Skip to content
KernelIndex
Search⌘K
Sign in

submission 654068

lgc0338 · python · License unknown
Kernel source · 53 lines ↓

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 53 lines, June 9 Researcher Reciprocity License v1.0.

submission.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-moe-mxfp4-654068?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
185.3µs
#615 of 782
2026-03-28
Run →

Reported · How evidence levels are derived →

Source and license

sourceavailable
revision digestsha256:129271461d8dfa6201a892aaceb7d9605a7b1c57ea08b303526fb3d7f6ef83e7
license declaredunknown
license concludedunknown
authorslgc0338
imported2026-08-15

Techniques

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

persistent-kernel2. Pre-allocated persistent buffers — avoid torch.empty per call
split-k3. splitk for E=33 small batch decode

Kernel source

submission.py53 lines
#!POPCORN leaderboard amd-moe-mxfp4
#!POPCORN gpu MI355X

"""
Python-level extreme optimization on CK baseline:
1. @torch.inference_mode() — disable autograd
2. Pre-allocated persistent buffers — avoid torch.empty per call
3. splitk for E=33 small batch decode
4. Try doweight_stage1=True
5. moe_sorting_dispatch_policy variations
"""

import torch
from task import input_t, output_t

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

@torch.inference_mode()
def custom_kernel(data: input_t) -> output_t:
    (
        hidden_states,
        _guw, _dw, _gus, _ds,
        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"]

    return 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,
    )
scrolls · 53 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 648658.

  #!POPCORN leaderboard amd-moe-mxfp4
  #!POPCORN gpu MI355X
   
+ """
+ Python-level extreme optimization on CK baseline:
+ 1. @torch.inference_mode() — disable autograd
+ 2. Pre-allocated persistent buffers — avoid torch.empty per call
+ 3. splitk for E=33 small batch decode
+ 4. Try doweight_stage1=True
+ 5. moe_sorting_dispatch_policy variations
+ """
+  
+ import torch
  from task import input_t, output_t
   
  from aiter import ActivationType, QuantType
  from aiter.fused_moe import fused_moe
   
-  
+ @torch.inference_mode()
  def custom_kernel(data: input_t) -> output_t:
      (
          hidden_states,
-         _gate_up_weight,        # raw — unused
-         _down_weight,           # raw — unused
-         _gate_up_weight_scale,  # raw — unused
-         _down_weight_scale,     # raw — unused
+         _guw, _dw, _gus, _ds,
          gate_up_weight_shuffled,
          down_weight_shuffled,
          gate_up_weight_scale_shuffled,
scrolls · 31 diff lines total

Best evidence level for this revision: reported

JSON