Skip to main content
T1 Large companies and unicorns Xiaomi New grad

Xiaomi AI Infra campus hire

Xiaomi AI Infra campus-hire interview questions covering inference optimization

Inference optimization

Coding

  1. Hand-write MGA (Multi-head Group Attention) in the editor, including init, forward, and the full attention-layer code.
  2. If you cannot finish the full implementation, describe the full compute flow from input to output.

Inference optimization

  1. Have you read the low-level source of vLLM and SGLang? How do you understand their architectures?
  2. What ways are there to load KV Cache?
  3. What does Prefill-Decode (PD) disaggregation do, and how is it implemented?
  4. Have you deployed models with vLLM? What throughput did you measure?
  5. Introduce the principle and role of MTP (Multi-Token Prediction).