xingy555888/cancri-latent-relay
收藏Hugging Face2026-04-28 更新2026-05-03 收录
下载链接:
https://hf-mirror.com/datasets/xingy555888/cancri-latent-relay
下载链接
链接失效反馈官方服务:
资源简介:
Cancri是一种通过在检查点边界直接传递隐藏状态来链接语言模型检查点的机制——中间无需进行令牌解码。特点包括:峰值内存仅相当于1个模型(与链长度无关)、每个序列约160 KB的中继负载(L=20,d=2048,float32)、已在Qwen3.5-2B Base → Instruct上验证。关键结果:与Instruct基线(自回归)相比,困惑度比为0.949×;三向令牌一致率为47%;在20个令牌生成过程中未观察到退化现象。
A mechanism for chaining language model checkpoints by passing hidden states directly across checkpoint boundaries — no token decoding in between. Features: Peak memory = 1 model, regardless of chain length; ~160 KB relay payload per sequence (L=20, d=2048, float32); Verified on Qwen3.5-2B Base → Instruct. Key Results: 0.949× perplexity ratio vs Instruct baseline (autoregressive); 47% three-way token consistency rate; No degeneration observed over 20-token generation.
提供机构:
xingy555888



