[ BugFix ] Prompt Logprobs Detokenization

vllm/engine/output_processor/single_step.py

@@ -65,10 +65,9 @@ def process_prompt_logprob(self, seq_group: SequenceGroup,
                 self.detokenizer.decode_prompt_logprobs_inplace(
                     seq_group, prompt_logprobs)
             if not seq_group.prompt_logprobs:
-                # The first prompt token's logprob is None because it doesn't
-                # have tokens that are precedent.
-                seq_group.prompt_logprobs = [None]
-            seq_group.prompt_logprobs.extend(prompt_logprobs)
+                seq_group.prompt_logprobs = prompt_logprobs
+            else:
+                seq_group.prompt_logprobs.extend(prompt_logprobs)
 
     def _process_sequence_group_outputs(self, seq_group: SequenceGroup,
                                         outputs: SequenceGroupOutput) -> None:

vllm/model_executor/layers/sampler.py

@@ -836,7 +836,7 @@ def _get_prompt_logprob_if_needed(
     # Find prompt logprobs
     prompt_logprobs: Optional[PromptLogprobs] = None
     if is_prompt and sampling_params.prompt_logprobs is not None:
-        prompt_logprobs = []
+        prompt_logprobs = [None]
         num_logprobs = sampling_params.prompt_logprobs
         next_prompt_tokens = _get_next_prompt_tokens(seq_group)
         # Pre-select indexes and create a list. It is faster than calling .item