submission 688700

"""
Optimized MLA decode kernel for MI355X.

Strategy: Use aiter fp8 MLA decode with tuned parameters and minimized overhead.
Then iterate with custom Triton kernel if needed.
"""

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 = 576
V_HEAD_DIM = 512
SM_SCALE = 1.0 / (QK_HEAD_DIM ** 0.5)
PAGE_SIZE = 1
FP8_DTYPE = aiter_dtypes.fp8


def quantize_fp8(tensor: torch.Tensor) -> tuple[torch.Tensor, torch.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)


# Cache for metadata to avoid recomputation across calls with same config
_meta_cache = {}


def _get_num_kv_splits(batch_size: int, kv_seq_len: int) -> int:
    """Tune NUM_KV_SPLITS based on problem size for MI355X (304 CUs)."""
    total_work = batch_size * NUM_HEADS  # base parallelism
    if total_work >= 304:
        # Enough parallelism from batch*heads alone
        if kv_seq_len <= 1024:
            return 8
        else:
            return 16
    else:
        # Need more splits for parallelism
        if kv_seq_len <= 1024:
            return 16
        else:
            return 32


def _build_metadata(batch_size, q_seq_len, q_dtype, kv_dtype,
                    qo_indptr, kv_indptr, num_kv_splits):
    nq = NUM_HEADS
    nkv = NUM_KV_HEADS
    max_q_len = q_seq_len

    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_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_dtype,
    )

    return {
        "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_last_page_len


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"]

    # Quantize Q to fp8
    q_fp8, q_scale = quantize_fp8(q)

    # Use pre-quantized fp8 KV
    kv_buffer_fp8, kv_scale = kv_data["fp8"]

    # Tune splits for this config
    num_kv_splits = _get_num_kv_splits(batch_size, kv_seq_len)

    # Build metadata
    meta, kv_last_page_len = _build_metadata(
        batch_size, q_seq_len,
        q_fp8.dtype, kv_buffer_fp8.dtype,
        qo_indptr, kv_indptr, num_kv_splits,
    )

    total_kv_len = int(kv_indptr[-1].item())
    kv_indices = torch.arange(total_kv_len, dtype=torch.int32, device="cuda")

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

    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_kv_splits,
        q_scale=q_scale,
        kv_scale=kv_scale,
        intra_batch_mode=True,
        **meta,
    )
    return o