submission 654962

#!POPCORN leaderboard amd-mixed-mla
#!POPCORN gpu MI355X

# gpumode leaderboard reference
"""
Reference implementation for MLA (Multi-head Latent Attention) decode kernel.

Uses aiter MLA kernels (mla_decode_fwd) as the reference.
DeepSeek R1 forward_absorb MLA: absorbed q (576), compressed kv_buffer (576),
output v_head_dim = kv_lora_rank = 512.

The input provides:
  q:       (total_q, 16, 576) bfloat16 — absorbed query
  kv_data: dict with KV cache in three formats:
    "bf16":  Tensor  (total_kv, 1, 576)  bfloat16          — highest precision
    "fp8":   (Tensor, Tensor)  kv_buffer fp8 + scalar scale — per-tensor quantized
    "mxfp4": (Tensor, Tensor)  kv_buffer fp4x2 + fp8_e8m0  — block-32 quantized
  The reference quantizes Q to fp8 on-the-fly inside ref_kernel.

The reference kernel quantizes Q to fp8 on-the-fly and uses fp8 KV (a8w8 kernel),
which is ~2-3x faster than bf16 on MI355X with negligible accuracy loss.

Decode only — persistent mode with get_mla_metadata_v1.
"""

import os

os.environ.setdefault("PYTORCH_ROCM_ARCH", "gfx950")
os.environ.setdefault("CXX", "clang++")

import torch
import triton
import triton.language as tl
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
from aiter import per_tensor_quant_hip
from aiter.utility.fp4_utils import (
    dynamic_mxfp4_quant,
    mxfp4_to_f32,
    e8m0_to_f32,
)

# ---------------------------------------------------------------------------
# DeepSeek R1 latent MQA constants (forward_absorb path)
# https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/blob/main/config.json
# ---------------------------------------------------------------------------
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 = 8
NUM_KV_SPLITS = 8
PAGED_NUM_KV_SPLITS = None

# FP8 dtype (platform-specific via aiter)
FP8_DTYPE = aiter_dtypes.fp8

# Query dtype for the reference kernel: "fp8" or "bf16"
Q_DTYPE = "bf16"

# KV cache dtype for the reference kernel: "fp8" or "bf16"
KV_DTYPE = "bf16"

_MLA_METADATA_CACHE = {}
_KV_INDICES_CACHE = {}
_PAGED_KV_CACHE = {}


@triton.jit
def _mla_decode_paged_stage1_kernel(
    Q,
    KV,
    QO_INDPTR,
    KV_PAGE_INDPTR,
    KV_LAST_PAGE_LEN,
    MID_O,
    MID_LSE,
    stride_qs,
    stride_qh,
    stride_qd,
    stride_kvp,
    stride_kvt,
    stride_kvh,
    stride_kvd,
    stride_mob,
    stride_mos,
    stride_moh,
    stride_mod,
    stride_mlb,
    stride_mls,
    stride_mlh,
    sm_scale,
    num_v_blocks: tl.constexpr,
    num_kv_splits: tl.constexpr,
    qk_dim: tl.constexpr,
    v_dim: tl.constexpr,
    page_size: tl.constexpr,
    block_n: tl.constexpr,
    block_d: tl.constexpr,
    block_v: tl.constexpr,
):
    batch_idx = tl.program_id(0)
    head_idx = tl.program_id(1)
    split_v_idx = tl.program_id(2)
    split_idx = split_v_idx // num_v_blocks
    v_block_idx = split_v_idx % num_v_blocks

    q_start = tl.load(QO_INDPTR + batch_idx)
    q_end = tl.load(QO_INDPTR + batch_idx + 1)
    if q_end <= q_start:
        return

    page_start = tl.load(KV_PAGE_INDPTR + batch_idx)
    page_end = tl.load(KV_PAGE_INDPTR + batch_idx + 1)
    num_pages = page_end - page_start
    if num_pages <= 0:
        return

    q_row = q_start
    last_page_len = tl.load(KV_LAST_PAGE_LEN + batch_idx)
    seq_len = (num_pages - 1) * page_size + last_page_len
    kv_len_per_split = tl.cdiv(seq_len, num_kv_splits)
    split_kv_start = kv_len_per_split * split_idx
    split_kv_end = tl.minimum(split_kv_start + kv_len_per_split, seq_len)

    offs_v = v_block_idx * block_v + tl.arange(0, block_v)
    mask_v = offs_v < v_dim
    acc = tl.zeros((block_v,), dtype=tl.float32)
    e_max = -float("inf")
    e_sum = 0.0

    if split_kv_end > split_kv_start:
        for tok_start in range(split_kv_start, split_kv_end, block_n):
            offs_n = tok_start + tl.arange(0, block_n)
            mask_n = offs_n < split_kv_end
            page_rel = offs_n // page_size
            token_off = offs_n % page_size
            page_idx = page_start + page_rel

            scores = tl.zeros((block_n,), dtype=tl.float32)
            for d_start in range(0, qk_dim, block_d):
                offs_d = d_start + tl.arange(0, block_d)
                mask_d = offs_d < qk_dim
                q = tl.load(
                    Q + q_row * stride_qs + head_idx * stride_qh + offs_d * stride_qd,
                    mask=mask_d,
                    other=0.0,
                )
                kv = tl.load(
                    KV
                    + page_idx[:, None] * stride_kvp
                    + token_off[:, None] * stride_kvt
                    + offs_d[None, :] * stride_kvd,
                    mask=mask_n[:, None] & mask_d[None, :],
                    other=0.0,
                )
                scores += tl.sum(kv.to(tl.float32) * q[None, :].to(tl.float32), axis=1)

            scores = tl.where(mask_n, scores * sm_scale, -float("inf"))
            n_e_max = tl.maximum(e_max, tl.max(scores, axis=0))
            re_scale = tl.exp(e_max - n_e_max)
            probs = tl.exp(scores - n_e_max)

            values = tl.load(
                KV
                + page_idx[:, None] * stride_kvp
                + token_off[:, None] * stride_kvt
                + offs_v[None, :] * stride_kvd,
                mask=mask_n[:, None] & mask_v[None, :],
                other=0.0,
            )
            acc = acc * re_scale + tl.sum(values.to(tl.float32) * probs[:, None], axis=0)
            e_sum = e_sum * re_scale + tl.sum(probs, axis=0)
            e_max = n_e_max

        tl.store(
            MID_O
            + batch_idx * stride_mob
            + split_idx * stride_mos
            + head_idx * stride_moh
            + offs_v * stride_mod,
            acc / e_sum,
            mask=mask_v,
        )
        if v_block_idx == 0:
            tl.store(
                MID_LSE
                + batch_idx * stride_mlb
                + split_idx * stride_mls
                + head_idx * stride_mlh,
                e_max + tl.log(e_sum),
            )
    else:
        tl.store(
            MID_O
            + batch_idx * stride_mob
            + split_idx * stride_mos
            + head_idx * stride_moh
            + offs_v * stride_mod,
            0.0,
            mask=mask_v,
        )
        if v_block_idx == 0:
            tl.store(
                MID_LSE
                + batch_idx * stride_mlb
                + split_idx * stride_mls
                + head_idx * stride_mlh,
                -float("inf"),
            )


@triton.jit
def _mla_decode_paged_stage2_kernel(
    MID_O,
    MID_LSE,
    QO_INDPTR,
    KV_PAGE_INDPTR,
    O,
    stride_mob,
    stride_mos,
    stride_moh,
    stride_mod,
    stride_mlb,
    stride_mls,
    stride_mlh,
    stride_os,
    stride_oh,
    stride_od,
    num_kv_splits: tl.constexpr,
    v_dim: tl.constexpr,
    block_v: tl.constexpr,
):
    batch_idx = tl.program_id(0)
    head_idx = tl.program_id(1)
    v_block_idx = tl.program_id(2)

    q_start = tl.load(QO_INDPTR + batch_idx)
    q_end = tl.load(QO_INDPTR + batch_idx + 1)
    if q_end <= q_start:
        return

    page_start = tl.load(KV_PAGE_INDPTR + batch_idx)
    page_end = tl.load(KV_PAGE_INDPTR + batch_idx + 1)
    if page_end <= page_start:
        offs_v = v_block_idx * block_v + tl.arange(0, block_v)
        mask_v = offs_v < v_dim
        tl.store(
            O + q_start * stride_os + head_idx * stride_oh + offs_v * stride_od,
            0.0,
            mask=mask_v,
        )
        return

    q_row = q_start
    offs_v = v_block_idx * block_v + tl.arange(0, block_v)
    mask_v = offs_v < v_dim

