submission 746673 · KernelIndex

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 59 lines, June 9 Researcher Reciprocity License v1.0.

submission_v2.py

curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-moe-mxfp4-746673?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

178.2µs

#417 of 782

2026-04-06

Run →

Reported · How evidence levels are derived →

Source and license

sourceavailable

revision digestsha256:c196a091b473b4b1dd6d844d2659b53f96298f5efd90c4a5ad0d2fcb514f10fe

license declaredunknown

license concludedunknown

authorsbigmodel_wuzhigang

imported2026-08-15

Techniques

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

fp4MXFP4 MoE Optimization V2 - Optimized for MI355X.

Kernel source

submission_v2.py59 lines

#!POPCORN leaderboard amd-moe-mxfp4
#!POPCORN gpu MI355X
"""
MXFP4 MoE Optimization V2 - Optimized for MI355X.
Key optimizations:
1. Ensure proper padding and alignment
2. Minimize memory allocation
3. Optimize for DeepSeek-R1 architecture
"""
import torch
from typing import Dict
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:
    """
    Optimized MXFP4 MoE implementation.
    """
    (
        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"]

    # Use the pre-shuffled weights for best performance
    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 · 59 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 745875.

  #!POPCORN leaderboard amd-moe-mxfp4
  #!POPCORN gpu MI355X
+ """
+ MXFP4 MoE Optimization V2 - Optimized for MI355X.
+ Key optimizations:
+ 1. Ensure proper padding and alignment
+ 2. Minimize memory allocation
+ 3. Optimize for DeepSeek-R1 architecture
+ """
  import torch
  from typing import Dict
  from task import input_t, output_t
  ⋯ 4 unchanged lines
  
  def custom_kernel(data: input_t) -> output_t:
      """
-     Submission template for DeepSeek-R1 MXFP4 MoE kernel.
- 
-     Input data tuple:
-         hidden_states:                [M, d_hidden]                           bf16
-         gate_up_weight:               [E, 2*d_expert_pad, d_hidden_pad//2]    fp4x2  (raw)
-         down_weight:                  [E, d_hidden_pad, d_expert_pad//2]      fp4x2  (raw)
-         gate_up_weight_scale:         [E, 2*d_expert_pad, scale_K]            e8m0   (raw)
-         down_weight_scale:            [E, d_hidden_pad, scale_K]              e8m0   (raw)
-         gate_up_weight_shuffled:      [E, 2*d_expert_pad, d_hidden_pad//2]    fp4x2  (shuffled)
-         down_weight_shuffled:         [E, d_hidden_pad, d_expert_pad//2]      fp4x2  (shuffled)
-         gate_up_weight_scale_shuffled:[padded, flat]                          e8m0   (shuffled)
-         down_weight_scale_shuffled:   [padded, flat]                          e8m0   (shuffled)
-         topk_weights:                 [M, total_top_k]                        float32
-         topk_ids:                     [M, total_top_k]                        int32
-         config:                       dict
- 
-     Returns:
-         output: [M, d_hidden] bf16
+     Optimized MXFP4 MoE implementation.
      """
      (
          hidden_states,
  ⋯ 13 unchanged lines
      hidden_pad = config["d_hidden_pad"] - config["d_hidden"]
      intermediate_pad = config["d_expert_pad"] - config["d_expert"]
  
+     # Use the pre-shuffled weights for best performance
      output = fused_moe(
          hidden_states,
          gate_up_weight_shuffled,
  ⋯ 12 unchanged lines
          intermediate_pad=intermediate_pad,
      )
  
-     return output
+     return output
   No newline at end of file

scrolls · 53 diff lines total

Best evidence level for this revision: reported

JSON