Jan 15, 2026 Hybrid attention: why new LLMs keep full attention in only a few layers Jul 03, 2025 Implementing the Transformer in PyTorch: why attention divides by √d_k