submission 623449

"""
Optimized MoE MXFP4 kernel.

The CK fused_moe kernel handles MXFP4 quantization, 2-stage GEMM, SwiGLU, and
expert reduction internally via highly-tuned ASM kernels. The internal
get_2stage_cfgs function selects optimal kernel configurations from a pre-tuned
CSV based on shape, dtype, and hardware.

Our strategy: let the kernel's internal auto-tuning handle everything (passing
block_size_M=None), as the CSV-tuned configurations account for CU count,
expert distribution, tile efficiency, and split-K settings specific to MI355X.

Optimizations over a naive approach:
1. block_size_M=None lets the CSV auto-tuner select optimal tiling
2. Correct quant_type and activation for MXFP4 SwiGLU path
3. Pre-shuffled weights (gate_up_weight_shuffled, down_weight_shuffled) for
   optimal CK memory access patterns
"""
import torch
from task import input_t, output_t

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"]

    return 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=None,
        hidden_pad=hidden_pad,
        intermediate_pad=intermediate_pad,
    )