submission 675048

#!POPCORN leaderboard amd-mixed-mla
#!POPCORN gpu MI355X

"""
MLA Decode v5: Adaptive dtype selection.
- Small KV (kv_len * bs <= threshold): bf16 (skip quant overhead)
- Large KV: fp8 (bandwidth savings dominate)
"""

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
V_HEAD_DIM = KV_LORA_RANK
SM_SCALE = 1.0 / (QK_HEAD_DIM ** 0.5)

PAGE_SIZE = 1
NUM_KV_SPLITS = 32

FP8_DTYPE = aiter_dtypes.fp8
_FP8_MAX = torch.finfo(FP8_DTYPE).max
_FP8_MIN = torch.finfo(FP8_DTYPE).min

# Threshold: when total KV tokens > this, use fp8 for bandwidth savings
# Based on benchmark: fp8 wins when kv_seq_len >= 8192 AND bs >= 32
KV_THRESHOLD = 32 * 4096  # ~131k tokens


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"]
    total_kv_len = int(kv_indptr[-1].item())

    # Adaptive: choose bf16 or fp8 based on total KV volume
    use_fp8 = total_kv_len > KV_THRESHOLD

    if use_fp8:
        # fp8 path: quantize Q, use pre-quantized KV
        amax = q.abs().amax().clamp(min=1e-12)
        q_scale = (amax / _FP8_MAX).to(torch.float32).reshape(1)
        q_input = (q / q_scale).clamp(min=_FP8_MIN, max=_FP8_MAX).to(FP8_DTYPE)
        kv_buffer, kv_scale = kv_data["fp8"]
    else:
        # bf16 path: zero quant overhead
        q_input = q
        q_scale = None
        kv_buffer = kv_data["bf16"]
        kv_scale = None

    kv_4d = kv_buffer.view(kv_buffer.shape[0], PAGE_SIZE, NUM_KV_HEADS, kv_buffer.shape[-1])

    kv_indices = torch.arange(total_kv_len, dtype=torch.int32, device="cuda")
    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_input.dtype, kv_buffer.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,
        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_KV_SPLITS,
        intra_batch_mode=True,
        dtype_q=q_input.dtype,
        dtype_kv=kv_buffer.dtype,
    )

    o = torch.empty((q.shape[0], NUM_HEADS, V_HEAD_DIM), dtype=torch.bfloat16, device="cuda")
    mla_decode_fwd(
        q_input.view(-1, NUM_HEADS, 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,
        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