# Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/gated_delta_rule/fused_recurrent.py
# -*- coding: utf-8 -*-
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang

from typing import Optional, Tuple

import torch
import triton
import triton.language as tl

from sglang.srt.layers.attention.fla.op import exp
from sglang.srt.layers.attention.fla.utils import input_guard


@triton.jit(do_not_specialize=["T"])
def fused_recurrent_gated_delta_rule_fwd_kernel(
    q,
    k,
    v,
    g,
    beta,
    o,
    h0,
    ht,
    cu_seqlens,
    scale,
    T,
    B: tl.constexpr,
    H: tl.constexpr,
    HV: tl.constexpr,
    K: tl.constexpr,
    V: tl.constexpr,
    BK: tl.constexpr,
    BV: tl.constexpr,
    USE_INITIAL_STATE: tl.constexpr,  # whether to use initial state
    STORE_FINAL_STATE: tl.constexpr,  # whether to store final state
    IS_BETA_HEADWISE: tl.constexpr,  # whether beta is headwise vector or scalar,
    USE_QK_L2NORM_IN_KERNEL: tl.constexpr,
    IS_VARLEN: tl.constexpr,
    IS_KDA: tl.constexpr,
):
    i_k, i_v, i_nh = tl.program_id(0), tl.program_id(1), tl.program_id(2)
    i_n, i_hv = i_nh // HV, i_nh % HV
    i_h = i_hv // (HV // 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)
        all = T
        T = eos - bos
    else:
        bos, eos = i_n * T, i_n * T + T
        all = B * T
    o_k = i_k * BK + tl.arange(0, BK)
    o_v = i_v * BV + tl.arange(0, BV)

    p_q = q + (bos * H + i_h) * K + o_k
    p_k = k + (bos * H + i_h) * K + o_k
    p_v = v + (bos * HV + i_hv) * V + o_v
    if IS_BETA_HEADWISE:
        p_beta = beta + (bos * HV + i_hv) * V + o_v
    else:
        p_beta = beta + bos * HV + i_hv
    if not IS_KDA:
        p_g = g + bos * HV + i_hv
    else:
        p_gk = g + (bos * HV + i_hv) * K + o_k

    p_o = o + ((i_k * all + bos) * HV + i_hv) * V + o_v

    mask_k = o_k < K
    mask_v = o_v < V
    mask_h = mask_k[:, None] & mask_v[None, :]

    b_h = tl.zeros([BK, BV], dtype=tl.float32)
    if USE_INITIAL_STATE:
        p_h0 = h0 + i_nh * K * V + o_k[:, None] * V + o_v[None, :]
        b_h += tl.load(p_h0, mask=mask_h, other=0).to(tl.float32)

    for _ in range(0, T):
        b_q = tl.load(p_q, mask=mask_k, other=0).to(tl.float32)
        b_k = tl.load(p_k, mask=mask_k, other=0).to(tl.float32)
        b_v = tl.load(p_v, mask=mask_v, other=0).to(tl.float32)

        if USE_QK_L2NORM_IN_KERNEL:
            b_q = b_q / (tl.sqrt(tl.sum(b_q * b_q) + 1e-6))
            b_k = b_k / (tl.sqrt(tl.sum(b_k * b_k) + 1e-6))
        b_q = b_q * scale
        # [BK, BV]
        if not IS_KDA:
            b_g = tl.load(p_g).to(tl.float32)
            b_h *= exp(b_g)
        else:
            b_gk = tl.load(p_gk).to(tl.float32)
            b_h *= exp(b_gk[:, None])
        # [BV]
        b_v -= tl.sum(b_h * b_k[:, None], 0)
        if IS_BETA_HEADWISE:
            b_beta = tl.load(p_beta, mask=mask_v, other=0).to(tl.float32)
        else:
            b_beta = tl.load(p_beta).to(tl.float32)
        b_v *= b_beta
        # [BK, BV]
        b_h += b_k[:, None] * b_v[None, :]
        # [BV]
        b_o = tl.sum(b_h * b_q[:, None], 0)
        tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=mask_v)

        p_q += H * K
        p_k += H * K
        p_o += HV * V
        p_v += HV * V
        if not IS_KDA:
            p_g += HV
        else:
            p_gk += HV * K
        p_beta += HV * (V if IS_BETA_HEADWISE else 1)

    if STORE_FINAL_STATE:
        p_ht = ht + i_nh * K * V + o_k[:, None] * V + o_v[None, :]
        tl.store(p_ht, b_h.to(p_ht.dtype.element_ty), mask=mask_h)


def fused_recurrent_gated_delta_rule_fwd(
    q: torch.Tensor,
    k: torch.Tensor,
    v: 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: Optional[torch.LongTensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
    B, T, H, K, V = *k.shape, v.shape[-1]
    HV = v.shape[2]
    N = B if cu_seqlens is None else len(cu_seqlens) - 1
    BK, BV = triton.next_power_of_2(K), min(triton.next_power_of_2(V), 32)
    NK, NV = triton.cdiv(K, BK), triton.cdiv(V, BV)
    assert NK == 1, "NK > 1 is not supported yet"
    num_stages = 3
    num_warps = 1

    o = q.new_empty(NK, *v.shape)
    if output_final_state:
        final_state = q.new_empty(N, HV, K, V, dtype=torch.float32)
    else:
        final_state = None

    grid = (NK, NV, N * HV)
    fused_recurrent_gated_delta_rule_fwd_kernel[grid](
        q=q,
        k=k,
        v=v,
        g=g,
        beta=beta,
        o=o,
        h0=initial_state,
        ht=final_state,
        cu_seqlens=cu_seqlens,
        scale=scale,
        T=T,
        B=B,
        H=H,
        HV=HV,
        K=K,
        V=V,
        BK=BK,
        BV=BV,
        USE_INITIAL_STATE=initial_state is not None,
        STORE_FINAL_STATE=final_state is not None,
        IS_BETA_HEADWISE=beta.ndim == v.ndim,
        USE_QK_L2NORM_IN_KERNEL=use_qk_l2norm_in_kernel,
        IS_VARLEN=cu_seqlens is not None,
        IS_KDA=False,
        num_warps=num_warps,
        num_stages=num_stages,
    )
    o = o.squeeze(0)
    return o, final_state


class FusedRecurrentFunction(torch.autograd.Function):

    @staticmethod
    @input_guard
    def forward(
        ctx,
        q: torch.Tensor,
        k: torch.Tensor,
        v: torch.Tensor,
        g: torch.Tensor,
        beta: torch.Tensor,
        scale: float,
        initial_state: torch.Tensor,
        output_final_state: bool,
        cu_seqlens: Optional[torch.LongTensor] = None,
        use_qk_l2norm_in_kernel: bool = False,
    ):
        o, final_state = fused_recurrent_gated_delta_rule_fwd(
            q=q,
            k=k,
            v=v,
            g=g,
            beta=beta,
            scale=scale,
            initial_state=initial_state,
            output_final_state=output_final_state,
            use_qk_l2norm_in_kernel=use_qk_l2norm_in_kernel,
            cu_seqlens=cu_seqlens,
        )

        return o, final_state

    @staticmethod
    @input_guard
    def backward(ctx, do, dht):
        raise NotImplementedError(
            "Backward pass is not implemented yet and we do not have plans to implement it "
            "because we haven't figured out how to compute dg without materializing the full "
            "hidden states for all time steps."
        )


