Is Ray related to hardware resources? If so, how well does it support domestic AI accelerators such as Ascend and Muxi, and does it provide full Ray capabilities for these chips?
Yes, Ray is closely related to hardware resources—it provides native support for scheduling and isolating tasks and actors on various accelerators, including CPUs, GPUs, and other AI chips. Ray allows you to specify resource requirements (like CPUs, GPUs, or custom accelerators) for each task or actor, and manages resource allocation and isolation automatically. This enables efficient use of heterogeneous hardware in distributed ML and Python applications. See Accelerator Support and Ray Accelerator Support REP.
For domestic AI accelerators such as Huawei Ascend (NPU), Ray provides experimental support. You can schedule tasks and actors using the “NPU” resource, and Ray will set the appropriate environment variables (e.g., ASCEND_RT_VISIBLE_DEVICES) for device isolation. However, this support is marked as experimental and not fully tested by the Ray team; some advanced Ray features may not be as robust as on NVIDIA GPUs. There is no official mention of Muxi accelerator support in the provided data sources, so its support status is unclear. See Accelerator Support Table and NPU AcceleratorManager.
Would you like more detail on how to use Ray with Ascend NPUs or on the current limitations?
Sources:
Hint: Mention @RunLLM in the post for followups.