submission 746818

import torch
from task import input_t, output_t
from utils import make_match_reference

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
from aiter.utility.fp4_utils import (
    dynamic_mxfp4_quant,
    mxfp4_to_f32,
    e8m0_to_f32,
)

# DeepSeek R1 latent MQA 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  # 576
V_HEAD_DIM = KV_LORA_RANK  # 512
SM_SCALE = 1.0 / (QK_HEAD_DIM ** 0.5)

PAGE_SIZE = 1

# Try different split values - reduce overhead
# For smaller batch sizes, fewer splits may be better
NUM_KV_SPLITS = 16  # Reduced from 32 to reduce overhead

# Use fp8 for both Q and KV (best performance)
FP8_DTYPE = aiter_dtypes.fp8
Q_DTYPE = "fp8"
KV_DTYPE = "fp8"


def quantize_fp8(tensor: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
    """Dynamic per-tensor FP8 quantization."""
    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 quantize_mxfp4(tensor: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
    """MXFP4 block-wise quantization."""
    orig_shape = tensor.shape  # (B, M, N)
    B, M, N = orig_shape
    tensor_2d = tensor.reshape(B * M, N)
    fp4_data_2d, scale_e8m0 = dynamic_mxfp4_quant(tensor_2d)
    fp4_data = fp4_data_2d.view(B, M, N // 2)
    return fp4_data, scale_e8m0


def _make_mla_decode_metadata(
    batch_size: int,
    max_q_len: int,
    nhead: int,
    nhead_kv: int,
    q_dtype: torch.dtype,
    kv_dtype: torch.dtype,
    qo_indptr: torch.Tensor,
    kv_indptr: torch.Tensor,
    kv_last_page_len: torch.Tensor,
    num_kv_splits: int = NUM_KV_SPLITS,
    fast_mode: bool = True,
):
    """Allocate and populate work buffers for MLA decode."""
    info = get_mla_metadata_info_v1(
        batch_size, max_q_len, nhead, q_dtype, kv_dtype,
        is_sparse=False, fast_mode=fast_mode,
        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,
        nhead // nhead_kv, nhead_kv, 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=fast_mode,
        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,
    }


def _aiter_mla_decode(
    q: torch.Tensor,
    kv_buffer: torch.Tensor,
    qo_indptr: torch.Tensor,
    kv_indptr: torch.Tensor,
    config: dict,
    q_scale: torch.Tensor | None = None,
    kv_scale: torch.Tensor | None = None,
    num_kv_splits: int = NUM_KV_SPLITS,
) -> torch.Tensor:
    """MLA decode attention using aiter persistent-mode kernel."""
    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)
    
    # Enable fast_mode for better performance
    meta = _make_mla_decode_metadata(
        batch_size, max_q_len, nq, nkv,
        q.dtype, kv_buffer.dtype,
        qo_indptr, kv_indptr, kv_last_page_len,
        num_kv_splits=num_kv_splits,
        fast_mode=True,
    )

    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,
        **meta,
    )
    return o


def custom_kernel(data: input_t) -> output_t:
    """Reference MLA decode attention using aiter kernel."""
    q, kv_data, qo_indptr, kv_indptr, config = data

    # Get batch size to determine optimal num_kv_splits
    batch_size = config["batch_size"]
    kv_seq_len = config.get("kv_seq_len", 1024)
    
    # Adaptive num_kv_splits based on batch size and kv length
    # Fewer splits for small workloads to reduce overhead
    # More splits for large kv to improve parallelism
    if batch_size <= 4:
        num_kv_splits = 2 if kv_seq_len <= 1024 else 4
    elif batch_size <= 32:
        num_kv_splits = 4 if kv_seq_len <= 1024 else 8
    elif batch_size <= 64:
        num_kv_splits = 8 if kv_seq_len <= 1024 else 16
    else:
        num_kv_splits = 16

    # Quantize Q to fp8
    if Q_DTYPE == "fp8":
        q_input, q_scale = quantize_fp8(q)
    else:
        q_input, q_scale = q, None

    # Use fp8 KV
    if KV_DTYPE == "fp8":
        kv_buffer_fp8, kv_scale = kv_data["fp8"]
        kv_input = kv_buffer_fp8
    else:
        kv_input, kv_scale = kv_data["bf16"], None

    return _aiter_mla_decode(
        q_input, kv_input, qo_indptr, kv_indptr, config,
        q_scale=q_scale, kv_scale=kv_scale,
        num_kv_splits=num_kv_splits,
    )