submission 663193

"""
Optimized MLA decode kernel based on aiter fp8 (a8w8) reference.

Optimization strategy: Minimize Python-side overhead around the aiter kernel.
The aiter mla_decode_fwd kernel itself is hand-tuned assembly — we can't change it.
But we CAN optimize everything around it:

1. Use aiter's native scaled_fp8_quant for Q quantization (fused GPU kernel vs manual)
2. Pre-allocate and cache reusable tensors (kv_indices, output, metadata buffers)
3. Minimize tensor operations (avoid unnecessary .to(), .view(), .reshape())
4. Try fast_mode=True for metadata generation
5. Experiment with kv_granularity and other metadata parameters
"""

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

# ---------------------------------------------------------------------------
# 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   # 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

# ---------------------------------------------------------------------------
# Try to import aiter's optimized scaled_fp8_quant
# This is a fused CUDA/HIP kernel that does amax + scale + quantize in one pass
# Much faster than the manual 3-step approach in reference.py
# ---------------------------------------------------------------------------
try:
    from aiter import scaled_fp8_quant
    _HAS_AITER_FP8_QUANT = True
except ImportError:
    _HAS_AITER_FP8_QUANT = False

# ---------------------------------------------------------------------------
# Caches to avoid repeated allocations
# ---------------------------------------------------------------------------
_cache = {}


def _quantize_fp8_fast(tensor: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
    """FP8 quantization — use aiter fused kernel if available."""
    if _HAS_AITER_FP8_QUANT:
        return scaled_fp8_quant(tensor)
    # Fallback: manual (same as reference)
    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_cached(key, factory):
    """Get or create a cached tensor/object."""
    if key not in _cache:
        _cache[key] = factory()
    return _cache[key]


def custom_kernel(data: input_t) -> output_t:
    """Optimized MLA decode — minimize overhead around aiter a8w8 kernel."""
    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"]
    kv_seq_len = config["kv_seq_len"]

    total_q = q.shape[0]
    total_kv_len = batch_size * kv_seq_len  # uniform lengths

    # --- 1. Quantize Q to fp8 ---
    q_fp8, q_scale = _quantize_fp8_fast(q)

    # --- 2. Get pre-quantized fp8 KV (already in input, zero cost) ---
    kv_buffer_fp8, kv_scale = kv_data["fp8"]

    # --- 3. Reshape KV to 4D (just a view, no copy) ---
    kv_buffer_4d = kv_buffer_fp8.view(kv_buffer_fp8.shape[0], PAGE_SIZE, nkv, kv_buffer_fp8.shape[-1])

    # --- 4. Cached kv_indices ---
    cache_key_indices = ("kv_indices", total_kv_len)
    kv_indices = _get_cached(cache_key_indices,
        lambda: torch.arange(total_kv_len, dtype=torch.int32, device="cuda"))

    # --- 5. kv_last_page_len (simple subtraction) ---
    kv_last_page_len = (kv_indptr[1:] - kv_indptr[:-1]).to(torch.int32)

    # --- 6. Build metadata ---
    # The metadata depends on (batch_size, kv_seq_len, q_dtype, kv_dtype)
    # For uniform-length batches, we can cache the metadata buffer allocations
    max_q_len = q_seq_len

    cache_key_meta = ("meta_info", batch_size, kv_seq_len, q_fp8.dtype, kv_buffer_fp8.dtype)
    
    def _build_meta_buffers():
        info = get_mla_metadata_info_v1(
            batch_size, max_q_len, nq, q_fp8.dtype, kv_buffer_fp8.dtype,
            is_sparse=False, fast_mode=False,
            num_kv_splits=NUM_KV_SPLITS, intra_batch_mode=True,
        )
        return [torch.empty(s, dtype=t, device="cuda") for s, t in info]

    work = _get_cached(cache_key_meta, _build_meta_buffers)
    (work_metadata, work_indptr, work_info_set,
     reduce_indptr, reduce_final_map, reduce_partial_map) = work

    # Populate metadata (must be done every call since indptr may differ)
    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_fp8.dtype,
        dtype_kv=kv_buffer_fp8.dtype,
    )

    # --- 7. Cached output tensor ---
    cache_key_out = ("output", total_q, nq, dv)
    o = _get_cached(cache_key_out,
        lambda: torch.empty((total_q, nq, dv), dtype=torch.bfloat16, device="cuda"))
    # Ensure correct size (in case batch changes)
    if o.shape[0] != total_q:
        o = torch.empty((total_q, nq, dv), dtype=torch.bfloat16, device="cuda")
        _cache[cache_key_out] = o

    # --- 8. Call aiter kernel ---
    mla_decode_fwd(
        q_fp8.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,
        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,
    )
    return o