DeepSeek-V4.1-Flash prefill on a single RTX 5090: 1,400 โ 1,950 t/s (+39%) ๐
One idea: stop re-decoding busy experts' weights 3โ6ร per chunk -decode once to FP16, let cuBLAS do the math.
Shipped it upstream to the Strata repo โ PR #
4366#, keen to hear what you think
@coldniko