diff --git "a/TextDecoder.mlmodelc/model.mil" "b/TextDecoder.mlmodelc/model.mil" deleted file mode 100644--- "a/TextDecoder.mlmodelc/model.mil" +++ /dev/null @@ -1,700 +0,0 @@ -program(1.0) -[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] -{ - 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_24_axis_0 = const()[name = tensor("op_24_axis_0"), val = tensor(0)]; - tensor var_24_batch_dims_0 = const()[name = tensor("op_24_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_24_cast_fp16 = gather(axis = var_24_axis_0, batch_dims = var_24_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_24_cast_fp16")]; - tensor var_28_axis_0 = const()[name = tensor("op_28_axis_0"), val = tensor(0)]; - tensor var_28_batch_dims_0 = const()[name = tensor("op_28_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(132777088)))]; - tensor var_28_cast_fp16 = gather(axis = var_28_axis_0, batch_dims = var_28_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor("op_28_cast_fp16")]; - tensor hidden_states_1_cast_fp16 = add(x = var_24_cast_fp16, y = var_28_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; - tensor var_42_axes_0 = const()[name = tensor("op_42_axes_0"), val = tensor([2])]; - tensor var_42_cast_fp16 = expand_dims(axes = var_42_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_42_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_42_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; - tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([1280, 1280, 1280, 1280])]; - tensor var_47_axis_0 = const()[name = tensor("op_47_axis_0"), val = tensor(1)]; - tensor var_47_cast_fp16_0, tensor var_47_cast_fp16_1, tensor var_47_cast_fp16_2, tensor var_47_cast_fp16_3 = split(axis = var_47_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_47_cast_fp16")]; - tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([1280, 1280, 1280, 1280])]; - tensor var_54_axis_0 = const()[name = tensor("op_54_axis_0"), val = tensor(1)]; - tensor var_54_cast_fp16_0, tensor var_54_cast_fp16_1, tensor var_54_cast_fp16_2, tensor var_54_cast_fp16_3 = split(axis = var_54_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_54_cast_fp16")]; - tensor var_64 = const()[name = tensor("op_64"), val = tensor(3)]; - tensor out_1_axes_0 = const()[name = tensor("out_1_axes_0"), val = tensor([1])]; - tensor var_89_to_fp16 = const()[name = tensor("op_89_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_89_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(133924032)))]; - 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(133926656)))]; - 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(133929280)))]; - 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(133931904)))]; - 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(133934528)))]; - 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(137211392)))]; - 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(137214016)))]; - 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(140490880)))]; - 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(143767744)))]; - 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_124_axes_0 = const()[name = tensor("op_124_axes_0"), val = tensor([1])]; - tensor var_124_cast_fp16 = expand_dims(axes = var_124_axes_0, x = kv_cache_update_mask)[name = tensor("op_124_cast_fp16")]; - tensor var_125_axes_0 = const()[name = tensor("op_125_axes_0"), val = tensor([2])]; - tensor var_125_cast_fp16 = expand_dims(axes = var_125_axes_0, x = var_124_cast_fp16)[name = tensor("op_125_cast_fp16")]; - tensor var_65_to_fp16 = const()[name = tensor("op_65_to_fp16"), val = tensor(0x1p+0)]; - tensor var_127_cast_fp16 = sub(x = var_65_to_fp16, y = var_125_cast_fp16)[name = tensor("op_127_cast_fp16")]; - tensor var_128_cast_fp16 = mul(x = var_47_cast_fp16_0, y = var_127_cast_fp16)[name = tensor("op_128_cast_fp16")]; - tensor var_129_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_125_cast_fp16)[name = tensor("op_129_cast_fp16")]; - tensor key_1_cast_fp16 = add(x = var_128_cast_fp16, y = var_129_cast_fp16)[name = tensor("key_1_cast_fp16")]; - tensor var_132_cast_fp16 = mul(x = var_54_cast_fp16_0, y = var_127_cast_fp16)[name = tensor("op_132_cast_fp16")]; - tensor var_133_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_125_cast_fp16)[name = tensor("op_133_cast_fp16")]; - tensor value_1_cast_fp16 = add(x = var_132_cast_fp16, y = var_133_cast_fp16)[name = tensor("value_1_cast_fp16")]; - tensor var_137 = const()[name = tensor("op_137"), val = tensor([1, 20, 64, 1])]; - tensor mh_q_1_cast_fp16 = reshape(shape = var_137, x = query_1_cast_fp16)[name = tensor("mh_q_1_cast_fp16")]; - tensor