update VLLM_USE_TOPK_RENORM

vllm/envs.py

@@ -1683,7 +1683,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # vLLM will use optimized topk_softmax + renormalize
     "VLLM_USE_TOPK_RENORM":
         lambda:
-        (os.environ.get("VLLM_USE_TOPK_RENORM", "False").lower() in
+        (os.environ.get("VLLM_USE_TOPK_RENORM", "True").lower() in
                 ("true", "1")),
     # vLLM will use fused RMS + RoPE kernel
     "VLLM_USE_FUSED_RMS_ROPE":

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -1363,7 +1363,7 @@ 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:
+    if envs.VLLM_USE_TOPK_RENORM and renormalize is True:
         from lightop import op as op
         op.topk_softmax(
             topk_weights,