Skip to content

README.md

Latest commit

 

History

History
32 lines (23 loc) · 1.38 KB

README.md

File metadata and controls

32 lines (23 loc) · 1.38 KB

VisDiff API Servers

Design Choices

  1. All LLMs/VLMs/CLIPs serve as API with cache enabled, because loading a LLM/VLM/CLIP is expensive and we never modify them.
  2. LLM functions in utils_llm.py, VLM functions in utils_vlm.py, CLIP functions in utils_clip.py, and others in utils_general.py.
  3. Write unit tests to understand major functions.

LLM Server Configuration

  1. Set up OpenAI API key: export OPENAI_API_KEY='[your key]'
  2. Pip install environments: pip install vllm
  3. Configure global variables in global_vars.py
  4. Run python -m vllm.entrypoints.openai.api_server --model lmsys/vicuna-7b-v1.5
  5. Run python -m serve.utils_llm to test the LLM.

CLIP Server Configuration

  1. Pip install environments: pip install open-clip-torch flask
  2. Configure global variables in global_vars.py
  3. Run python serve/clip_server.py
  4. Run python -m serve.utils_clip to test the CLIP.

VLM Server Configuration

  1. Install environments:
  • BLIP: pip install salesforce-lavis
  • LLaVA: git clone git@github.com:haotian-liu/LLaVA.git; cd LLaVA; pip install -e .
  1. Configure global variables in global_vars.py
  2. Run python serve/vlm_server_[vlm].py. It takes a while to load the VLM, especially the first time to download the VLM. (Note: concurrency is disabled as it surprisingly leads to worse GPU utilization)
  3. Run python -m serve.utils_vlm to test the VLM.