CrackML by @ml.with.umang
GenAI interview guides / Speculative Decoding Interview Guide
GenAI / LLM

Speculative Decoding Interview Guide

Explain how speculative decoding reduces inter-token latency and when draft/verification overhead is worth it.

Mechanism

A cheaper proposer predicts future tokens and the target model verifies them in fewer expensive steps.

When it helps

Best gains are typically decode-bound, medium-to-low-QPS, memory-bound workloads with good acceptance.

Trade-offs

Low acceptance, proposer overhead, memory, tokenizer mismatch, or heavy batching can erase the benefit.

Serving view

Measure accepted tokens per target step plus end-to-end p50/p95 latency.