def fused_recurrent_gated_delta_rule(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    g: torch.Tensor,
    beta: torch.Tensor = None,
    scale: float = None,
    initial_state: torch.Tensor = None,
    output_final_state: bool = False,
    cu_seqlens: Optional[torch.LongTensor] = None,
    use_qk_l2norm_in_kernel: bool = False,
) -> Tuple[torch.Tensor, torch.Tensor]:
    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, HV, V]`.
            GVA is applied if `HV > H`.
        g (torch.Tensor):
            g (decays) of shape `[B, T, HV]`.
        beta (torch.Tensor):
            betas of shape `[B, T, HV]`.
        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, HV, 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, HV, K, V]`. 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, HV, V]`.
        final_state (torch.Tensor):
            Final state of shape `[N, HV, 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.gated_delta_rule import fused_recurrent_gated_delta_rule
        # inputs with equal lengths
        >>> B, T, H, HV, K, V = 4, 2048, 4, 8, 512, 512
        >>> q = torch.randn(B, T, H, K, device='cuda')
        >>> k = F.normalize(torch.randn(B, T, H, K, device='cuda'), p=2, dim=-1)
        >>> v = torch.randn(B, T, HV, V, device='cuda')
        >>> g = F.logsigmoid(torch.rand(B, T, HV, device='cuda'))
        >>> beta = torch.rand(B, T, HV, device='cuda').sigmoid()
        >>> h0 = torch.randn(B, HV, K, V, device='cuda')
        >>> o, ht = fused_gated_recurrent_delta_rule(
            q, k, v, 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, g, beta = map(lambda x: rearrange(x, 'b t ... -> 1 (b t) ...'), (q, k, v, g, beta))
        # 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 = fused_gated_recurrent_delta_rule(
            q, k, v, g, beta,
            initial_state=h0,
            output_final_state=True,
            cu_seqlens=cu_seqlens
        )
    """
    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
    else:
        assert scale > 0, "scale must be positive"
    if beta is None:
        beta = torch.ones_like(q[..., 0])
    o, final_state = FusedRecurrentFunction.apply(
        q,
        k,
        v,
        g,
        beta,
        scale,
        initial_state,
        output_final_state,
        cu_seqlens,
        use_qk_l2norm_in_kernel,
    )
    return o, final_state


# HAS_EAGLE_TREE_CUSTOM_ATTN_MASK is added to support eagle tree attention mask
# retrieve_parent_token_ptr: [N, NP2_T], retrieve_next_sibling_ptr: [N, NP2_T]
# e.g. for a sequence of length 4, the eagle tree attention structure is:
# retrieve_next_token=[1, 3, -1, -1] -> retrieve_next_token[i]: the 1st child token of token i
# retrieve_next_sibling=[-1, 2, -1, -1] -> retrieve_next_sibling[i]: the 1st tree sibling token of token i
# retrieve_parent_token=[n/a, 0, 0, 1] -> retrieve_parent_token[i]: the parent token of token i
# Tree:
#    0
#   / \
#  1   2
# /
# 3
# When calculating token 3's attention, it should attend to token 1 (parent) and token 0 (grand-parent)
# When calculating token 2's attention, it should attend to token 0 (parent)
@triton.jit(do_not_specialize=["T"])
def fused_recurrent_gated_delta_rule_update_fwd_kernel(
    q,
    k,
    v,
    g,
    beta,
    o,
    h0_source,
    h0_indices,
    cu_seqlens,
    scale,
    intermediate_states_buffer,
    intermediate_state_indices,
    cache_steps,
    retrieve_parent_token_ptr,
    stride_retrieve_parent_token_seq: tl.constexpr,
    stride_retrieve_parent_token_token: tl.constexpr,
    T,
    NP2_T: tl.constexpr,
    B: tl.constexpr,
    H: tl.constexpr,
    HV: tl.constexpr,
    K: tl.constexpr,
    V: tl.constexpr,
    BK: tl.constexpr,
    BV: tl.constexpr,
    USE_INITIAL_STATE: tl.constexpr,  # whether to use initial state
    IS_BETA_HEADWISE: tl.constexpr,  # whether beta is headwise vector or scalar,
    USE_QK_L2NORM_IN_KERNEL: tl.constexpr,
    IS_VARLEN: tl.constexpr,
    DISABLE_STATE_UPDATE: tl.constexpr,  # whether to disable final state update
    DISABLE_OUTPUT_CALCULATION: tl.constexpr,  # whether to disable output calculation
    CACHE_INTERMEDIATE_STATES: tl.constexpr,
    HAS_EAGLE_TREE_CUSTOM_ATTN_MASK: tl.constexpr,
):
    i_k, i_v, i_nh = tl.program_id(0), tl.program_id(1), tl.program_id(2)
    i_n, i_hv = i_nh // HV, i_nh % HV
    i_h = i_hv // (HV // 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)
        all = T
        T = eos - bos
    else:
        bos, eos = i_n * T, i_n * T + T
        all = B * T
    o_k = i_k * BK + tl.arange(0, BK)
    o_v = i_v * BV + tl.arange(0, BV)

