What's Changed

• Add llama-2-70b GGML support by @oobabooga in #3285
• Bump bitsandbytes to 0.41.0 by @jllllll in #3258 -- faster speeds
• Bump exllama module to 0.0.8 by @jllllll in #3256 -- expanded LoRA support

Bug fixes

• Add checks for ROCm and unsupported architectures to llama_cpp_cuda loading by @jllllll in #3225

Extensions

• [extensions/openai] Fixes for: embeddings, tokens, better errors. +Docs update, +Images, +logit_bias/logprobs, +more. by @matatonic in #3122

Changes

• Add missing EOS and BOS tokens to Llama-2 template
• Bump transformers for better Llama-2 support
• Bump llama-cpp-python for better unicode support (untested)

Changes

• Llama-v2: add instruction template, autodetect the truncation length, add conversion documentation
• [GGML] Support for customizable RoPE by @randoentity in #3083
• Optimize llamacpp_hf (a bit)
• Add Airoboros-v1.2 template
• Disable "Autoload the model" by default
• Disable auto-loading at startup when only one model is available by @jllllll in #3187
• Don't unset the LoRA menu when loading a model
• Bump accelerate to 0.21.0
• Bump bitsandbytes to 0.40.2 (Windows wheels provided by @jllllll in #3186)
• Bump AutoGPTQ to 0.3.0 (loading LoRAs is now supported out of the box)
• Update LLaMA-v1 documentation

Bug fixes

• Use 'torch.backends.mps.is_available' to check if mps is supported by @appe233 in #3164

Changes

• Create llamacpp_HF loader by @oobabooga in #3062
• Make it possible to evaluate exllama perplexity by @oobabooga in #3138
• Add support for logits processors in extensions by @cyberfox in #3029
• Bump bitsandbytes to 0.40.1.post1 by @jllllll in #3156
• Bump llama cpp version by @ofirkris in #3160
• Increase alpha value limit for NTK RoPE scaling for exllama/exllama_HF by @Panchovix in #3149
• Decrease download timeout

Bug fixes

• Fix reload screen background color in dark mode

Extensions

• Color tokens by probability and/or perplexity by @SeanScripts in #3078

Bug fixes

• Fix output path when downloading models through the UI

Changes

• Bump bitsandbytes Windows wheel by @jllllll in #3097 -- --load-in-4bit is now a lot faster
• Add support low vram mode on llama.cpp module by @gabriel-pena in #3076
• Add links/reference to new multimodal instructblip-pipeline in multimodal readme by @kjerk in #2947
• Add token authorization for downloading model by @fahadh4ilyas in #3067
• Add default environment variable values to docker compose file by @Josh-XT in #3102
• models/config.yaml: +platypus/gplatty, +longchat, +vicuna-33b, +Redmond-Hermes-Coder, +wizardcoder, +more by @matatonic in #2928
• Add context_instruct to API. Load default model instruction template … by @atriantafy in #2688
• Chat history download creates more detailed file names by @UnskilledWolf in #3051
• Disable wandb remote HTTP requests
• Add Feature to Log Sample of Training Dataset for Inspection by @practicaldreamer in #1711
• Add ability to load all text files from a subdirectory for training by @kizinfo in #1997
• Add Tensorboard/Weights and biases integration for training by @kabachuha in #2624
• Fix: Fixed the tokenization process of a raw dataset and improved its efficiency by @Nan-Do in #3035
• More robust and error prone training by @FartyPants in #3058

Bug fixes

• [Fixed] wbits and groupsize values from model not shown by @set-soft in #2977
• Fix API example for loading models by @vadi2 in #3101
• google flan T5 tokenizer download fix by @FartyPants in #3080
• Changed FormComponent to IOComponent by @ricardopinto in #3017
• respect model dir for downloads by @micsthepick in #3079

Extensions

• Fix send_pictures extension
• Elevenlabs tts fixes by @set-soft in #2959
• [extensions/openai]: Major openai extension updates & fixes by @matatonic in #3049
• substitu superboog Beatiful Soup Parser by @juhenriquez in #2996

Let's call it "version 1.0".

1-click installers for Windows, Linux, MacOS, and WSL. Just download the zip, extract it, and double click on "start". The web UI and all its dependencies will be installed in the same folder.

The source codes and more information can be found at: https://github.com/oobabooga/one-click-installers

This is now obsolete! The one-click installers have been merged into the repository.