|
- # Copyright 2023 PengCheng Laboratory and Huawei Technologies Co., Ltd
- #
- # Licensed under the Apache License, Version 2.0 (the "License");
- # you may not use this file except in compliance with the License.
- # You may obtain a copy of the License at
- #
- # http://www.apache.org/licenses/LICENSE-2.0
- #
- # Unless required by applicable law or agreed to in writing, software
- # distributed under the License is distributed on an "AS IS" BASIS,
- # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- # See the License for the specific language governing permissions and
- # limitations under the License.
- # ============================================================================
- """
- PengChengMind predict run
- """
- import json
- import os
- import requests
- import datetime
- import glob
-
- import numpy as np
-
- import mindspore.common.dtype as mstype
- import mindspore.communication.management as D
- import mindspore as ms
- from mindspore import context, Tensor
- from mindspore import export
- from mindspore.context import ParallelMode
- from mindspore.parallel import set_algo_parameters
- from mindspore.parallel._cost_model_context import _set_multi_subgraphs
- from mindspore.train.model import Model
- from mindspore.train.serialization import load_distributed_checkpoint, load_checkpoint
- # from mindspore.nn.transformer.transformer import TransformerOpParallelConfig
- from mindspore.parallel._transformer.transformer import TransformerOpParallelConfig
-
- from src.pengcheng_mind import EvalNet, PengChengMindModel, EvalNet_200B
- from src.pengcheng_mind_config import set_parse, PengChengMindConfig
- from src.utils import get_args
-
- from mindspore.common import Parameter
- from mindspore.common.tensor import Tensor
- from mindspore.train.serialization import load_checkpoint, load_param_into_net
- from src.utils import download_ckpt_from_obs
-
- def restore_checkpoint(args_param, network, cache_url='/cache/Ckpt/'):
- r"""
- Load checkpoint process.
- """
- restore_ranks = D.get_rank()
- print("======start single checkpoint", flush=True)
- ckpt_name = os.path.join(cache_url, f"rank_{restore_ranks}.ckpt")
-
- if not ckpt_name:
- print(f"There is no ckpt file in {ckpt_name}, "
- f"current ckpt_files found is {ckpt_name} "
- f"with pattern {ckpt_name}, so skip the loading.")
-
- time_stamp = datetime.datetime.now()
- print(f"time stamp {time_stamp.strftime('%Y.%m.%d-%H:%M:%S')} pre trained ckpt model {ckpt_name} loading",
- flush=True)
- # Load checkpoint files latest file
- print(f'Start to load from {ckpt_name}')
- param_dict = load_checkpoint(ckpt_name)
- # for k, v in param_dict.items():
- # print("rank: ", restore_ranks, k)
- load_param_into_net(network, param_dict, strict_load=False)
-
- def set_auto_parallel_context(args_opt):
- """Set the auto parallel context"""
- rank = 0
- device_num = 1
- context.reset_auto_parallel_context()
- # context.set_auto_parallel_context(
- # strategy_ckpt_load_file=args_opt.strategy_load_ckpt_path)
- if args_opt.distribute == "true":
- D.init()
- device_num = D.get_group_size()
- rank = D.get_rank()
- print("rank_id is {}, device_num is {}".format(rank, device_num))
- context.set_auto_parallel_context(
- parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL,
- gradients_mean=False,
- full_batch=True,
- loss_repeated_mean=True,
- enable_parallel_optimizer=False,
- strategy_ckpt_save_file=f'/cache/strategy_{rank}.ckpt',
- pipeline_stages=args_opt.stage_num)
- set_algo_parameters(elementwise_op_strategy_follow=True)
- _set_multi_subgraphs()
-
- return rank, device_num
-
- def load_model(args_opt):
- r"""
- The main function for load model
- """
- context.set_context(mode=context.GRAPH_MODE)
- # Set parallel context
