9.9 KiB
Introduction
This project dockerises the deployment of oobabooga/text-generation-webui and its variants. It provides a default configuration (corresponding to a vanilla deployment of the application) as well as pre-configured support for other set-ups (e.g., llama-cpp for CPU-only inferencing, the triton and cuda branches of GPTQ). The images are available on Docker Hub: https://hub.docker.com/r/atinoda/text-generation-webui
The goal of this project is to be to oobabooga/text-generation-webui, what AbdBarho/stable-diffusion-webui-docker is to AUTOMATIC1111/stable-diffusion-webui.
Usage
This project currently supports Linux as the deployment platform. It will also probably work using WSL2.
Pre-Requisites
- docker
- docker compose
- CUDA docker runtime
Docker Compose
This is the recommended deployment method (it is the easiest and quickest way to manage folders and settings through updates and reinstalls). The recommend variant is default (it is an enhanced version of the vanilla application).
Select variant
Each variant has the 'extras' included in default but has some changes made as described in the table. Tagged release versions are published on a regular basis - check hub.docker.com/r/atinoda/text-generation-webui for available tags. Pseudo-versions will be selected periodically from the main branch and uploaded with a date tag to establish more frequent stable milestones. Pulling an untagged variant will pull either the latest release version or latest pseudo-version, whichever is most recent. Bleeding-edge is available via nightly builds of each variant. Choose the desired variant by setting the image :tag in docker-compose.yml to one of the following options:
| Variant | Description |
|---|---|
default |
Implementation of the vanilla deployment from source. Plus pre-installed ExLlamaV2 library from turboderp/exllamav2, and CUDA GPU offloading enabled for llama-cpp. This version is recommended for most users. |
triton |
Updated GPTQ-for-llama using the latest triton branch from qwopqwop200/GPTQ-for-LLaMa. Suitable for Linux only. This version is accurate but a little slow. |
cuda |
Updated GPTQ-for-llama using the latest cuda branch from qwopqwop200/GPTQ-for-LLaMa. This version is very slow! |
llama-cpu |
GPU supported is REMOVED from llama-cpp. Suitable for systems without a CUDA-capable GPU. This is only for when GPU acceleration is not available and is a slower way to run models! |
monkey-patch |
Use LoRAs in 4-Bit GPTQ-for-llama mode. DEPRECATION WARNING: This version is outdated, but will remain for now. |
{VARIANT}-{VERSION} |
Build of each {VARIANT} tagged with the release {VERSION} of the text-generation-webui (e.g., default-v1.5). Visit obabooga/text-generation-webui/releases for release notes. Dated milestone pseudo-version tags are also available (e.g., default-2023.09.02). Visit hub.docker.com/r/atinoda/text-generation-webui to see the available milestones. |
{VARIANT}-nightly |
Automated nightly build of the {VARIANT}. These images are built and pushed automatically - they are untested and may be unstable. Suitable when more frequent updates are required and instability is not an issue. |
See: oobabooga/text-generation-webui/blob/main/docs/GPTQ-models-(4-bit-mode).md, obabooga/text-generation-webui/blob/main/docs/llama.cpp-models.md, and oobabooga/text-generation-webui/blob/main/docs/ExLlama.md for more information on variants.
Deploy
Deploy the service:
docker compose up
Remove
Remove the service:
docker compose down -v
Configuration
These configuration instructions describe the relevant details for this docker wrapper. Refer to oobabooga/text-generation-webui documentation for usage of the application itself.
Ports
Three commonly used ports are exposed:
| Port | Description | Configuration |
|---|---|---|
7860 |
Web UI port | Pre-configured and enabled in docker-compose.yml |
5000 |
API port | Enable by adding --api --extensions api to launch args then uncomment mapping in docker-compose.yml |
5005 |
Streaming port | Enable by adding --api --extensions api to launch args then uncomment mapping in docker-compose.yml |
Extensions may use additional ports - check the application documentation for more details.
Volumes
The provided example docker compose maps several volumes from the local config directory into the container: loras, models, presets, prompts, training, extensions. If these folders are empty, they will be initialised when the container is run.
Extensions will persist their state between container launches if you use a mapped folder - but they will not automatically update when a new image is released, so this feature is disabled by default. The whole extensions folder can be mapped (all extensions are persisted) or individual extensions can be mapped one at a time. Examples are given in the docker-compose.yml.
If you are getting an error about missing files, try clearing these folders and letting the service re-populate them.
Extra launch arguments
Extra launch arguments can be defined in the environment variable EXTRA_LAUNCH_ARGS (e.g., "--model MODEL_NAME", to load a model at launch). The provided default extra arguments are --verbose and --listen (which makes the webui available on your local network) and these are set in the docker-compose.yml.
Launch arguments should be defined as a space-separated list, just like writing them on the command line. These arguments are passed to the server.py module.
Runtime extension build
Extensions which should be built during startup can be defined in the environment variable BUILD_EXTENSIONS_LIVE (e.g., "silero_tts whisper_stt", will rebuild those extensions at launch). This feature may be useful if you are developing a third-party extension and need its dependencies to refresh at launch.
Startup times will be much slower if you use this feature, because it will rebuild the named extensions every time the container is started (i.e., don't use this feature unless you are certain that you need it.)
Extension names for runtime build should be defined as a space-separated list.
Updates
These projects are moving quickly! To update to the most recent version on Docker hub, pull the latest image:
docker compose pull
Then recreate the container:
docker compose up
When the container is launched, it will print out how many commits behind origin the current build is, so you can decide if you want to update it. Docker hub images will be periodically updated. The default-nightly image is built every day but it is not manually tested. If you need bleeding edge versions you must build locally.
Build (optional)
The provided docker-compose.yml.build shows how to build the image locally. You can use it as a reference to modify the original docker-compose.yml, or you can rename it and use it as-is. Choose the desired variant to build by setting the build target and then run:
docker compose build
To do a clean build and ensure the latest version:
docker compose build --no-cache
If you choose a different variant later, you must rebuild the image.
Developers / Advanced Users
The Dockerfile can be easily modified to compile and run the application from a local source folder. This is useful if you want to do some development or run a custom version. See the Dockerfile itself for instructions on how to do this.
Support is not provided for this deployment pathway. It is assumed that you are competent and willing to do your own debugging! Pro-tip: start by placing a text-generation-webui repo into the project folder.
Standalone Container
NOT recommended, instructions are included for completeness.
Run
Run a network accessible container (and destroy it upon completion):
docker run -it --rm -e EXTRA_LAUNCH_ARGS="--listen --verbose" --gpus all -p 7860:7860 atinoda/text-generation-webui:default
Build and run (optional)
Build the image for the default target and tag it as local :
docker build --target default -t text-generation-webui:local .
Run the local image with local network access (and destroy it upon completion):
docker run -it --rm -e EXTRA_LAUNCH_ARGS="--listen --verbose" --gpus all -p 7860:7860 text-generation-webui:local
Contributions
Contributions are welcomed - please feel free to submit a PR. More variants (e.g., AMD/ROC-M support) and Windows support can help lower the barrier to entry, make this technology accessible to as many people as possible, and push towards democratising the severe impacts that AI is having on our society.
Also - it's fun to code and LLMs are cool.
DISCLAIMER
THE SOFTWARE IS PROVIDED 'AS IS', WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.