From 3322e26420bc9ea22c5033e9199cd9fb8be5f424 Mon Sep 17 00:00:00 2001
From: "Li, Jiang" <jiang1.li@intel.com>
Date: Thu, 19 Mar 2026 18:24:39 +0800
Subject: [PATCH] [Bugfix] Avoid more OpenMP thread reallocation in CPU torch
 compile  (#37538)

Signed-off-by: jiang1.li <jiang1.li@intel.com>
---
 vllm/platforms/cpu.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/vllm/platforms/cpu.py b/vllm/platforms/cpu.py
index c1bcf5b55..f8fc3a38a 100644
--- a/vllm/platforms/cpu.py
+++ b/vllm/platforms/cpu.py
@@ -281,6 +281,9 @@ class CpuPlatform(Platform):
         # Disable multi-stream for shared experts as no Stream on CPU
         os.environ["VLLM_DISABLE_SHARED_EXPERTS_STREAM"] = "1"
 
+        # Avoid inductor generates num_thread() and breaks the thread binding
+        os.environ["TORCHINDUCTOR_CPP_DYNAMIC_THREADS"] = "1"
+
         # Intel OpenMP setting
         ld_preload_str = os.getenv("LD_PRELOAD", "")
         if "libiomp5.so" in ld_preload_str: