Skip to content

fix: disable dynamo on Ulysses all-to-all to prevent FSDP×compile NCCL deadlock #3195

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
samsja merged 1 commit into from
Aug 6, 2026
Merged

fix: disable dynamo on Ulysses all-to-all to prevent FSDP×compile NCCL deadlock #3195

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
8 changes: 8 additions & 0 deletions src/prime_rl/trainer/models/layers/ulysses_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,11 +68,18 @@ def _replicate_kv_heads(t: torch.Tensor, cp_size: int) -> torch.Tensor:
return t.repeat_interleave(cp_size // h, dim=1)


@torch._dynamo.disable
def _all_to_all_seq_to_head(t: torch.Tensor, cp_size: int, cp_group: dist.ProcessGroup) -> torch.Tensor:
"""Redistribute [S_local, H, D] -> [S_global, H_local, D].

Splits the head dim into cp_size groups and exchanges them so each rank
ends up with the full sequence but only H/cp_size heads. Differentiable.

Disabled for dynamo: the collective must execute eagerly so it is scheduled
by the autograd engine in a consistent order across ranks, not reordered by
the compiled backward graph. Otherwise FSDP's post_backward reduce-scatter
and the all-to-all backward can interleave differently per rank, deadlocking
NCCL (see PR #3186 job 1578).
"""
s_local, h, d = t.shape
assert h % cp_size == 0, (
Expand All @@ -90,6 +97,7 @@ def _all_to_all_seq_to_head(t: torch.Tensor, cp_size: int, cp_group: dist.Proces
return out.reshape(cp_size * s_local, h_local, d)


@torch._dynamo.disable
def _all_to_all_head_to_seq(t: torch.Tensor, cp_size: int, cp_group: dist.ProcessGroup) -> torch.Tensor:
"""Inverse of `_all_to_all_seq_to_head`: [S_global, H_local, D] -> [S_local, H, D]."""
s_global, h_local, d = t.shape
Expand Down