var_139_to_fp16 = const()[name = tensor("op_139_to_fp16"), val = tensor(0x1p-3)]; - tensor var_140_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_139_to_fp16)[name = tensor("op_140_cast_fp16")]; - tensor var_143 = const()[name = tensor("op_143"), val = tensor([1, 20, 64, 448])]; - tensor var_144_cast_fp16 = reshape(shape = var_143, x = key_1_cast_fp16)[name = tensor("op_144_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_140_cast_fp16, y = var_144_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; - tensor var_148_axes_0 = const()[name = tensor("op_148_axes_0"), val = tensor([1])]; - tensor var_148_cast_fp16 = expand_dims(axes = var_148_axes_0, x = decoder_key_padding_mask)[name = tensor("op_148_cast_fp16")]; - tensor var_149_axes_0 = const()[name = tensor("op_149_axes_0"), val = tensor([2])]; - tensor var_149_cast_fp16 = expand_dims(axes = var_149_axes_0, x = var_148_cast_fp16)[name = tensor("op_149_cast_fp16")]; - tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_149_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; - tensor var_152_cast_fp16 = softmax(axis = var_64, x = mh_w_3_cast_fp16)[name = tensor("op_152_cast_fp16")]; - tensor var_153 = const()[name = tensor("op_153"), val = tensor([1, 20, 64, 448])]; - tensor var_154_cast_fp16 = reshape(shape = var_153, x = value_1_cast_fp16)[name = tensor("op_154_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_154_cast_fp16, y = var_152_cast_fp16)[name = tensor("attn_1_cast_fp16")]; - tensor var_157 = const()[name = tensor("op_157"), val = tensor([1, 1280, 1, 1])]; - tensor input_1_cast_fp16 = reshape(shape = var_157, 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(143770368)))]; - 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(147047232)))]; - 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_179_to_fp16 = const()[name = tensor("op_179_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_179_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(147049856)))]; - 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(147052480)))]; - 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(147055104)))]; - 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(150331968)))]; - 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(150334592)))]; - 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(153611456)))]; - 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(156888320)))]; - 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_215 = const()[name = tensor("op_215"), val = tensor([1, 20, 64, 1])]; - tensor mh_q_3_cast_fp16 = reshape(shape = var_215, x = query_3_cast_fp16)[name = tensor("mh_q_3_cast_fp16")]; - tensor var_217_to_fp16 = const()[name = tensor("op_217_to_fp16"), val = tensor(0x1p-3)]; - tensor var_218_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_217_to_fp16)[name = tensor("op_218_cast_fp16")]; - tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 20, 64, 1500])]; - tensor var_222_cast_fp16 = reshape(shape = var_221, x = key_3_cast_fp16)[name = tensor("op_222_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_218_cast_fp16, y = var_222_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; - tensor obj_13_cast_fp16 = softmax(axis = var_64, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; - tensor var_226 = const()[name = tensor("op_226"), val = tensor([1, 20, 64, 1500])]; - tensor var_227_cast_fp16 = reshape(shape = var_226, x = value_3_cast_fp16)[name = tensor("op_227_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_227_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; - tensor var_230 = const()[name = tensor("op_230"), val = tensor([1, 1280, 1, 1])]; - tensor input_3_cast_fp16 = reshape(shape = var_230, 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(156890944)))]; - 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(160167808)))]; - 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_248_to_fp16 = const()[name = tensor("op_248_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_248_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(160170432)))]; - 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(160173056)))]; - 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(160175680)))]; - 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(173282944)))]; - 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(173293248)))]; - 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(186400512)))]; - 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_283 = const()[name = tensor("op_283"), val = tensor(3)]; - tensor out_7_axes_0 = const()[name = tensor("out_7_axes_0"), val = tensor([1])]; - tensor