Pull requests / #1340
#1340 core: retain fixed RAM originals during native async exchanges
open · @CC-David-CC · 0 评论 · 在 GitHub 查看
AMD / HIPNVIDIA / CUDADocumentation
描述
Experimental default-off resident-RAM option. Q8/FP16 KV on RTX PRO 6000 Blackwell, 32K/128K plain and MTP: retains approximately 8 GiB and avoids 13.1-14.0% of counterfactual eviction-download payload. TPS nearly unchanged; enabled outputs and work differ. No isolated speedup or quality-equivalence claim. Default-off exact tokens, ownership/ASan/UBSan and STOP checks passed. 
站内延伸阅读
链到安装、模型与版本说明,便于 SEO/GEO,非官方 issue 正文。