rocmroll-acceleration-libraries

ROCmRoll Acceleration Libraries

Prebuilt packages for Flash Attention 2 and SageAttention 2, compiled from source for AMD GPUs running ROCm on Windows.

This repository is maintained as part of the ROCmRoll ecosystem and provides optimized binary artifacts that ROCmRoll can consume during ComfyUI installation, update, and repair operations.

Overview

Building attention libraries from source on Windows can require a correctly configured compiler toolchain, compatible Python and PyTorch versions, ROCm development components, and architecture-specific build settings.

This repository centralizes prebuilt packages so ROCmRoll users can install supported attention implementations without compiling them locally.

The main objectives are:

• provide reproducible packages built from upstream source;
• reduce ROCmRoll installation time;
• avoid requiring build tools on end-user systems;
• maintain compatibility with ROCm-enabled PyTorch builds;
• provide packages optimized for supported AMD GPU architectures;
• allow ROCmRoll to automatically select the correct artifact.

Available Packages

Flash Attention 2

Flash Attention 2 provides optimized attention kernels designed to reduce memory usage and improve attention performance.

Packages published here are built from source against selected combinations of:

• Python;
• PyTorch;
• ROCm;
• Windows;
• AMD GPU architectures.

SageAttention 2

SageAttention 2 provides optimized and quantized attention kernels intended to improve inference performance while reducing memory bandwidth requirements.

Packages are compiled from source and validated for use with supported ROCmRoll environments.

ROCmRoll Integration

ROCmRoll is a Windows platform manager for creating, launching, updating, diagnosing, and repairing portable ComfyUI installations optimized for AMD GPUs with ROCm.

During an instance installation or repair, ROCmRoll can:

1. detect the configured Python version;
2. identify the installed PyTorch and ROCm versions;
3. detect the AMD GPU architecture;
4. resolve a compatible attention package from this repository;
5. download and install the selected artifact;
6. record the installed package in the instance state.

This repository is intended to act as an artifact source. Package resolution, compatibility validation, installation, and lifecycle management are handled by ROCmRoll.

Compatibility

Package compatibility depends on the complete runtime combination, including:

• operating system;
• Python version;
• PyTorch version;
• ROCm version;
• GPU architecture;
• package version;
• build configuration.

A package built for one PyTorch or ROCm version may not work with another version, even when the Python version is the same.

ROCmRoll should be used whenever possible because it validates the environment before selecting an artifact.

Manual Installation

Manual installation is intended primarily for testing and troubleshooting.

Download the wheel matching your exact environment and install it with the Python executable used by your ComfyUI instance:

& "C:\path\to\python.exe" -m pip install "C:\path\to\package.whl"

To replace an existing installation:

& "C:\path\to\python.exe" -m pip install `
    --force-reinstall `
    --no-deps `
    "C:\path\to\package.whl"

Verify the installed package:

& "C:\path\to\python.exe" -m pip show flash-attn

Or:

& "C:\path\to\python.exe" -m pip show sageattention

The import name may differ from the package distribution name. Refer to the corresponding upstream project documentation for library-specific usage.

Source and Licensing

Flash Attention 2 and SageAttention 2 are developed and maintained by their respective upstream projects.

This repository does not claim ownership of the upstream source code. Packages are redistributed according to the licenses of their respective projects.

Consult the upstream repositories for:

• source code;
• license terms;
• implementation documentation;
• supported features;
• known limitations.

ROCm, AMD, PyTorch, ComfyUI, Flash Attention, and SageAttention are trademarks or project names belonging to their respective owners.

Support

For issues involving package selection or installation through ROCmRoll, open an issue in the ROCmRoll repository:

https://github.com/tomsnunes/rocmroll/issues

When reporting an issue, include:

• ROCmRoll version;
• GPU model;
• GPU architecture;
• Windows version;
• Python version;
• PyTorch version;
• ROCm version;
• package filename;
• installation logs;
• complete error message.

For problems originating in the attention implementation itself, consult the corresponding upstream project.

Disclaimer

These packages are provided on a best-effort basis for compatible AMD ROCm environments.

They are not official builds from AMD, PyTorch, Flash Attention, SageAttention, ComfyUI, or Hugging Face.

Compatibility is limited to the runtime combinations explicitly documented for each artifact.