submission 549610
WangRongsheng · python · License unknown
Use it
Vendorable · source mirrored · license unknownView source →
No package. Vendor the mirrored source: 38 lines, June 9 Researcher Reciprocity License v1.0.
submission_fp8mm.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-mxfp4-mm-549610?include=source"interfacepython
Compatibility
measured onAMD Instinct MI355X
declared hardwareAMD Instinct MI355X
architecturesgfx950
dtypesbf16, mxfp4
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 digestsha256:8fdb19d133095acd5917f16213c991fa72a6f5487dac64993ce4985107a9858b
license declaredunknown
license concludedunknown
authorsWangRongsheng
imported2026-08-26
Kernel source
submission_fp8mm.py38 lines
# MXFP4 Quant + GEMM kernel — v6: env tuning
# Reference: aiter gemm_a4w4
#!POPCORN leaderboard amd-mxfp4-mm
#!POPCORN gpu MI355X
import os
os.environ["HIP_FORCE_DEV_KERNARG"] = "1"
os.environ["GPU_MAX_HW_QUEUES"] = "1"
os.environ["HSA_NO_SCRATCH_RECLAIM"] = "1"
import torch
import aiter
from task import input_t, output_t
from aiter import dtypes
from aiter.ops.triton.quant import dynamic_mxfp4_quant
from aiter.utility.fp4_utils import e8m0_shuffle
def _quant_mxfp4(x):
x_fp4, bs_e8m0 = dynamic_mxfp4_quant(x)
bs_e8m0 = e8m0_shuffle(bs_e8m0)
return x_fp4.view(dtypes.fp4x2), bs_e8m0.view(dtypes.fp8_e8m0)
def custom_kernel(data: input_t) -> output_t:
A, B, B_q, B_shuffle, B_scale_sh = data
if not A.is_contiguous():
A = A.contiguous()
A_q, A_scale_sh = _quant_mxfp4(A)
out = aiter.gemm_a4w4(
A_q, B_shuffle, A_scale_sh, B_scale_sh,
dtype=dtypes.bf16, bpreshuffle=True,
)
return out
scrolls · 38 lines total
Source code from GPU Mode and the KernelBot dataset · June 9 Researcher Reciprocity License v1.0
Best evidence level for this revision: reported
JSON