submission 693793

#!POPCORN leaderboard amd-mixed-mla
#!POPCORN gpu MI355X
# v10-compile: v9 + torch.compile fused Q quantization.
# Eliminates .item() sync AND fuses Q quant ops into fewer kernel launches.

import torch
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

NUM_HEADS = 16
NUM_KV_HEADS = 1
QK_HEAD_DIM = 576
V_HEAD_DIM = 512
SM_SCALE = 1.0 / (QK_HEAD_DIM ** 0.5)
PAGE_SIZE = 1
NUM_KV_SPLITS = 32
FP8_DTYPE = aiter_dtypes.fp8
_FP8_FINFO = torch.finfo(FP8_DTYPE)
_FP8_MAX = _FP8_FINFO.max
_FP8_MIN = _FP8_FINFO.min

_cache = {}


@torch.compile(fullgraph=True)
def _quantize_fp8_compiled(tensor):
    amax = tensor.abs().amax().clamp(min=1e-12)
    scale = amax / _FP8_MAX
    fp8_tensor = (tensor / scale).clamp(min=_FP8_MIN, max=_FP8_MAX).to(FP8_DTYPE)
    return fp8_tensor, scale.to(torch.float32).reshape(1)


def _get_or_build_cache(batch_size, total_kv, q_dtype, kv_dtype,
                        qo_indptr, kv_indptr):
    key = (batch_size, total_kv, q_dtype, kv_dtype)
    cached = _cache.get(key)
    if cached is not None:
        return cached

    max_q_len = 1
    nq, nkv = NUM_HEADS, NUM_KV_HEADS

    kv_last_page_len = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)
    kv_indices = torch.arange(total_kv, dtype=torch.int32, device="cuda")

    info = get_mla_metadata_info_v1(
        batch_size, max_q_len, nq, 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,
        nq // nkv, nkv, 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,
    )

    output_buf = torch.empty((batch_size, nq, V_HEAD_DIM),
                             dtype=torch.bfloat16, device="cuda")

    cached = {
        "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,
        },
        "kv_indices": kv_indices,
        "kv_last_page_len": kv_last_page_len,
        "output_buf": output_buf,
    }
    _cache[key] = cached
    return cached


def custom_kernel(data: input_t) -> output_t:
    q, kv_data, qo_indptr, kv_indptr, config = data
    batch_size = config["batch_size"]

    kv_buffer_fp8, kv_scale = kv_data["fp8"]
    q_fp8, q_scale = _quantize_fp8_compiled(q)

    total_kv = batch_size * config["kv_seq_len"]
    cached = _get_or_build_cache(
        batch_size, total_kv, q_fp8.dtype, kv_buffer_fp8.dtype,
        qo_indptr, kv_indptr,
    )

    kv_4d = kv_buffer_fp8.view(total_kv, PAGE_SIZE, NUM_KV_HEADS, QK_HEAD_DIM)
    o = cached["output_buf"]

    mla_decode_fwd(
        q_fp8.view(-1, NUM_HEADS, QK_HEAD_DIM),
        kv_4d,
        o,
        qo_indptr,
        kv_indptr,
        cached["kv_indices"],
        cached["kv_last_page_len"],
        1,
        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,
        **cached["meta"],
    )
    return o