submission 643826 · KernelIndex

Use it

Vendorable · source mirrored · license unknownView source →

No package. Vendor the mirrored source: 69 lines, June 9 Researcher Reciprocity License v1.0.

submission.py

curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-moe-mxfp4-643826?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

149.0µs

#171 of 782

2026-03-27

Run →

Reported · How evidence levels are derived →

Source and license

sourceavailable

revision digestsha256:002cab3f944d3811be0e9bda676f22a8d5c44bababe3a8f0fe12fe5897ca4ac0

license declaredunknown

license concludedunknown

authorsHamza

imported2026-08-15

Kernel source

submission.py69 lines

#!POPCORN leaderboard amd-moe-mxfp4
#!POPCORN gpu MI355X

import os
import sys
import torch
from task import input_t, output_t

# Import WITHOUT setting AITER_CONFIG_FMOE — let aiter use its defaults
from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe
import aiter.fused_moe as _fm

# Append shape 6 tuned entry (expert=33 not in DSV3 CSV)
_tune_file = _fm.AITER_CONFIGS.AITER_CONFIG_FMOE_FILE
try:
    with open(_tune_file, "a") as f:
        # Shape 6: 32exp, M=512, inter=512, block_m=32 — both kernels 0.0% error
        f.write("256,512,7168,512,33,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,0,moe_ck2stages_gemm1_256x32x128x128_1x4_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16,0.0%,0,moe_ck2stages_gemm2_64x32x32x128_1x1_MulABScaleExpertWeightShuffled_v1_Nswizzle0_Quant3_MulRoutedWeight1_FP4X2_FP4X2_B16,0.0%,0,0,0,0\n")
    print(f"[moe] Appended shape 6 entry to {_tune_file}", file=sys.stderr)
except Exception as e:
    print(f"[moe] Failed: {e}", file=sys.stderr)

_PAD_CACHE = {}


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

    M = topk_ids.shape[0]

    cfg_id = id(config)
    cached = _PAD_CACHE.get(cfg_id)
    if cached is not None:
        hidden_pad, intermediate_pad = cached
    else:
        hidden_pad = config["d_hidden_pad"] - config["d_hidden"]
        intermediate_pad = config["d_expert_pad"] - config["d_expert"]
        _PAD_CACHE[cfg_id] = (hidden_pad, intermediate_pad)

    if M <= 128:
        os.environ["AITER_KSPLIT"] = "2"
        os.environ["AITER_BYPASS_TUNE_CONFIG"] = "1"
    else:
        os.environ["AITER_KSPLIT"] = "0"
        os.environ["AITER_BYPASS_TUNE_CONFIG"] = "0"

    return fused_moe(
        hidden_states, gate_up_weight_shuffled, down_weight_shuffled,
        topk_weights, topk_ids,
        activation=ActivationType.Silu, quant_type=QuantType.per_1x32,
        w1_scale=gate_up_weight_scale_shuffled,
        w2_scale=down_weight_scale_shuffled,
        hidden_pad=hidden_pad, intermediate_pad=intermediate_pad,
    )

scrolls · 69 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 634359.

  ⋯ 1 unchanged lines
  #!POPCORN gpu MI355X
   
  import os
+ import sys
  import torch
  from task import input_t, output_t
   
+ # Import WITHOUT setting AITER_CONFIG_FMOE — let aiter use its defaults
  from aiter import ActivationType, QuantType
  from aiter.fused_moe import fused_moe
+ import aiter.fused_moe as _fm
   
+ # Append shape 6 tuned entry (expert=33 not in DSV3 CSV)
+ _tune_file = _fm.AITER_CONFIGS.AITER_CONFIG_FMOE_FILE
+ try:
+     with open(_tune_file, "a") as f:
+         # Shape 6: 32exp, M=512, inter=512, block_m=32 — both kernels 0.0% error
+         f.write("256,512,7168,512,33,9,ActivationType.Silu,torch.bfloat16,torch.float4_e2m1fn_x2,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,0,moe_ck2stages_gemm1_256x32x128x128_1x4_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16,0.0%,0,moe_ck2stages_gemm2_64x32x32x128_1x1_MulABScaleExpertWeightShuffled_v1_Nswizzle0_Quant3_MulRoutedWeight1_FP4X2_FP4X2_B16,0.0%,0,0,0,0\n")
+     print(f"[moe] Appended shape 6 entry to {_tune_file}", file=sys.stderr)
+ except Exception as e:
+     print(f"[moe] Failed: {e}", file=sys.stderr)
+  
  _PAD_CACHE = {}
   
   
  ⋯ 24 unchanged lines
          intermediate_pad = config["d_expert_pad"] - config["d_expert"]
          _PAD_CACHE[cfg_id] = (hidden_pad, intermediate_pad)
   
-     # Per-shape backend selection via env vars (LRU-cached on first call per shape).
-     # M<=128: cktile (ksplit=2) skips fp4 quantization entirely.
-     # M>128: default ck2stages with CSV-tuned kernels.
      if M <= 128:
          os.environ["AITER_KSPLIT"] = "2"
          os.environ["AITER_BYPASS_TUNE_CONFIG"] = "1"
-         # For M=128 with fewer experts (<=64), block_m=32 halves block count
-         # vs cktile default 16, improving CU utilization.
-         num_experts = gate_up_weight_shuffled.shape[0]
-         block_size_M = 32 if (M > 16 and num_experts <= 64) else None
      else:
          os.environ["AITER_KSPLIT"] = "0"
          os.environ["AITER_BYPASS_TUNE_CONFIG"] = "0"
-         block_size_M = None
   
      return fused_moe(
          hidden_states, gate_up_weight_shuffled, down_weight_shuffled,
  ⋯ 2 unchanged lines
          w1_scale=gate_up_weight_scale_shuffled,
          w2_scale=down_weight_scale_shuffled,
          hidden_pad=hidden_pad, intermediate_pad=intermediate_pad,
-         block_size_M=block_size_M,
      )

scrolls · 53 diff lines total

Best evidence level for this revision: reported

JSON