Skip to content

[checkpoint_io] Fix gather_state_dict_fast #6333

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

[checkpoint_io] Fix gather_state_dict_fast #6333

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
aedb8d1
Fix gather_state_dict_fast
pbelevich May 28, 2025
dd562d2
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] May 28, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 7 additions & 5 deletions colossalai/checkpoint_io/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1132,18 +1132,20 @@ def gather_state_dict_fast(
if rank == dst:
returned_state_dict = state_dict.copy()
dist.gather_object(metadata, all_meta_data, dst=dist.get_global_rank(group, rank), group=group)
ks, ops = [], []
for i, target_metadata in enumerate(all_meta_data):
if i == dst:
continue
ops = []
for k, shape, dtype in target_metadata:
buffer = torch.empty(shape, dtype=dtype, device=get_current_device())
returned_state_dict[k] = buffer
ks.append(k)
ops.append(dist.P2POp(dist.irecv, buffer, dist.get_global_rank(group, i), group))
reqs = dist.batch_isend_irecv(ops)
for req, (k, *_) in zip(reqs, target_metadata):
req.wait()
returned_state_dict[k] = returned_state_dict[k].to(device)
reqs = dist.batch_isend_irecv(ops)
for req in reqs: # len(reqs) maybe be different from len(ops) because of coalescing
req.wait()
for k in ks:
returned_state_dict[k] = returned_state_dict[k].to(device)
return returned_state_dict
else:
dist.gather_object(metadata, dst=dist.get_global_rank(group, dst), group=group)
Expand Down