LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding
read the original abstract
We present LayerSkip, an end-to-end solution to speed-up inference of large language models (LLMs). First, during training we apply layer dropout, with low dropout rates for earlier layers and higher dropout rates for later layers, and an early exit loss where all transformer layers share the same exit. Second, during inference, we show that this training recipe increases the accuracy of early exit at earlier layers, without adding any auxiliary layers or modules to the model. Third, we present a novel self-speculative decoding solution where we exit at early layers and verify and correct with remaining layers of the model. Our proposed self-speculative decoding approach has less memory footprint than other speculative decoding approaches and benefits from shared compute and activations of the draft and verification stages. We run experiments on different Llama model sizes on different types of training: pretraining from scratch, continual pretraining, finetuning on specific data domain, and finetuning on specific task. We implement our inference solution and show speedups of up to 2.16x on summarization for CNN/DM documents, 1.82x on coding, and 2.0x on TOPv2 semantic parsing task. We open source our code and checkpoints at https://github.com/facebookresearch/LayerSkip.
This paper has not been read by Pith yet.
Forward citations
Cited by 14 Pith papers
-
Tapered Language Models
Tapered Language Models monotonically decrease MLP width across depth with a cosine schedule, yielding better perplexity and downstream performance than uniform-width baselines across multiple architectures and scales...
-
Component-Aware Self-Speculative Decoding in Hybrid Language Models
Component-aware self-speculative decoding achieves high acceptance rates in parallel hybrid models like Falcon-H1 but fails in sequential ones like Qwen3.5, with the gap tied to how components are integrated.
-
FASER: Fine-Grained Phase Management for Speculative Decoding in Dynamic LLM Serving
FASER delivers up to 53% higher throughput and 1.92x lower latency in dynamic LLM serving by adjusting speculative lengths per request, early pruning of rejects, and overlapping draft/verification phases via frontiers.
-
Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
A recurrent-depth architecture enables language models to improve reasoning performance by iterating computation in latent space, achieving gains equivalent to much larger models on benchmarks.
-
Before Thinking, Learn to Decide: Proactive Routing for Efficient Visual Reasoning
PRP introduces proactive routing via Draft Rating Learning and Joint Rating Learning to route queries early between draft and target models for efficient multimodal reasoning.
-
The Context-Ready Transformer
Context-ready transformer adds a correction network to pre-contextualize tokens in a D-layer block, turning the model recurrent for inference while allowing K-step unrolled parallel training, with reported gains over ...
-
CLP: Collocation-Length Prediction for Zero-Loss Adaptive Multi-Token Inference
CLP is a lightweight linear predictor for safe multi-token spans in LLM decoding that delivers 1.14x-1.29x speedup on Qwen2.5 models with zero measured quality degradation.
-
Compute Where it Counts: Self Optimizing Language Models
SOL trains a policy to dynamically control multiple efficiency mechanisms per token via group-relative policy optimization on teacher-forced episodes, yielding better quality at matched average budget than static or r...
-
MCAP: Deployment-Time Layer Profiling for Memory-Constrained LLM Inference
MCAP uses load-time Monte Carlo profiling to estimate layer importance, enabling dynamic quantization (W4A8 vs W4A16) and memory tiering (GPU/RAM/SSD) that delivers 1.5-1.8x higher decode throughput than llama-cpp Q4_...
-
Dr.LLM: Dynamic Layer Routing in LLMs
Dr. LLM retrofits frozen LLMs with MCTS-supervised per-layer routers for skip/execute/repeat decisions, delivering up to +3.4% accuracy and 5-layer savings on reasoning tasks with strong out-of-domain generalization.
-
Speculative Verification: Exploiting Information Gain to Refine Speculative Decoding
Speculative Verification adds a companion model that estimates draft-target alignment via information gain to dynamically set verification length, delivering up to 2x speedup over standard speculative decoding across ...
-
CalexNet: Soft Cascade-Aligned Training and Calibration for Lightweight Early-Exit Branches
CalexNet aligns early-exit branch training and calibration to the cascade inference distribution via weighted sampling, survivor-subset calibration, and KL distillation to the backbone, matching or exceeding baselines...
-
Growing a Multi-head Twig via Distillation and Reinforcement Learning to Accelerate Large Vision-Language Models
TwigVLM adds a twig module to VLMs for twig-guided token pruning and self-speculative decoding, retaining 96% performance after pruning 88.9% visual tokens and delivering 154% speedup on long responses for LLaVA-1.5-7B.
-
DREAM-R: Multimodal Speculative Reasoning with RL-Based Refined Drafting, Precise Verification, and Fully Parallel Execution
DREAM-R introduces RL-based draft alignment, ratio-based verification, and parallel execution to accelerate speculative reasoning in multimodal models while preserving accuracy.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.