docs/README_DOCKER.md

# Run or Build h2oGPT Docker

* Install Docker for [Linux](https://docs.docker.com/engine/install/ubuntu/)
* Install Docker for [Windows](https://docs.docker.com/desktop/install/windows-install/)
* Install Docker for [MAC](https://docs.docker.com/desktop/install/mac-install/)

## Linux Ubuntu: Setup Docker for CPU Inference

No special docker instructions are required, just follow [these instructions](https://docs.docker.com/engine/install/ubuntu/) to get docker setup at all, i.e.:
```bash
sudo apt update
sudo apt install -y apt-transport-https ca-certificates curl software-properties-common
curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo apt-key add -
sudo add-apt-repository -y "deb [arch=amd64] https://download.docker.com/linux/ubuntu jammy stable"
apt-cache policy docker-ce
sudo apt install -y docker-ce
sudo systemctl status docker
```
replace `focal` (Ubuntu 20) with `jammy` for Ubuntu 22.

Add your user as part of `docker` group:
```bash
sudo usermod -aG docker $USER
```
exit shell, login back in, and run:
```bash
newgrp docker
```
which avoids having to reboot.  Or just reboot to have docker access.  If this cannot be done without entering root access, then edit the `/etc/group` and add your user to group `docker`.

## Linux Ubuntu: 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](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#docker) for more examples.
```bash
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/stable/deb/nvidia-container-toolkit.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 -y nvidia-container-runtime
sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker
```

Confirm runs nvidia-smi from within docker without errors:
```bash
sudo docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi
```

If running on A100's, might require [Installing Fabric Manager](INSTALL.md#install-and-run-nvidia-fabric-manager-on-systems-with-multiple-a100-or-h100-gpus) and [Installing GPU Manager](INSTALL.md#install-nvidia-gpu-manager-on-systems-with-multiple-a100-or-h100-gpus).

## Prebuild Docker for Windows/Linux x86

All available public h2oGPT docker images can be found in [Google Container Registry](https://console.cloud.google.com/gcr/images/vorvan/global/h2oai/h2ogpt-runtime).  These require cuda drivers that handle CUDA 12.1 or higher.

Ensure image is up-to-date by running:
```bash
docker pull gcr.io/vorvan/h2oai/h2ogpt-runtime:0.2.1
```

## Build Docker

The GCR contains nightly and released images for x86.

### x86

The default docker supports CUDA or CPU for x86, and HF models supported by torch on Metal M1/M2.

### MAC Metal or other architectures

Choose your llama_cpp_python options, by changing `CMAKE_ARGS` to whichever system you have according to [llama_cpp_python backend documentation](https://github.com/abetlen/llama-cpp-python?tab=readme-ov-file#supported-backends).

For example, for Metal M1/M2 support of llama.cpp GGUF files, one should change `CMAKE_ARGS` in [docker_build_script_ubuntu.sh](../docker_build_script_ubuntu.sh) to have:
```bash
export CMAKE_ARGS="-DLLAMA_METAL=on"
```
and remove `GGML_CUDA=1`, so that the docker image is Metal Compatible for llama.cpp GGUF files.  Otherwise, Torch supports Metal M1/M2 directly without changes.

