Can you reduce the kv head num of this model? "num_key_value_heads": 64, which requies a lots of kv cache

#1
by luchangli03 - opened
•
This comment has been hidden (marked as Resolved)
•
This comment has been hidden (marked as Resolved)

We will continue iterating on this going forward. Let's try training with this suggestion.

Sign up or log in to comment