_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224710528)))]; + tensor obj_53_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_bias_to_fp16, dilations = obj_53_dilations_0, groups = obj_53_groups_0, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = obj_53_strides_0, weight = layers_3_encoder_attn_o_proj_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("obj_53_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor out_23_axes_0 = const()[name = tensor("out_23_axes_0"), val = tensor([1])]; + tensor var_945_to_fp16 = const()[name = tensor("op_945_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_945_to_fp16, x = inputs_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; + tensor input_35_gamma_0_to_fp16 = const()[name = tensor("input_35_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224712640)))]; + tensor input_35_beta_0_to_fp16 = const()[name = tensor("input_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224714752)))]; + tensor input_35_epsilon_0_to_fp16 = const()[name = tensor("input_35_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor input_37_pad_type_0 = const()[name = tensor("input_37_pad_type_0"), val = tensor("valid")]; + tensor input_37_strides_0 = const()[name = tensor("input_37_strides_0"), val = tensor([1, 1])]; + tensor input_37_pad_0 = const()[name = tensor("input_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_37_dilations_0 = const()[name = tensor("input_37_dilations_0"), val = tensor([1, 1])]; + tensor input_37_groups_0 = const()[name = tensor("input_37_groups_0"), val = tensor(1)]; + tensor layers_3_fc1_weight_to_fp16 = const()[name = tensor("layers_3_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224716864)))]; + tensor layers_3_fc1_bias_to_fp16 = const()[name = tensor("layers_3_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233105536)))]; + tensor input_37_cast_fp16 = conv(bias = layers_3_fc1_bias_to_fp16, dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_3_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_mode_0 = const()[name = tensor("input_39_mode_0"), val = tensor("EXACT")]; + tensor input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor hidden_states_9_pad_type_0 = const()[name = tensor("hidden_states_9_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_9_strides_0 = const()[name = tensor("hidden_states_9_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_9_pad_0 = const()[name = tensor("hidden_states_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_9_dilations_0 = const()[name = tensor("hidden_states_9_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_9_groups_0 = const()[name = tensor("hidden_states_9_groups_0"), val = tensor(1)]; + tensor layers_3_fc2_weight_to_fp16 = const()[name = tensor("layers_3_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233113792)))]; + tensor layers_3_fc2_bias_to_fp16 = const()[name = tensor("layers_3_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241502464)))]; + tensor hidden_states_9_cast_fp16 = conv(bias = layers_3_fc2_bias_to_fp16, dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = layers_3_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("hidden_states_9_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor var_980 = const()[name = tensor("op_980"), val = tensor(3)]; + tensor out_25_axes_0 = const()[name = tensor("out_25_axes_0"), val = tensor([1])]; + tensor var_1005_to_fp16 = const()[name = tensor("op_1005_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_1005_to_fp16, x = inputs_25_cast_fp16)[name = tensor("out_25_cast_fp16")]; + tensor obj_57_gamma_0_to_fp16 = const()[name = tensor("obj_57_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241504576)))]; + tensor obj_57_beta_0_to_fp16 = const()[name = tensor("obj_57_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241506688)))]; + tensor obj_57_epsilon_0_to_fp16 = const()[name = tensor("obj_57_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_57_cast_fp16 = batch_norm(beta = obj_57_beta_0_to_fp16, epsilon = obj_57_epsilon_0_to_fp16, gamma = obj_57_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor("obj_57_cast_fp16")]; + tensor query_17_pad_type_0 = const()[name = tensor("query_17_pad_type_0"), val = tensor("valid")]; + tensor query_17_strides_0 = const()[name = tensor("query_17_strides_0"), val = tensor([1, 1])]; + tensor query_17_pad_0 = const()[name = tensor("query_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_17_dilations_0 = const()[name = tensor("query_17_dilations_0"), val = tensor([1, 1])]; + tensor query_17_groups_0 = const()[name = tensor("query_17_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(241508800)))]; + tensor layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243606016)))]; + tensor query_17_cast_fp16 = conv(bias = layers_4_self_attn_q_proj_bias_to_fp16, dilations = query_17_dilations_0, groups = query_17_groups_0, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = query_17_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("query_17_cast_fp16")]; + tensor current_key_9_pad_type_0 = const()[name = tensor("current_key_9_pad_type_0"), val = tensor("valid")]; + tensor current_key_9_strides_0 = const()[name = tensor("current_key_9_strides_0"), val = tensor([1, 1])]; + tensor current_key_9_pad_0 = const()[name = tensor("current_key_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_9_dilations_0 = const()[name = tensor("current_key_9_dilations_0"), val = tensor([1, 1])]; + tensor current_key_9_groups_0 = const()[name = tensor("current_key_9_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243608128)))]; + tensor current_key_9_cast_fp16 = conv(dilations = current_key_9_dilations_0, groups = current_key_9_groups_0, pad = current_key_9_pad_0, pad_type = current_key_9_pad_type_0, strides = current_key_9_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("current_key_9_cast_fp16")]; + tensor current_value_9_pad_type_0 = const()[name = tensor("current_value_9_pad_type_0"), val = tensor("valid")]; + tensor current_value_9_strides_0 = const()[name = tensor("current_value_9_strides_0"), val = tensor([1, 1])]; + tensor current_value_9_pad_0 = const()[name = tensor("current_value_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_9_dilations_0 = const()[name = tensor("current_value_9_dilations_0"), val = tensor([1, 1])]; + tensor current_value_9_groups_0 = const()[name = tensor("current_value_9_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(245705344)))]; + tensor layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247802560)))]; + tensor current_value_9_cast_fp16 = conv(bias = layers_4_self_attn_v_proj_bias_to_fp16, dilations = current_value_9_dilations_0, groups = current_value_9_groups_0, pad = current_value_9_pad_0, pad_type = current_value_9_pad_type_0, strides = current_value_9_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = obj_57_cast_fp16)[name = tensor("current_value_9_cast_fp16")]; + tensor var_1043_cast_fp16 = mul(x = current_key_9_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1043_cast_fp16")]; + tensor key_17_cast_fp16 = add(x = var_87_cast_fp16_4, y = var_1043_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor var_1045_cast_fp16 = mul(x = current_value_9_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1045_cast_fp16")]; + tensor value_17_cast_fp16 = add(x = var_114_cast_fp16_4, y = var_1045_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_1048 = const()[name = tensor("op_1048"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_17_cast_fp16 = reshape(shape = var_1048, x = query_17_cast_fp16)[name = tensor("mh_q_17_cast_fp16")]; + tensor var_1050_to_fp16 = const()[name = tensor("op_1050_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1051_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_1050_to_fp16)[name = tensor("op_1051_cast_fp16")]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor([1, 16, 64, -1])]; + tensor var_1053_cast_fp16 = reshape(shape = var_1052, x = key_17_cast_fp16)[name = tensor("op_1053_cast_fp16")]; + tensor mh_w_25_transpose_x_0 = const()[name = tensor("mh_w_25_transpose_x_0"), val = tensor(true)]; + tensor mh_w_25_transpose_y_0 = const()[name = tensor("mh_w_25_transpose_y_0"), val = tensor(false)]; + tensor mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1051_cast_fp16, y = var_1053_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_1061_cast_fp16 = softmax(axis = var_980, x = mh_w_27_cast_fp16)[name = tensor("op_1061_cast_fp16")]; + tensor var_1062 = const()[name = tensor("op_1062"), val = tensor([1, 16, 64, -1])]; + tensor var_1063_cast_fp16 = reshape(shape = var_1062, x = value_17_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(true)]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_1063_cast_fp16, y = var_1061_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_1066 = const()[name = tensor("op_1066"), val = tensor([1, 1024, 1, -1])]; + tensor input_41_cast_fp16 = reshape(shape = var_1066, x = attn_17_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor obj_63_pad_type_0 = const()[name = tensor("obj_63_pad_type_0"), val = tensor("valid")]; + tensor obj_63_strides_0 = const()[name = tensor("obj_63_strides_0"), val = tensor([1, 1])]; + tensor obj_63_pad_0 = const()[name = tensor("obj_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_63_dilations_0 = const()[name = tensor("obj_63_dilations_0"), val = tensor([1, 1])]; + tensor obj_63_groups_0 = const()[name = tensor("obj_63_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247804672)))]; + tensor layers_4_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249901888)))]; + tensor obj_63_cast_fp16 = conv(bias = layers_4_self_attn_o_proj_bias_to_fp16, dilations = obj_63_dilations_0, groups = obj_63_groups_0, pad = obj_63_pad_0, pad_type = obj_63_pad_type_0, strides = obj_63_strides_0, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("obj_63_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = obj_63_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([1])]; + tensor var_1088_to_fp16 = const()[name = tensor("op_1088_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_1088_to_fp16, x = inputs_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; + tensor obj_65_gamma_0_to_fp16 = const()[name = tensor("obj_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249904000)))]; + tensor obj_65_beta_0_to_fp16 = const()[name = tensor("obj_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249906112)))]; + tensor obj_65_epsilon_0_to_fp16 = const()[name = tensor("obj_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_65_cast_fp16 = batch_norm(beta = obj_65_beta_0_to_fp16, epsilon = obj_65_epsilon_0_to_fp16, gamma = obj_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor("obj_65_cast_fp16")]; + tensor query_19_pad_type_0 = const()[name = tensor("query_19_pad_type_0"), val = tensor("valid")]; + tensor query_19_strides_0 = const()[name = tensor("query_19_strides_0"), val = tensor([1, 1])]; + tensor query_19_pad_0 = const()[name = tensor("query_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_19_dilations_0 = const()[name = tensor("query_19_dilations_0"), val = tensor([1, 1])]; + tensor query_19_groups_0 = const()[name = tensor("query_19_groups_0"), val = tensor(1)]; + tensor layers_4_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249908224)))]; + tensor layers_4_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252005440)))]; + tensor query_19_cast_fp16 = conv(bias = layers_4_encoder_attn_q_proj_bias_to_fp16, dilations = query_19_dilations_0, groups = query_19_groups_0, pad = query_19_pad_0, pad_type = query_19_pad_type_0, strides = query_19_strides_0, weight = layers_4_encoder_attn_q_proj_weight_to_fp16, x = obj_65_cast_fp16)[name = tensor("query_19_cast_fp16")]; + tensor key_19_pad_type_0 = const()[name = tensor("key_19_pad_type_0"), val = tensor("valid")]; + tensor key_19_strides_0 = const()[name = tensor("key_19_strides_0"), val = tensor([1, 1])]; + tensor key_19_pad_0 = const()[name = tensor("key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_19_dilations_0 = const()[name = tensor("key_19_dilations_0"), val = tensor([1, 1])]; + tensor key_19_groups_0 = const()[name = tensor("key_19_groups_0"), val = tensor(1)]; + tensor layers_4_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252007552)))]; + tensor key_19_cast_fp16 = conv(dilations = key_19_dilations_0, groups = key_19_groups_0, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = key_19_strides_0, weight = layers_4_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_19_cast_fp16")]; + tensor value_19_pad_type_0 = const()[name = tensor("value_19_pad_type_0"), val = tensor("valid")]; + tensor value_19_strides_0 = const()[name = tensor("value_19_strides_0"), val = tensor([1, 1])]; + tensor value_19_pad_0 = const()[name = tensor("value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_19_dilations_0 = const()[name = tensor("value_19_dilations_0"), val = tensor([1, 1])]; + tensor value_19_groups_0 = const()[name = tensor("value_19_groups_0"), val = tensor(1)]; + tensor layers_4_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(254104768)))]; + tensor layers_4_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256201984)))]; + tensor value_19_cast_fp16 = conv(bias = layers_4_encoder_attn_v_proj_bias_to_fp16, dilations = value_19_dilations_0, groups = value_19_groups_0, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = value_19_strides_0, weight = layers_4_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_19_cast_fp16")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_19_cast_fp16 = reshape(shape = var_1123, x = query_19_cast_fp16)[name = tensor("mh_q_19_cast_fp16")]; + tensor var_1125_to_fp16 = const()[name = tensor("op_1125_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1126_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_1125_to_fp16)[name = tensor("op_1126_cast_fp16")]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([1, 16, 64, -1])]; + tensor var_1128_cast_fp16 = reshape(shape = var_1127, x = key_19_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor mh_w_29_transpose_x_0 = const()[name = tensor("mh_w_29_transpose_x_0"), val = tensor(true)]; + tensor mh_w_29_transpose_y_0 = const()[name = tensor("mh_w_29_transpose_y_0"), val = tensor(false)]; + tensor mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_1126_cast_fp16, y = var_1128_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor obj_69_cast_fp16 = softmax(axis = var_980, x = mh_w_29_cast_fp16)[name = tensor("obj_69_cast_fp16")]; + tensor var_1132 = const()[name = tensor("op_1132"), val = tensor([1, 16, 64, -1])]; + tensor var_1133_cast_fp16 = reshape(shape = var_1132, x = value_19_cast_fp16)[name = tensor("op_1133_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(true)]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_1133_cast_fp16, y = obj_69_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_1136 = const()[name = tensor("op_1136"), val = tensor([1, 1024, 1, -1])]; + tensor input_43_cast_fp16 = reshape(shape = var_1136, x = attn_19_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor obj_67_pad_type_0 = const()[name = tensor("obj_67_pad_type_0"), val = tensor("valid")]; + tensor obj_67_strides_0 = const()[name = tensor("obj_67_strides_0"), val = tensor([1, 1])]; + tensor obj_67_pad_0 = const()[name = tensor("obj_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_67_dilations_0 = const()[name = tensor("obj_67_dilations_0"), val = tensor([1, 1])]; + tensor obj_67_groups_0 = const()[name = tensor("obj_67_groups_0"), val = tensor(1)]; + tensor layers_4_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256204096)))]; + tensor layers_4_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_4_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258301312)))]; + tensor obj_67_cast_fp16 = conv(bias = layers_4_encoder_attn_o_proj_bias_to_fp16, dilations = obj_67_dilations_0, groups = obj_67_groups_0, pad = obj_67_pad_0, pad_type = obj_67_pad_type_0, strides = obj_67_strides_0, weight = layers_4_encoder_attn_o_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("obj_67_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = obj_67_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor out_29_axes_0 = const()[name = tensor("out_29_axes_0"), val = tensor([1])]; + tensor var_1154_to_fp16 = const()[name = tensor("op_1154_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_1154_to_fp16, x = inputs_29_cast_fp16)[name = tensor("out_29_cast_fp16")]; + tensor input_45_gamma_0_to_fp16 = const()[name = tensor("input_45_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258303424)))]; + tensor input_45_beta_0_to_fp16 = const()[name = tensor("input_45_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258305536)))]; + tensor input_45_epsilon_0_to_fp16 = const()[name = tensor("input_45_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = batch_norm(beta = input_45_beta_0_to_fp16, epsilon = input_45_epsilon_0_to_fp16, gamma = input_45_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_pad_type_0 = const()[name = tensor("input_47_pad_type_0"), val = tensor("valid")]; + tensor input_47_strides_0 = const()[name = tensor("input_47_strides_0"), val = tensor([1, 1])]; + tensor input_47_pad_0 = const()[name = tensor("input_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_47_dilations_0 = const()[name = tensor("input_47_dilations_0"), val = tensor([1, 1])]; + tensor input_47_groups_0 = const()[name = tensor("input_47_groups_0"), val = tensor(1)]; + tensor layers_4_fc1_weight_to_fp16 = const()[name = tensor("layers_4_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258307648)))]; + tensor layers_4_fc1_bias_to_fp16 = const()[name = tensor("layers_4_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266696320)))]; + tensor input_47_cast_fp16 = conv(bias = layers_4_fc1_bias_to_fp16, dilations = input_47_dilations_0, groups = input_47_groups_0, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = input_47_strides_0, weight = layers_4_fc1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor hidden_states_11_pad_type_0 = const()[name = tensor("hidden_states_11_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_11_strides_0 = const()[name = tensor("hidden_states_11_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_11_pad_0 = const()[name = tensor("hidden_states_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_11_dilations_0 = const()[name = tensor("hidden_states_11_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_11_groups_0 = const()[name = tensor("hidden_states_11_groups_0"), val = tensor(1)]; + tensor layers_4_fc2_weight_to_fp16 = const()[name = tensor("layers_4_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(266704576)))]; + tensor layers_4_fc2_bias_to_fp16 = const()[name = tensor("layers_4_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275093248)))]; + tensor hidden_states_11_cast_fp16 = conv(bias = layers_4_fc2_bias_to_fp16, dilations = hidden_states_11_dilations_0, groups = hidden_states_11_groups_0, pad = hidden_states_11_pad_0, pad_type = hidden_states_11_pad_type_0, strides = hidden_states_11_strides_0, weight = layers_4_fc2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("hidden_states_11_cast_fp16")]; + tensor inputs_31_cast_fp16 = add(x = inputs_29_cast_fp16, y = hidden_states_11_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor(3)]; + tensor out_31_axes_0 = const()[name = tensor("out_31_axes_0"), val = tensor([1])]; + tensor var_1214_to_fp16 = const()[name = tensor("op_1214_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1214_to_fp16, x = inputs_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; + tensor obj_71_gamma_0_to_fp16 = const()[name = tensor("obj_71_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275095360)))]; + tensor obj_71_beta_0_to_fp16 = const()[name = tensor("obj_71_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275097472)))]; + tensor obj_71_epsilon_0_to_fp16 = const()[name = tensor("obj_71_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor("obj_71_cast_fp16")]; + tensor query_21_pad_type_0 = const()[name = tensor("query_21_pad_type_0"), val = tensor("valid")]; + tensor query_21_strides_0 = const()[name = tensor("query_21_strides_0"), val = tensor([1, 1])]; + tensor query_21_pad_0 = const()[name = tensor("query_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_21_dilations_0 = const()[name = tensor("query_21_dilations_0"), val = tensor([1, 1])]; + tensor query_21_groups_0 = const()[name = tensor("query_21_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275099584)))]; + tensor layers_5_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277196800)))]; + tensor query_21_cast_fp16 = conv(bias = layers_5_self_attn_q_proj_bias_to_fp16, dilations = query_21_dilations_0, groups = query_21_groups_0, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = query_21_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("query_21_cast_fp16")]; + tensor current_key_11_pad_type_0 = const()[name = tensor("current_key_11_pad_type_0"), val = tensor("valid")]; + tensor current_key_11_strides_0 = const()[name = tensor("current_key_11_strides_0"), val = tensor([1, 1])]; + tensor current_key_11_pad_0 = const()[name = tensor("current_key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_11_dilations_0 = const()[name = tensor("current_key_11_dilations_0"), val = tensor([1, 1])]; + tensor current_key_11_groups_0 = const()[name = tensor("current_key_11_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277198912)))]; + tensor current_key_11_cast_fp16 = conv(dilations = current_key_11_dilations_0, groups = current_key_11_groups_0, pad = current_key_11_pad_0, pad_type = current_key_11_pad_type_0, strides = current_key_11_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_key_11_cast_fp16")]; + tensor current_value_11_pad_type_0 = const()[name = tensor("current_value_11_pad_type_0"), val = tensor("valid")]; + tensor current_value_11_strides_0 = const()[name = tensor("current_value_11_strides_0"), val = tensor([1, 1])]; + tensor current_value_11_pad_0 = const()[name = tensor("current_value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_11_dilations_0 = const()[name = tensor("current_value_11_dilations_0"), val = tensor([1, 1])]; + tensor current_value_11_groups_0 = const()[name = tensor("current_value_11_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279296128)))]; + tensor layers_5_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281393344)))]; + tensor current_value_11_cast_fp16 = conv(bias = layers_5_self_attn_v_proj_bias_to_fp16, dilations = current_value_11_dilations_0, groups = current_value_11_groups_0, pad = current_value_11_pad_0, pad_type = current_value_11_pad_type_0, strides = current_value_11_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = obj_71_cast_fp16)[name = tensor("current_value_11_cast_fp16")]; + tensor var_1252_cast_fp16 = mul(x = current_key_11_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1252_cast_fp16")]; + tensor key_21_cast_fp16 = add(x = var_87_cast_fp16_5, y = var_1252_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor var_1254_cast_fp16 = mul(x = current_value_11_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1254_cast_fp16")]; + tensor value_21_cast_fp16 = add(x = var_114_cast_fp16_5, y = var_1254_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_21_cast_fp16 = reshape(shape = var_1257, x = query_21_cast_fp16)[name = tensor("mh_q_21_cast_fp16")]; + tensor var_1259_to_fp16 = const()[name = tensor("op_1259_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1260_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_1259_to_fp16)[name = tensor("op_1260_cast_fp16")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([1, 16, 64, -1])]; + tensor var_1262_cast_fp16 = reshape(shape = var_1261, x = key_21_cast_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor mh_w_31_transpose_x_0 = const()[name = tensor("mh_w_31_transpose_x_0"), val = tensor(true)]; + tensor mh_w_31_transpose_y_0 = const()[name = tensor("mh_w_31_transpose_y_0"), val = tensor(false)]; + tensor mh_w_31_cast_fp16 = matmul(transpose_x = mh_w_31_transpose_x_0, transpose_y = mh_w_31_transpose_y_0, x = var_1260_cast_fp16, y = var_1262_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor mh_w_33_cast_fp16 = add(x = mh_w_31_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor var_1270_cast_fp16 = softmax(axis = var_1189, x = mh_w_33_cast_fp16)[name = tensor("op_1270_cast_fp16")]; + tensor var_1271 = const()[name = tensor("op_1271"), val = tensor([1, 16, 64, -1])]; + tensor var_1272_cast_fp16 = reshape(shape = var_1271, x = value_21_cast_fp16)[name = tensor("op_1272_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(true)]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_1272_cast_fp16, y = var_1270_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_1275 = const()[name = tensor("op_1275"), val = tensor([1, 1024, 1, -1])]; + tensor input_51_cast_fp16 = reshape(shape = var_1275, x = attn_21_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor obj_77_pad_type_0 = const()[name = tensor("obj_77_pad_type_0"), val = tensor("valid")]; + tensor obj_77_strides_0 = const()[name = tensor("obj_77_strides_0"), val = tensor([1, 1])]; + tensor obj_77_pad_0 = const()[name = tensor("obj_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_77_dilations_0 = const()[name = tensor("obj_77_dilations_0"), val = tensor([1, 1])]; + tensor obj_77_groups_0 = const()[name = tensor("obj_77_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281395456)))]; + tensor layers_5_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283492672)))]; + tensor obj_77_cast_fp16 = conv(bias = layers_5_self_attn_o_proj_bias_to_fp16, dilations = obj_77_dilations_0, groups = obj_77_groups_0, pad = obj_77_pad_0, pad_type = obj_77_pad_type_0, strides = obj_77_strides_0, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("obj_77_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = obj_77_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([1])]; + tensor var_1297_to_fp16 = const()[name = tensor("op_1297_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1297_to_fp16, x = inputs_33_cast_fp16)[name = tensor("out_33_cast_fp16")]; + tensor obj_79_gamma_0_to_fp16 = const()[name = tensor("obj_79_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283494784)))]; + tensor obj_79_beta_0_to_fp16 = const()[name = tensor("obj_79_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283496896)))]; + tensor obj_79_epsilon_0_to_fp16 = const()[name = tensor("obj_79_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor("obj_79_cast_fp16")]; + tensor query_23_pad_type_0 = const()[name = tensor("query_23_pad_type_0"), val = tensor("valid")]; + tensor query_23_strides_0 = const()[name = tensor("query_23_strides_0"), val = tensor([1, 1])]; + tensor query_23_pad_0 = const()[name = tensor("query_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_23_dilations_0 = const()[name = tensor("query_23_dilations_0"), val = tensor([1, 1])]; + tensor query_23_groups_0 = const()[name = tensor("query_23_groups_0"), val = tensor(1)]; + tensor layers_5_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283499008)))]; + tensor layers_5_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285596224)))]; + tensor query_23_cast_fp16 = conv(bias = layers_5_encoder_attn_q_proj_bias_to_fp16, dilations = query_23_dilations_0, groups = query_23_groups_0, pad = query_23_pad_0, pad_type = query_23_pad_type_0, strides = query_23_strides_0, weight = layers_5_encoder_attn_q_proj_weight_to_fp16, x = obj_79_cast_fp16)[name = tensor("query_23_cast_fp16")]; + tensor key_23_pad_type_0 = const()[name = tensor("key_23_pad_type_0"), val = tensor("valid")]; + tensor key_23_strides_0 = const()[name = tensor("key_23_strides_0"), val = tensor([1, 1])]; + tensor key_23_pad_0 = const()[name = tensor("key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_23_dilations_0 = const()[name = tensor("key_23_dilations_0"), val = tensor([1, 1])]; + tensor key_23_groups_0 = const()[name = tensor("key_23_groups_0"), val = tensor(1)]; + tensor layers_5_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285598336)))]; + tensor key_23_cast_fp16 = conv(dilations = key_23_dilations_0, groups = key_23_groups_0, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = key_23_strides_0, weight = layers_5_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_23_cast_fp16")]; + tensor value_23_pad_type_0 = const()[name = tensor("value_23_pad_type_0"), val = tensor("valid")]; + tensor value_23_strides_0 = const()[name = tensor("value_23_strides_0"), val = tensor([1, 1])]; + tensor value_23_pad_0 = const()[name = tensor("value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_23_dilations_0 = const()[name = tensor("value_23_dilations_0"), val = tensor([1, 1])]; + tensor value_23_groups_0 = const()[name = tensor("value_23_groups_0"), val = tensor(1)]; + tensor layers_5_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(287695552)))]; + tensor layers_5_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289792768)))]; + tensor value_23_cast_fp16 = conv(bias = layers_5_encoder_attn_v_proj_bias_to_fp16, dilations = value_23_dilations_0, groups = value_23_groups_0, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = value_23_strides_0, weight = layers_5_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_23_cast_fp16")]; + tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_23_cast_fp16 = reshape(shape = var_1332, x = query_23_cast_fp16)[name = tensor("mh_q_23_cast_fp16")]; + tensor var_1334_to_fp16 = const()[name = tensor("op_1334_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1335_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_1334_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor var_1336 = const()[name = tensor("op_1336"), val = tensor([1, 16, 64, -1])]; + tensor var_1337_cast_fp16 = reshape(shape = var_1336, x = key_23_cast_fp16)[name = tensor("op_1337_cast_fp16")]; + tensor mh_w_35_transpose_x_0 = const()[name = tensor("mh_w_35_transpose_x_0"), val = tensor(true)]; + tensor mh_w_35_transpose_y_0 = const()[name = tensor("mh_w_35_transpose_y_0"), val = tensor(false)]; + tensor mh_w_35_cast_fp16 = matmul(transpose_x = mh_w_35_transpose_x_0, transpose_y = mh_w_35_transpose_y_0, x = var_1335_cast_fp16, y = var_1337_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; + tensor obj_83_cast_fp16 = softmax(axis = var_1189, x = mh_w_35_cast_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor var_1341 = const()[name = tensor("op_1341"), val = tensor([1, 16, 64, -1])]; + tensor var_1342_cast_fp16 = reshape(shape = var_1341, x = value_23_cast_fp16)[name = tensor("op_1342_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(true)]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_1342_cast_fp16, y = obj_83_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_1345 = const()[name = tensor("op_1345"), val = tensor([1, 1024, 1, -1])]; + tensor input_53_cast_fp16 = reshape(shape = var_1345, x = attn_23_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor obj_81_pad_type_0 = const()[name = tensor("obj_81_pad_type_0"), val = tensor("valid")]; + tensor obj_81_strides_0 = const()[name = tensor("obj_81_strides_0"), val = tensor([1, 1])]; + tensor obj_81_pad_0 = const()[name = tensor("obj_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_81_dilations_0 = const()[name = tensor("obj_81_dilations_0"), val = tensor([1, 1])]; + tensor obj_81_groups_0 = const()[name = tensor("obj_81_groups_0"), val = tensor(1)]; + tensor layers_5_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289794880)))]; + tensor layers_5_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_5_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291892096)))]; + tensor obj_81_cast_fp16 = conv(bias = layers_5_encoder_attn_o_proj_bias_to_fp16, dilations = obj_81_dilations_0, groups = obj_81_groups_0, pad = obj_81_pad_0, pad_type = obj_81_pad_type_0, strides = obj_81_strides_0, weight = layers_5_encoder_attn_o_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("obj_81_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_81_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor out_35_axes_0 = const()[name = tensor("out_35_axes_0"), val = tensor([1])]; + tensor var_1363_to_fp16 = const()[name = tensor("op_1363_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1363_to_fp16, x = inputs_35_cast_fp16)[name = tensor("out_35_cast_fp16")]; + tensor input_55_gamma_0_to_fp16 = const()[name = tensor("input_55_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291894208)))]; + tensor input_55_beta_0_to_fp16 = const()[name = tensor("input_55_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291896320)))]; + tensor input_55_epsilon_0_to_fp16 = const()[name = tensor("input_55_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_55_cast_fp16 = batch_norm(beta = input_55_beta_0_to_fp16, epsilon = input_55_epsilon_0_to_fp16, gamma = input_55_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_pad_type_0 = const()[name = tensor("input_57_pad_type_0"), val = tensor("valid")]; + tensor input_57_strides_0 = const()[name = tensor("input_57_strides_0"), val = tensor([1, 1])]; + tensor input_57_pad_0 = const()[name = tensor("input_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_57_dilations_0 = const()[name = tensor("input_57_dilations_0"), val = tensor([1, 1])]; + tensor input_57_groups_0 = const()[name = tensor("input_57_groups_0"), val = tensor(1)]; + tensor layers_5_fc1_weight_to_fp16 = const()[name = tensor("layers_5_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(291898432)))]; + tensor layers_5_fc1_bias_to_fp16 = const()[name = tensor("layers_5_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(300287104)))]; + tensor input_57_cast_fp16 = conv(bias = layers_5_fc1_bias_to_fp16, dilations = input_57_dilations_0, groups = input_57_groups_0, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = input_57_strides_0, weight = layers_5_fc1_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor hidden_states_13_pad_type_0 = const()[name = tensor("hidden_states_13_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_13_strides_0 = const()[name = tensor("hidden_states_13_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_13_pad_0 = const()[name = tensor("hidden_states_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_13_dilations_0 = const()[name = tensor("hidden_states_13_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_13_groups_0 = const()[name = tensor("hidden_states_13_groups_0"), val = tensor(1)]; + tensor layers_5_fc2_weight_to_fp16 = const()[name = tensor("layers_5_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(300295360)))]; + tensor layers_5_fc2_bias_to_fp16 = const()[name = tensor("layers_5_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308684032)))]; + tensor hidden_states_13_cast_fp16 = conv(bias = layers_5_fc2_bias_to_fp16, dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = layers_5_fc2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("hidden_states_13_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = hidden_states_13_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor var_1398 = const()[name = tensor("op_1398"), val = tensor(3)]; + tensor out_37_axes_0 = const()[name = tensor("out_37_axes_0"), val = tensor([1])]; + tensor var_1423_to_fp16 = const()[name = tensor("op_1423_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1423_to_fp16, x = inputs_37_cast_fp16)[name = tensor("out_37_cast_fp16")]; + tensor obj_85_gamma_0_to_fp16 = const()[name = tensor("obj_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308686144)))]; + tensor obj_85_beta_0_to_fp16 = const()[name = tensor("obj_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308688256)))]; + tensor obj_85_epsilon_0_to_fp16 = const()[name = tensor("obj_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_85_cast_fp16 = batch_norm(beta = obj_85_beta_0_to_fp16, epsilon = obj_85_epsilon_0_to_fp16, gamma = obj_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor("obj_85_cast_fp16")]; + tensor query_25_pad_type_0 = const()[name = tensor("query_25_pad_type_0"), val = tensor("valid")]; + tensor query_25_strides_0 = const()[name = tensor("query_25_strides_0"), val = tensor([1, 1])]; + tensor query_25_pad_0 = const()[name = tensor("query_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_25_dilations_0 = const()[name = tensor("query_25_dilations_0"), val = tensor([1, 1])]; + tensor query_25_groups_0 = const()[name = tensor("query_25_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308690368)))]; + tensor layers_6_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310787584)))]; + tensor query_25_cast_fp16 = conv(bias = layers_6_self_attn_q_proj_bias_to_fp16, dilations = query_25_dilations_0, groups = query_25_groups_0, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = query_25_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("query_25_cast_fp16")]; + tensor current_key_13_pad_type_0 = const()[name = tensor("current_key_13_pad_type_0"), val = tensor("valid")]; + tensor current_key_13_strides_0 = const()[name = tensor("current_key_13_strides_0"), val = tensor([1, 1])]; + tensor current_key_13_pad_0 = const()[name = tensor("current_key_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_13_dilations_0 = const()[name = tensor("current_key_13_dilations_0"), val = tensor([1, 1])]; + tensor current_key_13_groups_0 = const()[name = tensor("current_key_13_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310789696)))]; + tensor current_key_13_cast_fp16 = conv(dilations = current_key_13_dilations_0, groups = current_key_13_groups_0, pad = current_key_13_pad_0, pad_type = current_key_13_pad_type_0, strides = current_key_13_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_key_13_cast_fp16")]; + tensor current_value_13_pad_type_0 = const()[name = tensor("current_value_13_pad_type_0"), val = tensor("valid")]; + tensor current_value_13_strides_0 = const()[name = tensor("current_value_13_strides_0"), val = tensor([1, 1])]; + tensor current_value_13_pad_0 = const()[name = tensor("current_value_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_13_dilations_0 = const()[name = tensor("current_value_13_dilations_0"), val = tensor([1, 1])]; + tensor current_value_13_groups_0 = const()[name = tensor("current_value_13_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312886912)))]; + tensor layers_6_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314984128)))]; + tensor current_value_13_cast_fp16 = conv(bias = layers_6_self_attn_v_proj_bias_to_fp16, dilations = current_value_13_dilations_0, groups = current_value_13_groups_0, pad = current_value_13_pad_0, pad_type = current_value_13_pad_type_0, strides = current_value_13_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = obj_85_cast_fp16)[name = tensor("current_value_13_cast_fp16")]; + tensor var_1461_cast_fp16 = mul(x = current_key_13_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1461_cast_fp16")]; + tensor key_25_cast_fp16 = add(x = var_87_cast_fp16_6, y = var_1461_cast_fp16)[name = tensor("key_25_cast_fp16")]; + tensor var_1463_cast_fp16 = mul(x = current_value_13_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1463_cast_fp16")]; + tensor value_25_cast_fp16 = add(x = var_114_cast_fp16_6, y = var_1463_cast_fp16)[name = tensor("value_25_cast_fp16")]; + tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_25_cast_fp16 = reshape(shape = var_1466, x = query_25_cast_fp16)[name = tensor("mh_q_25_cast_fp16")]; + tensor var_1468_to_fp16 = const()[name = tensor("op_1468_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1469_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_1468_to_fp16)[name = tensor("op_1469_cast_fp16")]; + tensor var_1470 = const()[name = tensor("op_1470"), val = tensor([1, 16, 64, -1])]; + tensor var_1471_cast_fp16 = reshape(shape = var_1470, x = key_25_cast_fp16)[name = tensor("op_1471_cast_fp16")]; + tensor mh_w_37_transpose_x_0 = const()[name = tensor("mh_w_37_transpose_x_0"), val = tensor(true)]; + tensor mh_w_37_transpose_y_0 = const()[name = tensor("mh_w_37_transpose_y_0"), val = tensor(false)]; + tensor mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_1469_cast_fp16, y = var_1471_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; + tensor mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; + tensor var_1479_cast_fp16 = softmax(axis = var_1398, x = mh_w_39_cast_fp16)[name = tensor("op_1479_cast_fp16")]; + tensor var_1480 = const()[name = tensor("op_1480"), val = tensor([1, 16, 64, -1])]; + tensor var_1481_cast_fp16 = reshape(shape = var_1480, x = value_25_cast_fp16)[name = tensor("op_1481_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(true)]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_1481_cast_fp16, y = var_1479_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_1484 = const()[name = tensor("op_1484"), val = tensor([1, 1024, 1, -1])]; + tensor input_61_cast_fp16 = reshape(shape = var_1484, x = attn_25_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor obj_91_pad_type_0 = const()[name = tensor("obj_91_pad_type_0"), val = tensor("valid")]; + tensor obj_91_strides_0 = const()[name = tensor("obj_91_strides_0"), val = tensor([1, 1])]; + tensor obj_91_pad_0 = const()[name = tensor("obj_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_91_dilations_0 = const()[name = tensor("obj_91_dilations_0"), val = tensor([1, 1])]; + tensor obj_91_groups_0 = const()[name = tensor("obj_91_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314986240)))]; + tensor layers_6_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_6_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317083456)))]; + tensor obj_91_cast_fp16 = conv(bias = layers_6_self_attn_o_proj_bias_to_fp16, dilations = obj_91_dilations_0, groups = obj_91_groups_0, pad = obj_91_pad_0, pad_type = obj_91_pad_type_0, strides = obj_91_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("obj_91_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = obj_91_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor out_39_axes_0 = const()[name = tensor("out_39_axes_0"), val = tensor([1])]; + tensor var_1506_to_fp16 = const()[name = tensor("op_1506_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1506_to_fp16, x = inputs_39_cast_fp16)[name = tensor("out_39_cast_fp16")]; + tensor obj_93_gamma_0_to_fp16 = const()[name = tensor("obj_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317085568)))]; + tensor obj_93_beta_0_to_fp16 = const()[name = tensor("obj_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317087680)))]; + tensor obj_93_epsilon_0_to_fp16 = const()[name = tensor("obj_93_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_93_cast_fp16 = batch_norm(beta = obj_93_beta_0_to_fp16, epsilon = obj_93_epsilon_0_to_fp16, gamma = obj_93_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor("obj_93_cast_fp16")]; + tensor query_27_pad_type_0 = const()[name = tensor("query_27_pad_type_0"), val = tensor("valid")]; + tensor query_27_strides_0 = const()[name = tensor("query_27_strides_0"), val = tensor([1, 1])]; + tensor query_27_pad_0 = const()[name = tensor("query_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_27_dilations_0 = const()[name = tensor("query_27_dilations_0"), val = tensor([1, 1])]; + tensor query_27_groups_0 = const()[name = tensor("query_27_groups_0"), val = tensor(1)]; + tensor layers_6_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317089792)))]; + tensor layers_6_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319187008)))]; + tensor query_27_cast_fp16 = conv(bias = layers_6_encoder_attn_q_proj_bias_to_fp16, dilations = query_27_dilations_0, groups = query_27_groups_0, pad = query_27_pad_0, pad_type = query_27_pad_type_0, strides = query_27_strides_0, weight = layers_6_encoder_attn_q_proj_weight_to_fp16, x = obj_93_cast_fp16)[name = tensor("query_27_cast_fp16")]; + tensor key_27_pad_type_0 = const()[name = tensor("key_27_pad_type_0"), val = tensor("valid")]; + tensor key_27_strides_0 = const()[name = tensor("key_27_strides_0"), val = tensor([1, 1])]; + tensor key_27_pad_0 = const()[name = tensor("key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_27_dilations_0 = const()[name = tensor("key_27_dilations_0"), val = tensor([1, 1])]; + tensor key_27_groups_0 = const()[name = tensor("key_27_groups_0"), val = tensor(1)]; + tensor layers_6_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319189120)))]; + tensor key_27_cast_fp16 = conv(dilations = key_27_dilations_0, groups = key_27_groups_0, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = key_27_strides_0, weight = layers_6_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_27_cast_fp16")]; + tensor value_27_pad_type_0 = const()[name = tensor("value_27_pad_type_0"), val = tensor("valid")]; + tensor value_27_strides_0 = const()[name = tensor("value_27_strides_0"), val = tensor([1, 1])]; + tensor value_27_pad_0 = const()[name = tensor("value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_27_dilations_0 = const()[name = tensor("value_27_dilations_0"), val = tensor([1, 1])]; + tensor value_27_groups_0 = const()[name = tensor("value_27_groups_0"), val = tensor(1)]; + tensor layers_6_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321286336)))]; + tensor layers_6_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323383552)))]; + tensor value_27_cast_fp16 = conv(bias = layers_6_encoder_attn_v_proj_bias_to_fp16, dilations = value_27_dilations_0, groups = value_27_groups_0, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = value_27_strides_0, weight = layers_6_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_27_cast_fp16")]; + tensor var_1541 = const()[name = tensor("op_1541"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_27_cast_fp16 = reshape(shape = var_1541, x = query_27_cast_fp16)[name = tensor("mh_q_27_cast_fp16")]; + tensor var_1543_to_fp16 = const()[name = tensor("op_1543_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1544_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_1543_to_fp16)[name = tensor("op_1544_cast_fp16")]; + tensor var_1545 = const()[name = tensor("op_1545"), val = tensor([1, 16, 64, -1])]; + tensor var_1546_cast_fp16 = reshape(shape = var_1545, x = key_27_cast_fp16)[name = tensor("op_1546_cast_fp16")]; + tensor mh_w_41_transpose_x_0 = const()[name = tensor("mh_w_41_transpose_x_0"), val = tensor(true)]; + tensor mh_w_41_transpose_y_0 = const()[name = tensor("mh_w_41_transpose_y_0"), val = tensor(false)]; + tensor mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_1544_cast_fp16, y = var_1546_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; + tensor obj_97_cast_fp16 = softmax(axis = var_1398, x = mh_w_41_cast_fp16)[name = tensor("obj_97_cast_fp16")]; + tensor var_1550 = const()[name = tensor("op_1550"), val = tensor([1, 16, 64, -1])]; + tensor var_1551_cast_fp16 = reshape(shape = var_1550, x = value_27_cast_fp16)[name = tensor("op_1551_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(true)]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_1551_cast_fp16, y = obj_97_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_1554 = const()[name = tensor("op_1554"), val = tensor([1, 1024, 1, -1])]; + tensor input_63_cast_fp16 = reshape(shape = var_1554, x = attn_27_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor obj_95_pad_type_0 = const()[name = tensor("obj_95_pad_type_0"), val = tensor("valid")]; + tensor obj_95_strides_0 = const()[name = tensor("obj_95_strides_0"), val = tensor([1, 1])]; + tensor obj_95_pad_0 = const()[name = tensor("obj_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_95_dilations_0 = const()[name = tensor("obj_95_dilations_0"), val = tensor([1, 1])]; + tensor obj_95_groups_0 = const()[name = tensor("obj_95_groups_0"), val = tensor(1)]; + tensor layers_6_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(323385664)))]; + tensor layers_6_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_6_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325482880)))]; + tensor obj_95_cast_fp16 = conv(bias = layers_6_encoder_attn_o_proj_bias_to_fp16, dilations = obj_95_dilations_0, groups = obj_95_groups_0, pad = obj_95_pad_0, pad_type = obj_95_pad_type_0, strides = obj_95_strides_0, weight = layers_6_encoder_attn_o_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("obj_95_cast_fp16")]; + tensor inputs_41_cast_fp16 = add(x = inputs_39_cast_fp16, y = obj_95_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor out_41_axes_0 = const()[name = tensor("out_41_axes_0"), val = tensor([1])]; + tensor var_1572_to_fp16 = const()[name = tensor("op_1572_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1572_to_fp16, x = inputs_41_cast_fp16)[name = tensor("out_41_cast_fp16")]; + tensor input_65_gamma_0_to_fp16 = const()[name = tensor("input_65_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325484992)))]; + tensor input_65_beta_0_to_fp16 = const()[name = tensor("input_65_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325487104)))]; + tensor input_65_epsilon_0_to_fp16 = const()[name = tensor("input_65_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = batch_norm(beta = input_65_beta_0_to_fp16, epsilon = input_65_epsilon_0_to_fp16, gamma = input_65_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor input_67_pad_type_0 = const()[name = tensor("input_67_pad_type_0"), val = tensor("valid")]; + tensor input_67_strides_0 = const()[name = tensor("input_67_strides_0"), val = tensor([1, 1])]; + tensor input_67_pad_0 = const()[name = tensor("input_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_67_dilations_0 = const()[name = tensor("input_67_dilations_0"), val = tensor([1, 1])]; + tensor input_67_groups_0 = const()[name = tensor("input_67_groups_0"), val = tensor(1)]; + tensor layers_6_fc1_weight_to_fp16 = const()[name = tensor("layers_6_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325489216)))]; + tensor layers_6_fc1_bias_to_fp16 = const()[name = tensor("layers_6_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333877888)))]; + tensor input_67_cast_fp16 = conv(bias = layers_6_fc1_bias_to_fp16, dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = layers_6_fc1_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_mode_0 = const()[name = tensor("input_69_mode_0"), val = tensor("EXACT")]; + tensor input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor hidden_states_15_pad_type_0 = const()[name = tensor("hidden_states_15_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_15_strides_0 = const()[name = tensor("hidden_states_15_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_15_pad_0 = const()[name = tensor("hidden_states_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_15_dilations_0 = const()[name = tensor("hidden_states_15_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_15_groups_0 = const()[name = tensor("hidden_states_15_groups_0"), val = tensor(1)]; + tensor layers_6_fc2_weight_to_fp16 = const()[name = tensor("layers_6_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333886144)))]; + tensor layers_6_fc2_bias_to_fp16 = const()[name = tensor("layers_6_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342274816)))]; + tensor hidden_states_15_cast_fp16 = conv(bias = layers_6_fc2_bias_to_fp16, dilations = hidden_states_15_dilations_0, groups = hidden_states_15_groups_0, pad = hidden_states_15_pad_0, pad_type = hidden_states_15_pad_type_0, strides = hidden_states_15_strides_0, weight = layers_6_fc2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("hidden_states_15_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = hidden_states_15_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor var_1607 = const()[name = tensor("op_1607"), val = tensor(3)]; + tensor out_43_axes_0 = const()[name = tensor("out_43_axes_0"), val = tensor([1])]; + tensor var_1632_to_fp16 = const()[name = tensor("op_1632_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1632_to_fp16, x = inputs_43_cast_fp16)[name = tensor("out_43_cast_fp16")]; + tensor obj_99_gamma_0_to_fp16 = const()[name = tensor("obj_99_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342276928)))]; + tensor obj_99_beta_0_to_fp16 = const()[name = tensor("obj_99_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342279040)))]; + tensor obj_99_epsilon_0_to_fp16 = const()[name = tensor("obj_99_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_99_cast_fp16 = batch_norm(beta = obj_99_beta_0_to_fp16, epsilon = obj_99_epsilon_0_to_fp16, gamma = obj_99_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor("obj_99_cast_fp16")]; + tensor query_29_pad_type_0 = const()[name = tensor("query_29_pad_type_0"), val = tensor("valid")]; + tensor query_29_strides_0 = const()[name = tensor("query_29_strides_0"), val = tensor([1, 1])]; + tensor query_29_pad_0 = const()[name = tensor("query_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_29_dilations_0 = const()[name = tensor("query_29_dilations_0"), val = tensor([1, 1])]; + tensor query_29_groups_0 = const()[name = tensor("query_29_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342281152)))]; + tensor layers_7_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344378368)))]; + tensor query_29_cast_fp16 = conv(bias = layers_7_self_attn_q_proj_bias_to_fp16, dilations = query_29_dilations_0, groups = query_29_groups_0, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = query_29_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("query_29_cast_fp16")]; + tensor current_key_15_pad_type_0 = const()[name = tensor("current_key_15_pad_type_0"), val = tensor("valid")]; + tensor current_key_15_strides_0 = const()[name = tensor("current_key_15_strides_0"), val = tensor([1, 1])]; + tensor current_key_15_pad_0 = const()[name = tensor("current_key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_15_dilations_0 = const()[name = tensor("current_key_15_dilations_0"), val = tensor([1, 1])]; + tensor current_key_15_groups_0 = const()[name = tensor("current_key_15_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344380480)))]; + tensor current_key_15_cast_fp16 = conv(dilations = current_key_15_dilations_0, groups = current_key_15_groups_0, pad = current_key_15_pad_0, pad_type = current_key_15_pad_type_0, strides = current_key_15_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("current_key_15_cast_fp16")]; + tensor current_value_15_pad_type_0 = const()[name = tensor("current_value_15_pad_type_0"), val = tensor("valid")]; + tensor current_value_15_strides_0 = const()[name = tensor("current_value_15_strides_0"), val = tensor([1, 1])]; + tensor current_value_15_pad_0 = const()[name = tensor("current_value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_15_dilations_0 = const()[name = tensor("current_value_15_dilations_0"), val = tensor([1, 1])]; + tensor current_value_15_groups_0 = const()[name = tensor("current_value_15_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(346477696)))]; + tensor layers_7_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348574912)))]; + tensor current_value_15_cast_fp16 = conv(bias = layers_7_self_attn_v_proj_bias_to_fp16, dilations = current_value_15_dilations_0, groups = current_value_15_groups_0, pad = current_value_15_pad_0, pad_type = current_value_15_pad_type_0, strides = current_value_15_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = obj_99_cast_fp16)[name = tensor("current_value_15_cast_fp16")]; + tensor var_1670_cast_fp16 = mul(x = current_key_15_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1670_cast_fp16")]; + tensor key_29_cast_fp16 = add(x = var_87_cast_fp16_7, y = var_1670_cast_fp16)[name = tensor("key_29_cast_fp16")]; + tensor var_1672_cast_fp16 = mul(x = current_value_15_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1672_cast_fp16")]; + tensor value_29_cast_fp16 = add(x = var_114_cast_fp16_7, y = var_1672_cast_fp16)[name = tensor("value_29_cast_fp16")]; + tensor var_1675 = const()[name = tensor("op_1675"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_29_cast_fp16 = reshape(shape = var_1675, x = query_29_cast_fp16)[name = tensor("mh_q_29_cast_fp16")]; + tensor var_1677_to_fp16 = const()[name = tensor("op_1677_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1678_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_1677_to_fp16)[name = tensor("op_1678_cast_fp16")]; + tensor var_1679 = const()[name = tensor("op_1679"), val = tensor([1, 16, 64, -1])]; + tensor var_1680_cast_fp16 = reshape(shape = var_1679, x = key_29_cast_fp16)[name = tensor("op_1680_cast_fp16")]; + tensor mh_w_43_transpose_x_0 = const()[name = tensor("mh_w_43_transpose_x_0"), val = tensor(true)]; + tensor mh_w_43_transpose_y_0 = const()[name = tensor("mh_w_43_transpose_y_0"), val = tensor(false)]; + tensor mh_w_43_cast_fp16 = matmul(transpose_x = mh_w_43_transpose_x_0, transpose_y = mh_w_43_transpose_y_0, x = var_1678_cast_fp16, y = var_1680_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; + tensor mh_w_45_cast_fp16 = add(x = mh_w_43_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; + tensor var_1688_cast_fp16 = softmax(axis = var_1607, x = mh_w_45_cast_fp16)[name = tensor("op_1688_cast_fp16")]; + tensor var_1689 = const()[name = tensor("op_1689"), val = tensor([1, 16, 64, -1])]; + tensor var_1690_cast_fp16 = reshape(shape = var_1689, x = value_29_cast_fp16)[name = tensor("op_1690_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(true)]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_1690_cast_fp16, y = var_1688_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_1693 = const()[name = tensor("op_1693"), val = tensor([1, 1024, 1, -1])]; + tensor input_71_cast_fp16 = reshape(shape = var_1693, x = attn_29_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor obj_105_pad_type_0 = const()[name = tensor("obj_105_pad_type_0"), val = tensor("valid")]; + tensor obj_105_strides_0 = const()[name = tensor("obj_105_strides_0"), val = tensor([1, 1])]; + tensor obj_105_pad_0 = const()[name = tensor("obj_105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_105_dilations_0 = const()[name = tensor("obj_105_dilations_0"), val = tensor([1, 1])]; + tensor obj_105_groups_0 = const()[name = tensor("obj_105_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(348577024)))]; + tensor layers_7_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_7_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350674240)))]; + tensor obj_105_cast_fp16 = conv(bias = layers_7_self_attn_o_proj_bias_to_fp16, dilations = obj_105_dilations_0, groups = obj_105_groups_0, pad = obj_105_pad_0, pad_type = obj_105_pad_type_0, strides = obj_105_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("obj_105_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_105_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor out_45_axes_0 = const()[name = tensor("out_45_axes_0"), val = tensor([1])]; + tensor var_1715_to_fp16 = const()[name = tensor("op_1715_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1715_to_fp16, x = inputs_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; + tensor obj_107_gamma_0_to_fp16 = const()[name = tensor("obj_107_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350676352)))]; + tensor obj_107_beta_0_to_fp16 = const()[name = tensor("obj_107_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350678464)))]; + tensor obj_107_epsilon_0_to_fp16 = const()[name = tensor("obj_107_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_107_cast_fp16 = batch_norm(beta = obj_107_beta_0_to_fp16, epsilon = obj_107_epsilon_0_to_fp16, gamma = obj_107_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor("obj_107_cast_fp16")]; + tensor query_31_pad_type_0 = const()[name = tensor("query_31_pad_type_0"), val = tensor("valid")]; + tensor query_31_strides_0 = const()[name = tensor("query_31_strides_0"), val = tensor([1, 1])]; + tensor query_31_pad_0 = const()[name = tensor("query_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_31_dilations_0 = const()[name = tensor("query_31_dilations_0"), val = tensor([1, 1])]; + tensor query_31_groups_0 = const()[name = tensor("query_31_groups_0"), val = tensor(1)]; + tensor layers_7_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350680576)))]; + tensor layers_7_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352777792)))]; + tensor query_31_cast_fp16 = conv(bias = layers_7_encoder_attn_q_proj_bias_to_fp16, dilations = query_31_dilations_0, groups = query_31_groups_0, pad = query_31_pad_0, pad_type = query_31_pad_type_0, strides = query_31_strides_0, weight = layers_7_encoder_attn_q_proj_weight_to_fp16, x = obj_107_cast_fp16)[name = tensor("query_31_cast_fp16")]; + tensor key_31_pad_type_0 = const()[name = tensor("key_31_pad_type_0"), val = tensor("valid")]; + tensor key_31_strides_0 = const()[name = tensor("key_31_strides_0"), val = tensor([1, 1])]; + tensor key_31_pad_0 = const()[name = tensor("key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_31_dilations_0 = const()[name = tensor("key_31_dilations_0"), val = tensor([1, 1])]; + tensor key_31_groups_0 = const()[name = tensor("key_31_groups_0"), val = tensor(1)]; + tensor layers_7_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352779904)))]; + tensor key_31_cast_fp16 = conv(dilations = key_31_dilations_0, groups = key_31_groups_0, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = key_31_strides_0, weight = layers_7_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_31_cast_fp16")]; + tensor value_31_pad_type_0 = const()[name = tensor("value_31_pad_type_0"), val = tensor("valid")]; + tensor value_31_strides_0 = const()[name = tensor("value_31_strides_0"), val = tensor([1, 1])]; + tensor value_31_pad_0 = const()[name = tensor("value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_31_dilations_0 = const()[name = tensor("value_31_dilations_0"), val = tensor([1, 1])]; + tensor value_31_groups_0 = const()[name = tensor("value_31_groups_0"), val = tensor(1)]; + tensor layers_7_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354877120)))]; + tensor layers_7_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356974336)))]; + tensor value_31_cast_fp16 = conv(bias = layers_7_encoder_attn_v_proj_bias_to_fp16, dilations = value_31_dilations_0, groups = value_31_groups_0, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = value_31_strides_0, weight = layers_7_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_31_cast_fp16")]; + tensor var_1750 = const()[name = tensor("op_1750"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_31_cast_fp16 = reshape(shape = var_1750, x = query_31_cast_fp16)[name = tensor("mh_q_31_cast_fp16")]; + tensor var_1752_to_fp16 = const()[name = tensor("op_1752_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1753_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_1752_to_fp16)[name = tensor("op_1753_cast_fp16")]; + tensor var_1754 = const()[name = tensor("op_1754"), val = tensor([1, 16, 64, -1])]; + tensor var_1755_cast_fp16 = reshape(shape = var_1754, x = key_31_cast_fp16)[name = tensor("op_1755_cast_fp16")]; + tensor mh_w_47_transpose_x_0 = const()[name = tensor("mh_w_47_transpose_x_0"), val = tensor(true)]; + tensor mh_w_47_transpose_y_0 = const()[name = tensor("mh_w_47_transpose_y_0"), val = tensor(false)]; + tensor mh_w_47_cast_fp16 = matmul(transpose_x = mh_w_47_transpose_x_0, transpose_y = mh_w_47_transpose_y_0, x = var_1753_cast_fp16, y = var_1755_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; + tensor obj_111_cast_fp16 = softmax(axis = var_1607, x = mh_w_47_cast_fp16)[name = tensor("obj_111_cast_fp16")]; + tensor var_1759 = const()[name = tensor("op_1759"), val = tensor([1, 16, 64, -1])]; + tensor var_1760_cast_fp16 = reshape(shape = var_1759, x = value_31_cast_fp16)[name = tensor("op_1760_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(true)]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_1760_cast_fp16, y = obj_111_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_1763 = const()[name = tensor("op_1763"), val = tensor([1, 1024, 1, -1])]; + tensor input_73_cast_fp16 = reshape(shape = var_1763, x = attn_31_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor obj_109_pad_type_0 = const()[name = tensor("obj_109_pad_type_0"), val = tensor("valid")]; + tensor obj_109_strides_0 = const()[name = tensor("obj_109_strides_0"), val = tensor([1, 1])]; + tensor obj_109_pad_0 = const()[name = tensor("obj_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_109_dilations_0 = const()[name = tensor("obj_109_dilations_0"), val = tensor([1, 1])]; + tensor obj_109_groups_0 = const()[name = tensor("obj_109_groups_0"), val = tensor(1)]; + tensor layers_7_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356976448)))]; + tensor layers_7_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_7_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359073664)))]; + tensor obj_109_cast_fp16 = conv(bias = layers_7_encoder_attn_o_proj_bias_to_fp16, dilations = obj_109_dilations_0, groups = obj_109_groups_0, pad = obj_109_pad_0, pad_type = obj_109_pad_type_0, strides = obj_109_strides_0, weight = layers_7_encoder_attn_o_proj_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("obj_109_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = obj_109_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor out_47_axes_0 = const()[name = tensor("out_47_axes_0"), val = tensor([1])]; + tensor var_1781_to_fp16 = const()[name = tensor("op_1781_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1781_to_fp16, x = inputs_47_cast_fp16)[name = tensor("out_47_cast_fp16")]; + tensor input_75_gamma_0_to_fp16 = const()[name = tensor("input_75_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359075776)))]; + tensor input_75_beta_0_to_fp16 = const()[name = tensor("input_75_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359077888)))]; + tensor input_75_epsilon_0_to_fp16 = const()[name = tensor("input_75_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = batch_norm(beta = input_75_beta_0_to_fp16, epsilon = input_75_epsilon_0_to_fp16, gamma = input_75_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_pad_type_0 = const()[name = tensor("input_77_pad_type_0"), val = tensor("valid")]; + tensor input_77_strides_0 = const()[name = tensor("input_77_strides_0"), val = tensor([1, 1])]; + tensor input_77_pad_0 = const()[name = tensor("input_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_77_dilations_0 = const()[name = tensor("input_77_dilations_0"), val = tensor([1, 1])]; + tensor input_77_groups_0 = const()[name = tensor("input_77_groups_0"), val = tensor(1)]; + tensor layers_7_fc1_weight_to_fp16 = const()[name = tensor("layers_7_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359080000)))]; + tensor layers_7_fc1_bias_to_fp16 = const()[name = tensor("layers_7_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(367468672)))]; + tensor input_77_cast_fp16 = conv(bias = layers_7_fc1_bias_to_fp16, dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = layers_7_fc1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_mode_0 = const()[name = tensor("input_79_mode_0"), val = tensor("EXACT")]; + tensor input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor hidden_states_17_pad_type_0 = const()[name = tensor("hidden_states_17_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_17_strides_0 = const()[name = tensor("hidden_states_17_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_17_pad_0 = const()[name = tensor("hidden_states_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_17_dilations_0 = const()[name = tensor("hidden_states_17_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_17_groups_0 = const()[name = tensor("hidden_states_17_groups_0"), val = tensor(1)]; + tensor layers_7_fc2_weight_to_fp16 = const()[name = tensor("layers_7_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(367476928)))]; + tensor layers_7_fc2_bias_to_fp16 = const()[name = tensor("layers_7_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375865600)))]; + tensor hidden_states_17_cast_fp16 = conv(bias = layers_7_fc2_bias_to_fp16, dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = layers_7_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("hidden_states_17_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = hidden_states_17_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor var_1816 = const()[name = tensor("op_1816"), val = tensor(3)]; + tensor out_49_axes_0 = const()[name = tensor("out_49_axes_0"), val = tensor([1])]; + tensor var_1841_to_fp16 = const()[name = tensor("op_1841_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1841_to_fp16, x = inputs_49_cast_fp16)[name = tensor("out_49_cast_fp16")]; + tensor obj_113_gamma_0_to_fp16 = const()[name = tensor("obj_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375867712)))]; + tensor obj_113_beta_0_to_fp16 = const()[name = tensor("obj_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375869824)))]; + tensor obj_113_epsilon_0_to_fp16 = const()[name = tensor("obj_113_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_113_cast_fp16 = batch_norm(beta = obj_113_beta_0_to_fp16, epsilon = obj_113_epsilon_0_to_fp16, gamma = obj_113_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor("obj_113_cast_fp16")]; + tensor query_33_pad_type_0 = const()[name = tensor("query_33_pad_type_0"), val = tensor("valid")]; + tensor query_33_strides_0 = const()[name = tensor("query_33_strides_0"), val = tensor([1, 1])]; + tensor query_33_pad_0 = const()[name = tensor("query_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_33_dilations_0 = const()[name = tensor("query_33_dilations_0"), val = tensor([1, 1])]; + tensor query_33_groups_0 = const()[name = tensor("query_33_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375871936)))]; + tensor layers_8_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377969152)))]; + tensor query_33_cast_fp16 = conv(bias = layers_8_self_attn_q_proj_bias_to_fp16, dilations = query_33_dilations_0, groups = query_33_groups_0, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = query_33_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("query_33_cast_fp16")]; + tensor current_key_17_pad_type_0 = const()[name = tensor("current_key_17_pad_type_0"), val = tensor("valid")]; + tensor current_key_17_strides_0 = const()[name = tensor("current_key_17_strides_0"), val = tensor([1, 1])]; + tensor current_key_17_pad_0 = const()[name = tensor("current_key_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_17_dilations_0 = const()[name = tensor("current_key_17_dilations_0"), val = tensor([1, 1])]; + tensor current_key_17_groups_0 = const()[name = tensor("current_key_17_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377971264)))]; + tensor current_key_17_cast_fp16 = conv(dilations = current_key_17_dilations_0, groups = current_key_17_groups_0, pad = current_key_17_pad_0, pad_type = current_key_17_pad_type_0, strides = current_key_17_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("current_key_17_cast_fp16")]; + tensor current_value_17_pad_type_0 = const()[name = tensor("current_value_17_pad_type_0"), val = tensor("valid")]; + tensor current_value_17_strides_0 = const()[name = tensor("current_value_17_strides_0"), val = tensor([1, 1])]; + tensor current_value_17_pad_0 = const()[name = tensor("current_value_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_17_dilations_0 = const()[name = tensor("current_value_17_dilations_0"), val = tensor([1, 1])]; + tensor current_value_17_groups_0 = const()[name = tensor("current_value_17_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380068480)))]; + tensor layers_8_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(382165696)))]; + tensor current_value_17_cast_fp16 = conv(bias = layers_8_self_attn_v_proj_bias_to_fp16, dilations = current_value_17_dilations_0, groups = current_value_17_groups_0, pad = current_value_17_pad_0, pad_type = current_value_17_pad_type_0, strides = current_value_17_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = obj_113_cast_fp16)[name = tensor("current_value_17_cast_fp16")]; + tensor var_1879_cast_fp16 = mul(x = current_key_17_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1879_cast_fp16")]; + tensor key_33_cast_fp16 = add(x = var_87_cast_fp16_8, y = var_1879_cast_fp16)[name = tensor("key_33_cast_fp16")]; + tensor var_1881_cast_fp16 = mul(x = current_value_17_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor value_33_cast_fp16 = add(x = var_114_cast_fp16_8, y = var_1881_cast_fp16)[name = tensor("value_33_cast_fp16")]; + tensor var_1884 = const()[name = tensor("op_1884"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_33_cast_fp16 = reshape(shape = var_1884, x = query_33_cast_fp16)[name = tensor("mh_q_33_cast_fp16")]; + tensor var_1886_to_fp16 = const()[name = tensor("op_1886_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1887_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_1886_to_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor var_1888 = const()[name = tensor("op_1888"), val = tensor([1, 16, 64, -1])]; + tensor var_1889_cast_fp16 = reshape(shape = var_1888, x = key_33_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor mh_w_49_transpose_x_0 = const()[name = tensor("mh_w_49_transpose_x_0"), val = tensor(true)]; + tensor mh_w_49_transpose_y_0 = const()[name = tensor("mh_w_49_transpose_y_0"), val = tensor(false)]; + tensor mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_1887_cast_fp16, y = var_1889_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; + tensor mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; + tensor var_1897_cast_fp16 = softmax(axis = var_1816, x = mh_w_51_cast_fp16)[name = tensor("op_1897_cast_fp16")]; + tensor var_1898 = const()[name = tensor("op_1898"), val = tensor([1, 16, 64, -1])]; + tensor var_1899_cast_fp16 = reshape(shape = var_1898, x = value_33_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(true)]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_1899_cast_fp16, y = var_1897_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_1902 = const()[name = tensor("op_1902"), val = tensor([1, 1024, 1, -1])]; + tensor input_81_cast_fp16 = reshape(shape = var_1902, x = attn_33_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor obj_119_pad_type_0 = const()[name = tensor("obj_119_pad_type_0"), val = tensor("valid")]; + tensor obj_119_strides_0 = const()[name = tensor("obj_119_strides_0"), val = tensor([1, 1])]; + tensor obj_119_pad_0 = const()[name = tensor("obj_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_119_dilations_0 = const()[name = tensor("obj_119_dilations_0"), val = tensor([1, 1])]; + tensor obj_119_groups_0 = const()[name = tensor("obj_119_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(382167808)))]; + tensor layers_8_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_8_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384265024)))]; + tensor obj_119_cast_fp16 = conv(bias = layers_8_self_attn_o_proj_bias_to_fp16, dilations = obj_119_dilations_0, groups = obj_119_groups_0, pad = obj_119_pad_0, pad_type = obj_119_pad_type_0, strides = obj_119_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("obj_119_cast_fp16")]; + tensor inputs_51_cast_fp16 = add(x = inputs_49_cast_fp16, y = obj_119_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor out_51_axes_0 = const()[name = tensor("out_51_axes_0"), val = tensor([1])]; + tensor var_1924_to_fp16 = const()[name = tensor("op_1924_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_1924_to_fp16, x = inputs_51_cast_fp16)[name = tensor("out_51_cast_fp16")]; + tensor obj_121_gamma_0_to_fp16 = const()[name = tensor("obj_121_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384267136)))]; + tensor obj_121_beta_0_to_fp16 = const()[name = tensor("obj_121_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384269248)))]; + tensor obj_121_epsilon_0_to_fp16 = const()[name = tensor("obj_121_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_121_cast_fp16 = batch_norm(beta = obj_121_beta_0_to_fp16, epsilon = obj_121_epsilon_0_to_fp16, gamma = obj_121_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor("obj_121_cast_fp16")]; + tensor query_35_pad_type_0 = const()[name = tensor("query_35_pad_type_0"), val = tensor("valid")]; + tensor query_35_strides_0 = const()[name = tensor("query_35_strides_0"), val = tensor([1, 1])]; + tensor query_35_pad_0 = const()[name = tensor("query_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_35_dilations_0 = const()[name = tensor("query_35_dilations_0"), val = tensor([1, 1])]; + tensor query_35_groups_0 = const()[name = tensor("query_35_groups_0"), val = tensor(1)]; + tensor layers_8_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384271360)))]; + tensor layers_8_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386368576)))]; + tensor query_35_cast_fp16 = conv(bias = layers_8_encoder_attn_q_proj_bias_to_fp16, dilations = query_35_dilations_0, groups = query_35_groups_0, pad = query_35_pad_0, pad_type = query_35_pad_type_0, strides = query_35_strides_0, weight = layers_8_encoder_attn_q_proj_weight_to_fp16, x = obj_121_cast_fp16)[name = tensor("query_35_cast_fp16")]; + tensor key_35_pad_type_0 = const()[name = tensor("key_35_pad_type_0"), val = tensor("valid")]; + tensor key_35_strides_0 = const()[name = tensor("key_35_strides_0"), val = tensor([1, 1])]; + tensor key_35_pad_0 = const()[name = tensor("key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_35_dilations_0 = const()[name = tensor("key_35_dilations_0"), val = tensor([1, 1])]; + tensor key_35_groups_0 = const()[name = tensor("key_35_groups_0"), val = tensor(1)]; + tensor layers_8_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386370688)))]; + tensor key_35_cast_fp16 = conv(dilations = key_35_dilations_0, groups = key_35_groups_0, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = key_35_strides_0, weight = layers_8_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_35_cast_fp16")]; + tensor value_35_pad_type_0 = const()[name = tensor("value_35_pad_type_0"), val = tensor("valid")]; + tensor value_35_strides_0 = const()[name = tensor("value_35_strides_0"), val = tensor([1, 1])]; + tensor value_35_pad_0 = const()[name = tensor("value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_35_dilations_0 = const()[name = tensor("value_35_dilations_0"), val = tensor([1, 1])]; + tensor value_35_groups_0 = const()[name = tensor("value_35_groups_0"), val = tensor(1)]; + tensor layers_8_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388467904)))]; + tensor layers_8_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390565120)))]; + tensor value_35_cast_fp16 = conv(bias = layers_8_encoder_attn_v_proj_bias_to_fp16, dilations = value_35_dilations_0, groups = value_35_groups_0, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = value_35_strides_0, weight = layers_8_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_35_cast_fp16")]; + tensor var_1959 = const()[name = tensor("op_1959"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_35_cast_fp16 = reshape(shape = var_1959, x = query_35_cast_fp16)[name = tensor("mh_q_35_cast_fp16")]; + tensor var_1961_to_fp16 = const()[name = tensor("op_1961_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1962_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_1961_to_fp16)[name = tensor("op_1962_cast_fp16")]; + tensor var_1963 = const()[name = tensor("op_1963"), val = tensor([1, 16, 64, -1])]; + tensor var_1964_cast_fp16 = reshape(shape = var_1963, x = key_35_cast_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor mh_w_53_transpose_x_0 = const()[name = tensor("mh_w_53_transpose_x_0"), val = tensor(true)]; + tensor mh_w_53_transpose_y_0 = const()[name = tensor("mh_w_53_transpose_y_0"), val = tensor(false)]; + tensor mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_1962_cast_fp16, y = var_1964_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; + tensor obj_125_cast_fp16 = softmax(axis = var_1816, x = mh_w_53_cast_fp16)[name = tensor("obj_125_cast_fp16")]; + tensor var_1968 = const()[name = tensor("op_1968"), val = tensor([1, 16, 64, -1])]; + tensor var_1969_cast_fp16 = reshape(shape = var_1968, x = value_35_cast_fp16)[name = tensor("op_1969_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(true)]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_1969_cast_fp16, y = obj_125_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_1972 = const()[name = tensor("op_1972"), val = tensor([1, 1024, 1, -1])]; + tensor input_83_cast_fp16 = reshape(shape = var_1972, x = attn_35_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor obj_123_pad_type_0 = const()[name = tensor("obj_123_pad_type_0"), val = tensor("valid")]; + tensor obj_123_strides_0 = const()[name = tensor("obj_123_strides_0"), val = tensor([1, 1])]; + tensor obj_123_pad_0 = const()[name = tensor("obj_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_123_dilations_0 = const()[name = tensor("obj_123_dilations_0"), val = tensor([1, 1])]; + tensor obj_123_groups_0 = const()[name = tensor("obj_123_groups_0"), val = tensor(1)]; + tensor layers_8_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390567232)))]; + tensor layers_8_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_8_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392664448)))]; + tensor obj_123_cast_fp16 = conv(bias = layers_8_encoder_attn_o_proj_bias_to_fp16, dilations = obj_123_dilations_0, groups = obj_123_groups_0, pad = obj_123_pad_0, pad_type = obj_123_pad_type_0, strides = obj_123_strides_0, weight = layers_8_encoder_attn_o_proj_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("obj_123_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = obj_123_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor out_53_axes_0 = const()[name = tensor("out_53_axes_0"), val = tensor([1])]; + tensor var_1990_to_fp16 = const()[name = tensor("op_1990_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_1990_to_fp16, x = inputs_53_cast_fp16)[name = tensor("out_53_cast_fp16")]; + tensor input_85_gamma_0_to_fp16 = const()[name = tensor("input_85_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392666560)))]; + tensor input_85_beta_0_to_fp16 = const()[name = tensor("input_85_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392668672)))]; + tensor input_85_epsilon_0_to_fp16 = const()[name = tensor("input_85_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_85_cast_fp16 = batch_norm(beta = input_85_beta_0_to_fp16, epsilon = input_85_epsilon_0_to_fp16, gamma = input_85_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_pad_type_0 = const()[name = tensor("input_87_pad_type_0"), val = tensor("valid")]; + tensor input_87_strides_0 = const()[name = tensor("input_87_strides_0"), val = tensor([1, 1])]; + tensor input_87_pad_0 = const()[name = tensor("input_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_87_dilations_0 = const()[name = tensor("input_87_dilations_0"), val = tensor([1, 1])]; + tensor input_87_groups_0 = const()[name = tensor("input_87_groups_0"), val = tensor(1)]; + tensor layers_8_fc1_weight_to_fp16 = const()[name = tensor("layers_8_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392670784)))]; + tensor layers_8_fc1_bias_to_fp16 = const()[name = tensor("layers_8_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(401059456)))]; + tensor input_87_cast_fp16 = conv(bias = layers_8_fc1_bias_to_fp16, dilations = input_87_dilations_0, groups = input_87_groups_0, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = input_87_strides_0, weight = layers_8_fc1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_mode_0 = const()[name = tensor("input_89_mode_0"), val = tensor("EXACT")]; + tensor input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor hidden_states_19_pad_type_0 = const()[name = tensor("hidden_states_19_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_19_strides_0 = const()[name = tensor("hidden_states_19_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_19_pad_0 = const()[name = tensor("hidden_states_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_19_dilations_0 = const()[name = tensor("hidden_states_19_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_19_groups_0 = const()[name = tensor("hidden_states_19_groups_0"), val = tensor(1)]; + tensor layers_8_fc2_weight_to_fp16 = const()[name = tensor("layers_8_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(401067712)))]; + tensor layers_8_fc2_bias_to_fp16 = const()[name = tensor("layers_8_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409456384)))]; + tensor hidden_states_19_cast_fp16 = conv(bias = layers_8_fc2_bias_to_fp16, dilations = hidden_states_19_dilations_0, groups = hidden_states_19_groups_0, pad = hidden_states_19_pad_0, pad_type = hidden_states_19_pad_type_0, strides = hidden_states_19_strides_0, weight = layers_8_fc2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("hidden_states_19_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = hidden_states_19_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor var_2025 = const()[name = tensor("op_2025"), val = tensor(3)]; + tensor out_55_axes_0 = const()[name = tensor("out_55_axes_0"), val = tensor([1])]; + tensor var_2050_to_fp16 = const()[name = tensor("op_2050_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_2050_to_fp16, x = inputs_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; + tensor obj_127_gamma_0_to_fp16 = const()[name = tensor("obj_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409458496)))]; + tensor obj_127_beta_0_to_fp16 = const()[name = tensor("obj_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409460608)))]; + tensor obj_127_epsilon_0_to_fp16 = const()[name = tensor("obj_127_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_127_cast_fp16 = batch_norm(beta = obj_127_beta_0_to_fp16, epsilon = obj_127_epsilon_0_to_fp16, gamma = obj_127_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor("obj_127_cast_fp16")]; + tensor query_37_pad_type_0 = const()[name = tensor("query_37_pad_type_0"), val = tensor("valid")]; + tensor query_37_strides_0 = const()[name = tensor("query_37_strides_0"), val = tensor([1, 1])]; + tensor query_37_pad_0 = const()[name = tensor("query_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_37_dilations_0 = const()[name = tensor("query_37_dilations_0"), val = tensor([1, 1])]; + tensor query_37_groups_0 = const()[name = tensor("query_37_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409462720)))]; + tensor layers_9_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411559936)))]; + tensor query_37_cast_fp16 = conv(bias = layers_9_self_attn_q_proj_bias_to_fp16, dilations = query_37_dilations_0, groups = query_37_groups_0, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = query_37_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("query_37_cast_fp16")]; + tensor current_key_19_pad_type_0 = const()[name = tensor("current_key_19_pad_type_0"), val = tensor("valid")]; + tensor current_key_19_strides_0 = const()[name = tensor("current_key_19_strides_0"), val = tensor([1, 1])]; + tensor current_key_19_pad_0 = const()[name = tensor("current_key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_19_dilations_0 = const()[name = tensor("current_key_19_dilations_0"), val = tensor([1, 1])]; + tensor current_key_19_groups_0 = const()[name = tensor("current_key_19_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411562048)))]; + tensor current_key_19_cast_fp16 = conv(dilations = current_key_19_dilations_0, groups = current_key_19_groups_0, pad = current_key_19_pad_0, pad_type = current_key_19_pad_type_0, strides = current_key_19_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("current_key_19_cast_fp16")]; + tensor current_value_19_pad_type_0 = const()[name = tensor("current_value_19_pad_type_0"), val = tensor("valid")]; + tensor current_value_19_strides_0 = const()[name = tensor("current_value_19_strides_0"), val = tensor([1, 1])]; + tensor current_value_19_pad_0 = const()[name = tensor("current_value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_19_dilations_0 = const()[name = tensor("current_value_19_dilations_0"), val = tensor([1, 1])]; + tensor current_value_19_groups_0 = const()[name = tensor("current_value_19_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413659264)))]; + tensor layers_9_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415756480)))]; + tensor current_value_19_cast_fp16 = conv(bias = layers_9_self_attn_v_proj_bias_to_fp16, dilations = current_value_19_dilations_0, groups = current_value_19_groups_0, pad = current_value_19_pad_0, pad_type = current_value_19_pad_type_0, strides = current_value_19_strides_0, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = obj_127_cast_fp16)[name = tensor("current_value_19_cast_fp16")]; + tensor var_2088_cast_fp16 = mul(x = current_key_19_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2088_cast_fp16")]; + tensor key_37_cast_fp16 = add(x = var_87_cast_fp16_9, y = var_2088_cast_fp16)[name = tensor("key_37_cast_fp16")]; + tensor var_2090_cast_fp16 = mul(x = current_value_19_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2090_cast_fp16")]; + tensor value_37_cast_fp16 = add(x = var_114_cast_fp16_9, y = var_2090_cast_fp16)[name = tensor("value_37_cast_fp16")]; + tensor var_2093 = const()[name = tensor("op_2093"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_37_cast_fp16 = reshape(shape = var_2093, x = query_37_cast_fp16)[name = tensor("mh_q_37_cast_fp16")]; + tensor var_2095_to_fp16 = const()[name = tensor("op_2095_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2096_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_2095_to_fp16)[name = tensor("op_2096_cast_fp16")]; + tensor var_2097 = const()[name = tensor("op_2097"), val = tensor([1, 16, 64, -1])]; + tensor var_2098_cast_fp16 = reshape(shape = var_2097, x = key_37_cast_fp16)[name = tensor("op_2098_cast_fp16")]; + tensor mh_w_55_transpose_x_0 = const()[name = tensor("mh_w_55_transpose_x_0"), val = tensor(true)]; + tensor mh_w_55_transpose_y_0 = const()[name = tensor("mh_w_55_transpose_y_0"), val = tensor(false)]; + tensor mh_w_55_cast_fp16 = matmul(transpose_x = mh_w_55_transpose_x_0, transpose_y = mh_w_55_transpose_y_0, x = var_2096_cast_fp16, y = var_2098_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; + tensor mh_w_57_cast_fp16 = add(x = mh_w_55_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; + tensor var_2106_cast_fp16 = softmax(axis = var_2025, x = mh_w_57_cast_fp16)[name = tensor("op_2106_cast_fp16")]; + tensor var_2107 = const()[name = tensor("op_2107"), val = tensor([1, 16, 64, -1])]; + tensor var_2108_cast_fp16 = reshape(shape = var_2107, x = value_37_cast_fp16)[name = tensor("op_2108_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(true)]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_2108_cast_fp16, y = var_2106_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_2111 = const()[name = tensor("op_2111"), val = tensor([1, 1024, 1, -1])]; + tensor input_91_cast_fp16 = reshape(shape = var_2111, x = attn_37_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor obj_133_pad_type_0 = const()[name = tensor("obj_133_pad_type_0"), val = tensor("valid")]; + tensor obj_133_strides_0 = const()[name = tensor("obj_133_strides_0"), val = tensor([1, 1])]; + tensor obj_133_pad_0 = const()[name = tensor("obj_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_133_dilations_0 = const()[name = tensor("obj_133_dilations_0"), val = tensor([1, 1])]; + tensor obj_133_groups_0 = const()[name = tensor("obj_133_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(415758592)))]; + tensor layers_9_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_9_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417855808)))]; + tensor obj_133_cast_fp16 = conv(bias = layers_9_self_attn_o_proj_bias_to_fp16, dilations = obj_133_dilations_0, groups = obj_133_groups_0, pad = obj_133_pad_0, pad_type = obj_133_pad_type_0, strides = obj_133_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("obj_133_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = obj_133_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor out_57_axes_0 = const()[name = tensor("out_57_axes_0"), val = tensor([1])]; + tensor var_2133_to_fp16 = const()[name = tensor("op_2133_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_2133_to_fp16, x = inputs_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; + tensor obj_135_gamma_0_to_fp16 = const()[name = tensor("obj_135_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417857920)))]; + tensor obj_135_beta_0_to_fp16 = const()[name = tensor("obj_135_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417860032)))]; + tensor obj_135_epsilon_0_to_fp16 = const()[name = tensor("obj_135_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_135_cast_fp16 = batch_norm(beta = obj_135_beta_0_to_fp16, epsilon = obj_135_epsilon_0_to_fp16, gamma = obj_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor("obj_135_cast_fp16")]; + tensor query_39_pad_type_0 = const()[name = tensor("query_39_pad_type_0"), val = tensor("valid")]; + tensor query_39_strides_0 = const()[name = tensor("query_39_strides_0"), val = tensor([1, 1])]; + tensor query_39_pad_0 = const()[name = tensor("query_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_39_dilations_0 = const()[name = tensor("query_39_dilations_0"), val = tensor([1, 1])]; + tensor query_39_groups_0 = const()[name = tensor("query_39_groups_0"), val = tensor(1)]; + tensor layers_9_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(417862144)))]; + tensor layers_9_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419959360)))]; + tensor query_39_cast_fp16 = conv(bias = layers_9_encoder_attn_q_proj_bias_to_fp16, dilations = query_39_dilations_0, groups = query_39_groups_0, pad = query_39_pad_0, pad_type = query_39_pad_type_0, strides = query_39_strides_0, weight = layers_9_encoder_attn_q_proj_weight_to_fp16, x = obj_135_cast_fp16)[name = tensor("query_39_cast_fp16")]; + tensor key_39_pad_type_0 = const()[name = tensor("key_39_pad_type_0"), val = tensor("valid")]; + tensor key_39_strides_0 = const()[name = tensor("key_39_strides_0"), val = tensor([1, 1])]; + tensor key_39_pad_0 = const()[name = tensor("key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_39_dilations_0 = const()[name = tensor("key_39_dilations_0"), val = tensor([1, 1])]; + tensor key_39_groups_0 = const()[name = tensor("key_39_groups_0"), val = tensor(1)]; + tensor layers_9_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419961472)))]; + tensor key_39_cast_fp16 = conv(dilations = key_39_dilations_0, groups = key_39_groups_0, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = key_39_strides_0, weight = layers_9_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_39_cast_fp16")]; + tensor value_39_pad_type_0 = const()[name = tensor("value_39_pad_type_0"), val = tensor("valid")]; + tensor value_39_strides_0 = const()[name = tensor("value_39_strides_0"), val = tensor([1, 1])]; + tensor value_39_pad_0 = const()[name = tensor("value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_39_dilations_0 = const()[name = tensor("value_39_dilations_0"), val = tensor([1, 1])]; + tensor value_39_groups_0 = const()[name = tensor("value_39_groups_0"), val = tensor(1)]; + tensor layers_9_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422058688)))]; + tensor layers_9_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424155904)))]; + tensor value_39_cast_fp16 = conv(bias = layers_9_encoder_attn_v_proj_bias_to_fp16, dilations = value_39_dilations_0, groups = value_39_groups_0, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = value_39_strides_0, weight = layers_9_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_39_cast_fp16")]; + tensor var_2168 = const()[name = tensor("op_2168"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_39_cast_fp16 = reshape(shape = var_2168, x = query_39_cast_fp16)[name = tensor("mh_q_39_cast_fp16")]; + tensor var_2170_to_fp16 = const()[name = tensor("op_2170_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2171_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_2170_to_fp16)[name = tensor("op_2171_cast_fp16")]; + tensor var_2172 = const()[name = tensor("op_2172"), val = tensor([1, 16, 64, -1])]; + tensor var_2173_cast_fp16 = reshape(shape = var_2172, x = key_39_cast_fp16)[name = tensor("op_2173_cast_fp16")]; + tensor mh_w_59_transpose_x_0 = const()[name = tensor("mh_w_59_transpose_x_0"), val = tensor(true)]; + tensor mh_w_59_transpose_y_0 = const()[name = tensor("mh_w_59_transpose_y_0"), val = tensor(false)]; + tensor mh_w_59_cast_fp16 = matmul(transpose_x = mh_w_59_transpose_x_0, transpose_y = mh_w_59_transpose_y_0, x = var_2171_cast_fp16, y = var_2173_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; + tensor obj_139_cast_fp16 = softmax(axis = var_2025, x = mh_w_59_cast_fp16)[name = tensor("obj_139_cast_fp16")]; + tensor var_2177 = const()[name = tensor("op_2177"), val = tensor([1, 16, 64, -1])]; + tensor var_2178_cast_fp16 = reshape(shape = var_2177, x = value_39_cast_fp16)[name = tensor("op_2178_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(true)]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_2178_cast_fp16, y = obj_139_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_2181 = const()[name = tensor("op_2181"), val = tensor([1, 1024, 1, -1])]; + tensor input_93_cast_fp16 = reshape(shape = var_2181, x = attn_39_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor obj_137_pad_type_0 = const()[name = tensor("obj_137_pad_type_0"), val = tensor("valid")]; + tensor obj_137_strides_0 = const()[name = tensor("obj_137_strides_0"), val = tensor([1, 1])]; + tensor obj_137_pad_0 = const()[name = tensor("obj_137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_137_dilations_0 = const()[name = tensor("obj_137_dilations_0"), val = tensor([1, 1])]; + tensor obj_137_groups_0 = const()[name = tensor("obj_137_groups_0"), val = tensor(1)]; + tensor layers_9_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(424158016)))]; + tensor layers_9_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_9_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426255232)))]; + tensor obj_137_cast_fp16 = conv(bias = layers_9_encoder_attn_o_proj_bias_to_fp16, dilations = obj_137_dilations_0, groups = obj_137_groups_0, pad = obj_137_pad_0, pad_type = obj_137_pad_type_0, strides = obj_137_strides_0, weight = layers_9_encoder_attn_o_proj_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("obj_137_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = obj_137_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor out_59_axes_0 = const()[name = tensor("out_59_axes_0"), val = tensor([1])]; + tensor var_2199_to_fp16 = const()[name = tensor("op_2199_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_2199_to_fp16, x = inputs_59_cast_fp16)[name = tensor("out_59_cast_fp16")]; + tensor input_95_gamma_0_to_fp16 = const()[name = tensor("input_95_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426257344)))]; + tensor input_95_beta_0_to_fp16 = const()[name = tensor("input_95_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426259456)))]; + tensor input_95_epsilon_0_to_fp16 = const()[name = tensor("input_95_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = batch_norm(beta = input_95_beta_0_to_fp16, epsilon = input_95_epsilon_0_to_fp16, gamma = input_95_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor input_97_pad_type_0 = const()[name = tensor("input_97_pad_type_0"), val = tensor("valid")]; + tensor input_97_strides_0 = const()[name = tensor("input_97_strides_0"), val = tensor([1, 1])]; + tensor input_97_pad_0 = const()[name = tensor("input_97_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_97_dilations_0 = const()[name = tensor("input_97_dilations_0"), val = tensor([1, 1])]; + tensor input_97_groups_0 = const()[name = tensor("input_97_groups_0"), val = tensor(1)]; + tensor layers_9_fc1_weight_to_fp16 = const()[name = tensor("layers_9_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426261568)))]; + tensor layers_9_fc1_bias_to_fp16 = const()[name = tensor("layers_9_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434650240)))]; + tensor input_97_cast_fp16 = conv(bias = layers_9_fc1_bias_to_fp16, dilations = input_97_dilations_0, groups = input_97_groups_0, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = input_97_strides_0, weight = layers_9_fc1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_mode_0 = const()[name = tensor("input_99_mode_0"), val = tensor("EXACT")]; + tensor input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor hidden_states_21_pad_type_0 = const()[name = tensor("hidden_states_21_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_21_strides_0 = const()[name = tensor("hidden_states_21_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_21_pad_0 = const()[name = tensor("hidden_states_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_21_dilations_0 = const()[name = tensor("hidden_states_21_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_21_groups_0 = const()[name = tensor("hidden_states_21_groups_0"), val = tensor(1)]; + tensor layers_9_fc2_weight_to_fp16 = const()[name = tensor("layers_9_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434658496)))]; + tensor layers_9_fc2_bias_to_fp16 = const()[name = tensor("layers_9_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(443047168)))]; + tensor hidden_states_21_cast_fp16 = conv(bias = layers_9_fc2_bias_to_fp16, dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = layers_9_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("hidden_states_21_cast_fp16")]; + tensor inputs_61_cast_fp16 = add(x = inputs_59_cast_fp16, y = hidden_states_21_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_2234 = const()[name = tensor("op_2234"), val = tensor(3)]; + tensor out_61_axes_0 = const()[name = tensor("out_61_axes_0"), val = tensor([1])]; + tensor var_2259_to_fp16 = const()[name = tensor("op_2259_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_2259_to_fp16, x = inputs_61_cast_fp16)[name = tensor("out_61_cast_fp16")]; + tensor obj_141_gamma_0_to_fp16 = const()[name = tensor("obj_141_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(443049280)))]; + tensor obj_141_beta_0_to_fp16 = const()[name = tensor("obj_141_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(443051392)))]; + tensor obj_141_epsilon_0_to_fp16 = const()[name = tensor("obj_141_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_141_cast_fp16 = batch_norm(beta = obj_141_beta_0_to_fp16, epsilon = obj_141_epsilon_0_to_fp16, gamma = obj_141_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor("obj_141_cast_fp16")]; + tensor query_41_pad_type_0 = const()[name = tensor("query_41_pad_type_0"), val = tensor("valid")]; + tensor query_41_strides_0 = const()[name = tensor("query_41_strides_0"), val = tensor([1, 1])]; + tensor query_41_pad_0 = const()[name = tensor("query_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_41_dilations_0 = const()[name = tensor("query_41_dilations_0"), val = tensor([1, 1])]; + tensor query_41_groups_0 = const()[name = tensor("query_41_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(443053504)))]; + tensor layers_10_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445150720)))]; + tensor query_41_cast_fp16 = conv(bias = layers_10_self_attn_q_proj_bias_to_fp16, dilations = query_41_dilations_0, groups = query_41_groups_0, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = query_41_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("query_41_cast_fp16")]; + tensor current_key_21_pad_type_0 = const()[name = tensor("current_key_21_pad_type_0"), val = tensor("valid")]; + tensor current_key_21_strides_0 = const()[name = tensor("current_key_21_strides_0"), val = tensor([1, 1])]; + tensor current_key_21_pad_0 = const()[name = tensor("current_key_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_21_dilations_0 = const()[name = tensor("current_key_21_dilations_0"), val = tensor([1, 1])]; + tensor current_key_21_groups_0 = const()[name = tensor("current_key_21_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445152832)))]; + tensor current_key_21_cast_fp16 = conv(dilations = current_key_21_dilations_0, groups = current_key_21_groups_0, pad = current_key_21_pad_0, pad_type = current_key_21_pad_type_0, strides = current_key_21_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("current_key_21_cast_fp16")]; + tensor current_value_21_pad_type_0 = const()[name = tensor("current_value_21_pad_type_0"), val = tensor("valid")]; + tensor current_value_21_strides_0 = const()[name = tensor("current_value_21_strides_0"), val = tensor([1, 1])]; + tensor current_value_21_pad_0 = const()[name = tensor("current_value_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_21_dilations_0 = const()[name = tensor("current_value_21_dilations_0"), val = tensor([1, 1])]; + tensor current_value_21_groups_0 = const()[name = tensor("current_value_21_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447250048)))]; + tensor layers_10_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449347264)))]; + tensor current_value_21_cast_fp16 = conv(bias = layers_10_self_attn_v_proj_bias_to_fp16, dilations = current_value_21_dilations_0, groups = current_value_21_groups_0, pad = current_value_21_pad_0, pad_type = current_value_21_pad_type_0, strides = current_value_21_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = obj_141_cast_fp16)[name = tensor("current_value_21_cast_fp16")]; + tensor var_2297_cast_fp16 = mul(x = current_key_21_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2297_cast_fp16")]; + tensor key_41_cast_fp16 = add(x = var_87_cast_fp16_10, y = var_2297_cast_fp16)[name = tensor("key_41_cast_fp16")]; + tensor var_2299_cast_fp16 = mul(x = current_value_21_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2299_cast_fp16")]; + tensor value_41_cast_fp16 = add(x = var_114_cast_fp16_10, y = var_2299_cast_fp16)[name = tensor("value_41_cast_fp16")]; + tensor var_2302 = const()[name = tensor("op_2302"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_41_cast_fp16 = reshape(shape = var_2302, x = query_41_cast_fp16)[name = tensor("mh_q_41_cast_fp16")]; + tensor var_2304_to_fp16 = const()[name = tensor("op_2304_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2305_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_2304_to_fp16)[name = tensor("op_2305_cast_fp16")]; + tensor var_2306 = const()[name = tensor("op_2306"), val = tensor([1, 16, 64, -1])]; + tensor var_2307_cast_fp16 = reshape(shape = var_2306, x = key_41_cast_fp16)[name = tensor("op_2307_cast_fp16")]; + tensor mh_w_61_transpose_x_0 = const()[name = tensor("mh_w_61_transpose_x_0"), val = tensor(true)]; + tensor mh_w_61_transpose_y_0 = const()[name = tensor("mh_w_61_transpose_y_0"), val = tensor(false)]; + tensor mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_2305_cast_fp16, y = var_2307_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; + tensor mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_63_cast_fp16")]; + tensor var_2315_cast_fp16 = softmax(axis = var_2234, x = mh_w_63_cast_fp16)[name = tensor("op_2315_cast_fp16")]; + tensor var_2316 = const()[name = tensor("op_2316"), val = tensor([1, 16, 64, -1])]; + tensor var_2317_cast_fp16 = reshape(shape = var_2316, x = value_41_cast_fp16)[name = tensor("op_2317_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(true)]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_2317_cast_fp16, y = var_2315_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_2320 = const()[name = tensor("op_2320"), val = tensor([1, 1024, 1, -1])]; + tensor input_101_cast_fp16 = reshape(shape = var_2320, x = attn_41_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor obj_147_pad_type_0 = const()[name = tensor("obj_147_pad_type_0"), val = tensor("valid")]; + tensor obj_147_strides_0 = const()[name = tensor("obj_147_strides_0"), val = tensor([1, 1])]; + tensor obj_147_pad_0 = const()[name = tensor("obj_147_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_147_dilations_0 = const()[name = tensor("obj_147_dilations_0"), val = tensor([1, 1])]; + tensor obj_147_groups_0 = const()[name = tensor("obj_147_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(449349376)))]; + tensor layers_10_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_10_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451446592)))]; + tensor obj_147_cast_fp16 = conv(bias = layers_10_self_attn_o_proj_bias_to_fp16, dilations = obj_147_dilations_0, groups = obj_147_groups_0, pad = obj_147_pad_0, pad_type = obj_147_pad_type_0, strides = obj_147_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("obj_147_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = obj_147_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor out_63_axes_0 = const()[name = tensor("out_63_axes_0"), val = tensor([1])]; + tensor var_2342_to_fp16 = const()[name = tensor("op_2342_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_2342_to_fp16, x = inputs_63_cast_fp16)[name = tensor("out_63_cast_fp16")]; + tensor obj_149_gamma_0_to_fp16 = const()[name = tensor("obj_149_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451448704)))]; + tensor obj_149_beta_0_to_fp16 = const()[name = tensor("obj_149_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451450816)))]; + tensor obj_149_epsilon_0_to_fp16 = const()[name = tensor("obj_149_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_149_cast_fp16 = batch_norm(beta = obj_149_beta_0_to_fp16, epsilon = obj_149_epsilon_0_to_fp16, gamma = obj_149_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor("obj_149_cast_fp16")]; + tensor query_43_pad_type_0 = const()[name = tensor("query_43_pad_type_0"), val = tensor("valid")]; + tensor query_43_strides_0 = const()[name = tensor("query_43_strides_0"), val = tensor([1, 1])]; + tensor query_43_pad_0 = const()[name = tensor("query_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_43_dilations_0 = const()[name = tensor("query_43_dilations_0"), val = tensor([1, 1])]; + tensor query_43_groups_0 = const()[name = tensor("query_43_groups_0"), val = tensor(1)]; + tensor layers_10_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451452928)))]; + tensor layers_10_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453550144)))]; + tensor query_43_cast_fp16 = conv(bias = layers_10_encoder_attn_q_proj_bias_to_fp16, dilations = query_43_dilations_0, groups = query_43_groups_0, pad = query_43_pad_0, pad_type = query_43_pad_type_0, strides = query_43_strides_0, weight = layers_10_encoder_attn_q_proj_weight_to_fp16, x = obj_149_cast_fp16)[name = tensor("query_43_cast_fp16")]; + tensor key_43_pad_type_0 = const()[name = tensor("key_43_pad_type_0"), val = tensor("valid")]; + tensor key_43_strides_0 = const()[name = tensor("key_43_strides_0"), val = tensor([1, 1])]; + tensor key_43_pad_0 = const()[name = tensor("key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_43_dilations_0 = const()[name = tensor("key_43_dilations_0"), val = tensor([1, 1])]; + tensor key_43_groups_0 = const()[name = tensor("key_43_groups_0"), val = tensor(1)]; + tensor layers_10_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453552256)))]; + tensor key_43_cast_fp16 = conv(dilations = key_43_dilations_0, groups = key_43_groups_0, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = key_43_strides_0, weight = layers_10_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_43_cast_fp16")]; + tensor value_43_pad_type_0 = const()[name = tensor("value_43_pad_type_0"), val = tensor("valid")]; + tensor value_43_strides_0 = const()[name = tensor("value_43_strides_0"), val = tensor([1, 1])]; + tensor value_43_pad_0 = const()[name = tensor("value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_43_dilations_0 = const()[name = tensor("value_43_dilations_0"), val = tensor([1, 1])]; + tensor value_43_groups_0 = const()[name = tensor("value_43_groups_0"), val = tensor(1)]; + tensor layers_10_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(455649472)))]; + tensor layers_10_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457746688)))]; + tensor value_43_cast_fp16 = conv(bias = layers_10_encoder_attn_v_proj_bias_to_fp16, dilations = value_43_dilations_0, groups = value_43_groups_0, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = value_43_strides_0, weight = layers_10_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_43_cast_fp16")]; + tensor var_2377 = const()[name = tensor("op_2377"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_43_cast_fp16 = reshape(shape = var_2377, x = query_43_cast_fp16)[name = tensor("mh_q_43_cast_fp16")]; + tensor var_2379_to_fp16 = const()[name = tensor("op_2379_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2380_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_2379_to_fp16)[name = tensor("op_2380_cast_fp16")]; + tensor var_2381 = const()[name = tensor("op_2381"), val = tensor([1, 16, 64, -1])]; + tensor var_2382_cast_fp16 = reshape(shape = var_2381, x = key_43_cast_fp16)[name = tensor("op_2382_cast_fp16")]; + tensor mh_w_65_transpose_x_0 = const()[name = tensor("mh_w_65_transpose_x_0"), val = tensor(true)]; + tensor mh_w_65_transpose_y_0 = const()[name = tensor("mh_w_65_transpose_y_0"), val = tensor(false)]; + tensor mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_2380_cast_fp16, y = var_2382_cast_fp16)[name = tensor("mh_w_65_cast_fp16")]; + tensor obj_153_cast_fp16 = softmax(axis = var_2234, x = mh_w_65_cast_fp16)[name = tensor("obj_153_cast_fp16")]; + tensor var_2386 = const()[name = tensor("op_2386"), val = tensor([1, 16, 64, -1])]; + tensor var_2387_cast_fp16 = reshape(shape = var_2386, x = value_43_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(true)]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_2387_cast_fp16, y = obj_153_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_2390 = const()[name = tensor("op_2390"), val = tensor([1, 1024, 1, -1])]; + tensor input_103_cast_fp16 = reshape(shape = var_2390, x = attn_43_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor obj_151_pad_type_0 = const()[name = tensor("obj_151_pad_type_0"), val = tensor("valid")]; + tensor obj_151_strides_0 = const()[name = tensor("obj_151_strides_0"), val = tensor([1, 1])]; + tensor obj_151_pad_0 = const()[name = tensor("obj_151_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_151_dilations_0 = const()[name = tensor("obj_151_dilations_0"), val = tensor([1, 1])]; + tensor obj_151_groups_0 = const()[name = tensor("obj_151_groups_0"), val = tensor(1)]; + tensor layers_10_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457748800)))]; + tensor layers_10_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_10_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459846016)))]; + tensor obj_151_cast_fp16 = conv(bias = layers_10_encoder_attn_o_proj_bias_to_fp16, dilations = obj_151_dilations_0, groups = obj_151_groups_0, pad = obj_151_pad_0, pad_type = obj_151_pad_type_0, strides = obj_151_strides_0, weight = layers_10_encoder_attn_o_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("obj_151_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_151_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor out_65_axes_0 = const()[name = tensor("out_65_axes_0"), val = tensor([1])]; + tensor var_2408_to_fp16 = const()[name = tensor("op_2408_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2408_to_fp16, x = inputs_65_cast_fp16)[name = tensor("out_65_cast_fp16")]; + tensor input_105_gamma_0_to_fp16 = const()[name = tensor("input_105_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459848128)))]; + tensor input_105_beta_0_to_fp16 = const()[name = tensor("input_105_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459850240)))]; + tensor input_105_epsilon_0_to_fp16 = const()[name = tensor("input_105_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = batch_norm(beta = input_105_beta_0_to_fp16, epsilon = input_105_epsilon_0_to_fp16, gamma = input_105_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_pad_type_0 = const()[name = tensor("input_107_pad_type_0"), val = tensor("valid")]; + tensor input_107_strides_0 = const()[name = tensor("input_107_strides_0"), val = tensor([1, 1])]; + tensor input_107_pad_0 = const()[name = tensor("input_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_107_dilations_0 = const()[name = tensor("input_107_dilations_0"), val = tensor([1, 1])]; + tensor input_107_groups_0 = const()[name = tensor("input_107_groups_0"), val = tensor(1)]; + tensor layers_10_fc1_weight_to_fp16 = const()[name = tensor("layers_10_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459852352)))]; + tensor layers_10_fc1_bias_to_fp16 = const()[name = tensor("layers_10_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468241024)))]; + tensor input_107_cast_fp16 = conv(bias = layers_10_fc1_bias_to_fp16, dilations = input_107_dilations_0, groups = input_107_groups_0, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = input_107_strides_0, weight = layers_10_fc1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor hidden_states_23_pad_type_0 = const()[name = tensor("hidden_states_23_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_23_strides_0 = const()[name = tensor("hidden_states_23_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_23_pad_0 = const()[name = tensor("hidden_states_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_23_dilations_0 = const()[name = tensor("hidden_states_23_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_23_groups_0 = const()[name = tensor("hidden_states_23_groups_0"), val = tensor(1)]; + tensor layers_10_fc2_weight_to_fp16 = const()[name = tensor("layers_10_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(468249280)))]; + tensor layers_10_fc2_bias_to_fp16 = const()[name = tensor("layers_10_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476637952)))]; + tensor hidden_states_23_cast_fp16 = conv(bias = layers_10_fc2_bias_to_fp16, dilations = hidden_states_23_dilations_0, groups = hidden_states_23_groups_0, pad = hidden_states_23_pad_0, pad_type = hidden_states_23_pad_type_0, strides = hidden_states_23_strides_0, weight = layers_10_fc2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("hidden_states_23_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = hidden_states_23_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor var_2443 = const()[name = tensor("op_2443"), val = tensor(3)]; + tensor out_67_axes_0 = const()[name = tensor("out_67_axes_0"), val = tensor([1])]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2468_to_fp16, x = inputs_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; + tensor obj_155_gamma_0_to_fp16 = const()[name = tensor("obj_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476640064)))]; + tensor obj_155_beta_0_to_fp16 = const()[name = tensor("obj_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476642176)))]; + tensor obj_155_epsilon_0_to_fp16 = const()[name = tensor("obj_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_155_cast_fp16 = batch_norm(beta = obj_155_beta_0_to_fp16, epsilon = obj_155_epsilon_0_to_fp16, gamma = obj_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor("obj_155_cast_fp16")]; + tensor query_45_pad_type_0 = const()[name = tensor("query_45_pad_type_0"), val = tensor("valid")]; + tensor query_45_strides_0 = const()[name = tensor("query_45_strides_0"), val = tensor([1, 1])]; + tensor query_45_pad_0 = const()[name = tensor("query_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_45_dilations_0 = const()[name = tensor("query_45_dilations_0"), val = tensor([1, 1])]; + tensor query_45_groups_0 = const()[name = tensor("query_45_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(476644288)))]; + tensor layers_11_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478741504)))]; + tensor query_45_cast_fp16 = conv(bias = layers_11_self_attn_q_proj_bias_to_fp16, dilations = query_45_dilations_0, groups = query_45_groups_0, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = query_45_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("query_45_cast_fp16")]; + tensor current_key_23_pad_type_0 = const()[name = tensor("current_key_23_pad_type_0"), val = tensor("valid")]; + tensor current_key_23_strides_0 = const()[name = tensor("current_key_23_strides_0"), val = tensor([1, 1])]; + tensor current_key_23_pad_0 = const()[name = tensor("current_key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_23_dilations_0 = const()[name = tensor("current_key_23_dilations_0"), val = tensor([1, 1])]; + tensor current_key_23_groups_0 = const()[name = tensor("current_key_23_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478743616)))]; + tensor current_key_23_cast_fp16 = conv(dilations = current_key_23_dilations_0, groups = current_key_23_groups_0, pad = current_key_23_pad_0, pad_type = current_key_23_pad_type_0, strides = current_key_23_strides_0, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("current_key_23_cast_fp16")]; + tensor current_value_23_pad_type_0 = const()[name = tensor("current_value_23_pad_type_0"), val = tensor("valid")]; + tensor current_value_23_strides_0 = const()[name = tensor("current_value_23_strides_0"), val = tensor([1, 1])]; + tensor current_value_23_pad_0 = const()[name = tensor("current_value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_23_dilations_0 = const()[name = tensor("current_value_23_dilations_0"), val = tensor([1, 1])]; + tensor current_value_23_groups_0 = const()[name = tensor("current_value_23_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(480840832)))]; + tensor layers_11_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482938048)))]; + tensor current_value_23_cast_fp16 = conv(bias = layers_11_self_attn_v_proj_bias_to_fp16, dilations = current_value_23_dilations_0, groups = current_value_23_groups_0, pad = current_value_23_pad_0, pad_type = current_value_23_pad_type_0, strides = current_value_23_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = obj_155_cast_fp16)[name = tensor("current_value_23_cast_fp16")]; + tensor var_2506_cast_fp16 = mul(x = current_key_23_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2506_cast_fp16")]; + tensor key_45_cast_fp16 = add(x = var_87_cast_fp16_11, y = var_2506_cast_fp16)[name = tensor("key_45_cast_fp16")]; + tensor var_2508_cast_fp16 = mul(x = current_value_23_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2508_cast_fp16")]; + tensor value_45_cast_fp16 = add(x = var_114_cast_fp16_11, y = var_2508_cast_fp16)[name = tensor("value_45_cast_fp16")]; + tensor var_2511 = const()[name = tensor("op_2511"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_45_cast_fp16 = reshape(shape = var_2511, x = query_45_cast_fp16)[name = tensor("mh_q_45_cast_fp16")]; + tensor var_2513_to_fp16 = const()[name = tensor("op_2513_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2514_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_2513_to_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor var_2515 = const()[name = tensor("op_2515"), val = tensor([1, 16, 64, -1])]; + tensor var_2516_cast_fp16 = reshape(shape = var_2515, x = key_45_cast_fp16)[name = tensor("op_2516_cast_fp16")]; + tensor mh_w_67_transpose_x_0 = const()[name = tensor("mh_w_67_transpose_x_0"), val = tensor(true)]; + tensor mh_w_67_transpose_y_0 = const()[name = tensor("mh_w_67_transpose_y_0"), val = tensor(false)]; + tensor mh_w_67_cast_fp16 = matmul(transpose_x = mh_w_67_transpose_x_0, transpose_y = mh_w_67_transpose_y_0, x = var_2514_cast_fp16, y = var_2516_cast_fp16)[name = tensor("mh_w_67_cast_fp16")]; + tensor mh_w_69_cast_fp16 = add(x = mh_w_67_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_69_cast_fp16")]; + tensor var_2524_cast_fp16 = softmax(axis = var_2443, x = mh_w_69_cast_fp16)[name = tensor("op_2524_cast_fp16")]; + tensor var_2525 = const()[name = tensor("op_2525"), val = tensor([1, 16, 64, -1])]; + tensor var_2526_cast_fp16 = reshape(shape = var_2525, x = value_45_cast_fp16)[name = tensor("op_2526_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(true)]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_2526_cast_fp16, y = var_2524_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_2529 = const()[name = tensor("op_2529"), val = tensor([1, 1024, 1, -1])]; + tensor input_111_cast_fp16 = reshape(shape = var_2529, x = attn_45_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor obj_161_pad_type_0 = const()[name = tensor("obj_161_pad_type_0"), val = tensor("valid")]; + tensor obj_161_strides_0 = const()[name = tensor("obj_161_strides_0"), val = tensor([1, 1])]; + tensor obj_161_pad_0 = const()[name = tensor("obj_161_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_161_dilations_0 = const()[name = tensor("obj_161_dilations_0"), val = tensor([1, 1])]; + tensor obj_161_groups_0 = const()[name = tensor("obj_161_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(482940160)))]; + tensor layers_11_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_11_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485037376)))]; + tensor obj_161_cast_fp16 = conv(bias = layers_11_self_attn_o_proj_bias_to_fp16, dilations = obj_161_dilations_0, groups = obj_161_groups_0, pad = obj_161_pad_0, pad_type = obj_161_pad_type_0, strides = obj_161_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("obj_161_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = obj_161_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor out_69_axes_0 = const()[name = tensor("out_69_axes_0"), val = tensor([1])]; + tensor var_2551_to_fp16 = const()[name = tensor("op_2551_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2551_to_fp16, x = inputs_69_cast_fp16)[name = tensor("out_69_cast_fp16")]; + tensor obj_163_gamma_0_to_fp16 = const()[name = tensor("obj_163_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485039488)))]; + tensor obj_163_beta_0_to_fp16 = const()[name = tensor("obj_163_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485041600)))]; + tensor obj_163_epsilon_0_to_fp16 = const()[name = tensor("obj_163_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_163_cast_fp16 = batch_norm(beta = obj_163_beta_0_to_fp16, epsilon = obj_163_epsilon_0_to_fp16, gamma = obj_163_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor("obj_163_cast_fp16")]; + tensor query_47_pad_type_0 = const()[name = tensor("query_47_pad_type_0"), val = tensor("valid")]; + tensor query_47_strides_0 = const()[name = tensor("query_47_strides_0"), val = tensor([1, 1])]; + tensor query_47_pad_0 = const()[name = tensor("query_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_47_dilations_0 = const()[name = tensor("query_47_dilations_0"), val = tensor([1, 1])]; + tensor query_47_groups_0 = const()[name = tensor("query_47_groups_0"), val = tensor(1)]; + tensor layers_11_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485043712)))]; + tensor layers_11_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487140928)))]; + tensor query_47_cast_fp16 = conv(bias = layers_11_encoder_attn_q_proj_bias_to_fp16, dilations = query_47_dilations_0, groups = query_47_groups_0, pad = query_47_pad_0, pad_type = query_47_pad_type_0, strides = query_47_strides_0, weight = layers_11_encoder_attn_q_proj_weight_to_fp16, x = obj_163_cast_fp16)[name = tensor("query_47_cast_fp16")]; + tensor key_47_pad_type_0 = const()[name = tensor("key_47_pad_type_0"), val = tensor("valid")]; + tensor key_47_strides_0 = const()[name = tensor("key_47_strides_0"), val = tensor([1, 1])]; + tensor key_47_pad_0 = const()[name = tensor("key_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_47_dilations_0 = const()[name = tensor("key_47_dilations_0"), val = tensor([1, 1])]; + tensor key_47_groups_0 = const()[name = tensor("key_47_groups_0"), val = tensor(1)]; + tensor layers_11_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487143040)))]; + tensor key_47_cast_fp16 = conv(dilations = key_47_dilations_0, groups = key_47_groups_0, pad = key_47_pad_0, pad_type = key_47_pad_type_0, strides = key_47_strides_0, weight = layers_11_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_47_cast_fp16")]; + tensor value_47_pad_type_0 = const()[name = tensor("value_47_pad_type_0"), val = tensor("valid")]; + tensor value_47_strides_0 = const()[name = tensor("value_47_strides_0"), val = tensor([1, 1])]; + tensor value_47_pad_0 = const()[name = tensor("value_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_47_dilations_0 = const()[name = tensor("value_47_dilations_0"), val = tensor([1, 1])]; + tensor value_47_groups_0 = const()[name = tensor("value_47_groups_0"), val = tensor(1)]; + tensor layers_11_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(489240256)))]; + tensor layers_11_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(491337472)))]; + tensor value_47_cast_fp16 = conv(bias = layers_11_encoder_attn_v_proj_bias_to_fp16, dilations = value_47_dilations_0, groups = value_47_groups_0, pad = value_47_pad_0, pad_type = value_47_pad_type_0, strides = value_47_strides_0, weight = layers_11_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_47_cast_fp16")]; + tensor var_2586 = const()[name = tensor("op_2586"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_47_cast_fp16 = reshape(shape = var_2586, x = query_47_cast_fp16)[name = tensor("mh_q_47_cast_fp16")]; + tensor var_2588_to_fp16 = const()[name = tensor("op_2588_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2589_cast_fp16 = mul(x = mh_q_47_cast_fp16, y = var_2588_to_fp16)[name = tensor("op_2589_cast_fp16")]; + tensor var_2590 = const()[name = tensor("op_2590"), val = tensor([1, 16, 64, -1])]; + tensor var_2591_cast_fp16 = reshape(shape = var_2590, x = key_47_cast_fp16)[name = tensor("op_2591_cast_fp16")]; + tensor mh_w_71_transpose_x_0 = const()[name = tensor("mh_w_71_transpose_x_0"), val = tensor(true)]; + tensor mh_w_71_transpose_y_0 = const()[name = tensor("mh_w_71_transpose_y_0"), val = tensor(false)]; + tensor mh_w_71_cast_fp16 = matmul(transpose_x = mh_w_71_transpose_x_0, transpose_y = mh_w_71_transpose_y_0, x = var_2589_cast_fp16, y = var_2591_cast_fp16)[name = tensor("mh_w_71_cast_fp16")]; + tensor obj_167_cast_fp16 = softmax(axis = var_2443, x = mh_w_71_cast_fp16)[name = tensor("obj_167_cast_fp16")]; + tensor var_2595 = const()[name = tensor("op_2595"), val = tensor([1, 16, 64, -1])]; + tensor var_2596_cast_fp16 = reshape(shape = var_2595, x = value_47_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor attn_47_transpose_x_0 = const()[name = tensor("attn_47_transpose_x_0"), val = tensor(false)]; + tensor attn_47_transpose_y_0 = const()[name = tensor("attn_47_transpose_y_0"), val = tensor(true)]; + tensor attn_47_cast_fp16 = matmul(transpose_x = attn_47_transpose_x_0, transpose_y = attn_47_transpose_y_0, x = var_2596_cast_fp16, y = obj_167_cast_fp16)[name = tensor("attn_47_cast_fp16")]; + tensor var_2599 = const()[name = tensor("op_2599"), val = tensor([1, 1024, 1, -1])]; + tensor input_113_cast_fp16 = reshape(shape = var_2599, x = attn_47_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor obj_165_pad_type_0 = const()[name = tensor("obj_165_pad_type_0"), val = tensor("valid")]; + tensor obj_165_strides_0 = const()[name = tensor("obj_165_strides_0"), val = tensor([1, 1])]; + tensor obj_165_pad_0 = const()[name = tensor("obj_165_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_165_dilations_0 = const()[name = tensor("obj_165_dilations_0"), val = tensor([1, 1])]; + tensor obj_165_groups_0 = const()[name = tensor("obj_165_groups_0"), val = tensor(1)]; + tensor layers_11_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(491339584)))]; + tensor layers_11_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_11_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493436800)))]; + tensor obj_165_cast_fp16 = conv(bias = layers_11_encoder_attn_o_proj_bias_to_fp16, dilations = obj_165_dilations_0, groups = obj_165_groups_0, pad = obj_165_pad_0, pad_type = obj_165_pad_type_0, strides = obj_165_strides_0, weight = layers_11_encoder_attn_o_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("obj_165_cast_fp16")]; + tensor inputs_71_cast_fp16 = add(x = inputs_69_cast_fp16, y = obj_165_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor out_71_axes_0 = const()[name = tensor("out_71_axes_0"), val = tensor([1])]; + tensor var_2617_to_fp16 = const()[name = tensor("op_2617_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2617_to_fp16, x = inputs_71_cast_fp16)[name = tensor("out_71_cast_fp16")]; + tensor input_115_gamma_0_to_fp16 = const()[name = tensor("input_115_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493438912)))]; + tensor input_115_beta_0_to_fp16 = const()[name = tensor("input_115_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493441024)))]; + tensor input_115_epsilon_0_to_fp16 = const()[name = tensor("input_115_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_115_cast_fp16 = batch_norm(beta = input_115_beta_0_to_fp16, epsilon = input_115_epsilon_0_to_fp16, gamma = input_115_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_pad_type_0 = const()[name = tensor("input_117_pad_type_0"), val = tensor("valid")]; + tensor input_117_strides_0 = const()[name = tensor("input_117_strides_0"), val = tensor([1, 1])]; + tensor input_117_pad_0 = const()[name = tensor("input_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_117_dilations_0 = const()[name = tensor("input_117_dilations_0"), val = tensor([1, 1])]; + tensor input_117_groups_0 = const()[name = tensor("input_117_groups_0"), val = tensor(1)]; + tensor layers_11_fc1_weight_to_fp16 = const()[name = tensor("layers_11_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(493443136)))]; + tensor layers_11_fc1_bias_to_fp16 = const()[name = tensor("layers_11_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(501831808)))]; + tensor input_117_cast_fp16 = conv(bias = layers_11_fc1_bias_to_fp16, dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = layers_11_fc1_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor hidden_states_25_pad_type_0 = const()[name = tensor("hidden_states_25_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_25_strides_0 = const()[name = tensor("hidden_states_25_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_25_pad_0 = const()[name = tensor("hidden_states_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_25_dilations_0 = const()[name = tensor("hidden_states_25_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_25_groups_0 = const()[name = tensor("hidden_states_25_groups_0"), val = tensor(1)]; + tensor layers_11_fc2_weight_to_fp16 = const()[name = tensor("layers_11_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(501840064)))]; + tensor layers_11_fc2_bias_to_fp16 = const()[name = tensor("layers_11_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510228736)))]; + tensor hidden_states_25_cast_fp16 = conv(bias = layers_11_fc2_bias_to_fp16, dilations = hidden_states_25_dilations_0, groups = hidden_states_25_groups_0, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = hidden_states_25_strides_0, weight = layers_11_fc2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("hidden_states_25_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = hidden_states_25_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor var_2652 = const()[name = tensor("op_2652"), val = tensor(3)]; + tensor out_73_axes_0 = const()[name = tensor("out_73_axes_0"), val = tensor([1])]; + tensor var_2677_to_fp16 = const()[name = tensor("op_2677_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_2677_to_fp16, x = inputs_73_cast_fp16)[name = tensor("out_73_cast_fp16")]; + tensor obj_169_gamma_0_to_fp16 = const()[name = tensor("obj_169_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510230848)))]; + tensor obj_169_beta_0_to_fp16 = const()[name = tensor("obj_169_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510232960)))]; + tensor obj_169_epsilon_0_to_fp16 = const()[name = tensor("obj_169_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_169_cast_fp16 = batch_norm(beta = obj_169_beta_0_to_fp16, epsilon = obj_169_epsilon_0_to_fp16, gamma = obj_169_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor("obj_169_cast_fp16")]; + tensor query_49_pad_type_0 = const()[name = tensor("query_49_pad_type_0"), val = tensor("valid")]; + tensor query_49_strides_0 = const()[name = tensor("query_49_strides_0"), val = tensor([1, 1])]; + tensor query_49_pad_0 = const()[name = tensor("query_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_49_dilations_0 = const()[name = tensor("query_49_dilations_0"), val = tensor([1, 1])]; + tensor query_49_groups_0 = const()[name = tensor("query_49_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510235072)))]; + tensor layers_12_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512332288)))]; + tensor query_49_cast_fp16 = conv(bias = layers_12_self_attn_q_proj_bias_to_fp16, dilations = query_49_dilations_0, groups = query_49_groups_0, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = query_49_strides_0, weight = layers_12_self_attn_q_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("query_49_cast_fp16")]; + tensor current_key_25_pad_type_0 = const()[name = tensor("current_key_25_pad_type_0"), val = tensor("valid")]; + tensor current_key_25_strides_0 = const()[name = tensor("current_key_25_strides_0"), val = tensor([1, 1])]; + tensor current_key_25_pad_0 = const()[name = tensor("current_key_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_25_dilations_0 = const()[name = tensor("current_key_25_dilations_0"), val = tensor([1, 1])]; + tensor current_key_25_groups_0 = const()[name = tensor("current_key_25_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512334400)))]; + tensor current_key_25_cast_fp16 = conv(dilations = current_key_25_dilations_0, groups = current_key_25_groups_0, pad = current_key_25_pad_0, pad_type = current_key_25_pad_type_0, strides = current_key_25_strides_0, weight = layers_12_self_attn_k_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("current_key_25_cast_fp16")]; + tensor current_value_25_pad_type_0 = const()[name = tensor("current_value_25_pad_type_0"), val = tensor("valid")]; + tensor current_value_25_strides_0 = const()[name = tensor("current_value_25_strides_0"), val = tensor([1, 1])]; + tensor current_value_25_pad_0 = const()[name = tensor("current_value_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_25_dilations_0 = const()[name = tensor("current_value_25_dilations_0"), val = tensor([1, 1])]; + tensor current_value_25_groups_0 = const()[name = tensor("current_value_25_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(514431616)))]; + tensor layers_12_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(516528832)))]; + tensor current_value_25_cast_fp16 = conv(bias = layers_12_self_attn_v_proj_bias_to_fp16, dilations = current_value_25_dilations_0, groups = current_value_25_groups_0, pad = current_value_25_pad_0, pad_type = current_value_25_pad_type_0, strides = current_value_25_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = obj_169_cast_fp16)[name = tensor("current_value_25_cast_fp16")]; + tensor var_2715_cast_fp16 = mul(x = current_key_25_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2715_cast_fp16")]; + tensor key_49_cast_fp16 = add(x = var_87_cast_fp16_12, y = var_2715_cast_fp16)[name = tensor("key_49_cast_fp16")]; + tensor var_2717_cast_fp16 = mul(x = current_value_25_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2717_cast_fp16")]; + tensor value_49_cast_fp16 = add(x = var_114_cast_fp16_12, y = var_2717_cast_fp16)[name = tensor("value_49_cast_fp16")]; + tensor var_2720 = const()[name = tensor("op_2720"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_49_cast_fp16 = reshape(shape = var_2720, x = query_49_cast_fp16)[name = tensor("mh_q_49_cast_fp16")]; + tensor var_2722_to_fp16 = const()[name = tensor("op_2722_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2723_cast_fp16 = mul(x = mh_q_49_cast_fp16, y = var_2722_to_fp16)[name = tensor("op_2723_cast_fp16")]; + tensor var_2724 = const()[name = tensor("op_2724"), val = tensor([1, 16, 64, -1])]; + tensor var_2725_cast_fp16 = reshape(shape = var_2724, x = key_49_cast_fp16)[name = tensor("op_2725_cast_fp16")]; + tensor mh_w_73_transpose_x_0 = const()[name = tensor("mh_w_73_transpose_x_0"), val = tensor(true)]; + tensor mh_w_73_transpose_y_0 = const()[name = tensor("mh_w_73_transpose_y_0"), val = tensor(false)]; + tensor mh_w_73_cast_fp16 = matmul(transpose_x = mh_w_73_transpose_x_0, transpose_y = mh_w_73_transpose_y_0, x = var_2723_cast_fp16, y = var_2725_cast_fp16)[name = tensor("mh_w_73_cast_fp16")]; + tensor mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_75_cast_fp16")]; + tensor var_2733_cast_fp16 = softmax(axis = var_2652, x = mh_w_75_cast_fp16)[name = tensor("op_2733_cast_fp16")]; + tensor var_2734 = const()[name = tensor("op_2734"), val = tensor([1, 16, 64, -1])]; + tensor var_2735_cast_fp16 = reshape(shape = var_2734, x = value_49_cast_fp16)[name = tensor("op_2735_cast_fp16")]; + tensor attn_49_transpose_x_0 = const()[name = tensor("attn_49_transpose_x_0"), val = tensor(false)]; + tensor attn_49_transpose_y_0 = const()[name = tensor("attn_49_transpose_y_0"), val = tensor(true)]; + tensor attn_49_cast_fp16 = matmul(transpose_x = attn_49_transpose_x_0, transpose_y = attn_49_transpose_y_0, x = var_2735_cast_fp16, y = var_2733_cast_fp16)[name = tensor("attn_49_cast_fp16")]; + tensor var_2738 = const()[name = tensor("op_2738"), val = tensor([1, 1024, 1, -1])]; + tensor input_121_cast_fp16 = reshape(shape = var_2738, x = attn_49_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor obj_175_pad_type_0 = const()[name = tensor("obj_175_pad_type_0"), val = tensor("valid")]; + tensor obj_175_strides_0 = const()[name = tensor("obj_175_strides_0"), val = tensor([1, 1])]; + tensor obj_175_pad_0 = const()[name = tensor("obj_175_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_175_dilations_0 = const()[name = tensor("obj_175_dilations_0"), val = tensor([1, 1])]; + tensor obj_175_groups_0 = const()[name = tensor("obj_175_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(516530944)))]; + tensor layers_12_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_12_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518628160)))]; + tensor obj_175_cast_fp16 = conv(bias = layers_12_self_attn_o_proj_bias_to_fp16, dilations = obj_175_dilations_0, groups = obj_175_groups_0, pad = obj_175_pad_0, pad_type = obj_175_pad_type_0, strides = obj_175_strides_0, weight = layers_12_self_attn_o_proj_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("obj_175_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_175_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor out_75_axes_0 = const()[name = tensor("out_75_axes_0"), val = tensor([1])]; + tensor var_2760_to_fp16 = const()[name = tensor("op_2760_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_2760_to_fp16, x = inputs_75_cast_fp16)[name = tensor("out_75_cast_fp16")]; + tensor obj_177_gamma_0_to_fp16 = const()[name = tensor("obj_177_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518630272)))]; + tensor obj_177_beta_0_to_fp16 = const()[name = tensor("obj_177_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518632384)))]; + tensor obj_177_epsilon_0_to_fp16 = const()[name = tensor("obj_177_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_177_cast_fp16 = batch_norm(beta = obj_177_beta_0_to_fp16, epsilon = obj_177_epsilon_0_to_fp16, gamma = obj_177_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor("obj_177_cast_fp16")]; + tensor query_51_pad_type_0 = const()[name = tensor("query_51_pad_type_0"), val = tensor("valid")]; + tensor query_51_strides_0 = const()[name = tensor("query_51_strides_0"), val = tensor([1, 1])]; + tensor query_51_pad_0 = const()[name = tensor("query_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_51_dilations_0 = const()[name = tensor("query_51_dilations_0"), val = tensor([1, 1])]; + tensor query_51_groups_0 = const()[name = tensor("query_51_groups_0"), val = tensor(1)]; + tensor layers_12_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(518634496)))]; + tensor layers_12_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520731712)))]; + tensor query_51_cast_fp16 = conv(bias = layers_12_encoder_attn_q_proj_bias_to_fp16, dilations = query_51_dilations_0, groups = query_51_groups_0, pad = query_51_pad_0, pad_type = query_51_pad_type_0, strides = query_51_strides_0, weight = layers_12_encoder_attn_q_proj_weight_to_fp16, x = obj_177_cast_fp16)[name = tensor("query_51_cast_fp16")]; + tensor key_51_pad_type_0 = const()[name = tensor("key_51_pad_type_0"), val = tensor("valid")]; + tensor key_51_strides_0 = const()[name = tensor("key_51_strides_0"), val = tensor([1, 1])]; + tensor key_51_pad_0 = const()[name = tensor("key_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_51_dilations_0 = const()[name = tensor("key_51_dilations_0"), val = tensor([1, 1])]; + tensor key_51_groups_0 = const()[name = tensor("key_51_groups_0"), val = tensor(1)]; + tensor layers_12_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520733824)))]; + tensor key_51_cast_fp16 = conv(dilations = key_51_dilations_0, groups = key_51_groups_0, pad = key_51_pad_0, pad_type = key_51_pad_type_0, strides = key_51_strides_0, weight = layers_12_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_51_cast_fp16")]; + tensor value_51_pad_type_0 = const()[name = tensor("value_51_pad_type_0"), val = tensor("valid")]; + tensor value_51_strides_0 = const()[name = tensor("value_51_strides_0"), val = tensor([1, 1])]; + tensor value_51_pad_0 = const()[name = tensor("value_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_51_dilations_0 = const()[name = tensor("value_51_dilations_0"), val = tensor([1, 1])]; + tensor value_51_groups_0 = const()[name = tensor("value_51_groups_0"), val = tensor(1)]; + tensor layers_12_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(522831040)))]; + tensor layers_12_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524928256)))]; + tensor value_51_cast_fp16 = conv(bias = layers_12_encoder_attn_v_proj_bias_to_fp16, dilations = value_51_dilations_0, groups = value_51_groups_0, pad = value_51_pad_0, pad_type = value_51_pad_type_0, strides = value_51_strides_0, weight = layers_12_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_51_cast_fp16")]; + tensor var_2795 = const()[name = tensor("op_2795"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_51_cast_fp16 = reshape(shape = var_2795, x = query_51_cast_fp16)[name = tensor("mh_q_51_cast_fp16")]; + tensor var_2797_to_fp16 = const()[name = tensor("op_2797_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2798_cast_fp16 = mul(x = mh_q_51_cast_fp16, y = var_2797_to_fp16)[name = tensor("op_2798_cast_fp16")]; + tensor var_2799 = const()[name = tensor("op_2799"), val = tensor([1, 16, 64, -1])]; + tensor var_2800_cast_fp16 = reshape(shape = var_2799, x = key_51_cast_fp16)[name = tensor("op_2800_cast_fp16")]; + tensor mh_w_77_transpose_x_0 = const()[name = tensor("mh_w_77_transpose_x_0"), val = tensor(true)]; + tensor mh_w_77_transpose_y_0 = const()[name = tensor("mh_w_77_transpose_y_0"), val = tensor(false)]; + tensor mh_w_77_cast_fp16 = matmul(transpose_x = mh_w_77_transpose_x_0, transpose_y = mh_w_77_transpose_y_0, x = var_2798_cast_fp16, y = var_2800_cast_fp16)[name = tensor("mh_w_77_cast_fp16")]; + tensor obj_181_cast_fp16 = softmax(axis = var_2652, x = mh_w_77_cast_fp16)[name = tensor("obj_181_cast_fp16")]; + tensor var_2804 = const()[name = tensor("op_2804"), val = tensor([1, 16, 64, -1])]; + tensor var_2805_cast_fp16 = reshape(shape = var_2804, x = value_51_cast_fp16)[name = tensor("op_2805_cast_fp16")]; + tensor attn_51_transpose_x_0 = const()[name = tensor("attn_51_transpose_x_0"), val = tensor(false)]; + tensor attn_51_transpose_y_0 = const()[name = tensor("attn_51_transpose_y_0"), val = tensor(true)]; + tensor attn_51_cast_fp16 = matmul(transpose_x = attn_51_transpose_x_0, transpose_y = attn_51_transpose_y_0, x = var_2805_cast_fp16, y = obj_181_cast_fp16)[name = tensor("attn_51_cast_fp16")]; + tensor var_2808 = const()[name = tensor("op_2808"), val = tensor([1, 1024, 1, -1])]; + tensor input_123_cast_fp16 = reshape(shape = var_2808, x = attn_51_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor obj_179_pad_type_0 = const()[name = tensor("obj_179_pad_type_0"), val = tensor("valid")]; + tensor obj_179_strides_0 = const()[name = tensor("obj_179_strides_0"), val = tensor([1, 1])]; + tensor obj_179_pad_0 = const()[name = tensor("obj_179_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_179_dilations_0 = const()[name = tensor("obj_179_dilations_0"), val = tensor([1, 1])]; + tensor obj_179_groups_0 = const()[name = tensor("obj_179_groups_0"), val = tensor(1)]; + tensor layers_12_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(524930368)))]; + tensor layers_12_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_12_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527027584)))]; + tensor obj_179_cast_fp16 = conv(bias = layers_12_encoder_attn_o_proj_bias_to_fp16, dilations = obj_179_dilations_0, groups = obj_179_groups_0, pad = obj_179_pad_0, pad_type = obj_179_pad_type_0, strides = obj_179_strides_0, weight = layers_12_encoder_attn_o_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("obj_179_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = obj_179_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor out_77_axes_0 = const()[name = tensor("out_77_axes_0"), val = tensor([1])]; + tensor var_2826_to_fp16 = const()[name = tensor("op_2826_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_2826_to_fp16, x = inputs_77_cast_fp16)[name = tensor("out_77_cast_fp16")]; + tensor input_125_gamma_0_to_fp16 = const()[name = tensor("input_125_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527029696)))]; + tensor input_125_beta_0_to_fp16 = const()[name = tensor("input_125_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527031808)))]; + tensor input_125_epsilon_0_to_fp16 = const()[name = tensor("input_125_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = batch_norm(beta = input_125_beta_0_to_fp16, epsilon = input_125_epsilon_0_to_fp16, gamma = input_125_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor input_127_pad_type_0 = const()[name = tensor("input_127_pad_type_0"), val = tensor("valid")]; + tensor input_127_strides_0 = const()[name = tensor("input_127_strides_0"), val = tensor([1, 1])]; + tensor input_127_pad_0 = const()[name = tensor("input_127_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_127_dilations_0 = const()[name = tensor("input_127_dilations_0"), val = tensor([1, 1])]; + tensor input_127_groups_0 = const()[name = tensor("input_127_groups_0"), val = tensor(1)]; + tensor layers_12_fc1_weight_to_fp16 = const()[name = tensor("layers_12_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(527033920)))]; + tensor layers_12_fc1_bias_to_fp16 = const()[name = tensor("layers_12_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535422592)))]; + tensor input_127_cast_fp16 = conv(bias = layers_12_fc1_bias_to_fp16, dilations = input_127_dilations_0, groups = input_127_groups_0, pad = input_127_pad_0, pad_type = input_127_pad_type_0, strides = input_127_strides_0, weight = layers_12_fc1_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_mode_0 = const()[name = tensor("input_129_mode_0"), val = tensor("EXACT")]; + tensor input_129_cast_fp16 = gelu(mode = input_129_mode_0, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor hidden_states_27_pad_type_0 = const()[name = tensor("hidden_states_27_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_27_strides_0 = const()[name = tensor("hidden_states_27_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_27_pad_0 = const()[name = tensor("hidden_states_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_27_dilations_0 = const()[name = tensor("hidden_states_27_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_27_groups_0 = const()[name = tensor("hidden_states_27_groups_0"), val = tensor(1)]; + tensor layers_12_fc2_weight_to_fp16 = const()[name = tensor("layers_12_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535430848)))]; + tensor layers_12_fc2_bias_to_fp16 = const()[name = tensor("layers_12_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543819520)))]; + tensor hidden_states_27_cast_fp16 = conv(bias = layers_12_fc2_bias_to_fp16, dilations = hidden_states_27_dilations_0, groups = hidden_states_27_groups_0, pad = hidden_states_27_pad_0, pad_type = hidden_states_27_pad_type_0, strides = hidden_states_27_strides_0, weight = layers_12_fc2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("hidden_states_27_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = hidden_states_27_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor var_2861 = const()[name = tensor("op_2861"), val = tensor(3)]; + tensor out_79_axes_0 = const()[name = tensor("out_79_axes_0"), val = tensor([1])]; + tensor var_2886_to_fp16 = const()[name = tensor("op_2886_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_2886_to_fp16, x = inputs_79_cast_fp16)[name = tensor("out_79_cast_fp16")]; + tensor obj_183_gamma_0_to_fp16 = const()[name = tensor("obj_183_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543821632)))]; + tensor obj_183_beta_0_to_fp16 = const()[name = tensor("obj_183_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543823744)))]; + tensor obj_183_epsilon_0_to_fp16 = const()[name = tensor("obj_183_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_183_cast_fp16 = batch_norm(beta = obj_183_beta_0_to_fp16, epsilon = obj_183_epsilon_0_to_fp16, gamma = obj_183_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor("obj_183_cast_fp16")]; + tensor query_53_pad_type_0 = const()[name = tensor("query_53_pad_type_0"), val = tensor("valid")]; + tensor query_53_strides_0 = const()[name = tensor("query_53_strides_0"), val = tensor([1, 1])]; + tensor query_53_pad_0 = const()[name = tensor("query_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_53_dilations_0 = const()[name = tensor("query_53_dilations_0"), val = tensor([1, 1])]; + tensor query_53_groups_0 = const()[name = tensor("query_53_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(543825856)))]; + tensor layers_13_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545923072)))]; + tensor query_53_cast_fp16 = conv(bias = layers_13_self_attn_q_proj_bias_to_fp16, dilations = query_53_dilations_0, groups = query_53_groups_0, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = query_53_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("query_53_cast_fp16")]; + tensor current_key_27_pad_type_0 = const()[name = tensor("current_key_27_pad_type_0"), val = tensor("valid")]; + tensor current_key_27_strides_0 = const()[name = tensor("current_key_27_strides_0"), val = tensor([1, 1])]; + tensor current_key_27_pad_0 = const()[name = tensor("current_key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_27_dilations_0 = const()[name = tensor("current_key_27_dilations_0"), val = tensor([1, 1])]; + tensor current_key_27_groups_0 = const()[name = tensor("current_key_27_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545925184)))]; + tensor current_key_27_cast_fp16 = conv(dilations = current_key_27_dilations_0, groups = current_key_27_groups_0, pad = current_key_27_pad_0, pad_type = current_key_27_pad_type_0, strides = current_key_27_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("current_key_27_cast_fp16")]; + tensor current_value_27_pad_type_0 = const()[name = tensor("current_value_27_pad_type_0"), val = tensor("valid")]; + tensor current_value_27_strides_0 = const()[name = tensor("current_value_27_strides_0"), val = tensor([1, 1])]; + tensor current_value_27_pad_0 = const()[name = tensor("current_value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_27_dilations_0 = const()[name = tensor("current_value_27_dilations_0"), val = tensor([1, 1])]; + tensor current_value_27_groups_0 = const()[name = tensor("current_value_27_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(548022400)))]; + tensor layers_13_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550119616)))]; + tensor current_value_27_cast_fp16 = conv(bias = layers_13_self_attn_v_proj_bias_to_fp16, dilations = current_value_27_dilations_0, groups = current_value_27_groups_0, pad = current_value_27_pad_0, pad_type = current_value_27_pad_type_0, strides = current_value_27_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = obj_183_cast_fp16)[name = tensor("current_value_27_cast_fp16")]; + tensor var_2924_cast_fp16 = mul(x = current_key_27_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor key_53_cast_fp16 = add(x = var_87_cast_fp16_13, y = var_2924_cast_fp16)[name = tensor("key_53_cast_fp16")]; + tensor var_2926_cast_fp16 = mul(x = current_value_27_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor value_53_cast_fp16 = add(x = var_114_cast_fp16_13, y = var_2926_cast_fp16)[name = tensor("value_53_cast_fp16")]; + tensor var_2929 = const()[name = tensor("op_2929"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_53_cast_fp16 = reshape(shape = var_2929, x = query_53_cast_fp16)[name = tensor("mh_q_53_cast_fp16")]; + tensor var_2931_to_fp16 = const()[name = tensor("op_2931_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2932_cast_fp16 = mul(x = mh_q_53_cast_fp16, y = var_2931_to_fp16)[name = tensor("op_2932_cast_fp16")]; + tensor var_2933 = const()[name = tensor("op_2933"), val = tensor([1, 16, 64, -1])]; + tensor var_2934_cast_fp16 = reshape(shape = var_2933, x = key_53_cast_fp16)[name = tensor("op_2934_cast_fp16")]; + tensor mh_w_79_transpose_x_0 = const()[name = tensor("mh_w_79_transpose_x_0"), val = tensor(true)]; + tensor mh_w_79_transpose_y_0 = const()[name = tensor("mh_w_79_transpose_y_0"), val = tensor(false)]; + tensor mh_w_79_cast_fp16 = matmul(transpose_x = mh_w_79_transpose_x_0, transpose_y = mh_w_79_transpose_y_0, x = var_2932_cast_fp16, y = var_2934_cast_fp16)[name = tensor("mh_w_79_cast_fp16")]; + tensor mh_w_81_cast_fp16 = add(x = mh_w_79_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_81_cast_fp16")]; + tensor var_2942_cast_fp16 = softmax(axis = var_2861, x = mh_w_81_cast_fp16)[name = tensor("op_2942_cast_fp16")]; + tensor var_2943 = const()[name = tensor("op_2943"), val = tensor([1, 16, 64, -1])]; + tensor var_2944_cast_fp16 = reshape(shape = var_2943, x = value_53_cast_fp16)[name = tensor("op_2944_cast_fp16")]; + tensor attn_53_transpose_x_0 = const()[name = tensor("attn_53_transpose_x_0"), val = tensor(false)]; + tensor attn_53_transpose_y_0 = const()[name = tensor("attn_53_transpose_y_0"), val = tensor(true)]; + tensor attn_53_cast_fp16 = matmul(transpose_x = attn_53_transpose_x_0, transpose_y = attn_53_transpose_y_0, x = var_2944_cast_fp16, y = var_2942_cast_fp16)[name = tensor("attn_53_cast_fp16")]; + tensor var_2947 = const()[name = tensor("op_2947"), val = tensor([1, 1024, 1, -1])]; + tensor input_131_cast_fp16 = reshape(shape = var_2947, x = attn_53_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor obj_189_pad_type_0 = const()[name = tensor("obj_189_pad_type_0"), val = tensor("valid")]; + tensor obj_189_strides_0 = const()[name = tensor("obj_189_strides_0"), val = tensor([1, 1])]; + tensor obj_189_pad_0 = const()[name = tensor("obj_189_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_189_dilations_0 = const()[name = tensor("obj_189_dilations_0"), val = tensor([1, 1])]; + tensor obj_189_groups_0 = const()[name = tensor("obj_189_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(550121728)))]; + tensor layers_13_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_13_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552218944)))]; + tensor obj_189_cast_fp16 = conv(bias = layers_13_self_attn_o_proj_bias_to_fp16, dilations = obj_189_dilations_0, groups = obj_189_groups_0, pad = obj_189_pad_0, pad_type = obj_189_pad_type_0, strides = obj_189_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("obj_189_cast_fp16")]; + tensor inputs_81_cast_fp16 = add(x = inputs_79_cast_fp16, y = obj_189_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor out_81_axes_0 = const()[name = tensor("out_81_axes_0"), val = tensor([1])]; + tensor var_2969_to_fp16 = const()[name = tensor("op_2969_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_2969_to_fp16, x = inputs_81_cast_fp16)[name = tensor("out_81_cast_fp16")]; + tensor obj_191_gamma_0_to_fp16 = const()[name = tensor("obj_191_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552221056)))]; + tensor obj_191_beta_0_to_fp16 = const()[name = tensor("obj_191_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552223168)))]; + tensor obj_191_epsilon_0_to_fp16 = const()[name = tensor("obj_191_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_191_cast_fp16 = batch_norm(beta = obj_191_beta_0_to_fp16, epsilon = obj_191_epsilon_0_to_fp16, gamma = obj_191_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor("obj_191_cast_fp16")]; + tensor query_55_pad_type_0 = const()[name = tensor("query_55_pad_type_0"), val = tensor("valid")]; + tensor query_55_strides_0 = const()[name = tensor("query_55_strides_0"), val = tensor([1, 1])]; + tensor query_55_pad_0 = const()[name = tensor("query_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_55_dilations_0 = const()[name = tensor("query_55_dilations_0"), val = tensor([1, 1])]; + tensor query_55_groups_0 = const()[name = tensor("query_55_groups_0"), val = tensor(1)]; + tensor layers_13_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(552225280)))]; + tensor layers_13_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554322496)))]; + tensor query_55_cast_fp16 = conv(bias = layers_13_encoder_attn_q_proj_bias_to_fp16, dilations = query_55_dilations_0, groups = query_55_groups_0, pad = query_55_pad_0, pad_type = query_55_pad_type_0, strides = query_55_strides_0, weight = layers_13_encoder_attn_q_proj_weight_to_fp16, x = obj_191_cast_fp16)[name = tensor("query_55_cast_fp16")]; + tensor key_55_pad_type_0 = const()[name = tensor("key_55_pad_type_0"), val = tensor("valid")]; + tensor key_55_strides_0 = const()[name = tensor("key_55_strides_0"), val = tensor([1, 1])]; + tensor key_55_pad_0 = const()[name = tensor("key_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_55_dilations_0 = const()[name = tensor("key_55_dilations_0"), val = tensor([1, 1])]; + tensor key_55_groups_0 = const()[name = tensor("key_55_groups_0"), val = tensor(1)]; + tensor layers_13_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554324608)))]; + tensor key_55_cast_fp16 = conv(dilations = key_55_dilations_0, groups = key_55_groups_0, pad = key_55_pad_0, pad_type = key_55_pad_type_0, strides = key_55_strides_0, weight = layers_13_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_55_cast_fp16")]; + tensor value_55_pad_type_0 = const()[name = tensor("value_55_pad_type_0"), val = tensor("valid")]; + tensor value_55_strides_0 = const()[name = tensor("value_55_strides_0"), val = tensor([1, 1])]; + tensor value_55_pad_0 = const()[name = tensor("value_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_55_dilations_0 = const()[name = tensor("value_55_dilations_0"), val = tensor([1, 1])]; + tensor value_55_groups_0 = const()[name = tensor("value_55_groups_0"), val = tensor(1)]; + tensor layers_13_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(556421824)))]; + tensor layers_13_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(558519040)))]; + tensor value_55_cast_fp16 = conv(bias = layers_13_encoder_attn_v_proj_bias_to_fp16, dilations = value_55_dilations_0, groups = value_55_groups_0, pad = value_55_pad_0, pad_type = value_55_pad_type_0, strides = value_55_strides_0, weight = layers_13_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_55_cast_fp16")]; + tensor var_3004 = const()[name = tensor("op_3004"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_55_cast_fp16 = reshape(shape = var_3004, x = query_55_cast_fp16)[name = tensor("mh_q_55_cast_fp16")]; + tensor var_3006_to_fp16 = const()[name = tensor("op_3006_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3007_cast_fp16 = mul(x = mh_q_55_cast_fp16, y = var_3006_to_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor var_3008 = const()[name = tensor("op_3008"), val = tensor([1, 16, 64, -1])]; + tensor var_3009_cast_fp16 = reshape(shape = var_3008, x = key_55_cast_fp16)[name = tensor("op_3009_cast_fp16")]; + tensor mh_w_83_transpose_x_0 = const()[name = tensor("mh_w_83_transpose_x_0"), val = tensor(true)]; + tensor mh_w_83_transpose_y_0 = const()[name = tensor("mh_w_83_transpose_y_0"), val = tensor(false)]; + tensor mh_w_83_cast_fp16 = matmul(transpose_x = mh_w_83_transpose_x_0, transpose_y = mh_w_83_transpose_y_0, x = var_3007_cast_fp16, y = var_3009_cast_fp16)[name = tensor("mh_w_83_cast_fp16")]; + tensor obj_195_cast_fp16 = softmax(axis = var_2861, x = mh_w_83_cast_fp16)[name = tensor("obj_195_cast_fp16")]; + tensor var_3013 = const()[name = tensor("op_3013"), val = tensor([1, 16, 64, -1])]; + tensor var_3014_cast_fp16 = reshape(shape = var_3013, x = value_55_cast_fp16)[name = tensor("op_3014_cast_fp16")]; + tensor attn_55_transpose_x_0 = const()[name = tensor("attn_55_transpose_x_0"), val = tensor(false)]; + tensor attn_55_transpose_y_0 = const()[name = tensor("attn_55_transpose_y_0"), val = tensor(true)]; + tensor attn_55_cast_fp16 = matmul(transpose_x = attn_55_transpose_x_0, transpose_y = attn_55_transpose_y_0, x = var_3014_cast_fp16, y = obj_195_cast_fp16)[name = tensor("attn_55_cast_fp16")]; + tensor var_3017 = const()[name = tensor("op_3017"), val = tensor([1, 1024, 1, -1])]; + tensor input_133_cast_fp16 = reshape(shape = var_3017, x = attn_55_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor obj_193_pad_type_0 = const()[name = tensor("obj_193_pad_type_0"), val = tensor("valid")]; + tensor obj_193_strides_0 = const()[name = tensor("obj_193_strides_0"), val = tensor([1, 1])]; + tensor obj_193_pad_0 = const()[name = tensor("obj_193_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_193_dilations_0 = const()[name = tensor("obj_193_dilations_0"), val = tensor([1, 1])]; + tensor obj_193_groups_0 = const()[name = tensor("obj_193_groups_0"), val = tensor(1)]; + tensor layers_13_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(558521152)))]; + tensor layers_13_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_13_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560618368)))]; + tensor obj_193_cast_fp16 = conv(bias = layers_13_encoder_attn_o_proj_bias_to_fp16, dilations = obj_193_dilations_0, groups = obj_193_groups_0, pad = obj_193_pad_0, pad_type = obj_193_pad_type_0, strides = obj_193_strides_0, weight = layers_13_encoder_attn_o_proj_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("obj_193_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = obj_193_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor out_83_axes_0 = const()[name = tensor("out_83_axes_0"), val = tensor([1])]; + tensor var_3038_to_fp16 = const()[name = tensor("op_3038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_3038_to_fp16, x = inputs_83_cast_fp16)[name = tensor("out_83_cast_fp16")]; + tensor input_135_gamma_0_to_fp16 = const()[name = tensor("input_135_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560620480)))]; + tensor input_135_beta_0_to_fp16 = const()[name = tensor("input_135_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560622592)))]; + tensor input_135_epsilon_0_to_fp16 = const()[name = tensor("input_135_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = batch_norm(beta = input_135_beta_0_to_fp16, epsilon = input_135_epsilon_0_to_fp16, gamma = input_135_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_pad_type_0 = const()[name = tensor("input_137_pad_type_0"), val = tensor("valid")]; + tensor input_137_strides_0 = const()[name = tensor("input_137_strides_0"), val = tensor([1, 1])]; + tensor input_137_pad_0 = const()[name = tensor("input_137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_137_dilations_0 = const()[name = tensor("input_137_dilations_0"), val = tensor([1, 1])]; + tensor input_137_groups_0 = const()[name = tensor("input_137_groups_0"), val = tensor(1)]; + tensor layers_13_fc1_weight_to_fp16 = const()[name = tensor("layers_13_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560624704)))]; + tensor layers_13_fc1_bias_to_fp16 = const()[name = tensor("layers_13_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569013376)))]; + tensor input_137_cast_fp16 = conv(bias = layers_13_fc1_bias_to_fp16, dilations = input_137_dilations_0, groups = input_137_groups_0, pad = input_137_pad_0, pad_type = input_137_pad_type_0, strides = input_137_strides_0, weight = layers_13_fc1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_mode_0 = const()[name = tensor("input_139_mode_0"), val = tensor("EXACT")]; + tensor input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor hidden_states_29_pad_type_0 = const()[name = tensor("hidden_states_29_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_29_strides_0 = const()[name = tensor("hidden_states_29_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_29_pad_0 = const()[name = tensor("hidden_states_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_29_dilations_0 = const()[name = tensor("hidden_states_29_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_29_groups_0 = const()[name = tensor("hidden_states_29_groups_0"), val = tensor(1)]; + tensor layers_13_fc2_weight_to_fp16 = const()[name = tensor("layers_13_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(569021632)))]; + tensor layers_13_fc2_bias_to_fp16 = const()[name = tensor("layers_13_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577410304)))]; + tensor hidden_states_29_cast_fp16 = conv(bias = layers_13_fc2_bias_to_fp16, dilations = hidden_states_29_dilations_0, groups = hidden_states_29_groups_0, pad = hidden_states_29_pad_0, pad_type = hidden_states_29_pad_type_0, strides = hidden_states_29_strides_0, weight = layers_13_fc2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("hidden_states_29_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = hidden_states_29_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor var_3074 = const()[name = tensor("op_3074"), val = tensor(3)]; + tensor out_85_axes_0 = const()[name = tensor("out_85_axes_0"), val = tensor([1])]; + tensor var_3099_to_fp16 = const()[name = tensor("op_3099_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_3099_to_fp16, x = inputs_85_cast_fp16)[name = tensor("out_85_cast_fp16")]; + tensor obj_197_gamma_0_to_fp16 = const()[name = tensor("obj_197_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577412416)))]; + tensor obj_197_beta_0_to_fp16 = const()[name = tensor("obj_197_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577414528)))]; + tensor obj_197_epsilon_0_to_fp16 = const()[name = tensor("obj_197_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_197_cast_fp16 = batch_norm(beta = obj_197_beta_0_to_fp16, epsilon = obj_197_epsilon_0_to_fp16, gamma = obj_197_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor("obj_197_cast_fp16")]; + tensor query_57_pad_type_0 = const()[name = tensor("query_57_pad_type_0"), val = tensor("valid")]; + tensor query_57_strides_0 = const()[name = tensor("query_57_strides_0"), val = tensor([1, 1])]; + tensor query_57_pad_0 = const()[name = tensor("query_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_57_dilations_0 = const()[name = tensor("query_57_dilations_0"), val = tensor([1, 1])]; + tensor query_57_groups_0 = const()[name = tensor("query_57_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(577416640)))]; + tensor layers_14_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579513856)))]; + tensor query_57_cast_fp16 = conv(bias = layers_14_self_attn_q_proj_bias_to_fp16, dilations = query_57_dilations_0, groups = query_57_groups_0, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = query_57_strides_0, weight = layers_14_self_attn_q_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("query_57_cast_fp16")]; + tensor current_key_29_pad_type_0 = const()[name = tensor("current_key_29_pad_type_0"), val = tensor("valid")]; + tensor current_key_29_strides_0 = const()[name = tensor("current_key_29_strides_0"), val = tensor([1, 1])]; + tensor current_key_29_pad_0 = const()[name = tensor("current_key_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_29_dilations_0 = const()[name = tensor("current_key_29_dilations_0"), val = tensor([1, 1])]; + tensor current_key_29_groups_0 = const()[name = tensor("current_key_29_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579515968)))]; + tensor current_key_29_cast_fp16 = conv(dilations = current_key_29_dilations_0, groups = current_key_29_groups_0, pad = current_key_29_pad_0, pad_type = current_key_29_pad_type_0, strides = current_key_29_strides_0, weight = layers_14_self_attn_k_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("current_key_29_cast_fp16")]; + tensor current_value_29_pad_type_0 = const()[name = tensor("current_value_29_pad_type_0"), val = tensor("valid")]; + tensor current_value_29_strides_0 = const()[name = tensor("current_value_29_strides_0"), val = tensor([1, 1])]; + tensor current_value_29_pad_0 = const()[name = tensor("current_value_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_29_dilations_0 = const()[name = tensor("current_value_29_dilations_0"), val = tensor([1, 1])]; + tensor current_value_29_groups_0 = const()[name = tensor("current_value_29_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(581613184)))]; + tensor layers_14_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(583710400)))]; + tensor current_value_29_cast_fp16 = conv(bias = layers_14_self_attn_v_proj_bias_to_fp16, dilations = current_value_29_dilations_0, groups = current_value_29_groups_0, pad = current_value_29_pad_0, pad_type = current_value_29_pad_type_0, strides = current_value_29_strides_0, weight = layers_14_self_attn_v_proj_weight_to_fp16, x = obj_197_cast_fp16)[name = tensor("current_value_29_cast_fp16")]; + tensor var_3137_cast_fp16 = mul(x = current_key_29_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3137_cast_fp16")]; + tensor key_57_cast_fp16 = add(x = var_87_cast_fp16_14, y = var_3137_cast_fp16)[name = tensor("key_57_cast_fp16")]; + tensor var_3139_cast_fp16 = mul(x = current_value_29_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3139_cast_fp16")]; + tensor value_57_cast_fp16 = add(x = var_114_cast_fp16_14, y = var_3139_cast_fp16)[name = tensor("value_57_cast_fp16")]; + tensor var_3142 = const()[name = tensor("op_3142"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_57_cast_fp16 = reshape(shape = var_3142, x = query_57_cast_fp16)[name = tensor("mh_q_57_cast_fp16")]; + tensor var_3144_to_fp16 = const()[name = tensor("op_3144_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3145_cast_fp16 = mul(x = mh_q_57_cast_fp16, y = var_3144_to_fp16)[name = tensor("op_3145_cast_fp16")]; + tensor var_3146 = const()[name = tensor("op_3146"), val = tensor([1, 16, 64, -1])]; + tensor var_3147_cast_fp16 = reshape(shape = var_3146, x = key_57_cast_fp16)[name = tensor("op_3147_cast_fp16")]; + tensor mh_w_85_transpose_x_0 = const()[name = tensor("mh_w_85_transpose_x_0"), val = tensor(true)]; + tensor mh_w_85_transpose_y_0 = const()[name = tensor("mh_w_85_transpose_y_0"), val = tensor(false)]; + tensor mh_w_85_cast_fp16 = matmul(transpose_x = mh_w_85_transpose_x_0, transpose_y = mh_w_85_transpose_y_0, x = var_3145_cast_fp16, y = var_3147_cast_fp16)[name = tensor("mh_w_85_cast_fp16")]; + tensor mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_87_cast_fp16")]; + tensor var_3155_cast_fp16 = softmax(axis = var_3074, x = mh_w_87_cast_fp16)[name = tensor("op_3155_cast_fp16")]; + tensor var_3156 = const()[name = tensor("op_3156"), val = tensor([1, 16, 64, -1])]; + tensor var_3157_cast_fp16 = reshape(shape = var_3156, x = value_57_cast_fp16)[name = tensor("op_3157_cast_fp16")]; + tensor attn_57_transpose_x_0 = const()[name = tensor("attn_57_transpose_x_0"), val = tensor(false)]; + tensor attn_57_transpose_y_0 = const()[name = tensor("attn_57_transpose_y_0"), val = tensor(true)]; + tensor attn_57_cast_fp16 = matmul(transpose_x = attn_57_transpose_x_0, transpose_y = attn_57_transpose_y_0, x = var_3157_cast_fp16, y = var_3155_cast_fp16)[name = tensor("attn_57_cast_fp16")]; + tensor var_3160 = const()[name = tensor("op_3160"), val = tensor([1, 1024, 1, -1])]; + tensor input_141_cast_fp16 = reshape(shape = var_3160, x = attn_57_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor obj_203_pad_type_0 = const()[name = tensor("obj_203_pad_type_0"), val = tensor("valid")]; + tensor obj_203_strides_0 = const()[name = tensor("obj_203_strides_0"), val = tensor([1, 1])]; + tensor obj_203_pad_0 = const()[name = tensor("obj_203_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_203_dilations_0 = const()[name = tensor("obj_203_dilations_0"), val = tensor([1, 1])]; + tensor obj_203_groups_0 = const()[name = tensor("obj_203_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(583712512)))]; + tensor layers_14_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_14_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585809728)))]; + tensor obj_203_cast_fp16 = conv(bias = layers_14_self_attn_o_proj_bias_to_fp16, dilations = obj_203_dilations_0, groups = obj_203_groups_0, pad = obj_203_pad_0, pad_type = obj_203_pad_type_0, strides = obj_203_strides_0, weight = layers_14_self_attn_o_proj_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("obj_203_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = obj_203_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor out_87_axes_0 = const()[name = tensor("out_87_axes_0"), val = tensor([1])]; + tensor var_3182_to_fp16 = const()[name = tensor("op_3182_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_3182_to_fp16, x = inputs_87_cast_fp16)[name = tensor("out_87_cast_fp16")]; + tensor obj_205_gamma_0_to_fp16 = const()[name = tensor("obj_205_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585811840)))]; + tensor obj_205_beta_0_to_fp16 = const()[name = tensor("obj_205_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585813952)))]; + tensor obj_205_epsilon_0_to_fp16 = const()[name = tensor("obj_205_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_205_cast_fp16 = batch_norm(beta = obj_205_beta_0_to_fp16, epsilon = obj_205_epsilon_0_to_fp16, gamma = obj_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor("obj_205_cast_fp16")]; + tensor query_59_pad_type_0 = const()[name = tensor("query_59_pad_type_0"), val = tensor("valid")]; + tensor query_59_strides_0 = const()[name = tensor("query_59_strides_0"), val = tensor([1, 1])]; + tensor query_59_pad_0 = const()[name = tensor("query_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_59_dilations_0 = const()[name = tensor("query_59_dilations_0"), val = tensor([1, 1])]; + tensor query_59_groups_0 = const()[name = tensor("query_59_groups_0"), val = tensor(1)]; + tensor layers_14_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585816064)))]; + tensor layers_14_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587913280)))]; + tensor query_59_cast_fp16 = conv(bias = layers_14_encoder_attn_q_proj_bias_to_fp16, dilations = query_59_dilations_0, groups = query_59_groups_0, pad = query_59_pad_0, pad_type = query_59_pad_type_0, strides = query_59_strides_0, weight = layers_14_encoder_attn_q_proj_weight_to_fp16, x = obj_205_cast_fp16)[name = tensor("query_59_cast_fp16")]; + tensor key_59_pad_type_0 = const()[name = tensor("key_59_pad_type_0"), val = tensor("valid")]; + tensor key_59_strides_0 = const()[name = tensor("key_59_strides_0"), val = tensor([1, 1])]; + tensor key_59_pad_0 = const()[name = tensor("key_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_59_dilations_0 = const()[name = tensor("key_59_dilations_0"), val = tensor([1, 1])]; + tensor key_59_groups_0 = const()[name = tensor("key_59_groups_0"), val = tensor(1)]; + tensor layers_14_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587915392)))]; + tensor key_59_cast_fp16 = conv(dilations = key_59_dilations_0, groups = key_59_groups_0, pad = key_59_pad_0, pad_type = key_59_pad_type_0, strides = key_59_strides_0, weight = layers_14_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_59_cast_fp16")]; + tensor value_59_pad_type_0 = const()[name = tensor("value_59_pad_type_0"), val = tensor("valid")]; + tensor value_59_strides_0 = const()[name = tensor("value_59_strides_0"), val = tensor([1, 1])]; + tensor value_59_pad_0 = const()[name = tensor("value_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_59_dilations_0 = const()[name = tensor("value_59_dilations_0"), val = tensor([1, 1])]; + tensor value_59_groups_0 = const()[name = tensor("value_59_groups_0"), val = tensor(1)]; + tensor layers_14_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(590012608)))]; + tensor layers_14_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592109824)))]; + tensor value_59_cast_fp16 = conv(bias = layers_14_encoder_attn_v_proj_bias_to_fp16, dilations = value_59_dilations_0, groups = value_59_groups_0, pad = value_59_pad_0, pad_type = value_59_pad_type_0, strides = value_59_strides_0, weight = layers_14_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_59_cast_fp16")]; + tensor var_3217 = const()[name = tensor("op_3217"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_59_cast_fp16 = reshape(shape = var_3217, x = query_59_cast_fp16)[name = tensor("mh_q_59_cast_fp16")]; + tensor var_3219_to_fp16 = const()[name = tensor("op_3219_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3220_cast_fp16 = mul(x = mh_q_59_cast_fp16, y = var_3219_to_fp16)[name = tensor("op_3220_cast_fp16")]; + tensor var_3221 = const()[name = tensor("op_3221"), val = tensor([1, 16, 64, -1])]; + tensor var_3222_cast_fp16 = reshape(shape = var_3221, x = key_59_cast_fp16)[name = tensor("op_3222_cast_fp16")]; + tensor mh_w_89_transpose_x_0 = const()[name = tensor("mh_w_89_transpose_x_0"), val = tensor(true)]; + tensor mh_w_89_transpose_y_0 = const()[name = tensor("mh_w_89_transpose_y_0"), val = tensor(false)]; + tensor mh_w_89_cast_fp16 = matmul(transpose_x = mh_w_89_transpose_x_0, transpose_y = mh_w_89_transpose_y_0, x = var_3220_cast_fp16, y = var_3222_cast_fp16)[name = tensor("mh_w_89_cast_fp16")]; + tensor obj_209_cast_fp16 = softmax(axis = var_3074, x = mh_w_89_cast_fp16)[name = tensor("obj_209_cast_fp16")]; + tensor var_3226 = const()[name = tensor("op_3226"), val = tensor([1, 16, 64, -1])]; + tensor var_3227_cast_fp16 = reshape(shape = var_3226, x = value_59_cast_fp16)[name = tensor("op_3227_cast_fp16")]; + tensor attn_59_transpose_x_0 = const()[name = tensor("attn_59_transpose_x_0"), val = tensor(false)]; + tensor attn_59_transpose_y_0 = const()[name = tensor("attn_59_transpose_y_0"), val = tensor(true)]; + tensor attn_59_cast_fp16 = matmul(transpose_x = attn_59_transpose_x_0, transpose_y = attn_59_transpose_y_0, x = var_3227_cast_fp16, y = obj_209_cast_fp16)[name = tensor("attn_59_cast_fp16")]; + tensor var_3230 = const()[name = tensor("op_3230"), val = tensor([1, 1024, 1, -1])]; + tensor input_143_cast_fp16 = reshape(shape = var_3230, x = attn_59_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor obj_207_pad_type_0 = const()[name = tensor("obj_207_pad_type_0"), val = tensor("valid")]; + tensor obj_207_strides_0 = const()[name = tensor("obj_207_strides_0"), val = tensor([1, 1])]; + tensor obj_207_pad_0 = const()[name = tensor("obj_207_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_207_dilations_0 = const()[name = tensor("obj_207_dilations_0"), val = tensor([1, 1])]; + tensor obj_207_groups_0 = const()[name = tensor("obj_207_groups_0"), val = tensor(1)]; + tensor layers_14_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(592111936)))]; + tensor layers_14_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_14_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(594209152)))]; + tensor obj_207_cast_fp16 = conv(bias = layers_14_encoder_attn_o_proj_bias_to_fp16, dilations = obj_207_dilations_0, groups = obj_207_groups_0, pad = obj_207_pad_0, pad_type = obj_207_pad_type_0, strides = obj_207_strides_0, weight = layers_14_encoder_attn_o_proj_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("obj_207_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = obj_207_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor out_89_axes_0 = const()[name = tensor("out_89_axes_0"), val = tensor([1])]; + tensor var_3248_to_fp16 = const()[name = tensor("op_3248_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_3248_to_fp16, x = inputs_89_cast_fp16)[name = tensor("out_89_cast_fp16")]; + tensor input_145_gamma_0_to_fp16 = const()[name = tensor("input_145_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(594211264)))]; + tensor input_145_beta_0_to_fp16 = const()[name = tensor("input_145_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(594213376)))]; + tensor input_145_epsilon_0_to_fp16 = const()[name = tensor("input_145_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_145_cast_fp16 = batch_norm(beta = input_145_beta_0_to_fp16, epsilon = input_145_epsilon_0_to_fp16, gamma = input_145_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_pad_type_0 = const()[name = tensor("input_147_pad_type_0"), val = tensor("valid")]; + tensor input_147_strides_0 = const()[name = tensor("input_147_strides_0"), val = tensor([1, 1])]; + tensor input_147_pad_0 = const()[name = tensor("input_147_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_147_dilations_0 = const()[name = tensor("input_147_dilations_0"), val = tensor([1, 1])]; + tensor input_147_groups_0 = const()[name = tensor("input_147_groups_0"), val = tensor(1)]; + tensor layers_14_fc1_weight_to_fp16 = const()[name = tensor("layers_14_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(594215488)))]; + tensor layers_14_fc1_bias_to_fp16 = const()[name = tensor("layers_14_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(602604160)))]; + tensor input_147_cast_fp16 = conv(bias = layers_14_fc1_bias_to_fp16, dilations = input_147_dilations_0, groups = input_147_groups_0, pad = input_147_pad_0, pad_type = input_147_pad_type_0, strides = input_147_strides_0, weight = layers_14_fc1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_mode_0 = const()[name = tensor("input_149_mode_0"), val = tensor("EXACT")]; + tensor input_149_cast_fp16 = gelu(mode = input_149_mode_0, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor hidden_states_31_pad_type_0 = const()[name = tensor("hidden_states_31_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_31_strides_0 = const()[name = tensor("hidden_states_31_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_31_pad_0 = const()[name = tensor("hidden_states_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_31_dilations_0 = const()[name = tensor("hidden_states_31_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_31_groups_0 = const()[name = tensor("hidden_states_31_groups_0"), val = tensor(1)]; + tensor layers_14_fc2_weight_to_fp16 = const()[name = tensor("layers_14_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(602612416)))]; + tensor layers_14_fc2_bias_to_fp16 = const()[name = tensor("layers_14_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611001088)))]; + tensor hidden_states_31_cast_fp16 = conv(bias = layers_14_fc2_bias_to_fp16, dilations = hidden_states_31_dilations_0, groups = hidden_states_31_groups_0, pad = hidden_states_31_pad_0, pad_type = hidden_states_31_pad_type_0, strides = hidden_states_31_strides_0, weight = layers_14_fc2_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("hidden_states_31_cast_fp16")]; + tensor inputs_91_cast_fp16 = add(x = inputs_89_cast_fp16, y = hidden_states_31_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor var_3283 = const()[name = tensor("op_3283"), val = tensor(3)]; + tensor out_91_axes_0 = const()[name = tensor("out_91_axes_0"), val = tensor([1])]; + tensor var_3308_to_fp16 = const()[name = tensor("op_3308_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_3308_to_fp16, x = inputs_91_cast_fp16)[name = tensor("out_91_cast_fp16")]; + tensor obj_211_gamma_0_to_fp16 = const()[name = tensor("obj_211_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611003200)))]; + tensor obj_211_beta_0_to_fp16 = const()[name = tensor("obj_211_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611005312)))]; + tensor obj_211_epsilon_0_to_fp16 = const()[name = tensor("obj_211_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_211_cast_fp16 = batch_norm(beta = obj_211_beta_0_to_fp16, epsilon = obj_211_epsilon_0_to_fp16, gamma = obj_211_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor("obj_211_cast_fp16")]; + tensor query_61_pad_type_0 = const()[name = tensor("query_61_pad_type_0"), val = tensor("valid")]; + tensor query_61_strides_0 = const()[name = tensor("query_61_strides_0"), val = tensor([1, 1])]; + tensor query_61_pad_0 = const()[name = tensor("query_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_61_dilations_0 = const()[name = tensor("query_61_dilations_0"), val = tensor([1, 1])]; + tensor query_61_groups_0 = const()[name = tensor("query_61_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(611007424)))]; + tensor layers_15_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(613104640)))]; + tensor query_61_cast_fp16 = conv(bias = layers_15_self_attn_q_proj_bias_to_fp16, dilations = query_61_dilations_0, groups = query_61_groups_0, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = query_61_strides_0, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("query_61_cast_fp16")]; + tensor current_key_31_pad_type_0 = const()[name = tensor("current_key_31_pad_type_0"), val = tensor("valid")]; + tensor current_key_31_strides_0 = const()[name = tensor("current_key_31_strides_0"), val = tensor([1, 1])]; + tensor current_key_31_pad_0 = const()[name = tensor("current_key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_31_dilations_0 = const()[name = tensor("current_key_31_dilations_0"), val = tensor([1, 1])]; + tensor current_key_31_groups_0 = const()[name = tensor("current_key_31_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(613106752)))]; + tensor current_key_31_cast_fp16 = conv(dilations = current_key_31_dilations_0, groups = current_key_31_groups_0, pad = current_key_31_pad_0, pad_type = current_key_31_pad_type_0, strides = current_key_31_strides_0, weight = layers_15_self_attn_k_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("current_key_31_cast_fp16")]; + tensor current_value_31_pad_type_0 = const()[name = tensor("current_value_31_pad_type_0"), val = tensor("valid")]; + tensor current_value_31_strides_0 = const()[name = tensor("current_value_31_strides_0"), val = tensor([1, 1])]; + tensor current_value_31_pad_0 = const()[name = tensor("current_value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_31_dilations_0 = const()[name = tensor("current_value_31_dilations_0"), val = tensor([1, 1])]; + tensor current_value_31_groups_0 = const()[name = tensor("current_value_31_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(615203968)))]; + tensor layers_15_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617301184)))]; + tensor current_value_31_cast_fp16 = conv(bias = layers_15_self_attn_v_proj_bias_to_fp16, dilations = current_value_31_dilations_0, groups = current_value_31_groups_0, pad = current_value_31_pad_0, pad_type = current_value_31_pad_type_0, strides = current_value_31_strides_0, weight = layers_15_self_attn_v_proj_weight_to_fp16, x = obj_211_cast_fp16)[name = tensor("current_value_31_cast_fp16")]; + tensor var_3346_cast_fp16 = mul(x = current_key_31_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3346_cast_fp16")]; + tensor key_61_cast_fp16 = add(x = var_87_cast_fp16_15, y = var_3346_cast_fp16)[name = tensor("key_61_cast_fp16")]; + tensor var_3348_cast_fp16 = mul(x = current_value_31_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3348_cast_fp16")]; + tensor value_61_cast_fp16 = add(x = var_114_cast_fp16_15, y = var_3348_cast_fp16)[name = tensor("value_61_cast_fp16")]; + tensor var_3351 = const()[name = tensor("op_3351"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_61_cast_fp16 = reshape(shape = var_3351, x = query_61_cast_fp16)[name = tensor("mh_q_61_cast_fp16")]; + tensor var_3353_to_fp16 = const()[name = tensor("op_3353_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3354_cast_fp16 = mul(x = mh_q_61_cast_fp16, y = var_3353_to_fp16)[name = tensor("op_3354_cast_fp16")]; + tensor var_3355 = const()[name = tensor("op_3355"), val = tensor([1, 16, 64, -1])]; + tensor var_3356_cast_fp16 = reshape(shape = var_3355, x = key_61_cast_fp16)[name = tensor("op_3356_cast_fp16")]; + tensor mh_w_91_transpose_x_0 = const()[name = tensor("mh_w_91_transpose_x_0"), val = tensor(true)]; + tensor mh_w_91_transpose_y_0 = const()[name = tensor("mh_w_91_transpose_y_0"), val = tensor(false)]; + tensor mh_w_91_cast_fp16 = matmul(transpose_x = mh_w_91_transpose_x_0, transpose_y = mh_w_91_transpose_y_0, x = var_3354_cast_fp16, y = var_3356_cast_fp16)[name = tensor("mh_w_91_cast_fp16")]; + tensor mh_w_93_cast_fp16 = add(x = mh_w_91_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_93_cast_fp16")]; + tensor var_3364_cast_fp16 = softmax(axis = var_3283, x = mh_w_93_cast_fp16)[name = tensor("op_3364_cast_fp16")]; + tensor var_3365 = const()[name = tensor("op_3365"), val = tensor([1, 16, 64, -1])]; + tensor var_3366_cast_fp16 = reshape(shape = var_3365, x = value_61_cast_fp16)[name = tensor("op_3366_cast_fp16")]; + tensor attn_61_transpose_x_0 = const()[name = tensor("attn_61_transpose_x_0"), val = tensor(false)]; + tensor attn_61_transpose_y_0 = const()[name = tensor("attn_61_transpose_y_0"), val = tensor(true)]; + tensor attn_61_cast_fp16 = matmul(transpose_x = attn_61_transpose_x_0, transpose_y = attn_61_transpose_y_0, x = var_3366_cast_fp16, y = var_3364_cast_fp16)[name = tensor("attn_61_cast_fp16")]; + tensor var_3369 = const()[name = tensor("op_3369"), val = tensor([1, 1024, 1, -1])]; + tensor input_151_cast_fp16 = reshape(shape = var_3369, x = attn_61_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor obj_217_pad_type_0 = const()[name = tensor("obj_217_pad_type_0"), val = tensor("valid")]; + tensor obj_217_strides_0 = const()[name = tensor("obj_217_strides_0"), val = tensor([1, 1])]; + tensor obj_217_pad_0 = const()[name = tensor("obj_217_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_217_dilations_0 = const()[name = tensor("obj_217_dilations_0"), val = tensor([1, 1])]; + tensor obj_217_groups_0 = const()[name = tensor("obj_217_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(617303296)))]; + tensor layers_15_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_15_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619400512)))]; + tensor obj_217_cast_fp16 = conv(bias = layers_15_self_attn_o_proj_bias_to_fp16, dilations = obj_217_dilations_0, groups = obj_217_groups_0, pad = obj_217_pad_0, pad_type = obj_217_pad_type_0, strides = obj_217_strides_0, weight = layers_15_self_attn_o_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("obj_217_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = obj_217_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor out_93_axes_0 = const()[name = tensor("out_93_axes_0"), val = tensor([1])]; + tensor var_3391_to_fp16 = const()[name = tensor("op_3391_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_3391_to_fp16, x = inputs_93_cast_fp16)[name = tensor("out_93_cast_fp16")]; + tensor obj_219_gamma_0_to_fp16 = const()[name = tensor("obj_219_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619402624)))]; + tensor obj_219_beta_0_to_fp16 = const()[name = tensor("obj_219_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619404736)))]; + tensor obj_219_epsilon_0_to_fp16 = const()[name = tensor("obj_219_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_219_cast_fp16 = batch_norm(beta = obj_219_beta_0_to_fp16, epsilon = obj_219_epsilon_0_to_fp16, gamma = obj_219_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor("obj_219_cast_fp16")]; + tensor query_63_pad_type_0 = const()[name = tensor("query_63_pad_type_0"), val = tensor("valid")]; + tensor query_63_strides_0 = const()[name = tensor("query_63_strides_0"), val = tensor([1, 1])]; + tensor query_63_pad_0 = const()[name = tensor("query_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_63_dilations_0 = const()[name = tensor("query_63_dilations_0"), val = tensor([1, 1])]; + tensor query_63_groups_0 = const()[name = tensor("query_63_groups_0"), val = tensor(1)]; + tensor layers_15_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(619406848)))]; + tensor layers_15_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(621504064)))]; + tensor query_63_cast_fp16 = conv(bias = layers_15_encoder_attn_q_proj_bias_to_fp16, dilations = query_63_dilations_0, groups = query_63_groups_0, pad = query_63_pad_0, pad_type = query_63_pad_type_0, strides = query_63_strides_0, weight = layers_15_encoder_attn_q_proj_weight_to_fp16, x = obj_219_cast_fp16)[name = tensor("query_63_cast_fp16")]; + tensor key_63_pad_type_0 = const()[name = tensor("key_63_pad_type_0"), val = tensor("valid")]; + tensor key_63_strides_0 = const()[name = tensor("key_63_strides_0"), val = tensor([1, 1])]; + tensor key_63_pad_0 = const()[name = tensor("key_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_63_dilations_0 = const()[name = tensor("key_63_dilations_0"), val = tensor([1, 1])]; + tensor key_63_groups_0 = const()[name = tensor("key_63_groups_0"), val = tensor(1)]; + tensor layers_15_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(621506176)))]; + tensor key_63_cast_fp16 = conv(dilations = key_63_dilations_0, groups = key_63_groups_0, pad = key_63_pad_0, pad_type = key_63_pad_type_0, strides = key_63_strides_0, weight = layers_15_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_63_cast_fp16")]; + tensor value_63_pad_type_0 = const()[name = tensor("value_63_pad_type_0"), val = tensor("valid")]; + tensor value_63_strides_0 = const()[name = tensor("value_63_strides_0"), val = tensor([1, 1])]; + tensor value_63_pad_0 = const()[name = tensor("value_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_63_dilations_0 = const()[name = tensor("value_63_dilations_0"), val = tensor([1, 1])]; + tensor value_63_groups_0 = const()[name = tensor("value_63_groups_0"), val = tensor(1)]; + tensor layers_15_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(623603392)))]; + tensor layers_15_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(625700608)))]; + tensor value_63_cast_fp16 = conv(bias = layers_15_encoder_attn_v_proj_bias_to_fp16, dilations = value_63_dilations_0, groups = value_63_groups_0, pad = value_63_pad_0, pad_type = value_63_pad_type_0, strides = value_63_strides_0, weight = layers_15_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_63_cast_fp16")]; + tensor var_3426 = const()[name = tensor("op_3426"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_63_cast_fp16 = reshape(shape = var_3426, x = query_63_cast_fp16)[name = tensor("mh_q_63_cast_fp16")]; + tensor var_3428_to_fp16 = const()[name = tensor("op_3428_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3429_cast_fp16 = mul(x = mh_q_63_cast_fp16, y = var_3428_to_fp16)[name = tensor("op_3429_cast_fp16")]; + tensor var_3430 = const()[name = tensor("op_3430"), val = tensor([1, 16, 64, -1])]; + tensor var_3431_cast_fp16 = reshape(shape = var_3430, x = key_63_cast_fp16)[name = tensor("op_3431_cast_fp16")]; + tensor mh_w_95_transpose_x_0 = const()[name = tensor("mh_w_95_transpose_x_0"), val = tensor(true)]; + tensor mh_w_95_transpose_y_0 = const()[name = tensor("mh_w_95_transpose_y_0"), val = tensor(false)]; + tensor mh_w_95_cast_fp16 = matmul(transpose_x = mh_w_95_transpose_x_0, transpose_y = mh_w_95_transpose_y_0, x = var_3429_cast_fp16, y = var_3431_cast_fp16)[name = tensor("mh_w_95_cast_fp16")]; + tensor obj_223_cast_fp16 = softmax(axis = var_3283, x = mh_w_95_cast_fp16)[name = tensor("obj_223_cast_fp16")]; + tensor var_3435 = const()[name = tensor("op_3435"), val = tensor([1, 16, 64, -1])]; + tensor var_3436_cast_fp16 = reshape(shape = var_3435, x = value_63_cast_fp16)[name = tensor("op_3436_cast_fp16")]; + tensor attn_63_transpose_x_0 = const()[name = tensor("attn_63_transpose_x_0"), val = tensor(false)]; + tensor attn_63_transpose_y_0 = const()[name = tensor("attn_63_transpose_y_0"), val = tensor(true)]; + tensor attn_63_cast_fp16 = matmul(transpose_x = attn_63_transpose_x_0, transpose_y = attn_63_transpose_y_0, x = var_3436_cast_fp16, y = obj_223_cast_fp16)[name = tensor("attn_63_cast_fp16")]; + tensor var_3439 = const()[name = tensor("op_3439"), val = tensor([1, 1024, 1, -1])]; + tensor input_153_cast_fp16 = reshape(shape = var_3439, x = attn_63_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor obj_221_pad_type_0 = const()[name = tensor("obj_221_pad_type_0"), val = tensor("valid")]; + tensor obj_221_strides_0 = const()[name = tensor("obj_221_strides_0"), val = tensor([1, 1])]; + tensor obj_221_pad_0 = const()[name = tensor("obj_221_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_221_dilations_0 = const()[name = tensor("obj_221_dilations_0"), val = tensor([1, 1])]; + tensor obj_221_groups_0 = const()[name = tensor("obj_221_groups_0"), val = tensor(1)]; + tensor layers_15_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(625702720)))]; + tensor layers_15_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_15_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(627799936)))]; + tensor obj_221_cast_fp16 = conv(bias = layers_15_encoder_attn_o_proj_bias_to_fp16, dilations = obj_221_dilations_0, groups = obj_221_groups_0, pad = obj_221_pad_0, pad_type = obj_221_pad_type_0, strides = obj_221_strides_0, weight = layers_15_encoder_attn_o_proj_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("obj_221_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_221_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor out_95_axes_0 = const()[name = tensor("out_95_axes_0"), val = tensor([1])]; + tensor var_3460_to_fp16 = const()[name = tensor("op_3460_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_3460_to_fp16, x = inputs_95_cast_fp16)[name = tensor("out_95_cast_fp16")]; + tensor input_155_gamma_0_to_fp16 = const()[name = tensor("input_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(627802048)))]; + tensor input_155_beta_0_to_fp16 = const()[name = tensor("input_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(627804160)))]; + tensor input_155_epsilon_0_to_fp16 = const()[name = tensor("input_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = batch_norm(beta = input_155_beta_0_to_fp16, epsilon = input_155_epsilon_0_to_fp16, gamma = input_155_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor input_157_pad_type_0 = const()[name = tensor("input_157_pad_type_0"), val = tensor("valid")]; + tensor input_157_strides_0 = const()[name = tensor("input_157_strides_0"), val = tensor([1, 1])]; + tensor input_157_pad_0 = const()[name = tensor("input_157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_157_dilations_0 = const()[name = tensor("input_157_dilations_0"), val = tensor([1, 1])]; + tensor input_157_groups_0 = const()[name = tensor("input_157_groups_0"), val = tensor(1)]; + tensor layers_15_fc1_weight_to_fp16 = const()[name = tensor("layers_15_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(627806272)))]; + tensor layers_15_fc1_bias_to_fp16 = const()[name = tensor("layers_15_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(636194944)))]; + tensor input_157_cast_fp16 = conv(bias = layers_15_fc1_bias_to_fp16, dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = layers_15_fc1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_mode_0 = const()[name = tensor("input_159_mode_0"), val = tensor("EXACT")]; + tensor input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor hidden_states_33_pad_type_0 = const()[name = tensor("hidden_states_33_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_33_strides_0 = const()[name = tensor("hidden_states_33_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_33_pad_0 = const()[name = tensor("hidden_states_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_33_dilations_0 = const()[name = tensor("hidden_states_33_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_33_groups_0 = const()[name = tensor("hidden_states_33_groups_0"), val = tensor(1)]; + tensor layers_15_fc2_weight_to_fp16 = const()[name = tensor("layers_15_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(636203200)))]; + tensor layers_15_fc2_bias_to_fp16 = const()[name = tensor("layers_15_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644591872)))]; + tensor hidden_states_33_cast_fp16 = conv(bias = layers_15_fc2_bias_to_fp16, dilations = hidden_states_33_dilations_0, groups = hidden_states_33_groups_0, pad = hidden_states_33_pad_0, pad_type = hidden_states_33_pad_type_0, strides = hidden_states_33_strides_0, weight = layers_15_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("hidden_states_33_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = hidden_states_33_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; + tensor var_3496 = const()[name = tensor("op_3496"), val = tensor(3)]; + tensor out_97_axes_0 = const()[name = tensor("out_97_axes_0"), val = tensor([1])]; + tensor var_3521_to_fp16 = const()[name = tensor("op_3521_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_3521_to_fp16, x = inputs_97_cast_fp16)[name = tensor("out_97_cast_fp16")]; + tensor obj_225_gamma_0_to_fp16 = const()[name = tensor("obj_225_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644593984)))]; + tensor obj_225_beta_0_to_fp16 = const()[name = tensor("obj_225_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644596096)))]; + tensor obj_225_epsilon_0_to_fp16 = const()[name = tensor("obj_225_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_225_cast_fp16 = batch_norm(beta = obj_225_beta_0_to_fp16, epsilon = obj_225_epsilon_0_to_fp16, gamma = obj_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor("obj_225_cast_fp16")]; + tensor query_65_pad_type_0 = const()[name = tensor("query_65_pad_type_0"), val = tensor("valid")]; + tensor query_65_strides_0 = const()[name = tensor("query_65_strides_0"), val = tensor([1, 1])]; + tensor query_65_pad_0 = const()[name = tensor("query_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_65_dilations_0 = const()[name = tensor("query_65_dilations_0"), val = tensor([1, 1])]; + tensor query_65_groups_0 = const()[name = tensor("query_65_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(644598208)))]; + tensor layers_16_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(646695424)))]; + tensor query_65_cast_fp16 = conv(bias = layers_16_self_attn_q_proj_bias_to_fp16, dilations = query_65_dilations_0, groups = query_65_groups_0, pad = query_65_pad_0, pad_type = query_65_pad_type_0, strides = query_65_strides_0, weight = layers_16_self_attn_q_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("query_65_cast_fp16")]; + tensor current_key_33_pad_type_0 = const()[name = tensor("current_key_33_pad_type_0"), val = tensor("valid")]; + tensor current_key_33_strides_0 = const()[name = tensor("current_key_33_strides_0"), val = tensor([1, 1])]; + tensor current_key_33_pad_0 = const()[name = tensor("current_key_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_33_dilations_0 = const()[name = tensor("current_key_33_dilations_0"), val = tensor([1, 1])]; + tensor current_key_33_groups_0 = const()[name = tensor("current_key_33_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(646697536)))]; + tensor current_key_33_cast_fp16 = conv(dilations = current_key_33_dilations_0, groups = current_key_33_groups_0, pad = current_key_33_pad_0, pad_type = current_key_33_pad_type_0, strides = current_key_33_strides_0, weight = layers_16_self_attn_k_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("current_key_33_cast_fp16")]; + tensor current_value_33_pad_type_0 = const()[name = tensor("current_value_33_pad_type_0"), val = tensor("valid")]; + tensor current_value_33_strides_0 = const()[name = tensor("current_value_33_strides_0"), val = tensor([1, 1])]; + tensor current_value_33_pad_0 = const()[name = tensor("current_value_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_33_dilations_0 = const()[name = tensor("current_value_33_dilations_0"), val = tensor([1, 1])]; + tensor current_value_33_groups_0 = const()[name = tensor("current_value_33_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(648794752)))]; + tensor layers_16_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650891968)))]; + tensor current_value_33_cast_fp16 = conv(bias = layers_16_self_attn_v_proj_bias_to_fp16, dilations = current_value_33_dilations_0, groups = current_value_33_groups_0, pad = current_value_33_pad_0, pad_type = current_value_33_pad_type_0, strides = current_value_33_strides_0, weight = layers_16_self_attn_v_proj_weight_to_fp16, x = obj_225_cast_fp16)[name = tensor("current_value_33_cast_fp16")]; + tensor var_3559_cast_fp16 = mul(x = current_key_33_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3559_cast_fp16")]; + tensor key_65_cast_fp16 = add(x = var_87_cast_fp16_16, y = var_3559_cast_fp16)[name = tensor("key_65_cast_fp16")]; + tensor var_3561_cast_fp16 = mul(x = current_value_33_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3561_cast_fp16")]; + tensor value_65_cast_fp16 = add(x = var_114_cast_fp16_16, y = var_3561_cast_fp16)[name = tensor("value_65_cast_fp16")]; + tensor var_3564 = const()[name = tensor("op_3564"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_65_cast_fp16 = reshape(shape = var_3564, x = query_65_cast_fp16)[name = tensor("mh_q_65_cast_fp16")]; + tensor var_3566_to_fp16 = const()[name = tensor("op_3566_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3567_cast_fp16 = mul(x = mh_q_65_cast_fp16, y = var_3566_to_fp16)[name = tensor("op_3567_cast_fp16")]; + tensor var_3568 = const()[name = tensor("op_3568"), val = tensor([1, 16, 64, -1])]; + tensor var_3569_cast_fp16 = reshape(shape = var_3568, x = key_65_cast_fp16)[name = tensor("op_3569_cast_fp16")]; + tensor mh_w_97_transpose_x_0 = const()[name = tensor("mh_w_97_transpose_x_0"), val = tensor(true)]; + tensor mh_w_97_transpose_y_0 = const()[name = tensor("mh_w_97_transpose_y_0"), val = tensor(false)]; + tensor mh_w_97_cast_fp16 = matmul(transpose_x = mh_w_97_transpose_x_0, transpose_y = mh_w_97_transpose_y_0, x = var_3567_cast_fp16, y = var_3569_cast_fp16)[name = tensor("mh_w_97_cast_fp16")]; + tensor mh_w_99_cast_fp16 = add(x = mh_w_97_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_99_cast_fp16")]; + tensor var_3577_cast_fp16 = softmax(axis = var_3496, x = mh_w_99_cast_fp16)[name = tensor("op_3577_cast_fp16")]; + tensor var_3578 = const()[name = tensor("op_3578"), val = tensor([1, 16, 64, -1])]; + tensor var_3579_cast_fp16 = reshape(shape = var_3578, x = value_65_cast_fp16)[name = tensor("op_3579_cast_fp16")]; + tensor attn_65_transpose_x_0 = const()[name = tensor("attn_65_transpose_x_0"), val = tensor(false)]; + tensor attn_65_transpose_y_0 = const()[name = tensor("attn_65_transpose_y_0"), val = tensor(true)]; + tensor attn_65_cast_fp16 = matmul(transpose_x = attn_65_transpose_x_0, transpose_y = attn_65_transpose_y_0, x = var_3579_cast_fp16, y = var_3577_cast_fp16)[name = tensor("attn_65_cast_fp16")]; + tensor var_3582 = const()[name = tensor("op_3582"), val = tensor([1, 1024, 1, -1])]; + tensor input_161_cast_fp16 = reshape(shape = var_3582, x = attn_65_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor obj_231_pad_type_0 = const()[name = tensor("obj_231_pad_type_0"), val = tensor("valid")]; + tensor obj_231_strides_0 = const()[name = tensor("obj_231_strides_0"), val = tensor([1, 1])]; + tensor obj_231_pad_0 = const()[name = tensor("obj_231_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_231_dilations_0 = const()[name = tensor("obj_231_dilations_0"), val = tensor([1, 1])]; + tensor obj_231_groups_0 = const()[name = tensor("obj_231_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(650894080)))]; + tensor layers_16_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_16_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(652991296)))]; + tensor obj_231_cast_fp16 = conv(bias = layers_16_self_attn_o_proj_bias_to_fp16, dilations = obj_231_dilations_0, groups = obj_231_groups_0, pad = obj_231_pad_0, pad_type = obj_231_pad_type_0, strides = obj_231_strides_0, weight = layers_16_self_attn_o_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("obj_231_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = obj_231_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; + tensor out_99_axes_0 = const()[name = tensor("out_99_axes_0"), val = tensor([1])]; + tensor var_3604_to_fp16 = const()[name = tensor("op_3604_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_3604_to_fp16, x = inputs_99_cast_fp16)[name = tensor("out_99_cast_fp16")]; + tensor obj_233_gamma_0_to_fp16 = const()[name = tensor("obj_233_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(652993408)))]; + tensor obj_233_beta_0_to_fp16 = const()[name = tensor("obj_233_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(652995520)))]; + tensor obj_233_epsilon_0_to_fp16 = const()[name = tensor("obj_233_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_233_cast_fp16 = batch_norm(beta = obj_233_beta_0_to_fp16, epsilon = obj_233_epsilon_0_to_fp16, gamma = obj_233_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor("obj_233_cast_fp16")]; + tensor query_67_pad_type_0 = const()[name = tensor("query_67_pad_type_0"), val = tensor("valid")]; + tensor query_67_strides_0 = const()[name = tensor("query_67_strides_0"), val = tensor([1, 1])]; + tensor query_67_pad_0 = const()[name = tensor("query_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_67_dilations_0 = const()[name = tensor("query_67_dilations_0"), val = tensor([1, 1])]; + tensor query_67_groups_0 = const()[name = tensor("query_67_groups_0"), val = tensor(1)]; + tensor layers_16_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(652997632)))]; + tensor layers_16_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(655094848)))]; + tensor query_67_cast_fp16 = conv(bias = layers_16_encoder_attn_q_proj_bias_to_fp16, dilations = query_67_dilations_0, groups = query_67_groups_0, pad = query_67_pad_0, pad_type = query_67_pad_type_0, strides = query_67_strides_0, weight = layers_16_encoder_attn_q_proj_weight_to_fp16, x = obj_233_cast_fp16)[name = tensor("query_67_cast_fp16")]; + tensor key_67_pad_type_0 = const()[name = tensor("key_67_pad_type_0"), val = tensor("valid")]; + tensor key_67_strides_0 = const()[name = tensor("key_67_strides_0"), val = tensor([1, 1])]; + tensor key_67_pad_0 = const()[name = tensor("key_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_67_dilations_0 = const()[name = tensor("key_67_dilations_0"), val = tensor([1, 1])]; + tensor key_67_groups_0 = const()[name = tensor("key_67_groups_0"), val = tensor(1)]; + tensor layers_16_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(655096960)))]; + tensor key_67_cast_fp16 = conv(dilations = key_67_dilations_0, groups = key_67_groups_0, pad = key_67_pad_0, pad_type = key_67_pad_type_0, strides = key_67_strides_0, weight = layers_16_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_67_cast_fp16")]; + tensor value_67_pad_type_0 = const()[name = tensor("value_67_pad_type_0"), val = tensor("valid")]; + tensor value_67_strides_0 = const()[name = tensor("value_67_strides_0"), val = tensor([1, 1])]; + tensor value_67_pad_0 = const()[name = tensor("value_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_67_dilations_0 = const()[name = tensor("value_67_dilations_0"), val = tensor([1, 1])]; + tensor value_67_groups_0 = const()[name = tensor("value_67_groups_0"), val = tensor(1)]; + tensor layers_16_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(657194176)))]; + tensor layers_16_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(659291392)))]; + tensor value_67_cast_fp16 = conv(bias = layers_16_encoder_attn_v_proj_bias_to_fp16, dilations = value_67_dilations_0, groups = value_67_groups_0, pad = value_67_pad_0, pad_type = value_67_pad_type_0, strides = value_67_strides_0, weight = layers_16_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_67_cast_fp16")]; + tensor var_3639 = const()[name = tensor("op_3639"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_67_cast_fp16 = reshape(shape = var_3639, x = query_67_cast_fp16)[name = tensor("mh_q_67_cast_fp16")]; + tensor var_3641_to_fp16 = const()[name = tensor("op_3641_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3642_cast_fp16 = mul(x = mh_q_67_cast_fp16, y = var_3641_to_fp16)[name = tensor("op_3642_cast_fp16")]; + tensor var_3643 = const()[name = tensor("op_3643"), val = tensor([1, 16, 64, -1])]; + tensor var_3644_cast_fp16 = reshape(shape = var_3643, x = key_67_cast_fp16)[name = tensor("op_3644_cast_fp16")]; + tensor mh_w_101_transpose_x_0 = const()[name = tensor("mh_w_101_transpose_x_0"), val = tensor(true)]; + tensor mh_w_101_transpose_y_0 = const()[name = tensor("mh_w_101_transpose_y_0"), val = tensor(false)]; + tensor mh_w_101_cast_fp16 = matmul(transpose_x = mh_w_101_transpose_x_0, transpose_y = mh_w_101_transpose_y_0, x = var_3642_cast_fp16, y = var_3644_cast_fp16)[name = tensor("mh_w_101_cast_fp16")]; + tensor obj_237_cast_fp16 = softmax(axis = var_3496, x = mh_w_101_cast_fp16)[name = tensor("obj_237_cast_fp16")]; + tensor var_3648 = const()[name = tensor("op_3648"), val = tensor([1, 16, 64, -1])]; + tensor var_3649_cast_fp16 = reshape(shape = var_3648, x = value_67_cast_fp16)[name = tensor("op_3649_cast_fp16")]; + tensor attn_67_transpose_x_0 = const()[name = tensor("attn_67_transpose_x_0"), val = tensor(false)]; + tensor attn_67_transpose_y_0 = const()[name = tensor("attn_67_transpose_y_0"), val = tensor(true)]; + tensor attn_67_cast_fp16 = matmul(transpose_x = attn_67_transpose_x_0, transpose_y = attn_67_transpose_y_0, x = var_3649_cast_fp16, y = obj_237_cast_fp16)[name = tensor("attn_67_cast_fp16")]; + tensor var_3652 = const()[name = tensor("op_3652"), val = tensor([1, 1024, 1, -1])]; + tensor input_163_cast_fp16 = reshape(shape = var_3652, x = attn_67_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor obj_235_pad_type_0 = const()[name = tensor("obj_235_pad_type_0"), val = tensor("valid")]; + tensor obj_235_strides_0 = const()[name = tensor("obj_235_strides_0"), val = tensor([1, 1])]; + tensor obj_235_pad_0 = const()[name = tensor("obj_235_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_235_dilations_0 = const()[name = tensor("obj_235_dilations_0"), val = tensor([1, 1])]; + tensor obj_235_groups_0 = const()[name = tensor("obj_235_groups_0"), val = tensor(1)]; + tensor layers_16_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(659293504)))]; + tensor layers_16_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_16_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(661390720)))]; + tensor obj_235_cast_fp16 = conv(bias = layers_16_encoder_attn_o_proj_bias_to_fp16, dilations = obj_235_dilations_0, groups = obj_235_groups_0, pad = obj_235_pad_0, pad_type = obj_235_pad_type_0, strides = obj_235_strides_0, weight = layers_16_encoder_attn_o_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("obj_235_cast_fp16")]; + tensor inputs_101_cast_fp16 = add(x = inputs_99_cast_fp16, y = obj_235_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor out_101_axes_0 = const()[name = tensor("out_101_axes_0"), val = tensor([1])]; + tensor var_3673_to_fp16 = const()[name = tensor("op_3673_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_3673_to_fp16, x = inputs_101_cast_fp16)[name = tensor("out_101_cast_fp16")]; + tensor input_165_gamma_0_to_fp16 = const()[name = tensor("input_165_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(661392832)))]; + tensor input_165_beta_0_to_fp16 = const()[name = tensor("input_165_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(661394944)))]; + tensor input_165_epsilon_0_to_fp16 = const()[name = tensor("input_165_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = batch_norm(beta = input_165_beta_0_to_fp16, epsilon = input_165_epsilon_0_to_fp16, gamma = input_165_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_pad_type_0 = const()[name = tensor("input_167_pad_type_0"), val = tensor("valid")]; + tensor input_167_strides_0 = const()[name = tensor("input_167_strides_0"), val = tensor([1, 1])]; + tensor input_167_pad_0 = const()[name = tensor("input_167_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_167_dilations_0 = const()[name = tensor("input_167_dilations_0"), val = tensor([1, 1])]; + tensor input_167_groups_0 = const()[name = tensor("input_167_groups_0"), val = tensor(1)]; + tensor layers_16_fc1_weight_to_fp16 = const()[name = tensor("layers_16_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(661397056)))]; + tensor layers_16_fc1_bias_to_fp16 = const()[name = tensor("layers_16_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(669785728)))]; + tensor input_167_cast_fp16 = conv(bias = layers_16_fc1_bias_to_fp16, dilations = input_167_dilations_0, groups = input_167_groups_0, pad = input_167_pad_0, pad_type = input_167_pad_type_0, strides = input_167_strides_0, weight = layers_16_fc1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor hidden_states_35_pad_type_0 = const()[name = tensor("hidden_states_35_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_35_strides_0 = const()[name = tensor("hidden_states_35_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_35_pad_0 = const()[name = tensor("hidden_states_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_35_dilations_0 = const()[name = tensor("hidden_states_35_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_35_groups_0 = const()[name = tensor("hidden_states_35_groups_0"), val = tensor(1)]; + tensor layers_16_fc2_weight_to_fp16 = const()[name = tensor("layers_16_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(669793984)))]; + tensor layers_16_fc2_bias_to_fp16 = const()[name = tensor("layers_16_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(678182656)))]; + tensor hidden_states_35_cast_fp16 = conv(bias = layers_16_fc2_bias_to_fp16, dilations = hidden_states_35_dilations_0, groups = hidden_states_35_groups_0, pad = hidden_states_35_pad_0, pad_type = hidden_states_35_pad_type_0, strides = hidden_states_35_strides_0, weight = layers_16_fc2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("hidden_states_35_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = hidden_states_35_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; + tensor var_3709 = const()[name = tensor("op_3709"), val = tensor(3)]; + tensor out_103_axes_0 = const()[name = tensor("out_103_axes_0"), val = tensor([1])]; + tensor var_3734_to_fp16 = const()[name = tensor("op_3734_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_3734_to_fp16, x = inputs_103_cast_fp16)[name = tensor("out_103_cast_fp16")]; + tensor obj_239_gamma_0_to_fp16 = const()[name = tensor("obj_239_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(678184768)))]; + tensor obj_239_beta_0_to_fp16 = const()[name = tensor("obj_239_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(678186880)))]; + tensor obj_239_epsilon_0_to_fp16 = const()[name = tensor("obj_239_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_239_cast_fp16 = batch_norm(beta = obj_239_beta_0_to_fp16, epsilon = obj_239_epsilon_0_to_fp16, gamma = obj_239_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor("obj_239_cast_fp16")]; + tensor query_69_pad_type_0 = const()[name = tensor("query_69_pad_type_0"), val = tensor("valid")]; + tensor query_69_strides_0 = const()[name = tensor("query_69_strides_0"), val = tensor([1, 1])]; + tensor query_69_pad_0 = const()[name = tensor("query_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_69_dilations_0 = const()[name = tensor("query_69_dilations_0"), val = tensor([1, 1])]; + tensor query_69_groups_0 = const()[name = tensor("query_69_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(678188992)))]; + tensor layers_17_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(680286208)))]; + tensor query_69_cast_fp16 = conv(bias = layers_17_self_attn_q_proj_bias_to_fp16, dilations = query_69_dilations_0, groups = query_69_groups_0, pad = query_69_pad_0, pad_type = query_69_pad_type_0, strides = query_69_strides_0, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("query_69_cast_fp16")]; + tensor current_key_35_pad_type_0 = const()[name = tensor("current_key_35_pad_type_0"), val = tensor("valid")]; + tensor current_key_35_strides_0 = const()[name = tensor("current_key_35_strides_0"), val = tensor([1, 1])]; + tensor current_key_35_pad_0 = const()[name = tensor("current_key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_35_dilations_0 = const()[name = tensor("current_key_35_dilations_0"), val = tensor([1, 1])]; + tensor current_key_35_groups_0 = const()[name = tensor("current_key_35_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(680288320)))]; + tensor current_key_35_cast_fp16 = conv(dilations = current_key_35_dilations_0, groups = current_key_35_groups_0, pad = current_key_35_pad_0, pad_type = current_key_35_pad_type_0, strides = current_key_35_strides_0, weight = layers_17_self_attn_k_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("current_key_35_cast_fp16")]; + tensor current_value_35_pad_type_0 = const()[name = tensor("current_value_35_pad_type_0"), val = tensor("valid")]; + tensor current_value_35_strides_0 = const()[name = tensor("current_value_35_strides_0"), val = tensor([1, 1])]; + tensor current_value_35_pad_0 = const()[name = tensor("current_value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_35_dilations_0 = const()[name = tensor("current_value_35_dilations_0"), val = tensor([1, 1])]; + tensor current_value_35_groups_0 = const()[name = tensor("current_value_35_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(682385536)))]; + tensor layers_17_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(684482752)))]; + tensor current_value_35_cast_fp16 = conv(bias = layers_17_self_attn_v_proj_bias_to_fp16, dilations = current_value_35_dilations_0, groups = current_value_35_groups_0, pad = current_value_35_pad_0, pad_type = current_value_35_pad_type_0, strides = current_value_35_strides_0, weight = layers_17_self_attn_v_proj_weight_to_fp16, x = obj_239_cast_fp16)[name = tensor("current_value_35_cast_fp16")]; + tensor var_3772_cast_fp16 = mul(x = current_key_35_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3772_cast_fp16")]; + tensor key_69_cast_fp16 = add(x = var_87_cast_fp16_17, y = var_3772_cast_fp16)[name = tensor("key_69_cast_fp16")]; + tensor var_3774_cast_fp16 = mul(x = current_value_35_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3774_cast_fp16")]; + tensor value_69_cast_fp16 = add(x = var_114_cast_fp16_17, y = var_3774_cast_fp16)[name = tensor("value_69_cast_fp16")]; + tensor var_3777 = const()[name = tensor("op_3777"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_69_cast_fp16 = reshape(shape = var_3777, x = query_69_cast_fp16)[name = tensor("mh_q_69_cast_fp16")]; + tensor var_3779_to_fp16 = const()[name = tensor("op_3779_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3780_cast_fp16 = mul(x = mh_q_69_cast_fp16, y = var_3779_to_fp16)[name = tensor("op_3780_cast_fp16")]; + tensor var_3781 = const()[name = tensor("op_3781"), val = tensor([1, 16, 64, -1])]; + tensor var_3782_cast_fp16 = reshape(shape = var_3781, x = key_69_cast_fp16)[name = tensor("op_3782_cast_fp16")]; + tensor mh_w_103_transpose_x_0 = const()[name = tensor("mh_w_103_transpose_x_0"), val = tensor(true)]; + tensor mh_w_103_transpose_y_0 = const()[name = tensor("mh_w_103_transpose_y_0"), val = tensor(false)]; + tensor mh_w_103_cast_fp16 = matmul(transpose_x = mh_w_103_transpose_x_0, transpose_y = mh_w_103_transpose_y_0, x = var_3780_cast_fp16, y = var_3782_cast_fp16)[name = tensor("mh_w_103_cast_fp16")]; + tensor mh_w_105_cast_fp16 = add(x = mh_w_103_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_105_cast_fp16")]; + tensor var_3790_cast_fp16 = softmax(axis = var_3709, x = mh_w_105_cast_fp16)[name = tensor("op_3790_cast_fp16")]; + tensor var_3791 = const()[name = tensor("op_3791"), val = tensor([1, 16, 64, -1])]; + tensor var_3792_cast_fp16 = reshape(shape = var_3791, x = value_69_cast_fp16)[name = tensor("op_3792_cast_fp16")]; + tensor attn_69_transpose_x_0 = const()[name = tensor("attn_69_transpose_x_0"), val = tensor(false)]; + tensor attn_69_transpose_y_0 = const()[name = tensor("attn_69_transpose_y_0"), val = tensor(true)]; + tensor attn_69_cast_fp16 = matmul(transpose_x = attn_69_transpose_x_0, transpose_y = attn_69_transpose_y_0, x = var_3792_cast_fp16, y = var_3790_cast_fp16)[name = tensor("attn_69_cast_fp16")]; + tensor var_3795 = const()[name = tensor("op_3795"), val = tensor([1, 1024, 1, -1])]; + tensor input_171_cast_fp16 = reshape(shape = var_3795, x = attn_69_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor obj_245_pad_type_0 = const()[name = tensor("obj_245_pad_type_0"), val = tensor("valid")]; + tensor obj_245_strides_0 = const()[name = tensor("obj_245_strides_0"), val = tensor([1, 1])]; + tensor obj_245_pad_0 = const()[name = tensor("obj_245_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_245_dilations_0 = const()[name = tensor("obj_245_dilations_0"), val = tensor([1, 1])]; + tensor obj_245_groups_0 = const()[name = tensor("obj_245_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(684484864)))]; + tensor layers_17_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_17_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686582080)))]; + tensor obj_245_cast_fp16 = conv(bias = layers_17_self_attn_o_proj_bias_to_fp16, dilations = obj_245_dilations_0, groups = obj_245_groups_0, pad = obj_245_pad_0, pad_type = obj_245_pad_type_0, strides = obj_245_strides_0, weight = layers_17_self_attn_o_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("obj_245_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = obj_245_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; + tensor out_105_axes_0 = const()[name = tensor("out_105_axes_0"), val = tensor([1])]; + tensor var_3817_to_fp16 = const()[name = tensor("op_3817_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_3817_to_fp16, x = inputs_105_cast_fp16)[name = tensor("out_105_cast_fp16")]; + tensor obj_247_gamma_0_to_fp16 = const()[name = tensor("obj_247_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686584192)))]; + tensor obj_247_beta_0_to_fp16 = const()[name = tensor("obj_247_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686586304)))]; + tensor obj_247_epsilon_0_to_fp16 = const()[name = tensor("obj_247_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_247_cast_fp16 = batch_norm(beta = obj_247_beta_0_to_fp16, epsilon = obj_247_epsilon_0_to_fp16, gamma = obj_247_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor("obj_247_cast_fp16")]; + tensor query_71_pad_type_0 = const()[name = tensor("query_71_pad_type_0"), val = tensor("valid")]; + tensor query_71_strides_0 = const()[name = tensor("query_71_strides_0"), val = tensor([1, 1])]; + tensor query_71_pad_0 = const()[name = tensor("query_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_71_dilations_0 = const()[name = tensor("query_71_dilations_0"), val = tensor([1, 1])]; + tensor query_71_groups_0 = const()[name = tensor("query_71_groups_0"), val = tensor(1)]; + tensor layers_17_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(686588416)))]; + tensor layers_17_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(688685632)))]; + tensor query_71_cast_fp16 = conv(bias = layers_17_encoder_attn_q_proj_bias_to_fp16, dilations = query_71_dilations_0, groups = query_71_groups_0, pad = query_71_pad_0, pad_type = query_71_pad_type_0, strides = query_71_strides_0, weight = layers_17_encoder_attn_q_proj_weight_to_fp16, x = obj_247_cast_fp16)[name = tensor("query_71_cast_fp16")]; + tensor key_71_pad_type_0 = const()[name = tensor("key_71_pad_type_0"), val = tensor("valid")]; + tensor key_71_strides_0 = const()[name = tensor("key_71_strides_0"), val = tensor([1, 1])]; + tensor key_71_pad_0 = const()[name = tensor("key_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_71_dilations_0 = const()[name = tensor("key_71_dilations_0"), val = tensor([1, 1])]; + tensor key_71_groups_0 = const()[name = tensor("key_71_groups_0"), val = tensor(1)]; + tensor layers_17_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(688687744)))]; + tensor key_71_cast_fp16 = conv(dilations = key_71_dilations_0, groups = key_71_groups_0, pad = key_71_pad_0, pad_type = key_71_pad_type_0, strides = key_71_strides_0, weight = layers_17_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_71_cast_fp16")]; + tensor value_71_pad_type_0 = const()[name = tensor("value_71_pad_type_0"), val = tensor("valid")]; + tensor value_71_strides_0 = const()[name = tensor("value_71_strides_0"), val = tensor([1, 1])]; + tensor value_71_pad_0 = const()[name = tensor("value_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_71_dilations_0 = const()[name = tensor("value_71_dilations_0"), val = tensor([1, 1])]; + tensor value_71_groups_0 = const()[name = tensor("value_71_groups_0"), val = tensor(1)]; + tensor layers_17_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(690784960)))]; + tensor layers_17_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(692882176)))]; + tensor value_71_cast_fp16 = conv(bias = layers_17_encoder_attn_v_proj_bias_to_fp16, dilations = value_71_dilations_0, groups = value_71_groups_0, pad = value_71_pad_0, pad_type = value_71_pad_type_0, strides = value_71_strides_0, weight = layers_17_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_71_cast_fp16")]; + tensor var_3852 = const()[name = tensor("op_3852"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_71_cast_fp16 = reshape(shape = var_3852, x = query_71_cast_fp16)[name = tensor("mh_q_71_cast_fp16")]; + tensor var_3854_to_fp16 = const()[name = tensor("op_3854_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3855_cast_fp16 = mul(x = mh_q_71_cast_fp16, y = var_3854_to_fp16)[name = tensor("op_3855_cast_fp16")]; + tensor var_3856 = const()[name = tensor("op_3856"), val = tensor([1, 16, 64, -1])]; + tensor var_3857_cast_fp16 = reshape(shape = var_3856, x = key_71_cast_fp16)[name = tensor("op_3857_cast_fp16")]; + tensor mh_w_107_transpose_x_0 = const()[name = tensor("mh_w_107_transpose_x_0"), val = tensor(true)]; + tensor mh_w_107_transpose_y_0 = const()[name = tensor("mh_w_107_transpose_y_0"), val = tensor(false)]; + tensor mh_w_107_cast_fp16 = matmul(transpose_x = mh_w_107_transpose_x_0, transpose_y = mh_w_107_transpose_y_0, x = var_3855_cast_fp16, y = var_3857_cast_fp16)[name = tensor("mh_w_107_cast_fp16")]; + tensor obj_251_cast_fp16 = softmax(axis = var_3709, x = mh_w_107_cast_fp16)[name = tensor("obj_251_cast_fp16")]; + tensor var_3861 = const()[name = tensor("op_3861"), val = tensor([1, 16, 64, -1])]; + tensor var_3862_cast_fp16 = reshape(shape = var_3861, x = value_71_cast_fp16)[name = tensor("op_3862_cast_fp16")]; + tensor attn_71_transpose_x_0 = const()[name = tensor("attn_71_transpose_x_0"), val = tensor(false)]; + tensor attn_71_transpose_y_0 = const()[name = tensor("attn_71_transpose_y_0"), val = tensor(true)]; + tensor attn_71_cast_fp16 = matmul(transpose_x = attn_71_transpose_x_0, transpose_y = attn_71_transpose_y_0, x = var_3862_cast_fp16, y = obj_251_cast_fp16)[name = tensor("attn_71_cast_fp16")]; + tensor var_3865 = const()[name = tensor("op_3865"), val = tensor([1, 1024, 1, -1])]; + tensor input_173_cast_fp16 = reshape(shape = var_3865, x = attn_71_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor obj_249_pad_type_0 = const()[name = tensor("obj_249_pad_type_0"), val = tensor("valid")]; + tensor obj_249_strides_0 = const()[name = tensor("obj_249_strides_0"), val = tensor([1, 1])]; + tensor obj_249_pad_0 = const()[name = tensor("obj_249_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_249_dilations_0 = const()[name = tensor("obj_249_dilations_0"), val = tensor([1, 1])]; + tensor obj_249_groups_0 = const()[name = tensor("obj_249_groups_0"), val = tensor(1)]; + tensor layers_17_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(692884288)))]; + tensor layers_17_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_17_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(694981504)))]; + tensor obj_249_cast_fp16 = conv(bias = layers_17_encoder_attn_o_proj_bias_to_fp16, dilations = obj_249_dilations_0, groups = obj_249_groups_0, pad = obj_249_pad_0, pad_type = obj_249_pad_type_0, strides = obj_249_strides_0, weight = layers_17_encoder_attn_o_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("obj_249_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = obj_249_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; + tensor out_107_axes_0 = const()[name = tensor("out_107_axes_0"), val = tensor([1])]; + tensor var_3883_to_fp16 = const()[name = tensor("op_3883_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_3883_to_fp16, x = inputs_107_cast_fp16)[name = tensor("out_107_cast_fp16")]; + tensor input_175_gamma_0_to_fp16 = const()[name = tensor("input_175_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(694983616)))]; + tensor input_175_beta_0_to_fp16 = const()[name = tensor("input_175_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(694985728)))]; + tensor input_175_epsilon_0_to_fp16 = const()[name = tensor("input_175_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_175_cast_fp16 = batch_norm(beta = input_175_beta_0_to_fp16, epsilon = input_175_epsilon_0_to_fp16, gamma = input_175_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_pad_type_0 = const()[name = tensor("input_177_pad_type_0"), val = tensor("valid")]; + tensor input_177_strides_0 = const()[name = tensor("input_177_strides_0"), val = tensor([1, 1])]; + tensor input_177_pad_0 = const()[name = tensor("input_177_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_177_dilations_0 = const()[name = tensor("input_177_dilations_0"), val = tensor([1, 1])]; + tensor input_177_groups_0 = const()[name = tensor("input_177_groups_0"), val = tensor(1)]; + tensor layers_17_fc1_weight_to_fp16 = const()[name = tensor("layers_17_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(694987840)))]; + tensor layers_17_fc1_bias_to_fp16 = const()[name = tensor("layers_17_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(703376512)))]; + tensor input_177_cast_fp16 = conv(bias = layers_17_fc1_bias_to_fp16, dilations = input_177_dilations_0, groups = input_177_groups_0, pad = input_177_pad_0, pad_type = input_177_pad_type_0, strides = input_177_strides_0, weight = layers_17_fc1_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor hidden_states_37_pad_type_0 = const()[name = tensor("hidden_states_37_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_37_strides_0 = const()[name = tensor("hidden_states_37_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_37_pad_0 = const()[name = tensor("hidden_states_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_37_dilations_0 = const()[name = tensor("hidden_states_37_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_37_groups_0 = const()[name = tensor("hidden_states_37_groups_0"), val = tensor(1)]; + tensor layers_17_fc2_weight_to_fp16 = const()[name = tensor("layers_17_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(703384768)))]; + tensor layers_17_fc2_bias_to_fp16 = const()[name = tensor("layers_17_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711773440)))]; + tensor hidden_states_37_cast_fp16 = conv(bias = layers_17_fc2_bias_to_fp16, dilations = hidden_states_37_dilations_0, groups = hidden_states_37_groups_0, pad = hidden_states_37_pad_0, pad_type = hidden_states_37_pad_type_0, strides = hidden_states_37_strides_0, weight = layers_17_fc2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("hidden_states_37_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = hidden_states_37_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; + tensor var_3918 = const()[name = tensor("op_3918"), val = tensor(3)]; + tensor out_109_axes_0 = const()[name = tensor("out_109_axes_0"), val = tensor([1])]; + tensor var_3943_to_fp16 = const()[name = tensor("op_3943_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_3943_to_fp16, x = inputs_109_cast_fp16)[name = tensor("out_109_cast_fp16")]; + tensor obj_253_gamma_0_to_fp16 = const()[name = tensor("obj_253_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711775552)))]; + tensor obj_253_beta_0_to_fp16 = const()[name = tensor("obj_253_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711777664)))]; + tensor obj_253_epsilon_0_to_fp16 = const()[name = tensor("obj_253_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_253_cast_fp16 = batch_norm(beta = obj_253_beta_0_to_fp16, epsilon = obj_253_epsilon_0_to_fp16, gamma = obj_253_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor("obj_253_cast_fp16")]; + tensor query_73_pad_type_0 = const()[name = tensor("query_73_pad_type_0"), val = tensor("valid")]; + tensor query_73_strides_0 = const()[name = tensor("query_73_strides_0"), val = tensor([1, 1])]; + tensor query_73_pad_0 = const()[name = tensor("query_73_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_73_dilations_0 = const()[name = tensor("query_73_dilations_0"), val = tensor([1, 1])]; + tensor query_73_groups_0 = const()[name = tensor("query_73_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(711779776)))]; + tensor layers_18_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(713876992)))]; + tensor query_73_cast_fp16 = conv(bias = layers_18_self_attn_q_proj_bias_to_fp16, dilations = query_73_dilations_0, groups = query_73_groups_0, pad = query_73_pad_0, pad_type = query_73_pad_type_0, strides = query_73_strides_0, weight = layers_18_self_attn_q_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("query_73_cast_fp16")]; + tensor current_key_37_pad_type_0 = const()[name = tensor("current_key_37_pad_type_0"), val = tensor("valid")]; + tensor current_key_37_strides_0 = const()[name = tensor("current_key_37_strides_0"), val = tensor([1, 1])]; + tensor current_key_37_pad_0 = const()[name = tensor("current_key_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_37_dilations_0 = const()[name = tensor("current_key_37_dilations_0"), val = tensor([1, 1])]; + tensor current_key_37_groups_0 = const()[name = tensor("current_key_37_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(713879104)))]; + tensor current_key_37_cast_fp16 = conv(dilations = current_key_37_dilations_0, groups = current_key_37_groups_0, pad = current_key_37_pad_0, pad_type = current_key_37_pad_type_0, strides = current_key_37_strides_0, weight = layers_18_self_attn_k_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("current_key_37_cast_fp16")]; + tensor current_value_37_pad_type_0 = const()[name = tensor("current_value_37_pad_type_0"), val = tensor("valid")]; + tensor current_value_37_strides_0 = const()[name = tensor("current_value_37_strides_0"), val = tensor([1, 1])]; + tensor current_value_37_pad_0 = const()[name = tensor("current_value_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_37_dilations_0 = const()[name = tensor("current_value_37_dilations_0"), val = tensor([1, 1])]; + tensor current_value_37_groups_0 = const()[name = tensor("current_value_37_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(715976320)))]; + tensor layers_18_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(718073536)))]; + tensor current_value_37_cast_fp16 = conv(bias = layers_18_self_attn_v_proj_bias_to_fp16, dilations = current_value_37_dilations_0, groups = current_value_37_groups_0, pad = current_value_37_pad_0, pad_type = current_value_37_pad_type_0, strides = current_value_37_strides_0, weight = layers_18_self_attn_v_proj_weight_to_fp16, x = obj_253_cast_fp16)[name = tensor("current_value_37_cast_fp16")]; + tensor var_3981_cast_fp16 = mul(x = current_key_37_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3981_cast_fp16")]; + tensor key_73_cast_fp16 = add(x = var_87_cast_fp16_18, y = var_3981_cast_fp16)[name = tensor("key_73_cast_fp16")]; + tensor var_3983_cast_fp16 = mul(x = current_value_37_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_3983_cast_fp16")]; + tensor value_73_cast_fp16 = add(x = var_114_cast_fp16_18, y = var_3983_cast_fp16)[name = tensor("value_73_cast_fp16")]; + tensor var_3986 = const()[name = tensor("op_3986"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_73_cast_fp16 = reshape(shape = var_3986, x = query_73_cast_fp16)[name = tensor("mh_q_73_cast_fp16")]; + tensor var_3988_to_fp16 = const()[name = tensor("op_3988_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3989_cast_fp16 = mul(x = mh_q_73_cast_fp16, y = var_3988_to_fp16)[name = tensor("op_3989_cast_fp16")]; + tensor var_3990 = const()[name = tensor("op_3990"), val = tensor([1, 16, 64, -1])]; + tensor var_3991_cast_fp16 = reshape(shape = var_3990, x = key_73_cast_fp16)[name = tensor("op_3991_cast_fp16")]; + tensor mh_w_109_transpose_x_0 = const()[name = tensor("mh_w_109_transpose_x_0"), val = tensor(true)]; + tensor mh_w_109_transpose_y_0 = const()[name = tensor("mh_w_109_transpose_y_0"), val = tensor(false)]; + tensor mh_w_109_cast_fp16 = matmul(transpose_x = mh_w_109_transpose_x_0, transpose_y = mh_w_109_transpose_y_0, x = var_3989_cast_fp16, y = var_3991_cast_fp16)[name = tensor("mh_w_109_cast_fp16")]; + tensor mh_w_111_cast_fp16 = add(x = mh_w_109_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_111_cast_fp16")]; + tensor var_3999_cast_fp16 = softmax(axis = var_3918, x = mh_w_111_cast_fp16)[name = tensor("op_3999_cast_fp16")]; + tensor var_4000 = const()[name = tensor("op_4000"), val = tensor([1, 16, 64, -1])]; + tensor var_4001_cast_fp16 = reshape(shape = var_4000, x = value_73_cast_fp16)[name = tensor("op_4001_cast_fp16")]; + tensor attn_73_transpose_x_0 = const()[name = tensor("attn_73_transpose_x_0"), val = tensor(false)]; + tensor attn_73_transpose_y_0 = const()[name = tensor("attn_73_transpose_y_0"), val = tensor(true)]; + tensor attn_73_cast_fp16 = matmul(transpose_x = attn_73_transpose_x_0, transpose_y = attn_73_transpose_y_0, x = var_4001_cast_fp16, y = var_3999_cast_fp16)[name = tensor("attn_73_cast_fp16")]; + tensor var_4004 = const()[name = tensor("op_4004"), val = tensor([1, 1024, 1, -1])]; + tensor input_181_cast_fp16 = reshape(shape = var_4004, x = attn_73_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor obj_259_pad_type_0 = const()[name = tensor("obj_259_pad_type_0"), val = tensor("valid")]; + tensor obj_259_strides_0 = const()[name = tensor("obj_259_strides_0"), val = tensor([1, 1])]; + tensor obj_259_pad_0 = const()[name = tensor("obj_259_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_259_dilations_0 = const()[name = tensor("obj_259_dilations_0"), val = tensor([1, 1])]; + tensor obj_259_groups_0 = const()[name = tensor("obj_259_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(718075648)))]; + tensor layers_18_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_18_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720172864)))]; + tensor obj_259_cast_fp16 = conv(bias = layers_18_self_attn_o_proj_bias_to_fp16, dilations = obj_259_dilations_0, groups = obj_259_groups_0, pad = obj_259_pad_0, pad_type = obj_259_pad_type_0, strides = obj_259_strides_0, weight = layers_18_self_attn_o_proj_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("obj_259_cast_fp16")]; + tensor inputs_111_cast_fp16 = add(x = inputs_109_cast_fp16, y = obj_259_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor out_111_axes_0 = const()[name = tensor("out_111_axes_0"), val = tensor([1])]; + tensor var_4026_to_fp16 = const()[name = tensor("op_4026_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_4026_to_fp16, x = inputs_111_cast_fp16)[name = tensor("out_111_cast_fp16")]; + tensor obj_261_gamma_0_to_fp16 = const()[name = tensor("obj_261_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720174976)))]; + tensor obj_261_beta_0_to_fp16 = const()[name = tensor("obj_261_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720177088)))]; + tensor obj_261_epsilon_0_to_fp16 = const()[name = tensor("obj_261_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_261_cast_fp16 = batch_norm(beta = obj_261_beta_0_to_fp16, epsilon = obj_261_epsilon_0_to_fp16, gamma = obj_261_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor("obj_261_cast_fp16")]; + tensor query_75_pad_type_0 = const()[name = tensor("query_75_pad_type_0"), val = tensor("valid")]; + tensor query_75_strides_0 = const()[name = tensor("query_75_strides_0"), val = tensor([1, 1])]; + tensor query_75_pad_0 = const()[name = tensor("query_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_75_dilations_0 = const()[name = tensor("query_75_dilations_0"), val = tensor([1, 1])]; + tensor query_75_groups_0 = const()[name = tensor("query_75_groups_0"), val = tensor(1)]; + tensor layers_18_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(720179200)))]; + tensor layers_18_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722276416)))]; + tensor query_75_cast_fp16 = conv(bias = layers_18_encoder_attn_q_proj_bias_to_fp16, dilations = query_75_dilations_0, groups = query_75_groups_0, pad = query_75_pad_0, pad_type = query_75_pad_type_0, strides = query_75_strides_0, weight = layers_18_encoder_attn_q_proj_weight_to_fp16, x = obj_261_cast_fp16)[name = tensor("query_75_cast_fp16")]; + tensor key_75_pad_type_0 = const()[name = tensor("key_75_pad_type_0"), val = tensor("valid")]; + tensor key_75_strides_0 = const()[name = tensor("key_75_strides_0"), val = tensor([1, 1])]; + tensor key_75_pad_0 = const()[name = tensor("key_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_75_dilations_0 = const()[name = tensor("key_75_dilations_0"), val = tensor([1, 1])]; + tensor key_75_groups_0 = const()[name = tensor("key_75_groups_0"), val = tensor(1)]; + tensor layers_18_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(722278528)))]; + tensor key_75_cast_fp16 = conv(dilations = key_75_dilations_0, groups = key_75_groups_0, pad = key_75_pad_0, pad_type = key_75_pad_type_0, strides = key_75_strides_0, weight = layers_18_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_75_cast_fp16")]; + tensor value_75_pad_type_0 = const()[name = tensor("value_75_pad_type_0"), val = tensor("valid")]; + tensor value_75_strides_0 = const()[name = tensor("value_75_strides_0"), val = tensor([1, 1])]; + tensor value_75_pad_0 = const()[name = tensor("value_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_75_dilations_0 = const()[name = tensor("value_75_dilations_0"), val = tensor([1, 1])]; + tensor value_75_groups_0 = const()[name = tensor("value_75_groups_0"), val = tensor(1)]; + tensor layers_18_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(724375744)))]; + tensor layers_18_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(726472960)))]; + tensor value_75_cast_fp16 = conv(bias = layers_18_encoder_attn_v_proj_bias_to_fp16, dilations = value_75_dilations_0, groups = value_75_groups_0, pad = value_75_pad_0, pad_type = value_75_pad_type_0, strides = value_75_strides_0, weight = layers_18_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_75_cast_fp16")]; + tensor var_4061 = const()[name = tensor("op_4061"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_75_cast_fp16 = reshape(shape = var_4061, x = query_75_cast_fp16)[name = tensor("mh_q_75_cast_fp16")]; + tensor var_4063_to_fp16 = const()[name = tensor("op_4063_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4064_cast_fp16 = mul(x = mh_q_75_cast_fp16, y = var_4063_to_fp16)[name = tensor("op_4064_cast_fp16")]; + tensor var_4065 = const()[name = tensor("op_4065"), val = tensor([1, 16, 64, -1])]; + tensor var_4066_cast_fp16 = reshape(shape = var_4065, x = key_75_cast_fp16)[name = tensor("op_4066_cast_fp16")]; + tensor mh_w_113_transpose_x_0 = const()[name = tensor("mh_w_113_transpose_x_0"), val = tensor(true)]; + tensor mh_w_113_transpose_y_0 = const()[name = tensor("mh_w_113_transpose_y_0"), val = tensor(false)]; + tensor mh_w_113_cast_fp16 = matmul(transpose_x = mh_w_113_transpose_x_0, transpose_y = mh_w_113_transpose_y_0, x = var_4064_cast_fp16, y = var_4066_cast_fp16)[name = tensor("mh_w_113_cast_fp16")]; + tensor obj_265_cast_fp16 = softmax(axis = var_3918, x = mh_w_113_cast_fp16)[name = tensor("obj_265_cast_fp16")]; + tensor var_4070 = const()[name = tensor("op_4070"), val = tensor([1, 16, 64, -1])]; + tensor var_4071_cast_fp16 = reshape(shape = var_4070, x = value_75_cast_fp16)[name = tensor("op_4071_cast_fp16")]; + tensor attn_75_transpose_x_0 = const()[name = tensor("attn_75_transpose_x_0"), val = tensor(false)]; + tensor attn_75_transpose_y_0 = const()[name = tensor("attn_75_transpose_y_0"), val = tensor(true)]; + tensor attn_75_cast_fp16 = matmul(transpose_x = attn_75_transpose_x_0, transpose_y = attn_75_transpose_y_0, x = var_4071_cast_fp16, y = obj_265_cast_fp16)[name = tensor("attn_75_cast_fp16")]; + tensor var_4074 = const()[name = tensor("op_4074"), val = tensor([1, 1024, 1, -1])]; + tensor input_183_cast_fp16 = reshape(shape = var_4074, x = attn_75_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor obj_263_pad_type_0 = const()[name = tensor("obj_263_pad_type_0"), val = tensor("valid")]; + tensor obj_263_strides_0 = const()[name = tensor("obj_263_strides_0"), val = tensor([1, 1])]; + tensor obj_263_pad_0 = const()[name = tensor("obj_263_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_263_dilations_0 = const()[name = tensor("obj_263_dilations_0"), val = tensor([1, 1])]; + tensor obj_263_groups_0 = const()[name = tensor("obj_263_groups_0"), val = tensor(1)]; + tensor layers_18_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(726475072)))]; + tensor layers_18_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_18_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(728572288)))]; + tensor obj_263_cast_fp16 = conv(bias = layers_18_encoder_attn_o_proj_bias_to_fp16, dilations = obj_263_dilations_0, groups = obj_263_groups_0, pad = obj_263_pad_0, pad_type = obj_263_pad_type_0, strides = obj_263_strides_0, weight = layers_18_encoder_attn_o_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("obj_263_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = obj_263_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; + tensor out_113_axes_0 = const()[name = tensor("out_113_axes_0"), val = tensor([1])]; + tensor var_4092_to_fp16 = const()[name = tensor("op_4092_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_113_cast_fp16 = layer_norm(axes = out_113_axes_0, epsilon = var_4092_to_fp16, x = inputs_113_cast_fp16)[name = tensor("out_113_cast_fp16")]; + tensor input_185_gamma_0_to_fp16 = const()[name = tensor("input_185_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(728574400)))]; + tensor input_185_beta_0_to_fp16 = const()[name = tensor("input_185_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(728576512)))]; + tensor input_185_epsilon_0_to_fp16 = const()[name = tensor("input_185_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = batch_norm(beta = input_185_beta_0_to_fp16, epsilon = input_185_epsilon_0_to_fp16, gamma = input_185_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor input_187_pad_type_0 = const()[name = tensor("input_187_pad_type_0"), val = tensor("valid")]; + tensor input_187_strides_0 = const()[name = tensor("input_187_strides_0"), val = tensor([1, 1])]; + tensor input_187_pad_0 = const()[name = tensor("input_187_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_187_dilations_0 = const()[name = tensor("input_187_dilations_0"), val = tensor([1, 1])]; + tensor input_187_groups_0 = const()[name = tensor("input_187_groups_0"), val = tensor(1)]; + tensor layers_18_fc1_weight_to_fp16 = const()[name = tensor("layers_18_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(728578624)))]; + tensor layers_18_fc1_bias_to_fp16 = const()[name = tensor("layers_18_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736967296)))]; + tensor input_187_cast_fp16 = conv(bias = layers_18_fc1_bias_to_fp16, dilations = input_187_dilations_0, groups = input_187_groups_0, pad = input_187_pad_0, pad_type = input_187_pad_type_0, strides = input_187_strides_0, weight = layers_18_fc1_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_mode_0 = const()[name = tensor("input_189_mode_0"), val = tensor("EXACT")]; + tensor input_189_cast_fp16 = gelu(mode = input_189_mode_0, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor hidden_states_39_pad_type_0 = const()[name = tensor("hidden_states_39_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_39_strides_0 = const()[name = tensor("hidden_states_39_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_39_pad_0 = const()[name = tensor("hidden_states_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_39_dilations_0 = const()[name = tensor("hidden_states_39_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_39_groups_0 = const()[name = tensor("hidden_states_39_groups_0"), val = tensor(1)]; + tensor layers_18_fc2_weight_to_fp16 = const()[name = tensor("layers_18_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(736975552)))]; + tensor layers_18_fc2_bias_to_fp16 = const()[name = tensor("layers_18_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(745364224)))]; + tensor hidden_states_39_cast_fp16 = conv(bias = layers_18_fc2_bias_to_fp16, dilations = hidden_states_39_dilations_0, groups = hidden_states_39_groups_0, pad = hidden_states_39_pad_0, pad_type = hidden_states_39_pad_type_0, strides = hidden_states_39_strides_0, weight = layers_18_fc2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("hidden_states_39_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = hidden_states_39_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; + tensor var_4127 = const()[name = tensor("op_4127"), val = tensor(3)]; + tensor out_115_axes_0 = const()[name = tensor("out_115_axes_0"), val = tensor([1])]; + tensor var_4152_to_fp16 = const()[name = tensor("op_4152_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_115_cast_fp16 = layer_norm(axes = out_115_axes_0, epsilon = var_4152_to_fp16, x = inputs_115_cast_fp16)[name = tensor("out_115_cast_fp16")]; + tensor obj_267_gamma_0_to_fp16 = const()[name = tensor("obj_267_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(745366336)))]; + tensor obj_267_beta_0_to_fp16 = const()[name = tensor("obj_267_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(745368448)))]; + tensor obj_267_epsilon_0_to_fp16 = const()[name = tensor("obj_267_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_267_cast_fp16 = batch_norm(beta = obj_267_beta_0_to_fp16, epsilon = obj_267_epsilon_0_to_fp16, gamma = obj_267_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor("obj_267_cast_fp16")]; + tensor query_77_pad_type_0 = const()[name = tensor("query_77_pad_type_0"), val = tensor("valid")]; + tensor query_77_strides_0 = const()[name = tensor("query_77_strides_0"), val = tensor([1, 1])]; + tensor query_77_pad_0 = const()[name = tensor("query_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_77_dilations_0 = const()[name = tensor("query_77_dilations_0"), val = tensor([1, 1])]; + tensor query_77_groups_0 = const()[name = tensor("query_77_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(745370560)))]; + tensor layers_19_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(747467776)))]; + tensor query_77_cast_fp16 = conv(bias = layers_19_self_attn_q_proj_bias_to_fp16, dilations = query_77_dilations_0, groups = query_77_groups_0, pad = query_77_pad_0, pad_type = query_77_pad_type_0, strides = query_77_strides_0, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("query_77_cast_fp16")]; + tensor current_key_39_pad_type_0 = const()[name = tensor("current_key_39_pad_type_0"), val = tensor("valid")]; + tensor current_key_39_strides_0 = const()[name = tensor("current_key_39_strides_0"), val = tensor([1, 1])]; + tensor current_key_39_pad_0 = const()[name = tensor("current_key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_39_dilations_0 = const()[name = tensor("current_key_39_dilations_0"), val = tensor([1, 1])]; + tensor current_key_39_groups_0 = const()[name = tensor("current_key_39_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(747469888)))]; + tensor current_key_39_cast_fp16 = conv(dilations = current_key_39_dilations_0, groups = current_key_39_groups_0, pad = current_key_39_pad_0, pad_type = current_key_39_pad_type_0, strides = current_key_39_strides_0, weight = layers_19_self_attn_k_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("current_key_39_cast_fp16")]; + tensor current_value_39_pad_type_0 = const()[name = tensor("current_value_39_pad_type_0"), val = tensor("valid")]; + tensor current_value_39_strides_0 = const()[name = tensor("current_value_39_strides_0"), val = tensor([1, 1])]; + tensor current_value_39_pad_0 = const()[name = tensor("current_value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_39_dilations_0 = const()[name = tensor("current_value_39_dilations_0"), val = tensor([1, 1])]; + tensor current_value_39_groups_0 = const()[name = tensor("current_value_39_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(749567104)))]; + tensor layers_19_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(751664320)))]; + tensor current_value_39_cast_fp16 = conv(bias = layers_19_self_attn_v_proj_bias_to_fp16, dilations = current_value_39_dilations_0, groups = current_value_39_groups_0, pad = current_value_39_pad_0, pad_type = current_value_39_pad_type_0, strides = current_value_39_strides_0, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = obj_267_cast_fp16)[name = tensor("current_value_39_cast_fp16")]; + tensor var_4190_cast_fp16 = mul(x = current_key_39_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4190_cast_fp16")]; + tensor key_77_cast_fp16 = add(x = var_87_cast_fp16_19, y = var_4190_cast_fp16)[name = tensor("key_77_cast_fp16")]; + tensor var_4192_cast_fp16 = mul(x = current_value_39_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4192_cast_fp16")]; + tensor value_77_cast_fp16 = add(x = var_114_cast_fp16_19, y = var_4192_cast_fp16)[name = tensor("value_77_cast_fp16")]; + tensor var_4195 = const()[name = tensor("op_4195"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_77_cast_fp16 = reshape(shape = var_4195, x = query_77_cast_fp16)[name = tensor("mh_q_77_cast_fp16")]; + tensor var_4197_to_fp16 = const()[name = tensor("op_4197_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4198_cast_fp16 = mul(x = mh_q_77_cast_fp16, y = var_4197_to_fp16)[name = tensor("op_4198_cast_fp16")]; + tensor var_4199 = const()[name = tensor("op_4199"), val = tensor([1, 16, 64, -1])]; + tensor var_4200_cast_fp16 = reshape(shape = var_4199, x = key_77_cast_fp16)[name = tensor("op_4200_cast_fp16")]; + tensor mh_w_115_transpose_x_0 = const()[name = tensor("mh_w_115_transpose_x_0"), val = tensor(true)]; + tensor mh_w_115_transpose_y_0 = const()[name = tensor("mh_w_115_transpose_y_0"), val = tensor(false)]; + tensor mh_w_115_cast_fp16 = matmul(transpose_x = mh_w_115_transpose_x_0, transpose_y = mh_w_115_transpose_y_0, x = var_4198_cast_fp16, y = var_4200_cast_fp16)[name = tensor("mh_w_115_cast_fp16")]; + tensor mh_w_117_cast_fp16 = add(x = mh_w_115_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_117_cast_fp16")]; + tensor var_4208_cast_fp16 = softmax(axis = var_4127, x = mh_w_117_cast_fp16)[name = tensor("op_4208_cast_fp16")]; + tensor var_4209 = const()[name = tensor("op_4209"), val = tensor([1, 16, 64, -1])]; + tensor var_4210_cast_fp16 = reshape(shape = var_4209, x = value_77_cast_fp16)[name = tensor("op_4210_cast_fp16")]; + tensor attn_77_transpose_x_0 = const()[name = tensor("attn_77_transpose_x_0"), val = tensor(false)]; + tensor attn_77_transpose_y_0 = const()[name = tensor("attn_77_transpose_y_0"), val = tensor(true)]; + tensor attn_77_cast_fp16 = matmul(transpose_x = attn_77_transpose_x_0, transpose_y = attn_77_transpose_y_0, x = var_4210_cast_fp16, y = var_4208_cast_fp16)[name = tensor("attn_77_cast_fp16")]; + tensor var_4213 = const()[name = tensor("op_4213"), val = tensor([1, 1024, 1, -1])]; + tensor input_191_cast_fp16 = reshape(shape = var_4213, x = attn_77_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor obj_273_pad_type_0 = const()[name = tensor("obj_273_pad_type_0"), val = tensor("valid")]; + tensor obj_273_strides_0 = const()[name = tensor("obj_273_strides_0"), val = tensor([1, 1])]; + tensor obj_273_pad_0 = const()[name = tensor("obj_273_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_273_dilations_0 = const()[name = tensor("obj_273_dilations_0"), val = tensor([1, 1])]; + tensor obj_273_groups_0 = const()[name = tensor("obj_273_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(751666432)))]; + tensor layers_19_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_19_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(753763648)))]; + tensor obj_273_cast_fp16 = conv(bias = layers_19_self_attn_o_proj_bias_to_fp16, dilations = obj_273_dilations_0, groups = obj_273_groups_0, pad = obj_273_pad_0, pad_type = obj_273_pad_type_0, strides = obj_273_strides_0, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("obj_273_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = obj_273_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; + tensor out_117_axes_0 = const()[name = tensor("out_117_axes_0"), val = tensor([1])]; + tensor var_4235_to_fp16 = const()[name = tensor("op_4235_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_117_cast_fp16 = layer_norm(axes = out_117_axes_0, epsilon = var_4235_to_fp16, x = inputs_117_cast_fp16)[name = tensor("out_117_cast_fp16")]; + tensor obj_275_gamma_0_to_fp16 = const()[name = tensor("obj_275_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(753765760)))]; + tensor obj_275_beta_0_to_fp16 = const()[name = tensor("obj_275_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(753767872)))]; + tensor obj_275_epsilon_0_to_fp16 = const()[name = tensor("obj_275_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_275_cast_fp16 = batch_norm(beta = obj_275_beta_0_to_fp16, epsilon = obj_275_epsilon_0_to_fp16, gamma = obj_275_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor("obj_275_cast_fp16")]; + tensor query_79_pad_type_0 = const()[name = tensor("query_79_pad_type_0"), val = tensor("valid")]; + tensor query_79_strides_0 = const()[name = tensor("query_79_strides_0"), val = tensor([1, 1])]; + tensor query_79_pad_0 = const()[name = tensor("query_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_79_dilations_0 = const()[name = tensor("query_79_dilations_0"), val = tensor([1, 1])]; + tensor query_79_groups_0 = const()[name = tensor("query_79_groups_0"), val = tensor(1)]; + tensor layers_19_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(753769984)))]; + tensor layers_19_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(755867200)))]; + tensor query_79_cast_fp16 = conv(bias = layers_19_encoder_attn_q_proj_bias_to_fp16, dilations = query_79_dilations_0, groups = query_79_groups_0, pad = query_79_pad_0, pad_type = query_79_pad_type_0, strides = query_79_strides_0, weight = layers_19_encoder_attn_q_proj_weight_to_fp16, x = obj_275_cast_fp16)[name = tensor("query_79_cast_fp16")]; + tensor key_79_pad_type_0 = const()[name = tensor("key_79_pad_type_0"), val = tensor("valid")]; + tensor key_79_strides_0 = const()[name = tensor("key_79_strides_0"), val = tensor([1, 1])]; + tensor key_79_pad_0 = const()[name = tensor("key_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_79_dilations_0 = const()[name = tensor("key_79_dilations_0"), val = tensor([1, 1])]; + tensor key_79_groups_0 = const()[name = tensor("key_79_groups_0"), val = tensor(1)]; + tensor layers_19_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(755869312)))]; + tensor key_79_cast_fp16 = conv(dilations = key_79_dilations_0, groups = key_79_groups_0, pad = key_79_pad_0, pad_type = key_79_pad_type_0, strides = key_79_strides_0, weight = layers_19_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_79_cast_fp16")]; + tensor value_79_pad_type_0 = const()[name = tensor("value_79_pad_type_0"), val = tensor("valid")]; + tensor value_79_strides_0 = const()[name = tensor("value_79_strides_0"), val = tensor([1, 1])]; + tensor value_79_pad_0 = const()[name = tensor("value_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_79_dilations_0 = const()[name = tensor("value_79_dilations_0"), val = tensor([1, 1])]; + tensor value_79_groups_0 = const()[name = tensor("value_79_groups_0"), val = tensor(1)]; + tensor layers_19_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(757966528)))]; + tensor layers_19_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(760063744)))]; + tensor value_79_cast_fp16 = conv(bias = layers_19_encoder_attn_v_proj_bias_to_fp16, dilations = value_79_dilations_0, groups = value_79_groups_0, pad = value_79_pad_0, pad_type = value_79_pad_type_0, strides = value_79_strides_0, weight = layers_19_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_79_cast_fp16")]; + tensor var_4270 = const()[name = tensor("op_4270"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_79_cast_fp16 = reshape(shape = var_4270, x = query_79_cast_fp16)[name = tensor("mh_q_79_cast_fp16")]; + tensor var_4272_to_fp16 = const()[name = tensor("op_4272_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4273_cast_fp16 = mul(x = mh_q_79_cast_fp16, y = var_4272_to_fp16)[name = tensor("op_4273_cast_fp16")]; + tensor var_4274 = const()[name = tensor("op_4274"), val = tensor([1, 16, 64, -1])]; + tensor var_4275_cast_fp16 = reshape(shape = var_4274, x = key_79_cast_fp16)[name = tensor("op_4275_cast_fp16")]; + tensor mh_w_119_transpose_x_0 = const()[name = tensor("mh_w_119_transpose_x_0"), val = tensor(true)]; + tensor mh_w_119_transpose_y_0 = const()[name = tensor("mh_w_119_transpose_y_0"), val = tensor(false)]; + tensor mh_w_119_cast_fp16 = matmul(transpose_x = mh_w_119_transpose_x_0, transpose_y = mh_w_119_transpose_y_0, x = var_4273_cast_fp16, y = var_4275_cast_fp16)[name = tensor("mh_w_119_cast_fp16")]; + tensor obj_279_cast_fp16 = softmax(axis = var_4127, x = mh_w_119_cast_fp16)[name = tensor("obj_279_cast_fp16")]; + tensor var_4279 = const()[name = tensor("op_4279"), val = tensor([1, 16, 64, -1])]; + tensor var_4280_cast_fp16 = reshape(shape = var_4279, x = value_79_cast_fp16)[name = tensor("op_4280_cast_fp16")]; + tensor attn_79_transpose_x_0 = const()[name = tensor("attn_79_transpose_x_0"), val = tensor(false)]; + tensor attn_79_transpose_y_0 = const()[name = tensor("attn_79_transpose_y_0"), val = tensor(true)]; + tensor attn_79_cast_fp16 = matmul(transpose_x = attn_79_transpose_x_0, transpose_y = attn_79_transpose_y_0, x = var_4280_cast_fp16, y = obj_279_cast_fp16)[name = tensor("attn_79_cast_fp16")]; + tensor var_4283 = const()[name = tensor("op_4283"), val = tensor([1, 1024, 1, -1])]; + tensor input_193_cast_fp16 = reshape(shape = var_4283, x = attn_79_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor obj_277_pad_type_0 = const()[name = tensor("obj_277_pad_type_0"), val = tensor("valid")]; + tensor obj_277_strides_0 = const()[name = tensor("obj_277_strides_0"), val = tensor([1, 1])]; + tensor obj_277_pad_0 = const()[name = tensor("obj_277_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_277_dilations_0 = const()[name = tensor("obj_277_dilations_0"), val = tensor([1, 1])]; + tensor obj_277_groups_0 = const()[name = tensor("obj_277_groups_0"), val = tensor(1)]; + tensor layers_19_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(760065856)))]; + tensor layers_19_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_19_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762163072)))]; + tensor obj_277_cast_fp16 = conv(bias = layers_19_encoder_attn_o_proj_bias_to_fp16, dilations = obj_277_dilations_0, groups = obj_277_groups_0, pad = obj_277_pad_0, pad_type = obj_277_pad_type_0, strides = obj_277_strides_0, weight = layers_19_encoder_attn_o_proj_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("obj_277_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = obj_277_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; + tensor out_119_axes_0 = const()[name = tensor("out_119_axes_0"), val = tensor([1])]; + tensor var_4301_to_fp16 = const()[name = tensor("op_4301_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_119_cast_fp16 = layer_norm(axes = out_119_axes_0, epsilon = var_4301_to_fp16, x = inputs_119_cast_fp16)[name = tensor("out_119_cast_fp16")]; + tensor input_195_gamma_0_to_fp16 = const()[name = tensor("input_195_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762165184)))]; + tensor input_195_beta_0_to_fp16 = const()[name = tensor("input_195_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762167296)))]; + tensor input_195_epsilon_0_to_fp16 = const()[name = tensor("input_195_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = batch_norm(beta = input_195_beta_0_to_fp16, epsilon = input_195_epsilon_0_to_fp16, gamma = input_195_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_pad_type_0 = const()[name = tensor("input_197_pad_type_0"), val = tensor("valid")]; + tensor input_197_strides_0 = const()[name = tensor("input_197_strides_0"), val = tensor([1, 1])]; + tensor input_197_pad_0 = const()[name = tensor("input_197_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_197_dilations_0 = const()[name = tensor("input_197_dilations_0"), val = tensor([1, 1])]; + tensor input_197_groups_0 = const()[name = tensor("input_197_groups_0"), val = tensor(1)]; + tensor layers_19_fc1_weight_to_fp16 = const()[name = tensor("layers_19_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(762169408)))]; + tensor layers_19_fc1_bias_to_fp16 = const()[name = tensor("layers_19_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(770558080)))]; + tensor input_197_cast_fp16 = conv(bias = layers_19_fc1_bias_to_fp16, dilations = input_197_dilations_0, groups = input_197_groups_0, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = input_197_strides_0, weight = layers_19_fc1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_mode_0 = const()[name = tensor("input_199_mode_0"), val = tensor("EXACT")]; + tensor input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor hidden_states_41_pad_type_0 = const()[name = tensor("hidden_states_41_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_41_strides_0 = const()[name = tensor("hidden_states_41_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_41_pad_0 = const()[name = tensor("hidden_states_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_41_dilations_0 = const()[name = tensor("hidden_states_41_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_41_groups_0 = const()[name = tensor("hidden_states_41_groups_0"), val = tensor(1)]; + tensor layers_19_fc2_weight_to_fp16 = const()[name = tensor("layers_19_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(770566336)))]; + tensor layers_19_fc2_bias_to_fp16 = const()[name = tensor("layers_19_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(778955008)))]; + tensor hidden_states_41_cast_fp16 = conv(bias = layers_19_fc2_bias_to_fp16, dilations = hidden_states_41_dilations_0, groups = hidden_states_41_groups_0, pad = hidden_states_41_pad_0, pad_type = hidden_states_41_pad_type_0, strides = hidden_states_41_strides_0, weight = layers_19_fc2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("hidden_states_41_cast_fp16")]; + tensor inputs_121_cast_fp16 = add(x = inputs_119_cast_fp16, y = hidden_states_41_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_4336 = const()[name = tensor("op_4336"), val = tensor(3)]; + tensor out_121_axes_0 = const()[name = tensor("out_121_axes_0"), val = tensor([1])]; + tensor var_4361_to_fp16 = const()[name = tensor("op_4361_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_121_cast_fp16 = layer_norm(axes = out_121_axes_0, epsilon = var_4361_to_fp16, x = inputs_121_cast_fp16)[name = tensor("out_121_cast_fp16")]; + tensor obj_281_gamma_0_to_fp16 = const()[name = tensor("obj_281_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(778957120)))]; + tensor obj_281_beta_0_to_fp16 = const()[name = tensor("obj_281_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(778959232)))]; + tensor obj_281_epsilon_0_to_fp16 = const()[name = tensor("obj_281_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_281_cast_fp16 = batch_norm(beta = obj_281_beta_0_to_fp16, epsilon = obj_281_epsilon_0_to_fp16, gamma = obj_281_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor("obj_281_cast_fp16")]; + tensor query_81_pad_type_0 = const()[name = tensor("query_81_pad_type_0"), val = tensor("valid")]; + tensor query_81_strides_0 = const()[name = tensor("query_81_strides_0"), val = tensor([1, 1])]; + tensor query_81_pad_0 = const()[name = tensor("query_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_81_dilations_0 = const()[name = tensor("query_81_dilations_0"), val = tensor([1, 1])]; + tensor query_81_groups_0 = const()[name = tensor("query_81_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(778961344)))]; + tensor layers_20_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(781058560)))]; + tensor query_81_cast_fp16 = conv(bias = layers_20_self_attn_q_proj_bias_to_fp16, dilations = query_81_dilations_0, groups = query_81_groups_0, pad = query_81_pad_0, pad_type = query_81_pad_type_0, strides = query_81_strides_0, weight = layers_20_self_attn_q_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("query_81_cast_fp16")]; + tensor current_key_41_pad_type_0 = const()[name = tensor("current_key_41_pad_type_0"), val = tensor("valid")]; + tensor current_key_41_strides_0 = const()[name = tensor("current_key_41_strides_0"), val = tensor([1, 1])]; + tensor current_key_41_pad_0 = const()[name = tensor("current_key_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_41_dilations_0 = const()[name = tensor("current_key_41_dilations_0"), val = tensor([1, 1])]; + tensor current_key_41_groups_0 = const()[name = tensor("current_key_41_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(781060672)))]; + tensor current_key_41_cast_fp16 = conv(dilations = current_key_41_dilations_0, groups = current_key_41_groups_0, pad = current_key_41_pad_0, pad_type = current_key_41_pad_type_0, strides = current_key_41_strides_0, weight = layers_20_self_attn_k_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("current_key_41_cast_fp16")]; + tensor current_value_41_pad_type_0 = const()[name = tensor("current_value_41_pad_type_0"), val = tensor("valid")]; + tensor current_value_41_strides_0 = const()[name = tensor("current_value_41_strides_0"), val = tensor([1, 1])]; + tensor current_value_41_pad_0 = const()[name = tensor("current_value_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_41_dilations_0 = const()[name = tensor("current_value_41_dilations_0"), val = tensor([1, 1])]; + tensor current_value_41_groups_0 = const()[name = tensor("current_value_41_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(783157888)))]; + tensor layers_20_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(785255104)))]; + tensor current_value_41_cast_fp16 = conv(bias = layers_20_self_attn_v_proj_bias_to_fp16, dilations = current_value_41_dilations_0, groups = current_value_41_groups_0, pad = current_value_41_pad_0, pad_type = current_value_41_pad_type_0, strides = current_value_41_strides_0, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = obj_281_cast_fp16)[name = tensor("current_value_41_cast_fp16")]; + tensor var_4399_cast_fp16 = mul(x = current_key_41_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4399_cast_fp16")]; + tensor key_81_cast_fp16 = add(x = var_87_cast_fp16_20, y = var_4399_cast_fp16)[name = tensor("key_81_cast_fp16")]; + tensor var_4401_cast_fp16 = mul(x = current_value_41_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4401_cast_fp16")]; + tensor value_81_cast_fp16 = add(x = var_114_cast_fp16_20, y = var_4401_cast_fp16)[name = tensor("value_81_cast_fp16")]; + tensor var_4404 = const()[name = tensor("op_4404"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_81_cast_fp16 = reshape(shape = var_4404, x = query_81_cast_fp16)[name = tensor("mh_q_81_cast_fp16")]; + tensor var_4406_to_fp16 = const()[name = tensor("op_4406_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4407_cast_fp16 = mul(x = mh_q_81_cast_fp16, y = var_4406_to_fp16)[name = tensor("op_4407_cast_fp16")]; + tensor var_4408 = const()[name = tensor("op_4408"), val = tensor([1, 16, 64, -1])]; + tensor var_4409_cast_fp16 = reshape(shape = var_4408, x = key_81_cast_fp16)[name = tensor("op_4409_cast_fp16")]; + tensor mh_w_121_transpose_x_0 = const()[name = tensor("mh_w_121_transpose_x_0"), val = tensor(true)]; + tensor mh_w_121_transpose_y_0 = const()[name = tensor("mh_w_121_transpose_y_0"), val = tensor(false)]; + tensor mh_w_121_cast_fp16 = matmul(transpose_x = mh_w_121_transpose_x_0, transpose_y = mh_w_121_transpose_y_0, x = var_4407_cast_fp16, y = var_4409_cast_fp16)[name = tensor("mh_w_121_cast_fp16")]; + tensor mh_w_123_cast_fp16 = add(x = mh_w_121_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_123_cast_fp16")]; + tensor var_4417_cast_fp16 = softmax(axis = var_4336, x = mh_w_123_cast_fp16)[name = tensor("op_4417_cast_fp16")]; + tensor var_4418 = const()[name = tensor("op_4418"), val = tensor([1, 16, 64, -1])]; + tensor var_4419_cast_fp16 = reshape(shape = var_4418, x = value_81_cast_fp16)[name = tensor("op_4419_cast_fp16")]; + tensor attn_81_transpose_x_0 = const()[name = tensor("attn_81_transpose_x_0"), val = tensor(false)]; + tensor attn_81_transpose_y_0 = const()[name = tensor("attn_81_transpose_y_0"), val = tensor(true)]; + tensor attn_81_cast_fp16 = matmul(transpose_x = attn_81_transpose_x_0, transpose_y = attn_81_transpose_y_0, x = var_4419_cast_fp16, y = var_4417_cast_fp16)[name = tensor("attn_81_cast_fp16")]; + tensor var_4422 = const()[name = tensor("op_4422"), val = tensor([1, 1024, 1, -1])]; + tensor input_201_cast_fp16 = reshape(shape = var_4422, x = attn_81_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor obj_287_pad_type_0 = const()[name = tensor("obj_287_pad_type_0"), val = tensor("valid")]; + tensor obj_287_strides_0 = const()[name = tensor("obj_287_strides_0"), val = tensor([1, 1])]; + tensor obj_287_pad_0 = const()[name = tensor("obj_287_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_287_dilations_0 = const()[name = tensor("obj_287_dilations_0"), val = tensor([1, 1])]; + tensor obj_287_groups_0 = const()[name = tensor("obj_287_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(785257216)))]; + tensor layers_20_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_20_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(787354432)))]; + tensor obj_287_cast_fp16 = conv(bias = layers_20_self_attn_o_proj_bias_to_fp16, dilations = obj_287_dilations_0, groups = obj_287_groups_0, pad = obj_287_pad_0, pad_type = obj_287_pad_type_0, strides = obj_287_strides_0, weight = layers_20_self_attn_o_proj_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("obj_287_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = obj_287_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; + tensor out_123_axes_0 = const()[name = tensor("out_123_axes_0"), val = tensor([1])]; + tensor var_4444_to_fp16 = const()[name = tensor("op_4444_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_123_cast_fp16 = layer_norm(axes = out_123_axes_0, epsilon = var_4444_to_fp16, x = inputs_123_cast_fp16)[name = tensor("out_123_cast_fp16")]; + tensor obj_289_gamma_0_to_fp16 = const()[name = tensor("obj_289_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(787356544)))]; + tensor obj_289_beta_0_to_fp16 = const()[name = tensor("obj_289_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(787358656)))]; + tensor obj_289_epsilon_0_to_fp16 = const()[name = tensor("obj_289_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_289_cast_fp16 = batch_norm(beta = obj_289_beta_0_to_fp16, epsilon = obj_289_epsilon_0_to_fp16, gamma = obj_289_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor("obj_289_cast_fp16")]; + tensor query_83_pad_type_0 = const()[name = tensor("query_83_pad_type_0"), val = tensor("valid")]; + tensor query_83_strides_0 = const()[name = tensor("query_83_strides_0"), val = tensor([1, 1])]; + tensor query_83_pad_0 = const()[name = tensor("query_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_83_dilations_0 = const()[name = tensor("query_83_dilations_0"), val = tensor([1, 1])]; + tensor query_83_groups_0 = const()[name = tensor("query_83_groups_0"), val = tensor(1)]; + tensor layers_20_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(787360768)))]; + tensor layers_20_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(789457984)))]; + tensor query_83_cast_fp16 = conv(bias = layers_20_encoder_attn_q_proj_bias_to_fp16, dilations = query_83_dilations_0, groups = query_83_groups_0, pad = query_83_pad_0, pad_type = query_83_pad_type_0, strides = query_83_strides_0, weight = layers_20_encoder_attn_q_proj_weight_to_fp16, x = obj_289_cast_fp16)[name = tensor("query_83_cast_fp16")]; + tensor key_83_pad_type_0 = const()[name = tensor("key_83_pad_type_0"), val = tensor("valid")]; + tensor key_83_strides_0 = const()[name = tensor("key_83_strides_0"), val = tensor([1, 1])]; + tensor key_83_pad_0 = const()[name = tensor("key_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_83_dilations_0 = const()[name = tensor("key_83_dilations_0"), val = tensor([1, 1])]; + tensor key_83_groups_0 = const()[name = tensor("key_83_groups_0"), val = tensor(1)]; + tensor layers_20_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(789460096)))]; + tensor key_83_cast_fp16 = conv(dilations = key_83_dilations_0, groups = key_83_groups_0, pad = key_83_pad_0, pad_type = key_83_pad_type_0, strides = key_83_strides_0, weight = layers_20_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_83_cast_fp16")]; + tensor value_83_pad_type_0 = const()[name = tensor("value_83_pad_type_0"), val = tensor("valid")]; + tensor value_83_strides_0 = const()[name = tensor("value_83_strides_0"), val = tensor([1, 1])]; + tensor value_83_pad_0 = const()[name = tensor("value_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_83_dilations_0 = const()[name = tensor("value_83_dilations_0"), val = tensor([1, 1])]; + tensor value_83_groups_0 = const()[name = tensor("value_83_groups_0"), val = tensor(1)]; + tensor layers_20_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(791557312)))]; + tensor layers_20_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(793654528)))]; + tensor value_83_cast_fp16 = conv(bias = layers_20_encoder_attn_v_proj_bias_to_fp16, dilations = value_83_dilations_0, groups = value_83_groups_0, pad = value_83_pad_0, pad_type = value_83_pad_type_0, strides = value_83_strides_0, weight = layers_20_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_83_cast_fp16")]; + tensor var_4479 = const()[name = tensor("op_4479"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_83_cast_fp16 = reshape(shape = var_4479, x = query_83_cast_fp16)[name = tensor("mh_q_83_cast_fp16")]; + tensor var_4481_to_fp16 = const()[name = tensor("op_4481_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4482_cast_fp16 = mul(x = mh_q_83_cast_fp16, y = var_4481_to_fp16)[name = tensor("op_4482_cast_fp16")]; + tensor var_4483 = const()[name = tensor("op_4483"), val = tensor([1, 16, 64, -1])]; + tensor var_4484_cast_fp16 = reshape(shape = var_4483, x = key_83_cast_fp16)[name = tensor("op_4484_cast_fp16")]; + tensor mh_w_125_transpose_x_0 = const()[name = tensor("mh_w_125_transpose_x_0"), val = tensor(true)]; + tensor mh_w_125_transpose_y_0 = const()[name = tensor("mh_w_125_transpose_y_0"), val = tensor(false)]; + tensor mh_w_125_cast_fp16 = matmul(transpose_x = mh_w_125_transpose_x_0, transpose_y = mh_w_125_transpose_y_0, x = var_4482_cast_fp16, y = var_4484_cast_fp16)[name = tensor("mh_w_125_cast_fp16")]; + tensor obj_293_cast_fp16 = softmax(axis = var_4336, x = mh_w_125_cast_fp16)[name = tensor("obj_293_cast_fp16")]; + tensor var_4488 = const()[name = tensor("op_4488"), val = tensor([1, 16, 64, -1])]; + tensor var_4489_cast_fp16 = reshape(shape = var_4488, x = value_83_cast_fp16)[name = tensor("op_4489_cast_fp16")]; + tensor attn_83_transpose_x_0 = const()[name = tensor("attn_83_transpose_x_0"), val = tensor(false)]; + tensor attn_83_transpose_y_0 = const()[name = tensor("attn_83_transpose_y_0"), val = tensor(true)]; + tensor attn_83_cast_fp16 = matmul(transpose_x = attn_83_transpose_x_0, transpose_y = attn_83_transpose_y_0, x = var_4489_cast_fp16, y = obj_293_cast_fp16)[name = tensor("attn_83_cast_fp16")]; + tensor var_4492 = const()[name = tensor("op_4492"), val = tensor([1, 1024, 1, -1])]; + tensor input_203_cast_fp16 = reshape(shape = var_4492, x = attn_83_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor obj_291_pad_type_0 = const()[name = tensor("obj_291_pad_type_0"), val = tensor("valid")]; + tensor obj_291_strides_0 = const()[name = tensor("obj_291_strides_0"), val = tensor([1, 1])]; + tensor obj_291_pad_0 = const()[name = tensor("obj_291_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_291_dilations_0 = const()[name = tensor("obj_291_dilations_0"), val = tensor([1, 1])]; + tensor obj_291_groups_0 = const()[name = tensor("obj_291_groups_0"), val = tensor(1)]; + tensor layers_20_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(793656640)))]; + tensor layers_20_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_20_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(795753856)))]; + tensor obj_291_cast_fp16 = conv(bias = layers_20_encoder_attn_o_proj_bias_to_fp16, dilations = obj_291_dilations_0, groups = obj_291_groups_0, pad = obj_291_pad_0, pad_type = obj_291_pad_type_0, strides = obj_291_strides_0, weight = layers_20_encoder_attn_o_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("obj_291_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = obj_291_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; + tensor out_125_axes_0 = const()[name = tensor("out_125_axes_0"), val = tensor([1])]; + tensor var_4513_to_fp16 = const()[name = tensor("op_4513_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_125_cast_fp16 = layer_norm(axes = out_125_axes_0, epsilon = var_4513_to_fp16, x = inputs_125_cast_fp16)[name = tensor("out_125_cast_fp16")]; + tensor input_205_gamma_0_to_fp16 = const()[name = tensor("input_205_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(795755968)))]; + tensor input_205_beta_0_to_fp16 = const()[name = tensor("input_205_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(795758080)))]; + tensor input_205_epsilon_0_to_fp16 = const()[name = tensor("input_205_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_205_cast_fp16 = batch_norm(beta = input_205_beta_0_to_fp16, epsilon = input_205_epsilon_0_to_fp16, gamma = input_205_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_pad_type_0 = const()[name = tensor("input_207_pad_type_0"), val = tensor("valid")]; + tensor input_207_strides_0 = const()[name = tensor("input_207_strides_0"), val = tensor([1, 1])]; + tensor input_207_pad_0 = const()[name = tensor("input_207_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_207_dilations_0 = const()[name = tensor("input_207_dilations_0"), val = tensor([1, 1])]; + tensor input_207_groups_0 = const()[name = tensor("input_207_groups_0"), val = tensor(1)]; + tensor layers_20_fc1_weight_to_fp16 = const()[name = tensor("layers_20_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(795760192)))]; + tensor layers_20_fc1_bias_to_fp16 = const()[name = tensor("layers_20_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(804148864)))]; + tensor input_207_cast_fp16 = conv(bias = layers_20_fc1_bias_to_fp16, dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = layers_20_fc1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_mode_0 = const()[name = tensor("input_209_mode_0"), val = tensor("EXACT")]; + tensor input_209_cast_fp16 = gelu(mode = input_209_mode_0, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor hidden_states_43_pad_type_0 = const()[name = tensor("hidden_states_43_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_43_strides_0 = const()[name = tensor("hidden_states_43_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_43_pad_0 = const()[name = tensor("hidden_states_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_43_dilations_0 = const()[name = tensor("hidden_states_43_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_43_groups_0 = const()[name = tensor("hidden_states_43_groups_0"), val = tensor(1)]; + tensor layers_20_fc2_weight_to_fp16 = const()[name = tensor("layers_20_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(804157120)))]; + tensor layers_20_fc2_bias_to_fp16 = const()[name = tensor("layers_20_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(812545792)))]; + tensor hidden_states_43_cast_fp16 = conv(bias = layers_20_fc2_bias_to_fp16, dilations = hidden_states_43_dilations_0, groups = hidden_states_43_groups_0, pad = hidden_states_43_pad_0, pad_type = hidden_states_43_pad_type_0, strides = hidden_states_43_strides_0, weight = layers_20_fc2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("hidden_states_43_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = hidden_states_43_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; + tensor var_4549 = const()[name = tensor("op_4549"), val = tensor(3)]; + tensor out_127_axes_0 = const()[name = tensor("out_127_axes_0"), val = tensor([1])]; + tensor var_4574_to_fp16 = const()[name = tensor("op_4574_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_127_cast_fp16 = layer_norm(axes = out_127_axes_0, epsilon = var_4574_to_fp16, x = inputs_127_cast_fp16)[name = tensor("out_127_cast_fp16")]; + tensor obj_295_gamma_0_to_fp16 = const()[name = tensor("obj_295_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(812547904)))]; + tensor obj_295_beta_0_to_fp16 = const()[name = tensor("obj_295_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(812550016)))]; + tensor obj_295_epsilon_0_to_fp16 = const()[name = tensor("obj_295_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_295_cast_fp16 = batch_norm(beta = obj_295_beta_0_to_fp16, epsilon = obj_295_epsilon_0_to_fp16, gamma = obj_295_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor("obj_295_cast_fp16")]; + tensor query_85_pad_type_0 = const()[name = tensor("query_85_pad_type_0"), val = tensor("valid")]; + tensor query_85_strides_0 = const()[name = tensor("query_85_strides_0"), val = tensor([1, 1])]; + tensor query_85_pad_0 = const()[name = tensor("query_85_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_85_dilations_0 = const()[name = tensor("query_85_dilations_0"), val = tensor([1, 1])]; + tensor query_85_groups_0 = const()[name = tensor("query_85_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(812552128)))]; + tensor layers_21_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814649344)))]; + tensor query_85_cast_fp16 = conv(bias = layers_21_self_attn_q_proj_bias_to_fp16, dilations = query_85_dilations_0, groups = query_85_groups_0, pad = query_85_pad_0, pad_type = query_85_pad_type_0, strides = query_85_strides_0, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("query_85_cast_fp16")]; + tensor current_key_43_pad_type_0 = const()[name = tensor("current_key_43_pad_type_0"), val = tensor("valid")]; + tensor current_key_43_strides_0 = const()[name = tensor("current_key_43_strides_0"), val = tensor([1, 1])]; + tensor current_key_43_pad_0 = const()[name = tensor("current_key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_43_dilations_0 = const()[name = tensor("current_key_43_dilations_0"), val = tensor([1, 1])]; + tensor current_key_43_groups_0 = const()[name = tensor("current_key_43_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(814651456)))]; + tensor current_key_43_cast_fp16 = conv(dilations = current_key_43_dilations_0, groups = current_key_43_groups_0, pad = current_key_43_pad_0, pad_type = current_key_43_pad_type_0, strides = current_key_43_strides_0, weight = layers_21_self_attn_k_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("current_key_43_cast_fp16")]; + tensor current_value_43_pad_type_0 = const()[name = tensor("current_value_43_pad_type_0"), val = tensor("valid")]; + tensor current_value_43_strides_0 = const()[name = tensor("current_value_43_strides_0"), val = tensor([1, 1])]; + tensor current_value_43_pad_0 = const()[name = tensor("current_value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_43_dilations_0 = const()[name = tensor("current_value_43_dilations_0"), val = tensor([1, 1])]; + tensor current_value_43_groups_0 = const()[name = tensor("current_value_43_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(816748672)))]; + tensor layers_21_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(818845888)))]; + tensor current_value_43_cast_fp16 = conv(bias = layers_21_self_attn_v_proj_bias_to_fp16, dilations = current_value_43_dilations_0, groups = current_value_43_groups_0, pad = current_value_43_pad_0, pad_type = current_value_43_pad_type_0, strides = current_value_43_strides_0, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = obj_295_cast_fp16)[name = tensor("current_value_43_cast_fp16")]; + tensor var_4612_cast_fp16 = mul(x = current_key_43_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4612_cast_fp16")]; + tensor key_85_cast_fp16 = add(x = var_87_cast_fp16_21, y = var_4612_cast_fp16)[name = tensor("key_85_cast_fp16")]; + tensor var_4614_cast_fp16 = mul(x = current_value_43_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4614_cast_fp16")]; + tensor value_85_cast_fp16 = add(x = var_114_cast_fp16_21, y = var_4614_cast_fp16)[name = tensor("value_85_cast_fp16")]; + tensor var_4617 = const()[name = tensor("op_4617"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_85_cast_fp16 = reshape(shape = var_4617, x = query_85_cast_fp16)[name = tensor("mh_q_85_cast_fp16")]; + tensor var_4619_to_fp16 = const()[name = tensor("op_4619_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4620_cast_fp16 = mul(x = mh_q_85_cast_fp16, y = var_4619_to_fp16)[name = tensor("op_4620_cast_fp16")]; + tensor var_4621 = const()[name = tensor("op_4621"), val = tensor([1, 16, 64, -1])]; + tensor var_4622_cast_fp16 = reshape(shape = var_4621, x = key_85_cast_fp16)[name = tensor("op_4622_cast_fp16")]; + tensor mh_w_127_transpose_x_0 = const()[name = tensor("mh_w_127_transpose_x_0"), val = tensor(true)]; + tensor mh_w_127_transpose_y_0 = const()[name = tensor("mh_w_127_transpose_y_0"), val = tensor(false)]; + tensor mh_w_127_cast_fp16 = matmul(transpose_x = mh_w_127_transpose_x_0, transpose_y = mh_w_127_transpose_y_0, x = var_4620_cast_fp16, y = var_4622_cast_fp16)[name = tensor("mh_w_127_cast_fp16")]; + tensor mh_w_129_cast_fp16 = add(x = mh_w_127_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_129_cast_fp16")]; + tensor var_4630_cast_fp16 = softmax(axis = var_4549, x = mh_w_129_cast_fp16)[name = tensor("op_4630_cast_fp16")]; + tensor var_4631 = const()[name = tensor("op_4631"), val = tensor([1, 16, 64, -1])]; + tensor var_4632_cast_fp16 = reshape(shape = var_4631, x = value_85_cast_fp16)[name = tensor("op_4632_cast_fp16")]; + tensor attn_85_transpose_x_0 = const()[name = tensor("attn_85_transpose_x_0"), val = tensor(false)]; + tensor attn_85_transpose_y_0 = const()[name = tensor("attn_85_transpose_y_0"), val = tensor(true)]; + tensor attn_85_cast_fp16 = matmul(transpose_x = attn_85_transpose_x_0, transpose_y = attn_85_transpose_y_0, x = var_4632_cast_fp16, y = var_4630_cast_fp16)[name = tensor("attn_85_cast_fp16")]; + tensor var_4635 = const()[name = tensor("op_4635"), val = tensor([1, 1024, 1, -1])]; + tensor input_211_cast_fp16 = reshape(shape = var_4635, x = attn_85_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor obj_301_pad_type_0 = const()[name = tensor("obj_301_pad_type_0"), val = tensor("valid")]; + tensor obj_301_strides_0 = const()[name = tensor("obj_301_strides_0"), val = tensor([1, 1])]; + tensor obj_301_pad_0 = const()[name = tensor("obj_301_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_301_dilations_0 = const()[name = tensor("obj_301_dilations_0"), val = tensor([1, 1])]; + tensor obj_301_groups_0 = const()[name = tensor("obj_301_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(818848000)))]; + tensor layers_21_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_21_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(820945216)))]; + tensor obj_301_cast_fp16 = conv(bias = layers_21_self_attn_o_proj_bias_to_fp16, dilations = obj_301_dilations_0, groups = obj_301_groups_0, pad = obj_301_pad_0, pad_type = obj_301_pad_type_0, strides = obj_301_strides_0, weight = layers_21_self_attn_o_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("obj_301_cast_fp16")]; + tensor inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = obj_301_cast_fp16)[name = tensor("inputs_129_cast_fp16")]; + tensor out_129_axes_0 = const()[name = tensor("out_129_axes_0"), val = tensor([1])]; + tensor var_4657_to_fp16 = const()[name = tensor("op_4657_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_129_cast_fp16 = layer_norm(axes = out_129_axes_0, epsilon = var_4657_to_fp16, x = inputs_129_cast_fp16)[name = tensor("out_129_cast_fp16")]; + tensor obj_303_gamma_0_to_fp16 = const()[name = tensor("obj_303_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(820947328)))]; + tensor obj_303_beta_0_to_fp16 = const()[name = tensor("obj_303_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(820949440)))]; + tensor obj_303_epsilon_0_to_fp16 = const()[name = tensor("obj_303_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_303_cast_fp16 = batch_norm(beta = obj_303_beta_0_to_fp16, epsilon = obj_303_epsilon_0_to_fp16, gamma = obj_303_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_129_cast_fp16)[name = tensor("obj_303_cast_fp16")]; + tensor query_87_pad_type_0 = const()[name = tensor("query_87_pad_type_0"), val = tensor("valid")]; + tensor query_87_strides_0 = const()[name = tensor("query_87_strides_0"), val = tensor([1, 1])]; + tensor query_87_pad_0 = const()[name = tensor("query_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_87_dilations_0 = const()[name = tensor("query_87_dilations_0"), val = tensor([1, 1])]; + tensor query_87_groups_0 = const()[name = tensor("query_87_groups_0"), val = tensor(1)]; + tensor layers_21_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(820951552)))]; + tensor layers_21_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(823048768)))]; + tensor query_87_cast_fp16 = conv(bias = layers_21_encoder_attn_q_proj_bias_to_fp16, dilations = query_87_dilations_0, groups = query_87_groups_0, pad = query_87_pad_0, pad_type = query_87_pad_type_0, strides = query_87_strides_0, weight = layers_21_encoder_attn_q_proj_weight_to_fp16, x = obj_303_cast_fp16)[name = tensor("query_87_cast_fp16")]; + tensor key_87_pad_type_0 = const()[name = tensor("key_87_pad_type_0"), val = tensor("valid")]; + tensor key_87_strides_0 = const()[name = tensor("key_87_strides_0"), val = tensor([1, 1])]; + tensor key_87_pad_0 = const()[name = tensor("key_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_87_dilations_0 = const()[name = tensor("key_87_dilations_0"), val = tensor([1, 1])]; + tensor key_87_groups_0 = const()[name = tensor("key_87_groups_0"), val = tensor(1)]; + tensor layers_21_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(823050880)))]; + tensor key_87_cast_fp16 = conv(dilations = key_87_dilations_0, groups = key_87_groups_0, pad = key_87_pad_0, pad_type = key_87_pad_type_0, strides = key_87_strides_0, weight = layers_21_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_87_cast_fp16")]; + tensor value_87_pad_type_0 = const()[name = tensor("value_87_pad_type_0"), val = tensor("valid")]; + tensor value_87_strides_0 = const()[name = tensor("value_87_strides_0"), val = tensor([1, 1])]; + tensor value_87_pad_0 = const()[name = tensor("value_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_87_dilations_0 = const()[name = tensor("value_87_dilations_0"), val = tensor([1, 1])]; + tensor value_87_groups_0 = const()[name = tensor("value_87_groups_0"), val = tensor(1)]; + tensor layers_21_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(825148096)))]; + tensor layers_21_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827245312)))]; + tensor value_87_cast_fp16 = conv(bias = layers_21_encoder_attn_v_proj_bias_to_fp16, dilations = value_87_dilations_0, groups = value_87_groups_0, pad = value_87_pad_0, pad_type = value_87_pad_type_0, strides = value_87_strides_0, weight = layers_21_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_87_cast_fp16")]; + tensor var_4692 = const()[name = tensor("op_4692"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_87_cast_fp16 = reshape(shape = var_4692, x = query_87_cast_fp16)[name = tensor("mh_q_87_cast_fp16")]; + tensor var_4694_to_fp16 = const()[name = tensor("op_4694_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4695_cast_fp16 = mul(x = mh_q_87_cast_fp16, y = var_4694_to_fp16)[name = tensor("op_4695_cast_fp16")]; + tensor var_4696 = const()[name = tensor("op_4696"), val = tensor([1, 16, 64, -1])]; + tensor var_4697_cast_fp16 = reshape(shape = var_4696, x = key_87_cast_fp16)[name = tensor("op_4697_cast_fp16")]; + tensor mh_w_131_transpose_x_0 = const()[name = tensor("mh_w_131_transpose_x_0"), val = tensor(true)]; + tensor mh_w_131_transpose_y_0 = const()[name = tensor("mh_w_131_transpose_y_0"), val = tensor(false)]; + tensor mh_w_131_cast_fp16 = matmul(transpose_x = mh_w_131_transpose_x_0, transpose_y = mh_w_131_transpose_y_0, x = var_4695_cast_fp16, y = var_4697_cast_fp16)[name = tensor("mh_w_131_cast_fp16")]; + tensor obj_307_cast_fp16 = softmax(axis = var_4549, x = mh_w_131_cast_fp16)[name = tensor("obj_307_cast_fp16")]; + tensor var_4701 = const()[name = tensor("op_4701"), val = tensor([1, 16, 64, -1])]; + tensor var_4702_cast_fp16 = reshape(shape = var_4701, x = value_87_cast_fp16)[name = tensor("op_4702_cast_fp16")]; + tensor attn_87_transpose_x_0 = const()[name = tensor("attn_87_transpose_x_0"), val = tensor(false)]; + tensor attn_87_transpose_y_0 = const()[name = tensor("attn_87_transpose_y_0"), val = tensor(true)]; + tensor attn_87_cast_fp16 = matmul(transpose_x = attn_87_transpose_x_0, transpose_y = attn_87_transpose_y_0, x = var_4702_cast_fp16, y = obj_307_cast_fp16)[name = tensor("attn_87_cast_fp16")]; + tensor var_4705 = const()[name = tensor("op_4705"), val = tensor([1, 1024, 1, -1])]; + tensor input_213_cast_fp16 = reshape(shape = var_4705, x = attn_87_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor obj_305_pad_type_0 = const()[name = tensor("obj_305_pad_type_0"), val = tensor("valid")]; + tensor obj_305_strides_0 = const()[name = tensor("obj_305_strides_0"), val = tensor([1, 1])]; + tensor obj_305_pad_0 = const()[name = tensor("obj_305_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_305_dilations_0 = const()[name = tensor("obj_305_dilations_0"), val = tensor([1, 1])]; + tensor obj_305_groups_0 = const()[name = tensor("obj_305_groups_0"), val = tensor(1)]; + tensor layers_21_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(827247424)))]; + tensor layers_21_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_21_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829344640)))]; + tensor obj_305_cast_fp16 = conv(bias = layers_21_encoder_attn_o_proj_bias_to_fp16, dilations = obj_305_dilations_0, groups = obj_305_groups_0, pad = obj_305_pad_0, pad_type = obj_305_pad_type_0, strides = obj_305_strides_0, weight = layers_21_encoder_attn_o_proj_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("obj_305_cast_fp16")]; + tensor inputs_131_cast_fp16 = add(x = inputs_129_cast_fp16, y = obj_305_cast_fp16)[name = tensor("inputs_131_cast_fp16")]; + tensor out_131_axes_0 = const()[name = tensor("out_131_axes_0"), val = tensor([1])]; + tensor var_4723_to_fp16 = const()[name = tensor("op_4723_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_131_cast_fp16 = layer_norm(axes = out_131_axes_0, epsilon = var_4723_to_fp16, x = inputs_131_cast_fp16)[name = tensor("out_131_cast_fp16")]; + tensor input_215_gamma_0_to_fp16 = const()[name = tensor("input_215_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829346752)))]; + tensor input_215_beta_0_to_fp16 = const()[name = tensor("input_215_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829348864)))]; + tensor input_215_epsilon_0_to_fp16 = const()[name = tensor("input_215_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = batch_norm(beta = input_215_beta_0_to_fp16, epsilon = input_215_epsilon_0_to_fp16, gamma = input_215_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_131_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor input_217_pad_type_0 = const()[name = tensor("input_217_pad_type_0"), val = tensor("valid")]; + tensor input_217_strides_0 = const()[name = tensor("input_217_strides_0"), val = tensor([1, 1])]; + tensor input_217_pad_0 = const()[name = tensor("input_217_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_217_dilations_0 = const()[name = tensor("input_217_dilations_0"), val = tensor([1, 1])]; + tensor input_217_groups_0 = const()[name = tensor("input_217_groups_0"), val = tensor(1)]; + tensor layers_21_fc1_weight_to_fp16 = const()[name = tensor("layers_21_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(829350976)))]; + tensor layers_21_fc1_bias_to_fp16 = const()[name = tensor("layers_21_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(837739648)))]; + tensor input_217_cast_fp16 = conv(bias = layers_21_fc1_bias_to_fp16, dilations = input_217_dilations_0, groups = input_217_groups_0, pad = input_217_pad_0, pad_type = input_217_pad_type_0, strides = input_217_strides_0, weight = layers_21_fc1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_mode_0 = const()[name = tensor("input_219_mode_0"), val = tensor("EXACT")]; + tensor input_219_cast_fp16 = gelu(mode = input_219_mode_0, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor hidden_states_45_pad_type_0 = const()[name = tensor("hidden_states_45_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_45_strides_0 = const()[name = tensor("hidden_states_45_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_45_pad_0 = const()[name = tensor("hidden_states_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_45_dilations_0 = const()[name = tensor("hidden_states_45_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_45_groups_0 = const()[name = tensor("hidden_states_45_groups_0"), val = tensor(1)]; + tensor layers_21_fc2_weight_to_fp16 = const()[name = tensor("layers_21_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(837747904)))]; + tensor layers_21_fc2_bias_to_fp16 = const()[name = tensor("layers_21_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(846136576)))]; + tensor hidden_states_45_cast_fp16 = conv(bias = layers_21_fc2_bias_to_fp16, dilations = hidden_states_45_dilations_0, groups = hidden_states_45_groups_0, pad = hidden_states_45_pad_0, pad_type = hidden_states_45_pad_type_0, strides = hidden_states_45_strides_0, weight = layers_21_fc2_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("hidden_states_45_cast_fp16")]; + tensor inputs_133_cast_fp16 = add(x = inputs_131_cast_fp16, y = hidden_states_45_cast_fp16)[name = tensor("inputs_133_cast_fp16")]; + tensor var_4758 = const()[name = tensor("op_4758"), val = tensor(3)]; + tensor out_133_axes_0 = const()[name = tensor("out_133_axes_0"), val = tensor([1])]; + tensor var_4783_to_fp16 = const()[name = tensor("op_4783_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_133_cast_fp16 = layer_norm(axes = out_133_axes_0, epsilon = var_4783_to_fp16, x = inputs_133_cast_fp16)[name = tensor("out_133_cast_fp16")]; + tensor obj_309_gamma_0_to_fp16 = const()[name = tensor("obj_309_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(846138688)))]; + tensor obj_309_beta_0_to_fp16 = const()[name = tensor("obj_309_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(846140800)))]; + tensor obj_309_epsilon_0_to_fp16 = const()[name = tensor("obj_309_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_309_cast_fp16 = batch_norm(beta = obj_309_beta_0_to_fp16, epsilon = obj_309_epsilon_0_to_fp16, gamma = obj_309_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_133_cast_fp16)[name = tensor("obj_309_cast_fp16")]; + tensor