Skip to content
KernelIndex
Search⌘K

submission 643445

anairdrop · python · License unknown

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 58 lines, June 9 Researcher Reciprocity License v1.0.

submission_g3fast.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-moe-mxfp4-643445?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
177.9µs
#399 of 782
2026-03-27

Reported · How evidence levels are derived →

Source and license

sourceavailable
revision digestsha256:aa570ab6a4634b90e45b5fb78e1f826c6572ff497da7945df7108b8f72fd3ff4
license declaredunknown
license concludedunknown
authorsanairdrop
imported2026-08-26

Techniques

Extracted from the mirrored source by pattern, never inferred. Each row cites its line.

split-kfamily_config.update({"block_size_M": 64, "splitk": 2})

Kernel source

submission_g3fast.py58 lines
import torch
import sys
import os
from task import input_t, output_t
from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe

# Build Principle: Explicit Intent, Log Once
_families_seen = set()

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

    # 1. Shape Metadata
    bs = config["bs"]
    E = gate_up_weight_shuffled.shape[0]
    d_expert = config["d_expert"]
    est_m = (bs * config["total_top_k"]) / E
    
    family_config = {"moe_sorting_dispatch_policy": False}

    # 2. Specialized Tiered Dispatch
    if d_expert >= 2048:
        # TARGET: Blocker Family. Hypothesis: Tile-64 beats Tile-128.
        family_name = "BLOCKER_TILE_64"
        family_config.update({"block_size_M": 64, "splitk": 2})
    elif est_m < 50:
        # TARGET: Sparse family. Goal: Maintain 87us result.
        family_name = "SPARSE_LATENCY"
        family_config.update({"block_size_M": 32, "splitk": 2})
    else:
        # TARGET: Dense family. Standard occupancy.
        family_name = "DENSE_THROUGHPUT"
        family_config.update({"block_size_M": 64, "splitk": 2})

    # Log intent once to avoid I/O jitter
    if family_name not in _families_seen:
        _families_seen.add(family_name)
        print(f"[INTENT] {family_name} triggered for d_expert={d_expert}, est_m={est_m:.1f}", file=sys.stderr)

    return fused_moe(
        hidden_states, gate_up_weight_shuffled, down_weight_shuffled,
        topk_weights, topk_ids,
        activation=ActivationType.Silu,
        quant_type=QuantType.per_1x32,
        doweight_stage1=False,
        w1_scale=gate_up_weight_scale_shuffled,
        w2_scale=down_weight_scale_shuffled,
        hidden_pad=config["d_hidden_pad"] - config["d_hidden"],
        intermediate_pad=config["d_expert_pad"] - config["d_expert"],
        **family_config,
    )
scrolls · 58 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