On 2024-10-09 17:18, PICCA Frederic-Emmanuel wrote: > It is really nive to have this feature. thanks to your efforts. > Do you know if this service will integrate the official ci.debian.org at some > point ?
It's definitely a goal, but it'll take years to get there. This is because we have workers for many GPU architectures, but there is no way in debci to express this. On ci.rocm.debian.net, we abuse the architecture field by adding modifiers, eg: amd64+gfx1030 amd64+gfx1032 ... This probably also needs a Policy change, to express eg: "this package is only useful with a (certain) GPU." Then, we need to host the actual hardware. Finally, we need to figure out how to to this securely. > Is there a sort of plugin system in autopkgtest in order to customize a test > run. > This way rocm could put a special file somewhere and this file could > customize the run process. Taking care to check that pre-required are ok n > order to run a test. > like checking the /dev/kfd. I'll implement this sooner than later. > I just bought and RX 6400, I need to setup this and start to instrument for > now clpeak. > I need something special in order to run the test localy ? If you enable our APT repository [1], you will find two packages to help you with that: * rocm-qemu-support if you want to test in QEMU VMs * rocm-podman-support for rootless podman containers Note that with QEMU, you will not be able to use the GPU for anything else (it's assigned to VFIO), so I suggest podman. The command `rocm-podman-setup -u <your-user>` will tell you all the necessary steps for setting up your user for GPU-in-podman use. rocm-podman-create can be used to create a suitable image. With autopkgtest, you can then use this as $ autopkgtest ... - podman+rocm <image-name> Note the podman+rocm, which is our custom backend. Best, Christian [1]: https://apt.rocm.debian.net/
