Re: Building Magma with AMD GPU support

[Mo Zhou <lumin@debian.org>]

Hi Cordell,

Thanks for working on this.

As you may have noticed, src:pytorch (main) and src:pytorch-cuda 
(contrib) is the identical source but uploaded twice due to the 
difference in their sections. This is found to minimize my effort 
compared to maintaining two separate sources, especially when I need to 
apply the same logic to many other packages like src:gloo, 
src:tensorpipe, etc.

For magma I'd personally prefer my own approach. Maybe you can just 
refer to the debian/cudabuild.sh and debian/rocmbuild.sh from 
src:pytorch, and see whether this works for you. In that sense we can 
avoid duplicated working repository which is nothing but requiring 
double human effort. Namely, a debian/rocmbuild.sh conversion script, 
and a control.rocm file targeting at Section: main should be good to go.

That said, anybody is welcome to comment if there is any better approach 
to reduce human effort for such case.


On 12/18/24 23:53, Cordell Bloor wrote:
> Hi Mo,
>
> I was building PyTorch and noticed that Magma [1] is a dependency for 
> some configurations. There is a magma package with NVIDIA GPU support 
> in contrib [2], but we don't have an AMD GPU version packaged for 
> Debian. It took a bit of trial and error to successfully build the 
> library, so I thought I'd share instructions for building magma with 
> AMD GPU support:
>
> sudo apt-get -y install git build-essential libopenblas-dev gfortran hipcc librocblas-dev libhipblas-dev librocsparse-dev libhipsparse-dev
> git clonehttps://github.com/icl-utk-edu/magma.git
> cd magma
> git checkout v2.8.0
> echo -e 'BACKEND = hip\nFORT = true\nGPU_TARGET = gfx803 gfx900 gfx906 gfx908 gfx90a gfx1010 gfx1030 gfx1100 gfx1101 gfx1102' > make.inc
> sed -i '1s/python$/python3/' tools/codegen.py
> sed -i 's/hip::host/hip::device/' CMakeLists.txt
> make generate
> CXX=hipcc cmake -S. -Bbuild -DBLA_VENDOR=OpenBLAS -DAMDGPU_TARGETS="gfx803;gfx900;gfx906;gfx908;gfx90a;gfx1010;gfx1030;gfx1100;gfx1101;gfx1102" -DMAGMA_ENABLE_HIP=ON
> make -j16 -C build
>
> I've only tested on my local workstation, but the above commands 
> _should_ result in a magma library that runs on any discrete AMD GPU 
> since Vega (excluding MI300).
>
> This AMD GPU build takes a long time so it would be nice to provide a 
> binary package. I'd be happy to help maintain the magma package, but I 
> think I will need your help to get it started. In particular, it's not 
> clear to me how to organize the package sources to minimize duplicate 
> work between the NVIDIA and AMD variants. I'm also unsure of what 
> conventions to follow for package naming.
>
> Sincerely,
> Cory Bloor
>
> [1]: https://github.com/icl-utk-edu/magma
> [2]: https://tracker.debian.org/pkg/magma