    acc = tl.zeros((block_v,), dtype=tl.float32)
    e_max = -float("inf")
    e_sum = 0.0

    for split_idx in range(0, num_kv_splits):
        tlogic = tl.load(
            MID_LSE
            + batch_idx * stride_mlb
            + split_idx * stride_mls
            + head_idx * stride_mlh
        )
        tv = tl.load(
            MID_O
            + batch_idx * stride_mob
            + split_idx * stride_mos
            + head_idx * stride_moh
            + offs_v * stride_mod,
            mask=mask_v,
            other=0.0,
        )
        n_e_max = tl.maximum(tlogic, e_max)
        old_scale = tl.exp(e_max - n_e_max)
        exp_logic = tl.exp(tlogic - n_e_max)
        acc = acc * old_scale + exp_logic * tv
        e_sum = e_sum * old_scale + exp_logic
        e_max = n_e_max

    out = tl.where(e_sum > 0, acc / e_sum, 0.0)
    tl.store(
        O + q_row * stride_os + head_idx * stride_oh + offs_v * stride_od,
        out.to(tl.bfloat16),
        mask=mask_v,
    )


# ---------------------------------------------------------------------------
# FP8 quantization (sglang style: dynamic per-tensor)
# ---------------------------------------------------------------------------
def quantize_fp8(tensor: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
    return per_tensor_quant_hip(tensor, quant_dtype=FP8_DTYPE)


# ---------------------------------------------------------------------------
# MXFP4 quantization (aiter native: block-32, fp4x2 + fp8_e8m0 dtypes)
# Uses aiter.utility.fp4_utils.dynamic_mxfp4_quant
# ---------------------------------------------------------------------------

def quantize_mxfp4(tensor: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
    """
    MXFP4 block-wise quantization using aiter's dynamic_mxfp4_quant.

    Block size = 32. Each block gets an E8M0 scale factor.
    Two FP4 E2M1 values are packed per byte.

    Args:
        tensor: bf16 tensor of shape [B, M, N] (N must be divisible by 32)

    Returns:
        (fp4_data, scale_e8m0)
        - fp4_data:   shape [B, M, N//2] in aiter_dtypes.fp4x2
        - scale_e8m0: shape [B*M, ceil(N/32)] padded, in aiter_dtypes.fp8_e8m0
    """
    orig_shape = tensor.shape  # (B, M, N)
    B, M, N = orig_shape

    # dynamic_mxfp4_quant expects 2D: (B*M, N)
    tensor_2d = tensor.reshape(B * M, N)
    fp4_data_2d, scale_e8m0 = dynamic_mxfp4_quant(tensor_2d)

    # Reshape fp4_data back to 3D: (B, M, N//2)
    fp4_data = fp4_data_2d.view(B, M, N // 2)

    return fp4_data, scale_e8m0


def dequantize_mxfp4(
    fp4_data: torch.Tensor,
    scale_e8m0: torch.Tensor,
    orig_shape: tuple,
    dtype: torch.dtype = torch.bfloat16,
) -> torch.Tensor:
    """
    Dequantize MXFP4 tensor using aiter utilities.

    Note: dynamic_mxfp4_quant may pad both row and block dimensions in scale_e8m0.
    We trim scales to match the actual data dimensions.

    Args:
        fp4_data:   packed FP4 data, shape [B, M, N//2] in fp4x2 or uint8
        scale_e8m0: E8M0 block scale factors (possibly padded) in fp8_e8m0
        orig_shape: original (B, M, N) for reshaping
        dtype:      output dtype

    Returns:
        Dequantized tensor of shape orig_shape.
    """
    B, M, N = orig_shape
    num_rows = B * M
    block_size = 32
    num_blocks = N // block_size  # actual blocks needed (e.g. 576/32 = 18)

    # Unpack FP4 to float32: mxfp4_to_f32 expects (..., N//2) -> (..., N)
    fp4_data_2d = fp4_data.reshape(num_rows, N // 2)
    float_vals = mxfp4_to_f32(fp4_data_2d)  # (num_rows, N)

    # Convert E8M0 scales to float32 and trim padded dimensions
    scale_f32 = e8m0_to_f32(scale_e8m0)  # (padded_rows, padded_blocks)
    scale_f32 = scale_f32[:num_rows, :num_blocks]  # (num_rows, num_blocks)

