Skip to content

Paper 090 / Google DeepMind

Fast Inference from Transformers via Speculative Decoding

Speculative decoding lets a faster draft model propose token blocks that a larger target model verifies in parallel while preserving the target distribution.

UPDATED 16 Jul 2026SOURCE-LED REVIEWRESEARCH REVIEW
01

Paper, researchers, and primary source

Inference, evaluation & serving / inference_decoding

Speculative decoding lets a faster draft model propose token blocks that a larger target model verifies in parallel while preserving the target distribution.

CONTRIBUTION / 01

Contribution 1

Used a small approximation model to draft several future tokens.

CONTRIBUTION / 02

Contribution 2

Verified proposed tokens with parallel target-model scoring and rejection sampling.

CONTRIBUTION / 03

Contribution 3

Reduced sequential target-model calls without changing the sampled target distribution.

02

Research context

inference_decoding / 2022

Fast Inference from Transformers via Speculative Decoding places speculative decoding inside the broader inference decoding discussion at Google DeepMind, with draft model supplying a second analytical lens. The paper's through-line contains three reported moves: Used a small approximation model to draft several future tokens; Verified proposed tokens with parallel target-model scoring and rejection sampling; and Reduced sequential target-model calls without changing the sampled target distribution. That sequence keeps rejection sampling tied to the reported work instead of treating it as a free-standing promise. The practical stake is equally bounded: acceleration depends on draft acceptance, output length, batch size, hardware balance, and the cost of operating two models.

03

Methods and evidence reading

1 cataloged method notes

METHOD / 01

Method 1

The experimental design in Fast Inference from Transformers via Speculative Decoding tests used a small approximation model to draft several future tokens and verified proposed tokens with parallel target-model scoring and rejection sampling against the paper's documented baselines, datasets, model variants, or systems workloads.

How to read the evidence

Evidence for Fast Inference from Transformers via Speculative Decoding is best read as a scoped argument rather than a universal verdict. The source record contains two linked moves: Used a small approximation model to draft several future tokens; and Verified proposed tokens with parallel target-model scoring and rejection sampling. Its reported outcome is: Reduced sequential target-model calls without changing the sampled target distribution. The defensible takeaway remains acceleration depends on draft acceptance, output length, batch size, hardware balance, and the cost of operating two models. That conclusion must travel with the recorded boundary that evidence for speculative decoding in Fast Inference from Transformers via Speculative Decoding covers service-level objectives, accelerator hardware, request shapes, software revisions, reported models, comparison baselines, and numerical precision; behavior beyond that documented envelope remains untested. A replication should preserve the disclosed setup and test whether speculative decoding still holds when draft model conditions change. For a follow-on study of Fast Inference from Transformers via Speculative Decoding, pair speculative decoding measurements with draft model slices and preserve negative examples around Verified proposed tokens with parallel target-model scoring and rejection sampling as first-class evidence.

04

Findings in the source record

1 paper-specific findings

  1. The reported evidence in Fast Inference from Transformers via Speculative Decoding supports reduced sequential target-model calls without changing the sampled target distribution.
05

Practical implication for AI builders

Google DeepMind / 2022

06

Proposed BrokenGPT application

Research blueprint / proposed status

INTEGRATION POINT / 01

Proposed route placement / speculative decoding

For a proposed BrokenGPT experiment based on Fast Inference from Transformers via Speculative Decoding, pair version-locked draft and target endpoints, then measure exact distributional parity, acceptance rate, latency, memory, and failure behavior by request type. Keep the speculative decoding path isolated, versioned, and attributable to this research record.

VALIDATION METRIC / 02

Proposed acceptance test / draft model

Validate the proposed speculative decoding route against the paper's reported outcome: Reduced sequential target-model calls without changing the sampled target distribution. For the Fast Inference from Transformers via Speculative Decoding prototype, collect workload stability, latency, output parity, and accepted tokens and audit draft model slices independently before promoting the speculative decoding configuration.

TRADEOFF / 03

Proposed decision boundary / rejection sampling

Balance extra machinery, speed, and sampling fidelity before promoting the proposed rejection sampling design. Because for a later speculative decoding implementation, workload drift, authentication, networking overhead, tokenization, failure recovery, and safety checks define unresolved boundaries that require direct observation, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.

07

Limitations, verification, and source

Boundaries recorded with the paper

Limitations

  • Evidence for speculative decoding in Fast Inference from Transformers via Speculative Decoding covers service-level objectives, accelerator hardware, request shapes, software revisions, reported models, comparison baselines, and numerical precision; behavior beyond that documented envelope remains untested.
  • For a later speculative decoding implementation, workload drift, authentication, networking overhead, tokenization, failure recovery, and safety checks define unresolved boundaries that require direct observation.

PRIMARY SOURCES

  1. 01
    Fast Inference from Transformers via Speculative Decoding

    Google DeepMind — Primary primary arXiv paper / 30 November 2022 / Yaniv Leviathan, Matan Kalman, Yossi Matias

Related research reviews

View all 100 credited research papers

STRAIGHT ANSWERS

Frequently asked questions

01What does Fast Inference from Transformers via Speculative Decoding study?

Speculative decoding lets a faster draft model propose token blocks that a larger target model verifies in parallel while preserving the target distribution.

02Which methods does Fast Inference from Transformers via Speculative Decoding use?

The experimental design in Fast Inference from Transformers via Speculative Decoding tests used a small approximation model to draft several future tokens and verified proposed tokens with parallel target-model scoring and rejection sampling against the paper's documented baselines, datasets, model variants, or systems workloads.

03What does Fast Inference from Transformers via Speculative Decoding report?

The reported evidence in Fast Inference from Transformers via Speculative Decoding supports reduced sequential target-model calls without changing the sampled target distribution.

04What is the proposed BrokenGPT application for Fast Inference from Transformers via Speculative Decoding?

Proposed: pair version-locked draft and target endpoints, then measure exact distributional parity, acceptance rate, latency, memory, and failure behavior by request type.

INFERENCE, EVALUATION & SERVING / PAPER 090

Continue after Fast Inference from Transformers via Speculative Decoding

After Fast Inference from Transformers via Speculative Decoding, browse the full index for adjacent inference_decoding research and work from Google DeepMind.

Open the paper index