[Feature] Support ThinkingBudget Logits processor to control thinking content length

[jackyYang6]

Motivation

引入 ThinkingBudget logits processor 用于限制 段长度，同时补充中英文使用文档与单测。

Modifications

• 新增 ThinkingBudgetLogitsProcessor 并注册为内置 logits processor，用于控制思考长度。
• DataProcessor 阶段进行 <think> prompt 状态预计算，避免首步 GPU 扫描 prompt。
• 打通 Engine/SamplingParams 的参数透传链路。
• 新增 thinking budget 行为单测。
• 新增中英文功能文档并更新 mkdocs 导航。

Usage or Command

启动服务：

python -m fastdeploy.entrypoints.openai.api_server \
  --model Qwen/Qwen3-0.6B \
  --port 8180 \
  --metrics-port 8181 \
  --engine-worker-queue-port 8182 \
  --max-model-len 32768 \
  --max-num-seqs 32 \
  --logits-processors ThinkingBudgetLogitsProcessor

请求示例：

curl -X POST "http://0.0.0.0:8180/v1/chat/completions" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "你好！"}],
    "max_completion_tokens": 30,
    "logits_processors_args": {
      "thinking_budget": 20,
      "think_stop_sentence": "思考已达上限，开始回复"
    }
  }'

Benchmark

测试环境：

• GPU：A30 24GB（单卡）
• 模型：Qwen3-0.6B
• 数据集: filtered_sharedgpt_2000_input_1136_output_200_fd.json
• Baseline：不启用 --logits-processors
• ThinkingBudget：启用 ThinkingBudgetLogitsProcessor（thinking_budget=20000，不触发预算逻辑）

并发 64 / 500 请求

指标	Baseline	ThinkingBudget	变化
Output throughput (tok/s)	1947.53	1737.76	-10.77%
Mean Decode (tok/s)	45.64	44.50	-2.50%
Mean TTFT (ms)	279.41	286.99	+2.71%
Mean TPOT (ms)	21.99	22.56	+2.59%

并发 1 / 50 请求

指标	Baseline	ThinkingBudget	变化
Output throughput (tok/s)	212.18	202.43	-4.59%
Mean Decode (tok/s)	216.38	205.87	-4.86%
Mean TTFT (ms)	53.85	55.37	+2.82%
Mean TPOT (ms)	4.63	4.86	+4.97%

结论

• 最新单并发结果显示，TTFT 回退约 +2.8%，TPOT/ITL 回退约 +5%，属于可接受但有可优化空间。
• 高并发下吞吐仍有明显回退（Output/Total tok/s 约 -11%~-13%），这是当前方案的主要性能代价。

Accuracy Tests

不涉及 kernel 或模型 forward 逻辑变更，暂无精度回归需求。

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[FDConfig],[APIServer],[Engine], [Scheduler], [PD Disaggregation], [Executor], [Graph Optimization], [Speculative Decoding], [RL], [Models], [Quantization], [Loader], [OP], [KVCache], [DataProcessor], [BugFix], [Docs], [CI], [Optimization], [Feature], [Benchmark], [Others], [XPU], [HPU], [GCU], [DCU], [Iluvatar], [Metax]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no unit tests.
• Provide accuracy results.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch, then cherry-pick it to the release branch with the [Cherry-Pick] PR tag.

[paddle-bot]

Thanks for your contribution!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 84.73118% with 71 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@fd56d85). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...model_executor/logits_processor/thinking_budget.py	84.61%	14 Missing and 22 partials ⚠️
fastdeploy/input/v1/text_processor.py	81.52%	8 Missing and 9 partials ⚠️
fastdeploy/engine/common_engine.py	65.00%	2 Missing and 5 partials ⚠️
fastdeploy/engine/engine.py	65.00%	2 Missing and 5 partials ⚠️
fastdeploy/input/text_processor.py	95.83%	2 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #6367   +/-   ##
==========================================
  Coverage           ?   68.18%           
==========================================
  Files              ?      392           
  Lines              ?    52907           
  Branches           ?     8300           
==========================================
  Hits               ?    36076           
  Misses             ?    14179           
  Partials           ?     2652           

Flag	Coverage Δ
GPU	68.18% <84.73%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

该 PR 引入 ThinkingBudgetLogitsProcessor，用于在解码阶段对 <think>...</think> 思考段生成长度做硬性限制，并打通从 DataProcessor → Engine → Worker → ModelConfig 的相关参数链路，同时补充中英文文档与单测。

Changes:

• 新增 ThinkingBudgetLogitsProcessor 并注册为内置 logits processor，支持预算触发后强制换行与 </think>（可选强制 stop sentence）。
• 在 text_processor / v1/text_processor 侧预计算 prompt 的 <think> 状态，并补充 think_stop_sentence → token ids 的处理逻辑。
• Engine 启动 worker 时补齐 think_start_id / line_break_id 获取与透传，并新增文档与大规模单测覆盖。

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/model_executor/test_thinking_budget.py	覆盖 thinking budget 的关键状态机与 DataProcessor/Engine 相关分支的单测
mkdocs.yml	将 Thinking Budget 文档加入 mkdocs 导航
fastdeploy/worker/worker_process.py	worker 参数解析新增 --think_start_id
fastdeploy/model_executor/logits_processor/thinking_budget.py	新增 ThinkingBudgetLogitsProcessor 的核心实现
fastdeploy/model_executor/logits_processor/init.py	注册并导出 ThinkingBudgetLogitsProcessor
fastdeploy/input/v1/text_processor.py	v1 侧增加 prompt thinking 状态预计算与 stop sentence tokenization（并新增编码缓存）
fastdeploy/input/text_processor.py	非 v1 侧增加同等能力（并新增编码缓存与 get_eos_token_id import fallback）
fastdeploy/engine/engine.py	获取并透传 think_start_id，补齐 line_break_id 获取逻辑
fastdeploy/engine/common_engine.py	同 engine.py 的 worker 启动透传与 line_break_id 处理
fastdeploy/config.py	ModelConfig 增加 think_start_id 字段承接 worker 入参
docs/zh/features/thinking_budget.md	中文功能文档
docs/features/thinking_budget.md	英文功能文档