Skip to content
KernelIndex
Search⌘K
Sign in

submission 566632

Aniket Sadashiva · python · License unknown
Kernel source · 50 lines ↓

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 50 lines, June 9 Researcher Reciprocity License v1.0.

submission_v5.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-moe-mxfp4-566632?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
185.1µs
#592 of 782
2026-03-16
Run →

Reported · How evidence levels are derived →

Source and license

sourceavailable
revision digestsha256:aa2f7e6d866365ff5f0b5744e955b182e752ea72a590251c1631d48c53973aac
license declaredunknown
license concludedunknown
authorsAniket Sadashiva
imported2026-08-15

Kernel source

submission_v5.py50 lines
"""
v5: Safe approach - only env vars, no custom function params.
Shape branching with env var tuning only.
"""
import os
from task import input_t, output_t

# Set env vars BEFORE importing aiter (some are read at import time)
os.environ["AITER_USE_OPUS_MOE_SORTING"] = "1"
os.environ["AITER_USE_NT"] = "0"

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


def _call_default(data):
    (
        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"]

    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,
    )


def custom_kernel(data: input_t) -> output_t:
    return _call_default(data)
scrolls · 50 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 566630.

  """
- First baseline submission: directly call AITER fused_moe.
- This matches the reference implementation exactly.
+ v5: Safe approach - only env vars, no custom function params.
+ Shape branching with env var tuning only.
  """
+ import os
  from task import input_t, output_t
   
+ # Set env vars BEFORE importing aiter (some are read at import time)
+ os.environ["AITER_USE_OPUS_MOE_SORTING"] = "1"
+ os.environ["AITER_USE_NT"] = "0"
+  
  import aiter
  from aiter import ActivationType, QuantType
  from aiter.fused_moe import fused_moe
   
   
- def custom_kernel(data: input_t) -> output_t:
+ def _call_default(data):
      (
-         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,
+         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(
+     return fused_moe(
          hidden_states,
          gate_up_weight_shuffled,
          down_weight_shuffled,
  ⋯ 11 unchanged lines
          intermediate_pad=intermediate_pad,
      )
   
-     return output
+  
+ def custom_kernel(data: input_t) -> output_t:
+     return _call_default(data)
scrolls · 56 diff lines total

Best evidence level for this revision: reported

JSON