williamhtan/kernelsight
KernelSight v4 Per-timestep workload labels for GPU execution traces. KernelSight pairs every GPU workload trace with a dense, per-timestep workload labeling. Each snapshot is a [24, 512] counter image — 24 hardware-counter channels sampled across 512 equal-width time bins — paired with per-bin labels drawn from a two-level hierarchy of 12 coarse (L1) and 73 fine (L2) workload classes. The goal is to label what a kernel is doing at each instant (matmul, attention, reduction… See the full description on the dataset page: https://huggingface.co/datasets/williamhtan/kernelsight.
Add kernels/wgmma (3 snapshots' npz)
Add kernels/vector_add (60 snapshots' npz)
Add kernels/scatter (93 snapshots' npz)
Add kernels/reduction (48 snapshots' npz)
Add kernels/kernelbench (1440 snapshots' npz)
Add kernels/gather (51 snapshots' npz)
Add kernels/cutlass_ws_overlap (1416 snapshots' npz)
Add kernels/cutlass_sparse_gemm (54 snapshots' npz)
Add kernels/cutlass_grouped_gemm (36 snapshots' npz)
Add kernels/cutlass_gemm (834 snapshots' npz)
Add kernels/cutlass_fp8_gemm (42 snapshots' npz)
Add kernels/cutlass_fmha (255 snapshots' npz)
Add loader + taxonomy tools
Add split definitions
Add data_info.md
Add MANIFEST_v4.md
Add README.md
Store small per-snapshot .npz as regular git blobs (no LFS)
initial commit