submission 754413

"""
Optimized MLA Decode v9 for AMD MI355X.
- Try MXFP4 KV path (2x less bandwidth than fp8!) — kv_data["mxfp4"] is pre-quantized
- Fallback chain: mxfp4 KV → bf16 Q + fp8 KV (a16w8) → fp8 Q + fp8 KV (a8w8)
- Metadata caching + fast_mode=True + adaptive num_kv_splits
"""
from task import input_t, output_t
import torch
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

torch.set_grad_enabled(False)

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
FP8_DTYPE = aiter_dtypes.fp8

_FP8_FINFO = torch.finfo(FP8_DTYPE)
_FP8_MAX = _FP8_FINFO.max
_FP8_MIN = _FP8_FINFO.min

_cache = {}
# Path selection: None=untested, "mxfp4"=use mxfp4 KV, "a16w8"=bf16 Q+fp8 KV, "a8w8"=fp8 Q+fp8 KV
_best_path = None

_fused_fp8_quant = None
try:
    from aiter import scaled_fp8_quant as _fused_fp8_quant
except (ImportError, AttributeError):
    pass


def _quantize_fp8(tensor):
    if _fused_fp8_quant is not None:
        return _fused_fp8_quant(tensor)
    amax = tensor.abs().amax().clamp(min=1e-12)
    scale = amax / _FP8_MAX
    fp8 = (tensor / scale).clamp(min=_FP8_MIN, max=_FP8_MAX).to(FP8_DTYPE)
    return fp8, scale.to(torch.float32).reshape(1)


def _get_num_kv_splits(batch_size, kv_seq_len):
    total_kv = batch_size * kv_seq_len
    if total_kv <= 4096:
        return 8
    elif total_kv <= 32768:
        return 16
    else:
        return 32


def _build_and_cache(batch_size, q_seq_len, kv_seq_len, total_q,
                     q_dtype, kv_dtype, qo_indptr, kv_indptr, num_kv_splits):
    kv_last_page_len = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)
    total_kv_len = int(kv_indptr[-1].item())
    kv_indices = torch.arange(total_kv_len, dtype=torch.int32, device="cuda")

    info = get_mla_metadata_info_v1(
        batch_size, q_seq_len, NUM_HEADS, q_dtype, kv_dtype,
        is_sparse=False, fast_mode=True,
        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,
        NUM_HEADS // NUM_KV_HEADS, NUM_KV_HEADS, 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=q_seq_len, uni_seqlen_qo=q_seq_len,
        fast_mode=True, max_split_per_batch=num_kv_splits,
        intra_batch_mode=True, dtype_q=q_dtype, dtype_kv=kv_dtype,
    )

    o = torch.empty((total_q, NUM_HEADS, V_HEAD_DIM), dtype=torch.bfloat16, device="cuda")

    return {
        "kv_indices": kv_indices,
        "kv_last_page_len": kv_last_page_len,
        "o": o,
        "num_kv_splits": num_kv_splits,
        "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 _run_kernel(q_input, kv_4d, o, qo_indptr, kv_indptr, meta, q_scale, kv_scale, q_seq_len):
    mla_decode_fwd(
        q_input.view(-1, NUM_HEADS, QK_HEAD_DIM),
        kv_4d,
        o,
        qo_indptr,
        kv_indptr,
        meta["kv_indices"],
        meta["kv_last_page_len"],
        q_seq_len,
        page_size=PAGE_SIZE,
        nhead_kv=NUM_KV_HEADS,
        sm_scale=SM_SCALE,
        logit_cap=0.0,
        num_kv_splits=meta["num_kv_splits"],
        q_scale=q_scale,
        kv_scale=kv_scale,
        intra_batch_mode=True,
        work_meta_data=meta["work_meta_data"],
        work_indptr=meta["work_indptr"],
        work_info_set=meta["work_info_set"],
        reduce_indptr=meta["reduce_indptr"],
        reduce_final_map=meta["reduce_final_map"],
        reduce_partial_map=meta["reduce_partial_map"],
    )


def _try_mxfp4_path(q, kv_data, qo_indptr, kv_indptr, config):
    """Try MXFP4 KV path: bf16 Q + mxfp4 KV (2x less bandwidth than fp8)."""
    batch_size = config["batch_size"]
    q_seq_len = config["q_seq_len"]
    kv_seq_len = config["kv_seq_len"]

    kv_mxfp4, kv_scale_mxfp4 = kv_data["mxfp4"]
    # kv_mxfp4: (total_kv, 1, 288) fp4x2 — try viewing as uint8 for dispatch
    kv_as_uint8 = kv_mxfp4.view(torch.uint8)
    kv_4d = kv_as_uint8.view(kv_as_uint8.shape[0], PAGE_SIZE, NUM_KV_HEADS, -1)

