Hello! MAGI is excellent and amazing.
I'd like to ask if a longer kv_range is used during inference? The current configuration in the code only uses one additional range chunk for both the 4.5B model and the 24B model. It seems the remaining kv cache is not being utilized.
Thank you.
Hello! MAGI is excellent and amazing.
I'd like to ask if a longer kv_range is used during inference? The current configuration in the code only uses one additional range chunk for both the 4.5B model and the 24B model. It seems the remaining kv cache is not being utilized.
Thank you.