[Bugfix] Propagate default stop_token_ids to per-request SamplingParams (#35076)

Signed-off-by: sriganesh123 <arjulasriganesh@gmail.com>

tests/entrypoints/openai/test_stop_token_ids.py

@@ -0,0 +1,160 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+Unit tests for stop_token_ids propagation from default_sampling_params
+to SamplingParams in ChatCompletionRequest and CompletionRequest.
+
+Regression test for https://github.com/vllm-project/vllm/issues/22519
+where gpt-oss model stop tokens (e.g., </call> = 200012) were loaded into
+default_sampling_params at server startup but silently discarded on every
+request because to_sampling_params() never fell back to defaults.
+"""
+
+import pytest
+
+from vllm.entrypoints.openai.chat_completion.protocol import (
+    ChatCompletionRequest,
+)
+from vllm.entrypoints.openai.completion.protocol import (
+    CompletionRequest,
+)
+
+
+class TestChatCompletionStopTokenIds:
+    """Test stop_token_ids merging in ChatCompletionRequest.to_sampling_params()."""
+
+    @pytest.fixture
+    def minimal_chat_request(self):
+        return ChatCompletionRequest(
+            model="test-model",
+            messages=[{"role": "user", "content": "hello"}],
+        )
+
+    def test_default_stop_token_ids_applied(self, minimal_chat_request):
+        """Server-default stop_token_ids are applied when client sends none."""
+        default_sampling_params = {
+            "stop_token_ids": [200012, 200002],
+        }
+
+        sampling_params = minimal_chat_request.to_sampling_params(
+            max_tokens=100,
+            default_sampling_params=default_sampling_params,
+        )
+
+        assert set(sampling_params.stop_token_ids) == {200012, 200002}
+
+    def test_client_stop_token_ids_merged_with_defaults(self):
+        """Client-specified stop_token_ids are merged with server defaults."""
+        request = ChatCompletionRequest(
+            model="test-model",
+            messages=[{"role": "user", "content": "hello"}],
+            stop_token_ids=[99999],
+        )
+        default_sampling_params = {
+            "stop_token_ids": [200012, 200002],
+        }
+
+        sampling_params = request.to_sampling_params(
+            max_tokens=100,
+            default_sampling_params=default_sampling_params,
+        )
+
+        assert set(sampling_params.stop_token_ids) == {200012, 200002, 99999}
+        assert sampling_params.stop_token_ids == [99999, 200012, 200002]
+
+    def test_no_stop_token_ids_anywhere(self, minimal_chat_request):
+        """When neither client nor server specifies stop_token_ids, result is empty."""
+        sampling_params = minimal_chat_request.to_sampling_params(
+            max_tokens=100,
+            default_sampling_params={},
+        )
+
+        assert not sampling_params.stop_token_ids
+
+    def test_only_client_stop_token_ids(self):
+        """Client stop_token_ids work when no server defaults exist."""
+        request = ChatCompletionRequest(
+            model="test-model",
+            messages=[{"role": "user", "content": "hello"}],
+            stop_token_ids=[42, 43],
+        )
+
+        sampling_params = request.to_sampling_params(
+            max_tokens=100,
+            default_sampling_params={},
+        )
+
+        assert set(sampling_params.stop_token_ids) == {42, 43}
+
+    def test_duplicate_stop_token_ids_deduplicated(self):
+        """Overlapping stop_token_ids between client and server are deduplicated."""
+        request = ChatCompletionRequest(
+            model="test-model",
+            messages=[{"role": "user", "content": "hello"}],
+            stop_token_ids=[200012, 55555],
+        )
+        default_sampling_params = {
+            "stop_token_ids": [200012, 200002],
+        }
+
+        sampling_params = request.to_sampling_params(
+            max_tokens=100,
+            default_sampling_params=default_sampling_params,
+        )
+
+        assert set(sampling_params.stop_token_ids) == {200012, 200002, 55555}
+        assert sampling_params.stop_token_ids == [200012, 55555, 200002]
+        assert len(sampling_params.stop_token_ids) == 3
+
+
+class TestCompletionStopTokenIds:
+    """Test stop_token_ids merging in CompletionRequest.to_sampling_params()."""
+
+    @pytest.fixture
+    def minimal_completion_request(self):
+        return CompletionRequest(
+            model="test-model",
+            prompt="hello",
+        )
+
+    def test_default_stop_token_ids_applied(self, minimal_completion_request):
+        """Server-default stop_token_ids are applied when client sends none."""
+        default_sampling_params = {
+            "stop_token_ids": [200012, 200002],
+        }
+
+        sampling_params = minimal_completion_request.to_sampling_params(
+            max_tokens=100,
+            default_sampling_params=default_sampling_params,
+        )
+
+        assert set(sampling_params.stop_token_ids) == {200012, 200002}
+
+    def test_client_stop_token_ids_merged_with_defaults(self):
+        """Client-specified stop_token_ids are merged with server defaults."""
+        request = CompletionRequest(
+            model="test-model",
+            prompt="hello",
+            stop_token_ids=[99999],
+        )
+        default_sampling_params = {
+            "stop_token_ids": [200012, 200002],
+        }
+
+        sampling_params = request.to_sampling_params(
+            max_tokens=100,
+            default_sampling_params=default_sampling_params,
+        )
+
+        assert set(sampling_params.stop_token_ids) == {200012, 200002, 99999}
+        assert sampling_params.stop_token_ids == [99999, 200012, 200002]
+
+    def test_no_stop_token_ids_anywhere(self, minimal_completion_request):
+        """When neither client nor server specifies stop_token_ids, result is empty."""
+        sampling_params = minimal_completion_request.to_sampling_params(
+            max_tokens=100,
+            default_sampling_params={},
+        )
+
+        assert not sampling_params.stop_token_ids

