概览
dsh-plugin-c2c
Cache-to-Cache (C2C) for DeepSeek Harness: exposes the rosetta KV-cache fusion runtime (thu-nics/C2C) as the c2c_* tools through a local daemon
README / ZH
插件文档
目录摘要
Cache-to-Cache (C2C) for DeepSeek Harness: exposes the rosetta KV-cache fusion runtime (thu-nics/C2C) as the c2c_* tools through a local daemon
dsh.pub 核对固定版本的组合包契约、运行时事实与分发语义;完整 README 请查看源仓库。
在 GitHub 阅读完整 READMELIMITATIONS
已知限制
- Only the smallest pair is verified end to end. The 1.7B and 8B registry entries use the same code path but have not been run here. - One sharer per load. Upstream supports multi-sharer fusion (a bitmask in `kv_cache_index`); this plugin does not expose it yet. - No automated test suite. `python/verify_c2c.py` is a manual end-to-end check. - Training, multi-GPU, and the SGLang-backed evaluation harness are out of scope — they are CUDA-only and not reproduced.
