vLLM GB10 / DGX Spark

Custom vLLM build for NVIDIA GB10 (DGX Spark) that reuses NVIDIA’s Spark‑ready vLLM base image and builds vLLM from source to unlock optimized kernels on SM121 GPUs.

Why this exists

• DGX Spark needs CUDA 13.1+ toolchain and Spark‑specific LLVM/Triton patches.
• Stock vLLM images don’t fully support GB10 kernel compilation or the latest model adapters.

Quick start (Docker)

Pull (once published):

docker pull th0rgal/vllm-gb10-dgx-spark:latest

Run:

docker run -d --name step3-vl \
  --gpus all --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 \
  -p 8000:8000 -v /root/.cache/huggingface:/root/.cache/huggingface \
  th0rgal/vllm-gb10-dgx-spark:latest \
  vllm serve stepfun-ai/Step3-VL-10B \
    -tp 1 \
    --reasoning-parser deepseek_r1 \
    --enable-auto-tool-choice \
    --tool-call-parser hermes \
    --trust-remote-code

Build locally

docker build -t vllm-gb10:dev .

Pin a vLLM commit:

docker build --build-arg VLLM_COMMIT=<commit-or-tag> -t vllm-gb10:dev .

Notes

• Base image: nvcr.io/nvidia/vllm:25.12.post1-py3 (Spark‑ready CUDA/LLVM/Triton).
• Dockerfile installs vLLM deps without replacing torch/torchaudio/torchvision.