hao-ai-lab / LookaheadDecoding

Apache License 2.0
1.04k stars 63 forks source link

Why jacobi_sample_multilevel() fill window with argmax instead of also using sampling? #53

Open yangbohust opened 3 months ago

yangbohust commented 3 months ago

I have a question, Please ask the great author to help answer this question.

Why jacobi_sample_multilevel() fill window with argmax instead of also using sampling?

code location: https://github.com/hao-ai-lab/LookaheadDecoding/blob/9d50de4a81d1b473bfce104ace18fbbbb6dc3255/lade/decoding.py#L466 https://github.com/hao-ai-lab/LookaheadDecoding/blob/9d50de4a81d1b473bfce104ace18fbbbb6dc3255/lade/decoding.py#L478

Looking forward to your reply. Thanks