local-llm-server/README.md

# local-llm-server

_An HTTP API to serve local LLM Models._

The purpose of this server is to abstract your LLM backend from your frontend API. This enables you to make changes to (or even switch) your backend without affecting your clients.

### Install

1. `sudo apt install redis`
2. `python3 -m venv venv`
3. `source venv/bin/activate`
4. `pip install -r requirements.txt`
5. `wget https://git.evulid.cc/attachments/89c87201-58b1-4e28-b8fd-d0b323c810c4 -O /tmp/vllm_gptq-0.1.3-py3-none-any.whl && pip install /tmp/vllm_gptq-0.1.3-py3-none-any.whl && rm /tmp/vllm_gptq-0.1.3-py3-none-any.whl`
6. `python3 server.py`

An example systemctl service file is provided in `other/local-llm.service`.

### Configure

First, set up your LLM backend. Currently, only [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) is supported, but
eventually [huggingface/text-generation-inference](https://github.com/huggingface/text-generation-inference) will be the default.

Then, configure this server. The config file is located at `config/config.yml.sample` so copy it to `config/config.yml`.

1. Set `backend_url` to the base API URL of your backend.
2. Set `token_limit` to the configured token limit of the backend. This number is shown to clients and on the home page.

To set up token auth, add rows to the `token_auth` table in the SQLite database.

`token`: the token/password.

`type`: the type of token. Currently unused (maybe for a future web interface?) but required.

`priority`: the lower this value, the higher the priority. Higher priority tokens are bumped up in the queue line.

`uses`: how many responses this token has generated. Leave empty.

`max_uses`: how many responses this token is allowed to generate. Leave empty to leave unrestricted.

`expire`: UNIX timestamp of when this token expires and is not longer valid.

`disabled`: mark the token as disabled.

### Use

**DO NOT** lose your database. It's used for calculating the estimated wait time based on average TPS and response tokens and if you lose those stats your numbers will be inaccurate until the database fills back up again. If you change GPUs, you
should probably clear the `generation_time` time column in the `prompts` table.

### To Do

- Implement streaming
- Add `huggingface/text-generation-inference`
- Convince Oobabooga to implement concurrent generation
- Make sure stats work when starting from an empty database
- Make sure we're correctly canceling requests when the client cancels
- Implement auth and tokens on the websocket endpoint. Maybe add something to the instruct prompt and the remove it before proxying??