Reduce GPU to 3min/15steps to fit remaining Zero GPU quota

Re-enable Zero GPU with anti-crash measures: data download on CPU first, 10min timeout, 3L/96d model

Switch to CPU mode - remove GPU requirement, tiny model (2L/64d/20steps)

Increase GPU timeout to 5min, reduce model size (3L/128d) to avoid OOM

Run data download script to fetch tokenizer

Download parameter-golf data for tokenizer

Add Zero GPU decorator to activate GPU

Add sentencepiece + other missing deps

Add train_gpt_kernel.py directly + simplify app

Fix remote branch tracking + add direct download fallback

Fix branch checkout to use correct PR branch name

Fix streaming output + reduce model size for Zero GPU

Setup μ⁸ Kernel training for Parameter Golf

initial commit