You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Different servers have options to specify a chat template, which essentially wraps the user prompt is model specific tags, so that the LLM can process and respond better.
An example of how the tags can impact the accuracy of the model can be found here.
We temporarily updated our user docs to include the tags, but it's pretty garbage to ask users to add those tags themselves.
As can be seen by the link above, SGLang defaults their server to use a ChatTemplate, based on the model that was specified at startup. We should do the same.
If a user runs a llama model, we should default to use the llama-3 chat template, unless otherwise specified at startup. As we add more chat models, we should add more ChatTemplates
The text was updated successfully, but these errors were encountered:
Discussion for this here.
Different servers have options to specify a chat template, which essentially wraps the user prompt is model specific tags, so that the LLM can process and respond better.
An example of how the tags can impact the accuracy of the model can be found here.
We temporarily updated our user docs to include the tags, but it's pretty garbage to ask users to add those tags themselves.
As can be seen by the link above, SGLang defaults their server to use a
ChatTemplate
, based on the model that was specified at startup. We should do the same.If a user runs a llama model, we should default to use the
llama-3
chat template, unless otherwise specified at startup. As we add more chat models, we should add moreChatTemplates
The text was updated successfully, but these errors were encountered: