|
- # Copyright 2022 PCL
- #
- # 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.
- # ============================================================================
- """
- PanguAlpha train script
- """
-
- import datetime
- import json
- import glob
- import os
- import math, time
-
- from mindspore import context
- from mindspore.train.model import Model
- import mindspore.communication.management as D
- from mindspore.context import ParallelMode
- import mindspore.nn as nn
- from mindspore.train.callback import TimeMonitor
- from mindspore.nn.wrap.loss_scale import DynamicLossScaleUpdateCell
- import mindspore.common.dtype as mstype
- from mindspore.parallel import set_algo_parameters
- from mindspore.parallel._cost_model_context import _set_multi_subgraphs
- from mindspore.nn.wrap.cell_wrapper import PipelineCell, _VirtualDatasetCell, MicroBatchInterleaved
- from mindspore.nn.transformer import TransformerOpParallelConfig, CrossEntropyLoss, TransformerRecomputeConfig
- from mindspore.train.callback import ModelCheckpoint, CheckpointConfig
- from mindspore.train.serialization import load_distributed_checkpoint, load_checkpoint, load_param_into_net
-
- from src.adam import AdamWeightDecayOp
- from src.dataset import create_dataset_aisyn as create_dataset
- from src.pangu_alpha import PanGUAlphaWithLoss, PanguAlphaModel
- from src.pangu_alpha_wrapcell import PanguAlphaTrainOneStepWithLossScaleCell, PanguAlphaTrainPipelineWithLossScaleCell
- from src.pangu_alpha_config import set_parse, PanguAlphaConfig
- from src.utils import LearningRate, get_args, FP32StateAdamWeightDecay
- from src.utils import download_data
- from src.callbacks import EvalCallBack, LossCallBack
- from src.metrics import PPLMetric
-
- import AISyncore as asc
- import numpy as np
- from typing import Dict, List
- from mindspore import save_checkpoint
- from mindspore.common import Parameter
- import mindspore.common.tensor as Tensor
- from mindspore.ops import Print
-
- project_root = os.path.abspath(
- os.path.dirname(os.path.realpath(__file__)) + os.path.sep + "..")
- print('project_root:', project_root)
-
-
- def set_weight_decay(params):
- """
- Set weight decay coefficient, zero for bias and layernorm, 1e-1 for rest
- """
- decay_filter = lambda x: 'layernorm' not in x.name.lower() and "bias" not in x.name.lower()
- decay_params = list(filter(decay_filter, params))
- other_params = list(filter(lambda x: not decay_filter(x), params))
- group_params = [{
- 'params': decay_params,
- 'weight_decay': 1e-1
- }, {
- 'params': other_params,
- 'weight_decay': 0.0
- }, {
- 'order_params': params
- }]
- return group_params
-
-
- def add_checkpoint_callback_policy(args_param, callback, rank_id):
- r"""
- Add checkpoint policy to callback.
- """
- # if args_param.save_checkpoint:
- # # checkpoint store epoch_num and step_num info
- # ckpt_append_info = [{"epoch_num": args_param.has_trained_epoches, "step_num": args_param.has_trained_steps, "global_step": 0}]
- # ckpt_config = CheckpointConfig(save_checkpoint_steps=args_param.save_checkpoint_steps,
- # keep_checkpoint_max=args_param.keep_checkpoint_max,
- # integrated_save=False,
- # append_info=ckpt_append_info
- # )
-
- # # save checkpoint into rank directory
- # ckpoint_cb = ModelCheckpoint(prefix=args_param.ckpt_name_prefix + str(rank_id),
- # directory=os.path.join(args_param.save_checkpoint_path, f"rank_{rank_id}"),
- # config=ckpt_config)
-
- # callback.append(ckpoint_cb)
- if rank_id == 0:
- ckpt_append_info = [
- {"epoch_num": args_param.has_trained_epoches, "step_num": args_param.has_trained_steps, "global_step": 0}]
- ckpt_config = CheckpointConfig(save_checkpoint_steps=1000000,
- keep_checkpoint_max=1,
- integrated_save=True,
- append_info=ckpt_append_info
- )
-
- # save checkpoint into rank directory
- ckpoint_cb = ModelCheckpoint(prefix=args_param.ckpt_name_prefix + str(rank_id),
- directory="/cache/ckpt/",
- config=ckpt_config)
-
- callback.append(ckpoint_cb)
-
-
- def set_parallel_context(args_opt):
- r"""Set parallel context"""
- from mindspore.context import ParallelMode
- D.init()
- device_num = D.get_group_size()
- rank = D.get_rank()
- print("rank_id is {}, device_num is {}".format(rank, device_num))
- context.reset_auto_parallel_context()
- if args_opt.mode == "350M":
- if device_num == 1:
- context.set_auto_parallel_context(
- parallel_mode=ParallelMode.stand_alone,
- gradients_mean=False,
- full_batch=bool(args_opt.full_batch),
- enable_parallel_optimizer=False,
- enable_alltoall=False)
- else:
- context.set_auto_parallel_context(
- parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL,
- gradients_mean=False,
- full_batch=bool(args_opt.full_batch),
- enable_parallel_optimizer=False,
- enable_alltoall=False)
- else:
- context.set_auto_parallel_context(
- parallel_mode=args_opt.parallel_mode,
- gradients_mean=False,
- full_batch=bool(args_opt.full_batch),
- strategy_ckpt_load_file=args_opt.strategy_load_ckpt_path,
- enable_parallel_optimizer=bool(args_opt.optimizer_shard),
- strategy_ckpt_save_file='strategy.ckpt',
- enable_alltoall=bool(args_opt.enable_alltoall))
- set_algo_parameters(elementwise_op_strategy_follow=True)
- _set_multi_subgraphs()
- return rank, device_num
-
-
- def set_optimizer(optimizer, opt_offload, group_params, learning_rate, config):
- r"""Set optimizer"""
- if optimizer == "lamb":
- optimizer = nn.Lamb(group_params, learning_rate=learning_rate)
- elif opt_offload:
- optimizer = AdamWeightDecayOp(group_params, learning_rate=learning_rate, eps=1e-8, beta1=0.9, beta2=0.95,
- param_init_type=config.param_init_type)
- else:
- optimizer = FP32StateAdamWeightDecay(group_params, learning_rate=learning_rate, eps=1e-8, beta1=0.9, beta2=0.95)
- return optimizer
-
-
- def load_train_net(args_opt):
- r"""The main training process."""
- # Set execution mode
- context.set_context(mode=context.GRAPH_MODE,
- device_target=args_opt.device_target,
- variable_memory_max_size="30GB")
- # Set parallel context
- rank = 0
- device_num = 1
- if args_opt.distribute == "true":
- rank, device_num = set_parallel_context(args_opt)
- context.set_context(save_graphs=False, save_graphs_path="./graphs_of_device_id_" + str(rank))
- if args_opt.parallel_mode == "data_parallel":
- # in avoid of the loop call depth
- context.set_context(max_call_depth=10000)
-
- # env variable prepare
- group_info_file = os.getenv("GROUP_INFO_FILE")
- if group_info_file:
- with open(os.path.expanduser("job/code/group_info_env"), "a") as outfile:
- outfile.write(f"export GROUP_INFO_FILE_REFLECT={group_info_file}\n")
-
- # copy data from the cloud to the /cache/Data
- cache_url = '/cache/Data/'
- eval_cache_url = '/cache/EvalData/'
- if args_opt.offline:
- cache_url = args_opt.data_url
- eval_cache_url = args_opt.eval_data_url
- else:
- download_data(src_data_url=args_opt.data_url, tgt_data_path=cache_url, rank=rank)
- download_data(src_data_url=args_opt.eval_data_url, tgt_data_path=eval_cache_url, rank=rank)
- # Set model property
- model_parallel_num = args_opt.op_level_model_parallel_num
- data_parallel_num = int(device_num / model_parallel_num)
- batch_size = args_opt.per_batch_size * data_parallel_num if context.get_auto_parallel_context(
- "parallel_mode") != ParallelMode.DATA_PARALLEL else args_opt.per_batch_size
- micro_batch_interleaved = args_opt.micro_batch_interleaved
- recompute_config = TransformerRecomputeConfig(recompute=True,
- recompute_slice_activation=bool(args_opt.recompute_slice_activation))
- parallel_config = TransformerOpParallelConfig(data_parallel=data_parallel_num,
- model_parallel=model_parallel_num,
- expert_parallel=args_opt.expert_parallel_num,
- pipeline_stage=args_opt.stage_num,
- micro_batch_num=args_opt.micro_size,
- optimizer_shard=bool(args_opt.optimizer_shard),
- vocab_emb_dp=bool(args_opt.word_emb_dp), recompute=recompute_config,
- gradient_aggregation_group=args_opt.gradient_aggregation_group)
- config = PanguAlphaConfig(batch_size=batch_size // micro_batch_interleaved,
- num_heads=args_opt.num_heads,
- hidden_size=args_opt.embedding_size,
