SPADE Speculative Decoding
Draft-Model Acceptance/Rejection Analysis & Speculative Decoding Primitives
An experimental research repository for speculative decoding (SPADE) featuring M1 benchmark construction, M2 decoding baseline primitives, and M3 draft-model rejection analysis tools.
The Technical Bottleneck
Autoregressive LLM decoding is memory-bandwidth bound. Speculative decoding uses a small draft model to propose token sequences verified in parallel by a target model, but speedup collapses when draft tokens are frequently rejected.
Why Existing Baselines Fail
Understanding why and where draft models diverge from target distributions requires deterministic mathematical primitives and offline rejection diagnostics before running GPU kernel benchmarks.
Core Formulation & Insight
Modularizes speculative decoding research into three layers: M1 benchmark dataset preparation, M2 CPU-level speculative verification and acceptance-rate primitives, and M3 token-level rejection diagnostics.
Architectural Design
Three-module architecture (M1 Benchmark Construction, M2 Decoding Primitives, M3 Rejection Analysis) for speculative decoding research.
1. M1: Benchmark Construction
Prepares standardized prompt distributions for evaluating draft-target token alignment.
2. M2: Speculative Decoding Baseline Primitives
Implements exact rejection-sampling verification mathematics and acceptance tracking.
3. M3: Draft Rejection Diagnostics
Profiles token-position and entropy correlates of draft-model rejections.
Technical Stack & Systems Engineering
Implementation Highlights
- →Clean mathematical implementation of speculative verification and rejection profiling.
Datasets & Protocol
Datasets & Corpora
- M1 Offline Speculative Decoding Benchmark Suite
Evaluation Protocol
CPU-level mathematical verification of acceptance/rejection primitives (GPU model-backed speedup runs pending).
Tracked Evaluation Metrics
Completed M1–M3 mathematical primitives and offline rejection-analysis tooling; GPU model-backed speedup benchmarks are pending cache-safe adapter integration.
Benchmarking, Decoding Primitives, Rejection Analysis
GPU generation benchmarks explicitly marked pending
Empirical Analysis & Interpretation
Provides a clean, auditable foundation for analyzing draft-model rejection bottlenecks without conflating algorithmic acceptance rates with hardware-specific kernel variance.
What This System Contributes
- 01Developed the M1–M3 SPADE speculative decoding primitives and offline analysis suite.
Boundary Conditions & Future Work
- ·Live GPU model-backed generation and wall-clock speedup numbers are pending adapter validation and are not claimed.