Early-Exit Adapter Decoding for Low-Latency Token Verification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning models for sequence processing, such as large language models (LLMs) and vision-language models (VLMs), suffer from high computational and memory overhead due to the need for separate models for drafting and verification, leading to increased latency and resource utilization.

Innovation Solution

A single model architecture is modified with an adapter layer to perform fast drafting using lower layers and verification using higher layers, leveraging early exit adapters to reduce latency and computational overhead.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If separate models are used for drafting and verification, then output quality is improved, but computational overhead and latency increase

Engineering Contradiction:
Improveoutput qualityVSAvoidlatency
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent combines separate drafting and verification models into a single unified model architecture. The drafting module and verification module share the same underlying model parameters and processing infrastructure, eliminating the need to execute two distinct models while maintaining the functional separation needed for draft generation and quality verification.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The unified model serves multiple functions: it acts as both the drafting model that generates initial token sequences and the verification model that validates those sequences. The same model parameters are used for both draft generation and verification, allowing the system to perform multiple roles with a single model instance.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Measurement precision

If separate models are used for drafting and verification, then drafting accuracy is improved, but memory overhead increases

Engineering Contradiction:
Improvedrafting accuracyVSAvoidmemory overhead
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent merges separate drafting and verification models into one unified model, significantly reducing memory overhead. Instead of loading and maintaining two separate model instances in memory, the system uses a single model that can perform both drafting and verification functions, thereby halving the memory requirements for model storage.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The unified model is designed to be multi-functional, serving as both the drafting model and verification model. This universality allows the system to maintain high drafting accuracy while using only one model instance in memory, rather than requiring separate memory allocations for multiple specialized models.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If full model processing is used for each token, then verification accuracy is improved, but processing speed decreases

Engineering Contradiction:
Improveverification accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSSpeed

Solution Approach 1:

The patent implements early exit mechanisms that allow the verification process to terminate before completing full model processing for all tokens. When the drafting phase generates high-confidence token predictions, the verification can exit early without processing the entire model sequence, thereby maintaining accuracy for confident predictions while speeding up overall processing.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The verification process performs partial processing rather than complete full-model processing for every token. The system selectively applies verification based on confidence thresholds and token positions, performing full verification only when necessary while using lighter validation for routine tokens, thus balancing accuracy with processing speed.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4711981A1Efficient estimation & verification with early exits
Publication Date: 2026.03.18 GOOGLE LLC
  • EP4711981A1 patent drawingFigure 1
  • EP4711981A1 patent drawingFigure 2
  • EP4711981A1 patent drawingFigure 3

AI summary

One example aspect is directed to a computer-implemented method (400) for performing model decoding with reduced latency. The method includes obtaining (402) a pre-trained sequence processing model comprising a plurality of layers. The method includes modifying (404) the sequence processing model to contain an adapter layer (106) that is configured to receive and process an intermediate representation generated by a particular intermediate layer of the plurality of layers to predict an output token. The method includes training (406) the adapter layer while holding the plurality of layers of the sequence processing model frozen. The method includes deploying (408) the sequence processing model for speculative decoding in which the adapter layer, the particular intermediate layer, and the plurality of layers (104) that precede the particular intermediate layer perform speculative token decoding and the plurality of layers (108) that are subsequent to the particular intermediate layer perform token verification.