K3/kda/_fla/ops/utils/cumsum.py

# Copyright (c) 2023-2026, Songlin Yang, Yu Zhang, Zhiyuan Li
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
# For a list of all contributors, visit:
#   https://github.com/fla-org/flash-linear-attention/graphs/contributors

import torch
import triton
import triton.language as tl

from kda._fla.ops.backends import dispatch
from kda._fla.ops.utils.cache import fla_cache_autotune
from kda._fla.ops.utils.index import prepare_chunk_indices
from kda._fla.utils import autotune_cache_kwargs, check_shared_mem, input_guard

BS_LIST = [32, 64] if check_shared_mem() else [16, 32]


@triton.heuristics({
    'HAS_SCALE': lambda args: args['scale'] is not None,
    'IS_VARLEN': lambda args: args['cu_seqlens'] is not None,
})
@fla_cache_autotune(
    configs=[
        triton.Config({}, num_warps=num_warps)
        for num_warps in [1, 2, 4, 8]
    ],
    key=['B', 'H', 'BT', 'IS_VARLEN', 'REVERSE'],
    **autotune_cache_kwargs,
)
@triton.jit(do_not_specialize=['T'])
def chunk_local_cumsum_scalar_kernel(
    s,
    o,
    scale,
    cu_seqlens,
    chunk_indices,
    T,
    B: tl.constexpr,
    H: tl.constexpr,
    BT: tl.constexpr,
    REVERSE: tl.constexpr,
    HAS_SCALE: tl.constexpr,
    IS_VARLEN: tl.constexpr,
):
    i_t, i_bh = tl.program_id(0).to(tl.int64), tl.program_id(1).to(tl.int64)
    i_b, i_h = i_bh // H, i_bh % H
    if IS_VARLEN:
        i_n, i_t = tl.load(chunk_indices + i_t * 2).to(tl.int32), tl.load(chunk_indices + i_t * 2 + 1).to(tl.int64)
        bos, eos = tl.load(cu_seqlens + i_n).to(tl.int64), tl.load(cu_seqlens + i_n + 1).to(tl.int64)
        T = eos - bos
    else:
        bos, eos = i_b * T, i_b * T + T

    o_t = i_t * BT + tl.arange(0, BT)
    m_t = o_t < T
    p_s = s + bos*H + i_h + o_t * H
    p_o = o + bos*H + i_h + o_t * H
    # [BT]
    b_s = tl.load(p_s, mask=m_t, other=0.0).to(tl.float32)
    if REVERSE:
        b_o = tl.cumsum(b_s, axis=0, reverse=True)
    else:
        b_o = tl.cumsum(b_s, axis=0)
    if HAS_SCALE:
        b_o *= scale
    tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=m_t)


@triton.heuristics({
    'HAS_SCALE': lambda args: args['scale'] is not None,
    'IS_VARLEN': lambda args: args['cu_seqlens'] is not None,
})
@fla_cache_autotune(
    configs=[
        triton.Config({'BS': BS}, num_warps=num_warps)
        for BS in BS_LIST
        for num_warps in [2, 4, 8]
    ],
    key=['B', 'H', 'S', 'BT', 'IS_VARLEN', 'REVERSE'],
    **autotune_cache_kwargs,
)
@triton.jit(do_not_specialize=['T'])
def chunk_local_cumsum_vector_kernel(
    s,
    o,
    scale,
    cu_seqlens,
    chunk_indices,
    T,
    B: tl.constexpr,
    H: tl.constexpr,
    S: tl.constexpr,
    BT: tl.constexpr,
    BS: tl.constexpr,
    REVERSE: tl.constexpr,
    HAS_SCALE: tl.constexpr,
    IS_VARLEN: tl.constexpr,
):
    i_s, i_t, i_bh = tl.program_id(0), tl.program_id(1).to(tl.int64), tl.program_id(2).to(tl.int64)
    i_b, i_h = i_bh // H, i_bh % H
    if IS_VARLEN:
        i_n, i_t = tl.load(chunk_indices + i_t * 2).to(tl.int32), tl.load(chunk_indices + i_t * 2 + 1).to(tl.int64)
        bos, eos = tl.load(cu_seqlens + i_n).to(tl.int64), tl.load(cu_seqlens + i_n + 1).to(tl.int64)
        T = eos - bos
    else:
        bos, eos = i_b * T, i_b * T + T

