layer normalization

**Layer Normalization Pre-LN vs Post-LN Architecture** determines **where normalization occurs relative to residual connections in transformer blocks — Pre-LN (normalizing before sublayers) enabling training stability and better gradient flow for deep models while Post-LN (normalizing after additions) theoretically preserving more representational capacity**. **Post-LN (Original Transformer) Architecture:** - **Residual Block Structure**: input x → sublayer (attention/FFN) → LayerNorm → output: (x + sublayer(x)) normalized - **Mathematical Form**: y_i = LN(x_i + sublayer(x_i)) where LN(z) = (z - mean(z))/sqrt(var(z) + ε) — normalizes across feature dimension D - **Representational Capacity**: post-normalization preserves original residual amplitude — sublayer outputs retain original scale before normalization - **Training Challenges**: gradient magnitude inversely proportional to layer depth — deep networks (>24 layers) suffer vanishing gradients (0.1-0.01 gradient per layer) - **Stability Issues**: post-LN requires careful initialization (small embedding scale 0.1, attention scale √d_k) — training becomes brittle with learning rate sensitivity **Pre-LN (Modern Architecture) Architecture:** - **Residual Block Structure**: input x → LayerNorm → sublayer (attention/FFN) → output: x + sublayer(LN(x)) - **Mathematical Form**: y_i = x_i + sublayer(LN(x_i)) — normalization applied before transformation - **Gradient Flow**: residual connection carries constant gradient 1.0 throughout depth — enabling stable training of very deep models (100+ layers) - **Implicit Scaling**: normalized inputs restrict to unit variance, naturally scaling sublayer outputs — reduces initialization sensitivity - **Easier Optimization**: learning rate becomes less critical, wider range of hyperparameters work (LR 1e-4 to 1e-3) — robust training across model sizes **Technical Comparison:** - **Residual Learning**: post-LN preserves residual as original scale, pre-LN normalizes residual — mathematical difference with gradient implications - **Layer Skip Strength**: post-LN enables stronger skip connections (amplitude 1.5-2.0x), pre-LN weaker (amplitude ~1.0x) — affects information flow - **Output Distribution**: post-LN produces outputs with higher variance (std 1.5-2.0), pre-LN more constrained (std 1.0) — impacts downstream layer assumptions - **Initialization Dependency**: post-LN requires embedding scaling 0.1-0.2, pre-LN works with standard 1.0 — critical for stable training **Empirical Performance Data:** - **GPT-2 (Post-LN, 24 layers)**: requires LR 5e-5 with warmup schedule, trains unstably with LR 1e-3 — careful tuning needed - **GPT-3 (Post-LN, 96 layers)**: achieves 175B parameters despite depth, requires extensive grid search for hyperparameters - **Transformer-XL (Pre-LN)**: simplifies to relative position embeddings with pre-LN, trains stably without special initialization - **Llama 2 (Pre-LN)**: uses pre-LN throughout with RoPE, achieves 70B parameters with fewer training tricks — 20% fewer tokens needed for same performance **Practical Implications:** - **Depth Scaling**: pre-LN enables efficient scaling to 100+ layer models where post-LN becomes infeasible — key for retrieval-augmented and deep reasoning models - **Fine-tuning Stability**: pre-LN allows larger learning rates (5e-5 to 1e-4) without divergence — beneficial for parameter-efficient fine-tuning - **Batch Size Sensitivity**: post-LN training sensitive to batch size effects, pre-LN more robust — enables flexible batch sizing in distributed training - **Numerical Stability**: pre-LN naturally keeps activations near normal distribution — reduces overflow/underflow in mixed precision training (FP16, BF16) **Recent Architecture Trends:** - **RMSNorm Adoption**: simplifying layer normalization to RMS(z) × γ without centering — 5-10% speedup with pre-LN, used in Llama and PaLM - **Parallel Attention-FFN**: computing attention and FFN in parallel with pre-LN — enables faster training (1.5x throughput) in modern architectures - **ALiBi Integration**: combining pre-LN with Attention with Linear Biases (ALiBi) — avoids positional embedding learnable parameters while maintaining efficiency **Layer Normalization Pre-LN vs Post-LN Architecture is fundamental to transformer design — Pre-LN enabling stable training of deep models and becoming standard in modern architectures like Llama, PaLM, and recent foundation models.**