[Cherry-Pick][BugFix] fix consume_signals barrier deadlock in PD separation(#8034)

[kevincheng2]

Cherry-pick of #8034 to release/2.6.

Problem

In PD separation mode, consume_signals thread splits layer0 signals into ready_engine_signals and pending_engine_signals based on whether engine_idx exists in idx_cache_task_dict. Since _add_cache_task_thread fills idx_cache_task_dict asynchronously, different ranks may see different states at the same moment, causing mismatched finish_send_cache_barrier.wait() calls and deadlock.

Fix

Route all layer0 signals through pending_layer0_signals uniformly. The _add_cache_task_thread recovers them one-by-one after cache_info arrives, ensuring all ranks produce identical per-request signals into the queue.

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-06-10 17:47:58

📋 Review 摘要

PR 概述：修复 PD separation 下 consume_signals 在 cache_info 异步注册时按 rank 产生不同 layer0 signal 入队粒度导致的 barrier 死锁。
变更范围：fastdeploy/cache_manager/cache_messager.py
影响面 Tag：[KVCache] [PD Disaggregation]

问题

未发现阻塞性问题。PR 规范问题在下面章节报,不要在这里重复

历史 Findings 修复情况

Finding	问题	状态
F1	pending_layer0_signals 恢复时一次性批量入队导致不同 rank 的 cache_prefilled_engine_ids_queue 粒度仍可能不一致	✅ 已修复

📝 PR 规范检查

当前 PR 提交到 release/2.6 且描述为 cherry-pick，但标题缺少 [Cherry-Pick] 和原 PR 号；PR 描述也未包含 checklist §D2 要求的固定章节。

标题建议（可直接复制）：

• [Cherry-Pick][BugFix] fix consume_signals barrier deadlock in PD separation(#8034)

PR 描述建议（点击展开，可直接复制）

## Motivation
In PD separation mode, consume_signals may observe cache_info registration at different times across ranks. This can make layer0 signal enqueue granularity differ by rank and cause cache send barriers to deadlock.

## Modifications
- In fastdeploy/cache_manager/cache_messager.py, route layer0 signals through pending_layer0_signals and recover registered signals after cache_info is available.
- Cherry-pick of https://github.com/PaddlePaddle/FastDeploy/pull/8034 to release/2.6.

## Usage or Command
N/A

## Accuracy Tests
N/A

## Checklist

- [x] 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.
- [x] 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.

总体评价

本轮按风险优先审查了 consume_signals layer0 signal 入队、_add_cache_task_thread 恢复路径、prefill_layerwise_send_cache_thread 的 barrier/finished queue 下游，以及 EngineWorkerQueue 的 cache_info/finished queue 同步逻辑。当前 diff 将已注册 cache_info 的 layer0 signal 逐个 put([signal])，修复了历史指出的 batch 粒度不一致风险；未发现新的阻塞性问题。

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (release/2.6@21941f5). Learn more about missing BASE report.

Additional details and impacted files

@@              Coverage Diff               @@
##             release/2.6    #8035   +/-   ##
==============================================
  Coverage               ?   72.14%           
==============================================
  Files                  ?      386           
  Lines                  ?    55752           
  Branches               ?     8756           
==============================================
  Hits                   ?    40225           
  Misses                 ?    12691           
  Partials               ?     2836           

Flag	Coverage Δ
GPU	72.14% <100.00%> (?)

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

☔ View full report in Codecov by Harness.
📢 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.

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | ci_status_monitor | 2026-06-11 23:01:47

CI报告基于以下代码生成（30分钟更新一次）:
PR commit: 693d90c | Merge base: 21941f5 (branch: release/2.6)

---

1 Required任务 : 10/10 通过

总执行（rerun次数）	总任务	✅ 通过	❌ 失败	⏳ 运行中	⏸️ 等待中	跳过
37(1)	36	33	3	0	0	0

任务	错误类型	置信度	日志
无	无	无	无

2 失败详情

无

[paddle-bot]

Thanks for your contribution!