submission 616098

#!POPCORN leaderboard amd-mixed-mla
#!POPCORN gpu MI355X
"""
MLA Decode — Zhao Dylan
v9: 零GPU→CPU同步 + 动态Q dtype
bf16 Q省quant(小/中), fp8 Q高吞吐(大)
total_kv_len从config计算,避免.item()同步
"""
import torch
from task import input_t, output_t
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

PAGE_SIZE = 1
SM_SCALE = 1.0 / (576 ** 0.5)
FP8_DTYPE = aiter_dtypes.fp8
BF16_DTYPE = torch.bfloat16
_FP8_MAX = torch.finfo(FP8_DTYPE).max
_FP8_MIN = torch.finfo(FP8_DTYPE).min

_meta = {}


def _build_meta(bs, q_seq_len, nq, nkv, dv, total_kv_len, qo_indptr, kv_indptr, q_dtype):
    if total_kv_len <= 4096:
        nks = 8
    elif total_kv_len <= 16384:
        nks = 16
    elif total_kv_len <= 65536:
        nks = 32
    elif total_kv_len <= 524288:
        nks = 32
    else:
        nks = 64

    kv_lpl = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)
    kv_idx = torch.arange(total_kv_len, dtype=torch.int32, device="cuda")

    info = get_mla_metadata_info_v1(
        bs, q_seq_len, nq, q_dtype, FP8_DTYPE,
        is_sparse=False, fast_mode=False,
        num_kv_splits=nks, intra_batch_mode=True,
    )
    wm, wi, ws, ri, rf, rp = [torch.empty(s, dtype=t, device="cuda") for s, t in info]
    get_mla_metadata_v1(
        qo_indptr, kv_indptr, kv_lpl,
        nq // nkv, nkv, True,
        wm, ws, wi, ri, rf, rp,
        page_size=PAGE_SIZE, kv_granularity=16,
        max_seqlen_qo=q_seq_len, uni_seqlen_qo=q_seq_len,
        fast_mode=False, max_split_per_batch=nks,
        intra_batch_mode=True,
        dtype_q=q_dtype, dtype_kv=FP8_DTYPE,
    )

    return (
        torch.empty((bs * q_seq_len, nq, dv), dtype=torch.bfloat16, device="cuda"),
        q_seq_len, nks, kv_idx, kv_lpl, wm, wi, ws, ri, rf, rp,
    )


def custom_kernel(data: input_t) -> output_t:
    q, kv_data, qo_indptr, kv_indptr, config = data

    bs = config["batch_size"]
    nq = config["num_heads"]
    nkv = config["num_kv_heads"]
    dq = config["qk_head_dim"]
    dv = config["v_head_dim"]
    qsl = config["q_seq_len"]
    kvsl = config["kv_seq_len"]
    total_kv = bs * kvsl

    kv_fp8, kv_scale = kv_data["fp8"]
    kv_4d = kv_fp8.view(kv_fp8.shape[0], PAGE_SIZE, nkv, kv_fp8.shape[-1])

    use_fp8 = total_kv > 1_000_000
    ck = (bs, qsl, kvsl, use_fp8)
    mc = _meta.get(ck)
    if mc is None:
        qd = FP8_DTYPE if use_fp8 else BF16_DTYPE
        mc = _build_meta(bs, qsl, nq, nkv, dv, total_kv, qo_indptr, kv_indptr, qd)
        _meta[ck] = mc

    o, q_seq_len, nks, kv_idx, kv_lpl, wm, wi, ws, ri, rf, rp = mc

    if use_fp8:
        amax = q.abs().amax().clamp(min=1e-12)
        scale = amax / _FP8_MAX
        q_in = (q / scale).clamp(min=_FP8_MIN, max=_FP8_MAX).to(FP8_DTYPE).view(-1, nq, dq)
        q_sc = scale.to(torch.float32).reshape(1)
    else:
        q_in = q.view(-1, nq, dq)
        q_sc = None

    mla_decode_fwd(
        q_in, kv_4d, o,
        qo_indptr, kv_indptr, kv_idx, kv_lpl,
        q_seq_len,
        page_size=PAGE_SIZE, nhead_kv=nkv,
        sm_scale=SM_SCALE, logit_cap=0.0,
        num_kv_splits=nks,
        q_scale=q_sc, kv_scale=kv_scale,
        intra_batch_mode=True,
        work_meta_data=wm, work_indptr=wi, work_info_set=ws,
        reduce_indptr=ri, reduce_final_map=rf, reduce_partial_map=rp,
    )
    return o