upgrade triton moe config.

[xuanyuanminzheng]

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 | ci_status_monitor | 2026-06-03 15:35:39

CI报告基于以下代码生成（30分钟更新一次）:
PR commit: 3ca8136 | Merge base: 529ec9ef (branch: develop)

---

1 Required任务 : 9/10 通过

总执行（rerun次数）	总任务	✅ 通过	❌ 失败	⏳ 运行中	⏸️ 等待中	跳过
41(0)	41	36	5	0	0	0

任务	错误类型	置信度	日志
Run FastDeploy Unit Tests and Coverage / run_tests_with_coverage	PR问题：测试断言值过时，期望num_stages=4但实际返回5	高	Job

2 失败详情

🔴 Run FastDeploy Unit Tests and Coverage / run_tests_with_coverage — PR问题（置信度: 高）

错误类型: PR问题 | 置信度: 高
分析器: 通用分析(fallback)
失败用例:

用例	错误摘要
tests/layers/test_fused_moe_triton_backend.py::TestTritonMoEMethod::test_get_default_config_num_stages	num_stages断言失败：期望4，实际返回5

关键日志:

def test_get_default_config_num_stages(self):
    """M<=32 → num_stages=4; M>32 → num_stages=3."""
    method = backend.TritonMoEMethod()
    cfg32 = method._get_default_config(M=32, E=8)
>   assert cfg32["num_stages"] == 4
E   assert 5 == 4

tests/layers/test_fused_moe_triton_backend.py:1064: AssertionError

• 根因摘要: PR修改了_get_default_config，将M=32时的num_stages从4改为5，导致测试断言失败

PR将 _get_default_config 从基于简单M阈值的启发式规则替换为GPU感知的查表配置（SM100/SM90分档）。在新配置中，M=32对应 _SM100_CONFIGS[32]["num_stages"]=4 或 _SM90_CONFIGS[32]["num_stages"]=5。CI运行环境为SM90（H100），故返回5，而测试期望的是旧逻辑的4。

修复建议:

1. 更新 tests/layers/test_fused_moe_triton_backend.py 第1064行（及附近相关断言），将 == 4 改为与新配置逻辑一致的期望值（SM90: M=32对应num_stages=5，M>32对应视具体配置而定）
2. 或者在测试中通过mock get_sm_version() 固定SM版本，再分别断言SM90和SM100两种配置的期望值，使测试不依赖运行时GPU环境

关联变更: fastdeploy/model_executor/layers/moe/fused_moe_triton_backend.py 第1920-2237行，_get_default_config 方法完整替换

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@529ec9e). Learn more about missing BASE report.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #7980   +/-   ##
==========================================
  Coverage           ?   77.86%           
==========================================
  Files              ?      402           
  Lines              ?    57315           
  Branches           ?     9032           
==========================================
  Hits               ?    44630           
  Misses             ?     9839           
  Partials           ?     2846           

Flag	Coverage Δ
GPU	77.86% <100.00%> (?)

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.

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-06-03 20:47:45

📋 Review 摘要

PR 概述：为 Triton MoE 后端的 _get_default_config 增加 SM100 (B200) GPU 专用查找表，替代原来仅有的 vLLM 启发式逻辑。
变更范围：fastdeploy/model_executor/layers/moe/fused_moe_triton_backend.py、tests/layers/test_fused_moe_triton_backend.py
影响面 Tag：[OP]

问题

级别	文件	概述
🟡 建议	fused_moe_triton_backend.py:1932	热路径中的函数内 import 语句，建议移至模块顶部
🟡 建议	test_fused_moe_triton_backend.py:1119	测试 docstring 称"等距"但注释内数值显示并不等距，存在自相矛盾

历史 Findings 修复情况

Finding	问题	状态
F1	大型配置字典在方法体内每次调用时重新创建	⚠️ 仍存在 — _SM100_CONFIGS（约 130 行）仍定义在 if get_sm_version() >= 100: 分支内部，每次调用均重建。建议提升为模块级常量 _SM100_CONFIGS = {...}
F2	参数 E（专家数量）在新实现中完全未使用	⚠️ 仍存在（SM100 路径）— SM90 默认路径已使用 E（tokens_per_expert = M // max(E, 1)），但 SM100 路径完全忽略 E，与函数签名语义不符

📝 PR 规范检查

标题缺少官方 Tag，PR 描述各段落为空，不符合模板要求。

标题建议（可直接复制）：

• [OP] Upgrade triton MoE config with per-GPU tuned lookup tables

PR 描述建议（点击展开，可直接复制）

## Motivation

Replace the heuristic tile config in Triton MoE backend with GPU-aware (SM90/SM100) tuned lookup tables derived from SGLang's autotuning results, improving kernel performance on B200.

## Modifications

- Added SM100 (B200) lookup table `_SM100_CONFIGS` in `_get_default_config`, using nearest-M key selection
- Retained original vLLM-ported heuristic as fallback for non-SM100 GPUs
- Added `get_sm_version()` call to detect GPU architecture at runtime
- Updated unit tests: all existing SM90 tests patched with `_mock_sm90`; new SM100-specific test cases added

## Usage or Command

N/A

## Accuracy Tests

N/A

## 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.
- [x] Add unit tests. SM100 lookup table tests added in `test_fused_moe_triton_backend.py`.
- [ ] 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

PR 方向正确，为 B200 补充了专用 Triton 配置表并同步更新了测试。两个历史遗留问题（_SM100_CONFIGS 每次重建、SM100 路径忽略 E 参数）尚未修复，建议一并处理后合入。

[PaddlePaddle-bot]

🟡 建议 热路径中的函数内 import 语句

get_sm_version 在每次调用 _get_default_config 时都会触发一次模块字典查找（即使 Python 会缓存模块对象，from ... import 语句每次仍执行 sys.modules 查找 + 属性获取）。对于 MoE 推理热路径，建议将此 import 移到文件顶部模块级别：

# 文件顶部（其他 import 附近）
from fastdeploy.model_executor.utils import get_sm_version

然后在方法内直接使用 get_sm_version()。

[PaddlePaddle-bot]

🟡 建议 测试 docstring 与代码注释自相矛盾

docstring 写的是 "M=100 is equidistant between 96 and 128"（等距），但同一函数内的注释已正确计算 abs(96-100)=4、abs(128-100)=28，两者相差 7 倍，并不等距。

建议修正 docstring，避免误导后续维护者：

"""SM100: M=100 nearest key is 96 (abs_diff=4) not 128 (abs_diff=28)."""