Installation with ROCm
Aphrodite supports AMD GPUs using ROCm 6.1.
Requirements
- Linux (or WSL on Windows)
- Python 3.8 - 3.11
- GPU: MI200 (gfx90a), MI300 (gfx942), RX 7900 Series (gfx1100)
- ROCm 6.1
Installation with Docker
You can build Aphrodite Engine from source. First, build a docker image from the provided Dockerfile.rocm
, then launch a container from the image.
To build Aphrodite on high-end datacenter GPUs (e.g. MI300X), run this:
DOCKER_BUILDKIT=1 docker build -f Dockerfile.rocm -t aphrodite-rocm .
To build Aphrodite on NAVI GPUs (e.g. RTX 7900 XTX), run this:
DOCKER_BUILDKIT=1 docker build --build-arg BUILD_FA="0" -f Dockerfile.rocm -t aphrodite-rocm .
Then run your image:
docker run -it \ --network=host \ --group-add=video \ --ipc=host \ --cap-add=SYS_PTRACE \ --security-opt seccomp=unconfined \ --device /dev/kfd \ --device /dev/dri \ -v ~/.cache/huggingface/root/.cache/huggingface \ aphrodite-rocm \ bash
Installation from source
You can also build Aphrodite from source, but it’s more complicated, so we recommend Docker.
You will need the following installed beforehand:
Then install Triton for ROCm. You may also Install CK Flash Attention if needed.
You may need to downgrade ninja
version to 1.10.
Finally, build Aphrodite:
git clone https://github.com/PygmalionAI/aphrodite-engine.gitcd aphrodite-engine
chmod +x ./amdpatch.sh./amdpatch.shpip install -U -r requirements-rocm.txtpython setup.py develop # pip install -e . won't work for now