A small research prototype for demand-backed KV cache memory in SGLang. It reserves stable CUDA virtual addresses, commits physical GPU memory before a token slot is written, and returns empty chunks ...
Some results have been hidden because they may be inaccessible to you
