fix: no more multiple finish reasons in stream (#4154)

Signed-off-by: ayushag <ayushag@nvidia.com>

lib/llm/src/preprocessor.rs

@@ -764,7 +764,7 @@ impl OpenAIPreprocessor {
         let jail = JailedStream::builder()
             .tool_call_parser(tool_call_parser)
             .build();
-        jail.apply(stream)
+        jail.apply_with_finish_reason(stream)
     }
 
     // Motivation: Each transformation on the stream should be a separate step to allow for more flexibility

lib/llm/src/protocols/openai/chat_completions/jail.rs

@@ -13,6 +13,7 @@ use dynamo_parsers::tool_calling::{
 };
 use dynamo_runtime::protocols::annotated::Annotated;
 use futures::{Stream, StreamExt};
+use std::collections::HashMap;
 
 use crate::utils::{MarkerMatcher, MatchResult};
 
@@ -72,6 +73,8 @@ struct ChoiceJailState {
     accumulated_content: String,
     /// Buffer for partial marker matches across chunks
     partial_match_buffer: String,
+    /// Stream finish reason
+    stream_finish_reason: Option<FinishReason>,
 }
 
 fn create_choice_stream(
@@ -106,6 +109,7 @@ impl ChoiceJailState {
             is_jailed: false,
             accumulated_content: String::new(),
             partial_match_buffer: String::new(),
+            stream_finish_reason: None,
         }
     }
 
@@ -130,7 +134,6 @@ impl ChoiceJailState {
         jail_stream: &JailedStream,
     ) -> Vec<ChoiceEmission> {
         let mut emissions = Vec::new();
-
         if !self.is_jailed {
             // Use the marker matcher to detect complete/partial markers
             let match_result = jail_stream
@@ -152,7 +155,7 @@ impl ChoiceJailState {
                             choice.delta.role,
                             &prefix,
                             None,
-                            None,
+                            choice.finish_reason,
                             choice.logprobs.clone(),
                         );
                         emissions.push(ChoiceEmission::PassThrough(prefix_choice));
@@ -192,7 +195,7 @@ impl ChoiceJailState {
                                 choice.delta.role,
                                 trailing_part,
                                 None,
-                                None,
+                                choice.finish_reason,
                                 choice.logprobs.clone(),
                             );
                             emissions.push(ChoiceEmission::Trailing(trailing_choice));
@@ -224,7 +227,7 @@ impl ChoiceJailState {
                             choice.delta.role,
                             &prefix,
                             None,
-                            None,
+                            choice.finish_reason,
                             choice.logprobs.clone(),
                         );
                         emissions.push(ChoiceEmission::PassThrough(prefix_choice));
@@ -267,7 +270,7 @@ impl ChoiceJailState {
                                 choice.delta.role,
                                 &content,
                                 None,
-                                None,
+                                choice.finish_reason,
                                 choice.logprobs.clone(),
                             );
                             emissions.push(ChoiceEmission::PassThrough(pass_through_choice));
@@ -312,7 +315,7 @@ impl ChoiceJailState {
                         choice.delta.role,
                         trailing_part,
                         None,
-                        None,
+                        choice.finish_reason,
                         choice.logprobs.clone(),
                     );
                     emissions.push(ChoiceEmission::Trailing(trailing_choice));
@@ -323,7 +326,6 @@ impl ChoiceJailState {
             }
             // If not unjailing, don't emit anything (still accumulating)
         }
-
         emissions
     }
 
@@ -342,7 +344,7 @@ impl ChoiceJailState {
                 Some(Role::Assistant),
                 &self.accumulated_content,
                 None,
-                None,
+                self.stream_finish_reason, // For the accumulated content, assign the original stream finish reason, otherwise it will get lost
                 None,
             );
 
@@ -428,6 +430,19 @@ impl JailedStream {
         JailedStreamBuilder::new()
     }
 
