[Bugfix] Align thinking_budget behavior with ERNIE reasoning flow

[jackyYang6]

Motivation

Fix two behavior inconsistencies in ThinkingBudgetLogitsProcessor:

1. thinking_budget ended thinking with an extra newline before </think>, which was not aligned with the existing reasoning_max_tokens behavior.
2. For ERNIE models, thinking_budget could fail to take effect on GPU because prompt-side <think> state was not propagated through ERNIE-specific processors, and the runtime fallback only checked prompt_ids, which is not available on the current GPU path.

This PR aligns thinking_budget with the current ERNIE reasoning flow and removes the extra newline before </think>.

Modifications

1. Update ThinkingBudgetLogitsProcessor to terminate thinking by forcing </think> directly instead of \n + </think>.
2. Keep think_stop_sentence behavior as think_stop_sentence + </think>, and remove the implicit leading newline before the stop sentence.
3. Fix GPU runtime fallback in ThinkingBudgetLogitsProcessor:
   • keep prompt_ids scan for XPU/HPU compatibility
   • use token_ids_all + prompt_lens when prompt_ids is unavailable on GPU
4. Extract shared thinking-budget request-side preprocessing into common helpers and reuse it in:
   • fastdeploy/input/text_processor.py
   • fastdeploy/input/v1/text_processor.py
   • fastdeploy/input/ernie4_5_processor.py
   • fastdeploy/input/v1/ernie4_5_processor.py
   • fastdeploy/input/ernie4_5_vl_processor/ernie4_5_vl_processor.py
   • fastdeploy/input/v1/ernie4_5_vl_processor/ernie4_5_vl_processor.py
5. Align prompt-side counting semantics with operator-level reasoning length control:
   • prompt-side <think> scaffolding or prefilled thinking content does not consume thinking_budget
   • decode-time generated thinking tokens consume thinking_budget
   • think_stop_sentence still consumes thinking_budget
   • </think> does not consume thinking_budget
6. Update docs

Usage or Command

FD_USE_GET_SAVE_OUTPUT_V1=1 python -m fastdeploy.entrypoints.openai.api_server \
  --model PaddlePaddle/ERNIE-4.5-21B-A3B-Thinking \
  --host 0.0.0.0 \
  --port 8582 \
  --max-model-len 32768 \
  --max-num-seqs 8 \
  --kv-cache-ratio 0.75 \
  --enable-expert-parallel \
  --tensor-parallel-size 1 \
  --gpu-memory-utilization 0.9 \
  --enable-logprob \
  --max-logprobs 2 \
  --logprobs-mode processed_logprobs \
  --reasoning-parser ernie-x1 \
  --logits-processors ThinkingBudgetLogitsProcessor

Example request:

{
  "messages": [
    {
      "role": "user",
      "content": "你好"
    }
  ],
  "enable_thinking": true,
  "reasoning_max_tokens": 10,
  "logits_processors_args": {
    "thinking_budget": 10,
    "think_stop_sentence": "思考已结束，开始回复"
  },
  "logprobs": true,
  "top_logprobs": 1,
  "include_logprobs_decode_token": true,
  "temperature": 0,
  "max_tokens": 20,
  "return_token_ids": true
}

Accuracy Tests

This PR does not modify kernel math or model forward numerics. The validation focus is output behavior consistency for reasoning truncation.

Behavioral validation on ERNIE thinking model:

Case	Config	Result
thinking_budget only	thinking_budget=10	ThinkingBudgetLogitsProcessor takes effect on ERNIE and inserts </think> after 10 decode-time thinking tokens
thinking_budget + think_stop_sentence	thinking_budget=10, think_stop_sentence="思考已结束，开始回复"	stop sentence is emitted completely, followed by </think>
reasoning_max_tokens only	reasoning_max_tokens=10	</think> is inserted after 10 thinking tokens
both enabled	reasoning_max_tokens=10, thinking_budget=10, think_stop_sentence=...	stop sentence remains complete and total thinking tokens before </think> stay aligned at 10
newline check	all above cases	no extra newline before </think>

Checklist

• Add at least a tag in the PR title.
• 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 85.06787% with 33 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@3a4e139). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/input/v1/text_processor.py	76.74%	10 Missing and 10 partials ⚠️
fastdeploy/input/text_processor.py	90.47%	4 Missing and 4 partials ⚠️
...model_executor/logits_processor/thinking_budget.py	83.87%	3 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #6934   +/-   ##
==========================================
  Coverage           ?   73.78%           
==========================================
  Files              ?      399           
  Lines              ?    55697           
  Branches           ?     8784           
==========================================
  Hits               ?    41094           
  Misses             ?    11699           
  Partials           ?     2904           

Flag	Coverage Δ
GPU	73.78% <85.06%> (?)

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 旨在修复并统一 ThinkingBudgetLogitsProcessor 与 ERNIE 推理（reasoning）流程在“思考段截断”上的行为：去掉 </think> 前的隐式换行，并补齐 ERNIE 在 GPU 路径下 prompt 侧 <think> 状态无法传播导致 thinking_budget 失效的问题，同时将 request 侧的 thinking-budget 预处理逻辑抽取为通用 helper 以复用到多类 Processor 中。

Changes:

• 调整 ThinkingBudgetLogitsProcessor：budget 达到后直接强制输出 </think>（stop sentence 场景则先输出 stop sentence 再输出 </think>），并新增 prompt 状态扫描的 GPU fallback（token_ids_all + prompt_lens）。
• 在 text/ernie/v1/ernie_vl 等多个输入处理器中抽取并复用 thinking-budget 相关的 request-side 预处理 helper（stop sentence 编码、prompt <think> 状态更新等），对齐“prompt 侧不消耗 budget、decode 侧消耗 budget”的语义。
• 更新单测与中英文文档，覆盖新语义与 GPU fallback 行为。

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
fastdeploy/model_executor/logits_processor/thinking_budget.py	budget 达到后直接强制 </think>；stop sentence 优先；增加 prompt 扫描的 token_ids_all fallback。
fastdeploy/input/text_processor.py	抽取/新增 thinking-budget 通用 helper（stop sentence 编码、prompt 状态更新、literal 编码缓存）；tokenize cache 支持 lazy init。
fastdeploy/input/v1/text_processor.py	同上（v1 版本）。
fastdeploy/input/ernie4_5_processor.py	复用通用 helper，确保 ERNIE 文本处理器能写入 thinking-budget 所需的 prompt-side 状态与 stop sentence token ids。
fastdeploy/input/v1/ernie4_5_processor.py	同上（v1 版本）。
fastdeploy/input/ernie4_5_vl_processor/ernie4_5_vl_processor.py	复用通用 helper，确保 ERNIE-VL 路径也能准备 thinking-budget 参数与 prompt-side 状态。
fastdeploy/input/v1/ernie4_5_vl_processor/ernie4_5_vl_processor.py	同上（v1 版本）。
tests/model_executor/test_thinking_budget.py	更新/新增用例以匹配新语义（prompt 不消耗 budget、直接 </think>、stop sentence 行为、GPU fallback）。
docs/features/thinking_budget.md	更新英文文档描述与实践建议，对齐新行为。
docs/zh/features/thinking_budget.md	更新中文文档描述与实践建议，对齐新行为。