set VLLM_USE_FUSED_RMS_ROPE=1

vllm/envs.py

@@ -1687,7 +1687,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
                 ("true", "1")),
     # vLLM will use fused RMS + RoPE kernel
     "VLLM_USE_FUSED_RMS_ROPE":
-        lambda: (os.environ.get("VLLM_USE_FUSED_RMS_ROPE", "False").lower() in
+        lambda: (os.environ.get("VLLM_USE_FUSED_RMS_ROPE", "True").lower() in
                  ("true", "1")),
     # vLLM will use Marlin W16A16 kernel for MoE experts
     "VLLM_USE_MARLIN_W16A16_MOE":