Show HN: AI-related Jax module (I hate if) (github.com)

A new open-source JAX module, designed to optimize Layer-wise Learning Rate Decay (LLRD), has been released, targeting performance enhancements for developers using JAX/Optax. This module directly addresses the overhead caused by conventional Python conditional statements during weight tree path scanning, which can fragment GPU/TPU compilation graphs and lead to inefficient performance. By implementing innovative techniques like "getattr Duck-Typing Attribute Masking" and inline algebraic operations, the module achieves ultra-fast parameter routing, effectively eliminating branch stalls. Additionally, the module introduces a high-efficiency spatial macro-curvature measurement system that maintains static compilation integrity, addressing common runtime errors related to dynamic tensor reshaping. It locks configuration constants to prevent tracing contamination, allowing diverse input shapes to maximize performance by converting them into contiguous memory layouts for faster hardware processing. Another crucial feature of this release is a numerical robustness kernel that preserves gradients in transcendental operations, ensuring stability during training and addressing common pitfalls such as NaN explosions. The repository is licensed under Apache 2.0, promoting widespread use and modification within the community.