- Llama 2 models, including Llama 2 70B, are now fully supported
- Updated to latest text-generation-webui
requirements.txt - Removed the exllama pip package installed by text-generation-webui
- Therefore the ExLlama kernel will build automatically on first use
- This ensures that ExLlama is always up-to-date with any new ExLlama commits (which are pulled automatically on each boot)
- Added simple build script for building the Docker containers
- Updated to latest ExLlama code, fixing issue with SuperHOT GPTQs
- ExLlama now automaticaly updates on boot, like text-generation-webui already did
- This should result in the template automatically supporting new ExLlama features in future
- Major update to the template
- text-generation-webui is now integrated with:
- AutoGPTQ with support for all Runpod GPU types
- ExLlama, turbo-charged Llama GPTQ engine - performs 2x faster than AutoGPTQ (Llama 4bit GPTQs only)
- CUDA-accelerated GGML support, with support for all Runpod systems and GPUs.
- All text-generation-webui extensions are included and supported (Chat, SuperBooga, Whisper, etc).
- text-generation-webui is always up-to-date with the latest code and features.
- Automatic model download and loading via environment variable
MODEL. - Pass text-generation-webui parameters via environment variable
UI_ARGS.