[RL] MoE BF16 support m_grouped_bf16_gemm_nn_contiguous in EP

[ckl117]

Motivation

BF16-MoE在EP prefill使用m_grouped_bf16_gemm_nn_contiguous对齐训练精度。
依赖env：

FD_ATTENTION_BACKEND: "FLASH_ATTN"
FD_USE_PHI_MOE_TOPK: 1
FD_USE_PHI_MOE_PERMUTE: 1
FD_SiluAndMul_USE_PHI_SWIGLU: 1
FD_MOE_PROB_IN_ADVANCE: 1
FD_USE_PHI_RMSNORM: 1
FD_ENABLE_RL: 1

💡 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]

ð�¤� AI Code Review | 2026-04-21 16:12:39

� Review ��

PR æ¦�è¿°ï¼�å�¨ Cutlass MoE å��端中å¼�å�¥ m_grouped_bf16_gemm_nn_contiguousï¼�å�ºäº� DeepGemmï¼�æ�¿ä»£ paddle.incubate.nn.functional.batched_gemmï¼�并ä¿�ç�� fallback è·¯å¾�ã��
å��æ�´è��å�´ï¼�model_executor/layers/moe/fused_moe_cutlass_backend.py
å½±å��é�¢ Tagï¼�OP

� PR ����

PR æ��述中 Motivation / Modifications / Usage or Command ç­�å¿�å¡«ç« è��å��æ�ªå¡«å��ï¼�建议补å��å��æ�´å�¨æ�ºå��å�·ä½�ä¿®æ�¹å��容ã��æ �é¢�中 "del paddle.batch_gemm" 表述ä¸�å��ç¡®â��â��batched_gemm å®�é��ä¸�ä½�为 fallback 被ä¿�ç��è��é��å� é�¤ã��

��建议��������

• [RL] Cutlass MoE �端使� DeepGemm grouped_gemm �代 batched_gemm

**�述模�**��������

## Motivation
� Cutlass MoE �端� EP prefill 路�中�使� paddlefleet_ops ��� DeepGemm `m_grouped_bf16_gemm_nn_contiguous` �� `paddle.incubate.nn.functional.batched_gemm`�以���好� BF16 GEMM ���� paddlefleet_ops �������� batched_gemm �为 fallback�

## Modifications
1. �� `m_grouped_bf16_gemm_nn_contiguous` ������ DeepGemm ��
2. `moe_permute` �� `return_expert_indices=True` ��以�� `m_indices`
3. 两� `batched_gemm` ���为�件�����使� DeepGemm�fallback � batched_gemm

��

级�	�件	�述
� 建议	fused_moe_cutlass_backend.py:56	����������模���� backend ����
� ��	fused_moe_cutlass_backend.py:57	paddle.empty ��������确认��������

���价

��������件������好� fallback �容��主�建议补� PR �述并����������"

[PaddlePaddle-bot]

🟡 建议 此函数与 fused_moe_deepgemm_backend.py / fused_moe_blackwell_backend.py 中可能存在的类似封装有重复定义的风险。

建议将该工具函数提取到一个公共模块（如 moe/utils.py 或 moe/__init__.py）中，避免多处重复定义，方便后续统一维护。

[PaddlePaddle-bot]

❓ 疑问 paddle.empty 创建的输出张量内容未初始化。请确认底层 paddlefleet_ops.deep_gemm.m_grouped_bf16_gemm_nn_contiguous 会完整写入 out 的所有元素，否则在 token 数为 0 或 expert_idx_per_token 为空时可能返回包含未初始化内存的张量。

[codecov-commenter]

Codecov Report

❌ Patch coverage is 36.36364% with 7 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@7707be8). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...l_executor/layers/moe/fused_moe_cutlass_backend.py	36.36%	5 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #7527   +/-   ##
==========================================
  Coverage           ?   72.89%           
==========================================
  Files              ?      419           
  Lines              ?    57483           
  Branches           ?     9004           
==========================================
  Hits               ?    41902           
  Misses             ?    12755           
  Partials           ?     2826           

Flag	Coverage Δ
GPU	72.89% <36.36%> (?)

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.

[zoooo0820]

LGTM

[EmmonsCurse]

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