zerogpu-blackwell-wheels
收藏资源简介:
该数据集是一个专门为NVIDIA Blackwell (sm_120) GPU架构预编译的CUDA扩展包(wheel文件)集合。由于PyPI官方仓库目前未为大多数此类库提供针对sm_120架构的预编译包,且Hugging Face ZeroGPU Spaces在启动时无法进行本地编译,因此本数据集旨在为在这些环境中运行的项目提供可直接安装的二进制依赖。数据集涵盖了多个深度学习与计算机视觉领域常用的、需要编译的PyTorch扩展库,包括torchmcubes、nvdiffrast、diff_gaussian_rasterization、xformers、pytorch3d以及flash-attn的两个版本(2.8.3和2.7.4.post1)。这些包针对12种不同的PyTorch、CUDA和Python版本组合进行了构建(例如pt212-cu130-cp313代表PyTorch 2.12、CUDA 13.0、Python 3.13),确保了跨版本兼容性。每个构建单元不仅包含成功的wheel文件,还附带了完整的构建日志和源代码引用,以供审计。用户可以通过提供的URL模板,在项目的`requirements.txt`中直接指定对应版本的包链接进行安装。
This dataset is a collection of CUDA extension packages (wheel files) pre-compiled exclusively for the NVIDIA Blackwell (sm_120) GPU architecture. As the official PyPI repository currently does not offer pre-compiled packages for the sm_120 architecture for most such libraries, and Hugging Face ZeroGPU Spaces cannot perform local compilation upon startup, this dataset is designed to provide directly installable binary dependencies for projects running in these environments. The dataset covers a range of widely used PyTorch extension libraries in the fields of deep learning and computer vision that require compilation, including torchmcubes, nvdiffrast, diff_gaussian_rasterization, xformers, pytorch3d, and two versions of flash-attn (2.8.3 and 2.7.4.post1). These packages are built for 12 distinct combinations of PyTorch, CUDA, and Python versions (e.g., pt212-cu130-cp313 denotes PyTorch 2.12, CUDA 13.0, and Python 3.13), ensuring cross-version compatibility. Each build unit not only contains the successful wheel files but also includes complete build logs and source code references for auditing purposes. Users can directly specify the package link of the corresponding version in the project's `requirements.txt` for installation via the provided URL template.
ZeroGPU Blackwell (sm_120) Wheels 数据集详情
数据集简介
该数据集提供专为 Blackwell (sm_120) 架构构建的 CUDA 扩展 wheel 包。由于 PyPI 目前未发布适用于 sm_120 的 wheel,且 ZeroGPU Spaces 无法在启动时编译,因此 Spaces 应直接从该数据集的 URL 中引用 wheel 文件。
使用方式
URL 模板格式如下:
https://huggingface.co/datasets/multimodalart/zerogpu-blackwell-wheels/resolve/main/wheels/<TAG>/<wheel-filename>
示例(torch 2.12 / cu130 / py3.13 环境,需要 xformers):
https://huggingface.co/datasets/multimodalart/zerogpu-blackwell-wheels/resolve/main/wheels/pt212-cu130-cp313/xformers-0.0.34+3da0fc92.d20260528-cp39-abi3-linux_x86_64.whl
构建矩阵
数据集共包含 12 个标签(tag),每个标签对应特定的 PyTorch 版本、CUDA 版本和 Python 版本组合。每个标签下均构建以下 7 个库的 wheel 包(全部为 ✅ 成功状态):
| 标签 | torchmcubes | nvdiffrast | diff_gaussian_rasterization | xformers | pytorch3d | flash_attn 2.8.3 | flash_attn 2.7.4.post1 |
|---|---|---|---|---|---|---|---|
| pt212-cu130-cp313 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| pt28-cu128-cp310 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| pt212-cu130-cp312 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| pt212-cu130-cp310 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| pt211-cu130-cp313 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| pt211-cu130-cp312 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| pt210-cu130-cp312 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| pt210-cu128-cp310 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| pt291-cu128-cp310 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| pt28-cu128-cp311 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| pt28-cu128-cp312 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| pt212-cu130-cp311 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
注意事项
- flash-attn:每个标签同时提供
2.8.3和2.7.4.post1两个版本,编译时指定FLASH_ATTN_CUDA_ARCHS=120以跳过未使用的 sm 架构,构建时间缩短约 5 倍。 - xformers:针对 torch ≥ 2.10 使用
v0.0.34版本(稳定 ABI),针对 torch 2.8/2.9 回退至v0.0.33.post2(因缺乏torch/csrc/stable/*.h)。当前 master 分支在 OSS 路径上存在问题,提交c04f47b6将memory_efficient_attention限制在 Meta 内部包中。 - torchmcubes:使用
scikit-build-core后端,构建 venv 通过uv引导pip以确保后端正常解析。 - 跳过的标签:
pt212-cu128-cp310和pt211-cu128-cp310因上游 PyTorch 索引未发布 cu128 版本的 wheel(仅 2.8–2.10 支持 cu128),故未构建。 - 审计支持:每个标签文件夹还包含
<lib>.log(完整构建输出)和<lib>.ref(构建时源代码的 git 标签)文件,便于审查 wheel 的构建过程。





