submission 197871
gau.nernst · python · License unknown
Use it
Vendorable · source mirrored · license unknownView source →
No package. Vendor the mirrored source: 28 lines, June 9 Researcher Reciprocity License v1.0.
submission_ref.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-nvfp4-dual-gemm-197871?include=source"interfacepython
Compatibility
measured onNVIDIA B200
declared hardwareNVIDIA B200
architecturessm_100
dtypesfp8_e4m3, nvfp4
Benchmark evidence
1 measurement across 1 GPU, fastest first.
Operation / workload
Hardware
Latency
Rank
Observed
Reported · How evidence levels are derived →
Source and license
sourceavailable
revision digestsha256:a8181137d8e0bf9edb179c1a14e15c5292718dcd4752ca57bf483480dfe6194f
license declaredunknown
license concludedunknown
authorsgau.nernst
imported2026-08-15
Kernel source
submission_ref.py28 lines
#!POPCORN leaderboard nvfp4_dual_gemm
#!POPCORN gpu NVIDIA
import torch
import torch.nn.functional as F
from task import input_t, output_t
def custom_kernel(data: input_t) -> output_t:
a, b1, b2, _, _, _, sfa, sfb1, sfb2, _ = data
out1 = torch._scaled_mm(
a[..., 0],
b1[..., 0].T,
sfa.permute(5, 2, 4, 0, 1, 3).view(-1),
sfb1.permute(5, 2, 4, 0, 1, 3).view(-1),
out_dtype=torch.float32,
)
out2 = torch._scaled_mm(
a[..., 0],
b2[..., 0].T,
sfa.permute(5, 2, 4, 0, 1, 3).view(-1),
sfb2.permute(5, 2, 4, 0, 1, 3).view(-1),
out_dtype=torch.float32,
)
out = F.silu(out1) * out2
return out.half().unsqueeze(-1)
Source code from GPU Mode and the KernelBot dataset · June 9 Researcher Reciprocity License v1.0
Best evidence level for this revision: reported
JSON