[Cherry-Pick][BugFix] Fix clear_parameters hang issue in MTP during weight cleanup in RL (#7522)

[Deleter-D]

Motivation

Cleaning CUDA Graphs after releasing weights may lead to a hang issue.

Modifications

This PR resolves the problem by moving the CUDA Graph cleanup logic in MTP to occur before weight cleanup, preventing hangs observed in RL scenarios.

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.

[paddle-bot]

Thanks for your contribution!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 0% with 3 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (release/2.6@be2fd17). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/worker/gpu_model_runner.py	0.00%	3 Missing ⚠️

Additional details and impacted files

@@              Coverage Diff               @@
##             release/2.6    #7523   +/-   ##
==============================================
  Coverage               ?   73.55%           
==============================================
  Files                  ?      376           
  Lines                  ?    53033           
  Branches               ?     8288           
==============================================
  Hits                   ?    39011           
  Misses                 ?    11274           
  Partials               ?     2748           

Flag	Coverage Δ
GPU	73.55% <0.00%> (?)

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.

[gongshaotian]

LGTM

[PaddlePaddle-bot]

🤖 AI Code Review | 2026-04-21 21:05:12

📋 Review 摘要

PR 概述：将 MTP Speculative Decoding 场景中 CUDA Graph 的清理操作移至权重释放之前，修复 RL 场景下的 hang 问题。
变更范围：fastdeploy/worker/gpu_model_runner.py — clear_parameters 方法
影响面 Tag：[Speculative Decoding] [RL]

问题

级别	文件	概述
❓ 疑问	gpu_model_runner.py:2700	proposer.model.clear_graph_opt_backend() 嵌套在 if self.use_cudagraph: 内，当主模型未使用 cudagraph 但 draft model 启用了 cudagraph 时，可能存在资源未释放风险
🟡 建议	—	该 hang 问题缺乏回归测试，建议补充测试用例防止未来回退

总体评价

修复思路正确，将 proposer.model.clear_graph_opt_backend() 前移至 clear_parameters() 之前以避免 hang，同时新增 speculative_decoding 和 draft_model_use_cudagraph 条件守卫使逻辑更严谨。建议确认 use_cudagraph 与 draft_model_use_cudagraph 的配置约束关系，以及考虑补充回归测试。

[PaddlePaddle-bot]

❓ 疑问 proposer.model.clear_graph_opt_backend() 现在嵌套在 if self.use_cudagraph: 块内

原始代码中，MTP 的 proposer.model.clear_graph_opt_backend() 是在 if self.spec_method == SpecMethod.MTP: 条件下单独调用的，与主模型的 use_cudagraph 开关无关。

新代码将其嵌套在 if self.use_cudagraph: 内，意味着当主模型禁用 cudagraph（use_cudagraph=False），但 draft model 独立启用了 cudagraph（draft_model_use_cudagraph=True）时，proposer.model.clear_graph_opt_backend() 将永远不会被调用，可能导致 draft model 的 CUDA graph 资源泄漏。

请确认：draft_model_use_cudagraph=True 是否在设计上一定需要 use_cudagraph=True 才能生效？如果是，建议在配置层面添加约束或断言，使假设显式化。

[EmmonsCurse]

LGTM～ Skip coverage check as it mainly relies on tests with RL.