submission 744143

"""
MLA Decode Kernel - Optimized Direct Kernel Call
Strategy: Direct aiter kernel invocation with optimized 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
PAGE_SIZE = 1
SM_SCALE = 0.041666666666666664  # 1/sqrt(576)

# Optimized: reduce KV splits for smaller overhead
NUM_KV_SPLITS = 16  # Reduced from 32

FP8_DTYPE = aiter_dtypes.fp8


def quantize_fp8_fast(tensor: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
    """Optimized FP8 quantization with minimal overhead."""
    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 custom_kernel(data: input_t) -> output_t:
    """
    Optimized MLA decode with direct kernel call.
    
    Key optimizations:
    1. Reduced NUM_KV_SPLITS from 32 to 16 (less overhead)
    2. Pre-allocated output buffer
    3. Minimal metadata overhead
    4. Direct kernel invocation
    """
    q, kv_data, qo_indptr, kv_indptr, config = data
    
    batch_size = config["batch_size"]
    q_seq_len = config["q_seq_len"]
    
    # Use FP8 KV cache
    kv_buffer_fp8, kv_scale = kv_data["fp8"]
    
    # Fast FP8 quantization
    q_fp8, q_scale = quantize_fp8_fast(q)
    
    # Prepare KV buffer for aiter (4D format)
    total_kv_len = int(kv_indptr[-1].item())
    kv_buffer_4d = kv_buffer_fp8.view(total_kv_len, PAGE_SIZE, NUM_KV_HEADS, 576)
    
    # Create metadata (optimized with fewer splits)
    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_fp8.dtype, kv_buffer_fp8.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,  # is_causal
        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_fp8.dtype,
        dtype_kv=kv_buffer_fp8.dtype,
    )
    
    # Pre-allocate output
    o = torch.empty((q.shape[0], NUM_HEADS, 512), dtype=torch.bfloat16, device="cuda")
    
    # KV indices
    kv_indices = torch.arange(total_kv_len, dtype=torch.int32, device="cuda")
    
    # Direct kernel call
    mla_decode_fwd(
        q_fp8.view(-1, NUM_HEADS, 576),
        kv_buffer_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