update topk softmax (#985)

Author: shihaobai

lightllm/common/fused_moe/topk_select.py

@@ -41,19 +41,13 @@ def fused_topk(
 
     topk_weights = torch.empty(M, topk, dtype=torch.float32, device=hidden_states.device)
     topk_ids = torch.empty(M, topk, dtype=torch.int32, device=hidden_states.device)
-    token_expert_indicies = torch.empty(M, topk, dtype=torch.int32, device=hidden_states.device)
 
     sgl_ops.topk_softmax(
         topk_weights,
         topk_ids,
-        token_expert_indicies,
         gating_output.float(),  # TODO(woosuk): Optimize this.
+        renormalize=renormalize,
     )
-    del token_expert_indicies  # Not used. Will be used in the future.
-
-    if renormalize:
-        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
-
     return topk_weights, topk_ids