sparse-attention-reasoning

Learn sparse attention patterns for reasoning model decoding via self-distilled gating, achieving 9x speedup at 90% sparsity while maintaining reasoning quality.

Independent GitHub snapshot · Automated classification · Original documentation governs