Skip to content

Latest commit

 

History

History
180 lines (119 loc) · 9.17 KB

README.md

File metadata and controls

180 lines (119 loc) · 9.17 KB
Raw

Chat Application

This recipe helps developers start building their own custom LLM enabled chat applications. It consists of two main components: the Model Service and the AI Application.

There are a few options today for local Model Serving, but this recipe will use llama-cpp-python and their OpenAI compatible Model Service. There is a Containerfile provided that can be used to build this Model Service within the repo, model_servers/llamacpp_python/base/Containerfile.

The AI Application will connect to the Model Service via its OpenAI compatible API. The recipe relies on Langchain's python package to simplify communication with the Model Service and uses Streamlit for the UI layer. You can find an example of the chat application below.

Try the Chat Application

The Podman Desktop AI Lab Extension includes this recipe among others. To try it out, open Recipes Catalog -> Chatbot and follow the instructions to start the application.

Build the Application

The rest of this document will explain how to build and run the application from the terminal, and will go into greater detail on how each container in the Pod above is built, run, and what purpose it serves in the overall application. All the recipes use a central Makefile that includes variables populated with default values to simplify getting started. Please review the Makefile docs, to learn about further customizing your application.

This application requires a model, a model service and an AI inferencing application.

Quickstart

To run the application with pre-built images from quay.io/ai-lab, use make quadlet. This command builds the application's metadata and generates Kubernetes YAML at ./build/chatbot.yaml to spin up a Pod that can then be launched locally. Try it with:

make quadlet
podman kube play build/chatbot.yaml

This will take a few minutes if the model and model-server container images need to be downloaded. The Pod is named chatbot, so you may use Podman to manage the Pod and its containers:

podman pod list
podman ps

Once the Pod and its containers are running, the application can be accessed at http://localhost:8501. However, if you started the app via the podman desktop UI, a random port will be assigned instead of 8501. Please use the AI App Details Open AI App button to access it instead. Please refer to the section below for more details about interacting with the chatbot application.

To stop and remove the Pod, run:

podman pod stop chatbot
podman pod rm chatbot

Download a model

If you are just getting started, we recommend using granite-7b-lab. This is a well performant mid-sized model with an apache-2.0 license. In order to use it with our Model Service we need it converted and quantized into the GGUF format. There are a number of ways to get a GGUF version of granite-7b-lab, but the simplest is to download a pre-converted one from huggingface.co here: https://huggingface.co/instructlab/granite-7b-lab-GGUF.

The recommended model can be downloaded using the code snippet below:

cd ../../../models
curl -sLO https://huggingface.co/instructlab/granite-7b-lab-GGUF/resolve/main/granite-7b-lab-Q4_K_M.gguf
cd ../recipes/natural_language_processing/chatbot

A full list of supported open models is forthcoming.

Build the Model Service

The complete instructions for building and deploying the Model Service can be found in the llamacpp_python model-service document.

The Model Service can be built from make commands from the llamacpp_python directory.

# from path model_servers/llamacpp_python from repo containers/ai-lab-recipes
make build

Checkout the Makefile to get more details on different options for how to build.

Deploy the Model Service

The local Model Service relies on a volume mount to the localhost to access the model files. It also employs environment variables to dictate the model used and where its served. You can start your local Model Service using the following make command from model_servers/llamacpp_python set with reasonable defaults:

# from path model_servers/llamacpp_python from repo containers/ai-lab-recipes
make run

Build the AI Application

The AI Application can be built from the make command:

# Run this from the current directory (path recipes/natural_language_processing/chatbot from repo containers/ai-lab-recipes)
make build

Deploy the AI Application

Make sure the Model Service is up and running before starting this container image. When starting the AI Application container image we need to direct it to the correct MODEL_ENDPOINT. This could be any appropriately hosted Model Service (running locally or in the cloud) using an OpenAI compatible API. In our case the Model Service is running inside the Podman machine so we need to provide it with the appropriate address 10.88.0.1. To deploy the AI application use the following:

# Run this from the current directory (path recipes/natural_language_processing/chatbot from repo containers/ai-lab-recipes)
make run

Interact with the AI Application

Everything should now be up an running with the chat application available at http://localhost:8501. By using this recipe and getting this starting point established, users should now have an easier time customizing and building their own LLM enabled chatbot applications.

Embed the AI Application in a Bootable Container Image

To build a bootable container image that includes this sample chatbot workload as a service that starts when a system is booted, run: make -f Makefile bootc. You can optionally override the default image / tag you want to give the make command by specifying it as follows: make -f Makefile BOOTC_IMAGE=<your_bootc_image> bootc.

Substituting the bootc/Containerfile FROM command is simple using the Makefile FROM option.

make FROM=registry.redhat.io/rhel9/rhel-bootc:9.4 bootc

Selecting the ARCH for the bootc/Containerfile is simple using the Makefile ARCH= variable.

make ARCH=x86_64 bootc

The magic happens when you have a bootc enabled system running. If you do, and you'd like to update the operating system to the OS you just built with the chatbot application, it's as simple as ssh-ing into the bootc system and running:

bootc switch quay.io/ai-lab/chatbot-bootc:latest

Upon a reboot, you'll see that the chatbot service is running on the system. Check on the service with:

ssh user@bootc-system-ip
sudo systemctl status chatbot

What are bootable containers?

What's a bootable OCI container and what's it got to do with AI?

That's a good question! We think it's a good idea to embed AI workloads (or any workload!) into bootable images at build time rather than at runtime. This extends the benefits, such as portability and predictability, that containerizing applications provides to the operating system. Bootable OCI images bake exactly what you need to run your workloads into the operating system at build time by using your favorite containerization tools. Might I suggest podman?

Once installed, a bootc enabled system can be updated by providing an updated bootable OCI image from any OCI image registry with a single bootc command. This works especially well for fleets of devices that have fixed workloads - think factories or appliances. Who doesn't want to add a little AI to their appliance, am I right?

Bootable images lend toward immutable operating systems, and the more immutable an operating system is, the less that can go wrong at runtime!

Creating bootable disk images

You can convert a bootc image to a bootable disk image using the quay.io/centos-bootc/bootc-image-builder container image.

This container image allows you to build and deploy multiple disk image types from bootc container images.

Default image types can be set via the DISK_TYPE Makefile variable.

make bootc-image-builder DISK_TYPE=ami