Linux + ROCm: January 2026 Stable Configurations Update

Djip · January 23, 2026, 9:51pm

for me today;

on a fedora43 clean toolbox run on a silverblue-43: (so have to be the same on a clean fedora-43 install=

sudo dnf install python3-torch python3-torchaudio python3-torchvision
# [...]
$ python3 -c 'import torch' 2> /dev/null && echo 'Success' || echo 'Failure'
Success
$ python3 -c 'import torch; print(torch.cuda.is_available())'
True

So not that bad. Did you have simple test to know if it really work?

Edit:

git clone https://github.com/pytorch/examples.git
cd examples/mnist
python3 main.py

# look to work:
Train Epoch: 14 [57600/60000 (96%)]	Loss: 0.027002
Train Epoch: 14 [58240/60000 (97%)]	Loss: 0.007073
Train Epoch: 14 [58880/60000 (98%)]	Loss: 0.015837
Train Epoch: 14 [59520/60000 (99%)]	Loss: 0.009342

Test set: Average loss: 0.0287, Accuracy: 9909/10000 (99%)

sounds · January 24, 2026, 8:04pm

I’ll post a footnote, my Framework Desktop is working stably with:

Ubuntu 25.10 (add amd_iommu=off to get a 6% speedup source)

Kernel 6.17.7 (use the mainline ppa to install it)

Python 3.12 (use uv python install 3.12 and then use the python3.12 to do python3.12 -m venv yourdir)

Rocm 7.1 (when installing torch, torchvision, torchaudio wheels in the venv)

InvokeAI from pip (well, uv) - ok

ComfyUI from github - ok, no flash-attention 2

Python torch commands - ok

Lemonade - ok

llama.cpp - ok, I downloaded the latest build from Releases · ggml-org/llama.cpp · GitHub - find the Linux > Ubuntu > “Ubuntu x64 (Vulkan)” build

Nicolas_May · February 1, 2026, 2:06am

For anyone running Fedora 43 on Framework Desktop who just wants to know what commands get you a working PyTorch project, I created a GitHub gist: PyTorch on Framework Desktop with Fedora 43 - 2026-01-31 · GitHub

Lmk if comments/improvements. Good luck and godspeed.

Djip · February 1, 2026, 1:14pm

well… I think there is to much element installed.

for native fedora: fedora-43 is now back stable on latest upgrade:
fedora have build-in rocm-6.4.4 support..

sudo dnf install python3-torch python3-torchaudio python3-torchvision
# [...]
$ python3 -c 'import torch' 2> /dev/null && echo 'Success' || echo 'Failure'
Success
$ python3 -c 'import torch; print(torch.cuda.is_available())'
True

git clone https://github.com/pytorch/examples.git
cd examples/mnist
python3 main.py

# look to work:
Train Epoch: 14 [57600/60000 (96%)]	Loss: 0.027002
Train Epoch: 14 [58240/60000 (97%)]	Loss: 0.007073
Train Epoch: 14 [58880/60000 (98%)]	Loss: 0.015837
Train Epoch: 14 [59520/60000 (99%)]	Loss: 0.009342

Test set: Average loss: 0.0287, Accuracy: 9909/10000 (99%)

that’s all no need for rocm-repos install
In that case it use a python3.14 .

newer rocm / older python:
you can use the https://rocm.nightlies.amd.com/v2/gfx1151 like you did with your pyproject.toml… but do not install rocm from fedora they have 2 different rocm release (6.4.4 vs 7.2?) and it clash (python venv did not isolate correctly from native rocm…)

If you do not need pure romc7+ (like for llama.cpp build…) do not use Index of /rocm/el9/latest/main/ repos…
You can test toolbox with FC44 to get rocm-7.1 if you want.
Or some on the therock toolbox (or create one), but this need to add ROCM paths:

somthing like that

Containerfile.therock-nightly :

# build:
# from repos: https://therock-nightly-tarball.s3.amazonaws.com

# - podman build -t therock-toolbox:7.11 -f Containerfile.therock-nightly --build-arg THEROCK_VER=gfx1151-7.11.0a
# - podman build -t therock-toolbox:7.12 -f Containerfile.therock-nightly --build-arg THEROCK_VER=gfx1151-7.12.0a

# create:
# - toolbox create --image therock-toolbox:7.11 therock-devel-7.11
# - toolbox create --image therock-toolbox:7.12 therock-devel-7.12

