local-llm/README.md

# librechat-mistral

A quick prototype to self-host [LibreChat](https://github.com/danny-avila/LibreChat) with [Mistral](https://mistral.ai/news/announcing-mistral-7b/), and an OpenAI-like api provided by [LiteLLM](https://github.com/BerriAI/litellm) on the side.

## Goals

* Streamline deployment of a local LLM for experimentation purpose.
* Deploy a ChatGPT Clone for daily use.
* Deploy an OpenAI-like API for hacking on Generative AI using well-supported libraries.
* Use docker to prepare for an eventual deployment on a container orchestration platform like Kubernetes.

## Getting started

### Prerequisites

* Linux or WSL2
* docker

### Steps for NVIDIA GPU

1. Make sure your drivers are up to date.
2. Install the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html).
3. Clone the repo.
4. Copy the AMD compose spec to select it. `cp docker-compose.nvidia.yml docker.compose.yml`
5. Run `docker compose up`. Wait for a few minutes for the model to be downloaded and served.
6. Browse http://localhost:3080/
7. Create an admin account and start chatting!

### Steps for AMD GPU

**Warning: AMD was not tested on Windows and support seems to not be as good as on Linux.**

1. Make sure your drivers are up to date.
2. Clone the repo.
3. Copy the AMD compose spec to select it. `cp docker-compose.amd.yml docker.compose.yml`
4. If you are using an RX (consumer) series GPU, you *may* need to set `HSA_OVERRIDE_GFX_VERSION` to an appropriate value for the model of your GPU. You will need to look it up. The value can be set in *docker-compose.yml*,
5. Run `docker compose up`. Wait for a few minutes for the model to be downloaded and served.
6. Browse http://localhost:3080/
7. Create an admin account and start chatting!

### Steps for NO GPU (use CPU)

**Warning: This may be very slow depending on your CPU and may us a lot of RAM depending on the model**

1. Make sure your drivers are up to date.
2. Clone the repo.
3. Copy the CPU compose spec to select it. `cp docker-compose.cpu.yml docker.compose.yml`
4. Run `docker compose up`. Wait for a few minutes for the model to be downloaded and served.
5. Browse http://localhost:3080/
6. Create an admin account and start chatting!

## Configuring additional models

### SASS services

Read: https://docs.librechat.ai/install/configuration/dotenv.html#endpoints

**TL:DR**

Let say we want to configure an OpenAI API key.

1. Open the *.env* file.
2. Uncomment the line `# OPENAI_API_KEY=user_provided`.
3. Replace `user_provided` with your API key.
4. Restart LibreChat `docker compose restart librechat`.

Refer to the [LibreChat documentation](https://docs.librechat.ai/install/configuration/ai_setup.html#openai) for the full list of configuration options.

### Ollama (self-hosted)

Browse the [Ollama models library](https://ollama.ai/library) to find a model you wish to add. For this example we will add [mistral-openorca](https://ollama.ai/library/mistral-openorca)

1. Open the *docker compose.yml* file.
2. Find the `ollama` service. Find the `command:` option under the ollama sevice. Append the name of the model you wish to add at the end of the list (eg: `command: mistral mistral-openorca`).
3. Open the *litellm/config.yaml* file.
4. Add the following a the end of the file, replace {model_name} placeholders with the name of your model
``` yaml
  - model_name: {model_name}
    litellm_params:
      model: ollama/{model_name}
      api_base: http://ollama:11434
```
eg:
``` yaml
  - model_name: mistral-openorca
    litellm_params:
      model: ollama/mistral-openorca
      api_base: http://ollama:11434
```
5. Restart the stack `docker compose restart`. Wait for a few minutes for the model to be downloaded and served.

## Architecture components

* [LibreChat](https://github.com/danny-avila/LibreChat) is a ChatGPT clone with support for multiple AI endpoints. It's deployed alongside a [MongoDB](https://github.com/mongodb/mongo) database and [Meillisearch](https://github.com/meilisearch/meilisearch) for search. It's exposed on http://localhost:3080/.
* [LiteLLM](https://github.com/BerriAI/litellm) is an OpenAI-like API. It is exposed on http://localhost:8000/ without any authentication by default.
* [Ollama](https://github.com/ollama/ollama) manages and serve the local models.