- seq_length=args_opt.seq_length,
- vocab_size=args_opt.vocab_size,
- num_layers=args_opt.num_layers,
- eod_token=args_opt.eod_id,
- ffn_hidden_size=args_opt.embedding_size * 4,
- eod_reset=bool(args_opt.eod_reset),
- load_ckpt_path=args_opt.load_ckpt_path,
- expert_num=args_opt.expert_num,
- param_init_type=mstype.float32 if args_opt.param_init_type == 'fp32' else mstype.float16,
- enable_offload=bool(args_opt.opt_offload),
- use_moe=bool(args_opt.use_moe),
- per_token_num_experts_chosen=args_opt.per_token_num_experts_chosen,
- hidden_act='fast_gelu' if args_opt.device_target != "GPU" else 'gelu',
- parallel_config=parallel_config)
- print("===config is: ", config, flush=True)
- # Define network
- pangu_alpha = PanguAlphaModel(config=config)
- loss = CrossEntropyLoss(config.parallel_config.dp_mp_config)
- pangu_alpha_with_loss_net = MicroBatchInterleaved(PanGUAlphaWithLoss(config, pangu_alpha, loss),
- micro_batch_interleaved)
- pangu_alpha_with_loss = _VirtualDatasetCell(pangu_alpha_with_loss_net)
- print("=====args_opt is: ", args_opt, flush=True)
- # Warm-up and cosine decay learning rate
- lr = LearningRate(learning_rate=args_opt.start_lr,
- end_learning_rate=args_opt.end_lr,
- warmup_steps=args_opt.warmup_step,
- decay_steps=20000)
- params = pangu_alpha_with_loss.trainable_params()
- group_params = set_weight_decay(params)
- optimizer = set_optimizer(args_opt.optimizer, args_opt.opt_offload, group_params=group_params,
- learning_rate=lr, config=config)
- epoch_num = args_opt.epoch_size
- # Dataset loading mindrecord files
- # ds = create_dataset(config.batch_size * micro_batch_interleaved, data_path=cache_url, data_start_index=0,
- # eod_reset=config.eod_reset, full_batch=bool(args_opt.full_batch), eod_id=args_opt.eod_id,
- # device_num=device_num, rank=rank, column_name=args_opt.data_column_name, epoch=epoch_num)
- # step_per_epoch = ds.get_dataset_size()
- # actual_epoch_num = int(epoch_num * step_per_epoch / args_opt.sink_size)
-
- # loss_callback = LossCallBack(step_per_epoch, rank, 0, 0, micro_size=micro_batch_interleaved)
- loss_callback = LossCallBack(args_opt.sink_size, rank, 0, 0, micro_size=micro_batch_interleaved)
- callback = [TimeMonitor(args_opt.sink_size), loss_callback]
-
- update_cell = DynamicLossScaleUpdateCell(loss_scale_value=math.pow(2, 24), scale_factor=2, scale_window=1000)
- pangu_alpha_with_grads = PanguAlphaTrainOneStepWithLossScaleCell(
- pangu_alpha_with_loss, optimizer=optimizer, scale_update_cell=update_cell, enable_global_norm=True,
- config=config)
- if args_opt.train_and_eval_mode:
- ds_eval = create_dataset(config.batch_size * micro_batch_interleaved,
- data_path=eval_cache_url,
- data_start_index=0,
- eod_reset=config.eod_reset,
- full_batch=bool(args_opt.full_batch),
- eod_id=args_opt.eod_id,
- device_num=device_num,
- rank=rank,
- column_name=args_opt.data_column_name,
- epoch=epoch_num,
- num_samples=args_opt.eval_steps * config.batch_size)
- ppl_metric = PPLMetric(config.seq_length)
- model = Model(pangu_alpha_with_grads, eval_network=pangu_alpha_with_loss, metrics={"ppl": ppl_metric})
- callback.append(EvalCallBack(model, ds_eval, ppl_metric))
- else:
- model = Model(pangu_alpha_with_grads)
-
- add_checkpoint_callback_policy(args_opt, callback, rank)
- return pangu_alpha, args_opt, model, callback, config
-
-
- def run_train(args_opt):
- pangu_alpha, args_opt, model, ds, callback = load_train_net(args_opt)
- step_per_epoch = ds.get_dataset_size()
- actual_epoch_num = int(args_opt.epoch_size * step_per_epoch / args_opt.sink_size)
- print("===dataset size: ", ds.get_dataset_size, flush=True)
- print("===actual_epoch_num: ", actual_epoch_num, flush=True)
- print("dataset size:{}, actual_epoch_num:{} ".format(ds.get_dataset_size, actual_epoch_num), flush=True)
-
- if args_opt.incremental_training:
- strategy = model.infer_train_layout(train_dataset=ds, sink_size=args_opt.sink_size)
- print("======start load_distributed checkpoint", flush=True)
- # For 2.6B and 13B models, the number of ckpt files is 512.
