- How We Cold Start a 32B Model in 1.1s. How Lupine's checkpoint/restore engine moves live GPU state in and out of VRAM in about a second.
- What Is GPU Orchestration? What the scheduling layer does, and why pooling beats per-machine allocation.
- GPU Sizing for LLM Inference: How Many Do You Need? Working out real capacity from model size and request load.
- Self-Hosted LLM Cost vs. a Frontier API: The Real Math. What owning the hardware costs against per-token billing.
- What Is a Model in a Box? Where the compute layer sits in the rest of the stack.
On-demand GPUs.
Zero idle cost.
Lupine is Numerata's compute layer: GPUs that scale to zero when idle, aggregate across machines into one pool, and need no code changes to use. Private cloud by default, or fully air-gapped, so market data and models never leave your infrastructure.
PART OF THE NUMERATA STACK · lupine.sh ↗
HOW IT WORKS
A drop-in CUDA shim, not a new API to learn.
DROP-IN CUDA SHIM
Transparent driver shim that works with PyTorch, TensorFlow, and JAX out of the box. Supports cuBLAS, cuDNN, and NVML, so existing code runs unmodified.
MULTI-GPU AGGREGATION
Our coordinator exposes a unified device list across machines, turning scattered GPUs, on-prem or in the cloud, into one pool you can schedule against.
LOW OVERHEAD
CUDA calls are proxied transparently over a plain HTTP/2 protocol, with minimal serialization overhead. Bring your own TLS, auth, or load balancing.
QUICKSTART
Two containers. No SDK, no provisioning step.
SERVER: run where the GPU lives
docker run --rm --gpus all -p 14833:14833 \ ghcr.io/lupinemachines/lupine-server:cuda-13.1.0-ubuntu24.04
CLIENT: connect from any machine (mac, linux, arm, x86)
docker run --rm -it \ -e LUPINE_SERVER=<server>:14833 \ ghcr.io/lupinemachines/lupine-client:cuda-13.1.0-ubuntu24.04 \ nvidia-smi
USE CASES
Spin up GPUs for research sprints,
scale to zero between them.
Run risk and pricing models at scale
without managing servers.
Fine-tune signal models
without reserved capacity.
FURTHER READING
More on GPU scheduling and capacity.
STATUS