Guide-yakabatana mibvunzo · Pakati Nhanho

Medusa Decoding Heads Quiz

Ongorora Medusa's multi-token prediction heads, candidates trees, tree-attention verification, kureba-prefix kugamuchirwa, uye kugovera kuchengetedza.

Nzira dzekutungamira dzinoenderanaMedusa Decoding Misoro
Mubvunzo 1 ye 8

Misoro yekuwedzera yeMedusa inofanotaura chii?

Tarisa mubvunzo wese8 mibvunzo

Mhinduro uye tsananguro zvinoonekwa paunenge uchitora bvunzo.

  1. Misoro yekuwedzera yeMedusa inofanotaura chii?
    • Multiple ramangwana tokens panzvimbo dzakasiyana pamberi
    • Kurasikirwa kwekudzidzira kwemuenzaniso
    • Iyo GPU yekushandisa
    • Mufananidzo wechiratidzo chechiratidzo
  2. Under which Medusa configuration does the paper describe lossless inference acceleration?
    • Iyo inomanikidza maremu emuenzaniso zvakakwana
    • Medusa-1 with a frozen backbone and exact verification against the original model
    • Haimboiti zvikanganiso mugirama
    • Zvinobvisa kudiwa kwemazwi
  3. Ndechipi chimiro chinoenderera mberi chemumiriri weMedusa chakarongerwa kuti chionekwe?
    • Rondedzero yakati sandara yanwisa rimwe nerimwe
    • Tafura yehashi
    • Muti wakakohwewa mune imwe pass ine muti wekutarisisa mask
    • Mushonga unodzokororwa pamusoro pechiratidzo chimwe nechimwe
  4. Chii chakakosha mukana weMedusa pamusoro pedhizaini-modhi yekufungidzira decoding?
    • Inoda mhando mbiri dzakakura dzakasiyana
    • Inoshanda chete pamifananidzo
    • Inobvisa iyo base modhi mushure mekushandisa
    • Iyo inodzivirira kuda yakaparadzana dhizaini modhi nekuwedzera misoro kune iripo
  5. Zvichida kuti ndeipi kukurumidza kunoitwa naMedusa?
    • Around 2-3x
    • Inenge 0.5x (inononoka)
    • Chaizvo 100x
    • Hapana shanduko inoyerwa
  6. Ndeipi yega yega Medusa musoro, zvekuvaka?
    • A full duplicate transformer
    • Iyo diki yasara feed-forward (MLP) layer pamusoro peiyo base yakavanzika mamiriro
    • A convolutional image encoder
    • Imwe tokenizer yakasiyana
  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