Docker Desktop for Windows における GPU サポート

もくじ
メモ

現時点において Docker Desktop における GPU サポートは、WSL2 バックエンドを使った Windows 上においてのみ利用可能です。

Docker Desktop for Windows supports NVIDIA GPU Paravirtualization (GPU-PV) on NVIDIA GPUs, allowing containers to access GPU resources for compute-intensive workloads like AI, machine learning, or video processing.

Prerequisites

To enable WSL 2 GPU Paravirtualization, you need:

  • A Windows machine with an NVIDIA GPU
  • Up to date Windows 10 or Windows 11 installation
  • Up to date drivers from NVIDIA supporting WSL 2 GPU Paravirtualization
  • The latest version of the WSL 2 Linux kernel. Use wsl --update on the command line
  • To make sure the WSL 2 backend is turned on in Docker Desktop

Validate GPU support

To confirm GPU access is working inside Docker, run the following:

$ docker run --rm -it --gpus=all nvcr.io/nvidia/k8s/cuda-sample:nbody nbody -gpu -benchmark

This runs an n-body simulation benchmark on the GPU. The output will be similar to:

Run "nbody -benchmark [-numbodies=<numBodies>]" to measure performance.
        -fullscreen       (run n-body simulation in fullscreen mode)
        -fp64             (use double precision floating point values for simulation)
        -hostmem          (stores simulation data in host memory)
        -benchmark        (run benchmark to measure performance)
        -numbodies=<N>    (number of bodies (>= 1) to run in simulation)
        -device=<d>       (where d=0,1,2.... for the CUDA device to use)
        -numdevices=<i>   (where i=(number of CUDA devices > 0) to use for simulation)
        -compare          (compares simulation results running once on the default GPU and once on the CPU)
        -cpu              (run n-body simulation on the CPU)
        -tipsy=<file.bin> (load a tipsy model file for simulation)

> NOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.

> Windowed mode
> Simulation data stored in video memory
> Single precision floating point simulation
> 1 Devices used for simulation
MapSMtoCores for SM 7.5 is undefined.  Default to use 64 Cores/SM
GPU Device 0: "GeForce RTX 2060 with Max-Q Design" with compute capability 7.5

> Compute 7.5 CUDA device: [GeForce RTX 2060 with Max-Q Design]
30720 bodies, total time for 10 iterations: 69.280 ms
= 136.219 billion interactions per second
= 2724.379 single-precision GFLOP/s at 20 flops per interaction

Run a real-world model: SmolLM2 with Docker Model Runner

メモ

Docker Model Runner with vLLM for Windows with WSL2 is available starting with Docker Desktop 4.54.

Use Docker Model Runner to run the SmolLM2 LLM with vLLM and GPU acceleration:

$ docker model install-runner --backend vllm --gpu cuda

Check it's correctly installed:

$ docker status
Docker Model Runner is running

Status:
llama.cpp: running llama.cpp version: c22473b
vllm: running vllm version: 0.11.0

Run the model:

$ docker model un ai/smollm2-vllm hi
Hello! I'm sure everything goes smoothly here. How can I assist you today?