submission 598572

  #!POPCORN leaderboard amd-moe-mxfp4
  #!POPCORN gpu MI355X
- """MoE MXFP4 — aiter fused_moe with auto-tuned defaults."""
- from task import input_t, output_t
+  
+ """
+ v16: Monkeypatch cfg_2stages with per-shape CK/FlyDSL configs.
+ Based on dgavriloff/amd-structkernel pattern (rank 6, 130us).
+ Injects configs directly into in-memory dict, registers FlyDSL kernels,
+ and monkeypatches get_2stage_cfgs for NT override.
+ """
+ import os
+ import functools
  import torch
+ from task import input_t, output_t
+  
+ from aiter import ActivationType, QuantType
+ from aiter.fused_moe import fused_moe
+ import aiter.fused_moe as _fused_moe_module
  import aiter
- from aiter import dtypes
- from aiter.fused_moe import fused_moe, ActivationType, QuantType
+ import aiter.ops.flydsl.moe_kernels as _flydsl_moe_kernels
   
+ # Register FlyDSL tile_k=128 kernels that aren't in server's default registration
+ _flydsl_moe_kernels._KERNEL_PARAMS["flydsl_moe2_afp4_wfp4_bf16_t32x128x128_atomic"] = {
+     "stage": 2, "a_dtype": "fp4", "b_dtype": "fp4", "out_dtype": "bf16",
+     "tile_m": 32, "tile_n": 128, "tile_k": 128, "mode": "atomic", "MPerBlock": 32,
+ }
+ _flydsl_moe_kernels._KERNEL_PARAMS["flydsl_moe2_afp4_wfp4_bf16_t32x256x128_atomic"] = {
+     "stage": 2, "a_dtype": "fp4", "b_dtype": "fp4", "out_dtype": "bf16",
+     "tile_m": 32, "tile_n": 256, "tile_k": 128, "mode": "atomic", "MPerBlock": 32,
+ }
+ _flydsl_moe_kernels._KERNEL_PARAMS["flydsl_moe2_afp4_wfp4_bf16_t16x256x128_atomic"] = {
+     "stage": 2, "a_dtype": "fp4", "b_dtype": "fp4", "out_dtype": "bf16",
+     "tile_m": 16, "tile_n": 256, "tile_k": 128, "mode": "atomic", "MPerBlock": 16,
+ }
+ _flydsl_moe_kernels._KERNEL_PARAMS["flydsl_moe2_afp4_wfp4_bf16_t16x128x128_atomic"] = {
+     "stage": 2, "a_dtype": "fp4", "b_dtype": "fp4", "out_dtype": "bf16",
+     "tile_m": 16, "tile_n": 128, "tile_k": 128, "mode": "atomic", "MPerBlock": 16,
+ }
+  
+ # Custom configs per shape
+ _CUSTOM_CONFIGS = {}
+  
+ def _make_key(token, inter_dim, expert):
+     return (
+         256, token, 7168, inter_dim, expert, 9,
+         "ActivationType.Silu", "torch.bfloat16",
+         "torch.float4_e2m1fn_x2", "torch.float4_e2m1fn_x2",
+         "QuantType.per_1x32", True, False,
+     )
+  
+ # === E=33 shapes ===
+ _CUSTOM_CONFIGS[_make_key(16, 512, 33)] = {
+     "block_m": 32, "ksplit": 2,
+     "kernelName1": "", "kernelName2": "",
+     "run_1stage": False,
+ }
+  
+ _4WG_STAGE1_M128 = "moe_ck2stages_gemm1_256x128x128x128_1x4_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16"
+ _FLYDSL_STAGE2_M16_N128_K128 = "flydsl_moe2_afp4_wfp4_bf16_t16x128x128_atomic"
+  
+ _CUSTOM_CONFIGS[_make_key(128, 512, 33)] = {
+     "block_m": 64, "ksplit": 0,
+     "kernelName1": _4WG_STAGE1_M128,
+     "kernelName2": _FLYDSL_STAGE2_M16_N128_K128,
+     "run_1stage": False,
+ }
+  
+ _CUSTOM_CONFIGS[_make_key(512, 512, 33)] = {
+     "block_m": 64, "ksplit": 0,
+     "kernelName1": _4WG_STAGE1_M128,
+     "kernelName2": _FLYDSL_STAGE2_M16_N128_K128,
+     "run_1stage": False,
+ }
+  
+ _CUSTOM_CONFIGS[_make_key(512, 2048, 33)] = {
+     "block_m": 64, "ksplit": 0,
+     "kernelName1": _4WG_STAGE1_M128,
+     "kernelName2": _FLYDSL_STAGE2_M16_N128_K128,
+     "run_1stage": False,
+ }
+  
+ # === E=257 shapes ===
+ _CUSTOM_CONFIGS[_make_key(16, 256, 257)] = {
+     "block_m": 16, "ksplit": 2,
+     "kernelName1": "", "kernelName2": "",
+     "run_1stage": False,
+ }
+  
+ _CUSTOM_CONFIGS[_make_key(128, 256, 257)] = {
+     "block_m": 16, "ksplit": 2,
+     "kernelName1": "", "kernelName2": "",
+     "run_1stage": False,
+ }
+  
+ _4WG_STAGE1_M32 = "moe_ck2stages_gemm1_256x32x128x128_1x4_MulABScaleShuffled_v3_Nswizzle0_Quant3_MulRoutedWeight0_silu_FP4X2_FP4X2_B16"
+ _CUSTOM_CONFIGS[_make_key(512, 256, 257)] = {
+     "block_m": 32, "ksplit": 0,
+     "kernelName1": _4WG_STAGE1_M32,
+     "kernelName2": _FLYDSL_STAGE2_M16_N128_K128,
+     "run_1stage": False,
+     "use_non_temporal_load": True,
+ }
+  
+ _injected = False
+  
+ def _inject_configs():
+     global _injected
+     if _injected:
+         return
+     _injected = True
+  
+     if _fused_moe_module.cfg_2stages is None:
+         import pandas as pd
