I noticed that the ROCm version that gets installed via dnf is 6.3. But apparently all the good stuff for the Ryzen AI MAX+ 395 is in ROCm 6.4. Has anyone tried to install 6.4 on their Framework Desktop running Fedora 42? I would imagine this requires building from source, but maybe someone knows about pre-built binaries?
(trying to get ROCm 6.4 installed so I can run Pytorch with the ROCm backend)
Wait until Fedora 43 or use Fedora Rawhide if you can’t wait. These are probably the easiest options.
You can’t just install it on Fedora 42 since there are dependency issues with Pytorch and stuff.
I would look at just containerizing everything. You can check out GitHub - kyuz0/amd-strix-halo-toolboxes to start.
This way you can pick and choose ROCm/Vulkan versions without having to muck around with your base OS.
or you can use/test fedora_43 in a toolbox with :
> toolbox create --distro fedora --release f43
> toolbox enter fedora-toolbox-43
# and install rocm6.4 + ... from fedora repos
edited: thanks @Jonathan_Haas yes 43 not 36 
That doesn’t look right.
Toolbox might be worth a try, I’m just not sure that it will work since you can’t install/use kernel stuff or other hardware-adjacent stuff inside a toolbox container.
Haven’t used toolbox before. So I followed their directions there, created and entered the llama-rocm-6.4.3-rocwmma toolbox. I guess I’m surprised that there aren’t any devtools there - no python, no compilers, etc. So you basically have to install everything?
Yes, so you can install only what you need. it make easy to test what you need to create container.
Fedora 42 and 43 share the same kernel (within a few days) so not realy a probleme with that.
And I have already test the rocm-6.4 with it and llama.cpp it work. so yes it work.
oups yes… bad copy/past corrected, thanks.
Good to know. I know some tools like hwinfo don’t work properly inside toolbox, because they see the emulated hardware of the toolbox container and not the real hardware, so I want sure if that would work.
Just an update to tell how I got Pytorch working with ROCm on the Framework Desktop. I found this page helpful: PyTorch on ROCm — ROCm installation (Linux)
Installed this docker image which seemed the most up to date: rocm6.4.3_ubuntu24.04_py3.12_pytorch_release_2.6.0
(see this page: https://hub.docker.com/r/rocm/pytorch/tags )
Then inside of that image I cloned the pytorch examples repo and tried running an example - it failed with some kind of HIP error. After googling around some I determined that the prebuilt pytorch in the image was not built with support the GPU in Ryzen AI Max 395. So then I followed the instructions in the Pytorch on ROCm page above for building Pytorch (clone Pytorch, cd to pytorch, export PYTORCH_ROCM_ARCH=gfx1151, build with .ci/pytorch/build.sh ) then it seemed to work, except there were errors related to torchvision - turns out that I needed to build torchvision from source as well so that it would match torch. Now Pytorch is working using the ROCm backend.