set VLLM_REJECT_SAMPLE_OPT=1

vllm/envs.py

@@ -1228,7 +1228,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
 
     # vllm will use fused cat and mla
     "VLLM_REJECT_SAMPLE_OPT":
-        lambda: (os.getenv('VLLM_REJECT_SAMPLE_OPT', 'False').lower() in
+        lambda: (os.getenv('VLLM_REJECT_SAMPLE_OPT', 'True').lower() in
                  ("true", "1")),      
 
     # vLLM will use fused silu+mul kernel (fp16 + qwen3-30b)

vllm/model_executor/model_loader/utils.py

@@ -253,8 +253,8 @@ def get_model_architecture(
                     os.environ['VLLM_USE_OPT_CAT'] = '1'
                 if not envs.is_set("VLLM_USE_CAT_MLA"):
                     os.environ['VLLM_USE_CAT_MLA'] = '1'
-                if not envs.is_set("VLLM_REJECT_SAMPLE_OPT"):
-                    os.environ['VLLM_REJECT_SAMPLE_OPT'] = '1'
+                # if not envs.is_set("VLLM_REJECT_SAMPLE_OPT"):
+                #     os.environ['VLLM_REJECT_SAMPLE_OPT'] = '1'
                 if not envs.is_set("VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT"):
                     os.environ['VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT'] = '1'
                 if not envs.is_set("VLLM_SCHED_ENABLE_MINIMAL_INJECTION"):
@@ -298,8 +298,8 @@ def get_model_architecture(
                     os.environ['VLLM_USE_OPT_CAT'] = '1'
                 if not envs.is_set("VLLM_USE_CAT_MLA"):
                     os.environ['VLLM_USE_CAT_MLA'] = '1'
-                if not envs.is_set("VLLM_REJECT_SAMPLE_OPT"):
-                    os.environ['VLLM_REJECT_SAMPLE_OPT'] = '1'
+                # if not envs.is_set("VLLM_REJECT_SAMPLE_OPT"):
+                #     os.environ['VLLM_REJECT_SAMPLE_OPT'] = '1'
                 if not envs.is_set("VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT"):
                     os.environ['VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT'] = '1'
                 if not envs.is_set("VLLM_SCHED_ENABLE_MINIMAL_INJECTION"):