var_308_to_fp16 = const()[name = tensor("op_308_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_308_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(186403136)))]; - 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(186405760)))]; - 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(186408384)))]; - 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(189685248)))]; - 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(189687872)))]; - 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(192964736)))]; - 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(196241600)))]; - 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_347_cast_fp16 = mul(x = var_47_cast_fp16_1, y = var_127_cast_fp16)[name = tensor("op_347_cast_fp16")]; - tensor var_348_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_125_cast_fp16)[name = tensor("op_348_cast_fp16")]; - tensor key_5_cast_fp16 = add(x = var_347_cast_fp16, y = var_348_cast_fp16)[name = tensor("key_5_cast_fp16")]; - tensor var_351_cast_fp16 = mul(x = var_54_cast_fp16_1, y = var_127_cast_fp16)[name = tensor("op_351_cast_fp16")]; - tensor var_352_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_125_cast_fp16)[name = tensor("op_352_cast_fp16")]; - tensor value_5_cast_fp16 = add(x = var_351_cast_fp16, y = var_352_cast_fp16)[name = tensor("value_5_cast_fp16")]; - tensor var_356 = const()[name = tensor("op_356"), val = tensor([1, 20, 64, 1])]; - tensor mh_q_5_cast_fp16 = reshape(shape = var_356, x = query_5_cast_fp16)[name = tensor("mh_q_5_cast_fp16")]; - tensor var_358_to_fp16 = const()[name = tensor("op_358_to_fp16"), val = tensor(0x1p-3)]; - tensor var_359_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_358_to_fp16)[name = tensor("op_359_cast_fp16")]; - tensor var_362 = const()[name = tensor("op_362"), val = tensor([1, 20, 64, 448])]; - tensor var_363_cast_fp16 = reshape(shape = var_362, x = key_5_cast_fp16)[name = tensor("op_363_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_359_cast_fp16, y = var_363_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; - tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_149_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; - tensor var_371_cast_fp16 = softmax(axis = var_283, x = mh_w_9_cast_fp16)[name = tensor("op_371_cast_fp16")]; - tensor var_372 = const()[name = tensor("op_372"), val = tensor([1, 20, 64, 448])]; - tensor var_373_cast_fp16 = reshape(shape = var_372, x = value_5_cast_fp16)[name = tensor("op_373_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_373_cast_fp16, y = var_371_cast_fp16)[name = tensor("attn_5_cast_fp16")]; - tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 1280, 1, 1])]; - tensor input_11_cast_fp16 = reshape(shape = var_376, 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(196244224)))]; - 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(199521088)))]; - 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_398_to_fp16 = const()[name = tensor("op_398_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_398_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(199523712)))]; - 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(199526336)))]; - 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(199528960)))]; - 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(202805824)))]; - 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(202808448)))]; - 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(206085312)))]; - 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(209362176)))]; - 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_434 = const()[name = tensor("op_434"), val = tensor([1, 20, 64, 1])]; - tensor mh_q_7_cast_fp16 = reshape(shape = var_434, x = query_7_cast_fp16)[name = tensor("mh_q_7_cast_fp16")]; - tensor var_436_to_fp16 = const()[name = tensor("op_436_to_fp16"), val = tensor(0x1p-3)]; - tensor var_437_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_436_to_fp16)[name = tensor("op_437_cast_fp16")]; - tensor var_440 = const()[name = tensor("op_440"), val = tensor([1, 20, 64, 1500])]; - tensor var_441_cast_fp16 = reshape(shape = var_440, x = key_7_cast_fp16)[name = tensor("op_441_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_437_cast_fp16, y = var_441_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; - tensor obj_27_cast_fp16 = softmax(axis = var_283, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; - tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 20, 64, 1500])]; - tensor var_446_cast_fp16 = reshape(shape = var_445, x = value_7_cast_fp16)[name = tensor("op_446_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_446_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; - tensor var_449 = const()[name = tensor("op_449"), val = tensor([1, 1280, 1, 1])]; - tensor input_13_cast_fp16 = reshape(shape = var_449, 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(209364800)))]; - 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(212641664)))]; - 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_467_to_fp16 = const()[name = tensor("op_467_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_467_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(212644288)))]; - 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(212646912)))]; - 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(212649536)))]; - 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(225756800)))]; - 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(225767104)))]; - 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(238874368)))]; - 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_502 = const()[name = tensor("op_502"), val = tensor(3)]; - tensor out_13_axes_0 = const()[name = tensor("out_13_axes_0"), val = tensor([1])]; - tensor var_527_to_fp16 = const()[name = tensor("op_527_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_527_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(238876992)))]; - 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(238879616)))]; - 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(238882240)))]; - 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(242159104)))]; - 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(242161728)))]; - 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(245438592)))]; - 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(248715456)))]; - 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_566_cast_fp16 = mul(x = var_47_cast_fp16_2, y = var_127_cast_fp16)[name = tensor("op_566_cast_fp16")]; - tensor var_567_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_125_cast_fp16)[name = tensor("op_567_cast_fp16")]; - tensor key_9_cast_fp16 = add(x = var_566_cast_fp16, y = var_567_cast_fp16)[name = tensor("key_9_cast_fp16")]; - tensor var_570_cast_fp16 = mul(x = var_54_cast_fp16_2, y = var_127_cast_fp16)[name = tensor("op_570_cast_fp16")]; - tensor var_571_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_125_cast_fp16)[name = tensor("op_571_cast_fp16")]; - tensor value_9_cast_fp16 = add(x = var_570_cast_fp16, y = var_571_cast_fp16)[name = tensor("value_9_cast_fp16")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 20, 64, 1])]; - tensor mh_q_9_cast_fp16 = reshape(shape = var_575, x = query_9_cast_fp16)[name = tensor("mh_q_9_cast_fp16")]; - tensor var_577_to_fp16 = const()[name = tensor("op_577_to_fp16"), val = tensor(0x1p-3)]; - tensor var_578_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_577_to_fp16)[name = tensor("op_578_cast_fp16")]; - tensor var_581 = const()[name = tensor("op_581"), val = tensor([1, 20, 64, 448])]; - tensor var_582_cast_fp16 = reshape(shape = var_581, x = key_9_cast_fp16)[name = tensor("op_582_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_578_cast_fp16, y = var_582_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; - tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_149_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; - tensor var_590_cast_fp16 = softmax(axis = var_502, x = mh_w_15_cast_fp16)[name = tensor("op_590_cast_fp16")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 20, 64, 448])]; - tensor var_592_cast_fp16 = reshape(shape = var_591, x = value_9_cast_fp16)[name = tensor("op_592_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_592_cast_fp16, y = var_590_cast_fp16)[name = tensor("attn_9_cast_fp16")]; - tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1280, 1, 1])]; - tensor input_21_cast_fp16 = reshape(shape = var_595, 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(248718080)))]; - 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(251994944)))]; - 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_617_to_fp16 = const()[name = tensor("op_617_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_617_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(251997568)))]; - 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(252000192)))]; - 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(252002816)))]; - 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(255279680)))]; - 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(255282304)))]; - 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(258559168)))]; - 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(261836032)))]; - 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_653 = const()[name = tensor("op_653"), val = tensor([1, 20, 64, 1])]; - tensor mh_q_11_cast_fp16 = reshape(shape = var_653, x = query_11_cast_fp16)[name = tensor("mh_q_11_cast_fp16")]; - tensor var_655_to_fp16 = const()[name = tensor("op_655_to_fp16"), val = tensor(0x1p-3)]; - tensor var_656_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_655_to_fp16)[name = tensor("op_656_cast_fp16")]; - tensor var_659 = const()[name = tensor("op_659"), val = tensor([1, 20, 64, 1500])]; - tensor var_660_cast_fp16 = reshape(shape = var_659, x = key_11_cast_fp16)[name = tensor("op_660_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_656_cast_fp16, y = var_660_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; - tensor obj_41_cast_fp16 = softmax(axis = var_502, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; - tensor var_664 = const()[name = tensor("op_664"), val = tensor([1, 20, 64, 1500])]; - tensor var_665_cast_fp16 = reshape(shape = var_664, x = value_11_cast_fp16)[name = tensor("op_665_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_665_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; - tensor var_668 = const()[name = tensor("op_668"), val = tensor([1, 1280, 1, 1])]; - tensor input_23_cast_fp16 = reshape(shape = var_668, 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(261838656)))]; - 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(265115520)))]; - 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_689_to_fp16 = const()[name = tensor("op_689_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_689_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(265118144)))]; - 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(265120768)))]; - 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(265123392)))]; - 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(278230656)))]; - 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(278240960)))]; - 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(291348224)))]; - 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_725 = const()[name = tensor("op_725"), val = tensor(3)]; - tensor out_19_axes_0 = const()[name = tensor("out_19_axes_0"), val = tensor([1])]; - tensor var_750_to_fp16 = const()[name = tensor("op_750_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_750_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(291350848)))]; - 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(291353472)))]; - 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(291356096)))]; - 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(294632960)))]; - 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_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_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(294635584)))]; - 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_3_self_attn_k_proj_weight_to_fp16, x = obj_43_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_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(297912448)))]; - 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(301189312)))]; - tensor current_value_cast_fp16 = conv(bias = layers_3_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_3_self_attn_v_proj_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("current_value_cast_fp16")]; - tensor var_789_cast_fp16 = mul(x = var_47_cast_fp16_3, y = var_127_cast_fp16)[name = tensor("op_789_cast_fp16")]; - tensor var_790_cast_fp16 = mul(x = current_key_cast_fp16, y = var_125_cast_fp16)[name = tensor("op_790_cast_fp16")]; - tensor key_13_cast_fp16 = add(x = var_789_cast_fp16, y = var_790_cast_fp16)[name = tensor("key_13_cast_fp16")]; - tensor var_793_cast_fp16 = mul(x = var_54_cast_fp16_3, y = var_127_cast_fp16)[name = tensor("op_793_cast_fp16")]; - tensor var_794_cast_fp16 = mul(x = current_value_cast_fp16, y = var_125_cast_fp16)[name = tensor("op_794_cast_fp16")]; - tensor value_13_cast_fp16 = add(x = var_793_cast_fp16, y = var_794_cast_fp16)[name = tensor("value_13_cast_fp16")]; - tensor var_798 = const()[name = tensor("op_798"), val = tensor([1, 20, 64, 1])]; - tensor mh_q_13_cast_fp16 = reshape(shape = var_798, x = query_13_cast_fp16)[name = tensor("mh_q_13_cast_fp16")]; - tensor var_800_to_fp16 = const()[name = tensor("op_800_to_fp16"), val = tensor(0x1p-3)]; - tensor var_801_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_800_to_fp16)[name = tensor("op_801_cast_fp16")]; - tensor var_804 = const()[name = tensor("op_804"), val = tensor([1, 20, 64, 448])]; - tensor var_805_cast_fp16 = reshape(shape = var_804, x = key_13_cast_fp16)[name = tensor("op_805_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_801_cast_fp16, y = var_805_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; - tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_149_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; - tensor var_813_cast_fp16 = softmax(axis = var_725, x = mh_w_21_cast_fp16)[name = tensor("op_813_cast_fp16")]; - tensor var_814 = const()[name = tensor("op_814"), val = tensor([1, 20, 64, 448])]; - tensor var_815_cast_fp16 = reshape(shape = var_814, x = value_13_cast_fp16)[name = tensor("op_815_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_815_cast_fp16, y = var_813_cast_fp16)[name = tensor("attn_13_cast_fp16")]; - tensor var_818 = const()[name = tensor("op_818"), val = tensor([1, 1280, 1, 1])]; - tensor input_31_cast_fp16 = reshape(shape = var_818, 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(301191936)))]; - 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(304468800)))]; - 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_840_to_fp16 = const()[name = tensor("op_840_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_840_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(304471424)))]; - 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(304474048)))]; - 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_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_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(304476672)))]; - 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(307753536)))]; - tensor query_cast_fp16 = conv(bias = layers_3_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_3_encoder_attn_q_proj_weight_to_fp16, x = obj_51_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_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(307756160)))]; - 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_3_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_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(311033024)))]; - 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(314309888)))]; - tensor value_cast_fp16 = conv(bias = layers_3_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_3_encoder_attn_v_proj_weight_to_fp16, x = encoder_output_embeds)[name = tensor("value_cast_fp16")]; - tensor var_876 = const()[name = tensor("op_876"), val = tensor([1, 20, 64, 1])]; - tensor mh_q_cast_fp16 = reshape(shape = var_876, x = query_cast_fp16)[name = tensor("mh_q_cast_fp16")]; - tensor var_878_to_fp16 = const()[name = tensor("op_878_to_fp16"), val = tensor(0x1p-3)]; - tensor var_879_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_878_to_fp16)[name = tensor("op_879_cast_fp16")]; - tensor var_882 = const()[name = tensor("op_882"), val = tensor([1, 20, 64, 1500])]; - tensor var_883_cast_fp16 = reshape(shape = var_882, x = key_cast_fp16)[name = tensor("op_883_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_879_cast_fp16, y = var_883_cast_fp16)[name = tensor("mh_w_cast_fp16")]; - tensor obj_55_cast_fp16 = softmax(axis = var_725, x = mh_w_cast_fp16)[name = tensor("obj_55_cast_fp16")]; - tensor var_887 = const()[name = tensor("op_887"), val = tensor([1, 20, 64, 1500])]; - tensor var_888_cast_fp16 = reshape(shape = var_887, x = value_cast_fp16)[name = tensor("op_888_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_888_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_cast_fp16")]; - tensor var_891 = const()[name = tensor("op_891"), val = tensor([1, 1280, 1, 1])]; - tensor input_33_cast_fp16 = reshape(shape = var_891, x = attn_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(314312512)))]; - 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(317589376)))]; - 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_912_to_fp16 = const()[name = tensor("op_912_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_912_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(317592000)))]; - 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(317594624)))]; - 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(317597248)))]; - 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(330704512)))]; - 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_mode_0 = const()[name = tensor("input_mode_0"), val = tensor("EXACT")]; - tensor input_cast_fp16 = gelu(mode = input_mode_0, x = input_37_cast_fp16)[name = tensor("input_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(330714816)))]; - 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(343822080)))]; - 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_cast_fp16)[name = tensor("hidden_states_9_cast_fp16")]; - tensor inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_cast_fp16")]; - tensor out_axes_0 = const()[name = tensor("out_axes_0"), val = tensor([1])]; - tensor var_955_to_fp16 = const()[name = tensor("op_955_to_fp16"), val = tensor(0x1.5p-17)]; - tensor out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_955_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(343824704)))]; - 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(343827328)))]; - 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_966_axes_0 = const()[name = tensor("op_966_axes_0"), val = tensor([2])]; - tensor var_966_cast_fp16 = squeeze(axes = var_966_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_966_cast_fp16")]; - tensor var_969_perm_0 = const()[name = tensor("op_969_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(343829952)))]; - tensor var_969_cast_fp16 = transpose(perm = var_969_perm_0, x = var_966_cast_fp16)[name = tensor("transpose_0")]; - tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_969_cast_fp16)[name = tensor("linear_0_cast_fp16")]; - tensor var_973 = const()[name = tensor("op_973"), val = tensor(1)]; - tensor obj_59_interleave_0 = const()[name = tensor("obj_59_interleave_0"), val = tensor(false)]; - tensor key_cache_updates = concat(axis = var_973, interleave = obj_59_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_cast_fp16))[name = tensor("obj_59_cast_fp16")]; - tensor var_976 = const()[name = tensor("op_976"), val = tensor(1)]; - tensor obj_61_interleave_0 = const()[name = tensor("obj_61_interleave_0"), val = tensor(false)]; - tensor value_cache_updates = concat(axis = var_976, interleave = obj_61_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_cast_fp16))[name = tensor("obj_61_cast_fp16")]; - tensor var_987_begin_0 = const()[name = tensor("op_987_begin_0"), val = tensor([0, 4, 0, 0])]; - tensor var_987_end_0 = const()[name = tensor("op_987_end_0"), val = tensor([1, 5, 1, 1500])]; - tensor var_987_end_mask_0 = const()[name = tensor("op_987_end_mask_0"), val = tensor([true, false, true, true])]; - tensor var_987_cast_fp16 = slice_by_index(begin = var_987_begin_0, end = var_987_end_0, end_mask = var_987_end_mask_0, x = obj_41_cast_fp16)[name = tensor("op_987_cast_fp16")]; - tensor var_990_begin_0 = const()[name = tensor("op_990_begin_0"), val = tensor([0, 0, 0, 0])]; - tensor var_990_end_0 = const()[name = tensor("op_990_end_0"), val = tensor([1, 1, 1, 1500])]; - tensor var_990_end_mask_0 = const()[name = tensor("op_990_end_mask_0"), val = tensor([true, true, false, true])]; - tensor var_990_squeeze_mask_0 = const()[name = tensor("op_990_squeeze_mask_0"), val = tensor([false, false, true, false])]; - tensor var_990_cast_fp16 = slice_by_index(begin = var_990_begin_0, end = var_990_end_0, end_mask = var_990_end_mask_0, squeeze_mask = var_990_squeeze_mask_0, x = var_987_cast_fp16)[name = tensor("op_990_cast_fp16")]; - tensor var_1005_begin_0 = const()[name = tensor("op_1005_begin_0"), val = tensor([0, 11, 0, 0])]; - tensor var_1005_end_0 = const()[name = tensor("op_1005_end_0"), val = tensor([1, 12, 1, 1500])]; - tensor var_1005_end_mask_0 = const()[name = tensor("op_1005_end_mask_0"), val = tensor([true, false, true, true])]; - tensor var_1005_cast_fp16 = slice_by_index(begin = var_1005_begin_0, end = var_1005_end_0, end_mask = var_1005_end_mask_0, x = obj_41_cast_fp16)[name = tensor("op_1005_cast_fp16")]; - tensor var_1008_begin_0 = const()[name = tensor("op_1008_begin_0"), val = tensor([0, 0, 0, 0])]; - tensor var_1008_end_0 = const()[name = tensor("op_1008_end_0"), val = tensor([1, 1, 1, 1500])]; - tensor var_1008_end_mask_0 = const()[name = tensor("op_1008_end_mask_0"), val = tensor([true, true, false, true])]; - tensor var_1008_squeeze_mask_0 = const()[name = tensor("op_1008_squeeze_mask_0"), val = tensor([false, false, true, false])]; - tensor var_1008_cast_fp16 = slice_by_index(begin = var_1008_begin_0, end = var_1008_end_0, end_mask = var_1008_end_mask_0, squeeze_mask = var_1008_squeeze_mask_0, x = var_1005_cast_fp16)[name = tensor("op_1008_cast_fp16")]; - tensor var_1023_begin_0 = const()[name = tensor("op_1023_begin_0"), val = tensor([0, 3, 0, 0])]; - tensor var_1023_end_0 = const()[name = tensor("op_1023_end_0"), val = tensor([1, 4, 1, 1500])]; - tensor var_1023_end_mask_0 = const()[name = tensor("op_1023_end_mask_0"), val = tensor([true, false, true, true])]; - tensor var_1023_cast_fp16 = slice_by_index(begin = var_1023_begin_0, end = var_1023_end_0, end_mask = var_1023_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1023_cast_fp16")]; - tensor var_1026_begin_0 = const()[name = tensor("op_1026_begin_0"), val = tensor([0, 0, 0, 0])]; - tensor var_1026_end_0 = const()[name = tensor("op_1026_end_0"), val = tensor([1, 1, 1, 1500])]; - tensor var_1026_end_mask_0 = const()[name = tensor("op_1026_end_mask_0"), val = tensor([true, true, false, true])]; - tensor var_1026_squeeze_mask_0 = const()[name = tensor("op_1026_squeeze_mask_0"), val = tensor([false, false, true, false])]; - tensor var_1026_cast_fp16 = slice_by_index(begin = var_1026_begin_0, end = var_1026_end_0, end_mask = var_1026_end_mask_0, squeeze_mask = var_1026_squeeze_mask_0, x = var_1023_cast_fp16)[name = tensor("op_1026_cast_fp16")]; - tensor var_1041_begin_0 = const()[name = tensor("op_1041_begin_0"), val = tensor([0, 6, 0, 0])]; - tensor var_1041_end_0 = const()[name = tensor("op_1041_end_0"), val = tensor([1, 7, 1, 1500])]; - tensor var_1041_end_mask_0 = const()[name = tensor("op_1041_end_mask_0"), val = tensor([true, false, true, true])]; - tensor var_1041_cast_fp16 = slice_by_index(begin = var_1041_begin_0, end = var_1041_end_0, end_mask = var_1041_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1041_cast_fp16")]; - tensor var_1044_begin_0 = const()[name = tensor("op_1044_begin_0"), val = tensor([0, 0, 0, 0])]; - tensor var_1044_end_0 = const()[name = tensor("op_1044_end_0"), val = tensor([1, 1, 1, 1500])]; - tensor var_1044_end_mask_0 = const()[name = tensor("op_1044_end_mask_0"), val = tensor([true, true, false, true])]; - tensor var_1044_squeeze_mask_0 = const()[name = tensor("op_1044_squeeze_mask_0"), val = tensor([false, false, true, false])]; - tensor var_1044_cast_fp16 = slice_by_index(begin = var_1044_begin_0, end = var_1044_end_0, end_mask = var_1044_end_mask_0, squeeze_mask = var_1044_squeeze_mask_0, x = var_1041_cast_fp16)[name = tensor("op_1044_cast_fp16")]; - tensor var_1059_begin_0 = const()[name = tensor("op_1059_begin_0"), val = tensor([0, 11, 0, 0])]; - tensor var_1059_end_0 = const()[name = tensor("op_1059_end_0"), val = tensor([1, 12, 1, 1500])]; - tensor var_1059_end_mask_0 = const()[name = tensor("op_1059_end_mask_0"), val = tensor([true, false, true, true])]; - tensor var_1059_cast_fp16 = slice_by_index(begin = var_1059_begin_0, end = var_1059_end_0, end_mask = var_1059_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1059_cast_fp16")]; - tensor var_1062_begin_0 = const()[name = tensor("op_1062_begin_0"), val = tensor([0, 0, 0, 0])]; - tensor var_1062_end_0 = const()[name = tensor("op_1062_end_0"), val = tensor([1, 1, 1, 1500])]; - tensor var_1062_end_mask_0 = const()[name = tensor("op_1062_end_mask_0"), val = tensor([true, true, false, true])]; - tensor var_1062_squeeze_mask_0 = const()[name = tensor("op_1062_squeeze_mask_0"), val = tensor([false, false, true, false])]; - tensor var_1062_cast_fp16 = slice_by_index(begin = var_1062_begin_0, end = var_1062_end_0, end_mask = var_1062_end_mask_0, squeeze_mask = var_1062_squeeze_mask_0, x = var_1059_cast_fp16)[name = tensor("op_1062_cast_fp16")]; - tensor var_1077_begin_0 = const()[name = tensor("op_1077_begin_0"), val = tensor([0, 14, 0, 0])]; - tensor var_1077_end_0 = const()[name = tensor("op_1077_end_0"), val = tensor([1, 15, 1, 1500])]; - tensor var_1077_end_mask_0 = const()[name = tensor("op_1077_end_mask_0"), val = tensor([true, false, true, true])]; - tensor var_1077_cast_fp16 = slice_by_index(begin = var_1077_begin_0, end = var_1077_end_0, end_mask = var_1077_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1077_cast_fp16")]; - tensor var_1080_begin_0 = const()[name = tensor("op_1080_begin_0"), val = tensor([0, 0, 0, 0])]; - tensor var_1080_end_0 = const()[name = tensor("op_1080_end_0"), val = tensor([1, 1, 1, 1500])]; - tensor var_1080_end_mask_0 = const()[name = tensor("op_1080_end_mask_0"), val = tensor([true, true, false, true])]; - tensor var_1080_squeeze_mask_0 = const()[name = tensor("op_1080_squeeze_mask_0"), val = tensor([false, false, true, false])]; - tensor var_1080_cast_fp16 = slice_by_index(begin = var_1080_begin_0, end = var_1080_end_0, end_mask = var_1080_end_mask_0, squeeze_mask = var_1080_squeeze_mask_0, x = var_1077_cast_fp16)[name = tensor("op_1080_cast_fp16")]; - tensor var_1087 = const()[name = tensor("op_1087"), val = tensor(1)]; - tensor var_1088_interleave_0 = const()[name = tensor("op_1088_interleave_0"), val = tensor(false)]; - tensor var_1088_cast_fp16 = concat(axis = var_1087, interleave = var_1088_interleave_0, values = (var_990_cast_fp16, var_1008_cast_fp16, var_1026_cast_fp16, var_1044_cast_fp16, var_1062_cast_fp16, var_1080_cast_fp16))[name = tensor("op_1088_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_1088_cast_fp16)[name = tensor("obj_cast_fp16")]; - } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); -} \ No newline at end of file