CrackML by @ml.with.umang
GenAI interview guides / LLM Inference Optimization Interview Guide
GenAI / LLM

LLM Inference Optimization Interview Guide

Optimize LLM serving across throughput, time-to-first-token, inter-token latency, memory, and cost.

Split latency

Separate queueing, tokenization, prefill, first-token latency, decode, tool time, and network time.

Serving levers

Continuous batching, prefix caching, quantization, optimized kernels, parallelism, and prefill/decode scheduling target different bottlenecks.

Capacity

Use prompt/output length distributions and SLOs, not only requests per second. Track KV pressure, batch size, and tail latency.

Quality guardrail

Every cost/latency optimization needs quality and numerical regression tests.