submission 711646

import torch
import math
from aiter.mla import mla_decode_fwd
from aiter import get_mla_metadata_info_v1, get_mla_metadata_v1
from aiter import dtypes as aiter_dtypes

# ========== 常量（与参考实现保持一致）==========
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 / math.sqrt(QK_HEAD_DIM)

PAGE_SIZE = 1
NUM_KV_SPLITS = 32
FP8_DTYPE = aiter_dtypes.fp8

# ========== FP8 量化（与参考相同）==========
def quantize_fp8(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)

# ========== 构建 persistent 元数据（与参考相同）==========
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,
):
    info = get_mla_metadata_info_v1(
        batch_size, max_q_len, nhead, 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,
        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=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,
    }

# ========== 自定义内核入口（必须命名为 custom_kernel）==========
def custom_kernel(data):
    """
    data: (q, kv_data, qo_indptr, kv_indptr, config)
    q: (total_q, 16, 576) bfloat16
    kv_data: dict with key "fp8" -> (kv_buffer_fp8, kv_scale)
    """
    q, kv_data, qo_indptr, kv_indptr, config = data

    batch_size = config["batch_size"]
    q_seq_len = config.get("q_seq_len", 1)

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

    # 将 Q 量化为 FP8
    q_fp8, q_scale = quantize_fp8(q)

    # 使用 FP8 格式的 KV cache
    kv_buffer_fp8, kv_scale = kv_data["fp8"]

    # 重塑为 aiter 需要的 4D 格式: (total_kv, page_size, nhead_kv, dim)
    kv_buffer_4d = kv_buffer_fp8.view(kv_buffer_fp8.shape[0], PAGE_SIZE, NUM_KV_HEADS, kv_buffer_fp8.shape[-1])

    max_q_len = q_seq_len
    kv_last_page_len = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)

    meta = _make_mla_decode_metadata(
        batch_size, max_q_len, NUM_HEADS, NUM_KV_HEADS,
        q_fp8.dtype, kv_buffer_fp8.dtype,
        qo_indptr, kv_indptr, kv_last_page_len,
        num_kv_splits=NUM_KV_SPLITS,
    )

    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,
        max_q_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