NVFP4 GEMM, GEMV, linear layers and quantization for NVIDIA GPUs without native FP4 (sm_50–sm_90): every NVIDIA NVFP4 variant (ModelOpt, Transformer Engine 1x16 / 16x16, 4over6, swizzled scales) bit for bit, exact block sums, tensor cores on sm_75–sm_89, C API and PyTorch bindings
pascal deep-learning gpu cuda inference pytorch nvidia matrix-multiplication cuda-kernels quantization gemm 4-bit mixed-precision gemv llm fp4 nvfp4 transformer-engine dp4a e2m1
-
Updated
Nov 7, 2026 - Cuda