It's been a long time! And I'm excited to bring you a large update with significant improvements in many areas!
Long-term memory has been updated to V2!
This update replaces the LTM model with a much more efficient built-in summariser the constructs your memories 10x as fast! Previously, each memory ingestion could take anywhere from 1-5 minutes, now it takes around 5 seconds!
Your characters will now also utilise their memory and form an impression of you after chatting with them for a while. You can view their impression of you in your Profile page.
Re-designed Voice Chat
Hands-free mode has been completely re-designed as a new app: Voice Chat. It now works in a similar way as video chatting with your characters. Combine this with the new animated character backgrounds feature to bring your characters to life.
Offline Translator
Works similar to Google Translate, but works completely offline!
It can also translate your chats with LLM in real-time, so even if the model doesn't support multi-lingual, it does now with Layla!
Executorch Update
Significant updates to the Executorch backend, further stabilising and integrating it into Layla:
Executorch now supports edit & regen during chatting
This is made possible by the new KV structure in the compiled models.
IMPORTANT: To use the new Executorch edit & regen features, you must re-download your executorch models from my repo: https://huggingface.co/l3utterfly
Full changelog
New features:
Offline translator: similar to Google Translate but works completely offline!
Long-term memory update: your characters will now form an impression of you after chatting for a while, you can view this in your Profile page
LTM updated to v2! Improves conversation summary processing times significantly (2min -> 5s)
Redesigned Hands-free Mode: now moved to a dedicated app called "Voice Chat"
Added high quality TTS models: Piper High
Supports the latest Rope scaling implementation of Phi-3 from llama.cpp
Layla now supports custom animated character backgrounds and expressions!
Improvements:
added ability to view injected memories for your character during chatting
Layla will now self-diagnose and fix memory corruptions in Diagnostics mode
added actions to delete all unread messages or mark all messages as read
improved LTM ingestion for Llama3 models (pre-built Layla models still work much better because they are trained to take into account of memories)
added ability to exclude certain characters from LTM ingestion
added ability to choose BART or LLM ingestion for long-term memory
Executorch models now supports edit & regen!
Executorch now supports background inference!
added setting to toggle whether or not your characters recall their last memory when starting a new chat
clear chat history now shows a warning before proceeding
added ability edit memories during chatting
added ability to search voices when selecting
added better support for Native TTS voices
Bug fixes:
fixed bug where PDF tasks are not deleted when the PDF Digest app is uninstalled
fixed some stability issues when you have a lot of long-term memory tasks queued
fixed bug in Layla Cloud skipping random words during the message
fixed bug where Character Expression app uses too much memory
fixed bug where tool use was incorrectly using the new sentiment detector
fixed bug where Layla was outputting the EOS token
fixed bug in deleting all memories for a particular character
fixed related bugs that causes "failed to eval" error
fixed bug where PDF Digest seems to be queuing up infinite tasks
fixed bug where Executorch models were not using the full context length they were compiled with
fixed bug where character images were not deleted when you delete your characters
fixed bug where you cannot select the default Android/iOS voices after installing Layla's voices
Comments