[Bugfix] Fix vLLM UsageInfo and logprobs None AssertionError with empty token_ids (#9034)

Co-authored-by: Nick Hill <nickhill@us.ibm.com>

vllm/entrypoints/openai/serving_chat.py

@@ -435,6 +435,12 @@ class OpenAIServingChat(OpenAIServing):
                         logprobs = None
 
                     delta_text = output.text
+
+                    if not delta_text and not output.token_ids and \
+                        not previous_num_tokens[i]:
+                        # Chunked prefill case, don't return empty chunks
+                        continue
+
                     delta_message: Optional[DeltaMessage]
 
                     # handle streaming deltas for tools with named tool_choice

vllm/entrypoints/openai/serving_completion.py

@@ -274,8 +274,6 @@ class OpenAIServingCompletion(OpenAIServing):
 
                 for output in res.outputs:
                     i = output.index + prompt_idx * num_choices
-                    # TODO(simon): optimize the performance by avoiding full
-                    # text O(n^2) sending.
 
                     assert request.max_tokens is not None
                     if request.echo and request.max_tokens == 0:
@@ -307,6 +305,11 @@ class OpenAIServingCompletion(OpenAIServing):
                         delta_token_ids = output.token_ids
                         out_logprobs = output.logprobs
 
+                        if not delta_text and not delta_token_ids \
+                            and not previous_num_tokens[i]:
+                            # Chunked prefill case, don't return empty chunks
+                            continue
+
                     if request.logprobs is not None:
                         assert out_logprobs is not None, (
                             "Did not output logprobs")