submission 752051

"""
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 torch
import torch.nn.functional as F
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 (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 = 1
NUM_KV_SPLITS = 32

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

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

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


# ---------------------------------------------------------------------------
# 优化后的量化函数（统一处理 FP8 和 MXFP4）
# ---------------------------------------------------------------------------
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 = 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


# ---------------------------------------------------------------------------
# Persistent mode metadata helpers（保留原样，暂时不改）
# ---------------------------------------------------------------------------
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,
):
    """Allocate and populate work buffers for persistent mla_decode_fwd."""
    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,
    }


# ---------------------------------------------------------------------------
# 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 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_indices = torch.arange(total_kv_len, dtype=torch.int32, device="cuda")

    kv_buffer_4d = kv_buffer.view(kv_buffer.shape[0], PAGE_SIZE, nkv, kv_buffer.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, nq, nkv,
        q.dtype, kv_buffer.dtype,
        qo_indptr, kv_indptr, kv_last_page_len,
        num_kv_splits=NUM_KV_SPLITS,
    )

    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


# ---------------------------------------------------------------------------
# generate_input（保留原逻辑，只做小优化）
# ---------------------------------------------------------------------------
def generate_input(batchsize: int, qseqlen: int, kvseqlen: int, seed: int) -> input_t:
    """Generate absorbed q and compressed kv_buffer for MLA decode."""
    gen = torch.Generator(device="cuda")
    gen.manual_seed(seed)

    total_q = batchsize * qseqlen
    total_kv = batchsize * kvseqlen

    q = torch.randn(
        (total_q, NUM_HEADS, QK_HEAD_DIM),
        dtype=torch.bfloat16, device="cuda", generator=gen,
    )

    kv_buffer_bf16 = torch.randn(
        (total_kv, NUM_KV_HEADS, QK_HEAD_DIM),
        dtype=torch.bfloat16, device="cuda", generator=gen,
    )

    kv_buffer_fp8, kv_scale_fp8 = quantize_fp8(kv_buffer_bf16)
    kv_buffer_mxfp4, kv_scale_mxfp4 = quantize_mxfp4(kv_buffer_bf16)

    kv_data = {
        "bf16": kv_buffer_bf16,
        "fp8": (kv_buffer_fp8, kv_scale_fp8),
        "mxfp4": (kv_buffer_mxfp4, kv_scale_mxfp4),
    }

    qo_indptr = torch.arange(0, batchsize + 1, dtype=torch.int32, device="cuda") * qseqlen
    kv_indptr = torch.arange(0, batchsize + 1, dtype=torch.int32, device="cuda") * kvseqlen

    config = {
        "batch_size": batchsize,
        "num_heads": NUM_HEADS,
        "num_kv_heads": NUM_KV_HEADS,
        "qk_head_dim": QK_HEAD_DIM,
        "kv_lora_rank": KV_LORA_RANK,
        "qk_rope_head_dim": QK_ROPE_HEAD_DIM,
        "v_head_dim": V_HEAD_DIM,
        "q_seq_len": qseqlen,
        "kv_seq_len": kvseqlen,
        "sm_scale": SM_SCALE,
    }

    return (q, kv_data, qo_indptr, kv_indptr, config)


# ---------------------------------------------------------------------------
# 优化后的 ref_kernel（最终版）
# ---------------------------------------------------------------------------
def ref_kernel(data: input_t) -> output_t:
    """优化版 MLA Decode reference kernel"""
    q, kv_data, qo_indptr, kv_indptr, config = data

    # Resolve Q and KV based on dtype
    if Q_DTYPE == "fp8":
        q_input, q_scale = quantize_fp8(q)
    else:
        q_input, q_scale = q, None

    if KV_DTYPE == "fp8":
        kv_buffer, kv_scale = kv_data["fp8"]
        kv_input = kv_buffer
    else:
        kv_input, kv_scale = kv_data["bf16"], None

    # 内存友好处理
    q_input = q_input.contiguous()
    kv_input = kv_input.contiguous()
    qo_indptr = qo_indptr.contiguous()
    kv_indptr = kv_indptr.contiguous()

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

custom_kernel = ref_kernel

check_implementation = make_match_reference(ref_kernel, rtol=1e-01, atol=1e-01)