diff --git a/vllm/entrypoints/openai/chat_completion/protocol.py b/vllm/entrypoints/openai/chat_completion/protocol.py index 742f9cced6..03e4678c82 100644 --- a/vllm/entrypoints/openai/chat_completion/protocol.py +++ b/vllm/entrypoints/openai/chat_completion/protocol.py @@ -111,6 +111,9 @@ class ChatCompletionResponse(OpenAIBaseModel): # vLLM-specific fields that are not in OpenAI spec prompt_logprobs: list[dict[int, Logprob] | None] | None = None prompt_token_ids: list[int] | None = None + # Rendered prompt text from chat templating (only set when + # ``return_prompt_text=True`` on the request). + prompt_text: str | None = None kv_transfer_params: dict[str, Any] | None = Field( default=None, description="KVTransfer parameters." ) @@ -138,6 +141,9 @@ class ChatCompletionStreamResponse(OpenAIBaseModel): system_fingerprint: str | None = None # not part of the OpenAI spec but for tracing the tokens prompt_token_ids: list[int] | None = None + # Rendered prompt text from chat templating (only set when + # ``return_prompt_text=True`` on the request); only sent on the first chunk. + prompt_text: str | None = None class ChatCompletionToolsParam(OpenAIBaseModel): @@ -352,6 +358,15 @@ class ChatCompletionRequest(OpenAIBaseModel): "need to map generated text back to input tokens." ), ) + return_prompt_text: bool | None = Field( + default=None, + description=( + "If true, the response will include ``prompt_text`` containing the " + "prompt string produced by chat templating. In streaming mode it " + "is sent only on the first chunk. This is useful for inspecting " + "exactly what was fed into the model." + ), + ) cache_salt: str | None = Field( default=None, diff --git a/vllm/entrypoints/openai/chat_completion/serving.py b/vllm/entrypoints/openai/chat_completion/serving.py index 375fe9794c..10bde68478 100644 --- a/vllm/entrypoints/openai/chat_completion/serving.py +++ b/vllm/entrypoints/openai/chat_completion/serving.py @@ -508,6 +508,9 @@ class OpenAIServingChat(OpenAIServing): # the role role = self.get_chat_request_role(request) + # ``res.prompt`` is the rendered chat-templated prompt + prompt_text = res.prompt if request.return_prompt_text else None + # NOTE num_choices defaults to 1 so this usually executes # once per request for i in range(num_choices): @@ -533,6 +536,7 @@ class OpenAIServingChat(OpenAIServing): if request.return_token_ids else None ), + prompt_text=prompt_text, ) # if continuous usage stats are requested, add it @@ -1371,6 +1375,9 @@ class OpenAIServingChat(OpenAIServing): if final_res.prompt_routed_experts is not None: prompt_routed_experts = final_res.prompt_routed_experts.tolist() + # ``final_res.prompt`` is the rendered chat-templated prompt text + prompt_text = final_res.prompt if request.return_prompt_text else None + response = ChatCompletionResponse( id=request_id, created=created_time, @@ -1382,6 +1389,7 @@ class OpenAIServingChat(OpenAIServing): prompt_token_ids=( final_res.prompt_token_ids if request.return_token_ids else None ), + prompt_text=prompt_text, kv_transfer_params=final_res.kv_transfer_params, prompt_routed_experts=prompt_routed_experts, )