submission 583919

"""Optimized MoE MXFP4 submission for AMD Instinct MI355X.

Uses AITER's fused_moe with per-config block_size_M tuning.
The block_size_M controls the tile size for the CK/ASM kernel dispatch:
- For many-expert configs (E=257): smaller block_m=32 reduces padding waste
- For fewer-expert configs (E=33): larger block_m balances CU utilization
"""
from task import input_t, output_t
import torch

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"]

    # Select block_size_M based on problem shape
    E_total = config["n_routed_experts"] + config["n_shared_experts"]
    bs = config["bs"]
    d_expert = config["d_expert"]

    # For many-expert configs, block_m=32 minimizes padding waste
    # For fewer-expert configs, heuristic auto-selection is usually good
    if E_total > 64:
        block_m = 32
    elif d_expert >= 2048:
        block_m = 64
    elif bs >= 128:
        block_m = 128
    else:
        block_m = 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,
        block_size_M=block_m,
        hidden_pad=hidden_pad,
        intermediate_pad=intermediate_pad,
    )

    return output