feat: introduce GRPO GPU fallback support, enhance training script with warmstart tagging, and add learning rate parameter for improved training flexibility

1b6d30b

Humanlearning commited on 28 days ago