Decode · tok/s
114.3
Prefill
2141.2
tok/100W
19.9
RTX5090単機+Qwen38 Flash Next(UD-Q4_K_XL) llama.cppの魔改造版でpp 2,141.16 tok/s, tg 114.25 tok/s 出せたので満足。
First-person single 5090. Rank is the quoted 114.25 decode. Prefill 2,141.16 stays with the pair. Runtime is a modified llama.cpp; Strata not used.
