Lyes Sbahi(@lyesbahi):DeepSeek-V4.1-Flash prefill on a single RTX 5090: 1,400 → 1,950 t/s (+39%) 🚀 One idea: stop re-decoding busy experts' weights 3–6× per chunk -decode once to FP16, let cuBLAS do the math. Shipped it upstream to the Strata repo → PR #4366, keen to hear what you think @coldniko

Lyes Sbahi

@lyesbahi

Président de passionné par l'innovation technologique.

가입 June 2010

88 팔로잉 중 128 팬

Lyes Sbahi@lyesbahi

2026.10.27 21:23

DeepSeek-V4.1-Flash prefill on a single RTX 5090: 1,400 → 1,950 t/s (+39%) 🚀 One idea: stop re-decoding busy experts' weights 3–6× per chunk -decode once to FP16, let cuBLAS do the math. Shipped it upstream to the Strata repo → PR #4366#, keen to hear what you think @coldniko