accelerate

a68e0d50 - Fix wait_for_everyone() crashing on MULTI_CPU when the machine has an accelerator (#4354)

Fix wait_for_everyone() crashing on MULTI_CPU when the machine has an accelerator (#4354) * Fix wait_for_everyone() crashing on MULTI_CPU when the machine has an accelerator `PartialState.wait_for_everyone()` passes `device_ids=[local_process_index]` to `torch.distributed.barrier()` for every distributed type, including MULTI_CPU. When `device_ids` is given, torch runs the barrier on the machine's accelerator (`torch._C._get_accelerator()`) instead of on the gloo/CPU process group. On a machine with a non-CUDA accelerator (e.g. Apple Silicon / MPS) this raises `NotImplementedError: The distributed operator 'c10d::barrier' is not currently implemented for the MPS device`, so any multi-process CPU run that calls `wait_for_everyone()` / `main_process_first()` crashes. Only pass `device_ids` when the distributed type is not MULTI_CPU; on CPU let torch pick the device of the process group, as it did before #3744. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * Use elif for MULTI_CPU barrier and drop mocked tests per review Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>