update VLLM_USE_FUSED_RMS_ROPE=0 (default)

for qwen3, VLLM_USE_FUSED_RMS_ROPE=1 (default)

vllm/envs.py

@@ -1309,7 +1309,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
         
     # vLLM will use fused RMS + RoPE kernel
     "VLLM_USE_FUSED_RMS_ROPE":
-        lambda: (os.environ.get("VLLM_USE_FUSED_RMS_ROPE", "True").lower() in
+        lambda: (os.environ.get("VLLM_USE_FUSED_RMS_ROPE", "False").lower() in
                  ("true", "1")),
     # vLLM will use Marlin W16A16 kernel for MoE experts
     "VLLM_USE_MARLIN_W16A16_MOE":

vllm/model_executor/model_loader/utils.py

@@ -287,6 +287,8 @@ def get_model_architecture(
                         os.environ['VLLM_USE_FUSE_SILU_AND_MUL'] = '1'
                     if not envs.is_set("VLLM_USE_OPT_RESHAPE_AND_CACHE"):
                         os.environ['VLLM_USE_OPT_RESHAPE_AND_CACHE'] = '1'
+                    if not envs.is_set("VLLM_USE_FUSED_RMS_ROPE"):
+                        os.environ['VLLM_USE_FUSED_RMS_ROPE'] = '1'
                 
             if architectures in [['DeepseekV32ForCausalLM']]:
                 if not envs.is_set("VLLM_USE_V32_ENCODE"):
@@ -334,6 +336,8 @@ def get_model_architecture(
                         os.environ['VLLM_USE_FUSE_SILU_AND_MUL'] = '1'
                     if not envs.is_set("VLLM_USE_OPT_RESHAPE_AND_CACHE"):
                         os.environ['VLLM_USE_OPT_RESHAPE_AND_CACHE'] = '1'
+                    if not envs.is_set("VLLM_USE_FUSED_RMS_ROPE"):
+                        os.environ['VLLM_USE_FUSED_RMS_ROPE'] = '1'
                  
             if architectures in [['DeepseekV32ForCausalLM']]:
                 if not envs.is_set("VLLM_USE_V32_ENCODE"):