Maswali yanayounganishwa na mwongozo · Kati Kiwango

Maswali ya Makini ya Flash

Angalia jinsi FlashAttention inavyopunguza trafiki ya kumbukumbu-bandwidth ya juu huku ikiweka umakini kamili, na maunzi gani hubadilisha malengo ya FlashAttention-3.

Njia za mwongozo zinazohusianaUangalifu
Swali 1 ya 8

Je, ni shabaha gani kuu za FlashAttention katika umakini wa kawaida?

Hakiki kila swali8 maswali

Majibu na maelezo huonekana unapojibu maswali.

  1. Je, ni shabaha gani kuu za FlashAttention katika umakini wa kawaida?
    • Kusoma na kuandika matrix kubwa ya umakini ili kupunguza kasi ya kumbukumbu ya GPU
    • Idadi ya kuzidisha kwa sehemu zinazoelea
    • Gharama ya kuweka alama kwenye maandishi ya pembejeo
    • Muda wa kusubiri wa mtandao kati ya GPU
  2. Matokeo ya FlashAttention yanalinganishwaje na umakini wa kawaida?
    • Ni makadirio ambayo hufanya biashara ya usahihi kwa kasi
    • Inafanana kihisabati (sawa)
    • Inafanya kazi kwa mlolongo mfupi tu
    • Inabadilisha uzito wa mfano
  3. FlashAttention hutumia kumbukumbu gani ya GPU kwa kuchakata data kwenye vigae?
    • Kumbukumbu kubwa lakini polepole ya juu-bandwidth (HBM)
    • RAM ya CPU
    • SRAM ndogo lakini ya haraka sana kwenye chip
    • Hifadhi ya diski
  4. Ni mbinu gani huruhusu FlashAttention kukokotoa softmax bila kuona alama zote mara moja?
    • Kuruka softmax kabisa
    • Kuzungusha alama zote hadi sifuri
    • Inatayarisha softmax kwenye CPU
    • Mkondoni (wa nyongeza) softmax na kukimbia upya
  5. Kwa sababu matrix kamili ya umakini haihifadhiwi kamwe katika HBM, mizani ya utumiaji wa kumbukumbu takriban:
    • Linearly na urefu wa mlolongo
    • Mara nne na urefu wa mlolongo
    • Kwa kielelezo na urefu wa mlolongo
    • Kujitegemea kwa urefu wa mlolongo
  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