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
[Bugfix][CPU] Disable dual stream execution for experts on CPU (#27320)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Li, Jiang committed
Oct 22, 2025 at 19:02 UTC
843af7f7fc6c9395da02da0f2f5fa2cfdcf92e81
1 file changed
+3
vllm/platforms/cpu.py
+3
@@ -297,6 +297,9 @@ class CpuPlatform(Platform):
297
# Disable torch async compiling which won't work with daemonic processes
298
os.environ["TORCHINDUCTOR_COMPILE_THREADS"] = "1"
299
300
+ # Disable multi-stream for shared experts as no Stream on CPU
301
+ os.environ["VLLM_DISABLE_SHARED_EXPERTS_STREAM"] = "0"
302
+
303
# Intel OpenMP setting
304
ld_prealod_str = os.getenv("LD_PRELOAD", "")
305
if "libiomp5.so" in ld_prealod_str: