submission 748766

#!POPCORN leaderboard amd-mxfp4-mm
#!POPCORN gpu MI355X

"""
MXFP4 GEMM — optimized reference path.
Removes unnecessary B.contiguous(), pre-allocates output.
"""

import torch
from task import input_t, output_t

import aiter
from aiter import dtypes
from aiter.ops.triton.quant import dynamic_mxfp4_quant
from aiter.utility.fp4_utils import e8m0_shuffle


def _quant_mxfp4(x):
    x_fp4, bs_e8m0 = dynamic_mxfp4_quant(x)
    bs_e8m0 = e8m0_shuffle(bs_e8m0)
    return x_fp4.view(dtypes.fp4x2), bs_e8m0.view(dtypes.fp8_e8m0)


def custom_kernel(data: input_t) -> output_t:
    A, B, B_q, B_shuffle, B_scale_sh = data
    # Skip B.contiguous() — not needed for gemm_a4w4
    A = A.contiguous()

    A_q, A_scale_sh = _quant_mxfp4(A)

    return aiter.gemm_a4w4(
        A_q,
        B_shuffle,
        A_scale_sh,
        B_scale_sh,
        dtype=dtypes.bf16,
        bpreshuffle=True,
    )