Skip to content

Latest commit

 

History

History
112 lines (73 loc) · 6.14 KB

README.md

File metadata and controls

112 lines (73 loc) · 6.14 KB
Raw

Boost Your AMD GPU Performance with ROCmLibs for gfx1103 and Beyond!

**This repository initially created to share optimized ROCm Libraries specifically for the AMD 780M APU's gfx1103 architecture (due to limited official support), It has since grown to include more AMD GPU architectures using the same proven build methods to benefit the community, these libraries are designed to significantly boost performance in popular applications like AI models (e.g., Llama, Stable Diffusion) within the ZLUDA CUDA Wrapper and other ROCm-based environments.

All code is based on the ROCm Official Linux version as explained in the wiki, with additional tweaks and optimizations to address this gap in official support.

Here's a simplified guide on how to get started:

  1. Prerequisites: Ensure you have the HIP SDK installed for Windows or ROCm for Linux.
  2. Download Files: Grab the ROCmLibs-arch titles .zip or 7z file .For example the ROCmlibs for gfx1103 woud be rocm gfx1103 AMD780M phoenix V3 for hip sdk 5.7.7z (originally ROCmLibsgfx1103.zip) from this repository.
  3. Installation:
    • Unzip ROCmLibs-arch titles.zip or 7z. Place the extracted library folder into your %HIP_PATH%\bin\rocblas directory, renaming the existing "rocblas" folder (e.g., to "oldlibrary").
    • Copy the rocblas.dll file from the downloaded ROCmLibs-arch titles.zip or 7z into the rocblas\library folder within your %HIP_PATH%\bin directory (example: C:\Program Files\AMD\ROCm\5.7\bin). Replace the original file.
    • Rename your existing "rocblas\library" folder (e.g., to "origlibrary").
    • Extract the contents of rocm gfx1103 AMD780M phoenix V3 for hip sdk 5.7.7z into the same directory where you placed the ROCmLibs folder.
  4. Reboot: Restart your computer for the changes to take effect (not a must do).

That's it! Your AMD GPU is now ready to harness the power of optimized ROCm libraries.

Performance Advantages:

You can expect significant performance gains, often 2-3 times faster than DirectML, in applications like:

Support and Resources:

This repository also shares some custom-built ROCm Libraries for various AMD GPU architectures on Windows, including (gfx803; gfx902; gfx90c; gfx90c:xnack-; gfx906; gfx1010; gfx1010:xnack-; gfx1011; gfx1012; gfx1012:xnack-; gfx1031; gfx1032; gfx1034; gfx1035; gfx1036; gfx1103; gfx1150(experimental only)).

The builded ROCmlibs for the other arches listed above moved to released page .

Hip SDK 5.7

Hip SDK 6.1.2

Hip SDK 6.2.4

For detailed instructions on adding support for additional GPUs, please refer to the wiki page.

For more information on using ROCmLibs in LM Studio on unsupported AMD GPUs, visit our dedicated wiki: LM Studio

Additional Information & Installation Tips:

Extracting Archives:

You'll need a program like 7-Zip or WinRAR to open the compressed .zip and .7z files in this repository.

ROCm for Linux:

While ROCm is available for Linux on the releases page, recommend against using it directly. There are simpler and more effective methods for achieving compatibility. The easiest approach is using the HSA_OVERRIDE_GFX_VERSION environment variable. For example, setting export HSA_OVERRIDE_GFX_VERSION=11.0.0 can override your system's default and provide support for gfx1103 and similar GPUs.

ROCm for Windows: For the AMD 780M APU on Windows, recommend using one of these files:

  • rocm gfx1103 AMD 780M phoenix V2.0 for hip sdk 5.7.7z: Compatible with HIP SDK 5.7.1
  • rocm gfx1103 AMD780M phoenix V3 for hip sdk 5.7.7z: Compatible with HIP SDK 5.7.1
  • rocm gfx1103 AMD 780M phoenix V4.0 for hip sdk 6.1.2.7z: Compatible with HIP SDK 6.1.2
  • rocm gfx1103 AMD 780M phoenix V5.0 for hip sdk 6.2.4.7z: Compatible with HIP SDK 6.2.4

Important:

  • Always choose the file version that matches your installed HIP SDK.

Selecting the Right ROCm Files:

Choose the appropriate .7z file based on your AMD GPU architecture. Refer to the release page for a complete list of supported architectures.

Important

This 7z archive, rocBLAS-Custom-Logic-Files-for-rx580-vega8-90c-navi10-navi12-navi14-navi22-navi23-navi24-rembrandt-navi26-phoenix.7z, contains logic files optimized for various AMD GPUs (Rx 580, Vega series, Navi 10-26, Rembrandt, Phoenix). To build rocBLAS and utilize these custom logic files, please refer to the detailed guide provided in the wiki

Creating Issues:

Please refrain from creating issues in this repository for unrelated topics like ollama-for-amd. Maintaining a well-organized project benefits everyone. Therefore, please submit issues related to ollama-for-amd to the correct repository. This helps keep things clear and streamlined.

Much appreciate your understanding and cooperation!

Let me know if you have any further questions or need assistance with your ROCm setup!