Skip to content

v0.7.10
Compare
Choose a tag to compare
@Vali-98 Vali-98 released this 27 Aug 13:46
· 344 commits to master since this release
v0.7.10
902bf75

v0.7.10

This is a cummulative update for all recent changes from 0.7.9a-0.7.9g with a few additions.

Features:

  • Added the Gemma 2 instruct template to defaults
  • Added DRY Sampler to koboldcpp
  • Added XTC sampler to local inferencing
  • Added support for Gemma 2, Nemotron, Minitron and Minitron-Width to local inferencing
  • Added new supported API's: Generic Text Completions, Cohere
  • Added a default card; AI Bot

Changes:

  • Changed local inferencing to properly use tokenizer from currently used model
  • Changed icons to allow for adaptive and monochrome colors on Android
  • Default app hue to blue
  • Changed chat editor modal buttons to be bigger and less finicky to press
  • Changed several default Sampler options to be disabled, this was causing deterministic outputs on specific APIs
  • top_k can now be set to 0, which disables it for local inferencing and several other APIs (note, some APIs will not accept 0 here)

Fixes:

  • Fixed instruct suffixes never being added
  • Fixed Tokenizer calculation being very inaccurate, causing many issues with overflowing local models, causing context to fallback to halving the cache which also leads to massive reprocessing times. This also fixes issues with context shifitng failing.
  • Example messages should never be readded once at least the first message has been shifted out of context: this fixes issues with examples being readded causing massive reprocessing for local models.
  • Issues with TTS Engine when using custom sherpa-onnx models
  • Fixed specific popups to properly shift with keyboard instead of remaining in place
  • Fixed initial default preset always being broken

Dev:

  • Massive refactors of several screens, though rerendering is pretty minimal, this further improves it by sectioning more components and behaviors, as well as reducing zustand selectors in chat items.
  • Added sqlite-vec library to expo-sqlite in preparation for future embedding models
  • Added fixes to cui-llama.rn for embedding support
  • Finally removed a lot of deprecated screens and todos.
  • Changed file structure to fix broken interactions between zustand and Fast Refresh. If you are adding a new component, be sure it is contained within the /app directory.
Assets 3
Loading