Skip to content

[Bugfix] Disable allreduce_rms_fusion when pipeline_parallel_size > 1 #43616

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
zixi-qi wants to merge 1 commit into
base: main
Choose a base branch
from
+8 −1
Open

[Bugfix] Disable allreduce_rms_fusion when pipeline_parallel_size > 1 #43616

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
9 changes: 8 additions & 1 deletion vllm/config/vllm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -121,7 +121,13 @@ def enable_act_fusion(cfg: "VllmConfig") -> bool:


def enable_allreduce_rms_fusion(cfg: "VllmConfig") -> bool:
"""Enable if TP > 1 and Hopper/Blackwell and flashinfer installed."""
"""Enable if TP > 1, PP == 1, Hopper/Blackwell, and flashinfer installed.

Gated off for PP > 1: the fused op's GPU-side peer-signal spin-wait
assumes byte-identical kernel launches across TP peers, but concurrent
independent warmup of multiple TP subgroups lets ranks pick divergent
FlashInfer launch configs and deadlock.
"""
from vllm.platforms import current_platform
from vllm.utils.flashinfer import has_flashinfer

Expand All @@ -134,6 +140,7 @@ def enable_allreduce_rms_fusion(cfg: "VllmConfig") -> bool:

return (
cfg.parallel_config.tensor_parallel_size > 1
and cfg.parallel_config.pipeline_parallel_size == 1
and current_platform.is_cuda()
and has_flashinfer()
and (
Expand Down
Loading