Skip to content

【Hackathon 10th Spring No.47】【RFC】新增 MiniMax-M1 模型设计文档#1252

Closed
cloudforge1 wants to merge 26 commits into
PaddlePaddle:masterfrom
CloudForge-Solutions:task/047-rfc-minimax-m1
Closed

【Hackathon 10th Spring No.47】【RFC】新增 MiniMax-M1 模型设计文档#1252
cloudforge1 wants to merge 26 commits into
PaddlePaddle:masterfrom
CloudForge-Solutions:task/047-rfc-minimax-m1

Conversation

@cloudforge1
Copy link
Copy Markdown
Contributor

@cloudforge1 cloudforge1 commented Mar 20, 2026

概述

本 RFC 设计 FastDeploy 新增 MiniMax-M1 模型的推理支持方案。

核心挑战

MiniMax-M1 采用 80 层混合架构（标准 GQA + 线性注意力/Mamba），需要：

  1. Lightning Attention CUDA 内核开发：5 个 vLLM Triton kernel → CUDA C++ 翻译
  2. MambaBackend 注意力后端：新增 MambaBackend(AttentionBackend) 支持线性注意力
  3. 混合注意力调度：逐层通过 attn_type_list 选择 GQA 或 Mamba 后端
  4. MoE + SharedExpert + DeepNorm 集成

主要工作量分布

  • CUDA 内核翻译：2-3 周（核心难点）
  • MambaBackend + 算子包装：1 周
  • 模型组网 + 权重加载：1 周
  • 测试调优：1 周
  • 预计合计 ~5-6 周

复用与创新

  • 复用：GQA 注意力、MoE 分派、RoPE、并行线性层、量化框架
  • 新增：MambaBackend（可复用于 Jamba、Zamba 等后续 Mamba 模型）

相关链接

YqGe585 and others added 24 commits March 9, 2026 15:49
@YqGe585
add guide (#1221)
c071d41
@openvino-book @mattheliu
update: 🚀 PaddleOCR + ERNIE × Open-Source Ecosystem 高价值开源工具/应用/项目...征集 (
978de2b 
#1223)

* Create readme.md

* Add files via upload

* Create readme.md

* Add files via upload

* Delete pfcc/call-for-contributions/happy_os directory

* Add contribution invitation for PaddleOCR and ERNIE

This document invites contributions to the PaddleOCR and ERNIE project, outlining the benefits of participation and project structure.

* Revise Milvus community release status

Updated Milvus community release note to indicate pending release.

* Delete pfcc/call-for-contributions/OCR+ERNIE/PaddleOCR + ERNIE × Open-Source Ecosystem 高价值开源项目案例征集

* Update PaddleOCR + ERNIE × Open-Source Ecosystem 高价值开源项目案例征集.md

* Update project timeline from 30 days to 2 weeks

* Update frequency of high-quality open-source results

* fix: polish docs based on review

* Update Milvus community release link

* Revise PaddleOCR + ERNIE project details and goals

Updated the project description and requirements for the PaddleOCR + ERNIE open-source initiative, emphasizing collaboration and community involvement.

---------

Co-authored-by: mattheliu <leonliuzx@outlook.com>
@luotao1
Create 硬件生态任务合集.md (#1222)
de1af93 
* Create 硬件生态任务合集.md

* Update and rename 硬件生态任务合集.md to 【Hackathon_10th】硬件生态任务合集.md

* Create 打卡任务模版.md
@zhuo-yoyowz @luotao1
OpenVINO Update 【Hackathon_10th】硬件生态任务合集.md (#1227)
f14f131 
* Intel Update 【Hackathon_10th】硬件生态任务合集.md

* Update 【Hackathon_10th】硬件生态任务合集.md

* Update OpenVINO【Hackathon_10th】硬件生态任务合集.md

* Update 【Hackathon_10th】硬件生态任务合集.md

* Update OpenVINO【Hackathon_10th】硬件生态任务合集.md

---------

Co-authored-by: Tao Luo <luotao02@baidu.com>
@w-y-w @luotao1
瑞芯微-PaddleOCR-VL-RKNN3转换-打卡任务文档 (#1228)
0ac9dec 
* 新增瑞芯微-PaddleOCR-VL-RKNN3转换-打卡任务文档

* 移除代码文件

* Update 瑞芯微-PaddleOCR-VL-RKNN3转换-打卡任务.md

---------

Co-authored-by: Tao Luo <luotao02@baidu.com>
@deepin-mozart
feat: [Hackathon_10th] 添加统信相关的任务 (#1225)
a8a5267 
添加打卡任务、最终任务
@openvino-book @mattheliu
update 评价标准 (#1224)
42e9fb2 
* Create readme.md

* Add files via upload

* Create readme.md

* Add files via upload

* Delete pfcc/call-for-contributions/happy_os directory

* Add contribution invitation for PaddleOCR and ERNIE

This document invites contributions to the PaddleOCR and ERNIE project, outlining the benefits of participation and project structure.

