streambp-efficient-backprop

Enables 2.8-5.5x longer sequences during LLM training via linear decomposition of chain rule along sequence dimension, maintaining exact gradients with lower memory cost.

Independent GitHub snapshot · Automated classification · Original documentation governs