    # Apply block scales
    float_vals_blocked = float_vals.view(num_rows, num_blocks, block_size)
    scaled = float_vals_blocked * scale_f32.unsqueeze(-1)

    return scaled.view(B, M, N).to(dtype)


# ---------------------------------------------------------------------------
# Persistent mode metadata helpers
# ---------------------------------------------------------------------------

def _get_kv_indices(total_kv_len: int, device: torch.device) -> torch.Tensor:
    cache_key = (total_kv_len, str(device))
    kv_indices = _KV_INDICES_CACHE.get(cache_key)
    if kv_indices is None:
        kv_indices = torch.arange(total_kv_len, dtype=torch.int32, device=device)
        _KV_INDICES_CACHE[cache_key] = kv_indices
    return kv_indices


def _small_tensor_signature(tensor: torch.Tensor) -> tuple[int, ...]:
    return tuple(int(v) for v in tensor.to(device="cpu", dtype=torch.int32).tolist())


def _get_paged_kv_inputs(
    kv_buffer: torch.Tensor,
    kv_indptr: torch.Tensor,
    nhead_kv: int,
):
    kv_indptr_sig = _small_tensor_signature(kv_indptr)
    cache_key = (
        kv_buffer.data_ptr(),
        kv_indptr_sig,
        tuple(kv_buffer.shape),
        str(kv_buffer.dtype),
        str(kv_buffer.device),
        PAGE_SIZE,
    )
    state = _PAGED_KV_CACHE.get(cache_key)
    current_total_kv = int(kv_indptr[-1].item())
    if state is not None and state["total_kv"] == current_total_kv:
        return (
            state["kv_buffer_4d"],
            state["kv_page_indptr"],
            state["kv_last_page_len"],
            state["kv_indices"],
        )

    seq_lens = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)
    dim = kv_buffer.shape[-1]

    if torch.all(seq_lens % PAGE_SIZE == 0):
        kv_page_counts = torch.div(seq_lens, PAGE_SIZE, rounding_mode="floor")
        kv_buffer_4d = kv_buffer.view(-1, PAGE_SIZE, nhead_kv, dim)
        kv_last_page_len = torch.full_like(seq_lens, PAGE_SIZE)
    else:
        kv_page_counts = torch.div(
            seq_lens + (PAGE_SIZE - 1), PAGE_SIZE, rounding_mode="floor"
        )
        total_pages = int(kv_page_counts.sum().item())
        kv_buffer_4d = torch.empty(
            (total_pages, PAGE_SIZE, nhead_kv, dim),
            dtype=kv_buffer.dtype,
            device=kv_buffer.device,
        )
        kv_buffer_4d.zero_()
        flat_paged_kv = kv_buffer_4d.view(total_pages * PAGE_SIZE, nhead_kv, dim)
        seq_lens_list = seq_lens.tolist()
        kv_indptr_list = kv_indptr.tolist()
        kv_page_counts_list = kv_page_counts.tolist()
        page_start = 0
        for batch_idx, seq_len in enumerate(seq_lens_list):
            token_start = kv_indptr_list[batch_idx]
            token_end = kv_indptr_list[batch_idx + 1]
            flat_start = page_start * PAGE_SIZE
            flat_paged_kv[flat_start : flat_start + seq_len].copy_(
                kv_buffer[token_start:token_end]
            )
            page_start += kv_page_counts_list[batch_idx]

        kv_last_page_len = seq_lens.remainder(PAGE_SIZE)
        kv_last_page_len.masked_fill_(kv_last_page_len == 0, PAGE_SIZE)

    kv_page_indptr = torch.empty_like(kv_indptr)
    kv_page_indptr[0] = 0
    kv_page_indptr[1:] = torch.cumsum(kv_page_counts, dim=0)
    total_pages = int(kv_page_indptr[-1].item())

    state = {
        "kv_buffer_4d": kv_buffer_4d,
        "kv_page_indptr": kv_page_indptr,
        "kv_last_page_len": kv_last_page_len,
        "kv_indices": _get_kv_indices(total_pages, kv_buffer.device),
        "total_kv": current_total_kv,
    }
    _PAGED_KV_CACHE[cache_key] = state
    return (
        state["kv_buffer_4d"],
        state["kv_page_indptr"],
        state["kv_last_page_len"],
        state["kv_indices"],
    )


