feat: auto-detect Apple Silicon (MPS) and keep Triton CUDA-only

[Berkkirik]

Fixes #21

Builds on #17 (which added --device auto with cuda > cpu fallback). Extends the auto chain to include Apple Silicon (MPS) so Mac users get GPU acceleration by default instead of falling back to CPU.

Problem

After #17, an Apple Silicon Mac with no CUDA device falls back to CPU. That's correct but leaves performance on the table — MPS is available on M1/M2/M3 Macs. Naively adding MPS to the auto chain is unsafe, though: the Triton-backed MoE kernels are CUDA-only (Triton does not target Metal), and the current code auto-enables Triton on any non-CPU device. Picking mps as default would silently crash once the MoE layer is hit.

Fix

Two coordinated changes:

1. opf/_common/device.py — auto now picks cuda > mps > cpu. Each fallback emits an info line on stderr so the user always knows which backend was selected.
2. opf/_model/model.py — narrow the Triton auto-enable to device.type == "cuda". MPS and CPU both use the torch-ops path unless the user explicitly sets OPF_MOE_TRITON=1.
3. opf/_train/runner.py — mirror the same CUDA-only gate when the training runner sets OPF_MOE_TRITON=1 on behalf of the user (previously set for any non-CPU device, which would silently enable Triton on mps).
4. opf/_cli/common.py — expand --device help text to list the full backend order.

Verified on this machine

macOS 26.4 on Apple Silicon (M-series), Python 3.14, torch 2.11.0:

$ python3 -c "from opf._common.device import resolve_device; print(resolve_device('auto'))"
info: no CUDA device detected; using Apple Metal (MPS).
mps

Resolver cases:

Case	Result
resolve_device("auto") on Apple Silicon	mps + stderr info line ✅
resolve_device("mps")	mps ✅
resolve_device("cpu")	cpu ✅
resolve_device("cuda") on non-CUDA machine	Returns cuda object, fails at tensor alloc — same as today when user asks for cuda explicitly ✅

Low-level MPS op smoke test (ops the inference path depends on), all pass on MPS:

• torch.nn.Embedding forward (the op that crashed in on MacBook: Torch not compiled with CUDA enabled #21/Odd defaults: fails if no NVIDIA driver found on a system with no nvidia device at all #12)
• Attention-like matmul + softmax
• log_softmax (inference logprob path)
• topk (MoE expert routing)
• argsort (MoE packing)
• bincount (MoE expert counts — historically flaky on MPS, works here)

End-to-end inference on MPS was not run here (no checkpoint locally). A maintainer on Apple Silicon can verify with:

opf "Alice was born on 1990-01-02."
# should print an "info: ... using Apple Metal (MPS)" line and redact correctly

Backwards compatibility

• --device cuda / --device cpu / --device mps: unchanged (all pass through resolve_device as-is).
• On CUDA machines: auto still picks cuda first — no behavior change.
• Triton path: was auto-enabled on any non-CPU device, now auto-enabled on CUDA only. Users who were running training on CUDA get the same behavior; CPU users get the same behavior; only an MPS user (who would have crashed before) now gets the torch-ops path.
• No public Python API change.

Depends on

#17 — This PR builds directly on the resolve_device helper introduced there. If #17 is merged first, this rebases to a no-conflict diff; if the maintainer prefers to squash both into one, I'm happy to close this and post a combined patch.

[clement-heliot]

FTR, I currently do the following:
OPF_MOE_TRITON=0 opf --device mps "Alice was born on 1990-01-02."

[Berkkirik]

FTR, I currently do the following: OPF_MOE_TRITON=0 opf --device mps "Alice was born on 1990-01-02."

Thanks for the data point — that's exactly the workflow this PR automates. After merge, OPF_MOE_TRITON=0 and --device mps both become implicit on Apple Silicon; plain opf "..." should give you the same result.

[Berkkirik]

FTR, I currently do the following: OPF_MOE_TRITON=0 opf --device mps "Alice was born on 1990-01-02."

Thanks for the data point — that's exactly the workflow this PR automates. After merge, OPF_MOE_TRITON=0 and --device mps both become implicit on Apple Silicon; plain opf "..." should give you the same result.

I couldn't see your messages, can you write it again ?

[losoy88]

FTR, I currently do the following: OPF_MOE_TRITON=0 opf --device mps "Alice was born on 1990-01-02."

Thanks for the data point — that's exactly the workflow this PR automates. After merge, OPF_MOE_TRITON=0 and --device mps both become implicit on Apple Silicon; plain opf "..." should give you the same result.

I couldn't see your messages, can you write it again ?