Skip to content

Megatron-FSDP: Fix insufficient double buffers during gradient reduce #4054

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

Merged
cspades merged 3 commits into from
Apr 2, 2026
Merged

Megatron-FSDP: Fix insufficient double buffers during gradient reduce #4054

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
3697f00
Megatron-FSDP: Fix insufficient double buffers during gradient reduce
shjwudp Mar 28, 2026
2888056
Merge branch 'main' into mfsdp_fix_double_buffer
cspades Apr 1, 2026
74edeeb
Merge branch 'main' into mfsdp_fix_double_buffer
cspades Apr 1, 2026
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
6 changes: 4 additions & 2 deletions megatron/core/distributed/fsdp/src/megatron_fsdp/param_and_grad_buffer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3351,9 +3351,11 @@ def _enforce_double_buffer_limit(self, add_buckets):
for _, _, bucket_id in reversed(self.grad_reduce_queue):
fsdp_unit_id = param_groups[bucket_id].fsdp_unit_id
double_buf_units.add(fsdp_unit_id)
if len(double_buf_units) > 2:
if len(double_buf_units) > 1:
keep_n -= 1
self.wait_for_previous_grad_reduce(keep_n)

with torch.cuda.stream(self.rs_stream):
self.wait_for_previous_grad_reduce(keep_n)

def get_ready_bucket_group_for_reduction(self, bucket_id: int) -> Optional[List[int]]:
"""Checks if all buckets in the bucket group containing the given bucket_id
Expand Down
Loading