fix_moe_learable-score111

[chang-wenbin]

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.

[paddle-bot]

Thanks for your contribution!

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | ci_status_monitor | 2026-05-25 14:53:42

CI报告基于以下代码生成（30分钟更新一次）:

• PR commit: 4273e4e
• Merge base: e85f0e2 (branch: develop)
• 查看完整 Diff
• CI 详情

---

1 任务总览

Required 任务存在 2 个失败，均为 XPU case 日志上传阶段的 GitHub Actions Artifact 存储配额问题；PR 主测试任务 Run FastDeploy Unit Tests and Coverage / run_tests_with_coverage 已通过。建议等待/清理 Artifact 配额后 rerun 失败任务。

总执行（rerun次数）	总任务	✅ 通过	❌ 失败	⏳ 运行中	⏸️ 等待中	跳过
55(13)	42	36	6	0	0	0

2 任务状态汇总

2.1 Required任务 : 8/10 通过

必选任务阻塞合并，失败需优先处理。

状态	任务	耗时	根因	修复建议	日志	重跑
❌	xpu_4cards_case_test / run_xpu_4cards_cases	27m12s	环境问题：Artifact 存储配额耗尽	清理/等待配额恢复后 rerun	Job	🔄×1
❌	xpu_8cards_case_test / run_xpu_8cards_cases	9m42s	环境问题：Artifact 存储配额耗尽	清理/等待配额恢复后 rerun	Job	🔄×1
✅	其余 8 个必选任务通过	-	-	-	-	-

2.2 可选任务 — 28/32 通过

可选任务不阻塞合并，失败仅供参考。

状态	任务	耗时	日志	重跑
❌	Run iluvatar Tests / run_iluvatar_cases	2m5s	Job	-
❌	Check PR Template	19s	Job	-
❌	CI_HPU	1h4m	Job	-
❌	Trigger Jenkins for PR	14m52s	Job	-
✅	其余 28 个可选任务通过	-	-	-

---

3 失败详情（仅 required）

xpu_4cards_case_test / run_xpu_4cards_cases — 基础设施（置信度: 高）

xpu_4cards_case_test / run_xpu_4cards_cases

• 状态: ❌ 失败
• 错误类型: 基础设施
• 置信度: 高
• 根因摘要: Artifact 存储配额耗尽
• 分析器: 通用分析(fallback)

根因详情:
日志显示 Run CI unittest 步骤成功完成，4 卡 cases 14 passed，失败发生在后续 Upload case logs 步骤。actions/upload-artifact@v6 创建 Artifact 时提示组织/仓库 Artifact storage quota 已耗尽，属于 CI 平台配额问题，与本 PR 修改的 MoE 代码无直接关联。

关键日志:

======================= 14 passed in 1484.02s (0:24:44) ========================
============================4卡cases测试通过!============================
##[error]Failed to CreateArtifact: Artifact storage quota has been hit. Unable to upload any new artifacts.
Usage is recalculated every 6-12 hours.

修复建议:

1. 清理 GitHub Actions Artifact 存储或等待配额 6-12 小时重新计算后，rerun 该任务。
2. 若短期频繁发生，可临时减少 case_logs 上传量或调整 Artifact 保留策略。

修复建议摘要: 清理/等待配额恢复后 rerun

关联变更: 未发现与 fastdeploy/model_executor/layers/moe/fused_moe_cutlass_backend.py 改动直接相关
链接: 查看日志

xpu_8cards_case_test / run_xpu_8cards_cases — 基础设施（置信度: 高）

xpu_8cards_case_test / run_xpu_8cards_cases

• 状态: ❌ 失败
• 错误类型: 基础设施
• 置信度: 高
• 根因摘要: Artifact 存储配额耗尽
• 分析器: 通用分析(fallback)

根因详情:
日志显示 Run CI unittest 步骤成功完成，8 卡 cases 4 passed，失败发生在后续 Upload case logs 步骤。actions/upload-artifact@v6 创建 Artifact 时提示 Artifact storage quota 已耗尽，属于 CI 平台配额问题，与本 PR 修改的 MoE 代码无直接关联。

关键日志:

======================== 4 passed in 395.85s (0:06:35) =========================
============================8卡cases测试通过!============================
##[error]Failed to CreateArtifact: Artifact storage quota has been hit. Unable to upload any new artifacts.
Usage is recalculated every 6-12 hours.

修复建议:

1. 清理 GitHub Actions Artifact 存储或等待配额 6-12 小时重新计算后，rerun 该任务。
2. 若短期频繁发生，可临时减少 case_logs 上传量或调整 Artifact 保留策略。

修复建议摘要: 清理/等待配额恢复后 rerun

关联变更: 未发现与 fastdeploy/model_executor/layers/moe/fused_moe_cutlass_backend.py 改动直接相关
链接: 查看日志

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-05-25 13:21:22

📋 Review 摘要

PR 概述：修复 noaux_tc 分支中可学习路由缩放因子（routed_scaling_factor_learnable）因被 moe_expert_dispatch 返回值覆盖而实际未生效的 bug
变更范围：fastdeploy/model_executor/layers/moe/fused_moe_cutlass_backend.py
影响面 Tag：[OP] [BugFix]

问题

级别	文件	概述
🟡 建议	fused_moe_cutlass_backend.py	其他 MoE backend（deepgemm / triton）中 routed_scaling_factor_learnable 缩放时序是否存在同类问题，需确认
📝 PR 规范	—	标题缺少官方 Tag，PR body 各 section 均为空

📝 PR 规范检查

标题 fix_moe_learable-score111 缺少官方 Tag，且 PR body 各段落均未填写。

标题建议（可直接复制）：

• [BugFix] Fix routed_scaling_factor_learnable applied to stale topk_weights in apply_tp noaux_tc branch

PR 描述建议（可直接复制，必须复刻 checklist §D2 模板的完整结构）：

## Motivation
修复 `fused_moe_cutlass_backend.py` 中 `apply_tp` 的 `noaux_tc` 路径下，可学习缩放因子（`routed_scaling_factor_learnable`）实际未生效的 bug。原实现将缩放应用于 `get_moe_scores` 的返回值，但其后 `moe_expert_dispatch` 会重新输出新的 `topk_weights` 和 `topk_idx`，导致之前的缩放被覆盖丢失。

## Modifications
- `fastdeploy/model_executor/layers/moe/fused_moe_cutlass_backend.py`：将 `apply_tp` 的 `noaux_tc` 分支中可学习缩放因子的应用位置，从 `get_moe_scores` 调用后移至 `moe_expert_dispatch` 调用后，确保缩放作用于最终实际使用的 `topk_weights`。

## 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

本 PR 修复思路正确，将可学习缩放因子的应用时机从 get_moe_scores 之后调整到 moe_expert_dispatch 之后，逻辑上解决了缩放未实际生效的问题。建议同步确认 deepgemm / triton 等其他 MoE backend 的相同路径是否存在同类缩放时序问题。

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@e85f0e2). Learn more about missing BASE report.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #7903   +/-   ##
==========================================
  Coverage           ?   63.69%           
==========================================
  Files              ?      463           
  Lines              ?    64729           
  Branches           ?     9932           
==========================================
  Hits               ?    41226           
  Misses             ?    20714           
  Partials           ?     2789           

Flag	Coverage Δ
GPU	72.79% <100.00%> (?)
XPU	7.10% <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.