[Bugfix] logprobs is not compatible with the OpenAI spec #4795 (#5031)

Author: Etelis

vllm/entrypoints/openai/protocol.py

@@ -109,7 +109,7 @@ class ChatCompletionRequest(OpenAIBaseModel):
     frequency_penalty: Optional[float] = 0.0
     logit_bias: Optional[Dict[str, float]] = None
     logprobs: Optional[bool] = False
-    top_logprobs: Optional[int] = None
+    top_logprobs: Optional[int] = 0
     max_tokens: Optional[int] = None
     n: Optional[int] = 1
     presence_penalty: Optional[float] = 0.0
@@ -192,8 +192,7 @@ class ChatCompletionRequest(OpenAIBaseModel):
     # doc: end-chat-completion-extra-params
 
     def to_sampling_params(self) -> SamplingParams:
-        if self.logprobs and not self.top_logprobs:
-            raise ValueError("Top logprobs must be set when logprobs is.")
+        # We now allow logprobs being true without top_logrobs.
 
         logits_processors = None
         if self.logit_bias:

vllm/entrypoints/openai/serving_chat.py

@@ -286,7 +286,7 @@ async def chat_completion_stream_generator(
                         logprobs = self._create_logprobs(
                             token_ids=delta_token_ids,
                             top_logprobs=top_logprobs,
-                            num_output_top_logprobs=request.logprobs,
+                            num_output_top_logprobs=request.top_logprobs,
                             initial_text_offset=len(previous_texts[i]),
                         )
                     else:
@@ -373,7 +373,7 @@ async def chat_completion_full_generator(
                 logprobs = self._create_logprobs(
                     token_ids=token_ids,
                     top_logprobs=top_logprobs,
-                    num_output_top_logprobs=request.logprobs,
+                    num_output_top_logprobs=request.top_logprobs,
                 )
             else:
                 logprobs = None