### Build

To build the docker image after any local changes (to support Metal for GGUF files, etc.):
```bash
# 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:0.2.1` with `h2ogpt:latest` in any docker run commands.

## Linux: Run h2oGPT using Docker

An example running h2oGPT via docker using Zephyr 7B Beta model is:
```bash
mkdir -p ~/.cache/huggingface/hub/
mkdir -p ~/.triton/cache/
mkdir -p ~/.config/vllm/
mkdir -p ~/.cache
mkdir -p ~/save
mkdir -p ~/user_path
mkdir -p ~/db_dir_UserData
mkdir -p ~/users
mkdir -p ~/db_nonusers
mkdir -p ~/llamacpp_path
mkdir -p ~/h2ogpt_auth
echo '["key1","key2"]' > ~/h2ogpt_auth/h2ogpt_api_keys.json
export GRADIO_SERVER_PORT=7860
export OPENAI_SERVER_PORT=5000
docker run \
       --gpus all \
       --runtime=nvidia \
       --shm-size=2g \
       -p $GRADIO_SERVER_PORT:$GRADIO_SERVER_PORT \
       -p $OPENAI_SERVER_PORT:$OPENAI_SERVER_PORT \
       --rm --init \
       --network host \
       -v /etc/passwd:/etc/passwd:ro \
       -v /etc/group:/etc/group:ro \
       -u `id -u`:`id -g` \
       -v "${HOME}"/.cache/huggingface/hub/:/workspace/.cache/huggingface/hub \
       -v "${HOME}"/.config:/workspace/.config/ \
       -v "${HOME}"/.triton:/workspace/.triton/  \
       -v "${HOME}"/save:/workspace/save \
       -v "${HOME}"/user_path:/workspace/user_path \
       -v "${HOME}"/db_dir_UserData:/workspace/db_dir_UserData \
       -v "${HOME}"/users:/workspace/users \
       -v "${HOME}"/db_nonusers:/workspace/db_nonusers \
       -v "${HOME}"/llamacpp_path:/workspace/llamacpp_path \
       -v "${HOME}"/h2ogpt_auth:/workspace/h2ogpt_auth \
       -e GRADIO_SERVER_PORT=$GRADIO_SERVER_PORT \
       gcr.io/vorvan/h2oai/h2ogpt-runtime:0.2.1 /workspace/generate.py \
          --base_model=HuggingFaceH4/zephyr-7b-beta \
          --use_safetensors=True \
          --prompt_type=zephyr \
          --save_dir='/workspace/save/' \
          --auth_filename='/workspace/h2ogpt_auth/auth.db' \
          --h2ogpt_api_keys='/workspace/h2ogpt_auth/h2ogpt_api_keys.json' \
          --auth='/workspace/h2ogpt_auth/h2ogpt_api_keys.json' \
          --use_gpu_id=False \
          --user_path=/workspace/user_path \
          --langchain_mode="LLM" \
          --langchain_modes="['UserData', 'LLM']" \
          --score_model=None \
          --max_max_new_tokens=2048 \
          --max_new_tokens=1024 \
          --use_auth_token="${HUGGING_FACE_HUB_TOKEN}" \
          --openai_port=$OPENAI_SERVER_PORT
```
Use `docker run -d` to run in detached background. Then go to http://localhost:7860/ or http://127.0.0.1:7860/.  For authentication, if use `--auth=/workspace/h2ogpt_auth/auth.json` instead, then do not need to use `--auth_filename`.  For keyed access, change key1 and key2 for `h2ogpt_api_keys` or for open-access remove `--h2ogpt_api_keys` line.

If one does not need access to private repo, can remove `--use_auth_token` line, else set env `HUGGING_FACE_HUB_TOKEN` so h2oGPT gets the token.

For single GPU use `--gpus '"device=0"'` or for 2 GPUs use `--gpus '"device=0,1"'` instead of `--gpus all`.

See [README_GPU](README_GPU.md) for more details about what to run.

## Linux: Run h2oGPT in docker offline:

Ensure $HOME/users and $HOME/db_nonusers are writeable by user running docker, then run:
```bash

