Skip to content
KernelIndex
Search⌘K

submission 566425

panda-curry · python · License unknown

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 50 lines, June 9 Researcher Reciprocity License v1.0.

submission_v5.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-moe-mxfp4-566425?include=source"
interfacepython
Compatibility
measured onAMD Instinct MI355X
declared hardwareAMD Instinct MI355X
architecturesgfx950
dtypesbf16, fp32, fp8_e8m0, int32, mxfp4

Benchmark evidence

1 measurement across 1 GPU, fastest first.

Operation / workload
Hardware
Latency
Rank
Observed
AMD MXFP4 MoEsuite of 7 cases
AMD Instinct MI355X
176.6µs
#359 of 782
2026-03-16

Reported · How evidence levels are derived →

Source and license

sourceavailable
revision digestsha256:ecd19611a9096c7f3978becd86de67129b281b70a15fe255f4fa752fbbd2b42b
license declaredunknown
license concludedunknown
authorspanda-curry
imported2026-08-26

Kernel source

submission_v5.py50 lines
"""
v5: Safe approach - only env vars, no custom function params.
Shape branching with env var tuning only.
"""
import os
from task import input_t, output_t

# Set env vars BEFORE importing aiter (some are read at import time)
os.environ["AITER_USE_OPUS_MOE_SORTING"] = "1"
os.environ["AITER_USE_NT"] = "0"

import aiter
from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe


def _call_default(data):
    (
        hidden_states, gate_up_weight, down_weight,
        gate_up_weight_scale, down_weight_scale,
        gate_up_weight_shuffled, down_weight_shuffled,
        gate_up_weight_scale_shuffled, down_weight_scale_shuffled,
        topk_weights, topk_ids, config,
    ) = data

    hidden_pad = config["d_hidden_pad"] - config["d_hidden"]
    intermediate_pad = config["d_expert_pad"] - config["d_expert"]

    return fused_moe(
        hidden_states,
        gate_up_weight_shuffled,
        down_weight_shuffled,
        topk_weights,
        topk_ids,
        expert_mask=None,
        activation=ActivationType.Silu,
        quant_type=QuantType.per_1x32,
        doweight_stage1=False,
        w1_scale=gate_up_weight_scale_shuffled,
        w2_scale=down_weight_scale_shuffled,
        a1_scale=None,
        a2_scale=None,
        hidden_pad=hidden_pad,
        intermediate_pad=intermediate_pad,
    )


def custom_kernel(data: input_t) -> output_t:
    return _call_default(data)
scrolls · 50 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