[Bug] Qwen3 Reasoning Parser 解析错误

[QwertyJack]

Checklist

• 1. I have searched related issues but cannot get the expected help.
• 2. The bug has not been fixed in the latest version.
• 3. Please note that if the bug-related issue you submitted lacks corresponding environment info and a minimal reproducible demo, it will be challenging for us to reproduce and resolve the issue, reducing the likelihood of receiving feedback.

Describe the bug

lmdeploy 0.9.0 部署 Qwen3 reasoning_content 错误。

Reproduction

启动服务: lmdeploy serve api_server /models/Qwen3-8B --reasoning-parser deepseek-r1

以下两种情况存在bug：

Case 1 流式输出：第一段 chunk content 包含 <think>

$ curl ... -d '{..., "stream": true}'

data: {"id":"1","object":"chat.completion.chunk","created":1750741881,"model":"/models/Qwen3-8B","choices":[{"index":0,"delta":{"role":"assistant","content":"<think>","reasoning_content":null,"tool_calls":[]},"logprobs":null,"finish_reason":null}],"usage":null}

data: {"id":"1","object":"chat.completion.chunk","created":1750741881,"model":"/models/Qwen3-8B","choices":[{"index":0,"delta":{"role":"assistant","content":null,"reasoning_content":"\n","tool_calls":[]},"logprobs":null,"finish_reason":null}],"usage":null}

data: {"id":"1","object":"chat.completion.chunk","created":1750741881,"model":"/models/Qwen3-8B","choices":[{"index":0,"delta":{"role":"assistant","content":null,"reasoning_content":"好的","tool_calls":[]},"logprobs":null,"finish_reason":null}],"usage":null}

...

Case 2 非流式输出，限制 max_tokens：输出包含 <think>

$ curl ... -d '{..., "stream": false, "max_tokens": 8}'
{"id":"2","object":"chat.completion","created":1750742086,"model":"/models/Qwen3-8B","choices":[{"index":0,"message":{"role":"assistant","content":null,"reasoning_content":"<think>\n好的，用户发来的是“你好!”，我","tool_calls":null},"logprobs":null,"finish_reason":"length"}],"usage":{"prompt_tokens":16,"total_tokens":29,"completion_tokens":13}}

Environment

TorchVision: 0.21.0+cu124
LMDeploy: 0.9.0+unknown
transformers: 4.52.4
gradio: Not Found
fastapi: 0.115.13
pydantic: 2.11.7
triton: 3.2.0

Error traceback

[RunningLeon]

@QwertyJack hi, for streaming mode, you can use --reasoning-parser qwen-qwq. As for the other case, the current logic is to parse reasoning content only if the thinking end token is in output. Otherwise, no end thinking token, then the output is no use and should be dropped.

lmdeploy/lmdeploy/serve/openai/reasoning_parser/deepseek_r1_reasoning_parser.py

Line 125 in 51ab631

if self.think_end_token not in model_output:

[QwertyJack]

Many thanks, @RunningLeon !
With the latest main branch, streaming works like a charm!

[github-actions]

This issue is marked as stale because it has been marked as invalid or awaiting response for 7 days without any further response. It will be closed in 5 days if the stale label is not removed or if there is no further response.

[ywx217]

Hi, I found an issue when stream=true and think is disabled. There is only reasoning_content in delta message, which is not expected behavior.

Will submit a PR for this fix.

[github-actions]

This issue is marked as stale because it has been marked as invalid or awaiting response for 7 days without any further response. It will be closed in 5 days if the stale label is not removed or if there is no further response.

[QwertyJack]

no stale

[github-actions]

This issue is marked as stale because it has been marked as invalid or awaiting response for 7 days without any further response. It will be closed in 5 days if the stale label is not removed or if there is no further response.