* Revise Milvus community release status

Updated Milvus community release note to indicate pending release.

* Delete pfcc/call-for-contributions/OCR+ERNIE/PaddleOCR + ERNIE × Open-Source Ecosystem 高价值开源项目案例征集

* Update PaddleOCR + ERNIE × Open-Source Ecosystem 高价值开源项目案例征集.md

* Update project timeline from 30 days to 2 weeks

* Update frequency of high-quality open-source results

* fix: polish docs based on review

* Update Milvus community release link

* Revise PaddleOCR + ERNIE project details and goals

Updated the project description and requirements for the PaddleOCR + ERNIE open-source initiative, emphasizing collaboration and community involvement.

* Expand project evaluation and contribution guidelines

Added project evaluation criteria and guidelines for contributions.

---------

Co-authored-by: mattheliu <leonliuzx@outlook.com>
@mattheliu @claude
Add hackathon 10th code camp formal batch project details (#1237)
b9ba94d 
* Add hackathon 10th code camp formal batch project details

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Remove project 2 (PaddleOCR+ERNIE) and renumber projects

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
@cixjohn26
此芯 × 解锁基于 Armv9 架构硬件平台的端侧推理-打卡任务文档 (#1233)
6559e34
@zhuo-yoyowz @luotao1
OpenVINO Update 打卡任务模版.md (#1226)
938299d 
* OpenVINO Update 打卡任务模版.md

* Create OpenVINO打卡任务.md

* Update 打卡任务模版.md

* Update and rename OpenVINO打卡任务.md to OpenVINO-Notebook-快速上手打卡任务.md

---------

Co-authored-by: Tao Luo <luotao02@baidu.com>
@luotao1
[Iluvatar] Add Iluvatar task (#1230)
4f81e54 
* [Iluvatar] Add Iluvatar task

* [Iluvatar] Update task

* [Iluvatar] Update task

* Update 【Hackathon_10th】硬件生态任务合集.md

* Rename build.png to 天数-build.png

* Rename demo.png to 天数-demo.png

* Rename install.png to 天数-install.png

* Update and rename pfcc/paddle-hardware/iluvatar/打卡任务-天数.md to pfcc/paddle-hardware/基于天数智芯-BI-150S-的-FastDeploy-编译与打包.md

---------

Co-authored-by: Tao Luo <luotao02@baidu.com>
@AIwork4me
Update PaddleOCR + ERNIE × Open-Source Ecosystem 高价值开源项目案例征集.md (#1246)
92345f7
@openvino-book @mattheliu
update (#1244)
ad21ffc 
* Create readme.md

* Add files via upload

* Create readme.md

* Add files via upload

* Delete pfcc/call-for-contributions/happy_os directory

* Add contribution invitation for PaddleOCR and ERNIE

This document invites contributions to the PaddleOCR and ERNIE project, outlining the benefits of participation and project structure.

* Revise Milvus community release status

Updated Milvus community release note to indicate pending release.

* Delete pfcc/call-for-contributions/OCR+ERNIE/PaddleOCR + ERNIE × Open-Source Ecosystem 高价值开源项目案例征集

* Update PaddleOCR + ERNIE × Open-Source Ecosystem 高价值开源项目案例征集.md

* Update project timeline from 30 days to 2 weeks

* Update frequency of high-quality open-source results

* fix: polish docs based on review

* Update Milvus community release link

* Revise PaddleOCR + ERNIE project details and goals

Updated the project description and requirements for the PaddleOCR + ERNIE open-source initiative, emphasizing collaboration and community involvement.

* Expand project evaluation and contribution guidelines

Added project evaluation criteria and guidelines for contributions.

* Revise PaddleOCR + ERNIE contribution guidelines

Updated the project description and structure for PaddleOCR + ERNIE open-source contributions.

---------

Co-authored-by: mattheliu <leonliuzx@outlook.com>
@onecatcn @luotao1
【Hackathon 10th】Add Qualcomm PaddleOCR-VL task to hardware collection (…
ac9f3ce 
…#1241)

* Add Hackathon 10th Qualcomm PaddleOCR-VL task

* Add Qualcomm PaddleOCR-VL task to hardware hackathon tasks

* Remove standalone qualcomm task file, keep only the entry in hardware collection

* Update 【Hackathon_10th】硬件生态任务合集.md

---------

Co-authored-by: Tao Luo <luotao02@baidu.com>
@zhuo-yoyowz @luotao1
Update 【Hackathon_10th】硬件生态任务合集 OpenVINO任务第一阶段提交RFC方案.md (#1243)
3db6fb7 
* Update 【Hackathon_10th】硬件生态任务合集 OpenVINO任务第一阶段提交RFC方案.md

* Update 【Hackathon_10th】硬件生态任务合集.md

---------

Co-authored-by: Tao Luo <luotao02@baidu.com>
@w-y-w @luotao1
瑞芯微-第十届飞桨黑客松进阶任务说明 (#1229)
09d5621 
* 新增瑞芯微-第十届黑客松进阶任务说明

* remove >

* 更新进阶任务筛选依据

* Update 【Hackathon_10th】硬件生态任务合集.md

* Rename pfcc/paddle-hardware/paddleocr-vl-rk1820-infer-result.png to pfcc/paddle-hardware/images/瑞芯微-paddleocr-vl-rk1820-infer-result.png

* Rename pfcc/paddle-hardware/paddleocr-vl-simular-infer-result.png to pfcc/paddle-hardware/images/瑞芯微-paddleocr-vl-simular-infer-result.png

* Update 瑞芯微-PaddleOCR-VL-RKNN3转换-打卡任务.md

---------

Co-authored-by: Tao Luo <luotao02@baidu.com>
@gyshi @luotao1
【PaddlePaddle Hackathon 10】 增加燧原科技的打卡任务和进阶任务 (#1234)
d87849d 
* dd enflame gcu 燧原打卡任务

* little fix.

* 增加燧原任务二。

* update

* update 1

* add 燧原科技 email.

* Update 【Hackathon_10th】硬件生态任务合集.md

* Update and rename 燧原科技-打卡任务.md to 燧原科技-基于FastDeploy跑通ERNIE-4.5-0.3B-Paddle 打卡任务.md

* Update 燧原科技-基于FastDeploy跑通ERNIE-4.5-0.3B-Paddle 打卡任务.md

---------

Co-authored-by: Tao Luo <luotao02@baidu.com>
@Kane2011 @luotao1
[Metax] upload the task for Metax GPU platform (#1236)
631d5cf 
* [Metax] upload the task for Metax GPU platform

* [Metax] upload the task for Metax GPU platform

* [Metax] upload the task for Metax GPU platform

* [Metax] upload the task for Metax GPU platform

* Rename pfcc/paddle-hardware/fastdeploy-Metax C5xx-infer-result2.png to pfcc/paddle-hardware/images/fastdeploy-Metax-C5xx-infer-result2.png

* Rename pfcc/paddle-hardware/fastdeploy-Metax C5xx-infer-result1.png to pfcc/paddle-hardware/images/fastdeploy-Metax-C5xx-infer-result1.png

* Update 沐曦-FastDeploy-编译验证-打卡任务.md

* Update 【Hackathon_10th】硬件生态任务合集.md

---------

Co-authored-by: Tao Luo <luotao02@baidu.com>
@zongwave @onecatcn
[Phytium] Phytium fast deploy task (#1245)
3834719 
* [Phytium] Phytium fast deploy task

* Update ERNIE model name

* Update model name

---------

Co-authored-by: onecatcn <kaiwang85@qq.com>
@Yun1Liu @luotao1
海光黑客松 初赛 / 打卡 赛题 (#1239)
272959b 
* 海光黑客松 初赛 / 打卡 赛题

* Update as review

* A fix

* Update as review

* Fix a reference cmd issue

* Update 海光DCU-PaddleOCR应用-打卡任务.md

---------

Co-authored-by: Tao Luo <luotao02@baidu.com>
@cixjohn26 @luotao1
进阶任务文档：此芯-PaddleOCR-VL-1.5在此芯P1芯片上的端侧部署与优化 (#1238)
a9ea657 
* 此芯进阶任务提交：PaddleOCR-VL-1.5 在此芯 P1 芯片上的端侧部署与优化

此芯进阶任务提交

* 进阶任务：PaddleOCR-VL-1.5 在此芯 P1 芯片上的端侧部署与优化

此芯进阶任务提交

* Update 【Hackathon_10th】硬件生态任务合集.md

* Update 【Hackathon_10th】硬件生态任务合集.md

* Update 【Hackathon_10th】硬件生态任务合集.md

* Update 【Hackathon_10th】硬件生态任务合集.md

* Update 【Hackathon_10th】硬件生态任务合集.md

---------

Co-authored-by: Tao Luo <luotao02@baidu.com>
@ownia @luotao1
Update CIX PaddleOCR-VL-1.5 deployment tasks and guidelines (#1249)
fbd81b1 
* Update CIX PaddleOCR-VL-1.5 deployment tasks and guidelines

Signed-off-by: Weizhao Ouyang <weizhao.ouyang@arm.com>

* Update 【Hackathon_10th】硬件生态任务合集.md

* Update 【Hackathon_10th】硬件生态任务合集.md

* Update 【Hackathon_10th】硬件生态任务合集.md

---------

Signed-off-by: Weizhao Ouyang <weizhao.ouyang@arm.com>
Co-authored-by: Tao Luo <luotao02@baidu.com>
@sunzhongkai588
add 65th pfcc meeting agenda (#1250)
7aeb707
@r-cloudforge
Merge remote-tracking branch 'upstream/master'
a8ab6d5
@paddle-bot
Copy link
Copy Markdown

paddle-bot Bot commented Mar 20, 2026

你的PR提交成功，感谢你对开源项目的贡献!
请检查PR提交格式和内容是否完备，具体请参考示例和模版。
Your PR has been submitted. Thanks for your contribution!
Please check its format and content. For this, you can refer to Template and Demo.

@luotao1 luotao1 mentioned this pull request Mar 20, 2026
Open
@luotao1 luotao1 self-assigned this Mar 24, 2026
@luotao1
Copy link
Copy Markdown
Collaborator

luotao1 commented Mar 24, 2026

@chang-wenbin

@cloudforge1 cloudforge1 mentioned this pull request Mar 24, 2026
Closed
10 tasks
@luotao1 luotao1 mentioned this pull request Mar 27, 2026
Merged
5 tasks
@cloudforge1 cloudforge1 mentioned this pull request Mar 30, 2026
Open
@luotao1
Copy link
Copy Markdown
Collaborator

luotao1 commented Mar 30, 2026

已合并的 H9 RFC: #1156

如果觉得原来的RFC没问题，可以直接提代码PR。如果觉得RFC还需要修改，在原来的基础上修改即可，无需提一个新的文件。

@cloudforge1 cloudforge1 force-pushed the task/047-rfc-minimax-m1 branch from 792fb53 to 7abad1f Compare March 30, 2026 10:42
@cloudforge1
Copy link
Copy Markdown
Contributor Author

cloudforge1 commented Mar 30, 2026

了解，将调整为更新原 RFC 文件，不新增文件。代码 PR 同步推进中：PaddlePaddle/FastDeploy#6994

@cloudforge1
Copy link
Copy Markdown
Contributor Author

cloudforge1 commented Mar 30, 2026

已更新，改为直接修改原 RFC 文件。请查看。

@cloudforge1 cloudforge1 force-pushed the task/047-rfc-minimax-m1 branch from 7abad1f to 940b245 Compare March 30, 2026 10:52
@cloudforge1 cloudforge1 force-pushed the task/047-rfc-minimax-m1 branch from 940b245 to 8210271 Compare March 30, 2026 11:37
@cloudforge1 cloudforge1 mentioned this pull request Mar 30, 2026
Merged
@cloudforge1
Copy link
Copy Markdown
Contributor Author

cloudforge1 commented Mar 30, 2026

已按实际实现全面更新 RFC（同一文件名，190 行更新）。主要变更：Triton 算子替代 CUDA C++ 内核、移除 MambaBackend 抽象、补充量化支持与权重加载双路径。原始作者已在元数据中标注。

@r-cloudforge r-cloudforge mentioned this pull request Apr 10, 2026
Closed
10 tasks
@r-cloudforge r-cloudforge force-pushed the task/047-rfc-minimax-m1 branch from 8210271 to 84f4edb Compare April 13, 2026 17:48
@CloudForge-Solutions CloudForge-Solutions closed this by deleting the head repository Apr 19, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

@luotao1 luotao1

Labels

contributor

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.