    o_t = i_t * BT + tl.arange(0, BT)
    o_s = i_s * BS + tl.arange(0, BS)
    m_s = (o_t[:, None] < T) & (o_s[None, :] < S)
    p_s = s + (bos * H + i_h) * S + o_t[:, None] * (H*S) + o_s[None, :]
    p_o = o + (bos * H + i_h) * S + o_t[:, None] * (H*S) + o_s[None, :]
    # [BT, BS]
    b_s = tl.load(p_s, mask=m_s, other=0.0).to(tl.float32)
    if REVERSE:
        b_o = tl.cumsum(b_s, axis=0, reverse=True)
    else:
        b_o = tl.cumsum(b_s, axis=0)
    if HAS_SCALE:
        b_o *= scale
    tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=m_s)


@triton.heuristics({
    'HAS_SCALE': lambda args: args['scale'] is not None,
    'IS_VARLEN': lambda args: args['cu_seqlens'] is not None,
})
@triton.autotune(
    configs=[
        triton.Config({'BT': BT}, num_warps=num_warps, num_stages=num_stages)
        for BT in [32, 64, 128, 256]
        for num_warps in [2, 4, 8]
        for num_stages in [1, 2, 3, 4]
    ],
    key=['B', 'H', 'IS_VARLEN', 'REVERSE'],
    **autotune_cache_kwargs,
)
@triton.jit(do_not_specialize=['T'])
def chunk_global_cumsum_scalar_kernel(
    s,
    o,
    scale,
    cu_seqlens,
    T,
    B: tl.constexpr,
    H: tl.constexpr,
    BT: tl.constexpr,
    REVERSE: tl.constexpr,
    HAS_SCALE: tl.constexpr,
    IS_VARLEN: tl.constexpr,
):
    i_nh = tl.program_id(0).to(tl.int64)
    i_n, i_h = i_nh // H, i_nh % H
    if IS_VARLEN:
        bos, eos = tl.load(cu_seqlens + i_n).to(tl.int64), tl.load(cu_seqlens + i_n + 1).to(tl.int64)
    else:
        bos, eos = i_n * T, i_n * T + T
    T = eos - bos

    b_z = tl.zeros([], dtype=tl.float32)
    NT = tl.cdiv(T, BT)
    for i_c in range(NT):
        i_t = NT - 1 - i_c if REVERSE else i_c
        o_t = i_t * BT + tl.arange(0, BT)
        m_t = o_t < T
        p_s = s + bos*H + i_h + o_t * H
        p_o = o + bos*H + i_h + o_t * H
        b_s = tl.load(p_s, mask=m_t, other=0.0).to(tl.float32)
        if REVERSE:
            b_o = tl.cumsum(b_s, axis=0, reverse=True)
        else:
            b_o = tl.cumsum(b_s, axis=0)
        b_ss = tl.sum(b_s, 0)
        b_o += b_z
        if i_c >= 0:
            b_z += b_ss
        if HAS_SCALE:
            b_o *= scale
        tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=m_t)


@triton.heuristics({
    'HAS_SCALE': lambda args: args['scale'] is not None,
    'IS_VARLEN': lambda args: args['cu_seqlens'] is not None,
})
@triton.autotune(
    configs=[
        triton.Config({'BT': BT}, num_warps=num_warps, num_stages=num_stages)
        for BT in [16, 32, 64, 128]
        for num_warps in [2, 4, 8]
        for num_stages in [1, 2, 3, 4]
    ],
    key=['B', 'H', 'S', 'IS_VARLEN', 'REVERSE'],
    **autotune_cache_kwargs,
)
@triton.jit(do_not_specialize=['T'])
def chunk_global_cumsum_vector_kernel(
    s,
    o,
    scale,
    cu_seqlens,
    T,
    B: tl.constexpr,
    H: tl.constexpr,
    S: tl.constexpr,
    BT: tl.constexpr,
    BS: tl.constexpr,
    REVERSE: tl.constexpr,
    HAS_SCALE: tl.constexpr,
    IS_VARLEN: tl.constexpr,
):
    i_s, i_nh = tl.program_id(0), tl.program_id(1).to(tl.int64)
    i_n, i_h = i_nh // H, i_nh % H
    if IS_VARLEN:
        bos, eos = tl.load(cu_seqlens + i_n).to(tl.int64), tl.load(cu_seqlens + i_n + 1).to(tl.int64)
    else:
        bos, eos = i_n * T, i_n * T + T
    T = eos - bos

