Skip to main content
T1 Large companies and unicorns Kuaishou Internship

Kuaishou AI Infra internship

Real Kuaishou AI Infra internship interview questions covering inference optimization and operator optimization

Inference optimization Operator optimization

Theory

  1. Besides the quantization plan used in the project, which other quantization methods do you know?
  2. From a quantization-error angle, what data-distribution conditions are more ideal?
  3. With outliers, more uniform distributions usually quantize better. That holds for INT8, FP8, and INT4 — why does it stop holding for FP4?
  4. Describe CUDA GEMM optimization methods, and how you choose Shared Memory size during the compute.

Project deep dive

  1. Introduce the main internship work.
  2. Detail the project’s technical design and implementation.

Live coding

  1. Write a Norm operator in CUDA. Input: x[N]; output: norm = (x - u) / o, where u = mean(x[N]), o = sqrt(sum((x - u)^2) / N).