convert gpus to DeviceRequests with implicit "gpu" capability

[ndeloof]

What I did

This adds support for gpus to align with docker CLI
Coucou Guillaume 🥰

Related issue
compose-spec/compose-go#693

(not mandatory) A picture of a cute animal, if possible in relation to what you did

[glours]

LGTM

[polarathene]

Was this meant to have documentation at https://docs.docker.com/reference/compose-file/services ?

I assume gpus: all is valid? (to match --gpus all), not able to test that any time soon.

The CLI docs also specifically refer to --gpus as only for nvidia, I'm not sure if that's outdated or a limitation (I'm aware of AMD just relying on devices field, rather than deploy.resources.reservations.devices, but maybe that also works with --gpus all these days 🤷‍♂️).

[ndeloof]

gpus: all is not allowed as model uses a deviceRequest struct, but we could introduce such a shortcut notation.
AFAIK docker engine only supports nvidia GPUs, but from compose point of view this is vendor-neutral: we just forward device capabilities request to engine, which has to find adequate device to match.

[polarathene]

we could introduce such a shortcut notation

I assumed that was one of the main conveniences in supporting this alternative 😅 (to have parity with the docker CLI --gpus all example that is commonly documented)

---

AFAIK docker engine only supports nvidia GPUs, but from compose point of view this is vendor-neutral: we just forward device capabilities request to engine, which has to find adequate device to match.

From what I've seen with AMD (and presumably Intel), those two are stuck with the more verbose deploy config, however CDI with --device / device: appears to be the preferred way going forward once the vendors better support it (I've seen third-party tools to generate the CDI configs needed on the host).

Presently it might only be nvidia with proper support (I've not tried the CDI support yet, but have read there might be some extra changes required, and Docker Desktop / WSL2 may also affect compatibility). So at the same time I'm not sure how beneficial gpus: all is to users of Compose (Podman and k8s both adopted CDI).

I imagine for consistency with docker CLI that gpus: all should be supported, but docs should likewise direct users towards CDI (I have no clue how long that will remain behind a daemon feature flag to use though, which adds some friction).