Quick Start
Install GPU Shards with one command and deploy your first sharded GPU container in minutes.
Go from a bare Ubuntu host to your first memory-capped GPU container in a few minutes.
Assumes Ubuntu 22.04+ with a working NVIDIA driver (verify with
nvidia-smi).
1. Install
Run the one-line installer. It wires up Docker, the NVIDIA Container Toolkit, the
libvgpu image, and the management panel for you:
curl -fsSL https://gpushards.com/install.sh | bash
Prefer to do each step by hand? Follow the manual installation guide instead.
2. Open the panel
Visit http://localhost:3000. You will see the available GPU instances and a workload
builder.
3. Pick an instance and allocate a shard
Choose how much GPU memory the container may use. The bar shows your card's total memory; the slice you select becomes the container's hard limit.
4. Configure the container
Provide a container image (any CUDA-compatible image works) and, optionally, a private registry login and an exposed port.
Image: pytorch/pytorch:latest
Port: 8080
5. Deploy
Click Deploy Workload. GPU Shards launches the container with the libvgpu
library preloaded and the memory cap applied.
Verify the cap from the CLI
You can confirm the limit works with a stock CUDA image. The reported memory should be your shard size, not the full card:
docker run --rm --gpus all \
-e LD_PRELOAD=/libvgpu/build/libvgpu.so \
-e CUDA_DEVICE_MEMORY_LIMIT=4096m \
hami-core-demo:latest nvidia-smi
Next, learn what is happening under the hood in How GPU Sharing Works.