[Cherry-Pick][OP][Feature] 统一 limit_thinking_content_length CUDA 算子，支持回复长度限制与注入序列

[yuanlehome]

Motivation

将 develop 上 PR #6493 的核心能力迁移到 release/2.5：统一 thinking 截断 CUDA 算子，并引入 response_max_tokens 端到端支持。该迁移同时对 release 分支现有目录与调用差异做了对齐，保持行为与上游改动一致。

💡 If this PR is a Cherry Pick, the PR title needs to follow the format by adding the [Cherry-Pick] label at the very beginning and appending the original PR ID at the end. For example, [Cherry-Pick][CI] Add check trigger and logic(#5191)

💡 如若此PR是Cherry Pick，PR标题需遵循格式，在最开始加上[Cherry-Pick]标签，以及最后面加上原PR ID，例如[Cherry-Pick][CI] Add check trigger and logic(#5191)

Modifications

• 算子统一（OP / SpecDecode）

  • 合并 limit_thinking_content_length_v1/v2 为统一 GPU 算子 limit_thinking_content_length。
  • 合并 speculate_limit_thinking_content_length_v1/v2 为统一推测解码算子 speculate_limit_thinking_content_length。
  • 更新 cpp_extensions.cc 导出符号与 setup_ops.py 编译源，移除旧 v1/v2 文件。

• 推理后处理链路改造（Executor / Worker）

  • pre_and_post_process.py 移除按 limit_strategy 分发的 Python 包装逻辑，直接调用统一算子。
  • gpu_model_runner.py 新增并维护：
    • max_reply_lens
    • inject_token_ids
    • splitwise_role_is_decode 传递
  • thinking/reply 限制在 normal/speculative 两条路径统一生效。

• response_max_tokens 端到端支持（APIServer / DataProcessor / Engine）

  • OpenAI 协议与采样参数新增 response_max_tokens 字段。
  • engine_client 增加参数校验（response_max_tokens > 0）。
  • 多个输入处理器统一 max_tokens 上限裁剪逻辑，并在 enable_thinking=False 时应用 response_max_tokens 约束。
  • reasoning_max_tokens 校验口径同步（允许 0），并移除 D 侧 reasoning_max_tokens -= 1 的历史修正。

• 配置与启动参数透传

  • 新增 think_truncate_prompt_ids 配置读取、worker 启动参数传递与共享输入注入，支持任意长度注入序列。

• 测试同步

  • 删除与旧 v1/v2 算子强绑定的两份 operator 测试文件。
  • 同步更新受影响的输入处理与 E2E 用例断言。

示例（统一算子调用形态）：

limit_thinking_content_length(
    sampled_token_ids,
    share_inputs["max_think_lens"],
    share_inputs["max_reply_lens"],
    share_inputs["step_idx"],
    share_inputs["limit_think_status"],
    share_inputs["stop_flags"],
    share_inputs["eos_token_id"],
    share_inputs["inject_token_ids"],
    think_end_id,
    splitwise_role_is_decode,
)

Usage or Command

新增请求参数（OpenAI 兼容）：

{
  "max_tokens": 256,
  "reasoning_max_tokens": 128,
  "response_max_tokens": 64
}

语义：

• reasoning_max_tokens：限制思考阶段 token 数；
• response_max_tokens：限制回复阶段 token 数（thinking 关闭时直接约束总回复长度）。

Accuracy Tests

该 PR 涉及 kernel 与后处理路径重构，准确性结果待补充（与 develop 对齐策略）。

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.

Warning

Firewall rules blocked me from connecting to one or more addresses (expand for details)

I tried to connect to the following addresses, but was blocked by firewall rules:

• https://api.github.com/repos/PaddlePaddle/FastDeploy/pulls/6493/files
  • Triggering command: /home/REDACTED/work/_temp/ghcca-node/node/bin/node /home/REDACTED/work/_temp/ghcca-node/node/bin/node --enable-source-maps /home/REDACTED/work/_temp/copilot-developer-action-main/dist/index.js (http block)

If you need me to access, download, or install something from one of these locations, you can either:

• Configure Actions setup steps to set up my environment, which run before the firewall is enabled
• Add the appropriate URLs or hosts to the custom allowlist in this repository's Copilot coding agent settings (admins only)

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.

[paddle-bot]

Thanks for your contribution!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 65.04854% with 36 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (release/2.5@a368856). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/input/v1/ernie4_5_processor.py	41.66%	4 Missing and 3 partials ⚠️
.../v1/ernie4_5_vl_processor/ernie4_5_vl_processor.py	25.00%	4 Missing and 2 partials ⚠️
fastdeploy/entrypoints/engine_client.py	0.00%	3 Missing and 2 partials ⚠️
fastdeploy/input/v1/text_processor.py	57.14%	3 Missing ⚠️
fastdeploy/input/ernie4_5_processor.py	77.77%	1 Missing and 1 partial ⚠️
...put/ernie4_5_vl_processor/ernie4_5_vl_processor.py	75.00%	1 Missing and 1 partial ⚠️
fastdeploy/input/text_processor.py	75.00%	1 Missing and 1 partial ⚠️
...1/paddleocr_vl_processor/paddleocr_vl_processor.py	50.00%	1 Missing and 1 partial ⚠️
.../input/v1/qwen3_vl_processor/qwen3_vl_processor.py	50.00%	1 Missing and 1 partial ⚠️
...oy/input/v1/qwen_vl_processor/qwen_vl_processor.py	50.00%	1 Missing and 1 partial ⚠️
... and 3 more

Additional details and impacted files

@@              Coverage Diff               @@
##             release/2.5    #6511   +/-   ##
==============================================
  Coverage               ?   68.49%           
==============================================
  Files                  ?      391           
  Lines                  ?    52802           
  Branches               ?     8220           
==============================================
  Hits                   ?    36167           
  Misses                 ?    14004           
  Partials               ?     2631           

Flag	Coverage Δ
GPU	68.49% <65.04%> (?)

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.