[Bugfix][Frontend] Honor stop_token_ids on the Responses API (#56592)

Signed-off-by: YeonwooSung <neos960518@gmail.com>
Signed-off-by: sfeng33 <4florafeng@gmail.com>
Co-authored-by: Flora Feng <4florafeng@gmail.com>

tests/entrypoints/openai/responses/test_sampling_params.py

@@ -47,6 +47,7 @@ class TestResponsesRequestSamplingParams:
             repetition_penalty=1.2,
             seed=42,
             stop=["END", "STOP"],
+            stop_token_ids=[42, 43],
             ignore_eos=True,
             vllm_xargs={"custom": "value"},
         )
@@ -56,6 +57,7 @@ class TestResponsesRequestSamplingParams:
         assert sampling_params.repetition_penalty == 1.2
         assert sampling_params.seed == 42
         assert sampling_params.stop == ["END", "STOP"]
+        assert sampling_params.stop_token_ids == [42, 43]
         assert sampling_params.ignore_eos is True
         assert sampling_params.extra_args == {"custom": "value"}
 

vllm/entrypoints/openai/responses/protocol.py

@@ -363,6 +363,7 @@ class ResponsesRequest(OpenAIBaseModel):
     repetition_penalty: float | None = None
     seed: int | None = Field(None, ge=_INT64_MIN, le=_INT64_MAX)
     stop: StopParam = []
+    stop_token_ids: list[int] | None = []
     ignore_eos: bool = False
     vllm_xargs: dict[str, str | int | float | list[str | int | float]] | None = Field(
         default=None,
@@ -521,6 +522,7 @@ class ResponsesRequest(OpenAIBaseModel):
             max_tokens=max_tokens,
             logprobs=self.top_logprobs if self.is_include_output_logprobs() else None,
             stop=stop,
+            stop_token_ids=self.stop_token_ids,
             frequency_penalty=frequency_penalty,
             presence_penalty=presence_penalty,
             repetition_penalty=repetition_penalty,