query_89_pad_type_0 = const()[name = tensor("query_89_pad_type_0"), val = tensor("valid")]; + tensor query_89_strides_0 = const()[name = tensor("query_89_strides_0"), val = tensor([1, 1])]; + tensor query_89_pad_0 = const()[name = tensor("query_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_89_dilations_0 = const()[name = tensor("query_89_dilations_0"), val = tensor([1, 1])]; + tensor query_89_groups_0 = const()[name = tensor("query_89_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(846142912)))]; + tensor layers_22_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(848240128)))]; + tensor query_89_cast_fp16 = conv(bias = layers_22_self_attn_q_proj_bias_to_fp16, dilations = query_89_dilations_0, groups = query_89_groups_0, pad = query_89_pad_0, pad_type = query_89_pad_type_0, strides = query_89_strides_0, weight = layers_22_self_attn_q_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("query_89_cast_fp16")]; + tensor current_key_45_pad_type_0 = const()[name = tensor("current_key_45_pad_type_0"), val = tensor("valid")]; + tensor current_key_45_strides_0 = const()[name = tensor("current_key_45_strides_0"), val = tensor([1, 1])]; + tensor current_key_45_pad_0 = const()[name = tensor("current_key_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_45_dilations_0 = const()[name = tensor("current_key_45_dilations_0"), val = tensor([1, 1])]; + tensor current_key_45_groups_0 = const()[name = tensor("current_key_45_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(848242240)))]; + tensor current_key_45_cast_fp16 = conv(dilations = current_key_45_dilations_0, groups = current_key_45_groups_0, pad = current_key_45_pad_0, pad_type = current_key_45_pad_type_0, strides = current_key_45_strides_0, weight = layers_22_self_attn_k_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("current_key_45_cast_fp16")]; + tensor current_value_45_pad_type_0 = const()[name = tensor("current_value_45_pad_type_0"), val = tensor("valid")]; + tensor current_value_45_strides_0 = const()[name = tensor("current_value_45_strides_0"), val = tensor([1, 1])]; + tensor current_value_45_pad_0 = const()[name = tensor("current_value_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_45_dilations_0 = const()[name = tensor("current_value_45_dilations_0"), val = tensor([1, 1])]; + tensor current_value_45_groups_0 = const()[name = tensor("current_value_45_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(850339456)))]; + tensor layers_22_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(852436672)))]; + tensor current_value_45_cast_fp16 = conv(bias = layers_22_self_attn_v_proj_bias_to_fp16, dilations = current_value_45_dilations_0, groups = current_value_45_groups_0, pad = current_value_45_pad_0, pad_type = current_value_45_pad_type_0, strides = current_value_45_strides_0, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = obj_309_cast_fp16)[name = tensor("current_value_45_cast_fp16")]; + tensor var_4821_cast_fp16 = mul(x = current_key_45_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4821_cast_fp16")]; + tensor key_89_cast_fp16 = add(x = var_87_cast_fp16_22, y = var_4821_cast_fp16)[name = tensor("key_89_cast_fp16")]; + tensor var_4823_cast_fp16 = mul(x = current_value_45_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_4823_cast_fp16")]; + tensor value_89_cast_fp16 = add(x = var_114_cast_fp16_22, y = var_4823_cast_fp16)[name = tensor("value_89_cast_fp16")]; + tensor var_4826 = const()[name = tensor("op_4826"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_89_cast_fp16 = reshape(shape = var_4826, x = query_89_cast_fp16)[name = tensor("mh_q_89_cast_fp16")]; + tensor var_4828_to_fp16 = const()[name = tensor("op_4828_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4829_cast_fp16 = mul(x = mh_q_89_cast_fp16, y = var_4828_to_fp16)[name = tensor("op_4829_cast_fp16")]; + tensor var_4830 = const()[name = tensor("op_4830"), val = tensor([1, 16, 64, -1])]; + tensor var_4831_cast_fp16 = reshape(shape = var_4830, x = key_89_cast_fp16)[name = tensor("op_4831_cast_fp16")]; + tensor mh_w_133_transpose_x_0 = const()[name = tensor("mh_w_133_transpose_x_0"), val = tensor(true)]; + tensor mh_w_133_transpose_y_0 = const()[name = tensor("mh_w_133_transpose_y_0"), val = tensor(false)]; + tensor mh_w_133_cast_fp16 = matmul(transpose_x = mh_w_133_transpose_x_0, transpose_y = mh_w_133_transpose_y_0, x = var_4829_cast_fp16, y = var_4831_cast_fp16)[name = tensor("mh_w_133_cast_fp16")]; + tensor mh_w_135_cast_fp16 = add(x = mh_w_133_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_135_cast_fp16")]; + tensor var_4839_cast_fp16 = softmax(axis = var_4758, x = mh_w_135_cast_fp16)[name = tensor("op_4839_cast_fp16")]; + tensor var_4840 = const()[name = tensor("op_4840"), val = tensor([1, 16, 64, -1])]; + tensor var_4841_cast_fp16 = reshape(shape = var_4840, x = value_89_cast_fp16)[name = tensor("op_4841_cast_fp16")]; + tensor attn_89_transpose_x_0 = const()[name = tensor("attn_89_transpose_x_0"), val = tensor(false)]; + tensor attn_89_transpose_y_0 = const()[name = tensor("attn_89_transpose_y_0"), val = tensor(true)]; + tensor attn_89_cast_fp16 = matmul(transpose_x = attn_89_transpose_x_0, transpose_y = attn_89_transpose_y_0, x = var_4841_cast_fp16, y = var_4839_cast_fp16)[name = tensor("attn_89_cast_fp16")]; + tensor var_4844 = const()[name = tensor("op_4844"), val = tensor([1, 1024, 1, -1])]; + tensor input_221_cast_fp16 = reshape(shape = var_4844, x = attn_89_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor obj_315_pad_type_0 = const()[name = tensor("obj_315_pad_type_0"), val = tensor("valid")]; + tensor obj_315_strides_0 = const()[name = tensor("obj_315_strides_0"), val = tensor([1, 1])]; + tensor obj_315_pad_0 = const()[name = tensor("obj_315_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_315_dilations_0 = const()[name = tensor("obj_315_dilations_0"), val = tensor([1, 1])]; + tensor obj_315_groups_0 = const()[name = tensor("obj_315_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(852438784)))]; + tensor layers_22_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_22_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854536000)))]; + tensor obj_315_cast_fp16 = conv(bias = layers_22_self_attn_o_proj_bias_to_fp16, dilations = obj_315_dilations_0, groups = obj_315_groups_0, pad = obj_315_pad_0, pad_type = obj_315_pad_type_0, strides = obj_315_strides_0, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("obj_315_cast_fp16")]; + tensor inputs_135_cast_fp16 = add(x = inputs_133_cast_fp16, y = obj_315_cast_fp16)[name = tensor("inputs_135_cast_fp16")]; + tensor out_135_axes_0 = const()[name = tensor("out_135_axes_0"), val = tensor([1])]; + tensor var_4866_to_fp16 = const()[name = tensor("op_4866_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_135_cast_fp16 = layer_norm(axes = out_135_axes_0, epsilon = var_4866_to_fp16, x = inputs_135_cast_fp16)[name = tensor("out_135_cast_fp16")]; + tensor obj_317_gamma_0_to_fp16 = const()[name = tensor("obj_317_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854538112)))]; + tensor obj_317_beta_0_to_fp16 = const()[name = tensor("obj_317_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854540224)))]; + tensor obj_317_epsilon_0_to_fp16 = const()[name = tensor("obj_317_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_317_cast_fp16 = batch_norm(beta = obj_317_beta_0_to_fp16, epsilon = obj_317_epsilon_0_to_fp16, gamma = obj_317_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_135_cast_fp16)[name = tensor("obj_317_cast_fp16")]; + tensor query_91_pad_type_0 = const()[name = tensor("query_91_pad_type_0"), val = tensor("valid")]; + tensor query_91_strides_0 = const()[name = tensor("query_91_strides_0"), val = tensor([1, 1])]; + tensor query_91_pad_0 = const()[name = tensor("query_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_91_dilations_0 = const()[name = tensor("query_91_dilations_0"), val = tensor([1, 1])]; + tensor query_91_groups_0 = const()[name = tensor("query_91_groups_0"), val = tensor(1)]; + tensor layers_22_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(854542336)))]; + tensor layers_22_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(856639552)))]; + tensor query_91_cast_fp16 = conv(bias = layers_22_encoder_attn_q_proj_bias_to_fp16, dilations = query_91_dilations_0, groups = query_91_groups_0, pad = query_91_pad_0, pad_type = query_91_pad_type_0, strides = query_91_strides_0, weight = layers_22_encoder_attn_q_proj_weight_to_fp16, x = obj_317_cast_fp16)[name = tensor("query_91_cast_fp16")]; + tensor key_91_pad_type_0 = const()[name = tensor("key_91_pad_type_0"), val = tensor("valid")]; + tensor key_91_strides_0 = const()[name = tensor("key_91_strides_0"), val = tensor([1, 1])]; + tensor key_91_pad_0 = const()[name = tensor("key_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_91_dilations_0 = const()[name = tensor("key_91_dilations_0"), val = tensor([1, 1])]; + tensor key_91_groups_0 = const()[name = tensor("key_91_groups_0"), val = tensor(1)]; + tensor layers_22_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(856641664)))]; + tensor key_91_cast_fp16 = conv(dilations = key_91_dilations_0, groups = key_91_groups_0, pad = key_91_pad_0, pad_type = key_91_pad_type_0, strides = key_91_strides_0, weight = layers_22_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_91_cast_fp16")]; + tensor value_91_pad_type_0 = const()[name = tensor("value_91_pad_type_0"), val = tensor("valid")]; + tensor value_91_strides_0 = const()[name = tensor("value_91_strides_0"), val = tensor([1, 1])]; + tensor value_91_pad_0 = const()[name = tensor("value_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_91_dilations_0 = const()[name = tensor("value_91_dilations_0"), val = tensor([1, 1])]; + tensor value_91_groups_0 = const()[name = tensor("value_91_groups_0"), val = tensor(1)]; + tensor layers_22_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(858738880)))]; + tensor layers_22_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(860836096)))]; + tensor value_91_cast_fp16 = conv(bias = layers_22_encoder_attn_v_proj_bias_to_fp16, dilations = value_91_dilations_0, groups = value_91_groups_0, pad = value_91_pad_0, pad_type = value_91_pad_type_0, strides = value_91_strides_0, weight = layers_22_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_91_cast_fp16")]; + tensor var_4901 = const()[name = tensor("op_4901"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_91_cast_fp16 = reshape(shape = var_4901, x = query_91_cast_fp16)[name = tensor("mh_q_91_cast_fp16")]; + tensor var_4903_to_fp16 = const()[name = tensor("op_4903_to_fp16"), val = tensor(0x1p-3)]; + tensor var_4904_cast_fp16 = mul(x = mh_q_91_cast_fp16, y = var_4903_to_fp16)[name = tensor("op_4904_cast_fp16")]; + tensor var_4905 = const()[name = tensor("op_4905"), val = tensor([1, 16, 64, -1])]; + tensor var_4906_cast_fp16 = reshape(shape = var_4905, x = key_91_cast_fp16)[name = tensor("op_4906_cast_fp16")]; + tensor mh_w_137_transpose_x_0 = const()[name = tensor("mh_w_137_transpose_x_0"), val = tensor(true)]; + tensor mh_w_137_transpose_y_0 = const()[name = tensor("mh_w_137_transpose_y_0"), val = tensor(false)]; + tensor mh_w_137_cast_fp16 = matmul(transpose_x = mh_w_137_transpose_x_0, transpose_y = mh_w_137_transpose_y_0, x = var_4904_cast_fp16, y = var_4906_cast_fp16)[name = tensor("mh_w_137_cast_fp16")]; + tensor obj_321_cast_fp16 = softmax(axis = var_4758, x = mh_w_137_cast_fp16)[name = tensor("obj_321_cast_fp16")]; + tensor var_4910 = const()[name = tensor("op_4910"), val = tensor([1, 16, 64, -1])]; + tensor var_4911_cast_fp16 = reshape(shape = var_4910, x = value_91_cast_fp16)[name = tensor("op_4911_cast_fp16")]; + tensor attn_91_transpose_x_0 = const()[name = tensor("attn_91_transpose_x_0"), val = tensor(false)]; + tensor attn_91_transpose_y_0 = const()[name = tensor("attn_91_transpose_y_0"), val = tensor(true)]; + tensor attn_91_cast_fp16 = matmul(transpose_x = attn_91_transpose_x_0, transpose_y = attn_91_transpose_y_0, x = var_4911_cast_fp16, y = obj_321_cast_fp16)[name = tensor("attn_91_cast_fp16")]; + tensor var_4914 = const()[name = tensor("op_4914"), val = tensor([1, 1024, 1, -1])]; + tensor input_223_cast_fp16 = reshape(shape = var_4914, x = attn_91_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor obj_319_pad_type_0 = const()[name = tensor("obj_319_pad_type_0"), val = tensor("valid")]; + tensor obj_319_strides_0 = const()[name = tensor("obj_319_strides_0"), val = tensor([1, 1])]; + tensor obj_319_pad_0 = const()[name = tensor("obj_319_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_319_dilations_0 = const()[name = tensor("obj_319_dilations_0"), val = tensor([1, 1])]; + tensor obj_319_groups_0 = const()[name = tensor("obj_319_groups_0"), val = tensor(1)]; + tensor layers_22_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(860838208)))]; + tensor layers_22_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_22_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(862935424)))]; + tensor obj_319_cast_fp16 = conv(bias = layers_22_encoder_attn_o_proj_bias_to_fp16, dilations = obj_319_dilations_0, groups = obj_319_groups_0, pad = obj_319_pad_0, pad_type = obj_319_pad_type_0, strides = obj_319_strides_0, weight = layers_22_encoder_attn_o_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("obj_319_cast_fp16")]; + tensor inputs_137_cast_fp16 = add(x = inputs_135_cast_fp16, y = obj_319_cast_fp16)[name = tensor("inputs_137_cast_fp16")]; + tensor out_137_axes_0 = const()[name = tensor("out_137_axes_0"), val = tensor([1])]; + tensor var_4932_to_fp16 = const()[name = tensor("op_4932_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_137_cast_fp16 = layer_norm(axes = out_137_axes_0, epsilon = var_4932_to_fp16, x = inputs_137_cast_fp16)[name = tensor("out_137_cast_fp16")]; + tensor input_225_gamma_0_to_fp16 = const()[name = tensor("input_225_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(862937536)))]; + tensor input_225_beta_0_to_fp16 = const()[name = tensor("input_225_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(862939648)))]; + tensor input_225_epsilon_0_to_fp16 = const()[name = tensor("input_225_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = batch_norm(beta = input_225_beta_0_to_fp16, epsilon = input_225_epsilon_0_to_fp16, gamma = input_225_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_137_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_pad_type_0 = const()[name = tensor("input_227_pad_type_0"), val = tensor("valid")]; + tensor input_227_strides_0 = const()[name = tensor("input_227_strides_0"), val = tensor([1, 1])]; + tensor input_227_pad_0 = const()[name = tensor("input_227_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_227_dilations_0 = const()[name = tensor("input_227_dilations_0"), val = tensor([1, 1])]; + tensor input_227_groups_0 = const()[name = tensor("input_227_groups_0"), val = tensor(1)]; + tensor layers_22_fc1_weight_to_fp16 = const()[name = tensor("layers_22_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(862941760)))]; + tensor layers_22_fc1_bias_to_fp16 = const()[name = tensor("layers_22_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(871330432)))]; + tensor input_227_cast_fp16 = conv(bias = layers_22_fc1_bias_to_fp16, dilations = input_227_dilations_0, groups = input_227_groups_0, pad = input_227_pad_0, pad_type = input_227_pad_type_0, strides = input_227_strides_0, weight = layers_22_fc1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor hidden_states_47_pad_type_0 = const()[name = tensor("hidden_states_47_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_47_strides_0 = const()[name = tensor("hidden_states_47_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_47_pad_0 = const()[name = tensor("hidden_states_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_47_dilations_0 = const()[name = tensor("hidden_states_47_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_47_groups_0 = const()[name = tensor("hidden_states_47_groups_0"), val = tensor(1)]; + tensor layers_22_fc2_weight_to_fp16 = const()[name = tensor("layers_22_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(871338688)))]; + tensor layers_22_fc2_bias_to_fp16 = const()[name = tensor("layers_22_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(879727360)))]; + tensor hidden_states_47_cast_fp16 = conv(bias = layers_22_fc2_bias_to_fp16, dilations = hidden_states_47_dilations_0, groups = hidden_states_47_groups_0, pad = hidden_states_47_pad_0, pad_type = hidden_states_47_pad_type_0, strides = hidden_states_47_strides_0, weight = layers_22_fc2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("hidden_states_47_cast_fp16")]; + tensor inputs_139_cast_fp16 = add(x = inputs_137_cast_fp16, y = hidden_states_47_cast_fp16)[name = tensor("inputs_139_cast_fp16")]; + tensor var_4967 = const()[name = tensor("op_4967"), val = tensor(3)]; + tensor out_139_axes_0 = const()[name = tensor("out_139_axes_0"), val = tensor([1])]; + tensor var_4992_to_fp16 = const()[name = tensor("op_4992_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_139_cast_fp16 = layer_norm(axes = out_139_axes_0, epsilon = var_4992_to_fp16, x = inputs_139_cast_fp16)[name = tensor("out_139_cast_fp16")]; + tensor obj_323_gamma_0_to_fp16 = const()[name = tensor("obj_323_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(879729472)))]; + tensor obj_323_beta_0_to_fp16 = const()[name = tensor("obj_323_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(879731584)))]; + tensor obj_323_epsilon_0_to_fp16 = const()[name = tensor("obj_323_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_323_cast_fp16 = batch_norm(beta = obj_323_beta_0_to_fp16, epsilon = obj_323_epsilon_0_to_fp16, gamma = obj_323_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_139_cast_fp16)[name = tensor("obj_323_cast_fp16")]; + tensor query_93_pad_type_0 = const()[name = tensor("query_93_pad_type_0"), val = tensor("valid")]; + tensor query_93_strides_0 = const()[name = tensor("query_93_strides_0"), val = tensor([1, 1])]; + tensor query_93_pad_0 = const()[name = tensor("query_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_93_dilations_0 = const()[name = tensor("query_93_dilations_0"), val = tensor([1, 1])]; + tensor query_93_groups_0 = const()[name = tensor("query_93_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(879733696)))]; + tensor layers_23_self_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881830912)))]; + tensor query_93_cast_fp16 = conv(bias = layers_23_self_attn_q_proj_bias_to_fp16, dilations = query_93_dilations_0, groups = query_93_groups_0, pad = query_93_pad_0, pad_type = query_93_pad_type_0, strides = query_93_strides_0, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("query_93_cast_fp16")]; + tensor current_key_pad_type_0 = const()[name = tensor("current_key_pad_type_0"), val = tensor("valid")]; + tensor current_key_strides_0 = const()[name = tensor("current_key_strides_0"), val = tensor([1, 1])]; + tensor current_key_pad_0 = const()[name = tensor("current_key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_key_dilations_0 = const()[name = tensor("current_key_dilations_0"), val = tensor([1, 1])]; + tensor current_key_groups_0 = const()[name = tensor("current_key_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(881833024)))]; + tensor current_key_cast_fp16 = conv(dilations = current_key_dilations_0, groups = current_key_groups_0, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = current_key_strides_0, weight = layers_23_self_attn_k_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("current_key_cast_fp16")]; + tensor current_value_pad_type_0 = const()[name = tensor("current_value_pad_type_0"), val = tensor("valid")]; + tensor current_value_strides_0 = const()[name = tensor("current_value_strides_0"), val = tensor([1, 1])]; + tensor current_value_pad_0 = const()[name = tensor("current_value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor current_value_dilations_0 = const()[name = tensor("current_value_dilations_0"), val = tensor([1, 1])]; + tensor current_value_groups_0 = const()[name = tensor("current_value_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(883930240)))]; + tensor layers_23_self_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886027456)))]; + tensor current_value_cast_fp16 = conv(bias = layers_23_self_attn_v_proj_bias_to_fp16, dilations = current_value_dilations_0, groups = current_value_groups_0, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = current_value_strides_0, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = obj_323_cast_fp16)[name = tensor("current_value_cast_fp16")]; + tensor var_5030_cast_fp16 = mul(x = current_key_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_5030_cast_fp16")]; + tensor key_93_cast_fp16 = add(x = var_87_cast_fp16_23, y = var_5030_cast_fp16)[name = tensor("key_93_cast_fp16")]; + tensor var_5032_cast_fp16 = mul(x = current_value_cast_fp16, y = var_205_cast_fp16)[name = tensor("op_5032_cast_fp16")]; + tensor value_93_cast_fp16 = add(x = var_114_cast_fp16_23, y = var_5032_cast_fp16)[name = tensor("value_93_cast_fp16")]; + tensor var_5035 = const()[name = tensor("op_5035"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_93_cast_fp16 = reshape(shape = var_5035, x = query_93_cast_fp16)[name = tensor("mh_q_93_cast_fp16")]; + tensor var_5037_to_fp16 = const()[name = tensor("op_5037_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5038_cast_fp16 = mul(x = mh_q_93_cast_fp16, y = var_5037_to_fp16)[name = tensor("op_5038_cast_fp16")]; + tensor var_5039 = const()[name = tensor("op_5039"), val = tensor([1, 16, 64, -1])]; + tensor var_5040_cast_fp16 = reshape(shape = var_5039, x = key_93_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor mh_w_139_transpose_x_0 = const()[name = tensor("mh_w_139_transpose_x_0"), val = tensor(true)]; + tensor mh_w_139_transpose_y_0 = const()[name = tensor("mh_w_139_transpose_y_0"), val = tensor(false)]; + tensor mh_w_139_cast_fp16 = matmul(transpose_x = mh_w_139_transpose_x_0, transpose_y = mh_w_139_transpose_y_0, x = var_5038_cast_fp16, y = var_5040_cast_fp16)[name = tensor("mh_w_139_cast_fp16")]; + tensor mh_w_141_cast_fp16 = add(x = mh_w_139_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_141_cast_fp16")]; + tensor var_5048_cast_fp16 = softmax(axis = var_4967, x = mh_w_141_cast_fp16)[name = tensor("op_5048_cast_fp16")]; + tensor var_5049 = const()[name = tensor("op_5049"), val = tensor([1, 16, 64, -1])]; + tensor var_5050_cast_fp16 = reshape(shape = var_5049, x = value_93_cast_fp16)[name = tensor("op_5050_cast_fp16")]; + tensor attn_93_transpose_x_0 = const()[name = tensor("attn_93_transpose_x_0"), val = tensor(false)]; + tensor attn_93_transpose_y_0 = const()[name = tensor("attn_93_transpose_y_0"), val = tensor(true)]; + tensor attn_93_cast_fp16 = matmul(transpose_x = attn_93_transpose_x_0, transpose_y = attn_93_transpose_y_0, x = var_5050_cast_fp16, y = var_5048_cast_fp16)[name = tensor("attn_93_cast_fp16")]; + tensor var_5053 = const()[name = tensor("op_5053"), val = tensor([1, 1024, 1, -1])]; + tensor input_231_cast_fp16 = reshape(shape = var_5053, x = attn_93_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor obj_329_pad_type_0 = const()[name = tensor("obj_329_pad_type_0"), val = tensor("valid")]; + tensor obj_329_strides_0 = const()[name = tensor("obj_329_strides_0"), val = tensor([1, 1])]; + tensor obj_329_pad_0 = const()[name = tensor("obj_329_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_329_dilations_0 = const()[name = tensor("obj_329_dilations_0"), val = tensor([1, 1])]; + tensor obj_329_groups_0 = const()[name = tensor("obj_329_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886029568)))]; + tensor layers_23_self_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_23_self_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888126784)))]; + tensor obj_329_cast_fp16 = conv(bias = layers_23_self_attn_o_proj_bias_to_fp16, dilations = obj_329_dilations_0, groups = obj_329_groups_0, pad = obj_329_pad_0, pad_type = obj_329_pad_type_0, strides = obj_329_strides_0, weight = layers_23_self_attn_o_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("obj_329_cast_fp16")]; + tensor inputs_141_cast_fp16 = add(x = inputs_139_cast_fp16, y = obj_329_cast_fp16)[name = tensor("inputs_141_cast_fp16")]; + tensor out_141_axes_0 = const()[name = tensor("out_141_axes_0"), val = tensor([1])]; + tensor var_5075_to_fp16 = const()[name = tensor("op_5075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_141_cast_fp16 = layer_norm(axes = out_141_axes_0, epsilon = var_5075_to_fp16, x = inputs_141_cast_fp16)[name = tensor("out_141_cast_fp16")]; + tensor obj_331_gamma_0_to_fp16 = const()[name = tensor("obj_331_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888128896)))]; + tensor obj_331_beta_0_to_fp16 = const()[name = tensor("obj_331_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888131008)))]; + tensor obj_331_epsilon_0_to_fp16 = const()[name = tensor("obj_331_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_331_cast_fp16 = batch_norm(beta = obj_331_beta_0_to_fp16, epsilon = obj_331_epsilon_0_to_fp16, gamma = obj_331_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_141_cast_fp16)[name = tensor("obj_331_cast_fp16")]; + tensor query_pad_type_0 = const()[name = tensor("query_pad_type_0"), val = tensor("valid")]; + tensor query_strides_0 = const()[name = tensor("query_strides_0"), val = tensor([1, 1])]; + tensor query_pad_0 = const()[name = tensor("query_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_dilations_0 = const()[name = tensor("query_dilations_0"), val = tensor([1, 1])]; + tensor query_groups_0 = const()[name = tensor("query_groups_0"), val = tensor(1)]; + tensor layers_23_encoder_attn_q_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888133120)))]; + tensor layers_23_encoder_attn_q_proj_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(890230336)))]; + tensor query_cast_fp16 = conv(bias = layers_23_encoder_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_23_encoder_attn_q_proj_weight_to_fp16, x = obj_331_cast_fp16)[name = tensor("query_cast_fp16")]; + tensor key_pad_type_0 = const()[name = tensor("key_pad_type_0"), val = tensor("valid")]; + tensor key_strides_0 = const()[name = tensor("key_strides_0"), val = tensor([1, 1])]; + tensor key_pad_0 = const()[name = tensor("key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_dilations_0 = const()[name = tensor("key_dilations_0"), val = tensor([1, 1])]; + tensor key_groups_0 = const()[name = tensor("key_groups_0"), val = tensor(1)]; + tensor layers_23_encoder_attn_k_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(890232448)))]; + tensor key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_23_encoder_attn_k_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("key_cast_fp16")]; + tensor value_pad_type_0 = const()[name = tensor("value_pad_type_0"), val = tensor("valid")]; + tensor value_strides_0 = const()[name = tensor("value_strides_0"), val = tensor([1, 1])]; + tensor value_pad_0 = const()[name = tensor("value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_dilations_0 = const()[name = tensor("value_dilations_0"), val = tensor([1, 1])]; + tensor value_groups_0 = const()[name = tensor("value_groups_0"), val = tensor(1)]; + tensor layers_23_encoder_attn_v_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(892329664)))]; + tensor layers_23_encoder_attn_v_proj_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(894426880)))]; + tensor value_cast_fp16 = conv(bias = layers_23_encoder_attn_v_proj_bias_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_23_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; + tensor var_5110 = const()[name = tensor("op_5110"), val = tensor([1, 16, 64, -1])]; + tensor mh_q_cast_fp16 = reshape(shape = var_5110, x = query_cast_fp16)[name = tensor("mh_q_cast_fp16")]; + tensor var_5112_to_fp16 = const()[name = tensor("op_5112_to_fp16"), val = tensor(0x1p-3)]; + tensor var_5113_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_5112_to_fp16)[name = tensor("op_5113_cast_fp16")]; + tensor var_5114 = const()[name = tensor("op_5114"), val = tensor([1, 16, 64, -1])]; + tensor var_5115_cast_fp16 = reshape(shape = var_5114, x = key_cast_fp16)[name = tensor("op_5115_cast_fp16")]; + tensor mh_w_transpose_x_0 = const()[name = tensor("mh_w_transpose_x_0"), val = tensor(true)]; + tensor mh_w_transpose_y_0 = const()[name = tensor("mh_w_transpose_y_0"), val = tensor(false)]; + tensor mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_5113_cast_fp16, y = var_5115_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor obj_335_cast_fp16 = softmax(axis = var_4967, x = mh_w_cast_fp16)[name = tensor("obj_335_cast_fp16")]; + tensor var_5119 = const()[name = tensor("op_5119"), val = tensor([1, 16, 64, -1])]; + tensor var_5120_cast_fp16 = reshape(shape = var_5119, x = value_cast_fp16)[name = tensor("op_5120_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(true)]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_5120_cast_fp16, y = obj_335_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_5123 = const()[name = tensor("op_5123"), val = tensor([1, 1024, 1, -1])]; + tensor input_233_cast_fp16 = reshape(shape = var_5123, x = attn_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor obj_333_pad_type_0 = const()[name = tensor("obj_333_pad_type_0"), val = tensor("valid")]; + tensor obj_333_strides_0 = const()[name = tensor("obj_333_strides_0"), val = tensor([1, 1])]; + tensor obj_333_pad_0 = const()[name = tensor("obj_333_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_333_dilations_0 = const()[name = tensor("obj_333_dilations_0"), val = tensor([1, 1])]; + tensor obj_333_groups_0 = const()[name = tensor("obj_333_groups_0"), val = tensor(1)]; + tensor layers_23_encoder_attn_o_proj_weight_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(894428992)))]; + tensor layers_23_encoder_attn_o_proj_bias_to_fp16 = const()[name = tensor("layers_23_encoder_attn_o_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(896526208)))]; + tensor obj_333_cast_fp16 = conv(bias = layers_23_encoder_attn_o_proj_bias_to_fp16, dilations = obj_333_dilations_0, groups = obj_333_groups_0, pad = obj_333_pad_0, pad_type = obj_333_pad_type_0, strides = obj_333_strides_0, weight = layers_23_encoder_attn_o_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("obj_333_cast_fp16")]; + tensor inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = obj_333_cast_fp16)[name = tensor("inputs_143_cast_fp16")]; + tensor out_143_axes_0 = const()[name = tensor("out_143_axes_0"), val = tensor([1])]; + tensor var_5144_to_fp16 = const()[name = tensor("op_5144_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_143_cast_fp16 = layer_norm(axes = out_143_axes_0, epsilon = var_5144_to_fp16, x = inputs_143_cast_fp16)[name = tensor("out_143_cast_fp16")]; + tensor input_235_gamma_0_to_fp16 = const()[name = tensor("input_235_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(896528320)))]; + tensor input_235_beta_0_to_fp16 = const()[name = tensor("input_235_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(896530432)))]; + tensor input_235_epsilon_0_to_fp16 = const()[name = tensor("input_235_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_235_cast_fp16 = batch_norm(beta = input_235_beta_0_to_fp16, epsilon = input_235_epsilon_0_to_fp16, gamma = input_235_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_143_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_pad_type_0 = const()[name = tensor("input_237_pad_type_0"), val = tensor("valid")]; + tensor input_237_strides_0 = const()[name = tensor("input_237_strides_0"), val = tensor([1, 1])]; + tensor input_237_pad_0 = const()[name = tensor("input_237_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_237_dilations_0 = const()[name = tensor("input_237_dilations_0"), val = tensor([1, 1])]; + tensor input_237_groups_0 = const()[name = tensor("input_237_groups_0"), val = tensor(1)]; + tensor layers_23_fc1_weight_to_fp16 = const()[name = tensor("layers_23_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(896532544)))]; + tensor layers_23_fc1_bias_to_fp16 = const()[name = tensor("layers_23_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(904921216)))]; + tensor input_237_cast_fp16 = conv(bias = layers_23_fc1_bias_to_fp16, dilations = input_237_dilations_0, groups = input_237_groups_0, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = input_237_strides_0, weight = layers_23_fc1_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; + tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_237_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor hidden_states_49_pad_type_0 = const()[name = tensor("hidden_states_49_pad_type_0"), val = tensor("valid")]; + tensor hidden_states_49_strides_0 = const()[name = tensor("hidden_states_49_strides_0"), val = tensor([1, 1])]; + tensor hidden_states_49_pad_0 = const()[name = tensor("hidden_states_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor hidden_states_49_dilations_0 = const()[name = tensor("hidden_states_49_dilations_0"), val = tensor([1, 1])]; + tensor hidden_states_49_groups_0 = const()[name = tensor("hidden_states_49_groups_0"), val = tensor(1)]; + tensor layers_23_fc2_weight_to_fp16 = const()[name = tensor("layers_23_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(904929472)))]; + tensor layers_23_fc2_bias_to_fp16 = const()[name = tensor("layers_23_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(913318144)))]; + tensor hidden_states_49_cast_fp16 = conv(bias = layers_23_fc2_bias_to_fp16, dilations = hidden_states_49_dilations_0, groups = hidden_states_49_groups_0, pad = hidden_states_49_pad_0, pad_type = hidden_states_49_pad_type_0, strides = hidden_states_49_strides_0, weight = layers_23_fc2_weight_to_fp16, x = input_cast_fp16)[name = tensor("hidden_states_49_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_143_cast_fp16, y = hidden_states_49_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor out_axes_0 = const()[name = tensor("out_axes_0"), val = tensor([1])]; + tensor var_5187_to_fp16 = const()[name = tensor("op_5187_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_5187_to_fp16, x = inputs_cast_fp16)[name = tensor("out_cast_fp16")]; + tensor hidden_states_gamma_0_to_fp16 = const()[name = tensor("hidden_states_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(913320256)))]; + tensor hidden_states_beta_0_to_fp16 = const()[name = tensor("hidden_states_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(913322368)))]; + tensor hidden_states_epsilon_0_to_fp16 = const()[name = tensor("hidden_states_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor("hidden_states_cast_fp16")]; + tensor var_5198_axes_0 = const()[name = tensor("op_5198_axes_0"), val = tensor([2])]; + tensor var_5198_cast_fp16 = squeeze(axes = var_5198_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_5198_cast_fp16")]; + tensor var_5201_perm_0 = const()[name = tensor("op_5201_perm_0"), val = tensor([0, 2, 1])]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(913324480)))]; + tensor var_5201_cast_fp16 = transpose(perm = var_5201_perm_0, x = var_5198_cast_fp16)[name = tensor("transpose_0")]; + tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_5201_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_5205 = const()[name = tensor("op_5205"), val = tensor(1)]; + tensor obj_339_interleave_0 = const()[name = tensor("obj_339_interleave_0"), val = tensor(false)]; + tensor key_cache_updates = concat(axis = var_5205, interleave = obj_339_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_7_cast_fp16, current_key_9_cast_fp16, current_key_11_cast_fp16, current_key_13_cast_fp16, current_key_15_cast_fp16, current_key_17_cast_fp16, current_key_19_cast_fp16, current_key_21_cast_fp16, current_key_23_cast_fp16, current_key_25_cast_fp16, current_key_27_cast_fp16, current_key_29_cast_fp16, current_key_31_cast_fp16, current_key_33_cast_fp16, current_key_35_cast_fp16, current_key_37_cast_fp16, current_key_39_cast_fp16, current_key_41_cast_fp16, current_key_43_cast_fp16, current_key_45_cast_fp16, current_key_cast_fp16))[name = tensor("obj_339_cast_fp16")]; + tensor var_5208 = const()[name = tensor("op_5208"), val = tensor(1)]; + tensor obj_341_interleave_0 = const()[name = tensor("obj_341_interleave_0"), val = tensor(false)]; + tensor value_cache_updates = concat(axis = var_5208, interleave = obj_341_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_7_cast_fp16, current_value_9_cast_fp16, current_value_11_cast_fp16, current_value_13_cast_fp16, current_value_15_cast_fp16, current_value_17_cast_fp16, current_value_19_cast_fp16, current_value_21_cast_fp16, current_value_23_cast_fp16, current_value_25_cast_fp16, current_value_27_cast_fp16, current_value_29_cast_fp16, current_value_31_cast_fp16, current_value_33_cast_fp16, current_value_35_cast_fp16, current_value_37_cast_fp16, current_value_39_cast_fp16, current_value_41_cast_fp16, current_value_43_cast_fp16, current_value_45_cast_fp16, current_value_cast_fp16))[name = tensor("obj_341_cast_fp16")]; + tensor var_5219_begin_0 = const()[name = tensor("op_5219_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_5219_end_0 = const()[name = tensor("op_5219_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5219_end_mask_0 = const()[name = tensor("op_5219_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5219_cast_fp16 = slice_by_index(begin = var_5219_begin_0, end = var_5219_end_0, end_mask = var_5219_end_mask_0, x = obj_195_cast_fp16)[name = tensor("op_5219_cast_fp16")]; + tensor var_5222_begin_0 = const()[name = tensor("op_5222_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5222_end_0 = const()[name = tensor("op_5222_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5222_end_mask_0 = const()[name = tensor("op_5222_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5222_squeeze_mask_0 = const()[name = tensor("op_5222_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5222_cast_fp16 = slice_by_index(begin = var_5222_begin_0, end = var_5222_end_0, end_mask = var_5222_end_mask_0, squeeze_mask = var_5222_squeeze_mask_0, x = var_5219_cast_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor var_5237_begin_0 = const()[name = tensor("op_5237_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_5237_end_0 = const()[name = tensor("op_5237_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_5237_end_mask_0 = const()[name = tensor("op_5237_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5237_cast_fp16 = slice_by_index(begin = var_5237_begin_0, end = var_5237_end_0, end_mask = var_5237_end_mask_0, x = obj_223_cast_fp16)[name = tensor("op_5237_cast_fp16")]; + tensor var_5240_begin_0 = const()[name = tensor("op_5240_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5240_end_0 = const()[name = tensor("op_5240_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5240_end_mask_0 = const()[name = tensor("op_5240_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5240_squeeze_mask_0 = const()[name = tensor("op_5240_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5240_cast_fp16 = slice_by_index(begin = var_5240_begin_0, end = var_5240_end_0, end_mask = var_5240_end_mask_0, squeeze_mask = var_5240_squeeze_mask_0, x = var_5237_cast_fp16)[name = tensor("op_5240_cast_fp16")]; + tensor var_5255_begin_0 = const()[name = tensor("op_5255_begin_0"), val = tensor([0, 15, 0, 0])]; + tensor var_5255_end_0 = const()[name = tensor("op_5255_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5255_end_mask_0 = const()[name = tensor("op_5255_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5255_cast_fp16 = slice_by_index(begin = var_5255_begin_0, end = var_5255_end_0, end_mask = var_5255_end_mask_0, x = obj_223_cast_fp16)[name = tensor("op_5255_cast_fp16")]; + tensor var_5258_begin_0 = const()[name = tensor("op_5258_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5258_end_0 = const()[name = tensor("op_5258_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5258_end_mask_0 = const()[name = tensor("op_5258_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5258_squeeze_mask_0 = const()[name = tensor("op_5258_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5258_cast_fp16 = slice_by_index(begin = var_5258_begin_0, end = var_5258_end_0, end_mask = var_5258_end_mask_0, squeeze_mask = var_5258_squeeze_mask_0, x = var_5255_cast_fp16)[name = tensor("op_5258_cast_fp16")]; + tensor var_5273_begin_0 = const()[name = tensor("op_5273_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor var_5273_end_0 = const()[name = tensor("op_5273_end_0"), val = tensor([1, 2, 1, 1500])]; + tensor var_5273_end_mask_0 = const()[name = tensor("op_5273_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5273_cast_fp16 = slice_by_index(begin = var_5273_begin_0, end = var_5273_end_0, end_mask = var_5273_end_mask_0, x = obj_237_cast_fp16)[name = tensor("op_5273_cast_fp16")]; + tensor var_5276_begin_0 = const()[name = tensor("op_5276_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5276_end_0 = const()[name = tensor("op_5276_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5276_end_mask_0 = const()[name = tensor("op_5276_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5276_squeeze_mask_0 = const()[name = tensor("op_5276_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5276_cast_fp16 = slice_by_index(begin = var_5276_begin_0, end = var_5276_end_0, end_mask = var_5276_end_mask_0, squeeze_mask = var_5276_squeeze_mask_0, x = var_5273_cast_fp16)[name = tensor("op_5276_cast_fp16")]; + tensor var_5291_begin_0 = const()[name = tensor("op_5291_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5291_end_0 = const()[name = tensor("op_5291_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5291_end_mask_0 = const()[name = tensor("op_5291_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5291_cast_fp16 = slice_by_index(begin = var_5291_begin_0, end = var_5291_end_0, end_mask = var_5291_end_mask_0, x = obj_293_cast_fp16)[name = tensor("op_5291_cast_fp16")]; + tensor var_5294_begin_0 = const()[name = tensor("op_5294_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5294_end_0 = const()[name = tensor("op_5294_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5294_end_mask_0 = const()[name = tensor("op_5294_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5294_squeeze_mask_0 = const()[name = tensor("op_5294_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5294_cast_fp16 = slice_by_index(begin = var_5294_begin_0, end = var_5294_end_0, end_mask = var_5294_end_mask_0, squeeze_mask = var_5294_squeeze_mask_0, x = var_5291_cast_fp16)[name = tensor("op_5294_cast_fp16")]; + tensor var_5309_begin_0 = const()[name = tensor("op_5309_begin_0"), val = tensor([0, 4, 0, 0])]; + tensor var_5309_end_0 = const()[name = tensor("op_5309_end_0"), val = tensor([1, 5, 1, 1500])]; + tensor var_5309_end_mask_0 = const()[name = tensor("op_5309_end_mask_0"), val = tensor([true, false, true, true])]; + tensor var_5309_cast_fp16 = slice_by_index(begin = var_5309_begin_0, end = var_5309_end_0, end_mask = var_5309_end_mask_0, x = obj_335_cast_fp16)[name = tensor("op_5309_cast_fp16")]; + tensor var_5312_begin_0 = const()[name = tensor("op_5312_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5312_end_0 = const()[name = tensor("op_5312_end_0"), val = tensor([1, 1, 1, 1500])]; + tensor var_5312_end_mask_0 = const()[name = tensor("op_5312_end_mask_0"), val = tensor([true, true, false, true])]; + tensor var_5312_squeeze_mask_0 = const()[name = tensor("op_5312_squeeze_mask_0"), val = tensor([false, false, true, false])]; + tensor var_5312_cast_fp16 = slice_by_index(begin = var_5312_begin_0, end = var_5312_end_0, end_mask = var_5312_end_mask_0, squeeze_mask = var_5312_squeeze_mask_0, x = var_5309_cast_fp16)[name = tensor("op_5312_cast_fp16")]; + tensor var_5319 = const()[name = tensor("op_5319"), val = tensor(1)]; + tensor var_5320_interleave_0 = const()[name = tensor("op_5320_interleave_0"), val = tensor(false)]; + tensor var_5320_cast_fp16 = concat(axis = var_5319, interleave = var_5320_interleave_0, values = (var_5222_cast_fp16, var_5240_cast_fp16, var_5258_cast_fp16, var_5276_cast_fp16, var_5294_cast_fp16, var_5312_cast_fp16))[name = tensor("op_5320_cast_fp16")]; + tensor obj_axes_0 = const()[name = tensor("obj_axes_0"), val = tensor([1])]; + tensor obj_keep_dims_0 = const()[name = tensor("obj_keep_dims_0"), val = tensor(false)]; + tensor alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = obj_keep_dims_0, x = var_5320_cast_fp16)[name = tensor("obj_cast_fp16")]; + } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); +} \ No newline at end of file