add VLLM_USE_TOPK_RENORM to use optimized topk_softmax + renormalize

vllm/envs.py

@@ -188,6 +188,7 @@ if TYPE_CHECKING:
     VLLM_REJECT_SAMPLE_OPT: bool = False
     VLLM_USE_FUSE_SILU_AND_MUL: bool = False
     VLLM_USE_OPT_RESHAPE_AND_CACHE: bool = False
+    VLLM_USE_TOPK_RENORM: bool = False
 
 def get_default_cache_root():
     return os.getenv(
@@ -1230,6 +1231,13 @@ environment_variables: dict[str, Callable[[], Any]] = {
         lambda:
         (os.environ.get("VLLM_USE_OPT_RESHAPE_AND_CACHE", "False").lower() in
                 ("true", "1")),
+        
+    # vLLM will use optimized topk_softmax + renormalize
+    "VLLM_USE_TOPK_RENORM":
+        lambda:
+        (os.environ.get("VLLM_USE_TOPK_RENORM", "False").lower() in
+                ("true", "1")),
+        
 }
 
 # --8<-- [end:env-vars-definition]

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -1227,6 +1227,16 @@ def vllm_topk_softmax(topk_weights: torch.Tensor, topk_indices: torch.Tensor,
                       token_expert_indices: torch.Tensor,
                       gating_output: torch.Tensor,
                       renormalize: bool) -> tuple[torch.Tensor, ...]:
+    if envs.VLLM_USE_TOPK_RENORM:
+        from lightop import op as op
+        op.topk_softmax(
+            topk_weights,
+            topk_indices,
+            token_expert_indices,
+            gating_output,
+            Is_renormalize = True,
+        )
+    else:
         ops.topk_softmax(
             topk_weights,
             topk_indices,