Skip to content

Using Docker

Spyre base images

Base images containing the driver stack for IBM Spyre accelerators are available from the ibm-aiu organization on Quay. This includes the torch_sendnn package, which is required for using torch with Spyre cards.

Attention

These images contain an install of the torch package. The specific version installed is guaranteed to be compatible with torch_sendnn. Overwriting this install with a different version of torch may cause issues.

Using community built images

Community maintained images are also available on Quay, the latest x86 build is quay.io/ibm-aiu/vllm-spyre:latest.amd64.

Caution

These images are provided as a reference and come with no support guarantees.

Building vLLM Spyre's Docker Image from Source

You can build and run vLLM Spyre from source via the provided docker/Dockerfile.amd64. To build vLLM Spyre:

DOCKER_BUILDKIT=1 docker build . --target release --tag vllm/vllm-spyre --file docker/Dockerfile.amd64

Note

This Dockerfile currently only supports the x86 platform

Running vLLM Spyre in a Docker Container

To run your vLLM Spyre image on a host with Spyre cards installed:

$ docker run \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    -v /dev/vfio:/dev/vfio \
    -p 8000:8000 \
    --env "HUGGING_FACE_HUB_TOKEN=<secret>" \
    vllm/vllm-spyre <model> <args...>

To run your vLLM Spyre image on a host without Spyre cards installed:

$ docker run \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    -p 8000:8000 \
    --env "HUGGING_FACE_HUB_TOKEN=<secret>" \
    --env "VLLM_SPYRE_DYNAMO_BACKEND=eager" \
    vllm/vllm-spyre <model> <args...>