Mirror of `kernels-community/*` stopped syncing, and has no kernel version branches
维护者通常 2 天内回复
还没有人认领这个 Issue。
评估
- 难度
- 4/5
- 预计耗时
- 3-5 天
- 新手友好度
- 45/100
- Issue 类型
- 功能
- 描述清晰度
- 描述清楚
- 活跃度
- 活跃
- 技术栈
- python
调研方向
Start by reading the repository syncing logic and the modelscope kernels patch that maps list_repo_refs to branches and tags; the issue does not name their file paths. Trace how the Hugging Face kernel repo type and v* branches are discovered and mirrored, then verify that version=3 resolves to the expected branch. Done means syncing includes the version branches and ModelScope can serve the current v3 build.
由索引模型根据 Issue 内容生成。
描述
Describe the feature
kernels-community/flash-attn2 on ModelScope (https://www.modelscope.cn/models/kernels-community/flash-attn2) has stopped following the Hugging Face repo:
- The last kernel upload was 2026-07-15 ("Uploaded using
kernel-builder"). The only later commit is "Add model repository removal notice" (2026-09-03). Hugging Face moved kernels to the newkernelrepo type and announced the removal of the old model-type repos starting 2026-09-13, which looks like where the sync source disappeared. - The mirror has only
master. The Hugging Face repo hasv1,v2andv3version branches, andkernels.get_kernel(..., version=N)(also used by transformers'flash_attention_2fallback, which asks forversion=3) resolves them by branch name. - So the
v3torch-stable-ABI builds (any torch >= 2.10) are missing, including the 2026-10-20 rebuild that fixes the GQA backward crash (huggingface/kernels-community#1085).masterserves the older per-torch builds.
ms-swift already works around this (swift/utils/hub_utils.py): "ModelScope does not expose kernel generation branches (v1/v2/...). The requested generation is ignored and the default branch is used instead." modelscope's own kernels patch maps list_repo_refs to ModelScope branches and tags, so it would resolve version=3 once the branches exist.
Request: resume syncing the kernels-community kernels from the Hugging Face kernel repo type, including the v* version branches.
Motivation
We're adding ModelScope as an alternative source for the FlashAttention-2 kernel in TuFT (agentscope-ai/TuFT#161), for users who can't reach the Hugging Face Hub. Right now that source can only offer the older build line, not the one transformers loads from the Hub by default.
I'll raise PRs for this soon.
@tastelikefeet @wangxingjun778
- 主要语言
- Python
- 星标
- 9.2k
- 派生
- 977
- 平均合并
- 4 天 5 小时
- 30 天内合并 PR
- 10
环境准备
这个项目没有提供开发容器、Dockerfile 或贡献指南，环境需要你自己搭建：先看它的 README，通用步骤见我们的新手贡献指南。
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
modelscope/modelscope 的其他 Issue
-
希望创空间能增加变更代码推送自动重新部署功能可能已有人在做 @yingdachen 于 2 天前认领。 未关闭
modelscope/modelscope#1851 · 已指派 3 人 ·
维护者通常 2 天内回复
-
[API-Inference] GLM-5.3-Flash returns HTTP 401 while Qwen3.8-Flash-Next works with the same token未关闭
难度 4/5 3-5 天 新手友好度 45/100
modelscope/modelscope#1850 ·
维护者通常 2 天内回复
-
难度 4/5 3-5 天 新手友好度 35/100
modelscope/modelscope#1849 · 1 条评论 ·
维护者通常 2 天内回复
-
关于 model scope 的API 推理接口是否已支持了音频推理？可能已有人在做 @yingdachen 于 11 天前认领。 未关闭
modelscope/modelscope#1842 · 3 条评论 · 已指派 3 人 ·
维护者通常 2 天内回复
-
[API-Inference][modelscope.ai] Qwen3.8-Max Responses API returns HTTP 405 security block for inert code text可能已有人在做 @yingdachen 于 18 天前认领。 未关闭
modelscope/modelscope#1837 · 已指派 3 人 ·
维护者通常 2 天内回复
查看 modelscope/modelscope 的全部 Issue
相似的 Issue
-
enhancement
难度 2/5 1-3 小时 新手友好度 76/100
Sienna-Platform/SiennaSchemas#130 ·
维护者通常 3 天内回复
-
难度 2/5 1-3 小时 新手友好度 72/100
-
难度 1/5 1 小时以内 新手友好度 85/100
firedrakeproject/Irksome#293 ·
维护者通常 2 天内回复
-
难度 2/5 1-3 小时 新手友好度 78/100
ida-arbeitszeit/workers-control#1578 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 78/100
aws-samples/sample-per-model-bedrock#898 ·
维护者通常 1 天内回复