Skip to content

Latest commit

 

History

History
220 lines (189 loc) · 8.83 KB

README_DOCKER.md

File metadata and controls

220 lines (189 loc) · 8.83 KB
Raw

Run or Build h2oGPT Docker

Setup Docker for CPU Inference

No special docker instructions are required, just follow these instructions to get docker setup at all. Add your user as part of docker group, exit shell, login back in, and run:

newgrp docker

which avoids having to reboot. Or just reboot to have docker access.

Setup Docker for GPU Inference

Ensure docker installed and ready (requires sudo), can skip if system is already capable of running nvidia containers. Example here is for Ubuntu, see NVIDIA Containers for more examples.

distribution=$(. /etc/os-release;echo $ID$VERSION_ID) \
    && curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg \
    && curl -s -L https://nvidia.github.io/libnvidia-container/$distribution/libnvidia-container.list | \
        sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | \
        sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
sudo apt-get update && sudo apt-get install -y nvidia-container-toolkit-base
sudo apt install nvidia-container-runtime
sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker

If running on A100's, might require Installing Fabric Manager and Installing GPU Manager.

Run h2oGPT using Docker

All available public h2oGPT docker images can be found in Google Container Registry.

Ensure image is up-to-date by running:

docker pull gcr.io/vorvan/h2oai/h2ogpt-runtime:0.1.0

An example running h2oGPT via docker using LLaMa2 7B model is:

docker run \
       --gpus all \
       --runtime=nvidia \
       --shm-size=2g \
       -p 7860:7860 \
       --rm --init \
       -v "${HOME}"/.cache:/workspace/.cache \
       -v "${HOME}"/save:/workspace/save \
       gcr.io/vorvan/h2oai/h2ogpt-runtime:0.1.0 /workspace/generate.py \
          --base_model=h2oai/h2ogpt-4096-llama2-7b-chat \
          --use_safetensors=True \
          --prompt_type=llama2 \
          --save_dir='/workspace/save/' \
          --score_model=None \
          --max_max_new_tokens=2048 \
          --max_new_tokens=1024

then go to http://localhost:7860/ or http://127.0.0.1:7860/.

An example of running h2oGPT via docker using AutoGPTQ (4-bit, so using less GPU memory) with LLaMa2 7B model is:

docker run \
       --gpus all \
       --runtime=nvidia \
       --shm-size=2g \
       -p 7860:7860 \
       --rm --init \
       -v "${HOME}"/.cache:/workspace/.cache \
       -v "${HOME}"/save:/workspace/save \
       gcr.io/vorvan/h2oai/h2ogpt-runtime:0.1.0 /workspace/generate.py \
          --base_model=TheBloke/Llama-2-7b-Chat-GPTQ \
          --load_gptq="gptq_model-4bit-128g" \
          --use_safetensors=True \
          --prompt_type=llama2 \
          --save_dir='/workspace/save/' \
          --score_model=None \
          --max_max_new_tokens=2048 \
          --max_new_tokens=1024

then go to http://localhost:7860/ or http://127.0.0.1:7860/.

If one needs to use a Hugging Face token to access certain Hugging Face models like Meta version of LLaMa2, can run like:

export HUGGING_FACE_HUB_TOKEN=<hf_...>
docker run \
       --gpus all \
       --runtime=nvidia \
       --shm-size=2g \
       -p 7860:7860 \
       --rm --init \
       -v "${HOME}"/.cache:/workspace/.cache \
       -v "${HOME}"/save:/workspace/save \
       -e HUGGING_FACE_HUB_TOKEN=$HUGGING_FACE_HUB_TOKEN \
       gcr.io/vorvan/h2oai/h2ogpt-runtime:0.1.0 /workspace/generate.py \
          --base_model=meta-llama/Llama-2-7b-chat-hf \
          --prompt_type=llama2 \
          --save_dir='/workspace/save/' \
          --score_model=None \
          --max_max_new_tokens=2048 \
          --max_new_tokens=1024 \
          --use_auth_token=$HUGGING_FACE_HUB_TOKEN

for some token <hf_...>. See Hugging Face User Tokens for more details.

For GGML/GPT4All models, one should either download the file and map that path outsider docker to a pain told to h2oGPT for inside docker, or pass a URL that would download the model internally to docker.

See README_GPU for more details about what to run.

Run h2oGPT and TGI using Docker

One can run an inference server in one docker and h2oGPT in another docker.

