Summary

Fixes #1455. Running cellpose with --gpu_device mps on Apple silicon silently uses the CPU, while omitting the flag correctly uses MPS:

python -m cellpose --dir . --use_gpu --gpu_device mps ...
>>>> using CPU          # <- wrong, MPS is available

python -m cellpose --dir . --use_gpu ...
** TORCH MPS version installed and working. **
>>>> using GPU (MPS)    # <- works

Root cause

__main__ passes device=args.gpu_device (the string "mps") into assign_device, which keeps it as the string "mps" and calls use_gpu(gpu_number="mps"). _use_gpu_torch then treats gpu_number as a device index and builds:

torch.device("cuda:" + str(gpu_number))   # "cuda:mps" -> raises
torch.device("mps:"  + str(gpu_number))   # "mps:mps"  -> raises (invalid index)

Both raise, so _use_gpu_torch returns False and assign_device falls back to CPU. The default path works only because gpu_device defaults to "0", which becomes torch.device("mps:0").

Fix

In _use_gpu_torch, treat a non-integer gpu_number (i.e. the "mps" sentinel) as a request for the default MPS device torch.device("mps"), instead of building "mps:<gpu_number>". Integer indices keep their existing "mps:<n>" behaviour.

Verification

Reproduced on Apple silicon (MPS available):

from cellpose import core
core.assign_device(use_torch=True, gpu=True, device="mps")
# before: (device(type=cpu), False)
# after:  (device(type=mps), True)

Added a regression test test_assign_device_mps_string in tests/test_import.py (skipped when MPS is unavailable). It fails on the current code (assert gpu -> False) and passes with this change. The integer/default device paths are unchanged.