Skip to content
KernelIndex
Search⌘K

submission 572866

n8_gr8_ · python · License unknown

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 69 lines, June 9 Researcher Reciprocity License v1.0.

submission.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-moe-mxfp4-572866?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
155.1µs
#241 of 782
2026-03-17

Reported · How evidence levels are derived →

Source and license

sourceavailable
revision digestsha256:6c27595b9d9dc546c5757a3b94faa29f9b1eda03b715a885a81f8d1e48f92dbd
license declaredunknown
license concludedunknown
authorsn8_gr8_
imported2026-08-15

Kernel source

submission.py69 lines
#!POPCORN leaderboard amd-moe-mxfp4
#!POPCORN gpu MI355X
import os

os.environ["HIP_FORCE_DEV_KERNARG"] = "1"
os.environ["AITER_USE_OPUS_MOE_SORTING"] = "1"
os.environ["AITER_LOG_LEVEL"] = "ERROR"
os.environ["AITER_USE_NT"] = "1"

import torch
from task import input_t, output_t
from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe


@torch.inference_mode()
def custom_kernel(data: input_t) -> output_t:
    (
        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"]

    bs = config["bs"]
    n_routed = config["n_routed_experts"]
    d_expert = config["d_expert"]

    if bs >= 512:
        os.environ.pop("AITER_BYPASS_TUNE_CONFIG", None)
        if n_routed > 32:
            os.environ["AITER_KSPLIT"] = "2"
        elif d_expert >= 2048:
            os.environ["AITER_KSPLIT"] = "1"
        else:
            os.environ.pop("AITER_KSPLIT", None)
    elif bs >= 128:
        os.environ["AITER_BYPASS_TUNE_CONFIG"] = "1"
        if n_routed > 32:
            os.environ["AITER_KSPLIT"] = "4"
        else:
            os.environ["AITER_KSPLIT"] = "2"
    else:
        os.environ["AITER_BYPASS_TUNE_CONFIG"] = "1"
        if n_routed > 32:
            os.environ["AITER_KSPLIT"] = "4"
        else:
            os.environ["AITER_KSPLIT"] = "2"

    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,
    )
scrolls · 69 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