If you deployed a PyTorch model on a NVIDIA Desktop machine, torch.compile failed here, and the
cause was one missing file. /etc/ld.so.cache is now built at every boot by
pengyos-ldconfig.service, not only when a stopgap library is listed. Without that cache
ldconfig -p exited 1, so Python's ctypes.util.find_library returned None for every library, libc
included, and anything that probes that way died: Triton (torch.compile), numba and CuPy all failed
while looking for libcuda. This was found by deploying a Hugging Face model on Desktop NVIDIA at
26.11.11.
The GPU guests are now named where you choose them. The dx card prints pengyos-guest list, which
shows every declared guest whether or not it is created, and marks a guest that shares the NVIDIA
driver with [GPU]. pengyos.txt lists every profile in guests.d, so declaring a guest needs no
code. The PengyOS skill states that NVIDIA's libraries are in /usr/lib, not
/usr/lib/x86_64-linux-gnu. The gap this closes is real: an agent wiring up GPU work hand-wired
libcuda into the dev guest, which has no GPU, because nothing it read named the cuda guest.
A new conformance probe set checks the system conventions software probes for, and records the ones
left out on purpose with a reason. Run in a booted image by scripts/conformance-probes.py against
scripts/conformance-probes.toml, it found three absences, and all three now ship: getconf,
iconv, and /etc/ssl/cert.pem, the CA bundle OpenSSL looks for at that default path. file(1)
ships in every edition.
Nothing here is visible in the UI. What changes is that software built for other Linux systems finds the libraries, the tools and the certificates it asks for by name, and the machine with the GPU is the one it lands on.