submission 629332

#!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
FP8_DTYPE = aiter_dtypes.fp8

_cache = {}


def quantize_fp8(tensor):
    finfo = torch.finfo(FP8_DTYPE)
    amax = tensor.abs().amax().clamp(min=1e-12)
    scale = amax / finfo.max
    fp8_tensor = (tensor / scale).clamp(min=finfo.min, max=finfo.max).to(FP8_DTYPE)
    return fp8_tensor, scale.to(torch.float32).reshape(1)


def _get_cached(batch_size, kv_seq_len, q_dtype, kv_dtype, num_splits):
    key = (batch_size, kv_seq_len, str(q_dtype), str(kv_dtype), num_splits)
    if key in _cache:
        return _cache[key]

    qo_indptr = torch.arange(0, batch_size + 1, dtype=torch.int32, device="cuda")
    kv_indptr = torch.arange(0, batch_size + 1, dtype=torch.int32, device="cuda") * kv_seq_len
    total_kv = batch_size * kv_seq_len
    kv_indices = torch.arange(total_kv, dtype=torch.int32, device="cuda")
    kv_last_page_len = torch.full((batch_size,), kv_seq_len, dtype=torch.int32, device="cuda")

    info = get_mla_metadata_info_v1(
        batch_size, 1, NUM_HEADS, q_dtype, kv_dtype,
        is_sparse=False, fast_mode=False,
        num_kv_splits=num_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=1, uni_seqlen_qo=1, fast_mode=False,
        max_split_per_batch=num_splits, intra_batch_mode=True,
        dtype_q=q_dtype, dtype_kv=kv_dtype,
    )

    meta = {
        "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,
    }

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

    result = (qo_indptr, kv_indptr, kv_indices, kv_last_page_len, meta, o)
    _cache[key] = result
    return result


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

    batch_size = config["batch_size"]
    kv_seq_len = config["kv_seq_len"]

    kv_buffer_fp8, kv_scale = kv_data["fp8"]

    # Use a8w8 only when compute-bound (large batch + long seq)
    use_a8w8 = (kv_seq_len > 1024 and batch_size >= 64)

    if use_a8w8:
        q_input, q_scale_val = quantize_fp8(q)
        q_dtype = q_input.dtype
    else:
        q_input = q
        q_scale_val = None
        q_dtype = q.dtype

    # Tuned num_splits per case
    if kv_seq_len <= 1024:
        num_splits = 8 if batch_size <= 32 else 16
    else:
        num_splits = 32

    cached = _get_cached(batch_size, kv_seq_len, q_dtype, kv_buffer_fp8.dtype, num_splits)
    c_qo, c_kv, kv_indices, kv_last_page_len, meta, o = cached

    kv_4d = kv_buffer_fp8.view(-1, PAGE_SIZE, NUM_KV_HEADS, kv_buffer_fp8.shape[-1])

    mla_decode_fwd(
        q_input.view(-1, NUM_HEADS, QK_HEAD_DIM), kv_4d, o,
        c_qo, c_kv, kv_indices, kv_last_page_len, 1,
        page_size=PAGE_SIZE, nhead_kv=NUM_KV_HEADS,
        sm_scale=SM_SCALE, logit_cap=0.0, num_kv_splits=num_splits,
        q_scale=q_scale_val, kv_scale=kv_scale,
        intra_batch_mode=True, **meta,
    )
    return o