submission 637614

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,
)

# ---------------------------------------------------------------------------
# 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
V_HEAD_DIM = KV_LORA_RANK
SM_SCALE = 1.0 / (QK_HEAD_DIM ** 0.5)

PAGE_SIZE = 1
NUM_KV_SPLITS = 32

FP8_DTYPE = aiter_dtypes.fp8
Q_DTYPE = "fp8"
KV_DTYPE = "fp8"

# ---------------------------------------------------------------------------
# SAFE caches (only invariant things)
# ---------------------------------------------------------------------------
_KV_INDICES_CACHE = {}
_OUTPUT_CACHE = {}

# ---------------------------------------------------------------------------
# FP8 quantization
# ---------------------------------------------------------------------------
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)

# ---------------------------------------------------------------------------
# MXFP4 quantization
# ---------------------------------------------------------------------------
def quantize_mxfp4(tensor: torch.Tensor):
    B, M, N = tensor.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 dequantize_mxfp4(fp4_data, scale_e8m0, orig_shape, dtype=torch.bfloat16):
    B, M, N = orig_shape
    num_rows = B * M
    block_size = 32
    num_blocks = N // block_size

    fp4_data_2d = fp4_data.reshape(num_rows, N // 2)
    float_vals = mxfp4_to_f32(fp4_data_2d)

    scale_f32 = e8m0_to_f32(scale_e8m0)
    scale_f32 = scale_f32[:num_rows, :num_blocks]

    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)

# ---------------------------------------------------------------------------
# Metadata (NO caching — correctness critical)
# ---------------------------------------------------------------------------
def _make_mla_decode_metadata(
    batch_size,
    max_q_len,
    nhead,
    nhead_kv,
    q_dtype,
    kv_dtype,
    qo_indptr,
    kv_indptr,
    kv_last_page_len,
):
    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,
    }

# ---------------------------------------------------------------------------
# MLA decode (optimized safely)
# ---------------------------------------------------------------------------
def _aiter_mla_decode(
    q,
    kv_buffer,
    qo_indptr,
    kv_indptr,
    config,
    q_scale=None,
    kv_scale=None,
):
    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())

    # ---- SAFE cache: kv_indices ----
    if total_kv_len not in _KV_INDICES_CACHE:
        _KV_INDICES_CACHE[total_kv_len] = torch.arange(
            total_kv_len, dtype=torch.int32, device="cuda"
        )
    kv_indices = _KV_INDICES_CACHE[total_kv_len]

    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)

    # ---- NO CACHE: metadata must match structure ----
    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,
    )

    # ---- SAFE output reuse ----
    out_key = (q.shape[0], nq, dv)
    if out_key not in _OUTPUT_CACHE:
        _OUTPUT_CACHE[out_key] = torch.empty(
            out_key, dtype=torch.bfloat16, device="cuda"
        )

    o = _OUTPUT_CACHE[out_key]
    o.zero_()  # prevent stale memory

    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

# ---------------------------------------------------------------------------
# Input generator
# ---------------------------------------------------------------------------
def generate_input(batchsize, qseqlen, kvseqlen, seed):
    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_data = {
        "bf16": kv_buffer_bf16,
        "fp8": (kv_buffer_fp8, kv_scale_fp8),
    }

    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,
        "v_head_dim": V_HEAD_DIM,
        "q_seq_len": qseqlen,
        "kv_seq_len": kvseqlen,
    }

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

# ---------------------------------------------------------------------------
# Reference kernel
# ---------------------------------------------------------------------------
def ref_kernel(data: input_t) -> output_t:
    q, kv_data, qo_indptr, kv_indptr, config = data

    if Q_DTYPE == "fp8":
        q_input, q_scale = quantize_fp8(q)
    else:
        q_input, q_scale = q, None

    if KV_DTYPE == "fp8":
        kv_input, kv_scale = kv_data["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,
    )

# ---------------------------------------------------------------------------
# REQUIRED ENTRYPOINT
# ---------------------------------------------------------------------------
def custom_kernel(data: input_t) -> output_t:
    return ref_kernel(data)

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