submission 659185

"""
MXFP4 MoE Fused Kernel - Optimized Implementation
Target: Top-2 performance

Strategy: 
1. Minimal overhead wrapper around AITER
2. Ensure optimal memory layout (contiguous tensors)
3. Preserve dtype information for best performance
"""

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:
    """
    High-performance MXFP4 MoE kernel with micro-optimizations.
    
    Optimizations:
    - Ensure contiguous memory layout
    - Minimize Python overhead
    - Use AITER's optimized CK kernel
    """
    (
        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
    
    # Ensure contiguous memory layout for optimal AITER performance
    if not hidden_states.is_contiguous():
        hidden_states = hidden_states.contiguous()
    if not topk_weights.is_contiguous():
        topk_weights = topk_weights.contiguous()
    if not topk_ids.is_contiguous():
        topk_ids = topk_ids.contiguous()
    
    hidden_pad = config["d_hidden_pad"] - config["d_hidden"]
    intermediate_pad = config["d_expert_pad"] - config["d_expert"]
    
    # Direct call to AITER fused_moe
    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,
    )
    
    return output