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 14, 05:30 UTC
[ROCm][Bugfix] Only enable +rms_norm based on aiter if not explicitly disabled (#25275)
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Gregory Shtrasberg committed
Sep 24, 2025 at 11:24 UTC
487745ff493c365fe328f8d8bc8b29f50e04dfc2
1 file changed
+2
-1
vllm/platforms/rocm.py
+2
-1
@@ -340,7 +340,8 @@ class RocmPlatform(Platform):
340
else:
341
parallel_config.worker_cls = "vllm.worker.worker.Worker"
342
# Aiter rms norm perform best when CUDA Graph capture is enabled.
343
- if use_v1 and use_aiter_rms_norm and not is_eager_execution:
343
+ if (use_v1 and use_aiter_rms_norm and not is_eager_execution
344
+ and "-rms_norm" not in compilation_config.custom_ops):
345
compilation_config.custom_ops.append("+rms_norm")
346
347
@classmethod