ROCm 6.3
This document represents a proposed Change. As part of the Changes process, proposals are publicly announced in order to receive community feedback. This proposal will only be implemented if approved by the Fedora Engineering Steering Committee.
Summary
ROCm is an AMD GPU stack of math libraries. This is an update of the existing ROCm from 6.2 to 6.3
Owner
- Name: Tom Rix
- Email: Tom.Rix@amd.com
Current status
- Targeted release: Fedora Linux 42
- Last updated: 2024-11-27
- [Announced]
- [<will be assigned by the Wrangler> Discussion thread]
- FESCo issue: <will be assigned by the Wrangler>
- Tracker bug: <will be assigned by the Wrangler>
- Release notes tracker: <will be assigned by the Wrangler>
Detailed Description
ROCm is a set of math libraries that run on AMD GPUs. By running computations on a GPU, an application will have an order of magnitude performance improvement over CPU use and free the CPU to do other activities. ROCm is used by user AI applications as well as Fedora's pytorch and others.
In addition to updating ROCm in Fedora to the latest publicly available, there are several features in this release.
The development focus has been to enable ROCm for laptops by enabling APU's in the latest three generations, from newest to oldest, Strix Point, Phoenix and Rembrandt. This will allow our users to run AI and other GPU applications on more user friendly hardware.
The expansion of our supported GPU set is possible because of a rocm-llvm compiler feature to compress the GPU code in a binary. This removes the past release's need to split the libraries. There will be only one version of the library and it will be in the expected system locations.
Some new packages have been added.
- hiprt : HIP Raytracing https://gpuopen.com/hiprt/ . Hiprt is expected to be used by blender.
- rocdecode : ROC Decode https://rocm.docs.amd.com/projects/rocDecode/en/latest/ . rocdecode is for GPU assisted video decoding.
- rocm-rpp : ROCm Performance Primitives https://rocm.docs.amd.com/projects/rpp/en/latest/. rocm-rpp is part of the dependencies needed for MIVisionX, AMD's vision library.
AI ROCm accelerated packages
- python-torch
- llama-cpp
- ollama (WIP)
- whisper-cpp (WIP)
Feedback
Benefit to Fedora
Updates the ROCm stack to the latest public and enables users that are doing AI on laptops, desktops or data centers.
Scope
- Proposal owners:
This is another major update to ROCm, these have happened regularly since F39. Update ROCm and ensure the users can build and use the new stack as-is.
- Other developers:
Applications that use ROCm should test if the integration is working as it has worked in the past.
- Release engineering: #Releng issue number
- Policies and guidelines: N/A (not needed for this Change)
- Trademark approval: N/A (not needed for this Change)
- Alignment with the Fedora Strategy:
Upgrade/compatibility impact
Early Testing (Optional)
ROCm and it's dependencies are tracked in this COPR.
https://copr.fedorainfracloud.org/coprs/g/rocm-packagers-sig/RH/
How To Test
ROCm requires an AMD GPU to test. The COPR produces binaries for the Radeon 7900. To test with COPR, install the packages on desktop with a 7900 and run.
Each of the ROCm packages has an option -test subpackage. These can be built by adding the --with test to a local mock build. Install the -test package and run. Some instructions on how to run are the in the package spec files.
User Experience
No gpu were removed. Some new gpu's will be enabled, gfx1035, gfx1103, gfx1151, gfx1152 (likely)
Users doing module load rocm/gfx* to use the split libraries will no longer have to.
Package build times will improve because duplicate builds were removed.
Packages will no longer have gpu specific subpackages, they will look like normal packages.
Dependencies
These packages are tracked in the testing COPR
https://copr.fedorainfracloud.org/coprs/g/rocm-packagers-sig/RH/packages/
The most dependent are
- blender
- pytorch (owned by ROCm packaging)
Contingency Plan
Dependent packages should disable their ROCm feature if there is an issue that can not be worked out by the time of freeze.
Documentation
N/A (not a System Wide Change)