I’ve recently found a good setup for using locally powered LLMs in the IDEs I use every day. Specifically, I almost exclusively use Jetbrains IDEs like Rider, WebStorm, PyCharm, etc. After building a large chunk of my own LLM integration for these IDEs, I have finally found a setup that I wanted via the Open Source plugin called CodeGPT.

In this guide, I’ll walk you through the steps on how I integrated a locally running Large Language Model (LLM), like the fine-tuned Mistral-7B, into your JetBrains IDE setup. This approach offers the benefits of a powerful AI assistant right in your development environment.

This software can be very fiddly to setup, and is subject to a lot of environment specific issues. I highly recommend that if you are going to use text-generation-web-ui, you understand that this is rapidly changing project, so can be quite unstable.

Setting Up the Text Generation Web UI

The first step is to set up the text-generation-web-ui, a handy tool for running your LLM. Here’s how:

1. Installation: Head over to the text-generation-web-ui GitHub page and follow the instructions to install it. Remember not to use the --api option when starting it.

2. Running Locally: Docker is the simplest method for Linux users, but there are also shell and batch scripts for local development machines.

3. Configuration:

   • The default API port is 5000, as shown in the example .env file. Feel free to change this in your .env file if necessary. For instance, I use port 5010.
   • Load your desired model, like Mistral-7B-OpenOrca, to avoid common errors such as “No tokenizer is loaded.”

4. Session Setup: Under Session, choose the “openai” option, apply the changes, and restart.

Troubleshooting

I ran into several issues trying to get text-generation-web-ui to work with the OpenAI option selection, below are a few of the issues I hit and the steps I took to resolve.

1. Missing sentence_transformers and causes it to fail to restart with OpenAI support.
   • Update the requirements.txt and use docker compose up -d --build to rebuild and restart your container
2. Missing tiktoken, this can still come up looking like an issue with sentence_transformers
   • Update the requirements.txt and use docker compose up -d --build to rebuild and restart your container
3. Missing sse_starlette
   • Update the requirements.txt and use docker compose up -d --build to rebuild and restart your container

Be sure to update the requirements.txt rather than use docker exec -it {containerID} /bin/bash to install using pip install x as I found this just didn’t work.

The final requirements additions I had were adding the below list of python packages (minus all the specific git installs) at the top of my requirements.txt:

accelerate==0.24.*
colorama
datasets
einops
exllamav2==0.0.7; platform_system != "Darwin" and platform_machine != "x86_64"
gradio==3.50.*
markdown
numpy==1.24.*
optimum==1.13.1
pandas
peft==0.5.*
Pillow>=9.5.0
pyyaml
requests
safetensors==0.4.0
scipy
sentencepiece
sentence_transformers
sse_starlette
tiktoken
tensorboard
transformers==4.35.*
tqdm
wandb

If you are running a Mac or doing your development on Linux, CodeGPT can manage your Llama based models itself and you don’t need the above setup.

Integrating CodeGPT with JetBrains

The CodeGPT extension enhances your JetBrains IDEs with AI capabilities:

Installation

Download the CodeGPT from the JetBrains Marketplace.

Configuration

CodeGPT supports OpenAI but allows for local service configuration. This feature is especially useful if you’re running models on a separate machine.

You’ll want to override the Base host and provide a dummy API key. Neither the model selected or the API are used when using text-generation-webui, since that controls which model is loaded, and by default doesn’t have any authentication required.

For Linux and Mac Users

On Linux or Mac, CodeGPT can manage and install Llama-based models via Llama.Cpp, which might eliminate the need for text-generation-webui.

This setup is ideal for Mac users with M1, M2, or M3 CPUs and at least 24GB of RAM since even a 7B model quantized will take up ~4GB of memory.

Model Compatibility

Note that Mistral and similar models are not Llama-based and might not work with this setup. However, smaller Llama models are still effective coding assistants.