# vllm-gb10

A reproducible vLLM Docker image built specifically for the NVIDIA DGX Spark's GB10 chip, with every component in the stack pinned by commit or digest so the same version file always rebuilds the same image.

vllm-gb10 packages vLLM for the NVIDIA DGX Spark (GB10 SoC, sm_121a) as an arm64 Docker image, targeting that hardware only; it will not run on x86 machines or other GPU architectures. Every component of the stack, including the CUDA base image, PyTorch, NCCL and FlashInfer, is pinned by commit SHA or digest in a `versions.env` file and rebuilt from source where noted, with matching lockfiles under `locks/`.

A Rust frontend (`vllm-rs`) built on axum and PyO3 sits alongside the Python stack, and companion CUDA packages such as NVSHMEM, TVM-FFI and TileLang are pinned to exact versions.

Each release publishes four tags: a canonical immutable tag tied to the vLLM version and stack revision, a variant that also names the CUDA and PyTorch versions, a mutable `latest` tag, and a tag tied to the exact commit that produced the build; there is no bare, mutable version tag.

Website: https://technotim.com/posts/vllm-gb10-docker/