cumm-spconv

📦 Prebuilt Wheels for `cumm` and `spconv`

💡 This repository provides a PyPI simple index of prebuilt wheels for my custom cumm and spconv.

⚡ By using this index, you can conveniently install via pip install, saving time on local builds and configuration.

🆕 The new cumm v0.9.1 (and v0.7.14) and spconv v2.4.1 now include additional prebuilt wheels for CUDA 11.3 12.1 and 12.6 along with many fixes and enhancements.

💽 Installation

To install specific wheel(s), use the --extra-index-url with pip install to ensure packages are installed exclusively from this index; for example, for CUDA 13.0:

pip install --extra-index-url https://ratharog.github.io/cumm-spconv/ cumm-cu130 spconv-cu130

And for CPU-only variants:

pip install --extra-index-url https://ratharog.github.io/cumm-spconv/ cumm spconv

Note: --extra-index-url may not provide the right wheels because pip will still install packages from PyPI when the package name and version are the same (or higher) on PyPI, which cause it to ignore the custom wheels from this index. To make sure the correct wheels from this index are installed, you can use --index-url or -i; for example, for CUDA 13.0:

pip install --extra-index-url https://ratharog.github.io/cumm-spconv/ cumm-cu130 spconv-cu130
pip uninstall -y cumm-cu130 spconv-cu130
pip install -i https://ratharog.github.io/cumm-spconv/ cumm-cu130 spconv-cu130

🌟 Recommended version pairings for prebuilt wheels

The pairing table is updated as new versions arrive; always check here before upgrading either package to avoid incompatibility. Always use the recommended version pairs of cumm and spconv as below:

Recommended Version Pairs `cumm` ⟵ `spconv`	Available Prebuilt CUDA Wheels	Supported CUDA Versions (Source Build)
`0.9.1` (latest) ⟵ `2.4.1` (latest)	`cu126`, `cu128`, `cu130`	12.6+, 13.x
`0.9.0` ⟵ `2.4.0`	`cu128`, `cu130`	12.6+, 13.x
`0.8.5` ⟵ `2.3.11`	`cu128`, `cu130`	12.6+, 13.x
`0.8.4` ⟵ `2.3.10`	`cu128`, `cu130`	12.6+, 13.x
`0.8.3` ⟵ `2.3.9`	`cu128`	12.6+
`0.7.14` ⟵ `2.4.1` (latest)	`cu113`, `cu121`	11.x, 12.0-12.6

If you install without specifying versions, the versions you downloaded may not match the highest recommended pair. This often means a new spconv release is coming soon, and cumm release may already be ahead.
If there is no prebuilt wheel available for a specific minor version of CUDA, you can always build from source:
- For CUDA 12.6+, use the latest cumm-gpu and spconv-gpu.
- For CUDA 11.0-12.6, use cumm-gpu@v0.7.14 and the latest spconv-gpu.

For production or reproducible installs, explicitly use the recommended version pairs:

pip install cumm-cu130==0.9.0 spconv-cu130==2.4.0 --extra-index-url https://ratharog.github.io/cumm-spconv/

pip install cumm-cu128==0.9.0 spconv-cu128==2.4.0 --extra-index-url https://ratharog.github.io/cumm-spconv/

Notes:

cumm-gpu@v0.7.14 also supports CUDA 12.6; however, the prebuilt cumm-cu126 and spconv-cu126 wheels are based on the latest versions of cumm-gpu@v0.9.1 and spconv-gpu@v2.4.1.

🔭 Build scope

Wheels	Linux	Windows	macOS	Python Versions
`cumm`	✅	✅	✅	3.9-3.14
`spconv`	✅	❌	❌	3.9-3.14
`cumm-cu113`	✅	❌	❌	3.9-3.11
`spconv-cu113`	✅	❌	❌	3.9-3.11
`cumm-cu121`	✅	❌	❌	3.9-3.11
`spconv-cu121`	✅	❌	❌	3.9-3.11
`cumm-cu126`	✅	✅	❌	3.11-3.14
`spconv-cu126`	✅	✅	❌	3.11-3.14
`cumm-cu128`	✅	✅	❌	3.11-3.14
`spconv-cu128`	✅	✅	❌	3.11-3.14
`cumm-cu130`	✅	✅	❌	3.11-3.14
`spconv-cu130`	✅	✅	❌	3.11-3.14

^{* No ARM (aarch64) support for Linux}
^{* No ARM (ARM64) support for Windows}

📝 License

This site is open source. Improve this page.