A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 7, 15:15 UTC
[compile] Cleanup: Remove unnecessary +rms_norm forcing for sequence parallelism (#35410)
Signed-off-by: jasonlizhengjian <jasonlizhengjian@gmail.com>
Jason Li committed
Feb 27, 2026 at 05:36 UTC
66c1751d13b7b3c294418a88fbfbfe2ec49d5d3f
1 file changed
+1
-9
vllm/config/vllm.py
+1
-9
@@ -860,7 +860,7 @@ class VllmConfig:
860
self.compilation_config.pass_config.fuse_gemm_comms = False
861
else:
862
# Compute SP threshold early; disable if None (model too
863
- # small) before +rms_norm gets forced into custom_ops.
863
+ # small for SP to be beneficial).
864
pass_config = self.compilation_config.pass_config
865
if pass_config.sp_min_token_num is None:
866
from vllm.compilation.passes.fusion.sequence_parallelism import (
@@ -883,14 +883,6 @@ class VllmConfig:
883
self.compilation_config.pass_config.enable_sp = False
884
self.compilation_config.pass_config.fuse_gemm_comms = False
885
886
- if self.compilation_config.pass_config.enable_sp:
887
- if "-rms_norm" in self.compilation_config.custom_ops:
888
- logger.warning(
889
- "RMS norm force disabled, sequence parallelism might break"
890
- )
891
- else:
892
- self.compilation_config.custom_ops.append("+rms_norm")
893
-
886
if self.compilation_config.fast_moe_cold_start is None:
887
# resolve default behavior: try to be as safe as possible
888
# this config is unsafe if any spec decoding draft model has a MOE.