v0.0.3 Performance, Stability, Layout enhancements

This release features enhancements related to performance, stability, and layout.

• Fix crashes with the LLaMA backend related to model loading.
• Force directed layout to help with the node clutter.
• Ui enhancements like the ability to rename stories and a trash for deleted stories.
• App Icon is now fixed.

Full Changelog: v0.0.2...v0.0.3

Note: Linux build coming soon

v0.0.2 Settings Enhancements

This release exposes settings for drama_llama and fixes some crashes.

• (almost all) drama_llama settings are available in the settings tab. This includes the ability to mix and match any combination of sampling methods to your liking, N-gram based repetition penalties (not just unigrams). Left to add are: A UI for custom stop criteria as well as ignored n-grams for the purpose of repetition penalty. That will come next release. If you need custom stop strings, you can use the previous release to add them. The settings are backwards and forwards compatible. It was removed because I simply didn't have time to add them to drama_llama today.
• Fixed some crashes in the drama_llama sampling code.
• Improved defaults for OpenAI

v0.0.1 Generation Quality Improvements

This release improves quality of generation by:

• Changing the defaults sampling settings for LLaMA from greedy to locally typical sampling.
• Updating drama_llama and llama.cpp for BPE tokenizer changes. This will require regenerating any models. See the linked issue for scripts.
• Changing OpenAI sampling settings to better suit story generation.

Known Issues:

• This release has some crashes that are fixed in v0.0.2

v0.0.0

This is the first release of weave. It has only basic features but it supports:

Inference backends:

• Metal inference on MacOS.
• CUDA inference on Linux.
• OpenAI (all models supporting the chat completions API, including GPT-4o)

Settings:

• Settings for OpenAI are available in the GUI including sampling settings, system prompt, and an initial set of example messages to bootstrap the story writing.
• Settings for drama_llama are not yet available in the GUI. Right now they default to whatever the crate default for drama_llama is. As of writing that's greedy sampling with no repetition penalty. That's not great, so the defaults will be changed shortly and a GUI added.

Notes:

• The MacOS release is tested, including with LLaMA 3 70b, but the binary is not signed or notarized so it MacOS will complain until you tell Gatekeeper to allow it to run in Settings. Future releases will likely be signed and notarized so it will "just work".
• The Linux release is completely untested. I don't have Linux+NVIDIA machine that isn't headless available, so if you do, feel free to test it and file and issue if it doesn't work. It will likely also work via X11 forwarding if you have that setup. It will also likely work in WSL.

0.0.0 Prerelease

This is a prelease of Weave. It's been built locally (until I can get CI to work on all platforms).

Notes:

• Both Linux apps are untested
• The Mac App is signed but not notarized (yet) so MacOS will screech at you if you try to run it until you tell Gatekeeper to STFU in settings.
• All apps use the keyring crate to store and retrieve the OpenAI API key securely so you may be asked to authenticate the first time you run the app.
• Windows app is coming and this release will be updated when it's ready