[udsy19]

What does this PR do?

write_basic_config (used by accelerate config default) has no branch for MPS, so on Apple Silicon it falls through to the CPU-only branch and writes use_cpu: true. Any script launched with that config then runs on the CPU even though MPS is available. Launching without a config, or using the interactive accelerate config, correctly uses MPS.

Repro on an M-series Mac:

python -c "from accelerate.utils import write_basic_config; write_basic_config(save_location='cfg.yaml')"
accelerate launch --config_file cfg.yaml script.py   # Accelerator().device -> cpu (expected mps)

This adds an MPS branch that writes a single-process, non-CPU config, plus test_write_basic_config_mps in tests/test_cli.py, which mocks MPS availability so it also runs on CPU CI. The test fails on main (use_cpu is True) and passes with the fix; tests/test_cli.py passes (44 passed, 1 skipped).

Before submitting

• Did you read the contributor guideline?
• Did you write any new necessary tests?

Who can review?

@SunMarc

[mlzerotoheroyoutube]

I tested this on an Apple M5 Pro (macOS 26.6.2, Python 3.14, torch 2.15.0), PR head 4c8407a against main at f47439c.

	main	this PR
write_basic_config() writes	use_cpu: true	use_cpu: false
accelerate launch --config_file <that file> → Accelerator().device	cpu	mps
same with write_basic_config(mixed_precision="fp16")	cpu	mps
accelerate config default, then accelerate launch with no --config_file	not run	mps
pytest tests/test_cli.py	1 failed (test_write_basic_config_mps), 43 passed, 1 skipped	44 passed, 1 skipped

test_accelerate_test also runs on MPS with the new default config (Device: mps, all training checks pass). So the fix does what it says on real Apple Silicon, not only with the mocked MPS in the test. LGTM.

(Testing was done with help from Claude Code; I checked the results.)