AI 导读
快速恢复必须保证正确性。 CacheConfig 会校验模型身份、并行布局、量化、dtype、GPU 能力和 PyTorch 版本。不匹配则回退到磁盘加载。 IPC 目前已在未量化和 block-wise FP8 模型上验证。
正文
Fast recovery must preserve correctness.
CacheConfig validates model identity, parallel layout, quantization, dtype, GPU capability, and PyTorch version. A mismatch falls back to disk loading.
IPC is currently verified for unquantized and block-wise FP8 models.
来源：@AntLingAGI · x.com