New todaydeepseek-v4.1IMAGE
DeepSeek V4.1 Flash prefill on one RTX 5090: 1,400 to 1,950 t/s (+39%)
By decoding busy experts' weights once to FP16 per chunk and letting cuBLAS do the math, the author raised prefill on a single RTX 5090 from 1,400 to 1,950 t/s (+39%), shipped as PR #4366 to the Strata repo.
Lyes Sbahi
@lyesbahiArchived 10 hours agoOriginally posted 2026.10.28
Lyes Sbahi @lyesbahiOriginally posted 2026.10.28
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
x.com/lyesbahi/status/2115192611656855639Archived 2026.10.28 12:04