- ckpt_file_list = [os.path.join(args_opt.load_ckpt_path, f"filerted_{ckpt_rank}.ckpt") for ckpt_rank in
- range(0, 512)]
- print(f"Loading from path {ckpt_file_list[0]}", flush=True)
- load_distributed_checkpoint(model.train_network, ckpt_file_list, strategy)
- print("Dataset size: {}, actual_epoch_num: {}".format(step_per_epoch, actual_epoch_num), flush=True)
- model.train(actual_epoch_num, ds, callbacks=callback, sink_size=args_opt.sink_size, dataset_sink_mode=True)
-
-
- def run_train_asc(net, model, args_opt, callback, rank_id, config, device_num, epoch_idx):
- model_parallel_num = args_opt.op_level_model_parallel_num
- data_parallel_num = int(device_num / model_parallel_num)
- batch_size = args_opt.per_batch_size * data_parallel_num
- micro_batch_interleaved = args_opt.micro_batch_interleaved
-
- ds = create_dataset(batch_size * micro_batch_interleaved,
- data_path=args_opt.data_url,
- data_start_index=epoch_idx, # max(0, epoch_idx-1),
- eod_reset=args_opt.eod_reset,
- full_batch=bool(args_opt.full_batch),
- eod_id=args_opt.eod_id,
- device_num=device_num,
- rank=rank_id,
- column_name=args_opt.data_column_name,
- epoch=1)
- # actual_epoch_num = int(epoch_num * step_per_epoch / args_opt.sink_size)
- step_per_epoch = ds.get_dataset_size()
- callback_size = args_opt.sink_size
- steps_num = int(ds.get_dataset_size())
- # sink_mode=True: actual_epoch_num = int(epoch_num * step_per_epoch / args_opt.sink_size)
- # sink_mode=False: actual_epoch_num = 1
- actual_epoch_num = 1 # int(step_per_epoch / callback_size)
- print("===dataset size: ", int(ds.get_dataset_size()), flush=True)
- print("===actual_epoch_num: ", actual_epoch_num, flush=True)
-
- if epoch_idx >= 1:
- if rank_id == 0:
- model.train(actual_epoch_num, ds,
- callbacks=callback,
- sink_size=callback_size,
- dataset_sink_mode=False)
- else:
- # synchronization parameters
- local_ckpt_dir = f"/cache/ckpt/mPanGu-350M_epoch--{str(epoch_idx - 1)}.ckpt"
- ckpt_save_flag = f"./cache_file/Ckpt_save_flag--{str(epoch_idx - 1)}.txt"
-
- while not os.path.isfile(ckpt_save_flag):
- print(f"{ckpt_save_flag} not found, waitting 10s ...", flush=True)
- time.sleep(10)
- while not os.path.exists(local_ckpt_dir):
- print(f"{local_ckpt_dir} not found, break ...", flush=True)
- break
-
- print(os.path.getsize(local_ckpt_dir) / (1024 * 1024), local_ckpt_dir, '\n\n')
- state_dict = load_checkpoint(local_ckpt_dir)
-
- load_param_into_net(model.train_network, state_dict)
- model.train(actual_epoch_num, ds,
- callbacks=callback,
- sink_size=callback_size,
- dataset_sink_mode=False)
- else:
- # if args_opt.incremental_training:
- # strategy = model.infer_train_layout(train_dataset=ds, sink_size=args_opt.sink_size)
- # print("======start load_distributed checkpoint", flush=True)
- # # For 2.6B and 13B models, the number of ckpt files is 512.