For the TGI server run (e.g. to run on GPU 0)

export MODEL=meta-llama/Llama-2-7b-chat-hf
export HUGGING_FACE_HUB_TOKEN=<hf_...>
export CUDA_VISIBLE_DEVICES=0
docker run -d --gpus all \
       --shm-size 1g \
       -e CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES \
       -e HUGGING_FACE_HUB_TOKEN=$HUGGING_FACE_HUB_TOKEN \
       -e TRANSFORMERS_CACHE="/.cache/" \
       -p 6112:80 \
       -v $HOME/.cache:/.cache/ \
       -v $HOME/.cache/huggingface/hub/:/data ghcr.io/huggingface/text-generation-inference:0.9.3 \
       --model-id $MODEL \
       --max-input-length 4096 \
       --max-total-tokens 8192 \
       --max-stop-sequences 6 &>> logs.infserver.txt

Each docker can run on any system where network can reach or on same system on different GPUs. E.g. replace --gpus all with --gpus '"device=0,3"' to run on GPUs 0 and 3, and note the extra quotes, and then unset CUDA_VISIBLE_DEVICES and avoid passing that into the docker image. This multi-device format is required to avoid TGI server getting confused about which GPUs are available.

One a low-memory GPU system can add other options to limit batching, e.g.:

export MODEL=meta-llama/Llama-2-7b-chat-hf
export HUGGING_FACE_HUB_TOKEN=<hf_...>
unset CUDA_VISIBLE_DEVICES
docker run -d --gpus '"device=0"' \
        --shm-size 1g \
        -e HUGGING_FACE_HUB_TOKEN=$HUGGING_FACE_HUB_TOKEN \
        -e TRANSFORMERS_CACHE="/.cache/" \
        -p 6112:80 \
        -v $HOME/.cache:/.cache/ \
        -v $HOME/.cache/huggingface/hub/:/data ghcr.io/huggingface/text-generation-inference:0.9.3 \
        --model-id $MODEL \
        --max-input-length 1024 \
        --max-total-tokens 2048 \
        --max-batch-prefill-tokens 2048 \
        --max-batch-total-tokens 2048 \
        --max-stop-sequences 6 &>> logs.infserver.txt

then wait till it comes up (e.g. check docker logs for detatched container hash in logs.infserver.txt), about 30 seconds for 7B LLaMa2 on 1 GPU. Then for h2oGPT, just run one of the commands like the above, but add e.g. --inference_server=192.168.0.1:6112 to the docker command line. E.g. using same export's as above, run:

export GRADIO_SERVER_PORT=7860
export CUDA_VISIBLE_DEVICES=0
docker run -d \
       --gpus all \
       --runtime=nvidia \
       --shm-size=2g \
       -p $GRADIO_SERVER_PORT:7860 \
       --rm --init \
       --network host \
       -v "${HOME}"/.cache:/workspace/.cache \
       -v "${HOME}"/save:/workspace/save \
       -e HUGGING_FACE_HUB_TOKEN=$HUGGING_FACE_HUB_TOKEN \
       gcr.io/vorvan/h2oai/h2ogpt-runtime:0.1.0 /workspace/generate.py \
          --base_model=$MODEL \
          --inference_server=http://localhost:6112 \
          --prompt_type=llama2 \
          --save_dir='/workspace/save/' \
          --score_model=None \
          --max_max_new_tokens=4096 \
          --max_new_tokens=1024 \
          --use_auth_token="$HUGGING_FACE_HUB_TOKEN"

or change max_max_new_tokens to 2048 for low-memory case.

For maximal summarization performance when connecting to TGI server, auto-detection of file changes in --user_path every query, and maximum document filling of context, add these options:

          --num_async=10 \
          --top_k_docs=-1
          --detect_user_path_changes_every_query=True

When one is done with the docker instance, run docker ps and find the container ID's hash, then run docker stop <hash>.

Follow README_InferenceServers.md for similar (and more) examples of how to launch TGI server using docker.

Build Docker

# build auto-gptq
make docker_build_deps
# build image
touch build_info.txt
docker build -t h2ogpt .

then to run this version of the docker image, just replace gcr.io/vorvan/h2oai/h2ogpt-runtime:0.1.0 with h2ogpt:latest in above run command.

Docker Compose Setup & Inference

  1. (optional) Change desired model and weights under environment in the docker-compose.yml

  2. Build and run the container

    docker-compose up -d --build

  3. Open https://localhost:7860 in the browser

  4. See logs:

    docker-compose logs -f

  5. Clean everything up:

    docker-compose down --volumes --rmi all