biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[compile] Cleanup: Remove unnecessary +rms_norm forcing for sequence parallelism (#35410)

Browse Source 
Signed-off-by: jasonlizhengjian <jasonlizhengjian@gmail.com>
This commit is contained in:
Jason Li
2026-02-27 05:36:37 -08:00
committed by GitHub
parent 6467b635b6
commit 66c1751d13
1 changed files with 1 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/config/vllm.py
View File

@@ -860,7 +860,7 @@ class VllmConfig:
self.compilation_config.pass_config.fuse_gemm_comms = False
else:
# Compute SP threshold early; disable if None (model too
# small) before +rms_norm gets forced into custom_ops.
# small for SP to be beneficial).
pass_config = self.compilation_config.pass_config
if pass_config.sp_min_token_num is None:
from vllm.compilation.passes.fusion.sequence_parallelism import (
@@ -883,14 +883,6 @@ class VllmConfig:
self.compilation_config.pass_config.enable_sp = False
self.compilation_config.pass_config.fuse_gemm_comms = False
if self.compilation_config.pass_config.enable_sp:
if "-rms_norm" in self.compilation_config.custom_ops:
logger.warning(
"RMS norm force disabled, sequence parallelism might break"
)
else:
self.compilation_config.custom_ops.append("+rms_norm")
if self.compilation_config.fast_moe_cold_start is None:
# resolve default behavior: try to be as safe as possible
# this config is unsafe if any spec decoding draft model has a MOE.