github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-10-13 18:58:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

polish unit test

Browse Source
This commit is contained in:
ver217
2022-03-14 15:06:02 +08:00
parent 88804aee49
commit 54fd37f0e0
6 changed files with 64 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
tests/test_zero_data_parallel/test_init_context.py
View File

@@ -4,21 +4,20 @@
from functools import partial
import colossalai
from colossalai.utils.cuda import get_current_device
import pytest
import torch
import torch.multiprocessing as mp
from colossalai.utils import free_port
from colossalai.utils.cuda import get_current_device
from colossalai.utils.memory_tracer.allocator import GLOBAL_MODEL_DATA_TRACER
from colossalai.zero.init_ctx import ZeroInitContext
from colossalai.zero.shard_utils.tensor_shard_strategy import \
TensorShardStrategy
from colossalai.zero.shard_utils import (BucketTensorShardStrategy, TensorShardStrategy)
from tests.components_to_test.registry import non_distributed_component_funcs
from common import CONFIG
from colossalai.utils.memory_tracer.allocator import GLOBAL_MODEL_DATA_TRACER
def run_dist(rank, world_size, port, init_device):
def run_dist(rank, world_size, port, init_device, shard_strategy):
colossalai.launch(config=CONFIG, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
for get_components_func in non_distributed_component_funcs:
@@ -26,7 +25,7 @@ def run_dist(rank, world_size, port, init_device):
model_numel_tensor = torch.zeros(1, dtype=torch.int)
with ZeroInitContext(convert_fp16=True,
target_device=init_device,
shard_strategy=TensorShardStrategy(),
shard_strategy=shard_strategy(),
shard_param=True,
model_numel_tensor=model_numel_tensor):
model = model_builder(checkpoint=True)
@@ -50,11 +49,16 @@ def run_dist(rank, world_size, port, init_device):
@pytest.mark.dist
@pytest.mark.parametrize("world_size", [1, 4])
@pytest.mark.parametrize("init_device", [torch.device('cpu'), torch.device(f'cuda:{get_current_device()}')])
def test_zero_init_context(world_size, init_device):
run_func = partial(run_dist, world_size=world_size, port=free_port(), init_device=init_device)
@pytest.mark.parametrize("shard_strategy", [TensorShardStrategy, BucketTensorShardStrategy])
def test_zero_init_context(world_size, init_device, shard_strategy):
run_func = partial(run_dist,
world_size=world_size,
port=free_port(),
init_device=init_device,
shard_strategy=shard_strategy)
mp.spawn(run_func, nprocs=world_size)
if __name__ == '__main__':
test_zero_init_context(2, torch.device('cpu'))
test_zero_init_context(2, torch.device(f'cuda:{get_current_device()}'))
test_zero_init_context(2, torch.device('cpu'), TensorShardStrategy)
test_zero_init_context(2, torch.device(f'cuda:{get_current_device()}'), TensorShardStrategy)