[BugFix] fix group_topk (#8430)

Author: dsikka

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -410,6 +410,7 @@ def fused_topk(
 
     if renormalize:
         topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
+
     return topk_weights, topk_ids
 
 
@@ -443,7 +444,8 @@ def grouped_topk(hidden_states: torch.Tensor,
 
     if renormalize:
         topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
-    return topk_weights, topk_ids
+
+    return topk_weights, topk_ids.to(torch.int32)
 
 
 def get_config_dtype_str(dtype: torch.dtype,