submission 607984

import torch
import sys
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
import aiter

NUM_HEADS = 16
NUM_KV_HEADS = 1
QK_HEAD_DIM = 576
V_HEAD_DIM = 512
SM_SCALE = 1.0 / (QK_HEAD_DIM ** 0.5)
FP8_DTYPE = aiter_dtypes.fp8
_cache = {}
_direct_mode = None

def _quantize_fp8(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)


def _get_shape_config(batch_size, kv_seq_len):
    total = batch_size * kv_seq_len

    if kv_seq_len <= 1024:
        page_size = 2
        if batch_size <= 4:
            num_splits = 4
        elif batch_size <= 64:
            num_splits = 8
        else:
            num_splits = 16
    else:
        page_size = 8
        if batch_size <= 4:
            num_splits = 8
        elif batch_size <= 32:
            num_splits = 16
        else:
            num_splits = 32

    use_bf16_q = total < 500000
    return page_size, num_splits, use_bf16_q

def _setup(batch_size, q_seq_len, kv_seq_len, qo_indptr, kv_indptr,
           q_dtype, kv_dtype, page_size, num_splits):
    key = (batch_size, q_seq_len, kv_seq_len, q_dtype, kv_dtype, page_size, num_splits)
    if key in _cache:
        return _cache[key]

    total_kv = batch_size * kv_seq_len
    total_q = batch_size * q_seq_len
    num_pages = total_kv // page_size
    kv_indices = torch.arange(num_pages, dtype=torch.int32, device="cuda")
    kv_last_page_len = torch.full((batch_size,), page_size, dtype=torch.int32, device="cuda")
    kv_indptr_pages = kv_indptr // page_size
    info = get_mla_metadata_info_v1(
        batch_size, q_seq_len, NUM_HEADS, q_dtype, kv_dtype,
        is_sparse=False, fast_mode=False,
        num_kv_splits=num_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_pages, kv_last_page_len,
        NUM_HEADS // NUM_KV_HEADS, NUM_KV_HEADS, 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=q_seq_len, uni_seqlen_qo=q_seq_len,
        fast_mode=False, max_split_per_batch=num_splits,
        intra_batch_mode=True,
        dtype_q=q_dtype, dtype_kv=kv_dtype,
    )
    o = torch.empty((total_q, NUM_HEADS, V_HEAD_DIM), dtype=torch.bfloat16, device="cuda")
    entry = {
        "kv_indices": kv_indices,
        "kv_last_page_len": kv_last_page_len,
        "kv_indptr_pages": kv_indptr_pages,
        "num_kv_splits": num_splits,
        "page_size": page_size,
        "o": o,
        "meta": {
            "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,
        },
    }
    _cache[key] = entry
    return entry

def custom_kernel(data: input_t) -> output_t:
    q, kv_data, qo_indptr, kv_indptr, config = data
    batch_size = config["batch_size"]
    q_seq_len = config["q_seq_len"]
    kv_seq_len = config["kv_seq_len"]
    kv_buffer_fp8, kv_scale = kv_data["fp8"]
    page_size, num_splits, use_bf16_q = _get_shape_config(batch_size, kv_seq_len)
    if use_bf16_q:
        q_input, q_scale, q_dtype = q, None, q.dtype
    else:
        q_input, q_scale = _quantize_fp8(q)
        q_dtype = q_input.dtype
    setup = _setup(batch_size, q_seq_len, kv_seq_len,
                   qo_indptr, kv_indptr,
                   q_dtype, kv_buffer_fp8.dtype,
                   page_size, num_splits)

    kv_buffer_4d = kv_buffer_fp8.view(-1, page_size, NUM_KV_HEADS, QK_HEAD_DIM)
    o = setup["o"]
    mla_decode_fwd(
        q_input.view(-1, NUM_HEADS, QK_HEAD_DIM),
        kv_buffer_4d, o,
        qo_indptr, setup["kv_indptr_pages"],
        setup["kv_indices"], setup["kv_last_page_len"],
        q_seq_len,
        page_size=page_size, nhead_kv=NUM_KV_HEADS,
        sm_scale=SM_SCALE, logit_cap=0.0,
        num_kv_splits=setup["num_kv_splits"],
        q_scale=q_scale, kv_scale=kv_scale,
        intra_batch_mode=True,
        **setup["meta"],
    )
    return o

def ref_kernel(data: input_t) -> output_t:
    q, kv_data, qo_indptr, kv_indptr, config = data
    q_fp8, q_scale = _quantize_fp8(q)
    kv_buffer_fp8, kv_scale = kv_data["fp8"]
    batch_size = config["batch_size"]
    q_seq_len = config["q_seq_len"]
    total_kv = int(kv_indptr[-1].item())
    kv_indices = torch.arange(total_kv, dtype=torch.int32, device="cuda")
    kv_buffer_4d = kv_buffer_fp8.view(-1, 1, NUM_KV_HEADS, QK_HEAD_DIM)
    kv_last_page_len = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)
    info = get_mla_metadata_info_v1(
        batch_size, q_seq_len, NUM_HEADS, q_fp8.dtype, kv_buffer_fp8.dtype,
        is_sparse=False, fast_mode=False,
        num_kv_splits=32, intra_batch_mode=True,
    )
    work = [torch.empty(s, dtype=t, device="cuda") for s, t in info]
    (wm, wi, wis, ri, rfm, rpm) = work
    get_mla_metadata_v1(
        qo_indptr, kv_indptr, kv_last_page_len,
        NUM_HEADS // NUM_KV_HEADS, NUM_KV_HEADS, True,
        wm, wis, wi, ri, rfm, rpm,
        page_size=1, kv_granularity=16,
        max_seqlen_qo=q_seq_len, uni_seqlen_qo=q_seq_len,
        fast_mode=False, max_split_per_batch=32,
        intra_batch_mode=True,
        dtype_q=q_fp8.dtype, dtype_kv=kv_buffer_fp8.dtype,
    )
    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, q_seq_len,
        page_size=1, nhead_kv=NUM_KV_HEADS,
        sm_scale=SM_SCALE, logit_cap=0.0, num_kv_splits=32,
        q_scale=q_scale, kv_scale=kv_scale, intra_batch_mode=True,
        work_meta_data=wm, work_indptr=wi, work_info_set=wis,
        reduce_indptr=ri, reduce_final_map=rfm, reduce_partial_map=rpm,
    )
    return o

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