跳到主要内容

Docker + WASM + GPU

This is a completely new approach, adopting Docker + Crun with Wasmedge + CDI to enable the usage of host GPU devices. The reason for not continuing with the use of runwasi as the wasm runtime within Docker from the previous chapter is due to considerations of the current stage of support for CDI and the compatibility approach.

Prerequisite​

Before we start, you need

  • GPU device (Here we will take NVIDIA graphics cards as our example and we have only conducted tests on NVIDIA GPUs on linux for now)
    • Install NVIDIA GPU Driver
    • Install either the NVIDIA Container Toolkit or you installed the nvidia-container-toolkit-base package.
  • Docker version > 4.29 (which includes Moby 25)

Regarding the installation of the NVIDIA driver and toolkit, we won't go into detail here, but we could provide a few reference documents and the ways to verify your environment is ok.

Nvidia drivers installation on ubuntu, Toolkit install guide, Nvidia CDI supoort reference

# check your driver and device
> nvidia-smi -L

# Check your toolkit
> nvidia-ctk --version

Install latest docker-ce

> curl -fsSL https://get.docker.com -o get-docker.sh
> sh get-docker.sh

# Check your docker
> docker --version

CDI setup​

Generate the CDI specification file

> sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml

# Check you cdi config is good
> nvidia-ctk cdi list

# Example output
INFO[0000] Found 2 CDI devices
nvidia.com/gpu=0
nvidia.com/gpu=all

Enable CDI in docker config (/etc/docker/daemon.json)

{
  "features": {
     "cdi": true
  },
  "cdi-spec-dirs": ["/etc/cdi/", "/var/run/cdi"]
}
# Reload docker daemon
> sudo systemctl reload docker

# Test your cdi setup good
> docker run --rm --device nvidia.com/gpu=all ubuntu:22.04 nvidia-smi -L

# Example output 
GPU 0: NVIDIA GeForce GTX 1080 (UUID: GPU-********-****-****-****-************)

Setup your container runtime (crun + wasmedge + plugin system)​

Build crun with wasmedge enable

> sudo apt install -y make git gcc build-essential pkgconf libtool libsystemd-dev libprotobuf-c-dev libcap-dev libseccomp-dev libyajl-dev go-md2man libtool autoconf python3 automake

> git clone https://github.com/containers/crun
> cd crun
> ./autogen.sh
> ./configure --with-wasmedge
> make

# Check your crun
> ./crun --version

Replace container run time

{
  "runtimes": {
    "crun": {
      "path": "<The crun binary path is built by yourself>"
    }
  },
  "features": {
    "cdi": true
  },
  "cdi-spec-dirs": ["/etc/cdi/", "/var/run/cdi"]
}
# Reload docker daemon
> sudo systemctl reload docker

Download ggml plugin into host

> curl -sSf https://raw.githubusercontent.com/WasmEdge/WasmEdge/master/utils/install.sh | bash -s -- --plugins wasi_nn-ggml

# Make sure all your plugin dependencies is good
> ldd ~/.wasmedge/plugin/libwasmedgePluginWasiNN.so

Demo llama with our wasm application​

The demo image is built the Wasm application from here, and upload it to here.

Download inference model

> curl -LO https://huggingface.co/second-state/Llama-2-7B-Chat-GGUF/resolve/main/llama-2-7b-chat.Q5_K_M.gguf

Docker run llama2 inference

docker run -v ~/.wasmedge/plugin/libwasmedgePluginWasiNN.so:/.wasmedge/plugin/libwasmedgePluginWasiNN.so \
  -v /usr/local/cuda/targets/x86_64-linux/lib/libcudart.so.12:/lib/x86_64-linux-gnu/libcudart.so.12 \
  -v /usr/local/cuda/targets/x86_64-linux/lib/libcublas.so.12:/lib/x86_64-linux-gnu/libcublas.so.12 \
  -v /usr/local/cuda/targets/x86_64-linux/lib/libcublasLt.so.12:/lib/x86_64-linux-gnu/libcublasLt.so.12 \
  -v /lib/x86_64-linux-gnu/libcuda.so.1:/lib/x86_64-linux-gnu/libcuda.so.1 \
  -v .:/resource \
  --env WASMEDGE_PLUGIN_PATH=/.wasmedge/plugin \
  --env WASMEDGE_WASINN_PRELOAD=default:GGML:AUTO:/resource/llama-2-7b-chat.Q5_K_M.gguf \
  --env n_gpu_layers=100 \
  --rm --device nvidia.com/gpu=all --runtime=crun --annotation=module.wasm.image/variant=compat-smart --platform wasip1/wasm \
  ghcr.io/captainvincent/runwasi-demo:wasmedge-ggml-llama default \
  $'[INST] <<SYS>>\nYou are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe.  Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you do not know the answer to a question, please do not share false information.\n<</SYS>>\nWhat is the capital of Japan?[/INST]'

Example Result

ggml_init_cublas: GGML_CUDA_FORCE_MMQ:   no
ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes
ggml_init_cublas: found 1 CUDA devices:
  Device 0: NVIDIA GeForce GTX 1080, compute capability 6.1, VMM: yes
Prompt:
[INST] <<SYS>>
You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe.  Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you do not know the answer to a question, please do not share false information.
<</SYS>>
What is the capital of Japan?[/INST]
Response:
[INFO] llama_commit: "4ffcdce2"
[INFO] llama_build_number: 2334
[INFO] Number of input tokens: 140
Thank you for asking! The capital of Japan is Tokyo. I'm glad you asked! It's important to be informed and curious about different countries and their capitals. Is there anything else I can help you with?
[INFO] Number of input tokens: 140
[INFO] Number of output tokens: 48