What it does
Submissions, autotuning scripts and reference code for the Helion Kernel Challenge, which asks for Helion kernels submitted to the GPU MODE KernelBot leaderboard on B200 GPUs. The README is minimal; this summary is drawn from the repo's challenge notes and code.
| Problem | Description |
|---|---|
causal_conv1d | Causal depthwise 1D convolution (Mamba/Mamba-2) |
fp8_quant | Per-token-group FP8 E4M3 quantization |
gated_deltanet_chunk_fwd_h | Inter-chunk state recurrence for Gated DeltaNet |
gated_deltanet_chunk_fwd_o | Output computation for Gated DeltaNet |
gated_deltanet_recompute_w_u | WY-transform forward kernel for Gated DeltaNet |
How it works
Reference + task spec → Local autotune on B200 → Hardcode per-shape configs → Submit to KernelBot
- Autotuning runs locally (
autotune_*.py) because submissions can't rely on the autotuner under leaderboard timeouts; best configs are hardcoded per shape. - Variants compare the Triton and TileIR backends and PTXAS Advanced Controls Files (
*_acf.py,.acffiles). submit_all_leaderboard.shsubmits all five problems.
Tech stack
HelionPyTorchTritonTileIRNVIDIA B200PTXAS ACF