dask-hip

dask-hip extends Dask.distributed with one-worker-per-GPU scheduling for AMD Instinct™ GPUs, ported from the RAPIDS® dask-cuda project. It handles automatic GPU device visibility and CPU affinity, GPU memory spilling, and RMM pool integration, with UCX-based high-performance communication between workers for scaling AMD Data Science workloads across multiple GPUs.

Documentation Installation instructions, how-to guides, and API reference material are on the ROCm Documentation site. Github View the dask-hip source code on Github.