vllm/entrypoints/openai/chat_completion/protocol.py

@@ -610,6 +610,18 @@ class ChatCompletionRequest(OpenAIBaseModel):
                 "min_p", self._DEFAULT_SAMPLING_PARAMS["min_p"]
             )
 
+        # Merge server-default stop_token_ids (e.g., model-specific tokens
+        # like </call> for gpt-oss) with any request-specified ones
+        stop_token_ids = self.stop_token_ids
+        default_stop_ids = default_sampling_params.get("stop_token_ids")
+        if default_stop_ids:
+            if not stop_token_ids:
+                stop_token_ids = list(default_stop_ids)
+            else:
+                stop_token_ids = list(
+                    dict.fromkeys([*stop_token_ids, *default_stop_ids])
+                )
+
         prompt_logprobs = self.prompt_logprobs
         if prompt_logprobs is None and self.echo:
             prompt_logprobs = self.top_logprobs
@@ -661,7 +673,7 @@ class ChatCompletionRequest(OpenAIBaseModel):
             min_p=min_p,
             seed=self.seed,
             stop=self.stop,
-            stop_token_ids=self.stop_token_ids,
+            stop_token_ids=stop_token_ids,
             logprobs=self.top_logprobs if self.logprobs else None,
             prompt_logprobs=prompt_logprobs,
             ignore_eos=self.ignore_eos,

vllm/entrypoints/openai/completion/protocol.py

@@ -288,6 +288,18 @@ class CompletionRequest(OpenAIBaseModel):
                 "min_p", self._DEFAULT_SAMPLING_PARAMS["min_p"]
             )
 
+        # Merge server-default stop_token_ids (e.g., model-specific tokens
+        # like </call> for gpt-oss) with any request-specified ones
+        stop_token_ids = self.stop_token_ids
+        default_stop_ids = default_sampling_params.get("stop_token_ids")
+        if default_stop_ids:
+            if not stop_token_ids:
+                stop_token_ids = list(default_stop_ids)
+            else:
+                stop_token_ids = list(
+                    dict.fromkeys([*stop_token_ids, *default_stop_ids])
+                )
+
         prompt_logprobs = self.prompt_logprobs
         if prompt_logprobs is None and self.echo:
             prompt_logprobs = self.logprobs
@@ -341,7 +353,7 @@ class CompletionRequest(OpenAIBaseModel):
             min_p=min_p,
             seed=self.seed,
             stop=self.stop,
-            stop_token_ids=self.stop_token_ids,
+            stop_token_ids=stop_token_ids,
             logprobs=self.logprobs,
             ignore_eos=self.ignore_eos,
             max_tokens=max_tokens if not echo_without_generation else 1,