fix long dtype in topk sampling (#15049)

vllm/v1/sample/sampler.py

@@ -151,7 +151,7 @@ class Sampler(nn.Module):
                                                  dim=-1)
 
         # Get with the logprob of the prompt or sampled token.
-        token_ids = token_ids.unsqueeze(-1)
+        token_ids = token_ids.unsqueeze(-1).to(torch.long)
         token_logprobs = logprobs.gather(-1, token_ids)
 
         # Compute the ranks of the actual token.