RFD 1056 details: the abandoned workarounds, what the container carries, volumes, what it does not do

The abandoned workarounds

This branch first swapped EXLA for Torchx, and made the dependency conditional on the platform. Neither workaround is the answer. RFD 1019 selects EXLA because it compiles Nx.Defn graphs, while Torchx runs them one operation at a time.

What the container carries

Part	Why
Elixir 1.17.3	On Erlang 27, on Debian bookworm.
cmake and g++	EXLA compiles a NIF.
python3	The model image check runs it.
Docker in Docker	The model images build here.
CockroachDB	mix weftspun.crdb install fetches it.

Debian, and not Alpine. The XLA archive links against shared libraries that a musl base does not carry.

Two volumes, and not bind mounts

_build and deps each take a named volume. A bind mount from a Windows host makes both slow, because the BEAM writes many small files into them.

The source stays a bind mount. An edit on the host must reach the container without a copy.

What this does not do

It does not give the container a GPU. EXLA takes its host client here, and XLA_TARGET=cuda12 needs the NVIDIA runtime on the host.

RFD 1027 records that every model reaches a 24 GB card. That card is rented, and it is not this machine.