github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-08 04:24:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

[zero] support extra dp (#6123)

Browse Source 
* [zero] support extra dp

* [zero] update checkpoint

* fix bugs

* fix bugs
This commit is contained in:
Hongxin Liu
2024-11-12 11:20:46 +08:00
committed by GitHub
parent 30a9443132
commit a2596519fd
8 changed files with 238 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
tests/test_zero/test_low_level/test_zero_ckpt.py
View File

@@ -2,12 +2,14 @@ import copy
import pytest
import torch
import torch.distributed as dist
import torch.nn as nn
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.testing import assert_close
import colossalai
from colossalai.testing import rerun_if_address_is_in_use, spawn
from colossalai.cluster import ProcessGroupMesh
from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
from colossalai.testing.random import seed_all
from colossalai.zero import LowLevelZeroOptimizer
@@ -40,11 +42,19 @@ def loose_close(a, b, dtype: torch.dtype = torch.float32):
assert_close(a, b, rtol=rtol, atol=atol)
def exam_zero_1_torch_ddp_ckpt():
@parameterize("extra_dp_size", [1, 2])
def exam_zero_1_torch_ddp_ckpt(extra_dp_size: int):
"""
We examine the state_dict of zero and DDP.
Moreover, we examine the zero's loading checkpoint of a torch ckpt.
"""
if extra_dp_size > 1:
pg_mesh = ProcessGroupMesh(extra_dp_size, dist.get_world_size() // extra_dp_size)
extra_dp_group = pg_mesh.get_group_along_axis(0)
dp_group = pg_mesh.get_group_along_axis(1)
else:
dp_group = None
extra_dp_group = None
local_rank = torch.distributed.get_rank()
seed_all(1453)
@@ -60,7 +70,12 @@ def exam_zero_1_torch_ddp_ckpt():
# we only test stage 1 here
# the state dicts of stage 1 and stage 2 are the same
zero_optimizer = LowLevelZeroOptimizer(
zero_optimizer, overlap_communication=True, initial_scale=1, reduce_bucket_size=262144
zero_optimizer,
overlap_communication=True,
initial_scale=1,
reduce_bucket_size=262144,
dp_process_group=dp_group,
extra_dp_group=extra_dp_group,
)
torch_optimizer = torch.optim.Adam(torch_model.parameters(), lr=1)
@@ -111,7 +126,7 @@ def run_dist(rank, world_size, port):
@pytest.mark.dist
@rerun_if_address_is_in_use()
def test_zero_ckpt():
spawn(run_dist, 2)
spawn(run_dist, 4)
if __name__ == "__main__":