[Core] Disable fuse_allreduce_rms under VLLM_BATCH_INVARIANT (non-deterministic under TP) (#51292)

Signed-off-by: Don Tolley <tolleybot@gmail.com>
Signed-off-by: tolleybot <tolleybot@gmail.com>
Co-authored-by: yewentao256 <yewentao256@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Donald Tolley
2026-08-25 15:16:09 -04:00
committed by GitHub
co-authored by yewentao256 mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 0e30bd62fe
commit 80771bbbdd
+4
View File
@@ -183,6 +183,10 @@ def enable_allreduce_rms_fusion(cfg: "VllmConfig") -> bool:
from vllm.platforms import current_platform
from vllm.utils.flashinfer import has_flashinfer
# The fused all-reduce + RMSNorm path is not batch-invariant
if envs.VLLM_BATCH_INVARIANT:
return False
if current_platform.is_rocm():
from vllm._aiter_ops import rocm_aiter_ops