- rank, device_num = set_auto_parallel_context(args_opt)
-
- context.set_context(variable_memory_max_size="30GB")
- context.set_context(save_graphs=False,
- save_graphs_path="/cache/graphs_of_device_id_" + str(rank),
- device_target=args_opt.device_target)
-
- strategy_local_file = f"/cache/inference_strategy_100b_d8_mp8_dp1-{rank}.ckpt"
- ms.set_auto_parallel_context(strategy_ckpt_save_file=strategy_local_file)
-
- if args_opt.eval_task:
- use_past = False
- else:
- use_past = True if args_opt.export else (args_opt.use_past == "true")
- print('local_rank:{}, start to run...'.format(rank), flush=True)
-
- # Set model property, rewrite the model parallel
- if device_num < args_opt.op_level_model_parallel_num:
- print(f"The op_level_model_parallel_num {args_opt.op_level_model_parallel_num} is smaller than the device num,"
- f"so change it to the {device_num}", flush=True)
- args_opt.op_level_model_parallel_num = device_num
- model_parallel_num = args_opt.op_level_model_parallel_num
- data_parallel_num = int(device_num / (model_parallel_num*args_opt.stage_num))
-
- parallel_config = TransformerOpParallelConfig(data_parallel=data_parallel_num,
- model_parallel=model_parallel_num,
- pipeline_stage=args_opt.stage_num,
- micro_batch_num=args_opt.micro_size,
- vocab_emb_dp=False,
- recompute=False)
- # add sequence_parallel
- parallel_config.sequence_parallel = args_opt.sequence_parallel
- # add select_recompute
- parallel_config.select_recompute = args_opt.select_recompute
-
- per_batch_size = args_opt.per_batch_size
- batch_size = per_batch_size * data_parallel_num
- # Now only support single batch_size for predict
- if args_opt.run_type == "predict":
- batch_size = 1
-
- # download ckpt to local
- D.init()
- device_num = D.get_group_size()
- rank_id = D.get_rank()
-
- # if args_opt.softmax_compute_fp32 == "FP32":
- # softmax_compute_type = mstype.float32
- # elif args_opt.softmax_compute_fp32 == "FP16":
- # softmax_compute_type = mstype.float16
- # else:
- # raise ValueError(f"Unknown softmax_compute_fp32 {args_opt.softmax_compute_fp32}")
- #
- # if args_opt.top_query_softmax_fp32 == "FP32":
- # top_query_softmax = mstype.float32
- # elif args_opt.top_query_softmax_fp32 == "FP16":
- # top_query_softmax = mstype.float16
- # else:
- # raise ValueError(f"Unknown top_query_softmax_fp32 {args_opt.top_query_softmax_fp32}")
-
- softmax_compute_type = mstype.float16
- top_query_softmax = mstype.float16
- config = PengChengMindConfig(
- batch_size=batch_size,
- seq_length=args_opt.seq_length,
- vocab_size=args_opt.vocab_size,
- hidden_size=args_opt.embedding_size,
- num_layers=args_opt.num_layers,
- num_heads=args_opt.num_heads,
- post_layernorm_residual=False,
- dropout_rate=0.0,
- ffn_hidden_size=args_opt.embedding_size * 4,
- use_past=use_past,
- eod_reset=False,
- parallel_config=parallel_config,
- load_ckpt_path=None,
- run_type=args_opt.run_type,
- param_init_type=mstype.float32 if args_opt.param_init_type == 'fp32' else mstype.float16,
- use_rope=args_opt.use_rope,
- use_flash_attention=args_opt.use_flash_attention,
- pipeline_config_filename=args_opt.pipeline_config_filename)
-
- config.softmax_compute_fp32 = softmax_compute_type
- config.top_query_softmax_fp32 = top_query_softmax
- print("===config is: ", config, flush=True)
- print("=====args_opt is: ", args_opt, flush=True)
-
- # Define network
- pengcheng_mind = PengChengMindModel(config)
-
- # from src.pengcheng_mind import PengChengMindLossWithPrompt
- # from mindspore.nn.transformer import CrossEntropyLoss
- # loss = CrossEntropyLoss()
- # eval_net = PengChengMindLossWithPrompt(config, pengcheng_mind, loss)
- eval_net = EvalNet_200B(pengcheng_mind, pad_token=args_opt.padding_id)
- eval_net.set_train(False)
