Ikibazo gihuza ibibazo · Hagati Urwego

Ikibazo

Reba uburyo FlashAttention igabanya umuvuduko mwinshi-wibuka-traffic traffic mugihe ubara neza, hamwe nibyuma bihindura intego ya FlashAttention-3.

Inzira zijyanye no kuyoboraFlashattention
Ikibazo 1 Bya 8

Niyihe ntego nyamukuru FlashAttention yibitekerezo mubisanzwe?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Niyihe ntego nyamukuru FlashAttention yibitekerezo mubisanzwe?
    • Gusoma no kwandika matrix nini yo kwitondera kugabanya gahoro ya GPU
    • Umubare wo kureremba-ingingo kugwira
    • Igiciro cyo kwerekana ibimenyetso byinjira
    • Ubukererwe bwurusobe hagati ya GPUs
  2. Nigute umusaruro wa FlashAttention ugereranije no kwitabwaho bisanzwe?
    • Nukugereranya gucuruza neza kwihuta
    • Nibibare bisa (neza)
    • Cyakora gusa mugihe gito
    • Ihindura uburemere bwikitegererezo
  3. Nibihe bikoresho bya GPU FlashAttention ikoresha mugutunganya amakuru muri tile?
    • Ubunini bunini ariko buhoro buhoro bwibuke (HBM)
    • RAM ya CPU
    • Gitoya ariko byihuse cyane kuri chip SRAM
    • Ububiko bwa disiki
  4. Nubuhe buryo butuma FlashAttention ibara softmax utabonye amanota yose icyarimwe?
    • Kureka softmax rwose
    • Kuzenguruka amanota yose kuri zeru
    • Kubara softmax kuri CPU
    • Kumurongo (kwiyongera) softmax hamwe no gukuraho
  5. Kuberako matrike yuzuye itigeze ibikwa muri HBM, umunzani ukoresha wibuke:
    • Umurongo hamwe n'uburebure
    • Bikurikiranye hamwe n'uburebure bukurikirana
    • Byerekanwe n'uburebure bukurikirana
    • Ukurikije uburebure bwikurikiranya
  6. Which numeric format does FlashAttention-3 use in its Hopper-focused experiments to accelerate attention?
    • Only 64-bit floating point
    • FP8, with techniques to manage its numerical error
    • Integer text tokens
    • No reduced-precision format
  7. What work-partitioning change does FlashAttention-2 introduce to improve GPU parallelism?
    • It stores the complete attention matrix in HBM
    • It replaces attention with convolution
    • It reduces non-matrix-multiplication work and improves parallelization across thread blocks and warps
    • It uses only CPU execution
  8. In FlashAttention-3, what lets data movement overlap with computation?
    • Replacing attention with a lookup table
    • Computing all scores in host memory
    • Disabling the softmax step
    • Asynchronous execution on Hopper hardware