# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang

import torch

from fla.modules.l2norm import l2norm_bwd, l2norm_fwd
from fla.ops.comba.utils import chunk_comba_cumsum_scalar_bwd, chunk_comba_cumsum_scalar_fwd
from fla.ops.comba.wy_fast import chunk_scaled_dot_comba_pkt_fwd, prepare_wy_repr_bwd, recompute_w_u_fwd
from fla.ops.common.chunk_delta_h import chunk_gated_delta_rule_bwd_dhu, chunk_gated_delta_rule_fwd_h
from fla.ops.common.chunk_o import chunk_bwd_dqkwg, chunk_bwd_dv_local, chunk_fwd_o
from fla.ops.utils import chunk_local_cumsum, prepare_chunk_indices, solve_tril
from fla.ops.utils.constant import RCP_LN2
from fla.utils import autocast_custom_bwd, autocast_custom_fwd, input_guard


def chunk_comba_fwd(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    p: torch.Tensor,
    g: torch.Tensor,
    beta: torch.Tensor,
    scale: float,
    initial_state: torch.Tensor,
    output_final_state: bool,
    cu_seqlens: torch.LongTensor | None = None,
    chunk_indices: torch.LongTensor | None = None,
):
    g0, g = chunk_comba_cumsum_scalar_fwd(
        g,
        chunk_size=64,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        scale=RCP_LN2,
    )
    # obtain WY representation. u is actually the new v.
    A = chunk_scaled_dot_comba_pkt_fwd(
        k=k,
        p=p,
        beta=beta,
        g0=g0,
        g=g,
        cu_seqlens=cu_seqlens,
        output_dtype=torch.float32,
        chunk_indices=chunk_indices,
        use_exp2=True,
    )
    A = solve_tril(
        A=A,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        output_dtype=k.dtype,
    )
    w, u = recompute_w_u_fwd(
        k=p,
        v=v,
        beta=beta,
        A=A,
        g_cumsum=g0,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        use_exp2=True,
    )
    h, v_new, final_state = chunk_gated_delta_rule_fwd_h(
        k=k,
        w=w,
        u=u,
        g=g,
        initial_state=initial_state,
        output_final_state=output_final_state,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        use_exp2=True,
    )
    o = chunk_fwd_o(
        q=q,
        k=k,
        v=v_new,
        h=h,
        g=g,
        scale=scale,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        use_exp2=True,
    )
    return g0, g, o, A, final_state


def chunk_comba_bwd(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    p: torch.Tensor,
    g0: torch.Tensor,
    g: torch.Tensor,
    beta: torch.Tensor,
    A: torch.Tensor,
    scale: float,
    initial_state: torch.Tensor,
    do: torch.Tensor,
    dht: torch.Tensor,
    cu_seqlens: torch.LongTensor | None = None,
    chunk_indices: torch.LongTensor | None = None,
):
    w, u = recompute_w_u_fwd(
        k=p,
        v=v,
        beta=beta,
        A=A,
        g_cumsum=g0,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        use_exp2=True,
    )
    h, v_new, _ = chunk_gated_delta_rule_fwd_h(
        k=k,
        w=w,
        u=u,
        g=g,
        initial_state=initial_state,
        output_final_state=False,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        use_exp2=True,
    )
    dv = chunk_bwd_dv_local(
        q=q,
        k=k,
        g=g,
        do=do,
        scale=scale,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        use_exp2=True,
    )
    dh, dh0, dv = chunk_gated_delta_rule_bwd_dhu(
        q=q,
        k=k,
        w=w,
        g=g,
        h0=initial_state,
        dht=dht,
        do=do,
        dv=dv,
        scale=scale,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        use_exp2=True,
    )
    dq, dk, dw, dg = chunk_bwd_dqkwg(
        q=q,
        k=k,
        v=v_new,
        w=w,
        g=g,
        h=h,
        dv=dv,
        do=do,
        dh=dh,
        scale=scale,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        use_exp2=True,
    )
    dk2, dv, dp, db, dg0, dg2 = prepare_wy_repr_bwd(
        k=k,
        v=v,
        p=p,
        beta=beta,
        g0=g0,
        g=g,
        A=A,
        dw=dw,
        du=dv,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        use_exp2=True,
    )
    dk.add_(dk2)
    dg.add_(dg2)
    assert dg.dtype == torch.float32, "dg should be fp32"
    dg = chunk_local_cumsum(dg, chunk_size=64, reverse=True, cu_seqlens=cu_seqlens, chunk_indices=chunk_indices)
    # dg0 = d(g_cumsum - g)
    dg += chunk_comba_cumsum_scalar_bwd(dg0, chunk_size=64, cu_seqlens=cu_seqlens, chunk_indices=chunk_indices)
    return dq, dk, dv, dp, db, dg, dh0


