efficient-long-context-reasoning

Achieve 10× higher decoding throughput on long prompts by replacing 50% of cross-attention layers with gated memory units (GMUs) combining SSMs and attention. Maintains reasoning capability while reducing memory I/O bottleneck from O(d_kv·N) to O(d_h).

Independent GitHub snapshot · Automated classification · Original documentation governs