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 6, 10:45 UTC
[Bugfix] Avoid more OpenMP thread reallocation in CPU torch compile (#37538)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Li, Jiang committed
Mar 19, 2026 at 18:24 UTC
3322e26420bc9ea22c5033e9199cd9fb8be5f424
1 file changed
+3
vllm/platforms/cpu.py
+3
@@ -281,6 +281,9 @@ class CpuPlatform(Platform):
281
# Disable multi-stream for shared experts as no Stream on CPU
282
os.environ["VLLM_DISABLE_SHARED_EXPERTS_STREAM"] = "1"
283
284
+ # Avoid inductor generates num_thread() and breaks the thread binding
285
+ os.environ["TORCHINDUCTOR_CPP_DYNAMIC_THREADS"] = "1"
286
+
287
# Intel OpenMP setting
288
ld_preload_str = os.getenv("LD_PRELOAD", "")
289
if "libiomp5.so" in ld_preload_str: