Speculative-Decoding repo benchmarks proposer tradeoffs

Strong repo, because it turns speculative decoding from a buzzword into a controlled systems experiment. The useful lesson is that acceptance rate alone is not the metric that matters; verifier cost, proposer cost, batching, and cache behavior decide throughput.

• –One contract across methods makes the comparisons actually meaningful: same target model, same verifier path, same metrics schema, same baseline.
• –The repo makes the key distinction explicit between learned proposers like EAGLE/Medusa and draft-model speculation, which is usually blurred in casual explanations.
• –PARD is the best example of why lower acceptance can still win: cheaper parallel proposers can outrun a heavier autoregressive draft model even when they match fewer tokens.
• –N-gram and suffix decoding are a good reality check for repetitive prompts: they can look surprisingly strong when the prompt has reusable structure, but they are context-dependent rather than universal.
• –The small-slice benchmarks are useful for behavior and implementation debugging, but they should not be overread as general throughput rankings across workloads.