github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-11-03 07:33:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

update sharded optim and fix zero init ctx (#457)

Browse Source
This commit is contained in:
ver217
2022-03-18 15:44:47 +08:00
committed by GitHub
parent e2e9f82588
commit 642846d6f9
11 changed files with 162 additions and 162 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
tests/test_zero_data_parallel/test_shard_model_v2.py
View File

@@ -1,15 +1,13 @@
#!/usr/bin/env python
# -*- encoding: utf-8 -*-
import copy
from asyncio.log import logger
from functools import partial
import colossalai
import pytest
import torch
import torch.multiprocessing as mp
from colossalai.logging import get_dist_logger
from colossalai.testing import parameterize
from colossalai.utils import free_port
from colossalai.zero.init_ctx import ZeroInitContext
from colossalai.zero.shard_utils import (BucketTensorShardStrategy, TensorShardStrategy)
@@ -20,36 +18,30 @@ from tests.components_to_test.registry import non_distributed_component_funcs
from torch.nn.parallel import DistributedDataParallel as DDP
from common import CONFIG, check_grads_padding, run_fwd_bwd
from colossalai.testing import parameterize
@parameterize("enable_autocast", [True])
@parameterize("use_zero_init_ctx", [True])
@parameterize("shard_strategy", [TensorShardStrategy, BucketTensorShardStrategy])
def run_model_test(enable_autocast, use_zero_init_ctx, shard_strategy, logger):
@parameterize("shard_strategy_class", [TensorShardStrategy, BucketTensorShardStrategy])
def run_model_test(enable_autocast, shard_strategy_class):
test_models = ['repeated_computed_layers', 'resnet18', 'bert']
shard_strategy = shard_strategy()
shard_strategy = shard_strategy_class()
for model_name in test_models:
get_components_func = non_distributed_component_funcs.get_callable(model_name)
model_builder, train_dataloader, _, _, criterion = get_components_func()
rm_torch_payload_on_the_fly = False
if use_zero_init_ctx:
with ZeroInitContext(convert_fp16=True,
target_device=torch.device(f'cpu:0'),
shard_strategy=shard_strategy,
shard_param=True,
rm_torch_payload_on_the_fly=rm_torch_payload_on_the_fly):
zero_model = model_builder(checkpoint=True)
zero_model = ShardedModelV2(zero_model, shard_strategy, use_memory_tracer=True)
with ZeroInitContext(convert_fp16=True,
target_device=torch.cuda.current_device(),
shard_strategy=shard_strategy,
shard_param=True,
rm_torch_payload_on_the_fly=rm_torch_payload_on_the_fly):
zero_model = model_builder(checkpoint=True)
zero_model = ShardedModelV2(zero_model, shard_strategy, use_memory_tracer=True)
model = model_builder(checkpoint=True).half()
col_model_deepcopy(zero_model, model)
model = model.cuda()
else:
model = model_builder(checkpoint=True).half().cuda()
zero_model = ShardedModelV2(copy.deepcopy(model), shard_strategy)
model = model_builder(checkpoint=True).half()
col_model_deepcopy(zero_model, model)
model = model.cuda()
model = DDP(model)
@@ -63,15 +55,10 @@ def run_model_test(enable_autocast, use_zero_init_ctx, shard_strategy, logger):
check_grads_padding(model, zero_model, loose=True)
# logger.debug('overall cuda ', zero_model._memstats_collector._overall_cuda)
# logger.debug('model cuda ', zero_model._memstats_collector._model_data_cuda)
def run_dist(rank, world_size, port):
colossalai.launch(config=CONFIG, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
logger = get_dist_logger()
logger.set_level('DEBUG')
run_model_test(logger=logger)
run_model_test()
@pytest.mark.dist