fix

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

vllm/v1/worker/gpu/sampler.py

@@ -246,16 +246,15 @@ def compute_logprobs(
         logprob_token_ids = torch.cat(
             (sampled_token_ids.unsqueeze(-1), topk_indices), dim=1)
 
+    # NOTE(woosuk): Here, to save GPU memory, we do not materialize the full
+    # logprobs tensor. Instead, we only compute and return the logprobs of
+    # the topk + 1 tokens.
     logprobs = torch.empty(
         batch_size,
         num_logprobs + 1,
         dtype=torch.float32,
         device=logits.device,
     )
-
-    # NOTE(woosuk): Here, to save GPU memory, we do not materialize the full
-    # logprobs tensor. Instead, we only compute and return the logprobs of
-    # the topk + 1 tokens.
     BLOCK_SIZE = 1024
     _topk_logprobs_kernel[(batch_size, )](
         logprobs,