submission 625818

#!POPCORN leaderboard amd-mixed-mla
#!POPCORN gpu MI355X
"""
v27: Adaptive KV splits — fewer for short seqs (less reduce overhead),
more for long seqs (better parallelism).
"""
import torch
from task import input_t, output_t
from aiter import dtypes as aiter_dtypes, get_mla_metadata_info_v1, get_mla_metadata_v1
import aiter

FP8_DTYPE = aiter_dtypes.fp8
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

_full_cache = {}


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

    batch_size = config["batch_size"]
    nq = config["num_heads"]
    nkv = NUM_KV_HEADS
    dq = config["qk_head_dim"]
    dv = config["v_head_dim"]
    q_seq_len = config["q_seq_len"]

    kv_buffer_fp8, kv_scale = kv_data["fp8"]
    total_kv_len = int(kv_indptr[-1].item())

    use_fp8_q = total_kv_len > 1_000_000
    if use_fp8_q:
        finfo = torch.finfo(FP8_DTYPE)
        amax = q.abs().amax().clamp(min=1e-12)
        scale = amax / finfo.max
        q_input = (q / scale).clamp(min=finfo.min, max=finfo.max).to(FP8_DTYPE)
        q_scale = scale.to(torch.float32).reshape(1)
    else:
        q_input, q_scale = q, None

    # Adaptive splits: balance reduce overhead vs parallelism
    avg_kv_len = total_kv_len // max(batch_size, 1)
    if avg_kv_len <= 2048:
        num_kv_splits = 16
    elif avg_kv_len <= 4096:
        num_kv_splits = 24
    else:
        num_kv_splits = 32

    cache_key = (batch_size, total_kv_len, use_fp8_q, num_kv_splits)

    if cache_key not in _full_cache:
        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, q_seq_len, nq, q_input.dtype, kv_buffer_fp8.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=q_seq_len,
            uni_seqlen_qo=q_seq_len,
            fast_mode=False,
            max_split_per_batch=num_kv_splits,
            intra_batch_mode=True,
            dtype_q=q_input.dtype,
            dtype_kv=kv_buffer_fp8.dtype,
        )

        split_entries = reduce_partial_map.size(0) * q_seq_len
        split_data = torch.empty(
            (split_entries, 1, nq, dv), dtype=torch.float32, device="cuda"
        )
        split_lse = torch.empty(
            (split_entries, 1, nq, 1), dtype=torch.float32, device="cuda"
        )

        o = torch.empty((q.shape[0], nq, dv), dtype=torch.bfloat16, device="cuda")

        _full_cache[cache_key] = (
            kv_indices, kv_last_page_len,
            work_metadata, work_indptr, work_info_set,
            reduce_indptr, reduce_final_map, reduce_partial_map,
            split_data, split_lse, o,
        )

    (kv_indices, kv_last_page_len,
     work_metadata, work_indptr, work_info_set,
     reduce_indptr, reduce_final_map, reduce_partial_map,
     split_data, split_lse, o) = _full_cache[cache_key]

    kv_buffer_4d = kv_buffer_fp8.view(
        kv_buffer_fp8.shape[0], PAGE_SIZE, nkv, kv_buffer_fp8.shape[-1]
    )

    aiter.mla_decode_stage1_asm_fwd(
        q_input.view(-1, nq, dq),
        kv_buffer_4d,
        qo_indptr, kv_indptr, kv_indices, kv_last_page_len,
        None,
        work_metadata, work_indptr, work_info_set,
        q_seq_len, PAGE_SIZE, nkv, SM_SCALE,
        split_data, split_lse, o,
        q_scale, kv_scale,
    )

    aiter.mla_reduce_v1(
        split_data, split_lse,
        reduce_indptr, reduce_final_map, reduce_partial_map,
        q_seq_len, o, None,
    )

    return o