update VLLM_USE_FUSED_RMS_ROPE=0 (default)

for qwen3, VLLM_USE_FUSED_RMS_ROPE=1 (default)

vllm/envs.py

@@ -1686,7 +1686,7 @@ 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", "True").lower() in
+        lambda: (os.environ.get("VLLM_USE_FUSED_RMS_ROPE", "False").lower() in
                  ("true", "1")),
     # vLLM will use fast token id copy
     "VLLM_V1_FAST_TOKEN_ID_COPY":

vllm/model_executor/model_loader/utils.py

@@ -210,6 +210,8 @@ 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"):
@@ -237,6 +239,8 @@ 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"):