-
- # # 完整模型加载,要在构图之前
- # import time
- # time.sleep((rank % 8)*20)
- # load_checkpoint(local_ckpt_path, net=eval_net)
-
- model_predict = Model(eval_net)
- # Compile network and obtain tensor layout for loading ckpt
- inputs_np = Tensor(np.ones(shape=(config.batch_size, config.seq_length)), mstype.int32)
-
- if args_opt.distribute == "false":
- predict_layout = None
- else:
- # Compiling only needs the shape
- current_index = Tensor(np.array([0]), mstype.int32)
- predict_layout = model_predict.infer_predict_layout(inputs_np, current_index)
-
- # if args_opt.save_strategy_bucket_dir and args_opt.save_strategy_name:
- # import moxing as mox
- # mox.file.copy(src_url=strategy_local_file,
- # dst_url=args_opt.save_strategy_bucket_dir+args_opt.save_strategy_name.replace('*', str(rank_id)))
- # import moxing as mox
- # obs_graphs_path = f"obs://research-my/taoht-100b/graphs_tmp/"
- # mox.file.copy_parallel(src_url="/cache/", dst_url=obs_graphs_path)
- cache_url = '/cache/Ckpt/'
- download_ckpt_from_obs(args_opt, cache_url, rank=rank_id)
- restore_checkpoint(args_opt, eval_net, cache_url=cache_url)
- print("================load param ok=================", flush=True)
-
- return model_predict, config
-
- def run_predict(model_predict, config, args_opt):
- """run predict"""
- from src.generate import generate, generate_increment, generate_100b, generate_100b_task
- import time
-
- D.init()
- rank_id = D.get_rank()
- device_num = D.get_group_size()
-
- # generate_func = generate_increment if config.use_past else generate
- generate_func = generate_100b_task
- # Define tokenizer
- from transformers import LlamaTokenizer
- vocab_file = '/home/ma-user/modelarts/user-job-dir/PengCheng.Mind/tokenizer/llama_vocab/llama_zh_hf/tokenizer_2.model'
- tokenizer = LlamaTokenizer.from_pretrained(vocab_file)
-
- test_sample = ["中国的四大发明有哪些?",
- "请把‘We introduce Vicuna-13B, an open-source chatbot trained by fine-tuning LLaMA on user-shared conversations collected from ShareGPT. Preliminary evaluation using GPT-4 as a judge shows Vicuna-13B achieves more than 90% quality of OpenAI ChatGPT and Google Bard while outperforming other models like LLaMA and Stanford Alpaca in more than 90%* of cases. The cost of training Vicuna-13B is around $300. ‘翻译为中文:",
- "北极熊和灰熊的科学名称分别是什么?",
- "我要做蛋炒饭,请按步骤写一个详细的制作教程:",
- "中国有哪些好玩的城市?请选取其中一个城市详细介绍当地景点和游玩攻略。",]
-
- time_all = []
- time_word = []
- for idx, sample in enumerate(test_sample):
-
- # Tokenize input sentence to ids
- start_sentence = tokenizer.encode(sample, add_special_tokens=False)
- # start_sentence = tokenizer.encode(sample)
- input_ids = np.array(start_sentence).reshape(1, -1)
- time_start = time.time()
- output_ids = generate_func(model_predict,
- input_ids,
- args_opt,
- top_p=args_opt.top_p,
- top_k_num=args_opt.top_k_num,
- max_generate_length=args_opt.max_generate_length,
- duRepeate=args_opt.duRepeate)
- time_use = time.time() - time_start
- # Call inference
- if rank_id%8 == 0:
- # Decode output ids to sentence
- output_samples = tokenizer.decode(output_ids[input_ids.shape[-1]:].tolist(), skip_special_tokens=True)
- time_all.append(time_use/max(1, len(output_ids[len(input_ids):].tolist())))
- time_word.append(time_use/max(1, len(output_samples)))
- print(f"----------------------{idx}------------------------")
- print(f"Input is: {sample}\n")
- print(f'Output is: {output_samples}\n')
- print(f"Average time: {np.average(time_all[1:])} s/tokens")
-
- def main():
- """Main process for predict or export model"""
- opt = get_args(True)
- set_parse(opt)
- model_predict, config = load_model(opt)
-
- run_predict(model_predict, config, opt)
-
-
- if __name__ == "__main__":
- main()
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
|