[Fronten] Clean up stop_token_ids override for Harmony (#44009)

Signed-off-by: Yifan Zong <yzong@redhat.com>

vllm/entrypoints/openai/chat_completion/serving.py

@@ -57,7 +57,6 @@ from vllm.entrypoints.openai.engine.serving import (
 )
 from vllm.entrypoints.openai.models.serving import OpenAIServingModels
 from vllm.entrypoints.openai.parser.harmony_utils import (
-    get_stop_tokens_for_assistant_actions,
     get_streamable_parser_for_assistant,
     parse_chat_output,
 )
@@ -158,13 +157,6 @@ class OpenAIServingChat(OpenAIServing):
             else getattr(mc, "override_generation_config", {}).get("max_new_tokens")
         )
         self.use_harmony = self.model_config.hf_config.model_type == "gpt_oss"
-        if self.use_harmony:
-            if "stop_token_ids" not in self.default_sampling_params:
-                self.default_sampling_params["stop_token_ids"] = []
-            self.default_sampling_params["stop_token_ids"].extend(
-                get_stop_tokens_for_assistant_actions()
-            )
-
         self.tool_call_id_type = get_tool_call_id_type(self.model_config)
 
         # NOTE(woosuk): While OpenAI's chat completion API supports browsing

vllm/entrypoints/openai/parser/harmony_utils.py

@@ -365,10 +365,6 @@ def render_for_completion(messages: list[Message]) -> list[int]:
     return token_ids
 
 
-def get_stop_tokens_for_assistant_actions() -> list[int]:
-    return get_encoding().stop_tokens_for_assistant_actions()
-
-
 def get_streamable_parser_for_assistant() -> StreamableParser:
     return StreamableParser(get_encoding(), role=Role.ASSISTANT)
 

vllm/entrypoints/openai/responses/protocol.py

@@ -372,8 +372,6 @@ class ResponsesRequest(OpenAIBaseModel):
         if (frequency_penalty := self.frequency_penalty) is None:
             frequency_penalty = default_sampling_params.get("frequency_penalty", 0.0)
 
-        stop_token_ids = default_sampling_params.get("stop_token_ids")
-
         # Structured output
         structured_outputs = self.structured_outputs
 
@@ -409,7 +407,6 @@ class ResponsesRequest(OpenAIBaseModel):
             top_k=top_k,
             max_tokens=max_tokens,
             logprobs=self.top_logprobs if self.is_include_output_logprobs() else None,
-            stop_token_ids=stop_token_ids,
             stop=stop,
             frequency_penalty=frequency_penalty,
             presence_penalty=presence_penalty,

vllm/entrypoints/openai/responses/serving.py

@@ -46,7 +46,6 @@ from vllm.entrypoints.openai.engine.serving import (
 from vllm.entrypoints.openai.models.serving import OpenAIServingModels
 from vllm.entrypoints.openai.parser.harmony_utils import (
     get_developer_message,
-    get_stop_tokens_for_assistant_actions,
     get_system_message,
     get_user_message,
     has_custom_tools,
@@ -222,13 +221,6 @@ class OpenAIServingResponses(OpenAIServing):
                 "For gpt-oss, we ignore --enable-auto-tool-choice "
                 "and always enable tool use."
             )
-            # OpenAI models have two EOS-like tokens: <|return|> and <|call|>.
-            # We need to add them to the stop token ids.
-            if "stop_token_ids" not in self.default_sampling_params:
-                self.default_sampling_params["stop_token_ids"] = []
-            self.default_sampling_params["stop_token_ids"].extend(
-                get_stop_tokens_for_assistant_actions()
-            )
 
         self.tool_call_id_type = get_tool_call_id_type(self.model_config)