[Cherry-Pick][RL] Support moe_topk_select using Paddle native operators and Add fused stack-transpose-quant for BlockWiseFP8 MoE weight quantization and swiglu-fp8-quant op for DeepGemmFusedMoE for training alignment (#6850)

[DanielSun11]

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!

[Copilot]

Pull request overview

该 PR 主要围绕 MoE 的 TopK 选择与 FP8（含 UE8M0 scale）量化路径做增强：在 DeepGemmFusedMoE 侧引入基于 Paddle 原生算子的 moe_topk_select 选择逻辑，并新增/对齐 Fleet 侧的 fused 算子接入（stack+transpose+fp8 quant、以及带 routed prob 的 fused swiglu+fp8 quant），同时补充对应单测与环境开关。

Changes:

• 在 DeepGemm MoE 路径新增 moe_topk_select（Paddle native）并通过 FD_USE_PHI_TOPK 控制启用。
• 新增 fp8_utils.fused_stack_transpose_quant 并在 Triton MoE 的 UE8M0 量化权重处理里通过 FD_USE_FLEET_FP8_QUANT 切到 Fleet fused kernel。
• 新增 FD_MOE_PROB_IN_ADVANCE/FD_USE_FLEET_FP8_QUANT/FD_USE_PHI_TOPK 环境变量，并补充 deepgemm/fused-quant 相关测试用例。

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
tests/operators/test_noaux_tc_redundant.py	增加基于 moe_topk_select 的 group-topk 对齐测试
tests/layers/test_fp8_ue8m0.py	增加 Fleet fused stack+transpose+quant 及 fused_stack_transpose_quant 的单测覆盖
tests/layers/test_deepgemm_fused_moe.py	新增 DeepGemmFusedMoE 多路径（TP/EP、prob-in-advance、phi permute 等）对齐测试
fastdeploy/model_executor/layers/quantization/fp8_utils.py	新增 try_import、paddlefleet_ops 引入与 fused_stack_transpose_quant 等量化辅助能力
fastdeploy/model_executor/layers/moe/fused_moe_triton_backend.py	UE8M0 权重量化新增 Fleet fused 路径（chunk 化处理）
fastdeploy/model_executor/layers/moe/fused_moe_deepgemm_backend.py	引入 moe_topk_select 与 FD_MOE_PROB_IN_ADVANCE 下的 fused swiglu+fp8 quant 路径，并调整 unpermute combine 开关
fastdeploy/envs.py	新增 MoE TopK/Fleet FP8 quant/prob-in-advance 等环境变量开关

[codecov-commenter]

Codecov Report

❌ Patch coverage is 73.68421% with 25 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (release/2.5@d585b22). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
..._executor/layers/moe/fused_moe_deepgemm_backend.py	64.44%	9 Missing and 7 partials ⚠️
...oy/model_executor/layers/quantization/fp8_utils.py	65.38%	6 Missing and 3 partials ⚠️

Additional details and impacted files

@@              Coverage Diff               @@
##             release/2.5    #6935   +/-   ##
==============================================
  Coverage               ?   69.18%           
==============================================
  Files                  ?      389           
  Lines                  ?    53474           
  Branches               ?     8394           
==============================================
  Hits                   ?    36995           
  Misses                 ?    13812           
  Partials               ?     2667           

Flag	Coverage Δ
GPU	69.18% <73.68%> (?)

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.

[Copilot]

Pull request overview

该 PR（Cherry-pick）主要围绕 MoE 路由与 FP8 权重量化路径做增强：一方面在 DeepGemm MoE 中新增基于 Paddle 原生算子的 moe_topk_select（用于 noaux_tc 训练对齐），另一方面引入 Fleet 的 fused stack-transpose-fp8-quant / fused weighted swiglu-fp8-quant 相关能力，并补充/扩展对应单测覆盖。

Changes:

• 新增 moe_topk_select（Paddle 原生算子实现）并在 DeepGemm MoE 的 TP/EP prefill 路径按开关启用
• 在 BlockWiseFP8 MoE 权重量化中接入 fused_stack_transpose_quant（支持按 chunk 处理专家权重）
• 增加/扩展多份测试文件以覆盖 Triton/DeepGemm/ue8m0 分支与 fused quant 行为

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
tests/operators/test_noaux_tc_redundant.py	新增用例对比 moe_topk_select 与参考实现的一致性
tests/layers/test_fused_moe_triton_backend.py	新增 Triton MoE backend 的大量单测（mock/patch 覆盖多分支）
tests/layers/test_fp8_ue8m0.py	扩展 ue8m0 与 fused_stack_transpose_quant 等分支的测试覆盖
tests/layers/test_deepgemm_fused_moe.py	新增 DeepGemmFusedMoE 训练对齐相关路径的测试（含 EP/TP 分支）
fastdeploy/model_executor/layers/quantization/fp8_utils.py	增加 try_import、paddlefleet_ops 探测与 fused_stack_transpose_quant 封装
fastdeploy/model_executor/layers/moe/fused_moe_triton_backend.py	BlockWiseFP8 MoE 动态量化 ue8m0 分支中接入 fused stack-transpose-quant
fastdeploy/model_executor/layers/moe/fused_moe_deepgemm_backend.py	新增 moe_topk_select 并引入 prob-in-advance + fused swiglu-fp8-quant 的分支
fastdeploy/envs.py	新增/调整若干 env 开关（MoE TopK / prob-in-advance 等）

[zoooo0820]

这里是不是合并代码的时候漏了，同时加了2个topk环境变量

[DanielSun11]

解决conflict的时候漏删掉了。现在已经删掉了