quiz bu lëkkale ak tegtal · Dëgër Niveau

SmoothQuant ak quiz buy xayma limu nit ñi

Saytu li génn ci aktivaasioŋ bi, migraasioŋ ci eskaalu chaine bu nekk, notaasioŋ W8A8, pliage offline ak SmoothQuant's test inference kompromis.

Yooni njiit yu jëm ciSmoothquant ak jëfandikoo limu
Laaj 1 de 8

Ban jafe-jafe la SmoothQuant di wax ci inference bu woyof?

Xoolaatal laaj bu nekk8 laaj

Sooy def quiz bi, tontu yi ak leeral yi dañuy feeñ.

  1. Ban jafe-jafe la SmoothQuant di wax ci inference bu woyof?
    • Valeur yu sori lool ci chaine de activation
    • Poids yu tuuti lool ngir wane
    • Degrade biy réer ci diiru tasaare ci ginaaw
    • Dayo vokabileer token
  2. naka lay def ba SmoothQuant gëna yombal limu nit ñi?
    • Dafay dagg bépp aktivasioŋ bu weesu benn threshold buñ tëral
    • Dafay toxal jafe-jafe bi ci wàññi activation yi ak yokk poid yi ci benn facteur bi
    • Dafay tàggataat model bi ci 8-bit
    • Dafay soppi aktivaasioŋ yi ci poñ flotër kese
  3. Luy tekki ci notaasioŋ W8A8?
    • 8 couches de poids et 8 têtes d'attention
    • Decay de poids de 8 et alpha de 8
    • 8-bit poids ak 8-bit aktivasioŋ
    • 8 liggéeykat ak 8 gaawaaykat
  4. Lan moo waral scaling SmoothQuant du yokk benn coût ci kaw runtime bi?
    • Dafay dox ci GPU buñ jagleel
    • Dafay dox ludul ci diiru tàggat yaram, du musa jël doggal
    • Processeur yu bees yi duñu bàyyi xel ci fakteer yu mag yi
    • Balance yi dañu leen di xëcci ci poids yi ci wetam wala ci operation yu boole
  5. Ban cër la hiperparametru alfa di def ci SmoothQuant?
    • Dafay saytu ni outlier burden di toxalee ci aktivasioŋ yi dem ci poid yi
    • Dafay tëral tolluwaayu jàng ngir ajuste
    • Dafay wane limu bit kantite yi
    • Dafay tànn ban boppu bàyyi xel lañu wara dagg
  6. What result did the SmoothQuant paper report for its evaluated W8A8 large-language-model deployments?
    • A guaranteed 10× speedup on every GPU
    • About 1.56× speedup and roughly 2× memory savings with near-full-precision accuracy
    • No runtime or memory benefit
    • A 50% accuracy drop in every tested model
  7. Why does multiplying activation channels by 1/s and corresponding weights by s preserve a linear layer’s product?
    • The output is rounded to the nearest class
    • The model replaces multiplication with addition
    • The reciprocal scale factors cancel in the activation-weight multiplication
    • The weights are set to zero after scaling
  8. In SmoothQuant’s scale formula, what does increasing alpha generally do to outlier handling?
    • It removes the model’s weight matrices
    • It turns W8A8 into W16A16
    • It eliminates all quantization error
    • It shifts more scaling burden away from activations and onto weights