|
- # Copyright 2021 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.
- # ============================================================================
- """
- PanGu predict run
- """
- import json
- import os
- import requests
- import datetime
- import glob
-
- import numpy as np
- from tqdm import tqdm
-
- 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, TransformerRecomputeConfig
- from mindspore.parallel._transformer.loss import CrossEntropyLoss
- from src.generate import get_scores
- from src.pangu_alpha import EvalNet, PanguAlphaModel, EvalNet_200B, EvalNet_200B_pp
- from src.pangu_alpha_config import set_parse, PanguAlphaConfig
- 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
-
- import mindspore.dataset.transforms.c_transforms as C
- import mindspore.dataset as ds
-
- 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(f"{k}: ", v.shape)
- 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,
- pipeline_stages=args_opt.stage_num)
- set_algo_parameters(elementwise_op_strategy_follow=True)
- _set_multi_subgraphs()
- if rank == 0:
- strategy_local_file = os.path.join(args_opt.strategy_save_path, f"inference_strategy_200b_mp{device_num}_dp1.ckpt")
- context.set_auto_parallel_context(strategy_ckpt_save_file=strategy_local_file)
- return rank, device_num
-
- def load_model(args_opt):
- r"""
- The main function for load model
- """
- context.set_context(mode=context.GRAPH_MODE,
- save_graphs=False,
- device_target=args_opt.device_target,
- max_device_memory="57GB")
- # Set parallel context
- rank, device_num = set_auto_parallel_context(args_opt)
- context.set_context(save_graphs_path="/cache/graphs_of_device_id_" + str(rank))
- """
- 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()
-
- softmax_compute_type = mstype.float16
- top_query_softmax = mstype.float16
- layernorm_compute_type = mstype.float16
- config = PanguAlphaConfig(
- 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=args_opt.use_past,
- eod_reset=False,
- parallel_config=parallel_config,
- load_ckpt_path=None,
- run_type=args_opt.run_type,
- param_init_type=mstype.float16,
- use_rope=args_opt.use_rope,)
-
- config.softmax_compute_fp32 = softmax_compute_type
- config.top_query_softmax_fp32 = top_query_softmax
- config.layernorm_compute_fp32 = layernorm_compute_type
- print("===config is: ", config, flush=True)
- print("=====args_opt is: ", args_opt, flush=True)
-
- # Define network
- pangu_alpha = PanguAlphaModel(config)
-
- # loss = CrossEntropyLoss()
- # eval_net = PanGUAlphaLossWithPrompt(config, pangu_alpha, loss)
- # eval_net = EvalNet_200B(pangu_alpha, pad_token=args_opt.padding_id)
- eval_net = EvalNet_200B_pp(pangu_alpha, pad_token=args_opt.padding_id, seq_length=args_opt.seq_length)
- 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)
-
- dataset = init_dataset(100, args_opt.seq_length, args_opt.vocab_size)
- model_predict.build(train_dataset=dataset, epoch=1)
-
- return model_predict, config
-
- def init_dataset(data_size=100, seq_length=4096, vocab_size=10000):
- data1 = np.array(np.random.sample(size=(data_size, seq_length)) * vocab_size, dtype=np.uint32)
- input_position = np.array(seq_length).reshape([1, -1]).repeat(data_size, axis=0)
- type_cast_op = C.TypeCast(mstype.int32)
- # 加载数据集
- dataset = ds.NumpySlicesDataset((data1, input_position), ["input_ids", "position_id"])
- dataset = dataset.map(input_columns="input_ids", operations=type_cast_op)
- dataset = dataset.map(input_columns="position_id", operations=type_cast_op)
- # batch操作
- dataset = dataset.batch(batch_size=1)
- return dataset
-
- def get_model():
- opt = get_args(True)
- set_parse(opt)
- model_predict, config = load_model(opt)
- return model_predict, config, opt
-
- def main():
- """Main process for predict or export model"""
- opt = get_args(True)
- set_parse(opt)
- load_model(opt)
- print("策略文件 dst_strategy.ckpt 导出完毕!", flush=True)
-
- if __name__ == "__main__":
- main()
-
|