def _get_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,
):
    """Allocate and cache work buffers for persistent mla_decode_fwd."""
    paged_mode = PAGE_SIZE > 1
    metadata_fast_mode = paged_mode
    metadata_is_causal = False if paged_mode else True
    qo_indptr_sig = _small_tensor_signature(qo_indptr)
    kv_indptr_sig = _small_tensor_signature(kv_indptr)
    kv_last_page_sig = _small_tensor_signature(kv_last_page_len)
    cache_key = (
        batch_size,
        max_q_len,
        nhead,
        nhead_kv,
        str(q_dtype),
        str(kv_dtype),
        qo_indptr_sig,
        kv_indptr_sig,
        kv_last_page_sig,
        str(qo_indptr.device),
        PAGE_SIZE,
        num_kv_splits,
    )
    state = _MLA_METADATA_CACHE.get(cache_key)
    if state is None:
        info = get_mla_metadata_info_v1(
            batch_size,
            max_q_len,
            nhead,
            q_dtype,
            kv_dtype,
            is_sparse=False,
            fast_mode=metadata_fast_mode,
            num_kv_splits=num_kv_splits,
            intra_batch_mode=False,
        )
        work = [torch.empty(s, dtype=t, device=qo_indptr.device) for s, t in info]
        (work_metadata, work_indptr, work_info_set,
         reduce_indptr, reduce_final_map, reduce_partial_map) = work
        state = {
            "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": torch.empty_like(kv_last_page_len),
            "metadata_initialized": False,
        }
        _MLA_METADATA_CACHE[cache_key] = state

    state["kv_last_page_len"].copy_(kv_last_page_len)
    if not state["metadata_initialized"]:
        if paged_mode:
            get_mla_metadata_v1(
                qo_indptr,
                kv_indptr,
                state["kv_last_page_len"],
                nhead // nhead_kv,
                nhead_kv,
                metadata_is_causal,
                state["work_meta_data"],
                state["work_info_set"],
                state["work_indptr"],
                state["reduce_indptr"],
                state["reduce_final_map"],
                state["reduce_partial_map"],
                kv_granularity=max(PAGE_SIZE, 16),
                max_seqlen_qo=max_q_len,
                uni_seqlen_qo=max_q_len,
                fast_mode=metadata_fast_mode,
                max_split_per_batch=num_kv_splits,
                intra_batch_mode=False,
                dtype_q=q_dtype,
                dtype_kv=kv_dtype,
            )
        else:
            get_mla_metadata_v1(
                qo_indptr,
                kv_indptr,
                state["kv_last_page_len"],
                nhead // nhead_kv,
                nhead_kv,
                metadata_is_causal,
                state["work_meta_data"],
                state["work_info_set"],
                state["work_indptr"],
                state["reduce_indptr"],
                state["reduce_final_map"],
                state["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=metadata_fast_mode,
                max_split_per_batch=num_kv_splits,
                intra_batch_mode=False,
                dtype_q=q_dtype,
                dtype_kv=kv_dtype,
            )
        state["metadata_initialized"] = True

    return state


# ---------------------------------------------------------------------------
# Aiter reference kernel (decode only)
# ---------------------------------------------------------------------------

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,
) -> torch.Tensor:
    """
    MLA decode attention using aiter persistent-mode kernel.

    Supports multiple Q/KV dtype combinations:
      - Q_DTYPE="fp8":  fp8 Q + fp8 KV (a8w8) — fastest on MI355X
      - Q_DTYPE="bf16": bf16 Q + bf16 KV (a16w16) — highest precision

    q:          (total_q, num_heads, 576)  fp8 or bf16
    kv_buffer:  (total_kv, 1, 576)         fp8 or bf16
    q_scale:    scalar float32 (required for fp8 Q, None for bf16)
    kv_scale:   scalar float32 (required for fp8 KV, None for bf16)
    """
    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"]

