·
4 commits
to pytorch-transformers
since this release
- Various model input cleaning fixes
- Improved wordcloud
- Allows for disabling whisper_mic entirely (for sessions that don't have a window manage)
- Added a story illustrator (stable diffusion): ctrl+o to display
- Automatically enables cpu offloading when vRAM becomes insufficient
- Smart usage of KV caches to greatly reduce vRAM usage peaks, with a very small tradeoff in speed
- Added support for GGUF models (they're converted to pytorch when loaded)
- Gender detection for TTS speakers: you can customize the voice for 2 men, 2 women, 1 non-binary person, plus the narrator. (Same gender voices will alternate if their dialogues follow each other.)
- Removed hard cap of 6000 tokens input size, instead using the model's maximum, or 1.5x the sliding window's size for models that don't have a maximum (mistral).
Full Changelog: v2.4.2...v2.5.0