# create:
# - toolbox enter <> ...

#=======================================================================================================
FROM registry.fedoraproject.org/fedora-toolbox:43

# update OS
# + ajouter qq element pour les builds llama.cpp
RUN dnf -y --nodocs --setopt=install_weak_deps=False \
    install \
      curl tar xz git-lfs patch \
      make gcc-c++ cmake libcurl-devel ninja-build radeontop libatomic \
 && dnf clean all && rm -rf /var/cache/dnf/*

# find & fetch the latest Linux M.N.0rc tarball (gfx1151 / gfx110X-all)
WORKDIR /tmp
#ARG THEROCK_VER=gfx1151-7.11.0a
ARG THEROCK_VER=gfx1151-7.12.0a

RUN set -euo pipefail; \
    BASE="https://therock-nightly-tarball.s3.amazonaws.com"; \
    PREFIX="therock-dist-linux-${THEROCK_VER}"; \
    KEY="$(curl -s "${BASE}?list-type=2&prefix=${PREFIX}" \
      | grep -o "therock-dist-linux-${THEROCK_VER}[0-9]\{8\}\.tar\.gz" \
      | sort | tail -n1)"; \
    echo "Latest tarball: ${KEY}"; \
    curl -L --fail -o therock.tar.gz "${BASE}/${KEY}" \
    && mkdir -p /opt/rocm \
    && tar xzf therock.tar.gz -C /opt/rocm --strip-components=1 \
    && rm therock.tar.gz

# les env pour rocm
ENV ROCM_PATH=/opt/rocm \
    HIP_PLATFORM=amd \
    HIP_PATH=/opt/rocm \
    HIP_CLANG_PATH=/opt/rocm/llvm/bin \
    HIP_INCLUDE_PATH=/opt/rocm/include \
    HIP_LIB_PATH=/opt/rocm/lib \
    HIP_DEVICE_LIB_PATH=/opt/rocm/lib/llvm/amdgcn/bitcode \
    PATH=/opt/rocm/bin:/opt/rocm/llvm/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin \
    LD_LIBRARY_PATH=/opt/rocm/lib:/opt/rocm/lib64:/opt/rocm/llvm/lib \
    LIBRARY_PATH=/opt/rocm/lib:/opt/rocm/lib64 \
    CPATH=/opt/rocm/include \
    PKG_CONFIG_PATH=/opt/rocm/lib/pkgconfig
    
# good for llama.cpp hip build for APU:
ENV GGML_CUDA_ENABLE_UNIFIED_MEMORY=ON

# make /usr/local libs visible without touching env
RUN echo "/usr/local/lib"  > /etc/ld.so.conf.d/local.conf \
 && echo "/usr/local/lib64" >> /etc/ld.so.conf.d/local.conf \
 && ldconfig

Note: only rocm, not python/pytorch for this one.

And if some want to test someting on fedora43 advise to use toolbox, it won’t break you main OS, and have full OS isolation.

Djip · February 1, 2026, 7:23pm

for vllm there is a new build:

# in a venv ;)
pip install vllm==0.14.0+rocm700 --extra-index-url https://wheels.vllm.ai/rocm/0.14.0/rocm700

Note: did not test, I never have use vllm so not sure to know if it work

edf · February 23, 2026, 4:45am

What is the specific working configuration recommended by the video?

And while we’re at it, what’s the TLDR-equivalent acronym for “nope, not gonna watch a video”?

kyuz0 · February 23, 2026, 10:08am

I make videos, but I also publish ALL documentation on github and there’s even a landing page dedicated to the toolboxes and the known working configurations:

Topic		Replies	Views
Sadly, ROCm Remains VaporWare Framework Desktop	32	5309	June 8, 2025
AMD Strix Halo Llama.cpp Installation Guide for Fedora 42 Framework Desktop framework-desktop-ai-max-300 , ai	18	6937	January 14, 2026
Llama.cpp/vLLM Toolboxes for LLM inference on Strix Halo Framework Desktop	56	8560	February 2, 2026
AMD ROCm does not support the AMD Ryzen AI 300 Series GPUs Framework Laptop 13 framework-laptop-13-amd-ai-300 , ai	57	14764	December 9, 2025
Installing ROCm 6.4 on Fedora 42 Linux fedora	15	3263	October 20, 2025

Linux + ROCm: January 2026 Stable Configurations Update

Related topics