+    /// Apply jail stream transformation with finish_reason fix
+    /// This is a convenience method that applies both apply() and fix_finish_reason()
+    pub fn apply_with_finish_reason<S>(
+        self,
+        stream: S,
+    ) -> impl Stream<Item = Annotated<NvCreateChatCompletionStreamResponse>> + Send
+    where
+        S: Stream<Item = Annotated<NvCreateChatCompletionStreamResponse>> + Send + 'static,
+    {
+        let jailed_stream = self.apply(stream);
+        JailedStream::fix_finish_reason(jailed_stream)
+    }
+
     /// Apply the jail transformation to a stream of chat completion responses
     /// Consumes self and returns the transformed stream
     pub fn apply<S>(
@@ -449,6 +464,7 @@ impl JailedStream {
             // Pin the stream for iteration (stack pinning is more efficient)
             tokio::pin!(stream);
 
+
             // Process each item in the stream
             while let Some(response) = stream.next().await {
                 if let Some(chat_response) = response.data.as_ref() {
@@ -467,6 +483,9 @@ impl JailedStream {
                                     last_annotated_comment = response.comment.clone();
                                 }
 
+                            // Track actual stream finish reason in the choice state
+                            choice_state.stream_finish_reason = choice.finish_reason;
+
                             // Process this choice and get emissions
                             let emissions = choice_state.process_content(choice, content, &self).await;
                             all_emissions.extend(emissions);
@@ -707,16 +726,16 @@ impl JailedStream {
                     }),
                 })
                 .collect();
-
             // Create choice with tool calls
-            return create_choice_stream(
+            let choice = create_choice_stream(
                 choice_index,
                 Some(Role::Assistant),
                 normal_text.as_deref().unwrap_or(""),
                 Some(tool_call_chunks),
-                Some(FinishReason::ToolCalls),
+                None,
                 None,
             );
+            return choice;
         }
 
         // No tool calls found or parsing failed, return content choice
@@ -725,7 +744,7 @@ impl JailedStream {
             Some(Role::Assistant),
             accumulated_content,
             None,
-            None,
+            base_choice.finish_reason,
             base_choice.logprobs.clone(),
         )
     }
@@ -745,6 +764,44 @@ impl JailedStream {
         }
         false
     }
+
+    /// Post-processor that sets finish_reason to ToolCalls when tool calls were emitted
+    /// This should be called after apply() to fix the finish_reason for tool call chunks
+    pub fn fix_finish_reason<S>(
+        input_stream: S,
+    ) -> impl Stream<Item = Annotated<NvCreateChatCompletionStreamResponse>> + Send
+    where
+        S: Stream<Item = Annotated<NvCreateChatCompletionStreamResponse>> + Send + 'static,
+    {
+        stream! {
+            tokio::pin!(input_stream);
+            let mut has_tool_calls_per_choice: HashMap<u32, bool> = HashMap::new();
+
+            while let Some(mut response) = input_stream.next().await {
+                // Track if any choice emitted tool calls
+                if let Some(ref data) = response.data {
+                    for choice in &data.choices {
+                        if choice.delta.tool_calls.is_some() {
+                            has_tool_calls_per_choice.insert(choice.index, true);
+                        }
+                    }
+                }
+
+                // If this chunk has finish_reason and the choice had tool calls, override to ToolCalls
+                if let Some(ref mut data) = response.data {
+                    for choice in &mut data.choices {
+                        if choice.finish_reason.is_some() && choice.finish_reason == Some(FinishReason::Stop)
+                            && has_tool_calls_per_choice.get(&choice.index).copied().unwrap_or(false)
+                        {
+                            choice.finish_reason = Some(FinishReason::ToolCalls);
+                        }
+                    }
+                }
+
+                yield response;
+            }
+        }
+    }
 }
 
 /// Builder for configuring a JailedStream

lib/llm/tests/data/vllm/qwen3-0.6B/chat_completion_incomplete_tool.json

+{
+    "request_id": "8f33c28b-cb52-4272-9ac5-0cb9f80386d3",
+    "expected_output": {
+        "normal_content": " the requested format.\n</think>\n\n<tool_call>\n\n{\"name\":\"get"
+      },
+    "input_stream": [
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":" the","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":" requested","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":" format","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":".\n","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":"</think>","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":"\n\n","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":"<tool_call>","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":"\n","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":"{\"","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":"name","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":"\":","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":" \"","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":"get","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}, "finish_reason":"length"}]}}
+    ]
+}

