Skip to main content

LLM-Definitions

This documentation page contains information about the folder LLM-Definitions, which in turn contains information on how to add LLMs to the repository in the SAS Model Manager. Each model is packaged so that it can be deployed using the SAS Container Runtime (SCR). More on the SCR in the SAS Documentation.

Each subfolder there contains the definition for one specific LLM - the name of the folder specifies the LLM.

Tags

Tags are being used to provide additional information and filtering options around the LLMs inside of SAS Model Manager. Below you'll find a table with short description and its impact (if any):

TagDescriptionImpact
PythonIndicates that the model is implemented in PythonThis is required as the whole building process is setup around Python.
Open-SourceIndicates that the model has an open-source license-
ProprietaryIndicates that the model is proprietary-
deprecatedIndicates that the models is no longer supportedThe model will not show up in the Prompt Builder UI
smallIndicates the required resources for serving this model is smallCan be used for when publishing to SCR as a sizing indication
mediumIndicates the required resources for serving this model is mediumCan be used for when publishing to SCR as a sizing indication
largeIndicates the required resources for serving this model is largeCan be used for when publishing to SCR as a sizing indication
LLMIndicates that the model has more than 7 billion parameters-
SLMIndicates that the model has less than or equal to 7 billion parameters-

There is a lot more tags available like MIT-License, Apache-2, Google, etc. these are used to showcase the specific model license and the model provider but have no further impact and new once are added as the market evolves.

Models that require the Hugging Face token

Here is a list of models in this repository that are gated on Hugging Face and thus require you to first accept a license - this is sometimes also related to a waiting time until you are confirmed for access.

Model NameModel ProviderHugging Face LinkNote
Llama 3.1 405BMetahttps://huggingface.co/meta-llama/Llama-3.1-405B-InstructIt is recommended to use a hosting provider, instead of hosting it yourself.
Llama 3.2 1BMetahttps://huggingface.co/meta-llama/Llama-3.2-1B-Instruct
Llama 3.2 3BMetahttps://huggingface.co/meta-llama/Llama-3.2-3B-Instruct
Llama 3.3 70BMetahttps://huggingface.co/meta-llama/Llama-3.3-70B-InstructIt is recommended to use a hosting provider, instead of hosting it yourself.
Mistral NemoMistralhttps://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407While it runs on just CPU a hosting provider is recommended.