submission 680990

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

"""
Safe, benchmark-friendly submission for `amd-mxfp4-mm`.

Design goal:
- keep exactly the same fast path as the known-good baseline
- remove only two clearly unnecessary costs:
  1) per-call imports
  2) touching B / B_q even though gemm_a4w4 only needs A, B_shuffle, B_scale_sh

Anything more aggressive than this should be done as real AITER/CK kernel tuning,
not wrapper-level Python changes.
"""

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: torch.Tensor):
    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_shuffle, B_scale_sh = data
    if not A.is_contiguous():
        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,
    )