export TRANSFORMERS_OFFLINE=1
export GRADIO_SERVER_PORT=7860
export OPENAI_SERVER_PORT=5000
export HF_HUB_OFFLINE=1
docker run --gpus all \
--runtime=nvidia \
--shm-size=2g \
-e TRANSFORMERS_OFFLINE=$TRANSFORMERS_OFFLINE \
-e HUGGING_FACE_HUB_TOKEN=$HUGGING_FACE_HUB_TOKEN \
-e HF_HUB_OFFLINE=$HF_HUB_OFFLINE \
-e HF_HOME="/workspace/.cache/huggingface/" \
-p $GRADIO_SERVER_PORT:$GRADIO_SERVER_PORT \
-p $OPENAI_SERVER_PORT:$OPENAI_SERVER_PORT \
--rm --init \
--network host \
-v /etc/passwd:/etc/passwd:ro \
-v /etc/group:/etc/group:ro \
-u `id -u`:`id -g` \
-v "${HOME}"/.cache/huggingface/:/workspace/.cache/huggingface \
-v "${HOME}"/.cache/torch/:/workspace/.cache/torch \
-v "${HOME}"/.cache/transformers/:/workspace/.cache/transformers \
-v "${HOME}"/save:/workspace/save \
-v "${HOME}"/user_path:/workspace/user_path \
-v "${HOME}"/db_dir_UserData:/workspace/db_dir_UserData \
-v "${HOME}"/users:/workspace/users \
-v "${HOME}"/db_nonusers:/workspace/db_nonusers \
-v "${HOME}"/llamacpp_path:/workspace/llamacpp_path \
-e GRADIO_SERVER_PORT=$GRADIO_SERVER_PORT \
 gcr.io/vorvan/h2oai/h2ogpt-runtime:0.2.1 \
 /workspace/generate.py \
 --base_model=mistralai/Mistral-7B-Instruct-v0.2 \
 --use_safetensors=False \
 --prompt_type=mistral \
 --save_dir='/workspace/save/' \
 --use_gpu_id=False \
 --user_path=/workspace/user_path \
 --langchain_mode="LLM" \
 --langchain_modes="['UserData', 'MyData', 'LLM']" \
 --score_model=None \
 --max_max_new_tokens=2048 \
 --max_new_tokens=1024 \
 --visible_visible_models=False \
 --openai_port=$OPENAI_SERVER_PORT \
 --gradio_offline_level=2
```
Depending upon if use links, may require more specific mappings to direct location not linked location that cannot be used, e.g.
```bash
-v "${HOME}"/.cache/huggingface/hub:/workspace/.cache/huggingface/hub \
 -v "${HOME}"/.cache:/workspace/.cache \
```
You can also specify the cache location:
```bash
 -e TRANSFORMERS_CACHE="/workspace/.cache/" \
 ```


## Run h2oGPT +  vLLM or vLLM using Docker

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

For the vLLM server running on 2 GPUs using h2oai/h2ogpt-4096-llama2-7b-chat model, run:
```bash
unset CUDA_VISIBLE_DEVICES
mkdir -p $HOME/.cache/huggingface/hub
mkdir -p $HOME/.cache/huggingface/modules/
mkdir -p $HOME/.triton/cache/
mkdir -p $HOME/.config/vllm
docker run \
    --runtime=nvidia \
    --gpus '"device=0,1"' \
    --shm-size=10.24gb \
    -p 5000:5000 \
    --rm --init \
    -e NCCL_IGNORE_DISABLED_P2P=1 \
    -e HUGGING_FACE_HUB_TOKEN=$HUGGING_FACE_HUB_TOKEN \
    -e VLLM_NO_USAGE_STATS=1 \
    -e VLLM_NCCL_SO_PATH=/usr/local/lib/python3.10/dist-packages/nvidia/nccl/lib/libnccl.so.2 \
    -e DO_NOT_TRACK=1 \
    -e NUMBA_CACHE_DIR=/tmp/ \
    -v /etc/passwd:/etc/passwd:ro \
    -v /etc/group:/etc/group:ro \
    -u `id -u`:`id -g` \
    -v "${HOME}"/.cache:$HOME/.cache/ -v "${HOME}"/.config:$HOME/.config/   -v "${HOME}"/.triton:$HOME/.triton/  \
    --network host \
    vllm/vllm-openai:latest \
        --port=5000 \
        --host=0.0.0.0 \
        --model=h2oai/h2ogpt-4096-llama2-7b-chat \
        --tokenizer=hf-internal-testing/llama-tokenizer \
        --tensor-parallel-size=2 \
        --seed 1234 \
        --trust-remote-code \
        --download-dir=/workspace/.cache/huggingface/hub &>> logs.vllm_server.txt
```
Use `docker run -d` to run in detached background.

