submission 744922

#!POPCORN leaderboard amd-mixed-mla
#!POPCORN gpu MI355X
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

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
_fp8_max = torch.finfo(FP8_DTYPE).max
_fp8_min = torch.finfo(FP8_DTYPE).min
_work_cache = {}


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

    amax = q.abs().amax().clamp(min=1e-12)
    q_scale = amax / _fp8_max
    q_fp8 = (q / q_scale).clamp(min=_fp8_min, max=_fp8_max).to(FP8_DTYPE)
    q_scale = q_scale.to(torch.float32).reshape(1)

    kv_fp8, kv_scale = kv_data["fp8"]
    total_kv = int(kv_indptr[-1].item())
    kv_indices = torch.arange(total_kv, dtype=torch.int32, device="cuda")
    kv_4d = kv_fp8.view(kv_fp8.shape[0], 1, 1, kv_fp8.shape[-1])
    kv_last = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)

    if batch_size not in _work_cache:
        info = get_mla_metadata_info_v1(
            batch_size, 1, NUM_HEADS, q_fp8.dtype, kv_fp8.dtype,
            is_sparse=False, fast_mode=False,
            num_kv_splits=NUM_KV_SPLITS, intra_batch_mode=True,
        )
        _work_cache[batch_size] = [torch.empty(s, dtype=t, device="cuda") for s, t in info]

    (wm, wi, wis, ri, rfm, rpm) = _work_cache[batch_size]
    get_mla_metadata_v1(
        qo_indptr, kv_indptr, kv_last,
        NUM_HEADS, NUM_KV_HEADS, True,
        wm, wis, wi, ri, rfm, rpm,
        page_size=PAGE_SIZE, kv_granularity=16,
        max_seqlen_qo=1, uni_seqlen_qo=1,
        fast_mode=False, max_split_per_batch=NUM_KV_SPLITS,
        intra_batch_mode=True,
        dtype_q=q_fp8.dtype, dtype_kv=kv_fp8.dtype,
    )

    o = torch.empty((batch_size, NUM_HEADS, V_HEAD_DIM), dtype=torch.bfloat16, device="cuda")
    mla_decode_fwd(
        q_fp8.view(-1, NUM_HEADS, QK_HEAD_DIM), kv_4d, o,
        qo_indptr, kv_indptr, kv_indices, kv_last, 1,
        page_size=PAGE_SIZE, nhead_kv=NUM_KV_HEADS,
        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,
        work_meta_data=wm, work_indptr=wi, work_info_set=wis,
        reduce_indptr=ri, reduce_final_map=rfm, reduce_partial_map=rpm,
    )
    return o