submission 592660 · KernelIndex

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 32 lines, June 9 Researcher Reciprocity License v1.0.

submission.py

curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-moe-mxfp4-592660?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

179.6µs

#453 of 782

2026-03-19

Run →

Reported · How evidence levels are derived →

Source and license

sourceavailable

revision digestsha256:0fc9b30fb5c4d29419d07f1c8263c7842ad7bf3a7a1e6e720de931452417150d

license declaredunknown

license concludedunknown

authorsjohnny.t.shi

imported2026-08-15

Techniques

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

fp4"""MoE MXFP4 — aiter fused_moe with auto-tuned defaults."""

Kernel source

submission.py32 lines

#!POPCORN leaderboard amd-moe-mxfp4
#!POPCORN gpu MI355X
"""MoE MXFP4 — aiter fused_moe with auto-tuned defaults."""
from task import input_t, output_t
import torch
import aiter
from aiter import dtypes
from aiter.fused_moe import fused_moe, ActivationType, QuantType

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

    result = fused_moe(
        hidden_states=hidden_states,
        w1=gate_up_weight_shuffled,
        w2=down_weight_shuffled,
        topk_weight=topk_weights,
        topk_ids=topk_ids,
        w1_scale=gate_up_weight_scale_shuffled,
        w2_scale=down_weight_scale_shuffled,
        activation=ActivationType.Silu,
        quant_type=QuantType.per_1x32,
        dtype=torch.bfloat16,
    )
    return result

scrolls · 32 lines total

Source code from GPU Mode and the KernelBot dataset · June 9 Researcher Reciprocity License v1.0

Changes from previous submission

Against this author's previous submission submission 592373.

  #!POPCORN leaderboard amd-moe-mxfp4
  #!POPCORN gpu MI355X
- """MoE MXFP4 baseline — use aiter fused_moe directly."""
+ """MoE MXFP4 — aiter fused_moe with auto-tuned defaults."""
  from task import input_t, output_t
  import torch
  import aiter
  ⋯ 9 unchanged lines
       topk_weights, topk_ids,
       config) = data
   
-     M = hidden_states.shape[0]
-     d_hidden = config['d_hidden']
-     d_expert = config['d_expert']
-     n_routed = config['n_routed_experts']
-     n_shared = config['n_shared_experts']
-     total_top_k = config['total_top_k']
-  
-     # Use aiter's fused_moe with MXFP4 quantization
-     # gate_up_weight_shuffled is [E, 2*d_expert_pad, d_hidden_pad//2] fp4x2 (pre-shuffled)
-     # down_weight_shuffled is [E, d_hidden_pad, d_expert_pad//2] fp4x2 (pre-shuffled)
      result = fused_moe(
          hidden_states=hidden_states,
          w1=gate_up_weight_shuffled,

scrolls · 24 diff lines total

Best evidence level for this revision: reported

JSON