Skip to main content
T0 Big tech Baidu Round 1

Baidu AI Infra first round (3)

Real Baidu AI Infra first-round interview questions covering training optimization and high-performance computing

Training optimization High-performance computing

Reinforcement learning

  1. Relationship between Trust Region methods and PPO
  2. Is PPO on-policy or off-policy? Why introduce importance sampling?
  3. In PPO, how does the clip mechanism bound the upper and lower limits when advantage A is positive vs negative?
  4. How is the PPO loss computed? How is Generalized Advantage Estimation (GAE) computed, and how does lambda affect variance and bias?
  5. How GRPO computes loss; how sequence-level loss is assigned to each token; difference between sequence-level average and batch-level average
  6. Traits of other GRPO variants (such as DAPO, GSPO, GFPO)
  7. Basic idea and applications of Agentic RL

Distributed training

  1. During training, what share of GPU memory do optimizer states, gradients, and model parameters each take
  2. Compare FSDP with DeepSpeed ZeRO Stage 1/2/3

Project deep dive

  1. Project introduction and technical design; how you combine LLMs with concrete application scenarios; evaluation methods and metric results

Live coding

  1. Level-order traversal of a binary tree, also recording each node’s level (LeetCode 102 variant)