submission 608379

"""
Optimized MLA decode kernel for MI355X.
Key optimizations over reference:
- Cache persistent-mode metadata across calls (same shape → reuse buffers)
- Precompute constants at module level
- Minimize Python overhead in hot path
"""
from task import input_t, output_t
import torch

from aiter.mla import mla_decode_fwd
from aiter import dtypes as aiter_dtypes, get_mla_metadata_info_v1, get_mla_metadata_v1

# DeepSeek R1 MLA constants
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)
NUM_KV_HEADS = 1
PAGE_SIZE = 1
NUM_KV_SPLITS = 32
FP8_DTYPE = aiter_dtypes.fp8

# Cache metadata and work buffers per shape
_meta_cache = {}


def _get_cached_metadata(batch_size, max_q_len, nhead, q_dtype, kv_dtype, qo_indptr, kv_indptr):
    key = (batch_size, max_q_len, nhead, q_dtype, kv_dtype)
    if key not in _meta_cache:
        nkv = NUM_KV_HEADS
        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]
        _meta_cache[key] = work

    work = _meta_cache[key]
    (work_metadata, work_indptr, work_info_set,
     reduce_indptr, reduce_final_map, reduce_partial_map) = work

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

    get_mla_metadata_v1(
        qo_indptr, kv_indptr, kv_last_page_len,
        nhead // 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=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,
    }, kv_last_page_len


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

    batch_size = config["batch_size"]
    nhead = config["num_heads"]
    q_seq_len = config["q_seq_len"]
    total_kv_len = int(kv_indptr[-1].item())

    # FP8 quantize Q (dynamic per-tensor)
    finfo = torch.finfo(FP8_DTYPE)
    amax = q.abs().amax().clamp(min=1e-12)
    q_scale = (amax / finfo.max).to(torch.float32).reshape(1)
    q_fp8 = (q / q_scale).clamp(min=finfo.min, max=finfo.max).to(FP8_DTYPE)

    # FP8 KV
    kv_buffer_fp8, kv_scale = kv_data["fp8"]

    # Reshape KV to 4D: (total_kv, page_size=1, nkv=1, 576)
    kv_4d = kv_buffer_fp8.view(kv_buffer_fp8.shape[0], PAGE_SIZE, NUM_KV_HEADS, kv_buffer_fp8.shape[-1])

    # Cached metadata + KV indices
    meta, kv_last_page_len = _get_cached_metadata(
        batch_size, q_seq_len, nhead, q_fp8.dtype, kv_buffer_fp8.dtype,
        qo_indptr, kv_indptr,
    )

    kv_indices = torch.arange(total_kv_len, dtype=torch.int32, device="cuda")

    o = torch.empty((q.shape[0], nhead, V_HEAD_DIM), dtype=torch.bfloat16, device="cuda")
    mla_decode_fwd(
        q_fp8.view(-1, nhead, QK_HEAD_DIM),
        kv_4d,
        o,
        qo_indptr,
        kv_indptr,
        kv_indices,
        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=NUM_KV_SPLITS,
        q_scale=q_scale,
        kv_scale=kv_scale,
        intra_batch_mode=True,
        **meta,
    )
    return o