Two things
orbistoun-llm left undone
The crate exists (D212) and nothing calls it yet. Two gaps in it are worth recording rather than rediscovering.
Only NVIDIA accelerators report their memory.
host.rs shells out to nvidia-smi; AMD and
Intel report nothing, so those machines fall to the catalogue default
and get a smaller model than they could run. The better answer is
already in this repository - orbistoun-gpu-vulkan loads
Vulkan at runtime and can enumerate device-local heaps on any vendor -
and was left out to avoid coupling an isolated crate to ash
for one number. It becomes nearly free the moment anything else wants a
host description.
host.rs should not live there. It has
no knowledge of models and no dependency on the rest of the crate,
precisely so it can be lifted whole. The reason to lift it is not
tidiness: D046 requires a run report to embed its own inputs so that a
difference between two runs is attributable to the change rather than to
drift, and the machine is such an input. Two runs compared
across two machines are not comparable today and nothing says so.
Blocked on deciding where it goes - orbistoun-core is
IO-free by contract and probing is IO, so it is either a new crate low
in the spine or a module of orbistoun-report.