submission 594273

"""
MI355X MLA decode – Hybrid MXFP4 (long seq) + FP8 fallback (exact reference dequant)
"""
import torch
from task import input_t, output_t
from utils import make_match_reference
from aiter.mla import mla_decode_fwd
from aiter import dtypes as aiter_dtypes
from aiter import get_mla_metadata_info_v1, get_mla_metadata_v1
from aiter.utility.fp4_utils import mxfp4_to_f32, e8m0_to_f32

NUM_HEADS = 16
NUM_KV_HEADS = 1
QK_HEAD_DIM = 576
V_HEAD_DIM = 512
SM_SCALE = 1.0 / (QK_HEAD_DIM ** 0.5)
PAGE_SIZE = 1
NUM_KV_SPLITS = 32
FP8_DTYPE = aiter_dtypes.fp8

def quantize_fp8(tensor: torch.Tensor):
    finfo = torch.finfo(FP8_DTYPE)
    amax = tensor.abs().amax().clamp(min=1e-12)
    scale = amax / finfo.max
    fp8 = (tensor / scale).clamp(min=finfo.min, max=finfo.max).to(FP8_DTYPE)
    return fp8, scale.to(torch.float32).reshape(1)

def dequantize_mxfp4(fp4_data, scale_e8m0, orig_shape):
    """Exact reference dequant (verbatim from problem statement)"""
    B, M, N = orig_shape
    num_rows = B * M
    block_size = 32
    num_blocks = N // block_size

    fp4_2d = fp4_data.reshape(num_rows, N // 2)
    float_vals = mxfp4_to_f32(fp4_2d)

    scale_f32 = e8m0_to_f32(scale_e8m0)[:num_rows, :num_blocks]
    float_vals_blocked = float_vals.view(num_rows, num_blocks, block_size)
    scaled = float_vals_blocked * scale_f32.unsqueeze(-1)
    return scaled.view(B, M, N).to(torch.bfloat16)

def _make_mla_decode_metadata(
    batch_size: int,
    max_q_len: int,
    nhead: int,
    nhead_kv: int,
    q_dtype: torch.dtype,
    kv_dtype: torch.dtype,
    qo_indptr: torch.Tensor,
    kv_indptr: torch.Tensor,
    kv_last_page_len: torch.Tensor,
    num_kv_splits: int = NUM_KV_SPLITS,
):
    info = get_mla_metadata_info_v1(
        batch_size, max_q_len, nhead, q_dtype, kv_dtype,
        is_sparse=False, fast_mode=False,
        num_kv_splits=num_kv_splits, intra_batch_mode=True,
    )
    work = [torch.empty(s, dtype=t, device="cuda") for s, t in info]
    (work_metadata, work_indptr, work_info_set,
     reduce_indptr, reduce_final_map, reduce_partial_map) = work

    get_mla_metadata_v1(
        qo_indptr, kv_indptr, kv_last_page_len,
        nhead // nhead_kv,
        nhead_kv,
        True,
        work_metadata, work_info_set, work_indptr,
        reduce_indptr, reduce_final_map, reduce_partial_map,
        page_size=PAGE_SIZE,
        kv_granularity=max(PAGE_SIZE, 16),
        max_seqlen_qo=max_q_len,
        uni_seqlen_qo=max_q_len,
        fast_mode=False,
        max_split_per_batch=num_kv_splits,
        intra_batch_mode=True,
        dtype_q=q_dtype,
        dtype_kv=kv_dtype,
    )
    return {
        "work_meta_data": work_metadata,
        "work_indptr": work_indptr,
        "work_info_set": work_info_set,
        "reduce_indptr": reduce_indptr,
        "reduce_final_map": reduce_final_map,
        "reduce_partial_map": reduce_partial_map,
    }

def _aiter_mla_decode(
    q: torch.Tensor,
    kv_buffer: torch.Tensor,
    qo_indptr: torch.Tensor,
    kv_indptr: torch.Tensor,
    config: dict,
    q_scale: torch.Tensor | None = None,
    kv_scale: torch.Tensor | None = None,
) -> torch.Tensor:
    batch_size = config["batch_size"]
    nq = config["num_heads"]
    nkv = config["num_kv_heads"]
    dq = config["qk_head_dim"]
    dv = config["v_head_dim"]
    q_seq_len = config["q_seq_len"]

    total_kv_len = int(kv_indptr[-1].item())
    kv_indices = torch.arange(total_kv_len, dtype=torch.int32, device="cuda")

    kv_buffer_4d = kv_buffer.view(kv_buffer.shape[0], PAGE_SIZE, nkv, kv_buffer.shape[-1])

    max_q_len = q_seq_len
    kv_last_page_len = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)

    meta = _make_mla_decode_metadata(
        batch_size, max_q_len, nq, nkv,
        q.dtype, kv_buffer.dtype,
        qo_indptr, kv_indptr, kv_last_page_len,
        num_kv_splits=NUM_KV_SPLITS,
    )

    o = torch.empty((q.shape[0], nq, dv), dtype=torch.bfloat16, device="cuda")
    mla_decode_fwd(
        q.view(-1, nq, dq),
        kv_buffer_4d,
        o,
        qo_indptr,
        kv_indptr,
        kv_indices,
        kv_last_page_len,
        max_q_len,
        page_size=PAGE_SIZE,
        nhead_kv=nkv,
        sm_scale=SM_SCALE,
        logit_cap=0.0,
        num_kv_splits=NUM_KV_SPLITS,
        q_scale=q_scale,
        kv_scale=kv_scale,
        intra_batch_mode=True,
        **meta,
    )
    return o

def custom_kernel(data: input_t) -> output_t:
    q, kv_data, qo_indptr, kv_indptr, config = data
    kv_len = config.get("kv_seq_len", 0)

    if kv_len >= 4096 and "mxfp4" in kv_data:
        kv_fp4, kv_scale = kv_data["mxfp4"]
        total_kv = int(kv_indptr[-1].item())
        kv_bf16 = dequantize_mxfp4(kv_fp4, kv_scale, (total_kv, NUM_KV_HEADS, QK_HEAD_DIM))
        kv_fp8, kv_sc = quantize_fp8(kv_bf16)
    else:
        kv_fp8, kv_sc = kv_data["fp8"]

    q_fp8, q_sc = quantize_fp8(q)

    return _aiter_mla_decode(
        q_fp8, kv_fp8, qo_indptr, kv_indptr, config,
        q_scale=q_sc, kv_scale=kv_sc
    )

check_implementation = make_match_reference(custom_kernel, rtol=1e-01, atol=1e-01)