Skip to content

Latest commit

 

History

History
106 lines (69 loc) · 4.2 KB

CHANGELOG.md

File metadata and controls

106 lines (69 loc) · 4.2 KB
Raw

v0.6.0 [Unreleased]

Features

  • Add distribution support (running completion / chat model on different process / machine).
  • Add conversation history in chat playground.
  • Add /metrics endpoint for prometheus metrics collection.

Fixes and Improvements

  • Fix the slow repository indexing due to constraint memory arena in tantivy index writer.
  • Make --model optional, so users can create a chat only instance.
  • Add --parallelism to control the throughput and VRAM usage: TabbyML#727

v0.5.5 (11/09/2023)

Fixes and Improvements

Notice

  • llama.cpp backend (CPU, Metal) now requires a redownload of gguf model due to upstream format changes: TabbyML#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: TabbyML#638
  • add server.completion_timeout to control the code completion interface timeout: TabbyML#637
  • Cuda backend is switched to llama.cpp: TabbyML#656
  • Tokenizer implementation is switched to llama.cpp, so tabby no longer need to download additional tokenizer file: TabbyML#683
  • Fix deadlock issue reported in TabbyML#718

v0.4.0 (10/24/2023)

Features

  • Supports golang: TabbyML#553
  • Supports ruby: TabbyML#597
  • Supports using local directory for Repository.git_url: use file:///path/to/repo to specify a local directory.
  • A new UI design for webserver.

Fixes and Improvements

  • Improve snippets retrieval by dedup candidates to existing content + snippets: TabbyML#582

v0.3.1 (10/21/2023)

Fixes and improvements

v0.3.0 (10/13/2023)

Features

Retrieval-Augmented Code Completion Enabled by Default

The currently supported languages are:

  • Rust
  • Python
  • JavaScript / JSX
  • TypeScript / TSX

A blog series detailing the technical aspects of Retrieval-Augmented Code Completion will be published soon. Stay tuned!

Fixes and Improvements

  • Fix Issue #511 by marking ggml models as optional.
  • Improve stop words handling by combining RegexSet into Regex for efficiency.

v0.2.2 (10/09/2023)

Fixes and improvements

  • Fix a critical issue that might cause request dead locking in ctranslate2 backend (when loading is heavy)

v0.2.1 (10/03/2023)

Features

Chat Model & Web Interface

We have introduced a new argument, --chat-model, which allows you to specify the model for the chat playground located at http://localhost:8080/playground

To utilize this feature, use the following command in the terminal:

tabby serve --device metal --model TabbyML/StarCoder-1B --chat-model TabbyML/Mistral-7B

ModelScope Model Registry

Mainland Chinese users have been facing challenges accessing Hugging Face due to various reasons. The Tabby team is actively working to address this issue by mirroring models to a hosting provider in mainland China called modelscope.cn.

# Download from the Modelscope registry
TABBY_REGISTRY=modelscope tabby download --model TabbyML/WizardCoder-1B

Fixes and improvements

  • Implemented more accurate UTF-8 incremental decoding in the GitHub pull request.
  • Fixed the stop words implementation by utilizing RegexSet to isolate the stop word group.
  • Improved model downloading logic; now Tabby will attempt to fetch the latest model version if there's a remote change, and the local cache key becomes stale.
  • set default num_replicas_per_device for ctranslate2 backend to increase parallelism.