Checks the logs `logs.vllm_server.txt` to make sure server is running.
If ones sees similar output to below, then endpoint it up & running.
```bash
INFO:     Started server process [7]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:5000 (Press CTRL+C to quit
```

For LLaMa-2 70B AWQ in docker using vLLM run:
```bash
mkdir -p $HOME/.cache/huggingface/hub
mkdir -p $HOME/.cache/huggingface/modules/
mkdir -p $HOME/.triton/cache/
mkdir -p $HOME/.config/vllm
docker run -d \
    --runtime=nvidia \
    --gpus '"device=0,1"' \
    --shm-size=10.24gb \
    -p 5000:5000 \
    -e NCCL_IGNORE_DISABLED_P2P=1 \
    -e HUGGING_FACE_HUB_TOKEN=$HUGGING_FACE_HUB_TOKEN \
    -e VLLM_NO_USAGE_STATS=1 \
    -e VLLM_NCCL_SO_PATH=/usr/local/lib/python3.10/dist-packages/nvidia/nccl/lib/libnccl.so.2 \
    -e DO_NOT_TRACK=1 \
    -e NUMBA_CACHE_DIR=/tmp/ \
    -v /etc/passwd:/etc/passwd:ro \
    -v /etc/group:/etc/group:ro \
    -u `id -u`:`id -g` \
    -v "${HOME}"/.cache:$HOME/.cache/ -v "${HOME}"/.config:$HOME/.config/   -v "${HOME}"/.triton:$HOME/.triton/  \
    --network host \
    vllm/vllm-openai:latest \
        --port=5000 \
        --host=0.0.0.0 \
        --model=h2oai/h2ogpt-4096-llama2-70b-chat-4bit \
        --tensor-parallel-size=2 \
        --seed 1234 \
        --trust-remote-code \
	    --max-num-batched-tokens 8192 \
	    --quantization awq \
	    --worker-use-ray \
	    --enforce-eager \
        --download-dir=/workspace/.cache/huggingface/hub &>> logs.vllm_server.70b_awq.txt
```
for choice of port, IP,  model, some number of GPUs matching tensor-parallel-size, etc.
We add `--enforce-eager` to avoid excess memory usage by CUDA graphs.

For 4*A10G on AWS using LLaMa-2 70B AWQ run:
```bash
mkdir -p $HOME/.cache/huggingface/hub
mkdir -p $HOME/.cache/huggingface/modules/
mkdir -p $HOME/.triton/cache/
mkdir -p $HOME/.config/vllm
docker run -d \
    --runtime=nvidia \
    --gpus '"device=0,1,2,3"' \
    --shm-size=10.24gb \
    -p 5000:5000 \
    -e NCCL_IGNORE_DISABLED_P2P=1 \
    -e HUGGING_FACE_HUB_TOKEN=$HUGGING_FACE_HUB_TOKEN \
    -e VLLM_NO_USAGE_STATS=1 \
    -e VLLM_NCCL_SO_PATH=/usr/local/lib/python3.10/dist-packages/nvidia/nccl/lib/libnccl.so.2 \
    -e DO_NOT_TRACK=1 \
    -e NUMBA_CACHE_DIR=/tmp/ \
    -v /etc/passwd:/etc/passwd:ro \
    -v /etc/group:/etc/group:ro \
    -u `id -u`:`id -g` \
    -v "${HOME}"/.cache:$HOME/.cache/ -v "${HOME}"/.config:$HOME/.config/   -v "${HOME}"/.triton:$HOME/.triton/  \
    --network host \
    vllm/vllm-openai:latest \
        --port=5000 \
        --host=0.0.0.0 \
        --model=h2oai/h2ogpt-4096-llama2-70b-chat-4bit \
        --tensor-parallel-size=4 \
        --seed 1234 \
        --trust-remote-code \
	    --max-num-batched-tokens 8192 \
	    --max-num-seqs 256 \
	    --quantization awq \
	    --worker-use-ray \
	    --enforce-eager \
        --download-dir=/workspace/.cache/huggingface/hub &>> logs.vllm_server.70b_awq.txt
```
One can lower `--max-num-seqs` and `--max-num-batched-tokens` to reduce memory usage.

### Curl Test

One can also verify the endpoint by running following curl command.
```bash
curl http://localhost:5000/v1/completions \
    -H "Content-Type: application/json" \
    -d '{
    "model": "h2oai/h2ogpt-4096-llama2-7b-chat",
    "prompt": "San Francisco is a",
    "max_tokens": 7,
    "temperature": 0
    }'
```
If one sees similar output to below, then endpoint it up & running.

