FlagCX is now fully integrated into Paddle as an optional high-performance communication backend. This integration enables efficient distributed training on multiple hardware platforms, including support for heterogeneous training on Nvidia and Iluvatar GPUs.
Use the guides below to quickly get started with training models using Paddle + FlagCX.
Train on a single type of hardware platform:
| Hardware | User Guide |
|---|---|
| Nvidia GPU | Get Started |
| Kunlunxin XPU | Get Started |
| Iluvatar GPU | Get Started |
Train across different hardware platforms simultaneously:
| Hardware Combination | User Guide |
|---|---|
| Nvidia GPU + Iluvatar GPU | Get Started |