Releases: awslabs/LISA
Releases · awslabs/LISA
v3.1.0
Enhancements
Model Management Administration
- Supports customers updating a subset of model properties through the model management user interface (UI) or APIs
- These new model management features are also limited to users in the configured IDP LISA administration group
- This feature prevents customers from having to delete and re-create models every time they want to make changes to available models already deployed in the infrastructure
Other Enhancements
- Updated the chat UI to pull available models from the model management APIs instead of LiteLLM. This will allow the UI to pull all metadata that is stored about a model to properly enable/disable features, current model status is used to ensure users can only interact with
InServicemodels when chatting - Updated default Model Creation values, so that there are fewer fields that should need updating when creating a model through the UI
- Removed the unnecessary fields for ECS config in the properties file. LISA will be able to go and pull the weights with these optional values and if an internet connection is available
- Added the deployed LISA version in the UI profile dropdown so users understand what version of the software they are using
Bug fixes
- Updated naming prefixes if they are populated to prevent potential name clashes, customers can now more easily use prefix resource names with LISA
- Fixed an issue where a hard reload was not pulling in the latest models
- Resolved a deployment issue where the SSM deployment parameter was being retained
- Addressed an issue where users could interact with the chat API if a request was being processed by hitting the
Enterkey
Coming Soon
- Version 3.2.0 will simplify the deployment process by removing all but the key properties required for the deployment, and extracting constants into a separate file as optional items to override. This will make LISA's deployment process a lot easier to understand and manage.
Acknowledgements
Full Changelog: v3.0.1...v3.1.0
Contributors
dustins, petermuller, and estohlmann
Assets 2
v3.0.1
Bug fixes
- Updated our Lambda admin validation to work for no-auth if user has the admin secret token. This applies to model management APIs.
- State machine for create model was not reporting failed status
- Delete state machine could not delete models that weren't stored in LiteLLM DB
Enhancements
- Added units to the create model wizard to help with clarity
- Increased default timeouts to 10 minutes to enable large documentation processing without errors
- Updated ALB and Target group names to be lower cased by default to prevent networking issues
Coming Soon
- 3.1.0 will expand support for model management. Administrators will be able to modify, activate, and deactivate models through the UI or APIs. The following release we will continue to ease deployment steps for customers through a new deployment wizard and updated documentation.
Acknowledgements
Full Changelog: v3.0.0...v3.0.1
Contributors
dustins, petermuller, and estohlmann
Assets 2
v3.0.0
Key Features
Model Management Administration
- Supports customers creating and deleting models through a new model management user interface (UI), or APIs
- Our new Model Management access limits these privileges to users in the configured IDP LISA administration group
- This feature prevents customers from having to re-deploy every time they want to add or remove available models
Note
- These changes will require a redeployment of LISA
- Take note of your configuration file and the models you have previously configured. Upon deployment of LISA 3.0 these models will be deleted and will need to be added back via the new model management APIs or UI
- You can see breaking changes with migrating from 2.0 -> 3.0 in the README
Enhancements
- Updated our documentation to include more details and to account for model management
Coming Soon
- 3.0.1 will expand support for model management. Administrators will be able to modify, activate, and deactivate models through the UI or APIs. The following release we will continue to ease deployment steps for customers through a new deployment wizard and updated documentation.
Acknowledgements
Full Changelog: v2.0.1...v3.0.0
Contributors
dustins, petermuller, and 4 other contributors
Assets 2
v2.0.1
What's Changed
Since the release of v2.0.0 we have made a few bugfixes, added greater model support, and added user experience changes to help reduce the friction in sending chat prompts to a model in the Chat UI. The following are the highlights of changes since v2.0.0.
- Fixed support for self-signed certs across RAG and LISA Serve features
- Added support for vLLM embedding models
- Simplified API Token usage for those users who already use OpenAI clients. README and example notebook have been updated to reflect this change
- Fixed default values in the Chat UI so that TGI models are less likely to throw errors if the model kwargs are not modified before the first prompt
- Various dependency version bumps for security fixes
- Formally deprecated v1 routes. See the README for migrating to the v2 routes
v2.0.0
What's Changed
The LISA v2.0.0 release boasts wider support for the OpenAI spec for users making queries directly against the LISA Serve endpoint. The following is a list of changes that we have introduced since v1.2.0
- LiteLLM added as a model proxy in the REST API container. LiteLLM provides LISA's OpenAI API spec support
- SageMaker Endpoint and Bedrock Foundation Model support added via LiteLLM configuration
- Added support for vLLM containers in addition to the existing support for TGI and TEI containers
- Chat UI updates to support LISA's OpenAI API support
- Documentation updates
- Added example Jupyter notebook showing the OpenAI compatibility with the LISA Serve endpoint
- Deprecated v1 endpoints as they will be removed in a future release
v1.2.0
What's Changed
- Fix issue with deprecated FastAPI lifecycle events.
- Update model prep scripts to support huggingface login.
- Limited support for OpenAI API spec for chat completion as well as support for authenticating via tokens which are managed in dynamodb
- Remove the workaround of using s5cmd in ADC regions as mounts3 is now available in aws-iso and aws-isob partitions
- Minor UX improvements and bug fixes for the example chat bot
- Recontextualize question before searching rag so that it works better with follow-up questions ie asking 'What is MLSpace?' and then following it up with 'Why would I use it?' will recontextualize the second question to be 'Why would I use MLSpace?' and search the rag repository based on embedding the contextualized question as opposed to the raw question.
- Fixes an issue with chat context missing the most recent answer due to react render cycles and the way we were rebuilding the chain
- Moved RAG options to a separate component
- Moved file upload modals to a separate component and did a bit of refactoring/adding additional types
- Removed some unused dependencies
- Added CDK unit tests baseline
Full Changelog: https://github.com/awslabs/LISA/commits/v1.2.0