add(v1): RequestStatesStats to RequestOutput (#24947)

Signed-off-by: huijjj <huijong.jeong@squeezebits.com>

tests/entrypoints/llm/test_generate.py

@@ -86,3 +86,16 @@ def test_max_model_len():
         # It can be less if generation finishes due to other reasons (e.g., EOS)
         # before reaching the absolute model length limit.
         assert num_total_tokens <= max_model_len
+
+
+def test_log_stats():
+    llm = LLM(
+        model=MODEL_NAME,
+        disable_log_stats=False,
+        gpu_memory_utilization=0.10,
+        enforce_eager=True,  # reduce test time
+    )
+    outputs = llm.generate(PROMPTS, sampling_params=None)
+
+    # disable_log_stats is False, every output should have metrics
+    assert all(output.metrics is not None for output in outputs)

vllm/outputs.py

@@ -14,6 +14,7 @@ from vllm.logprobs import PromptLogprobs, SampleLogprobs
 from vllm.lora.request import LoRARequest
 from vllm.multimodal.inputs import MultiModalPlaceholderDict
 from vllm.sequence import RequestMetrics
+from vllm.v1.metrics.stats import RequestStateStats
 
 logger = init_logger(__name__)
 
@@ -108,7 +109,7 @@ class RequestOutput:
         prompt_logprobs: Optional[PromptLogprobs],
         outputs: list[CompletionOutput],
         finished: bool,
-        metrics: Optional[RequestMetrics] = None,
+        metrics: Optional[Union[RequestMetrics, RequestStateStats]] = None,
         lora_request: Optional[LoRARequest] = None,
         encoder_prompt: Optional[str] = None,
         encoder_prompt_token_ids: Optional[list[int]] = None,

vllm/v1/engine/output_processor.py

@@ -248,16 +248,15 @@ class RequestState:
         if prompt_token_ids is None and self.prompt_embeds is not None:
             prompt_token_ids = [0] * len(self.prompt_embeds)
 
-        return RequestOutput(
-            request_id=request_id,
-            prompt=self.prompt,
-            prompt_token_ids=prompt_token_ids,
-            prompt_logprobs=prompt_logprobs,
-            outputs=cast(list[CompletionOutput], outputs),
-            finished=finished,
-            kv_transfer_params=kv_transfer_params,
-            num_cached_tokens=self.num_cached_tokens,
-        )
+        return RequestOutput(request_id=request_id,
+                             prompt=self.prompt,
+                             prompt_token_ids=prompt_token_ids,
+                             prompt_logprobs=prompt_logprobs,
+                             outputs=cast(list[CompletionOutput], outputs),
+                             finished=finished,
+                             kv_transfer_params=kv_transfer_params,
+                             num_cached_tokens=self.num_cached_tokens,
+                             metrics=self.stats)
 
     def _new_completion_output(
         self,