thiagoakhe/custom-builds

Personal ComfyUI Wheels Backup

This repository serves as a personal backup for ComfyUI wheels, specifically compiled and optimized for bleeding-edge environments.

System Specifications & Compatibility:

• Python: 3.14
• PyTorch: 2.1.1 and 2.1.2
• CUDA Toolkit: 13.0 and 13.2
• Compute Capability: sm_89 (NVIDIA RTX 40-series / Ada Lovelace)
• OS: Kubuntu 26.04

Hardware used for compilation:

• CPU: AMD Ryzen 7 5700X (8-Core)
• RAM: 32GB DDR4
• GPU: RTX 40-series (sm_89)

Included Wheels:

1. Flash Attention 2 (v2.8.4): Locally compiled from source.
2. Xformers (v0.0.35): Locally compiled from source.
3. Nunchaku (1.3.0.dev20260513): Locally compiled from source.
4. Nunchaku (1.3.0.dev20260518): Locally compiled from source.
5. Sage Attention 2.2 (torch 2.12 - cuda132): Locally compiled from source.
6. SpargeAttn (0.1.0) - Locally compiled from source.
7. Insightface (0.7.3): Locally compiled from source.
8. Sage Attention 2.2: Sourced from @rozenkristall's Hugging Face repository.

How to install packages via PIP

To install any of these files directly from the cloud using pip, you must use the direct download link (raw) from Hugging Face.

Important: Copying the link from your browser's address bar (which contains the /blob/ path) will result in an error because the server will attempt to install the HTML file instead of the binary. The pip installer requires the URL to contain the /resolve/ path.

Example Installation Command Format:

pip install 'https://huggingface.co/thiagoakhe/custom-builds/resolve/main/nunchaku-1.3.0.dev20260518+cu13.2torch2.12.sm89-cp314-cp314-linux_x86_64.whl'