AIAny
Icon for item

Don't Drop Dropout: Optimizing Layer Sparsity for Efficient LLM Training and Inference

Evaluates layer dropout (stochastic depth) in LLM pretraining, finds optimal layer distributions and time schedules that save ~20–25% training FLOPs at comparable or lower validation loss, and enables zero-shot layer skipping and self-speculative decoding for up to ~1.4–1.5× inference speedup.

Introduction

Layer dropout vanished from many modern LLM recipes because of concerns it harms accuracy. This paper shows the opposite is possible: with careful choices of layer-density distribution, time schedule, and a few optimizer/hyperparameter adjustments, layer dropout reduces training compute and yields models that can be elastically accelerated at inference time with minimal accuracy loss.

Key Findings
  • Best-practice configuration (Increasing Layer Distribution + Decreasing Time Schedule, “ILD+DTS”) produces models that match or beat dense baselines at the same training FLOPs, enabling roughly 20–25% non-embedding FLOPs savings so training can be cheaper or faster.
  • Practical hyperparameter rules: scale activations during training by 1/ρ (ρ = layer density) to keep activation magnitudes stable; prefer per-sequence, block-level dropout granularity for better generalization; tune optimizer settings when increasing dropout to avoid degradation.
  • Post-training elasticity: models trained with layer dropout are robust to zero-shot layer skipping, early-exit strategies, and self-speculative (draft-and-verify) decoding, producing up to ~1.4–1.5× end-to-end inference speedups with negligible loss on standard benchmarks.
  • Empirical scale: conclusions hold across a large sweep (≈2400 runs) spanning models from 271M to 8.2B parameters and datasets up to 160B tokens, with all pretraining experiments run on Cerebras CS-3 hardware.
Who it's for and trade-offs

Great fit if you train foundation or production LLMs and need to reduce training FLOPs or want inference-time elasticity without post-hoc retraining. The approach requires extra hyperparameter tuning (layer density schedule, optimizer tweaks) and careful validation; earlier negative reports about dropout reflect poor configuration rather than an inherent impossibility. Not ideal if you require single-epoch pretraining regimes where dropout has been shown to hurt, or if you want plug-and-play results without any retuning. The paper does not fully explore attention-head or neuron-level dropping, non-transformer architectures (e.g., MoE), or produce a formal scaling law for maximum safe dropout rates.

Information

  • Websitearxiv.org
  • OrganizationsCerebras Systems, MBZUAI
  • AuthorsMostafa Elhoushi, Alex Pretko, Nolan Dey, Bin Claire Zhang, Gavia Gray, Gurpreet Gosal, Abdulrahman Mahmoud, Shane Bergsma, Joel Hestness
  • Published date2026/09/04

More Items

Presents two LLM-based search agents (Iris-mini and Iris-pro) trained by alternating supervised fine-tuning and reinforcement learning against live web search. Key features: web-graph-derived multi-hop tasks with entity abstraction, SFT–RL climbing, inference-time context management, and state-of-the-art open-source benchmark results.

Decides when past post-training updates should be reused for autonomous LLM adaptation by introducing Boundary-Calibrated Intervention Transfer (BCIT). BCIT binds effects to source context, checks applicability and hard conflicts, and runs bounded trials to obtain current-state evidence—reducing harmful updates and improving equal-budget final-model quality.

Converts a natural-language function specification into a reusable local neural function by using teacher models to synthesize examples and finetuning a small adapter for a compact interpreter. Achieves higher semantic accuracy (83.6% on FuzzyBench-Hard) at the cost of roughly one minute compile time; produces versionable PAW artifacts for local deployment.