Model deprecations
SambaStudio provides a wide selection of the latest and most capable models. As new models are made available in the Model Hub, older models are retired. When we retire a model, it is known as a Deprecated model. Keep in mind the following regarding deprecated models.
-
SambaNova will never immediately deprecate a model. There will always be a deprecation process, typically two months, accompanied by corresponding notifications.
-
When a model is deprecated, we will suggest a suitable replacement model to use if one is available.
-
Identified replacement models will be made available during the deprecation window, which will allow you to test endpoint workflows using the replacement model.
-
Fine-tuned models that you have created using a model that is being deprecated will not be affected. You will be able to continue further fine-tuning on these models.
Deprecation process
When a model is being deprecated in SambaStudio, it will never immediately be removed from the platform. It goes through the deprecation process described below.
To be deprecated
Before a model will be deprecated, it will be identified in SambaStudio with the tag To be deprecated. Hovering over the tag will display the date the model will be deprecated and a suggested replacement model to use. The tag will appear in the following locations:
-
The model card of a model that will be deprecated.
-
Any Train job details page that uses the model that will be deprecated.
-
Any Endpoint window that uses the model that will be deprecated.
-
The model or checkpoint fields displayed when using the SambaNova API (snapi) CLI commands
snapi job infoorsnapi endpoint info.-
Example model field:
GPT_1.5B_GT_Finetuned(To be Deprecated)
-
The SambaStudio Notifications panel will display an alert describing that a model will be deprecated, along with the date the model will be deprecated and a suggested replacement model to use.
Deprecated
When a model has reached its deprecation date, it will no longer be accessible in the SambaStudio GUI and CLI. This includes the Model Hub, Train job creation, and Endpoint creation workflows. Keep in mind the following once a model has reached its deprecation date.
-
The SambaStudio Notifications panel will display an alert describing that the model has been deprecated, along with the date the model was deprecated and a suggested replacement model to use.
-
Any Train job details page that uses a deprecated model will display the tag Deprecated.
-
Any Endpoint window that uses a deprecated model will display the tag Deprecated.
-
Hovering over the Deprecated tag will display the date the model was deprecated and a suggested replacement model to use.
Deprecation list
The table below describes the models identified to be deprecated, their target dates, and suggested replacements.
December 22, 2024 deprecations
| ML App | Model Name | Deprecation Target Date | Suggested Replacement |
|---|---|---|---|
Dialog Act Classification |
GPT_1.5B_Dialog_Act_Classification_Finetuned |
December 22, 2024 |
|
Deepseek 6.7B single socket |
Deepseek_6.7B_Base Deepseek_6.7B_Instruct |
December 22, 2024 December 22, 2024 |
|
Llama 2 70B with dynamic batching |
Llama-2-70b-chat-hf-dynamic-batching |
December 22, 2024 |
|
Llama 2 with dynamic batching |
Llama-2-7b-hf-dynamic-batching Llama-2-13b-hf-dynamic-batching |
December 22, 2024 December 22, 2024 |
|
Named Entity Recognition |
GPT_1.5B_NER_Finetuned |
December 22, 2024 |