deka-ai-coding/registry-tabby

Tabby Registry

This repository uses ollama library to download models. You must add TABBY_DOWNLOAD_HOST=registry.ollama.ai in tabby's environment in order to use this registry.

Passive maintenance notice

From Tabby 0.12 onward (starting with commit), it is possible to connect your Tabby instance to Ollama. Furthermore, you can enable automatic model downloading with the TABBY_OLLAMA_ALLOW_PULL=1 environment variable. Ollama will handle chat templates and stop words for you; however, you still need to set a prompt template for FIM. If you want to use models from Ollama, consider deploying Ollama (as a sidecar, for example) and connecting Tabby to it.

With that in mind, I am moving this repository into passive maintenance. You can open issues and pull requests to add new models. While I will review them, I will not actively develop this repository further. It's possible that I will add new FIM models solely to store prompt templates for them.

Warning

The model list is auto generated. Some models may not work properly.

If a model haves instruct in tag name you should prefer use it. Chat templates are set for instruct models and can work incorrectly with others.

Models

This repository generates a registry for Tabby but uses ollama library instead a official manual filled registry. The models in registry must be included in allow list: meta/models.yaml.

Model naming

All models named using this pattern:

<name>:<tag>

Where:

• <name> - name of model
• <tag> - tag as specified by ollama library. Note: it is required

Example:

deepseek-coder:6.7b-base-q8_0
^              ^
Name           Tag

Current models

Name	Chat?	Completion?	Comments
starcoder	No	Yes
starcoder2	Yes	Yes
deepseek-coder	Yes	Yes
gemma	Yes	No	Requires Tabby 0.10+ or after commit
codegemma	Yes	Yes	Requires Tabby 0.10+ or after commit
mistral	Yes	No
mixtral	Yes	No
wizardcoder	Yes	No
codellama	Yes	Yes	Prefer code tags for completion, instruct for chat
openhermes	Yes	No
nous-hermes2	Yes	No
nous-hermes2-mixtral	Yes	No
codeqwen	Yes	Yes	chat for chat, code for completion, others for chat
llama3	Yes	No	Unsupported stop words, so looping will happen
llama3-chatqa	Yes	No	Short answers, does not use code blocks even if you ask to do so
codestral	Yes	Yes
deepseek-coder-v2	Yes	Yes	base for completion, instruct for chat

registry-tabby's Issues

The gemma chat broken and panics

New chat template seems broken. It will be reverted when I have a time. Do not use gemma models for now.

I want add DeepseekCoder-v2-15.7B-instruct model , any tips or guide?

I always use your model registry, thank you !

but now , I want add DeepseekCoder-v2-15.7B-instruct model , any tips or guide?