mm_fp4_nvfp4_flashinfer_n4096_k4096
FlashInfer-Bench baselines · python · Apache-2.0
Kernel source · 17 lines ↓holds 3 records
Use it
Vendorable · source mirrored · Apache-2.0View source →
No package. Vendor the mirrored source: 17 lines, Apache-2.0, pinned at da91508.
main.py
curl "https://kernelindex.com/api/v1/implementations/flashinfer-mm-fp4-nvfp4-flashinfer-n4096-k4096?include=source"interfacepython
revisionda915083d4c7
symbolrun
pathmain.py
Compatibility
measured onNVIDIA B200
declared hardwareNVIDIA B200
architecturessm_100
dtypesfp32, int8
Benchmark evidence
9 measurements across 1 GPU, fastest first.
Operation / workload
Hardware
Latency
Rank
Observed
Reported · How evidence levels are derived →
Source and license
sourcehttps://huggingface.co/datasets/flashinfer-ai/flashinfer-trace
commitda915083d4c7c5e61aa3005e3d17ae488e0fc71c
revision digestsha256:42b3c7cb46511ceff0e03b3bf3faa70972965f7f6b40657f085b5b6065c0acd6
license declaredApache-2.0
license concludedApache-2.0
authorsbaseline
imported2026-08-16
Kernel source
main.py17 lines
import torch
import flashinfer
def run(A_fp4, A_scale, B_fp4, B_scale, alpha):
m = A_fp4.shape[0]
n = B_fp4.shape[0]
out = torch.empty(m, n, device=A_fp4.device, dtype=torch.bfloat16)
alpha_t = torch.tensor(alpha, device=A_fp4.device, dtype=torch.float32)
flashinfer.mm_fp4(
A_fp4.view(torch.uint8), B_fp4.view(torch.uint8).T,
A_scale.view(torch.uint8), B_scale.view(torch.uint8).T,
alpha_t, torch.bfloat16, out,
block_size=16, use_8x4_sf_layout=False, backend="cutlass", use_nvfp4=True,
)
return out
Source code from FlashInfer-Bench (flashinfer-ai/flashinfer-trace) · Apache-2.0
Best evidence level for this revision: reported
JSON