[RL] enforce cuda graph clearing and rebuilding in rsync weight update by rdma

[liyonghua0910]

Motivation

💡 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

Usage or Command

Accuracy Tests

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.

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-06-29 15:19:09

📋 Review 摘要

PR 概述：调整 rsync/RDMA 权重更新前后的 CUDA Graph、KV cache 与输入缓存重建流程。
变更范围：fastdeploy/worker/gpu_model_runner.py
影响面 Tag：[RL] [Graph Optimization] [KVCache]

问题

级别	文件	概述
🔴 Bug	fastdeploy/worker/gpu_model_runner.py:3313	非 MTP 场景下 GDR release cache 重建会访问空的 self.proposer

📝 PR 规范检查

标题符合规范；PR 描述包含模板章节，但 Motivation、Modifications、Usage or Command、Accuracy Tests 仍为空/占位内容，建议替换为可读的完整描述。

PR 描述建议（点击展开，可直接复制）

## Motivation
修复 rsync/RDMA 权重更新过程中 CUDA Graph 未按权重变化及时清理并重建的问题，避免权重更新后继续复用旧图状态。

## Modifications
在 `fastdeploy/worker/gpu_model_runner.py` 中抽取权重更新前后的内存处理逻辑：
- GDR release cache 路径在权重更新前后清理/重建 KV cache 与 CUDA Graph。
- RDMA 权重更新路径在更新前清理 CUDA Graph，并在更新后重建 CUDA Graph。
- 权重更新后重置 `share_inputs`、MTP model inputs 以及缓存的 model/sampler 输出状态。

## 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.
- [ ] 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

当前实现有一个会阻断 GDR release cache 权重更新的空对象访问问题，需要先恢复 MTP 条件保护后再合入。

[PaddlePaddle-bot]

🔴 Bug 非 MTP 场景下这里会无条件访问空的 self.proposer。

_init_speculative_proposer() 在 spec_method is None 时将 self.proposer 设为 None，但这段重建逻辑只受 rebuild_kv_cache 和 enable_cache_manager_v1 控制。只要使用 GDR 权重更新并开启 gdr_release_cache，普通非投机解码模型在 cache manager v1 关闭时就会在重建阶段抛 AttributeError，导致本次权重更新失败且 KV cache 状态不能恢复到正常服务。

建议修复方式：保留旧逻辑中的 MTP 条件，只在 self.spec_method == SpecMethod.MTP 时初始化 proposer cache；主模型 cache 仍然独立初始化。

if self.spec_method == SpecMethod.MTP:
    if not self.enable_cache_manager_v1:
        self.proposer.initialize_kv_cache(main_model_num_blocks=self.num_gpu_blocks)
self.initialize_kv_cache()

[codecov-commenter]

Codecov Report

❌ Patch coverage is 6.45161% with 29 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@f4eda5a). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/worker/gpu_model_runner.py	6.45%	29 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #8085   +/-   ##
==========================================
  Coverage           ?   67.55%           
==========================================
  Files              ?      475           
  Lines              ?    66913           
  Branches           ?    10323           
==========================================
  Hits               ?    45202           
  Misses             ?    18836           
  Partials           ?     2875           

Flag	Coverage Δ
GPU	77.57% <6.45%> (?)
XPU	6.95% <0.00%> (?)

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.