Skip to content
#

gradient-checking

Here are 8 public repositories matching this topic...

Language: Python
Filter by language

LoRA implemented from the paper math in NumPy with finite-difference gradient checks. Rank-2 hits 99.7% accuracy training 0.4% of params with 226x less optimizer memory; a paired experiment shows exactly when LoRA cannot work (no domain coverage: 0.66 vs 1.0).

  • Updated Jul 31, 2026
  • Python

Add this topic to your repo

To associate your repository with the gradient-checking topic, visit your repo's landing page and select "manage topics."

Learn more