Ray-Ascend¶
Welcome to the official documentation for Ray-Ascend — a high-performance extension for Ray that enables efficient tensor transport and collective communication on Huawei Ascend AI accelerators.
Overview¶
Ray-Ascend provides two core capabilities:
- Direct Tensor Transport: Zero-copy, low-latency tensor transfer between Ray actors/workers using Ascend-native mechanisms.
- HCCL-based Collective Communication: Scalable all-reduce, broadcast, and other collective operations via Huawei's HCCL backend.
This documentation is divided into two parts:
- User Guide: For end users who want to install and use Ray-Ascend in their applications.
- Developer Guide: For contributors who want to understand, extend, or debug the codebase.
Note: This project assumes you have a working Ascend environment (e.g., Atlas 900) with CANN installed.