class ChunkCombaFunction(torch.autograd.Function):

    @staticmethod
    @input_guard
    @autocast_custom_fwd
    def forward(
        ctx,
        q: torch.Tensor,
        k: torch.Tensor,
        v: torch.Tensor,
        p: torch.Tensor,
        g: torch.Tensor,
        beta: torch.Tensor,
        scale: float,
        initial_state: torch.Tensor,
        output_final_state: bool,
        use_qk_l2norm_in_kernel: bool = False,
        cu_seqlens: torch.LongTensor | None = None,
        cu_seqlens_cpu: torch.LongTensor | None = None,
    ):
        if use_qk_l2norm_in_kernel:
            q, q_rstd = l2norm_fwd(q)
            k, k_rstd = l2norm_fwd(k)
            p, p_rstd = l2norm_fwd(p)
        else:
            q_rstd, k_rstd, p_rstd = None, None, None

        chunk_indices = prepare_chunk_indices(
            cu_seqlens, 64, cu_seqlens_cpu=cu_seqlens_cpu) if cu_seqlens is not None else None

        g0, g, o, A, final_state = chunk_comba_fwd(
            q=q,
            k=k,
            v=v,
            p=p,
            g=g,
            beta=beta,
            scale=scale,
            initial_state=initial_state,
            output_final_state=output_final_state,
            cu_seqlens=cu_seqlens,
            chunk_indices=chunk_indices,
        )
        ctx.save_for_backward(q, q_rstd, k, k_rstd, p, p_rstd, v, g0, g, beta, A, initial_state, cu_seqlens,
                              chunk_indices)
        ctx.scale = scale
        ctx.use_qk_l2norm_in_kernel = use_qk_l2norm_in_kernel
        return o.to(q.dtype), final_state

    @staticmethod
    @input_guard
    @autocast_custom_bwd
    def backward(
        ctx,
        do: torch.Tensor,
        dht: torch.Tensor,
    ):
        q, q_rstd, k, k_rstd, p, p_rstd, v, g0, g, beta, A, initial_state, cu_seqlens, chunk_indices = (
            ctx.saved_tensors
        )
        dq, dk, dv, dp, db, dg, dh0 = chunk_comba_bwd(
            q=q,
            k=k,
            v=v,
            p=p,
            g0=g0,
            g=g,
            beta=beta,
            A=A,
            scale=ctx.scale,
            initial_state=initial_state,
            do=do,
            dht=dht,
            cu_seqlens=cu_seqlens,
            chunk_indices=chunk_indices,
        )
        if ctx.use_qk_l2norm_in_kernel:
            dq = l2norm_bwd(q, q_rstd, dq)
            dk = l2norm_bwd(k, k_rstd, dk)
            dp = l2norm_bwd(p, p_rstd, dp)
        return dq.to(q), dk.to(k), dv.to(v), dp.to(p), dg.to(g), db.to(beta), None, dh0, None, None, None, None


