submission 754755

#!POPCORN gpu MI355X
"""
Kernel B: MLA Decode — Multi-Head Latent Attention for DeepSeek-R1
Inner attention kernel from the forward_absorb MLA path.

Target: Beat AITER a8w8 MLA decode reference on AMD MI355X (CDNA 4)

Architecture (DeepSeek R1 forward_absorb MLA):
  - num_heads = 16 (query heads, after TP split)
  - num_kv_heads = 1 (shared latent KV head — MQA)
  - kv_lora_rank = 512
  - qk_rope_head_dim = 64
  - qk_head_dim = 576 (512 + 64, absorbed Q/K dim)
  - v_head_dim = 512 (= kv_lora_rank, output dim)
  - sm_scale = 1/sqrt(576)
  - Decode only: q_seq_len=1, kv_seq_len up to 8192

Optimizations over reference:
  1. Tuned NUM_KV_SPLITS=64 (from 32) for better SM utilization at large KV lengths
  2. Removed unused imports to reduce module load time
"""
import subprocess, sys
for _pkg in ["aiter"]:
    try:
        __import__(_pkg)
    except ImportError:
        subprocess.check_call([sys.executable, "-m", "pip", "install", _pkg])

import torch
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

# ── MLA 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 = 64  # Tuned up from 32 for better SM utilization

FP8_DTYPE = aiter_dtypes.fp8


# ── FP8 Quantization ──────────────────────────────────────────────
def quantize_fp8(tensor: torch.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)


# ── Metadata Builder ──────────────────────────────────────────────
def _make_mla_decode_metadata(
    batch_size, max_q_len, nhead, nhead_kv,
    q_dtype, kv_dtype,
    qo_indptr, kv_indptr, kv_last_page_len,
    num_kv_splits=NUM_KV_SPLITS,
):
    """Allocate and populate persistent-mode work buffers for mla_decode_fwd."""
    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]
    (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,
        nhead // nhead_kv,
        nhead_kv,
        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=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):
    """
    MLA Decode attention — DeepSeek R1 forward_absorb path.

    Input tuple (5 elements):
        q:          (total_q, 16, 576)   bfloat16 — absorbed query
        kv_data:    dict with three KV cache formats:
                      "bf16":  (total_kv, 1, 576)         bfloat16
                      "fp8":   (Tensor, Tensor)           fp8 + scalar scale
                      "mxfp4": (Tensor, Tensor)           fp4x2 + e8m0 scale
        qo_indptr:  (batch_size + 1,) int32
        kv_indptr:  (batch_size + 1,) int32
        config:     dict with MLA parameters

    Output: (total_q, 16, 512) bfloat16
    """
    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"]

    # Quantize Q to FP8
    q_fp8, q_scale = quantize_fp8(q)

    # Use FP8 KV path (matches reference, proven correctness)
    kv_buffer_fp8, kv_scale = kv_data["fp8"]

    total_kv_len = int(kv_indptr[-1].item())
    kv_indices = torch.arange(total_kv_len, dtype=torch.int32, device="cuda")

    # Reshape KV buffer: (total_kv, page_size, nkv, dim)
    kv_buffer_4d = kv_buffer_fp8.view(kv_buffer_fp8.shape[0], PAGE_SIZE, nkv, kv_buffer_fp8.shape[-1])

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

    # Build persistent-mode metadata
    meta = _make_mla_decode_metadata(
        batch_size, max_q_len, nq, nkv,
        q_fp8.dtype, kv_buffer_fp8.dtype,
        qo_indptr, kv_indptr, kv_last_page_len,
        num_kv_splits=NUM_KV_SPLITS,
    )

    o = torch.empty((q.shape[0], nq, dv), dtype=torch.bfloat16, device="cuda")
    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,
        **meta,
    )
    return o