submission 650567

"""Optimized submission: per-shape tuned block_size_M."""
from task import input_t, output_t
import torch
import aiter
from aiter import ActivationType, QuantType
from aiter.fused_moe import fused_moe


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"]

    bs = config["bs"]
    n_routed = config["n_routed_experts"]

    # Per-shape block_size_M tuning (verified on online benchmark)
    if n_routed >= 256:
        block_m = 32  # E=256: only 32 works (64/128 hang or fail)
    elif bs >= 512:
        block_m = 128  # E=32, large bs: 128 is best
    else:
        block_m = 32  # E=32, small/medium bs: 32 slightly better than heuristic's 64

    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