```json
{
    "id": "cmpl-4b9584f743ff4dc590f0c168f82b063b",
    "object": "text_completion",
    "created": 1692796549,
    "model": "h2oai/h2ogpt-4096-llama2-7b-chat",
    "choices": [
        {
            "index": 0,
            "text": "city in Northern California that is known",
            "logprobs": null,
            "finish_reason": "length"
        }
    ],
    "usage": {
        "prompt_tokens": 5,
        "total_tokens": 12,
        "completion_tokens": 7
    }
}
```

If one needs to only setup vLLM one can stop here.

### Run h2oGPT
Just add to the above docker run command:
```bash
        --inference_server="vllm:0.0.0.0:5000"
```
where `--base_model` should match for how ran vLLM and h2oGPT. Make sure to set `--inference_server` argument to the correct vllm endpoint.

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](README_InferenceServers.md) for more information on how to setup vLLM.

## 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)
```bash
export MODEL=h2oai/h2ogpt-4096-llama2-7b-chat
docker run -d --gpus '"device=0"' \
       --shm-size 1g \
       --network host \
       -p 6112:80 \
       -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.  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.:
```bash
mkdir -p $HOME/.cache/huggingface/hub/
mkdir -p $HOME/.cache/huggingface/modules/
export MODEL=h2oai/h2ogpt-4096-llama2-7b-chat
docker run -d --gpus '"device=0"' \
        --shm-size 1g \
        -p 6112:80 \
        -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 detached 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 to the docker run line:
```bash
    --inference_server=http://localhost:6112
````
Note the h2oGPT container has `--network host` with same port inside and outside so the other container on same host can see it.  Otherwise use actual IP addersses if on separate hosts.

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

## Make UserData db for generate.py using Docker

To make UserData db for generate.py, put pdfs, etc. into path user_path and run:
```bash
mkdir -p ~/.cache
mkdir -p ~/save
mkdir -p ~/user_path
mkdir -p ~/db_dir_UserData
docker run \
       --gpus all \
       --runtime=nvidia \
       --shm-size=2g \
       --rm --init \
       --network host \
       -v /etc/passwd:/etc/passwd:ro \
       -v /etc/group:/etc/group:ro \
       -u `id -u`:`id -g` \
       -v "${HOME}"/.cache:/workspace/.cache \
       -v "${HOME}"/save:/workspace/save \
       -v "${HOME}"/user_path:/workspace/user_path \
       -v "${HOME}"/db_dir_UserData:/workspace/db_dir_UserData \
       gcr.io/vorvan/h2oai/h2ogpt-runtime:0.2.1 /workspace/src/make_db.py
```

Once db is made, can use in generate.py like:
```bash
mkdir -p ~/.cache
mkdir -p ~/save
mkdir -p ~/user_path
mkdir -p ~/db_dir_UserData
mkdir -p ~/users
mkdir -p ~/db_nonusers
mkdir -p ~/llamacpp_path
docker run \
       --gpus '"device=0"' \
       --runtime=nvidia \
       --shm-size=2g \
       -p 7860:7860 \
       --rm --init \
       --network host \
       -v /etc/passwd:/etc/passwd:ro \
       -v /etc/group:/etc/group:ro \
       -u `id -u`:`id -g` \
       -v "${HOME}"/.cache:/workspace/.cache \
       -v "${HOME}"/save:/workspace/save \
       -v "${HOME}"/user_path:/workspace/user_path \
       -v "${HOME}"/db_dir_UserData:/workspace/db_dir_UserData \
       -v "${HOME}"/users:/workspace/users \
       -v "${HOME}"/db_nonusers:/workspace/db_nonusers \
       -v "${HOME}"/llamacpp_path:/workspace/llamacpp_path \
       gcr.io/vorvan/h2oai/h2ogpt-runtime:0.2.1 /workspace/generate.py \
          --base_model=h2oai/h2ogpt-4096-llama2-7b-chat \
          --use_safetensors=True \
          --prompt_type=llama2 \
          --save_dir='/workspace/save/' \
          --use_gpu_id=False \
          --score_model=None \
          --max_max_new_tokens=2048 \
          --max_new_tokens=1024 \
          --langchain_mode=LLM
```

For a more detailed description of other parameters of the make_db script, checkout the definition in this file: https://github.com/h2oai/h2ogpt/blob/main/src/make_db.py




## Docker Compose Setup & Inference

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

2. Build and run the container

    ```bash
    docker-compose up -d --build
    ```

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

4. See logs:

    ```bash
    docker-compose logs -f
    ```

5. Clean everything up:

    ```bash
    docker-compose down --volumes --rmi all
    ```