    b_z = tl.zeros([BS], dtype=tl.float32)
    NT = tl.cdiv(T, BT)
    for i_c in range(NT):
        i_t = NT - 1 - i_c if REVERSE else i_c
        o_t = i_t * BT + tl.arange(0, BT)
        o_s = i_s * BS + tl.arange(0, BS)
        m_s = (o_t[:, None] < T) & (o_s[None, :] < S)
        p_s = s + (bos * H + i_h) * S + o_t[:, None] * (H*S) + o_s[None, :]
        p_o = o + (bos * H + i_h) * S + o_t[:, None] * (H*S) + o_s[None, :]
        # [BT, BS]
        b_s = tl.load(p_s, mask=m_s, other=0.0).to(tl.float32)
        if REVERSE:
            b_c = b_z[None, :] + tl.cumsum(b_s, axis=0, reverse=True)
        else:
            b_c = b_z[None, :] + tl.cumsum(b_s, axis=0)
        if HAS_SCALE:
            b_c *= scale
        tl.store(p_o, b_c.to(p_o.dtype.element_ty), mask=m_s)
        b_z += tl.sum(b_s, 0)


def chunk_local_cumsum_scalar(
    g: torch.Tensor,
    chunk_size: int,
    reverse: bool = False,
    scale: float = None,
    cu_seqlens: torch.Tensor | None = None,
    output_dtype: torch.dtype | None = torch.float,
    chunk_indices: torch.LongTensor | None = None,
    **kwargs,
) -> torch.Tensor:
    if 'head_first' in kwargs:
        raise DeprecationWarning(
            "head_first has been removed. Inputs must be in `[B, T, H, ...]` format.",
        )
    B, T, H = g.shape
    assert chunk_size == 2**(chunk_size.bit_length()-1), "chunk_size must be a power of 2"
    BT = chunk_size
    if chunk_indices is None and cu_seqlens is not None:
        chunk_indices = prepare_chunk_indices(cu_seqlens, BT)
    NT = triton.cdiv(T, BT) if cu_seqlens is None else len(chunk_indices)
    g_org, g = g, torch.empty_like(g, dtype=output_dtype or g.dtype)
    grid = (NT, B * H)
    chunk_local_cumsum_scalar_kernel[grid](
        s=g_org,
        o=g,
        scale=scale,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        T=T,
        B=B,
        H=H,
        BT=BT,
        REVERSE=reverse,
    )
    return g


def chunk_local_cumsum_vector(
    g: torch.Tensor,
    chunk_size: int,
    reverse: bool = False,
    scale: float = None,
    cu_seqlens: torch.Tensor | None = None,
    output_dtype: torch.dtype | None = torch.float,
    chunk_indices: torch.LongTensor | None = None,
    **kwargs,
) -> torch.Tensor:
    if 'head_first' in kwargs:
        raise DeprecationWarning(
            "head_first has been removed. Inputs must be in `[B, T, H, ...]` format.",
        )
    B, T, H, S = g.shape
    BT = chunk_size
    if chunk_indices is None and cu_seqlens is not None:
        chunk_indices = prepare_chunk_indices(cu_seqlens, BT)
    NT = triton.cdiv(T, BT) if cu_seqlens is None else len(chunk_indices)
    assert chunk_size == 2**(chunk_size.bit_length()-1), "chunk_size must be a power of 2"

    g_org, g = g, torch.empty_like(g, dtype=output_dtype or g.dtype)
    def grid(meta): return (triton.cdiv(meta['S'], meta['BS']), NT, B * H)
    # keep cummulative normalizer in fp32
    # this kernel is equivalent to
    # g = g.view(B, H, NT, BT, -1).cumsum(-2).view(B, H, T, -1)
    chunk_local_cumsum_vector_kernel[grid](
        s=g_org,
        o=g,
        scale=scale,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        T=T,
        B=B,
        H=H,
        S=S,
        BT=BT,
        REVERSE=reverse,
    )
    return g


