submission 614253
KingOfZhao · python · License unknown
Use it
Vendorable · source mirrored · license unknownView source →
No package. Vendor the mirrored source: 26 lines, June 9 Researcher Reciprocity License v1.0.
mx_fp4_mm_optimized.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-mxfp4-mm-614253?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:ed476c3c27d93bcf569830180c7c39d864572620ad35da0ee0e7bca1d6ab8f7e
license declaredunknown
license concludedunknown
authorsKingOfZhao
imported2026-08-26
Techniques
Extracted from the mirrored source by pattern, never inferred. Each row cites its line.
fp4
MXFP4 GEMM — Zhao DylanKernel source
mx_fp4_mm_optimized.py26 lines
#!POPCORN leaderboard amd-mxfp4-mm
#!POPCORN gpu MI355X
"""
MXFP4 GEMM — Zhao Dylan
优化: 模块级import, 消除冗余contiguous/变量, 内联quant, 最小Python开销
"""
from task import input_t, output_t
import aiter
from aiter import dtypes
from aiter.ops.triton.quant import dynamic_mxfp4_quant
from aiter.utility.fp4_utils import e8m0_shuffle
def custom_kernel(data: input_t) -> output_t:
A, B, B_q, B_shuffle, B_scale_sh = data
# MXFP4 动态量化 A (Triton kernel)
A_fp4, A_scale = dynamic_mxfp4_quant(A)
# dtype转换 + scale shuffle (为CK GEMM tile对齐)
A_q = A_fp4.view(dtypes.fp4x2)
A_scale_sh = e8m0_shuffle(A_scale).view(dtypes.fp8_e8m0)
# CK gemm_a4w4: MXFP4×MXFP4 → bf16
return aiter.gemm_a4w4(
A_q, B_shuffle, A_scale_sh, B_scale_sh,
dtype=dtypes.bf16, bpreshuffle=True,
)
Source code from GPU Mode and the KernelBot dataset · June 9 Researcher Reciprocity License v1.0
Best evidence level for this revision: reported
JSON