AdamWClip adds adaptive clipping to AdamW

This is a practical training-stability tweak that could save teams from brittle manual clipping thresholds, but it still needs broader benchmarking to prove durability.

• –The GitHub README positions it as a drop-in AdamW replacement with the same setup flow plus adaptive clipping controls.
• –Its key claim is using AdamW’s existing second-moment state to avoid additional memory costs.
• –Evidence so far is preliminary and mostly shared in-thread, so reproducible benchmarks across LLM and non-vision workloads are the next credibility milestone.
• –If results hold, it could become a low-friction default for fine-tuning pipelines that currently rely on hand-tuned grad clipping.