Re: Finishing the ROCm 6 update and few other stuffs I explored

[Christian BAYLE <bayle@debian.org>]

Hello,

will be happy to try to help, coming back to make more work on debian, 
since I have more time for that.

I have got a RX7600XT (16Gb VRAM) board and had some trouble with rocm6 
an pytorch

I tried to build (and mostly managed) to build pytorch-rocm filling the 
holes in the debian/rules files on salsa git
at https://salsa.debian.org/deeplearning-team/pytorch.git
but hit some issues with cuda support not available.

It's quite difficult to understand how all this works, thank you for 
your talk at Ubuntu Summit and pointing it.

I also tried and mostly managed to build rocm from amd source code
from https://github.com/ROCm/ROCm
creating bookworm and trixie images here :
https://github.com/cbayle/ROCm/tree/debian-build/tools/rocm-build/docker


I had to make few changes in the build scripts
	modified:   ROCm.mk
	modified:   build_aqlprofile.sh
	modified:   build_composable_kernel.sh
	modified:   build_hipblaslt.sh
	modified:   build_miopen-deps.sh
	modified:   build_opencl_icd_loader.sh
	modified:   build_openmp_extras.sh
	modified:   build_rocal.sh
	modified:   build_rocblas.sh
	modified:   envsetup.sh
	modified:   post_inst_pkg.sh

You can see here:
https://github.com/ROCm/ROCm/commit/7863614c3cf24c54b10b099848391059a42df59b

not sure this is worth to merge request since build is moving a lot, and 
may not work. I'm interested in your opinion about this.

This gives some information on what is missing in debian to build 6.3

This give some python package missing in trixie:
- argparse
- lit
- cppheaderparser
- Optionnaly rocm-docs-core

I had to downgrade to gcc13 g++13

also lapack grpc turbojpeg need some rebuild

There is also some source code missing, and some place doing pip downloads.

I also tested many Stable Diffusion software that works rather well
in virtualenv with trixie/sid and debian only amd package (kernel and 
rocm), that I think should be interesting to package.


I was wondering if you have some coordination meeting on IRC,
or a kind of debian-ai WNPP or priority to help for trixie


I asked to bridge debian-ai to matrix
https://salsa.debian.org/debiansocial-team/sysadmin/accounts/-/issues/144

As I couldn't see it at:
https://app.element.io/#/room/#debian-space:matrix.debian.social
and that it's convenient to follow live informations

Cheers

Christian B. (chris38 on IRC)