@torch.compiler.disable
def chunk_comba(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    p: torch.Tensor,
    g: torch.Tensor,
    beta: torch.Tensor = None,
    scale: float = None,
    initial_state: torch.Tensor = None,
    output_final_state: bool = False,
    use_qk_l2norm_in_kernel: bool = False,
    cu_seqlens: torch.LongTensor | None = None,
    cu_seqlens_cpu: torch.LongTensor | None = None,
):
    r"""
    Args:
        q (torch.Tensor):
            queries of shape `[B, T, H, K]`.
        k (torch.Tensor):
            keys of shape `[B, T, H, K]`.
        v (torch.Tensor):
            values of shape `[B, T, H, V]`.
        p (torch.Tensor):
            auxiliary keys of shape `[B, T, H, K]`.
        g (torch.Tensor):
            (forget) gating tensor (in log space!) of shape `[B, T, H]`.
        beta (torch.Tensor):
            betas of shape `[B, T, H]`.
        scale (Optional[int]):
            Scale factor for the RetNet attention scores.
            If not provided, it will default to `1 / sqrt(K)`. Default: `None`.
        initial_state (Optional[torch.Tensor]):
            Initial state of shape `[N, H, K, V]` for `N` input sequences.
            For equal-length input sequences, `N` equals the batch size `B`.
            Default: `None`.
        output_final_state (Optional[bool]):
            Whether to output the final state of shape `[N, H, K, V]`. Default: `False`.
        use_qk_l2norm_in_kernel (bool):
            Whether to apply L2norm to the q/k tensor internally. Default: `False`.
        cu_seqlens (torch.LongTensor):
            Cumulative sequence lengths of shape `[N+1]` used for variable-length training,
            consistent with the FlashAttention API.

    Returns:
        o (torch.Tensor):
            Outputs of shape `[B, T, H, V]`.
        final_state (torch.Tensor):
            Final state of shape `[N, H, K, V]` if `output_final_state=True` else `None`.

    Examples::
        >>> import torch
        >>> import torch.nn.functional as F
        >>> from einops import rearrange
        >>> from fla.ops.comba import chunk_comba
        # inputs with equal lengths
        >>> B, T, H, K, V = 4, 2048, 4, 512, 512
        >>> q = torch.randn(B, T, H, K, dtype=torch.bfloat16, device='cuda')
        >>> k = F.normalize(torch.randn(B, T, H, K, dtype=torch.bfloat16, device='cuda'), p=2, dim=-1)
        >>> v = torch.randn(B, T, H, V, dtype=torch.bfloat16, device='cuda')
        >>> b = torch.rand(H, dtype=torch.bfloat16, device='cuda').sigmoid()
        >>> p = k * b[:, None]
        >>> beta = torch.rand(B, T, H, dtype=torch.bfloat16, device='cuda').sigmoid()
        >>> g = F.logsigmoid(torch.rand(B, T, H, dtype=torch.bfloat16, device='cuda'))
        >>> h0 = torch.randn(B, H, K, V, dtype=torch.bfloat16, device='cuda')
        >>> o, ht = chunk_comba(
            q, k, v, p, g, beta,
            initial_state=h0,
            output_final_state=True
        )
        # for variable-length inputs, the batch size `B` is expected to be 1 and `cu_seqlens` is required
        >>> q, k, v, beta, g = map(lambda x: rearrange(x, 'b t ... -> 1 (b t) ...'), (q, k, v, beta, g))
        # for a batch with 4 sequences, `cu_seqlens` with 5 start/end positions are expected
        >>> cu_seqlens = q.new_tensor([0, 2048, 4096, 6144, 8192], dtype=torch.long)
        >>> o_var, ht_var = chunk_comba(
            q, k, v, p, g, beta,
            initial_state=h0,
            output_final_state=True,
            cu_seqlens=cu_seqlens
        )
    """
    if p is None:
        p = k
    if cu_seqlens is not None:
        if q.shape[0] != 1:
            raise ValueError(
                f"The batch size is expected to be 1 rather than {q.shape[0]} when using `cu_seqlens`."
                f"Please flatten variable-length inputs before processing.",
            )
        if initial_state is not None and initial_state.shape[0] != len(cu_seqlens) - 1:
            raise ValueError(
                f"The number of initial states is expected to be equal to the number of input sequences, "
                f"i.e., {len(cu_seqlens) - 1} rather than {initial_state.shape[0]}.",
            )
    if scale is None:
        scale = k.shape[-1] ** -0.5
    o, final_state = ChunkCombaFunction.apply(
        q,
        k,
        v,
        p,
        g,
        beta,
        scale,
        initial_state,
        output_final_state,
        use_qk_l2norm_in_kernel,
        cu_seqlens,
        cu_seqlens_cpu,
    )
    return o, final_state