Fix completion API echo and logprob combo (#1992)

Author: simon-mo

vllm/entrypoints/openai/api_server.py

@@ -567,17 +567,22 @@ async def completion_stream_generator() -> AsyncGenerator[str, None]:
                 i = output.index
                 delta_text = output.text[len(previous_texts[i]):]
                 token_ids = output.token_ids[previous_num_tokens[i]:]
-                top_logprobs = output.logprobs[previous_num_tokens[i]:]
+                if request.logprobs is not None:
+                    top_logprobs = output.logprobs[previous_num_tokens[i]:]
+                else:
+                    top_logprobs = None
                 offsets = len(previous_texts[i])
                 if request.echo and not has_echoed[i]:
                     if not echo_without_generation:
                         delta_text = res.prompt + delta_text
                         token_ids = res.prompt_token_ids + token_ids
-                        top_logprobs = res.prompt_logprobs + top_logprobs
-                    else:
+                        if top_logprobs:
+                            top_logprobs = res.prompt_logprobs + top_logprobs
+                    else:  # only just return the prompt
                         delta_text = res.prompt
                         token_ids = res.prompt_token_ids
-                        top_logprobs = res.prompt_logprobs
+                        if top_logprobs:
+                            top_logprobs = res.prompt_logprobs
                     has_echoed[i] = True
                 if request.logprobs is not None:
                     logprobs = create_logprobs(