PengyOS has shipped AMD graphics on the Server for a while, but nothing could use the GPU for
compute. The kernel had no CONFIG_HSA_AMD, so /dev/kfd never existed and no ROCm workload
ran on any edition. The symbol has no Kconfig default: it was simply never on. Now it is, along
with HSA_AMD_SVM and device-private memory. ROCm stays inside the container, the same way the
CUDA toolkit does on Server NVIDIA, so the image still builds from source and the vendor stack
stays where the vendor put it.
Three things had to move together for that to be usable. podman now runs containers with
crun, because rootless --group-add keep-groups needs it to hand the render and video groups
into the container. The 8 MiB locked-memory default is gone (DefaultLimitMEMLOCK=infinity),
which is what --ulimit memlock=-1 actually requires. And pengyos-gpu-memory status | set SIZE
| reset raises how much system RAM the GPU may map, from the kernel's default of half of RAM up
to RAM minus 4 GiB, written into /etc/modprobe.d for the next boot. If you are running a big
model on something like a Strix Halo, that's the knob: a 70-120 GB model needs the aperture, not
just the GPU.
Every image also has /usr/share/pengyos.txt now: one page, written at build time from what
that image actually holds, covering the read-only system and A/B updates, installing software,
the dev guest and dx, timers instead of cron, the container lanes, the firewall, and the
package list. The login banner's first line points at it, as do /etc/os-release, the stubs
that print "this isn't installed, use this instead", and a PENGYOS_README variable. The point
is that a new machine should answer "how do I do things here?" without anyone hunting.