Ace the LLM Coding Interview

Modern LLM interviews are no longer just algorithm drills. You may be asked to implement attention masks, a KV cache, grouped-query attention, RoPE, a decoder-only Transformer, sampling strategies, memory estimates, or preference-optimization losses - from scratch, under time pressure.
Ace the LLM Coding Interview is a focused, hands-on guide to the problems that matter most in LLM coding rounds at frontier AI companies and AI-native startups. Every problem uses one compact format: intuition, plain-words math (every formula decoded in plain English), a concrete coding task, a clean Python + NumPy solution, and the follow-up questions interviewers actually ask.
Inside: the attention family (masks, scaled dot-product attention, multi-head attention, KV cache, GQA/MQA, RoPE); the full modern architecture (RMSNorm, SwiGLU, pre-norm blocks, a decoder-only mini-GPT, and the generate loop); tokenization and decoding (BPE, temperature, top-k, top-p, beam search); training mechanics (backprop, AdamW, LR schedules, gradient clipping and accumulation); scaling and systems math (KV-cache memory, prefill vs decode, parameter counts, mixed precision, speculative decoding); alignment and RL (SFT, reward models, DPO, PPO/RLHF, GRPO and RLVR); and evaluation (perplexity, pass@k, agent evals, and how interviewers grade you).
Each chapter ends with a last-day review sheet: formulas to memorize, code templates, shape invariants, and the common mistakes that break solutions.
No survey filler. No framework dependency. Just the core implementations you should be able to reproduce on a whiteboard.