Research Papers 论文研究 4d ago Updated 3d ago 更新于 3天前 48

Forward Pass Domain Adaptation (Without Cross-Layer Backpropagation) 前向传播域适应(无需跨层反向传播)

Forward-Pass-Only MLP training (FPO) adapts LLMs without a backward pass through the model body, achieving 2.7–3.2x throughput and ~40% less peak training memory versus standard fine-tuning The core empirical finding: at late transformer layers, the output-layer prediction error approximates the true gradient with cosine similarity of 0.47–0.59 across six surveyed models A two-minute diagnostic is introduced to quantify this gradient approximation per layer, identifying where late-layer adaptati FPO(Forward-Pass-Only MLP Training)是一种无需跨层反向传播即可适配大语言模型的新方法,吞吐量达标准微调的2.7-3.2倍 该方法峰值训练内存降低约40%,且离域基准测试性能保持在种子噪声范围内,这是全网络微调无法可靠实现的特性 核心发现:在transformer后期层,输出层预测误差与真实梯度的余弦相似度达0.47-0.59,跨越六个公开模型验证 提出两分钟诊断工具量化各层近似程度,指导目标层选择;在OLMo-2-7B、Qwen3-8B、Falcon3-7B上验证有效 将SFT定位到FPO目标层可行,但墙钟成本是FPO的2.2倍

65
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • Forward-Pass-Only MLP training (FPO) adapts LLMs without a backward pass through the model body, achieving 2.7–3.2x throughput and ~40% less peak training memory versus standard fine-tuning
  • The core empirical finding: at late transformer layers, the output-layer prediction error approximates the true gradient with cosine similarity of 0.47–0.59 across six surveyed models
  • A two-minute diagnostic is introduced to quantify this gradient approximation per layer, identifying where late-layer adaptation is viable for any given model
  • FPO computes a single error signal at the output and broadcasts it to each target layer with no inter-layer signal propagation and no autograd graph construction
  • Evaluated on OLMo-2-7B, Qwen3-8B, and Falcon3-7B: FPO improves in-domain perplexity while keeping MMLU, ARC-Challenge, HellaSwag, and Winogrande within seed-noise of baseline

Why It Matters

FPO challenges the assumption that cross-layer backpropagation is necessary for effective domain adaptation, offering a dramatically cheaper alternative that preserves off-domain generalization—a persistent pain point in fine-tuning. For practitioners constrained by compute or memory, this method could make adaptation accessible at scale without the catastrophic forgetting that often accompanies full-network fine-tuning.

Technical Details

  • Core mechanism: FPO computes a single prediction error at the output layer and applies it directly to each target late layer, bypassing the autograd graph entirely. No gradients flow between layers during adaptation.
  • Empirical basis: The method is grounded in the observation that output-layer prediction error correlates with true gradients (cosine similarity 0.47–0.59) at deep transformer layers, validated across six public models.
  • Diagnostic tool: A lightweight two-minute per-layer diagnostic quantifies the gradient approximation quality, enabling practitioners to identify which layers are suitable for FPO adaptation on any model.
  • Benchmarks: Tested on OLMo-2-7B, Qwen3-8B, and Falcon3-7B across in-domain perplexity and four off-domain benchmarks (MMLU, ARC-Challenge, HellaSwag, Winogrande).
  • SFT localization: Restricting standard supervised fine-tuning to FPO's identified target layers is feasible but incurs 2.2x the wall-clock cost of pure FPO, highlighting FPO's efficiency advantage.

Industry Insight

  • FPO could become a standard first-pass adaptation strategy for resource-constrained teams, enabling rapid domain fine-tuning on commodity hardware before committing to expensive full fine-tuning.
  • The per-layer diagnostic tool offers a practical framework for understanding gradient flow dynamics in any transformer, potentially guiding architectural decisions beyond just adaptation.
  • As compute costs continue to rise, forward-pass-only methods may shift the fine-tuning landscape toward lighter adaptation regimes, reducing the barrier to entry for domain-specific model customization.

TL;DR

  • FPO(Forward-Pass-Only MLP Training)是一种无需跨层反向传播即可适配大语言模型的新方法,吞吐量达标准微调的2.7-3.2倍
  • 该方法峰值训练内存降低约40%,且离域基准测试性能保持在种子噪声范围内,这是全网络微调无法可靠实现的特性
  • 核心发现:在transformer后期层,输出层预测误差与真实梯度的余弦相似度达0.47-0.59,跨越六个公开模型验证
  • 提出两分钟诊断工具量化各层近似程度,指导目标层选择;在OLMo-2-7B、Qwen3-8B、Falcon3-7B上验证有效
  • 将SFT定位到FPO目标层可行,但墙钟成本是FPO的2.2倍

为什么值得看

本文提出了一种突破性的模型适配范式,通过消除反向传播大幅降低计算成本,为资源受限场景下的大模型微调提供了新路径。其"输出误差近似梯度"的发现挑战了传统训练范式,具有理论创新价值。

技术解析

  • 核心假设:在transformer后期层,输出层预测误差可近似真实梯度(余弦相似度0.47-0.59),无需构建autograd图或跨层信号传播
  • 诊断工具:两分钟快速诊断可量化任意模型各层的梯度近似程度,识别适合后期层适配的模型
  • 模型评估:在OLMo-2-7B、Qwen3-8B、Falcon3-7B三个模型系列上验证,域内困惑度改善,MMLU、ARC-Challenge、HellaSwag、Winogrande保持在种子噪声范围内
  • 性能对比:吞吐量2.7-3.2x,峰值内存-40%;若将SFT定位到FPO目标层,墙钟成本为FPO的2.2倍

行业启示

  • 为边缘设备和资源受限场景的大模型适配提供了可行方案,降低部署门槛
  • "前向传播-only"范式可能开启新的模型优化研究方向,值得跟踪验证
  • 建议AI团队关注此类轻量级适配技术,评估其在特定业务场景中的适用性

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Fine-tuning 微调 Training 训练 Research 科学研究