submission 754460

"""
MLA v4: Hybrid bf16/fp8 Q approach.
- Small workloads: bf16 Q + fp8 KV (skip quantize_fp8, use a16w8 kernel) — 36% faster
- Large workloads: fp8 Q + fp8 KV (a8w8 kernel) — avoids 2x Q bandwidth penalty

From v3 benchmarks:
  bf16 Q wins on bs<=64 (92-224µs vs baseline 145-260µs)
  fp8 Q wins on bs=256,kv=8192 (401µs vs bf16's 662µs)

Metadata buffer caching, output pre-allocation, kv_indices caching.
"""
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
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 = aiter_dtypes.fp8

# Threshold: if total_kv_len > this, use fp8 Q (a8w8 kernel)
# bs=256 * kv=8192 = 2M → fp8 is better
# bs=64 * kv=8192 = 524K → bf16 is still better
_FP8_Q_THRESHOLD = 1000000

# Caches
_meta_cache = {}
_out_cache = {}
_idx_cache = {}


def _quantize_fp8(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 _get_metadata_cached(batch_size, max_q_len, nq, nkv, q_dtype, kv_dtype,
                         qo_indptr, kv_indptr, kv_last_page_len):
    key = (batch_size, int(kv_indptr[-1].item()), q_dtype, kv_dtype)
    if key not in _meta_cache:
        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,
        )
        _meta_cache[key] = [torch.empty(s, dtype=t, device="cuda") for s, t in info]

    work = _meta_cache[key]
    (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,
    )

    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,
    }


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

    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())

    # Decide Q dtype based on workload size
    use_fp8_q = total_kv_len > _FP8_Q_THRESHOLD

    if use_fp8_q:
        # Large workload: fp8 Q + fp8 KV (a8w8 kernel)
        q_input, q_scale = _quantize_fp8(q)
    else:
        # Small workload: bf16 Q + fp8 KV (a16w8 kernel, skip quant)
        q_input = q
        q_scale = None

    # fp8 KV always
    kv_buffer_fp8, kv_scale = kv_data["fp8"]
    kv_buffer_4d = kv_buffer_fp8.view(kv_buffer_fp8.shape[0], PAGE_SIZE, nkv, kv_buffer_fp8.shape[-1])

    # Cached kv_indices
    idx_key = total_kv_len
    if idx_key not in _idx_cache:
        _idx_cache[idx_key] = torch.arange(total_kv_len, dtype=torch.int32, device="cuda")
    kv_indices = _idx_cache[idx_key]

    kv_last_page_len = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)

    meta = _get_metadata_cached(
        batch_size, q_seq_len, nq, nkv,
        q_input.dtype, kv_buffer_fp8.dtype,
        qo_indptr, kv_indptr, kv_last_page_len,
    )

    # Cached output tensor
    out_key = (q.shape[0], nq, dv)
    if out_key not in _out_cache:
        _out_cache[out_key] = torch.empty(out_key, dtype=torch.bfloat16, device="cuda")
    o = _out_cache[out_key]

    mla_decode_fwd(
        q_input.view(-1, nq, dq),
        kv_buffer_4d,
        o,
        qo_indptr,
        kv_indptr,
        kv_indices,
        kv_last_page_len,
        q_seq_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