[BUG] Two phase pause to prevent deadlock (#39366)

Signed-off-by: ahao-anyscale <ahao@anyscale.com>
Signed-off-by: Aaron Hao <ahao@anyscale.com>
Co-authored-by: Junjie Zhang <junj.jay.zhang@gmail.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Aaron Hao
2026-04-29 17:51:03 -04:00
committed by GitHub
co-authored by Junjie Zhang Nick Hill
parent 944e138bcf
commit 0335316a9b
3 changed files with 232 additions and 20 deletions
+27
View File
@@ -663,6 +663,33 @@ class ParallelConfig:
aggregated_has_unfinished = bool(tensor.item())
return aggregated_has_unfinished
@staticmethod
def sync_dp_state(
dp_group: ProcessGroup, has_unfinished: bool, pending_pause: bool
) -> tuple[bool, bool]:
"""Combined all-reduce for DP state synchronization.
Uses a single SUM all-reduce on a 2-element tensor:
[0] = 1 if this rank has unfinished work, else 0.
SUM > 0 ≡ logical OR across ranks → any rank has work.
[1] = 1 if this rank has a pending pause request, else 0.
SUM == dp_size ≡ all ranks reached pause consensus.
has_unfinished_global is true if any rank has unfinished work,
or if some ranks are waiting for a pause consensus.
Returns:
(has_unfinished_global, pause_consensus)
"""
tensor = torch.tensor(
[int(has_unfinished), int(pending_pause)], dtype=torch.int32, device="cpu"
)
torch.distributed.all_reduce(tensor, op=ReduceOp.SUM, group=dp_group)
dp_size = dp_group.size()
pause_count = tensor[1].item()
has_unfinished_global = tensor[0].item() > 0 or pause_count % dp_size != 0
return has_unfinished_global, pause_count == dp_size
@staticmethod
def sync_kv_cache_memory_size(dp_group: ProcessGroup, kv_cache_memory: int) -> int:
if kv_cache_memory == -1: