xc-llm-kunlun/vllm_kunlun/ops/fla/chunk_o.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
#
# This file contains code copied from the flash-linear-attention project.
# The original source code was licensed under the MIT license and included
# the following copyright notice:
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang

# ruff: noqa: E501

from typing import Optional

import torch
from vllm.triton_utils import tl, triton

from .index import prepare_chunk_indices
from .utils import FLA_GDN_FIX_BT, check_shared_mem, is_nvidia_hopper

BKV_LIST = [64, 128] if check_shared_mem() else [32, 64]
NUM_WARPS = [2, 4] if is_nvidia_hopper else [2, 4, 8]


@triton.heuristics(
    {
        "USE_G": lambda args: args["g"] is not None,
        "IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
    }
)
# @triton.autotune(
#     configs=[
#         triton.Config({
#             'BK': BK,
#             'BV': BV
#         },
#                       num_warps=num_warps,
#                       num_stages=num_stages) for BK in BKV_LIST
#         for BV in BKV_LIST for num_warps in NUM_WARPS
#         for num_stages in [2, 3, 4]
#     ],
#     key=['H', 'K', 'V', 'BT'],
# )
@triton.jit(do_not_specialize=["T"])
def chunk_fwd_kernel_o(
    q,
    k,
    v,
    h,
    g,
    o,
    cu_seqlens,
    chunk_indices,
    scale,
    T,
    H: tl.constexpr,
    Hg: tl.constexpr,
    K: tl.constexpr,
    V: tl.constexpr,
    BT: tl.constexpr,
    BK: tl.constexpr,
    BV: tl.constexpr,
    USE_G: tl.constexpr,
    IS_VARLEN: tl.constexpr,
):
    i_v, i_t, i_bh = tl.program_id(0), tl.program_id(1), tl.program_id(2)
    i_b, i_h = i_bh // H, i_bh % H

    if IS_VARLEN:
        i_tg = i_t
        i_n, i_t = tl.load(chunk_indices + i_t * 2).to(tl.int32), tl.load(
            chunk_indices + i_t * 2 + 1
        ).to(tl.int32)
        bos, eos = tl.load(cu_seqlens + i_n).to(tl.int32), tl.load(
            cu_seqlens + i_n + 1
        ).to(tl.int32)
        T = eos - bos
        NT = tl.cdiv(T, BT)
    else:
        NT = tl.cdiv(T, BT)
        i_tg = i_b * NT + i_t
        bos, eos = i_b * T, i_b * T + T

    # offset calculation
    q += (bos * Hg + i_h // (H // Hg)) * K
    k += (bos * Hg + i_h // (H // Hg)) * K
    v += (bos * H + i_h) * V
    o += (bos * H + i_h) * V
    h += (i_tg * H + i_h).to(tl.int64) * K * V

    b_o = tl.zeros([BT, BV], dtype=tl.float32)
    b_A = tl.zeros([BT, BT], dtype=tl.float32)

    for i_k in range(tl.cdiv(K, BK)):
        p_q = tl.make_block_ptr(
            q, (T, K), (Hg * K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0)
        )
        p_k = tl.make_block_ptr(
            k, (K, T), (1, Hg * K), (i_k * BK, i_t * BT), (BK, BT), (0, 1)
        )
        p_h = tl.make_block_ptr(
            h, (K, V), (V, 1), (i_k * BK, i_v * BV), (BK, BV), (1, 0)
        )
        # [BT, BK]
        b_q = tl.load(p_q, boundary_check=(0, 1))
        # [BK, BT]
        b_k = tl.load(p_k, boundary_check=(0, 1))
        # [BK, BV]
        b_h = tl.load(p_h, boundary_check=(0, 1))

        # [BT, BK] @ [BK, BV] -> [BT, BV]
        b_o += tl.dot(b_q, b_h)
        # [BT, BK] @ [BK, BT] -> [BT, BT]
        b_A += tl.dot(b_q, b_k)

    if USE_G:
        g += bos * H + i_h
        p_g = tl.make_block_ptr(g, (T,), (H,), (i_t * BT,), (BT,), (0,))
        b_g = tl.load(p_g, boundary_check=(0,))
        b_o = b_o * tl.exp(b_g)[:, None]
        b_A = b_A * tl.exp(b_g[:, None] - b_g[None, :])

    o_t = i_t * BT + tl.arange(0, BT)
    # m_t = o_t < T
    # m_A = (o_t[:, None] >= o_t[None, :]) & (m_t[:, None] & m_t)
    # b_A = tl.where(m_A, b_A, 0)
    b_A = tl.where(o_t[:, None] >= o_t[None, :], b_A, 0)

    p_v = tl.make_block_ptr(
        v, (T, V), (H * V, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
    )
    p_o = tl.make_block_ptr(
        o, (T, V), (H * V, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
    )
    b_v = tl.load(p_v, boundary_check=(0, 1))

    # to fix mma -> mma layout conversion
    # already solved by triton v3.2 or higher
    b_o = b_o * scale + tl.dot(b_A.to(b_v.dtype), b_v) * scale
    tl.store(p_o, b_o.to(p_o.dtype.element_ty), boundary_check=(0, 1))


def chunk_fwd_o(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    h: torch.Tensor,
    g: Optional[torch.Tensor] = None,  # cumsum of log decay
    scale: Optional[float] = None,
    cu_seqlens: Optional[torch.LongTensor] = None,
    chunk_size: int = 64,
) -> torch.Tensor:
    _, T, _, _, _ = *q.shape, v.shape[-1]
    if FLA_GDN_FIX_BT:
        BT = 64
    else:
        BT = min(chunk_size, max(16, triton.next_power_of_2(T)))
    chunk_indices = (
        prepare_chunk_indices(cu_seqlens, BT) if cu_seqlens is not None else None
    )
    if scale is None:
        scale = k.shape[-1] ** -0.5

    o = torch.empty_like(v)

    o = torch.ops.xspeedgate_ops.chunk_fwd_o(
        q, k, v, h, g, scale, cu_seqlens, chunk_indices, chunk_size
    )
    return o