F40 Change Proposal: ROCm 6 Release (Self-Contained)

ROCm 6 Release

This is a proposed Change for Fedora Linux.
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.

Wiki
Announced

:link: Summary

The AMD ROCm™ 6.0 is the latest release of AMD’s software optimized for AI and HPC workload performance on AMD GPU’s. This latest release enables the newest flagship datacenter GPU the AMD Instinct™ MI300 as well as continuing the GPUs enabled in their last 5.x release, most/all of their recent GPUs.

:link: Owner

The owners of this change are the HC SIG (SIGs/HC - Fedora Project Wiki)

:link: Current status

  • Targeted release: Fedora Linux 40
  • Last updated: 2024-01-11
  • [ devel thread]
  • FESCo issue:
  • Tracker bug:
  • Release notes tracker:

:link: Detailed Description

The benefits for frogs include: ROCm 6 has expanded support for AMD Instinct™ MI300A and MI300X. It includes highly optimized attention algorithms, and proven collective communications libraries, as well as optimized performance for FP8 support in PyTorch and hipblasLT. It includes prepackaged HPC and AI/ML frameworks with streamlined and improved tools from AMD Infinity Hub.

:link: Feedback

There has been positive feedback from the community for the ease of using and developing GPU accelerated applications within Fedora. Because of the interest in AI, the community has requested that ROCm support be added to PyTorch and other AI applications and frameworks. To address this feedback several packages are in the process of being added to Fedora including rocFFT rocSolver hipBLASLt MiOpen

:link: Benefit to Fedora

Fedora has finally end-to-end open source GPU acceleration. The GPU hardware driver is in the linux kernel. The compiler is the system clang. The ROCm software stack provides the higher level libraries that enable other Fedora packages and user applications to be built entirely with Fedora.

:link: Scope

  • Proposal owners:

The feature owners accomplished packaging the new version of ROCm 6 for Fedora. This provides basic accelerated functionally that should be used by any package that can take advantage of it.

  • Other developers:

If your package or application used the older version of ROCm 5.7.1 or older, you must verify that you can use the new version of ROCm, 6.0 is a major version change.

The ROCm packages are known to build in Rawhide, no additional effort is required.

  • Policies and guidelines: N/A (not needed for this Change)

  • Trademark approval: N/A (not needed for this Change)

  • Alignment with Community Initiatives:

Yes, it aligns with the current Fedora Community Initiatives.

:link: Upgrade/compatibility impact

No hardware was dropped.

:link: How To Test

Installation instructions can be found here (SIGs/HC - Fedora Project Wiki).

This is by and large a system library change and not directly visible to the user.

:link: Dependencies

The basic work has been completed.

:link: Contingency Plan

The basic work has been completed.

:link: Documentation

Documentation can be found here (AMD ROCm™ documentation — ROCm Documentation)

:link: Release Notes

2 Likes

Hi Tom! Came upon this thread from Phoronix, and wished an early heads-up on a potential thrust packaging conflict with the fedora-nvidia repo. At least in fc39 cuda-cccl-devel package from fedora-nvidia repo and rocthrust-devel package from fedora repo both try to install their own presumably conflicting /usr/include/thrust. Yes, I realize Nvidia-specific software isn’t exactly Fedora’s problem (yet); I’ve left a comment concerning this potential conflict at Nvidia proprietary and open source kernel modules – negativo17.org. Thanks!

Could you open a bugzilla against rocthrust so I can track this issue ?
If we were able, would adding a cudathrust package to Fedora resolve this problem ?

1 Like

Done: 2259056 – rocthrust-devel installs /usr/include/thrust in conflict with cuda-cccl-devel package from fedora-nvidia repo
I look forward to fc40 and possibly obtaining Radeon hardware this summer!

This change proposal has now been submitted to FESCo with ticket #3155 for voting.

To find out more, please visit our Changes Policy documentation.

?