Pre-built AI wheels for Windows, no compiler required

What it does

AI-windows-whl is a curated index of pre-compiled Python wheels for high-performance AI libraries—Flash Attention, xFormers, SageAttention, NATTEN, Triton, and others—that are notorious for lacking official Windows builds. The repository centralizes direct download links for .whl files matched to specific Python, PyTorch, and CUDA versions, and it includes a searchable web page to navigate the growing matrix of binaries.

The interesting bit

Windows is the neglected stepchild of the CUDA ecosystem, and this repository treats “find a compatible binary” as the core problem. It hosts some wheels directly on Hugging Face while aggregating third-party builds, turning a scattered forum hunt into something resembling a package index.

Key highlights

• Covers libraries that usually require source builds: Flash Attention, xFormers, SageAttention, NATTEN, Triton (Windows fork), and bitsandbytes.
• Targets the ComfyUI and broader PyTorch Windows community.
• Provides a searchable web UI to filter wheels by Python, PyTorch, and CUDA version.
• Maintains compatibility matrices across multiple PyTorch stable and nightly releases.

Caveats

• The collection is a mix of self-hosted Hugging Face assets and third-party links, so provenance varies by package.
• Link rot is an acknowledged hazard; the README includes a “Report a Broken Link” issue template.
• The README is dominated by long, manually curated tables, so version coverage is best-effort.

Verdict

Bookmark this if you run ComfyUI or local PyTorch on Windows and would rather not spend time on complicated, error-prone compilation. Linux users and anyone who enjoys building from source can safely ignore it.