[Core] simplify logits resort in _apply_top_k_top_p (vllm-project#8619)

Author: hidva

vllm/model_executor/layers/sampler.py

@@ -433,12 +433,9 @@ def _apply_top_k_top_p(
     logits_sort.masked_fill_(top_p_mask, -float("inf"))
 
     # Re-sort the probabilities.
-    src = torch.arange(logits_idx.shape[-1],
-                       device=logits_idx.device).expand_as(logits_idx)
-    logits_idx_inv = torch.empty_like(logits_idx).scatter_(dim=-1,
-                                                           index=logits_idx,
-                                                           src=src)
-    logits = torch.gather(logits_sort, dim=-1, index=logits_idx_inv)
+    logits = torch.empty_like(logits_sort).scatter_(dim=-1,
+                                                    index=logits_idx,
+                                                    src=logits_sort)
     return logits