Tree-attention for medusa about text-generation-inference HOT 2 OPEN

Hey @hustxiayang, IIRC, applying the tree-attention results in significantly more compute than using argmax, so the throughput cost is too high to justify the value.

Don't quote me on this but IIRC without the tree the speedup is 2x, while with the tree there is a maximum of ~3x speedup but throughput significantly decreases (divided by 10-20x).

To confirm with @Narsil eventually, hope it sheds some light on the question in the meantime.

from text-generation-inference.

Hi, I suppose you observe speedups from one of:

text-generation-inference/gemma-7b-it-medusa
text-generation-inference/Mixtral-8x7B-Instruct-v0.1-medusa
text-generation-inference/Mistral-7B-Instruct-v0.2-medusa

If that is the case, which dataset(s) did you use?

Another question: do you plan to support typical sampling mentioned in their paper?

from text-generation-inference.