submission 632161
Hamza · python · License unknown
Use it
Vendorable · source mirrored · license unknownView source →
No package. Vendor the mirrored source: 64 lines, June 9 Researcher Reciprocity License v1.0.
submission.py
curl "https://kernelindex.com/api/v1/implementations/kernelbot-amd-moe-mxfp4-632161?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
Reported · How evidence levels are derived →
Source and license
sourceavailable
revision digestsha256:486926ed858c3648c7947d56e92bc283fdbe3f67a8a4e001a5594345e6a9eacb
license declaredunknown
license concludedunknown
authorsHamza
imported2026-08-15
Kernel source
submission.py64 lines
#!POPCORN leaderboard amd-moe-mxfp4
#!POPCORN gpu MI355X
import os
import torch
from task import input_t, output_t
from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe
_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]
# Per-shape kernel backend selection via env vars.
# get_2stage_cfgs() and get_ksplit() are LRU-cached per shape,
# so env vars only matter on first (warmup) call.
if M <= 128:
# Trigger cktile backend with split_k=2 for decode shapes.
# BYPASS_TUNE_CONFIG=1 skips CSV kernel selection, forcing
# the default heuristic path which respects AITER_KSPLIT.
os.environ["AITER_KSPLIT"] = "2"
os.environ["AITER_BYPASS_TUNE_CONFIG"] = "1"
else:
# Keep CSV-tuned ck2stages for large-M 256-expert shapes,
# and default ck backend for large-M 32-expert shapes.
os.environ["AITER_KSPLIT"] = "0"
os.environ["AITER_BYPASS_TUNE_CONFIG"] = "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)
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 · 64 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 630031.
⋯ 28 unchanged linesM = topk_ids.shape[0]- # For small batch sizes, ksplit=2 triggers the cktile backend- # which is ~30% faster for 32-expert shapes with M<=128.- # For large M, the default ck backend is better.- # get_ksplit() is LRU-cached per shape, so the env var only- # matters on the first (warmup) call for each unique shape.+ # Per-shape kernel backend selection via env vars.+ # get_2stage_cfgs() and get_ksplit() are LRU-cached per shape,+ # so env vars only matter on first (warmup) call.if M <= 128:+ # Trigger cktile backend with split_k=2 for decode shapes.+ # BYPASS_TUNE_CONFIG=1 skips CSV kernel selection, forcing+ # the default heuristic path which respects AITER_KSPLIT.os.environ["AITER_KSPLIT"] = "2"+ os.environ["AITER_BYPASS_TUNE_CONFIG"] = "1"else:+ # Keep CSV-tuned ck2stages for large-M 256-expert shapes,+ # and default ck backend for large-M 32-expert shapes.os.environ["AITER_KSPLIT"] = "0"+ os.environ["AITER_BYPASS_TUNE_CONFIG"] = "0"cfg_id = id(config)cached = _PAD_CACHE.get(cfg_id)
scrolls · 26 diff lines total
Best evidence level for this revision: reported
JSON