Skip to main content
T3 Chips and hardware Enflame Internship Round 1

Enflame AI Infra internship first round

Real Enflame AI Infra internship first-round interview questions covering inference optimization, operator optimization, and related topics

Inference optimization Operator optimization

Below is what the Enflame AI Infra internship first round covers.

Project experience

  1. Deep follow-up on internship experience
  2. Deep follow-up on project experience

Quantization

  1. How you choose a quantization strategy: why INT8; how A100 and H100 support different quantization precisions
  2. Do you quantize model weights or KV Cache, and how do you set the scale
  3. After quantization, did you evaluate accuracy loss

Operator development

  1. Triton operator implementation logic, including tiling and related strategies
  2. Which official baseline you compared against, and which dtypes
  3. Where the speedup numbers come from; dynamic tiling strategy and operator config
  4. Did you consider writing operators in CUDA instead of Triton, and why you chose Triton
  5. Did you profile; how did memory throughput and related metrics look; what next optimizations

Inference optimization

  1. What share of end-to-end system latency does the Attention module take
  2. Is decode compute-bound or memory-bound; which side does KV Cache quantization improve
  3. Theoretical peak memory bandwidth of A100