submission 646374 · KernelIndex

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 71 lines, June 9 Researcher Reciprocity License v1.0.

Submission_v1_fp8_baseline.py

curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-moe-mxfp4-646374?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

#454 of 782

2026-03-27

Run →

Reported · How evidence levels are derived →

Source and license

sourceavailable

revision digestsha256:72391c414930c5859a2cf8f990f92b5dd38d0f87936a99f6edea473fb6fba9da

license declaredunknown

license concludedunknown

authors.jonnss

imported2026-08-15

Techniques

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

fp4Submission template for DeepSeek-R1 MXFP4 MoE kernel.

Kernel source

Submission_v1_fp8_baseline.py71 lines

import os
import torch
from task import input_t, output_t

os.environ["VLLM_MOE_WTYPE"] = "fp8"
os.environ["VLLM_QUANT_OVERRIDE"] = "0"

from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe


@torch.inference_mode()
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
    """
    (
        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"]

    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 · 71 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 595191.

- #!POPCORN leaderboard amd-moe-mxfp4
+ import os
  import torch
- from typing import Dict
  from task import input_t, output_t
   
+ os.environ["VLLM_MOE_WTYPE"] = "fp8"
+ os.environ["VLLM_QUANT_OVERRIDE"] = "0"
+  
  from aiter import ActivationType, QuantType
  from aiter.fused_moe import fused_moe
   
   
+ @torch.inference_mode()
  def custom_kernel(data: input_t) -> output_t:
      """
-     Safe baseline: direct AITER fused_moe invocation with pre-shuffled MXFP4 weights.
+     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
      """
      (
          hidden_states,
  ⋯ 13 unchanged lines
      hidden_pad = config["d_hidden_pad"] - config["d_hidden"]
      intermediate_pad = config["d_expert_pad"] - config["d_expert"]
   
-     return fused_moe(
+     output = fused_moe(
          hidden_states,
          gate_up_weight_shuffled,
          down_weight_shuffled,
  ⋯ 10 unchanged lines
          hidden_pad=hidden_pad,
          intermediate_pad=intermediate_pad,
      )
+  
+     return output

scrolls · 53 diff lines total

Best evidence level for this revision: reported

JSON