Is your feature request related to a problem? Please describe.
We currently only support full logits; or top-k only if we have top-k logits already stored
Describe the solution you'd like
Add an option/arg to use only top_k logits for distillation instead of full logits
Is your feature request related to a problem? Please describe.
We currently only support full logits; or top-k only if we have top-k logits already stored
Describe the solution you'd like
Add an option/arg to use only top_k logits for distillation instead of full logits