v0.5.5

⚠️ Notice

• llama.cpp backend (CPU, Metal) now requires a redownload of gguf model due to upstream format changes: #645 ggerganov/llama.cpp#3252
• Due to indexing format changes, the ~/.tabby/index needs to be manually removed before any further runs of tabby scheduler.
• TABBY_REGISTRY is replaced with TABBY_DOWNLOAD_HOST for the github based registry implementation.

🚀 Features

• Improved dashboard UI.

🧰 Fixes and Improvements

• Cpu backend is switched to llama.cpp: #638
• add server.completion_timeout to control the code completion interface timeout: #637
• Cuda backend is switched to llama.cpp: #656
• Tokenizer implementation is switched to llama.cpp, so tabby no longer need to download additional tokenizer file: #683

💫 New Contributors

• @CrCs2O4 made their first contribution in #597
• @yusiwen made their first contribution in #620
• @gjedeer made their first contribution in #635
• @XpycT made their first contribution in #634
• @HKABIG made their first contribution in #662

Full Changelog: v0.4.0...v0.5.5