    max_q_len = q_seq_len
    kv_buffer_4d, kv_page_indptr, kv_last_page_len, kv_indices = _get_paged_kv_inputs(
        kv_buffer, kv_indptr, nkv
    )
    o = torch.empty((q.shape[0], nq, dv), dtype=torch.bfloat16, device=q.device)
    if PAGE_SIZE > 1:
        mla_decode_fwd(
            q.view(-1, nq, dq),
            kv_buffer_4d,
            o,
            qo_indptr,
            kv_page_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=PAGED_NUM_KV_SPLITS,
            q_scale=q_scale,
            kv_scale=kv_scale,
            intra_batch_mode=False,
        )
    else:
        meta = _get_mla_decode_metadata(
            batch_size,
            max_q_len,
            nq,
            nkv,
            q.dtype,
            kv_buffer_4d.dtype,
            qo_indptr,
            kv_page_indptr,
            kv_last_page_len,
            num_kv_splits=NUM_KV_SPLITS,
        )
        mla_meta = {
            "work_meta_data": meta["work_meta_data"],
            "work_indptr": meta["work_indptr"],
            "work_info_set": meta["work_info_set"],
            "reduce_indptr": meta["reduce_indptr"],
            "reduce_final_map": meta["reduce_final_map"],
            "reduce_partial_map": meta["reduce_partial_map"],
        }
        mla_decode_fwd(
            q.view(-1, nq, dq),
            kv_buffer_4d,
            o,
            qo_indptr,
            kv_page_indptr,
            kv_indices,
            meta["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=False,
            **mla_meta,
        )
    return o


def _triton_mla_decode(
    q: torch.Tensor,
    kv_buffer: torch.Tensor,
    qo_indptr: torch.Tensor,
    kv_indptr: torch.Tensor,
    config: dict,
) -> torch.Tensor:
    batch_size = config["batch_size"]
    nq = config["num_heads"]
    dv = config["v_head_dim"]
    q_seq_len = config["q_seq_len"]

    if q_seq_len != 1:
        raise RuntimeError(f"triton pg8 kernel only supports q_seq_len == 1, got {q_seq_len}")

    kv_buffer_4d, kv_page_indptr, kv_last_page_len, _ = _get_paged_kv_inputs(
        kv_buffer, kv_indptr, NUM_KV_HEADS
    )
    block_v = 128
    num_v_blocks = triton.cdiv(dv, block_v)
    mid_o = torch.empty(
        (batch_size, NUM_KV_SPLITS, nq, dv), dtype=torch.float32, device=q.device
    )
    mid_lse = torch.empty(
        (batch_size, NUM_KV_SPLITS, nq), dtype=torch.float32, device=q.device
    )
    o = torch.empty((q.shape[0], nq, dv), dtype=torch.bfloat16, device=q.device)

    stage1_grid = (batch_size, nq, NUM_KV_SPLITS * num_v_blocks)
    _mla_decode_paged_stage1_kernel[stage1_grid](
        q,
        kv_buffer_4d,
        qo_indptr,
        kv_page_indptr,
        kv_last_page_len,
        mid_o,
        mid_lse,
        q.stride(0),
        q.stride(1),
        q.stride(2),
        kv_buffer_4d.stride(0),
        kv_buffer_4d.stride(1),
        kv_buffer_4d.stride(2),
        kv_buffer_4d.stride(3),
        mid_o.stride(0),
        mid_o.stride(1),
        mid_o.stride(2),
        mid_o.stride(3),
        mid_lse.stride(0),
        mid_lse.stride(1),
        mid_lse.stride(2),
        SM_SCALE,
        num_v_blocks=num_v_blocks,
        num_kv_splits=NUM_KV_SPLITS,
        qk_dim=QK_HEAD_DIM,
        v_dim=V_HEAD_DIM,
        page_size=PAGE_SIZE,
        block_n=32,
        block_d=128,
        block_v=block_v,
        num_warps=4,
        num_stages=2,
    )

    stage2_grid = (batch_size, nq, num_v_blocks)
    _mla_decode_paged_stage2_kernel[stage2_grid](
        mid_o,
        mid_lse,
        qo_indptr,
        kv_page_indptr,
        o,
        mid_o.stride(0),
        mid_o.stride(1),
        mid_o.stride(2),
        mid_o.stride(3),
        mid_lse.stride(0),
        mid_lse.stride(1),
        mid_lse.stride(2),
        o.stride(0),
        o.stride(1),
        o.stride(2),
        num_kv_splits=NUM_KV_SPLITS,
        v_dim=V_HEAD_DIM,
        block_v=block_v,
        num_warps=4,
        num_stages=2,
    )
    return o

def custom_kernel(data: input_t) -> output_t:
    """Triton MLA decode over paged bf16 KV."""
    q, kv_data, qo_indptr, kv_indptr, config = data
    q = q.contiguous()
    kv_input = kv_data["bf16"]
    return _triton_mla_decode(q, kv_input, qo_indptr, kv_indptr, config)