[Misc] Minor code cleanup for _get_prompt_logprobs_dict (#23064)

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

Author: WoosukKwon

vllm/v1/worker/gpu_model_runner.py

@@ -1722,7 +1722,7 @@ def execute_model(
         # Compute prompt logprobs if needed.
         prompt_logprobs_dict = self._get_prompt_logprobs_dict(
             hidden_states[:num_scheduled_tokens],
-            scheduler_output,
+            scheduler_output.num_scheduled_tokens,
         )
 
         # Get the valid generated tokens.
@@ -2064,7 +2064,7 @@ def save_tensorized_model(
     def _get_prompt_logprobs_dict(
         self,
         hidden_states: torch.Tensor,
-        scheduler_output: "SchedulerOutput",
+        num_scheduled_tokens: dict[str, int],
     ) -> dict[str, Optional[LogprobsTensors]]:
         num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
         if not num_prompt_logprobs_dict:
@@ -2077,8 +2077,7 @@ def _get_prompt_logprobs_dict(
         # maintainable loop over optimal performance.
         completed_prefill_reqs = []
         for req_id, num_prompt_logprobs in num_prompt_logprobs_dict.items():
-
-            num_tokens = scheduler_output.num_scheduled_tokens[req_id]
+            num_tokens = num_scheduled_tokens[req_id]
 
             # Get metadata for this request.
             request = self.requests[req_id]