As a public community project the KitOps roadmap is always open to feedback. If you think a feature lower in the list is more important, or feel that we've missed an important feature please:
- Open an issue in the repo
- Tell us in the KitOps Discord channel
- Send us a message on X / Twitter
The main things we're working on are:
- CLI dev mode to simplify running LLMs locally for testing or app integrations
- Tutorials for using KitOps with LLM fine-tuning and RAG pipelines
- Attestation for ModelKits and their contents
- Signing of ModelKits
- Model chaining (foundational --> tuned)
- Deployment
- More CLI distribution options
- Documentation of the KitOps workflow
Users can run a ModelKit packaged LLM on localhost and interact with it via chat or prompt. Uses Llama CPP with a cleaned up and more user-friendly UI. Simplifies running, testing, and integratting LLMs with applications.
- Design UX for chat, prompt, and parameter interactions
- Code UI for interactions focused on desktop form factor
- Create ModelKits for popular LLMs
- Add
dev
command to Kit CLI on MacOS - Add
dev
command to Kit CLI on Windows - Add
dev
command to Kit CLI on Linux
Users can learn how to use ModelKits as part of a RAG pipeline, or while fine-tuning their LLM. Helps people get the most value out of ModelKits during their LLM development work.
- Tutorial: Fine-tuning LLMs using KitOps
- Tutorial: Doing RAG with KitOps
This feature will come in two parts: build attestation and self-attestation. Build attestations will be done by KitOps itself, adding a SLSA attestation about how the ModelKit was built. ModelKit creators can add an optional attestation for specific assets in the ModelKit, or the whole ModelKit. Users can include any 3rd party attestation Verification Summary Attestation (VSA). Additional attestations could be added as predicates. Attestation would be included as a separate layer in the ModelKit.
- Add attestation for KitOps ModelKit builds
- Add mechanism for adding attestations to ModelKits
- Update CLI to make use of attestations
- Add CLI warnings if attestation was expected and not found
Users can optionally sign their ModelKit using something like Cosign in order to add an extra layer of security to their packaging.
- Decide on a signing utility
- Add signing to Kit CLI
A ModelKit can refer to another ModelKit's model as its "base model." This is useful for situations where a user wants to take an existing model (e.g., Llama 2) but create their own model based on it. The user would create a kind of link between their ModelKit and the ModelKit for Llama 2 that they used. Right now we think this will be done by having a base model and model parts. The chain could theoretically be as long as desired (not limited to a single relationship).
- Determine mechanism for ModelKit chaining
- Add chaining to ModelKit schema
- Add support for chaining to relevant Kit CLI commands
Users want to be able to deploy their models through existing CI/CD/CT pipelines. Since KitOps doesn't know the details of a user's deployment pipeline or process we will unpack
a ModelKit into an appropriate directory structure for one of several deployment targets.
Users want to be able to get the Kit CLI from locations like Brew, Choco, and Conda.
To help users understand how ModelKits fit into their existing workflows we will create documentation explaining how CI/CD/CT, observability, and other tasks can be simplified with KitOps.