Pinned Loading

1. gate-fusion gate-fusion Public

   Measuring cuStateVec before writing a kernel to compete with it. The benchmark disproved the premise; the ADR records that instead of working around it.

   Python

2. int4-gemv int4-gemv Public

   Fused INT4 dequantisation GEMV for LLM decode — 85% of the memory roofline on sm_120, with the baselines and the caveats

   Cuda