Medusa Yiyan Awọn ori adanwo

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

1. Kini afikun awọn olori Medusa ṣe asọtẹlẹ?
   • Awọn ami ami iwaju lọpọlọpọ ni awọn ipo oriṣiriṣi wa niwaju
   • Ipadanu ikẹkọ awoṣe
   • Eyi ti GPU lati lo
   • Akọle aworan fun ami kan
2. Under which Medusa configuration does the paper describe lossless inference acceleration?
   • O compress awọn iwọn awoṣe daradara
   • Medusa-1 with a frozen backbone and exact verification against the original model
   • Ko ṣe awọn aṣiṣe ni girama
   • O yọ awọn nilo fun a fokabulari
3. Ilana wo ni a ṣeto awọn itesiwaju oludije Medusa sinu fun ijẹrisi?
   • A Building akojọ gba wọle ọkan nipa ọkan
   • A hash tabili
   • Igi ti o gba wọle ni igbasilẹ ẹyọkan pẹlu iboju akiyesi igi kan
   • Loop ti nwaye lori ami-ami kọọkan
4. Kini anfani bọtini kan ti Medusa lori apẹrẹ-awoṣe iyipada asọye?
   • O nilo awọn awoṣe nla meji lọtọ
   • O ṣiṣẹ nikan lori awọn aworan
   • O paarẹ awoṣe ipilẹ lẹhin lilo
   • O yago fun nilo awoṣe iyasilẹ lọtọ nipa fifi awọn ori kun ọkan ti o wa tẹlẹ
5. Ni aijọju kini iyara wo ni Medusa ṣe ijabọ deede?
   • Ni ayika 2-3x
   • Nipa 0.5x (lọra)
   • Gangan 100x
   • Ko si iyipada iwọnwọn
6. Kini ori Medusa kọọkan kọọkan, ti ayaworan?
   • A ni kikun pidánpidán transformer
   • A kekere aloku kikọ sii-siwaju (MLP) Layer lori oke ti ipilẹ farasin ipinle
   • A convolutional image kooduopo
   • A lọtọ tokenizer
7. What does Medusa accept from a verified candidate tree?
   • Every speculative branch regardless of verification
   • Only the shortest candidate token
   • The longest prefix consistent with the base model’s decoding result
   • A random leaf from the tree
8. How do Medusa’s prediction-head offsets differ?
   • Every head predicts the same next token
   • Each head predicts a different vocabulary
   • The heads predict tokens only after decoding is complete
   • Each head predicts a token at a different position ahead in the sequence