submission 678917

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

# -------------------------------
# Constants
# -------------------------------
NUM_HEADS = 16
NUM_KV_HEADS = 1
KV_LORA_RANK = 512
QK_ROPE_HEAD_DIM = 64
QK_HEAD_DIM = KV_LORA_RANK + QK_ROPE_HEAD_DIM
V_HEAD_DIM = KV_LORA_RANK
SM_SCALE = 1.0 / (QK_HEAD_DIM ** 0.5)

PAGE_SIZE = 1
NUM_KV_SPLITS = 32

# -------------------------------
# Aiter kernel (optimized baseline)
# -------------------------------
def _aiter_mla_decode(
    q,
    kv_buffer,
    qo_indptr,
    kv_indptr,
    config,
    q_scale=None,
    kv_scale=None,
):
    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_buffer_4d = kv_buffer.view(kv_buffer.shape[0], PAGE_SIZE, nkv, kv_buffer.shape[-1])

    max_q_len = q_seq_len
    kv_indices = torch.arange(total_kv_len, dtype=torch.int32, device="cuda")
    kv_last_page_len = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)

    info = get_mla_metadata_info_v1(
        batch_size, max_q_len, nq, q.dtype, kv_buffer.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,
        nq // nkv, nkv, 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_buffer.dtype,
    )

    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,
        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,
    )
    return o


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

    # 🔥 Use fp8 path (best with aiter)
    kv_buffer_fp8, kv_scale = kv_data["fp8"]

    return _aiter_mla_decode(
        q,
        kv_buffer_fp8,
        qo_indptr,
        kv_indptr,
        config,
        q_scale=None,
        kv_scale=kv_scale,
    )