Skip to content
KernelIndex
Search⌘K

submission 729944

thehimalayanleo · python · License unknown

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 72 lines, June 9 Researcher Reciprocity License v1.0.

solution.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-moe-mxfp4-729944?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
185.1µs
#596 of 782
2026-04-05

Reported · How evidence levels are derived →

Source and license

sourceavailable
revision digestsha256:ead3694635e0056daaf9fb8a3d1e5f1aea4e9d7fa7b41c46bec26bf02e2b5380
license declaredunknown
license concludedunknown
authorsthehimalayanleo
imported2026-08-26

Kernel source

solution.py72 lines
#!POPCORN leaderboard amd-moe-mxfp4
#!POPCORN gpu MI355X
import torch
from task import input_t, output_t
from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe

def custom_kernel(data: input_t) -> output_t:
    # Defensive unpack -- handles both 8-field and 12-field versions
    if len(data) == 12:
        (
            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
    elif len(data) == 8:
        (
            hidden_states,
            gate_up_weight_shuffled,
            down_weight_shuffled,
            gate_up_weight_scale_shuffled,
            down_weight_scale_shuffled,
            topk_weights,
            topk_ids,
            config,
        ) = data
        gate_up_weight = gate_up_weight_shuffled
        down_weight = down_weight_shuffled
        gate_up_weight_scale = gate_up_weight_scale_shuffled
        down_weight_scale = down_weight_scale_shuffled
    else:
        # Last resort: unpack whatever we have positionally
        hidden_states = data[0]
        gate_up_weight_shuffled = data[1]
        down_weight_shuffled = data[2]
        gate_up_weight_scale_shuffled = data[3]
        down_weight_scale_shuffled = data[4]
        topk_weights = data[-3]
        topk_ids = data[-2]
        config = data[-1]

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

    output = 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,
    )
    return output
scrolls · 72 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