diff --git a/rust/src/server/src/routes/openai/chat_completions.rs b/rust/src/server/src/routes/openai/chat_completions.rs index bd86a06136..8a1e9c4383 100644 --- a/rust/src/server/src/routes/openai/chat_completions.rs +++ b/rust/src/server/src/routes/openai/chat_completions.rs @@ -131,6 +131,7 @@ async fn collect_chat_completion( echo, return_token_ids, return_tokens_as_token_ids, + is_named_tool_choice, }: ResponseOptions, ) -> Result { let collected = stream.collect_message().await.map_err(|error| { @@ -157,7 +158,9 @@ async fn collect_chat_completion( // When reasoning is hidden, omit them rather than leaking hidden reasoning // tokens through per-token metadata. let include_output_metadata = include_reasoning || reasoning.is_none(); - let finish_reason = chat_finish_reason_to_openai(&finish_reason, saw_tool_calls)?.to_string(); + let finish_reason = + chat_finish_reason_to_openai(&finish_reason, saw_tool_calls && !is_named_tool_choice)? + .to_string(); let tool_calls = message .tool_calls() .map(|call| ToolCall { @@ -254,6 +257,7 @@ async fn chat_completion_chunk_stream( echo, return_token_ids, return_tokens_as_token_ids, + is_named_tool_choice, }: ResponseOptions, mut y: TryYielder, ) -> Result<(), ApiError> { @@ -454,7 +458,7 @@ async fn chat_completion_chunk_stream( &response_model, created, finish_reason, - saw_tool_calls, + saw_tool_calls && !is_named_tool_choice, ) { Ok(chunk) => yield_chunk!(chunk), Err(error) => { @@ -787,10 +791,10 @@ fn final_chunk( response_model: &str, created: u64, finish_reason: FinishReason, - saw_tool_calls: bool, + use_tool_calls_finish_reason: bool, ) -> Result { let stop_reason = finish_reason.as_stop_reason().map(stop_reason_to_json); - let finish_reason = chat_finish_reason_to_openai(&finish_reason, saw_tool_calls)?; + let finish_reason = chat_finish_reason_to_openai(&finish_reason, use_tool_calls_finish_reason)?; debug!( finish_reason = %finish_reason, @@ -809,10 +813,10 @@ fn final_chunk( fn chat_finish_reason_to_openai( finish_reason: &FinishReason, - saw_tool_calls: bool, + use_tool_calls_finish_reason: bool, ) -> Result<&'static str, ApiError> { match finish_reason { - FinishReason::Stop(_) if saw_tool_calls => Ok("tool_calls"), + FinishReason::Stop(_) if use_tool_calls_finish_reason => Ok("tool_calls"), FinishReason::Stop(_) => Ok("stop"), FinishReason::Length => Ok("length"), FinishReason::Abort => Ok("abort"), diff --git a/rust/src/server/src/routes/openai/chat_completions/convert.rs b/rust/src/server/src/routes/openai/chat_completions/convert.rs index 5b2fa3c19e..d75f3fab16 100644 --- a/rust/src/server/src/routes/openai/chat_completions/convert.rs +++ b/rust/src/server/src/routes/openai/chat_completions/convert.rs @@ -54,6 +54,8 @@ pub(super) struct ResponseOptions { pub return_token_ids: bool, /// Whether to format logprob tokens as `token_id:{id}`. pub return_tokens_as_token_ids: bool, + /// Whether the request forces one named function tool. + pub is_named_tool_choice: bool, } /// Validate and lower one OpenAI chat completion request into the internal chat @@ -98,6 +100,7 @@ pub(super) fn prepare_chat_request( .and_then(|options| options.continuous_usage_stats) .unwrap_or(false); let requested_logprobs = request.logprobs; + let is_named_tool_choice = matches!(&request.tool_choice, Some(ToolChoice::Function { .. })); // Auto-enable prompt logprobs for non-streaming echo, matching Python vLLM's // behavior. @@ -180,6 +183,7 @@ pub(super) fn prepare_chat_request( echo, return_token_ids: request.return_token_ids.unwrap_or(false), return_tokens_as_token_ids: request.return_tokens_as_token_ids.unwrap_or(false), + is_named_tool_choice, }, chat_request, }) @@ -1068,6 +1072,7 @@ mod tests { .expect("request is valid"); assert_eq!