Enhance/gpu support

[drcege]

GPU Model Loading

• Models are now loaded onto GPUs where possible to maximize efficiency.
  • Currently, this feature is exclusive to Hugging Face models.
  • Adherence to the CUDA_VISIBLE_DEVICES setting ensures only designated GPUs are utilized.
  • The num_proc parameter is intelligently adjusted for optimal performance with mixed CPU/GPU operations.

Multiprocessing Strategy

• Will automatically determine the most effective multiprocessing start method.
  • On GPU setups, forkserver is the default choice, switching to spawn if necessary.
  • For CPU-bound execution, fork provides the best compatibility and performance.
• Users retain control over multiprocessing behavior via the MP_START_METHOD environment variable, with permissible values including auto, fork, forkserver, and spawn.

New Operator Development

• When developing GPU-accelerated operators:
  • Declare GPU usage with self._accelerator = 'cuda'.
  • Both compute_stats and process methods should now include an optional rank parameter (rank=None) to support multi-GPU setups.

[HYLcool]

New Operator Development

• When developing GPU-accelerated operators:

  • Declare GPU usage with self._accelerator = 'cuda'.
  • Both compute_stats and process methods should now include an optional rank parameter (rank=None) to support multi-GPU setups.

How about adding these instructions to the DeveloperGuide docs as well?

[drcege]

Added GPU development guide.

[HYLcool]

LGTM 🎉