    num_kv_splits = _get_num_kv_splits(batch_size, kv_seq_len)
    key = ("mxfp4", batch_size, q_seq_len, kv_seq_len)

    if key not in _cache:
        _cache[key] = _build_and_cache(
            batch_size, q_seq_len, kv_seq_len, q.shape[0],
            q.dtype, torch.uint8, qo_indptr, kv_indptr,
            num_kv_splits,
        )
    meta = _cache[key]
    o = meta["o"]
    _run_kernel(q, kv_4d, o, qo_indptr, kv_indptr, meta,
                q_scale=None, kv_scale=kv_scale_mxfp4, q_seq_len=q_seq_len)
    return o


def _try_a16w8_path(q, kv_data, qo_indptr, kv_indptr, config):
    """Try a16w8 path: bf16 Q + fp8 KV (skip Q quantization)."""
    batch_size = config["batch_size"]
    q_seq_len = config["q_seq_len"]
    kv_seq_len = config["kv_seq_len"]

    kv_buffer_fp8, kv_scale = kv_data["fp8"]
    kv_4d = kv_buffer_fp8.view(kv_buffer_fp8.shape[0], PAGE_SIZE, NUM_KV_HEADS, -1)

    num_kv_splits = _get_num_kv_splits(batch_size, kv_seq_len)
    key = ("a16w8", batch_size, q_seq_len, kv_seq_len)

    if key not in _cache:
        _cache[key] = _build_and_cache(
            batch_size, q_seq_len, kv_seq_len, q.shape[0],
            q.dtype, kv_buffer_fp8.dtype, qo_indptr, kv_indptr,
            num_kv_splits,
        )
    meta = _cache[key]
    o = meta["o"]
    _run_kernel(q, kv_4d, o, qo_indptr, kv_indptr, meta,
                q_scale=None, kv_scale=kv_scale, q_seq_len=q_seq_len)
    return o


def _run_a8w8_path(q, kv_data, qo_indptr, kv_indptr, config):
    """a8w8 path: fp8 Q + fp8 KV (always works)."""
    batch_size = config["batch_size"]
    q_seq_len = config["q_seq_len"]
    kv_seq_len = config["kv_seq_len"]

    kv_buffer_fp8, kv_scale = kv_data["fp8"]
    kv_4d = kv_buffer_fp8.view(kv_buffer_fp8.shape[0], PAGE_SIZE, NUM_KV_HEADS, -1)

    q_fp8, q_scale = _quantize_fp8(q)

    num_kv_splits = _get_num_kv_splits(batch_size, kv_seq_len)
    key = ("a8w8", batch_size, q_seq_len, kv_seq_len)

    if key not in _cache:
        _cache[key] = _build_and_cache(
            batch_size, q_seq_len, kv_seq_len, q.shape[0],
            q_fp8.dtype, kv_buffer_fp8.dtype, qo_indptr, kv_indptr,
            num_kv_splits,
        )
    meta = _cache[key]
    o = meta["o"]
    _run_kernel(q_fp8, kv_4d, o, qo_indptr, kv_indptr, meta,
                q_scale=q_scale, kv_scale=kv_scale, q_seq_len=q_seq_len)
    return o


@torch.inference_mode()
def custom_kernel(data: input_t) -> output_t:
    global _best_path
    q, kv_data, qo_indptr, kv_indptr, config = data

    if _best_path is None:
        # Try MXFP4 first (2x less bandwidth)
        try:
            result = _try_mxfp4_path(q, kv_data, qo_indptr, kv_indptr, config)
            _best_path = "mxfp4"
            return result
        except Exception:
            _cache.clear()

        # Try a16w8 (bf16 Q + fp8 KV, no Q quant overhead)
        try:
            result = _try_a16w8_path(q, kv_data, qo_indptr, kv_indptr, config)
            _best_path = "a16w8"
            return result
        except Exception:
            _cache.clear()

        # Fall back to a8w8
        result = _run_a8w8_path(q, kv_data, qo_indptr, kv_indptr, config)
        _best_path = "a8w8"
        return result

    if _best_path == "mxfp4":
        return _try_mxfp4_path(q, kv_data, qo_indptr, kv_indptr, config)
    elif _best_path == "a16w8":
        return _try_a16w8_path(q, kv_data, qo_indptr, kv_indptr, config)
    else:
        return _run_a8w8_path(q, kv_data, qo_indptr, kv_indptr, config)