lib/llm/tests/data/vllm/qwen3-0.6B/chat_completion_stream_finish_length.json

+{
+    "request_id": "8f33c28b-cb52-4272-9ac5-0cb9f80386d3",
+    "expected_output": {
+        "normal_content": "<think>\nOkay, the user is asking for the weather in San Francisco in"
+      },
+    "input_stream": [
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":"<think>","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":"\n","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":"Okay","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":",","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":" the","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":" user","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":" is","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":" asking","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":" for","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":" the","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":" weather","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null}}]}},
+      {"data":{"id":"chatcmpl-8f33c28b-cb52-4272-9ac5-0cb9f80386d3","choices":[{"index":0,"delta":{"content":" in","function_call":null,"tool_calls":null,"role":"assistant","refusal":null,"reasoning_content":null},"finish_reason":"length"}]}}
+    ]
+}

lib/llm/tests/test_streaming_tool_parsers.rs

@@ -26,7 +26,7 @@ across backends.
 
 */
 
-use dynamo_async_openai::types::ChatChoiceStream;
+use dynamo_async_openai::types::{ChatChoiceStream, FinishReason};
 use dynamo_llm::preprocessor::OpenAIPreprocessor;
 use dynamo_llm::protocols::openai::chat_completions::NvCreateChatCompletionStreamResponse;
 use dynamo_runtime::protocols::annotated::Annotated;
@@ -251,6 +251,71 @@ fn aggregate_content_from_chunks(
     }
 }
 
+/// Helper function to validate finish_reason in the stream
+/// Returns true if:
+/// 1. There is exactly one finish_reason in the entire stream
+/// 2. The finish_reason is in the last chunk
+/// 3. The finish_reason matches the expected value
+fn validate_finish_reason(
+    chunks: &[Annotated<NvCreateChatCompletionStreamResponse>],
+    expected_finish_reason: FinishReason,
+) -> bool {
+    let mut finish_reason_count = 0;
+    let mut last_chunk_index = None;
+    let mut finish_reason_value = None;
+
+    // Count finish_reason occurrences and track position
+    for (idx, chunk) in chunks.iter().enumerate() {
+        if let Some(ref response_data) = chunk.data {
+            for choice in &response_data.choices {
+                if let Some(reason) = choice.finish_reason {
+                    finish_reason_count += 1;
+                    last_chunk_index = Some(idx);
+                    finish_reason_value = Some(reason);
+                }
+            }
+        }
+    }
+
+    // Validate:
+    // 1. Exactly one finish_reason in the stream
+    if finish_reason_count != 1 {
+        eprintln!(
+            "Expected exactly 1 finish_reason, but found {}",
+            finish_reason_count
+        );
+        return false;
+    }
+
+    // 2. finish_reason is in the last chunk
+    if let Some(idx) = last_chunk_index {
+        if idx != chunks.len() - 1 {
+            eprintln!(
+                "Expected finish_reason in last chunk (index {}), but found at index {}",
+                chunks.len() - 1,
+                idx
+            );
+            return false;
+        }
+    } else {
+        eprintln!("No finish_reason found in stream");
+        return false;
+    }
+
+    // 3. finish_reason matches expected value
+    if let Some(reason) = finish_reason_value
+        && reason != expected_finish_reason
+    {
+        eprintln!(
+            "Expected finish_reason {:?}, but found {:?}",
+            expected_finish_reason, reason
+        );
+        return false;
+    }
+
+    true
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -304,6 +369,12 @@ mod tests {
             aggregated.has_tool_calls, expected_has_tool_calls,
             "Tool calls presence should match expected value"
         );
+
+        // Verify finish_reason is valid: exactly one occurrence, in last chunk, and is Stop
+        assert!(
+            validate_finish_reason(&output_chunks, FinishReason::Stop),
+            "finish_reason validation failed for non-tool call case"
+        );
     }
 
     #[tokio::test]
@@ -360,6 +431,12 @@ mod tests {
 
         // Verify tool calls
         assert_tool_calls(&aggregated.tool_calls, &test_data.expected_tool_calls);
+
+        // Verify finish_reason is valid: exactly one occurrence, in last chunk, and is ToolCalls
+        assert!(
+            validate_finish_reason(&output_chunks, FinishReason::ToolCalls),
+            "finish_reason validation failed for tool call case"
+        );
     }
 
