The vLLM Spyre plugin (vllm-spyre
) is a dedicated backend extension that enables seamless integration of IBM Spyre Accelerator with vLLM. It follows the architecture describes in vLLM's Plugin System, making it easy to integrate IBM's advanced AI acceleration into existing vLLM workflows.
First, download vllm-spyre
git clone https://github.com/vllm-project/vllm-spyre.git
cd vllm-spyre
Build image from source
docker build . -f Dockerfile.spyre -t vllm-spyre
docker run -it --rm vllm-spyre bash
We use the uv package manager to manage the
installation of the plugin and its dependencies. uv
provides advanced
dependency resolution which is required to properly install dependencies like
vllm
without overwriting critical dependencies like torch
.
# Install uv
pip install uv
# Install vllm-spyre
git clone https://github.com/vllm-project/vllm-spyre.git
cd vllm-spyre
VLLM_TARGET_DEVICE=empty uv pip install -e .