Gefen: Optimized Stochastic Optimizer

In three linesGefen is a memory-efficient optimizer reducing AdamW's memory footprint by ~8x (6.5 GiB per billion parameters) through shared second-moment estimates across parameter blocks and learned codebook quantization of first moments. Maintains AdamW-level performance while enabling larger microbatches in distributed training.

Summary generated by Claude — human-verified