submission 727032

"""Optimized MLA decode: cached metadata + tuned KV splits + output pre-alloc."""
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
KV_LORA_RANK = 512
QK_ROPE_HEAD_DIM = 64
QK_HEAD_DIM = KV_LORA_RANK + QK_ROPE_HEAD_DIM  # 576
V_HEAD_DIM = KV_LORA_RANK  # 512
SM_SCALE = 1.0 / (QK_HEAD_DIM ** 0.5)
PAGE_SIZE = 1
FP8_DTYPE = aiter_dtypes.fp8

# Cache for metadata and output tensors
_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_meta(batch_size, q_seq_len, kv_seq_len, q_dtype, kv_dtype,
                     qo_indptr, kv_indptr):
    """Cache metadata + work buffers for repeated calls with same shape."""
    key = (batch_size, q_seq_len, kv_seq_len, q_dtype, kv_dtype)
    if key in _cache:
        return _cache[key]

    total_kv = batch_size * kv_seq_len
    kv_last_page_len = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)

    # Try different num_kv_splits based on shape
    # Smaller batches benefit from fewer splits (less reduction overhead)
    if batch_size <= 4:
        num_splits = 16
    elif batch_size <= 32:
        num_splits = 32
    else:
        num_splits = 32

    info = get_mla_metadata_info_v1(
        batch_size, q_seq_len, 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=q_seq_len,
        uni_seqlen_qo=q_seq_len,
        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,
    }

    kv_indices = torch.arange(total_kv, dtype=torch.int32, device="cuda")

    result = (meta, kv_indices, kv_last_page_len, num_splits)
    _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"]
    q_seq_len = config["q_seq_len"]
    kv_seq_len = config["kv_seq_len"]

    # FP8 quantize Q on-the-fly
    q_fp8, q_scale = quantize_fp8(q)

    # Use FP8 KV cache
    kv_buffer_fp8, kv_scale = kv_data["fp8"]
    kv_buffer_4d = kv_buffer_fp8.view(kv_buffer_fp8.shape[0], PAGE_SIZE, NUM_KV_HEADS, kv_buffer_fp8.shape[-1])

    # Get cached metadata
    meta, kv_indices, kv_last_page_len, num_splits = _get_cached_meta(
        batch_size, q_seq_len, kv_seq_len,
        q_fp8.dtype, kv_buffer_fp8.dtype,
        qo_indptr, kv_indptr,
    )

    # Pre-allocate output
    o = torch.empty((q.shape[0], NUM_HEADS, V_HEAD_DIM), dtype=torch.bfloat16, device="cuda")

    mla_decode_fwd(
        q_fp8.view(-1, NUM_HEADS, QK_HEAD_DIM),
        kv_buffer_4d,
        o,
        qo_indptr, kv_indptr, kv_indices, kv_last_page_len,
        q_seq_len,
        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, kv_scale=kv_scale,
        intra_batch_mode=True,
        **meta,
    )
    return o