1. Synced codebase to recent llama.cpp. 2. Added support for gemma-3 and mistral_small_2504 and other older LLMs. (Image input support is coming next). 3. New param `--n_keep 128` (to keep n tokens from the beginning when shifting context with context shift. 4. `-fa` (flash attention) Some bugs may still exist. More testing is needed. Current release is for avx2 CPU's. For older CPU's without AVX2 - wait or use previous versions.

1. Added support for gemma-2 and mistral-nemo. 2. Added multiple gpu support. Don't set those 3 params if you have just 1 gpu. `--main-gpu 0` - set main gpu id with kv-cache: 0, 1, ... `--split-mode none` - `none` or `layer`. split-mode tensor is not supported `--tensor-split 0.5,0.5` - how to split layers or tensors per gpus, array of floats. 3. Added instruct mode with presets. It is optional and experimental. There are still some bugs. `--instruct-preset gemma` where gemma is the name of the file `\instruct_presets\gemma.json` Instruct mode helps to make responses longer and smarter. You can find correct instruct-preset for each model at the model card on huggingface or in sillytavern - formatting - instruct mode sequences. Example dialogue in assisttant.txt should also be formatted using instruct mode tags. I added gemma and mistral instruct presets. And added some bats to run gemma and nemo in instruct mode. 4. Added `-debug` to print whole context dialogue after each LLM response. Useful to see if there's something wrong with formatting.

Added `--min_p` sampler param. I recommend using `--min_p 0.10` for Russian.

- Added `--push-to-talk` option: hold "Alt" key to speak (useful with loudspeakers without headphones). Turned off by default. - And now you can use Cyrillic letters in bat files. Save them using Cyrillic "OEM 866" encoding, notepad++ supports it. (В bat файлах теперь можно использовать кириллицу. Для этого сохраните bat файл в кодировке "OEM 866" в приложении notepad++: Encoding -> Character sets -> Cyrillic -> OEM 866).

-bug fix with start prompt: start prompt was not written correctly into context when running with default `--batch-size 64` parameter or without it. Llama couldn't remember anything from the start prompt (just first 64 tokens). This bug first came in v0.1.3 and no one noticed.

New features: - Keyboard input (finally you can type messages using keyboard now). - Hotkeys: Stop(Ctrl+Space), Regenerate(Ctrl+Right), Delete(Ctrl+Delete), Reset(Ctrl+R). - Bug fix: `Reset` command works fine now, no bugs with long context. New bugs: - Sometimes when you type fast, the first letter of your message is not typed. (Type a little slower. I have to investigate more, what is causing this bug).

- removed --xtts-control-path param. Now it is not needed anymore. xtts control file is stored in temp dir. No need to set xtts_play_allowed_path it in extras. - added missing default.wav voice No other changes. To make this version work - please update (git pull) xtts_api_server, tts, and wav2lip if you have older versions.

Now it's using 2 condas to install updated bats