     #[tokio::test]
@@ -403,6 +480,12 @@ mod tests {
             aggregated.has_tool_calls, expected_has_tool_calls,
             "Tool calls presence should match expected value"
         );
+
+        // Verify finish_reason is valid: exactly one occurrence, in last chunk, and is Stop
+        assert!(
+            validate_finish_reason(&output_chunks, FinishReason::Stop),
+            "finish_reason validation failed for non-tool call case"
+        );
     }
 
     #[tokio::test]
@@ -455,6 +538,12 @@ mod tests {
 
         // Verify tool calls
         assert_tool_calls(&aggregated.tool_calls, &test_data.expected_tool_calls);
+
+        // Verify finish_reason is valid: exactly one occurrence, in last chunk, and is ToolCalls
+        assert!(
+            validate_finish_reason(&output_chunks, FinishReason::ToolCalls),
+            "finish_reason validation failed for tool call case"
+        );
     }
 
     #[tokio::test]
@@ -511,6 +600,12 @@ mod tests {
         );
 
         assert_tool_calls(&aggregated.tool_calls, &test_data.expected_tool_calls);
+
+        // Verify finish_reason is valid: exactly one occurrence, in last chunk, and is Stop
+        assert!(
+            validate_finish_reason(&output_chunks, FinishReason::Stop),
+            "finish_reason validation failed for non-tool call case"
+        );
     }
 
     #[tokio::test]
@@ -567,6 +662,12 @@ mod tests {
         );
 
         assert_tool_calls(&aggregated.tool_calls, &test_data.expected_tool_calls);
+
+        // Verify finish_reason is valid: exactly one occurrence, in last chunk, and is ToolCalls
+        assert!(
+            validate_finish_reason(&output_chunks, FinishReason::ToolCalls),
+            "finish_reason validation failed for tool call case"
+        );
     }
 
     #[tokio::test]
@@ -620,6 +721,12 @@ mod tests {
         );
 
         assert_tool_calls(&aggregated.tool_calls, &test_data.expected_tool_calls);
+
+        // Verify finish_reason is valid: exactly one occurrence, in last chunk, and is Stop
+        assert!(
+            validate_finish_reason(&output_chunks, FinishReason::Stop),
+            "finish_reason validation failed for non-tool call case"
+        );
     }
 
     #[tokio::test]
@@ -674,6 +781,12 @@ mod tests {
             "Tool calls presence should match expected value"
         );
         assert_tool_calls(&aggregated.tool_calls, &test_data.expected_tool_calls);
+
+        // Verify finish_reason is valid: exactly one occurrence, in last chunk, and is ToolCalls
+        assert!(
+            validate_finish_reason(&output_chunks, FinishReason::ToolCalls),
+            "finish_reason validation failed for tool call case"
+        );
     }
 
     #[tokio::test]
@@ -726,5 +839,46 @@ mod tests {
 
         // Verify tool calls
         assert_tool_calls(&aggregated.tool_calls, &test_data.expected_tool_calls);
+
+        // Verify finish_reason is valid: exactly one occurrence, in last chunk, and is ToolCalls
+        assert!(
+            validate_finish_reason(&output_chunks, FinishReason::ToolCalls),
+            "finish_reason validation failed for tool call case"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_qwen_finish_reason_length_vllm() {
+        let file_paths = vec![
+            format!(
+                "{}/vllm/qwen3-0.6B/chat_completion_stream_finish_length.json",
+                DATA_ROOT_PATH
+            ),
+            format!(
+                "{}/vllm/qwen3-0.6B/chat_completion_incomplete_tool.json",
+                DATA_ROOT_PATH
+            ),
+        ];
+
+        for file_path in file_paths {
+            let test_data = load_test_data(&file_path);
+
+            // Create a stream from the mock chunks
+            let input_stream = stream::iter(test_data.stream_chunks);
+
+            // Parse the response stream with tool parsing enabled
+            let output_chunks =
+                parse_response_stream(input_stream, true, false, Some("hermes".to_string()), None)
+                    .await;
+
+            // Verify we got output chunks
+            assert!(!output_chunks.is_empty(), "Should have output chunks");
+
+            // Verify finish_reason is valid: exactly one occurrence, in last chunk, and is Length
+            assert!(
+                validate_finish_reason(&output_chunks, FinishReason::Length),
+                "finish_reason validation failed for length finish case"
+            );
+        }
     }
 }