- # ckpt_file_list = [os.path.join(args_opt.load_ckpt_path, f"filerted_{ckpt_rank}.ckpt") for ckpt_rank in range(0, 512)]
- # print(f"Loading from path {ckpt_file_list[0]}", flush=True)
- # load_distributed_checkpoint(model.train_network, ckpt_file_list, strategy)
- model.train(actual_epoch_num, ds,
- callbacks=callback,
- sink_size=callback_size,
- dataset_sink_mode=False)
- return steps_num
-
-
- def main(opt):
- net, args_opt, model, callback, config = load_train_net(opt)
- D.init()
- device_num = D.get_group_size()
- rank_id = D.get_rank()
- round_num = 10
-
- class Node_MindSpore_NPU(asc.client.NumPyClient):
- def __init__(self):
- self.local_ckpt_dir = f"/cache/ckpt/mPanGu-350M_epoch--0.ckpt"
- self.tmp_ckpt_dir = f"/cache/ckpt/tmp_ckpt.ckpt"
- self.local_ckpt_save_flag = f"./cache_file/Ckpt_save_flag--0.txt"
- self.count = 0
-
- if not os.path.exists("/cache/ckpt/"):
- os.makedirs("/cache/ckpt/")
- if not os.path.exists("./cache_file/ckpt/"):
- os.makedirs("./cache_file/ckpt/")
-
- def get_parameters(self): # get local model parameters
- # get network parameters
- print(f"get_parameters...{self.count}....\n" * 5, flush=True)
-
- save_checkpoint(net, self.tmp_ckpt_dir, integrated_save=True)
- # align parameters
- state_dict = load_checkpoint(self.tmp_ckpt_dir)
- keys = [key for key in state_dict.keys() if 'adam' not in key]
- return [np.array(state_dict[i].asnumpy(), dtype=np.float32) for i in keys]
-
- def set_parameters(self, parameters):
- # init network parameters
- print(f"set_parameters....{self.count}..." * 4, flush=True)
-
- state_dict = {}
- for idx, key in enumerate(net.parameters_dict().keys()):
- # align parameters
- state_dict[key] = Parameter(np.array(parameters[idx], dtype=np.float32))
-
- load_param_into_net(net, state_dict)
-
- params = [{'name': k, 'data': v} for k, v in state_dict.items()]
- local_ckpt_dir = self.local_ckpt_dir.split("--")[0] + f"--{str(self.count)}.ckpt"
- save_checkpoint(params, local_ckpt_dir)
-
- while not os.path.exists(local_ckpt_dir):
- print(f"Save ckpt again at round {self.count}\n\n", flush=True)
- save_checkpoint(params, local_ckpt_dir)
- time.sleep(5)
-
- if os.path.exists(local_ckpt_dir):
- f = open(self.local_ckpt_save_flag.split("--")[0] + f"--{str(self.count)}.txt", 'w')
- f.close()
-
- def fit(self, parameters, config): # globle model parameters
- print(f"AiSynergy training fit at round {str(self.count)}...\n\n", flush=True)
- run_train_asc(net, model, args_opt, callback, rank_id, config, device_num, epoch_idx=self.count)
- return self.get_parameters(), 5, {}
-
- def evaluate(self, parameters, config):
- self.set_parameters(parameters)
- # mPanGu don't use evaluate mode
- loss = 1.0
- self.count += 1
- return float(loss), 5, {"acc: ": float(0.9)}
-
- if rank_id == 0:
- SERVER_IP = '*.*.*.*'
- port = [30000]
- # AiSynergy 2.0
- print(f"nginx_port:{args_opt.ng_port}")
- asc.client.run_numpy_client(f"218.17.115.229:30003", args_opt.ng_port,
- client=Node_MindSpore_NPU(),
- grpc_max_message_length=2 * 1024 * 1024 * 1024 - 1)
- # AiSynergy 1.0
- # asc.client.run_numpy_client(f"218.17.115.229:{args_opt.ng_port}",
- # client=Node_MindSpore_NPU(),
- # grpc_max_message_length=2 * 1024 * 1024 * 1024 - 1)
- else:
- for i in range(round_num + 1):
- _ = run_train_asc(net, model, args_opt, callback, rank_id, config, device_num, epoch_idx=i)
- if i == round_num:
- break
-
-
- if __name__ == "__main__":
- opt = get_args()
- set_parse(opt)
- if opt.per_batch_size == 0:
- raise ValueError("The per_batch_size has not been configured.")
- if bool(opt.enable_alltoall) is True and bool(opt.use_moe) is False:
- raise ValueError("The alltoall communication is only effective when applying moe")
- os.environ['HCCL_CONNECT_TIMEOUT'] = str(opt.hccl_connect_time)
- if opt.stage_num > 1:
- raise ValueError("The pipline parallel is only effective when applying more then 64 device number")
- else:
- # run_train(opt)
- main(opt)
|