USE_FUSED_SILU_MUL_QUANT默认改成开启

vllm/envs.py

@@ -173,7 +173,7 @@ if TYPE_CHECKING:
     VLLM_USE_LIGHTOP_MOE_ALIGN: bool = False
     VLLM_USE_MERGE_ATTN_STATES_OPT: bool = False
     USE_FUSED_RMS_QUANT: bool = False
-    USE_FUSED_SILU_MUL_QUANT: bool = False
+    USE_FUSED_SILU_MUL_QUANT: bool = True
     VLLM_P2P_ASYNC: bool = False
     VLLM_P2P_BUF_TOKENS: int = 30000
     VLLM_SCHED_ENABLE_MINIMAL_INJECTION: bool = False
@@ -1141,10 +1141,12 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "USE_FUSED_RMS_QUANT": 
     lambda: (os.getenv('USE_FUSED_RMS_QUANT', '0').lower() in
              ("true", "1")),
-    # vllm will use silu_mul_quant fused op 
-    "USE_FUSED_SILU_MUL_QUANT": 
-    lambda: (os.getenv('USE_FUSED_SILU_MUL_QUANT', '0').lower() in
-             ("true", "1")),
+    # vllm will use silu_mul_quant fused op,
+    # This variable has a default value of true, 
+    # but it is still controlled by CRQ and RQ.
+    "USE_FUSED_SILU_MUL_QUANT":
+    lambda: bool(int(os.getenv("USE_FUSED_SILU_MUL_QUANT", "1"))),
+    
     # vllm pd separation will be used async
     "VLLM_P2P_ASYNC":
     lambda: bool(int(os.getenv("VLLM_P2P_ASYNC", "0"))),