LLaVA-server

Serves LLaVA inference using an HTTP server. Supports batched inference and caches the embeddings for each image in order to produce multiple responses per image more efficiently.

Usage

gunicorn "app:create_app()"

You must modify gunicorn.conf.py to change the number of GPUs.