@input_guard
def chunk_global_cumsum_scalar(
    s: torch.Tensor,
    reverse: bool = False,
    cu_seqlens: torch.Tensor | None = None,
    scale: float = None,
    output_dtype: torch.dtype | None = torch.float,
    **kwargs,
) -> torch.Tensor:
    if 'head_first' in kwargs:
        raise DeprecationWarning(
            "head_first has been removed. Inputs must be in `[B, T, H, ...]` format.",
        )
    B, T, H = s.shape
    N = len(cu_seqlens) - 1 if cu_seqlens is not None else B

    z = torch.empty_like(s, dtype=output_dtype or s.dtype)
    grid = (N * H,)
    chunk_global_cumsum_scalar_kernel[grid](
        s=s,
        o=z,
        scale=scale,
        cu_seqlens=cu_seqlens,
        T=T,
        B=B,
        H=H,
        REVERSE=reverse,
    )
    return z


@input_guard
def chunk_global_cumsum_vector(
    s: torch.Tensor,
    reverse: bool = False,
    cu_seqlens: torch.Tensor | None = None,
    scale: float = None,
    output_dtype: torch.dtype | None = torch.float,
    **kwargs,
) -> torch.Tensor:
    if 'head_first' in kwargs:
        raise DeprecationWarning(
            "head_first has been removed. Inputs must be in `[B, T, H, ...]` format.",
        )
    B, T, H, S = s.shape
    N = len(cu_seqlens) - 1 if cu_seqlens is not None else B
    BS = min(32, triton.next_power_of_2(S))

    z = torch.empty_like(s, dtype=output_dtype or s.dtype)
    grid = (triton.cdiv(S, BS), N * H)
    chunk_global_cumsum_vector_kernel[grid](
        s=s,
        o=z,
        scale=scale,
        cu_seqlens=cu_seqlens,
        T=T,
        B=B,
        H=H,
        S=S,
        BS=BS,
        REVERSE=reverse,
    )
    return z


@input_guard
@dispatch('utils')
def chunk_global_cumsum(
    s: torch.Tensor,
    reverse: bool = False,
    cu_seqlens: torch.Tensor | None = None,
    scale: float = None,
    output_dtype: torch.dtype | None = torch.float,
    **kwargs,
) -> torch.Tensor:
    if 'head_first' in kwargs:
        raise DeprecationWarning(
            "head_first has been removed. Inputs must be in `[B, T, H, ...]` format.",
        )
    if cu_seqlens is not None:
        assert s.shape[0] == 1, "Only batch size 1 is supported when cu_seqlens are provided"
    if len(s.shape) == 3:
        return chunk_global_cumsum_scalar(
            s=s,
            reverse=reverse,
            cu_seqlens=cu_seqlens,
            scale=scale,
            output_dtype=output_dtype,
        )
    elif len(s.shape) == 4:
        return chunk_global_cumsum_vector(
            s=s,
            reverse=reverse,
            cu_seqlens=cu_seqlens,
            scale=scale,
            output_dtype=output_dtype,
        )
    else:
        raise ValueError(
            f"Unsupported input shape {s.shape}, "
            f"which should be [B, T, H] or [B, T, H, D]",
        )


@input_guard
@dispatch('utils')
def chunk_local_cumsum(
    g: torch.Tensor,
    chunk_size: int,
    reverse: bool = False,
    scale: float = None,
    cu_seqlens: torch.Tensor | None = None,
    output_dtype: torch.dtype | None = torch.float,
    chunk_indices: torch.LongTensor | None = None,
    **kwargs,
) -> torch.Tensor:
    if 'head_first' in kwargs:
        raise DeprecationWarning(
            "head_first has been removed. Inputs must be in `[B, T, H, ...]` format.",
        )
    if cu_seqlens is not None:
        assert g.shape[0] == 1, "Only batch size 1 is supported when cu_seqlens are provided"
    if len(g.shape) == 3:
        return chunk_local_cumsum_scalar(
            g=g,
            chunk_size=chunk_size,
            reverse=reverse,
            scale=scale,
            cu_seqlens=cu_seqlens,
            output_dtype=output_dtype,
            chunk_indices=chunk_indices,
        )
    elif len(g.shape) == 4:
        return chunk_local_cumsum_vector(
            g=g,
            chunk_size=chunk_size,
            reverse=reverse,
            scale=scale,
            cu_seqlens=cu_seqlens,
            output_dtype=output_dtype,
            chunk_indices=chunk_indices,
        )
    else:
        raise ValueError(
            f"Unsupported input shape {g.shape}, "
            f"which should be (B, T, H) or (B, T, H, D)",
        )