+         from aiter.jit.core import AITER_CONFIGS
+         tune_file = AITER_CONFIGS.AITER_CONFIG_FMOE_FILE
+         if os.path.exists(tune_file):
+             _INDEX_COLS = [
+                 "cu_num", "token", "model_dim", "inter_dim", "expert", "topk",
+                 "act_type", "dtype", "q_dtype_a", "q_dtype_w", "q_type",
+                 "use_g1u1", "doweight_stage1",
+             ]
+             df = pd.read_csv(tune_file)
+             if "_tag" in df.columns:
+                 df = df[df["_tag"].fillna("") == ""]
+             _fused_moe_module.cfg_2stages = df.set_index(_INDEX_COLS).to_dict("index")
+         else:
+             _fused_moe_module.cfg_2stages = {}
+  
+     _fused_moe_module.cfg_2stages.update(_CUSTOM_CONFIGS)
+  
+     # Monkeypatch get_2stage_cfgs to support use_non_temporal_load from config
+     _original_get_2stage_cfgs = _fused_moe_module.get_2stage_cfgs
+  
+     @functools.lru_cache(maxsize=2048)
+     def _patched_get_2stage_cfgs(
+         token, model_dim, inter_dim, expert, topk,
+         dtype, q_dtype_a, q_dtype_w, q_type, use_g1u1,
+         activation, doweight_stage1, hidden_pad, intermediate_pad, is_shuffled=True,
+     ):
+         metadata = _original_get_2stage_cfgs(
+             token, model_dim, inter_dim, expert, topk,
+             dtype, q_dtype_a, q_dtype_w, q_type, use_g1u1,
+             activation, doweight_stage1, hidden_pad, intermediate_pad, is_shuffled,
+         )
+  
+         from aiter.jit.utils.chip_info import get_cu_num
+         cu_num = get_cu_num()
+         keys = (
+             cu_num, token, model_dim, inter_dim, expert, topk,
+             str(activation), str(dtype), str(q_dtype_a), str(q_dtype_w),
+             str(q_type), use_g1u1, doweight_stage1,
+         )
+         cfg = _fused_moe_module.cfg_2stages.get(keys)
+         if cfg and cfg.get("use_non_temporal_load") is not None:
+             nt = cfg["use_non_temporal_load"]
+             old_s1 = metadata.stage1
+             if hasattr(old_s1, 'func') and old_s1.func is not None:
+                 if 'use_non_temporal_load' in (old_s1.keywords or {}):
+                     new_kw = dict(old_s1.keywords)
+                     new_kw['use_non_temporal_load'] = nt
+                     metadata = _fused_moe_module.MOEMetadata(
+                         functools.partial(old_s1.func, **{k: v for k, v in new_kw.items()}),
+                         metadata.stage2,
+                         metadata.block_m,
+                         metadata.ksplit,
+                         metadata.run_1stage,
+                         metadata.has_bias,
+                         nt,
+                     )
+                     old_s2 = metadata.stage2
+                     if old_s2 and hasattr(old_s2, 'keywords') and 'use_non_temporal_load' in (old_s2.keywords or {}):
+                         new_kw2 = dict(old_s2.keywords)
+                         new_kw2['use_non_temporal_load'] = nt
+                         metadata = _fused_moe_module.MOEMetadata(
+                             metadata.stage1,
+                             functools.partial(old_s2.func, **{k: v for k, v in new_kw2.items()}),
+                             metadata.block_m,
+                             metadata.ksplit,
+                             metadata.run_1stage,
+                             metadata.has_bias,
+                             nt,
+                         )
+         return metadata
+  
+     _fused_moe_module.get_2stage_cfgs = _patched_get_2stage_cfgs
+  
+  
  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
+     (
+         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
   
-     result = fused_moe(
-         hidden_states=hidden_states,
-         w1=gate_up_weight_shuffled,
-         w2=down_weight_shuffled,
-         topk_weight=topk_weights,
-         topk_ids=topk_ids,
+     _inject_configs()
+  
+     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,
-         activation=ActivationType.Silu,
-         quant_type=QuantType.per_1x32,
-         dtype=torch.bfloat16,
+         a1_scale=None, a2_scale=None,
+         hidden_pad=hidden_pad, intermediate_pad=intermediate_pad,
      )
-     return result
+  
+     return output