(prepared.chat_request.tool_choice, ChatToolChoice::Required); + assert!(!prepared.options.is_named_tool_choice); } #[test] @@ -1107,6 +1112,7 @@ mod tests { name: "get_weather".to_string(), } ); + assert!(prepared.options.is_named_tool_choice); } #[test] diff --git a/rust/src/server/src/routes/tests.rs b/rust/src/server/src/routes/tests.rs index 64e904b562..1c228f99f1 100644 --- a/rust/src/server/src/routes/tests.rs +++ b/rust/src/server/src/routes/tests.rs @@ -153,6 +153,20 @@ fn default_stream_output_specs() -> Vec<(Vec, Option Vec<(Vec, Option)> { + vec![ + (bytes_to_token_ids(b"Need tool."), None), + ( + bytes_to_token_ids(b"\n{\"name\":\"get_weather\", "), + None, + ), + ( + bytes_to_token_ids(b"\"arguments\":{\"city\":\"Paris\"}}\n"), + Some(EngineCoreFinishReason::Stop), + ), + ] +} + fn assert_adapter_a_lora_request(request: &EngineCoreRequest) { let lora = request.lora_request.as_ref().expect("lora request"); assert_eq!(lora.lora_name, "adapter-a"); @@ -4554,17 +4568,7 @@ async fn include_reasoning_false_suppresses_non_stream_output_metadata() { async fn tool_calls_are_mapped_to_tool_call_sse_chunks() { let (app, engine_task) = test_app_with_backend_and_stream_output_specs( Arc::new(FakeChatBackend::with_model_id("Qwen/Qwen3-0.6B")), - vec![ - (bytes_to_token_ids(b"Need tool."), None), - ( - bytes_to_token_ids(b"\n{\"name\":\"get_weather\", "), - None, - ), - ( - bytes_to_token_ids(b"\"arguments\":{\"city\":\"Paris\"}}\n"), - Some(EngineCoreFinishReason::Stop), - ), - ], + weather_tool_call_output_specs(), ) .await; @@ -4613,6 +4617,63 @@ async fn tool_calls_are_mapped_to_tool_call_sse_chunks() { assert!(text.contains("\"finish_reason\":\"tool_calls\""), "{text}"); } +#[tokio::test(flavor = "multi_thread", worker_threads = 2)] +#[serial] +async fn named_tool_choice_uses_stop_finish_reason() { + for stream in [false, true] { + let (app, engine_task) = test_app_with_backend_and_stream_output_specs( + Arc::new(FakeChatBackend::with_model_id("Qwen/Qwen3-0.6B")), + weather_tool_call_output_specs(), + ) + .await; + + let response = app + .clone() + .call( + Request::builder() + .method("POST") + .uri("/v1/chat/completions") + .header("content-type", "application/json") + .body(Body::from( + json!({ + "model": "Qwen/Qwen1.5-0.5B-Chat", + "stream": stream, + "messages": [{"role": "user", "content": "hello"}], + "tools": [{ + "type": "function", + "function": { + "name": "get_weather", + "description": "Get weather", + "parameters": { + "type": "object", + "properties": {"city": {"type": "string"}} + } + } + }], + "tool_choice": { + "type": "function", + "function": {"name": "get_weather"} + } + }) + .to_string(), + )) + .expect("build request"), + ) + .await + .expect("call app"); + + assert_eq!(response.status(), StatusCode::OK); + let body = to_bytes(response.into_body(), usize::MAX).await.expect("read body"); + engine_task.await.expect("mock engine task"); + let text = String::from_utf8(body.to_vec()).expect("utf8 body"); + + assert!(text.contains("\"tool_calls\":"), "{text}"); + assert!(text.contains("\"name\":\"get_weather\""), "{text}"); + assert!(text.contains("\"finish_reason\":\"stop\""), "{text}"); + assert!(!text.contains("\"finish_reason\":\"tool_calls\""), "{text}"); + } +} + #[tokio::test(flavor = "multi_thread", worker_threads = 2)] #[serial] async fn tool_call_sse_chunks_can_carry_logprobs() {