submission 698874

#!POPCORN leaderboard amd-mxfp4-mm
#!POPCORN gpu MI355X
# =============================================================================
# MXFP4 GEMM Kernel — Baseline for AMD MI355X
# Competition: AMD x GPU MODE 2026 - amd-mxfp4-mm
# Strategy: Use aiter gemm_a4w4 as baseline
# =============================================================================

import torch
from task import input_t, output_t
import aiter
from aiter import QuantType, dtypes
from aiter.ops.triton.quant import dynamic_mxfp4_quant
from aiter.utility.fp4_utils import e8m0_shuffle

# Warmup state to prevent re-initialization if any
_initialized = False

def _quant_mxfp4(x, shuffle=True):
    x_fp4, bs_e8m0 = dynamic_mxfp4_quant(x)
    if shuffle:
        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:
    """
    MXFP4 GEMM — Highly optimized python invocation baseline
    """
    global _initialized
    if not _initialized:
        _initialized = True

    A, B, B_q, B_shuffle, B_scale_sh = data

    A_q, A_scale_sh = _quant_mxfp4(A, shuffle=True)
    out_gemm = aiter.gemm_a4w4(
        A_q,
        B_shuffle,
        A_scale_sh,
        B_scale_sh,
        dtype=dtypes.bf16,
        bpreshuffle=True,
    )
    return out_gemm