vLLM is an inference engine designed to run large language models efficiently. The key idea is maximizing throughput and minimizing memory waste when serving LLMs.
It uses a memory-efficient attention algoritm called PagedAttention to handle long sequences without running out of GPU memory.
New requests can be added to a batch already in process through continuous batching to keep GPUs fully utilized.
It has an OpenAI-compatible API so applications built for the OpenAI API can switch to a vLLM backend with little or no modification.
What you'll accomplish
You'll set up vLLM high-throughput LLM serving on DGX Spark with Blackwell architecture,
either using a pre-built Docker container or building from source with custom LLVM/Triton
support for ARM64.
What to know before starting
Experience building and configuring containers with Docker
Familiarity with CUDA toolkit installation and version management
Understanding of Python virtual environments and package management
Knowledge of building software from source using CMake and Ninja
Experience with Git version control and patch management
Prerequisites
DGX Spark device with ARM64 processor and Blackwell GPU architecture
CUDA 13.0 toolkit installed: nvcc --version shows CUDA toolkit version.
Docker installed and configured: docker --version succeeds
The Phi-4-multimodal-instruct models require --trust-remote-code when launching vLLM.
NOTE
You can use the NVFP4 Quantization documentation to generate your own NVFP4-quantized checkpoints for your favorite models. This enables you to take advantage of the performance and memory benefits of NVFP4 quantization even for models not already published by NVIDIA.
Reminder: not all model architectures are supported for NVFP4 quantization.