submission 644487

#!POPCORN leaderboard amd-moe-mxfp4
#!POPCORN gpu MI355X
"""
MoE MXFP4 v3: aiter fused_moe with per-shape block_m tuning.

Optimal block_m found via self-bench sweep:
  E=33, bs=16,  de=512:  block_m=32  (matches heuristic)
  E=33, bs=128, de=512:  block_m=32  (heuristic chose 64, 3% faster)
  E=33, bs=512, de=512:  block_m=128 (heuristic chose 64, 3% faster)
  E=33, bs=512, de=2048: block_m=64  (heuristic chose 128, 3% faster)
  E=257 shapes: use CSV-tuned defaults (already optimal)

Note: CK kernel dispatch is static (no JIT). block_m selects pre-compiled variants.
"""
import torch
from task import input_t, output_t

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

# Per-shape block_m override table (from sweep results)
# Key: (n_routed_experts, bs, d_expert) -> optimal block_m
_BLOCK_M_TABLE = {
    (32, 16, 512): 32,    # matches heuristic
    (32, 128, 512): 32,   # heuristic=64, 3% faster
    (32, 512, 512): 128,  # heuristic=64, 3% faster
    (32, 512, 2048): 64,  # heuristic=128, 3% faster
}


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_pad = config["d_hidden_pad"] - config["d_hidden"]
    intermediate_pad = config["d_expert_pad"] - config["d_expert"]

    # Look up optimal block_m for this shape
    shape_key = (config["n_routed_experts"], config["bs"], config["d_expert"])
    block_m = _BLOCK_M_TABLE.get(shape_key, None)

    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,
        a1_scale=None,
        a2_scale=None,
        hidden_pad=hidden_pad,
        intermediate_pad=intermediate_pad,
        block_size_M=block_m,
    )

    return output