mirror-speculative-decoding

【效率与协作】Run draft and target models in parallel across GPU/NPU using bidirectional speculation: draft predicts forward continuations while target speculates correction paths simultaneously. Achieve 2.8x-5.8x wall-time speedup on 14B-66B models.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准