vLLM

Setting up vLLM

1. Download + install vLLM
2. Launch a vLLM OpenAI-compatible API server using the official vLLM documentation

For example, if we want to use the model dolphin-2.2.1-mistral-7b from HuggingFace, we would run:

1
python -m vllm.entrypoints.openai.api_server \
2
--model ehartford/dolphin-2.2.1-mistral-7b

vLLM will automatically download the model (if it’s not already downloaded) and store it in your HuggingFace cache directory.

Enabling vLLM as a provider

To enable the vLLM provider, you must set the VLLM_API_BASE environment variable. When this is set, Letta will use available LLM and embedding models running on vLLM.

Using the docker compose server with vLLM

Since vLLM is running on the host network, you will need to use host.docker.internal to connect to the vLLM server instead of localhost. You’ll also want to make sure to open the port 8000 (the default port for vLLM) on your host machine.

$
# replace `~/.letta/.persist/pgdata` with wherever you want to store your agent data
>
docker run \
>
  -v ~/.letta/.persist/pgdata:/var/lib/postgresql/data \
>
  -p 8283:8283 \
>
  -p 8000:8000 \
>
  -e VLLM_API_BASE="http://host.docker.internal:8000" \
>
  letta/letta:latest

$
export VLLM_API_BASE="http://localhost:8000"
>
letta run

$
export VLLM_API_BASE="http://localhost:8000"
>
letta server