    p_q = q + (bos * H + i_h) * K + o_k
    p_k = k + (bos * H + i_h) * K + o_k
    p_v = v + (bos * HV + i_hv) * V + o_v
    if IS_BETA_HEADWISE:
        p_beta = beta + (bos * HV + i_hv) * V + o_v
    else:
        p_beta = beta + bos * HV + i_hv
    p_g = g + bos * HV + i_hv
    p_o = o + ((i_k * all + bos) * HV + i_hv) * V + o_v

    if HAS_EAGLE_TREE_CUSTOM_ATTN_MASK:
        token_indices = tl.arange(0, NP2_T)
        mask_retrieve = token_indices < T
        retrieve_parent_token_base = (
            retrieve_parent_token_ptr
            + (i_n * stride_retrieve_parent_token_seq)
            + token_indices * stride_retrieve_parent_token_token
        )
        parent_idx_tokens = tl.load(retrieve_parent_token_base, mask_retrieve)

    mask_k = o_k < K
    mask_v = o_v < V
    mask_h = mask_k[:, None] & mask_v[None, :]

    b_h = tl.zeros([BK, BV], dtype=tl.float32)
    if USE_INITIAL_STATE:
        idx = tl.load(h0_indices + i_n)
        # Add bounds checking for idx
        if idx >= 0:  # Assuming negative indices are invalid
            p_h0 = (
                h0_source
                + idx * HV * K * V
                + i_hv * K * V
                + o_k[:, None] * V
                + o_v[None, :]
            )
            b_h += tl.load(p_h0, mask=mask_h, other=0).to(tl.float32)

    # Prepare intermediate state cache variables if enabled
    cache_idx = -1
    if CACHE_INTERMEDIATE_STATES:
        cache_idx = tl.load(intermediate_state_indices + i_n)

    step_idx = 0
    for _ in range(0, T):
        if HAS_EAGLE_TREE_CUSTOM_ATTN_MASK:
            # step_idx = 0 should use the b_h from USE_INITIAL_STATE
            if step_idx != 0 and cache_idx >= 0:
                # when calculating current step's attention, load the state from the parent token
                parent_step_idx = tl.sum(
                    tl.where(token_indices == step_idx, parent_idx_tokens, 0)
                )
                step_offset = parent_step_idx * HV * K * V
                cache_ptr = (
                    intermediate_states_buffer
                    + cache_idx * cache_steps * HV * K * V
                    + step_offset
                    + i_hv * K * V
                    + o_k[:, None] * V
                    + o_v[None, :]
                )
                b_h = tl.load(cache_ptr, mask=mask_h, other=0).to(tl.float32)

        b_q = tl.load(p_q, mask=mask_k, other=0).to(tl.float32)
        b_k = tl.load(p_k, mask=mask_k, other=0).to(tl.float32)
        b_v = tl.load(p_v, mask=mask_v, other=0).to(tl.float32)
        b_g = tl.load(p_g).to(tl.float32)

        if USE_QK_L2NORM_IN_KERNEL:
            b_q = b_q / (tl.sqrt(tl.sum(b_q * b_q) + 1e-6))
            b_k = b_k / (tl.sqrt(tl.sum(b_k * b_k) + 1e-6))
        b_q = b_q * scale
        # [BK, BV]
        b_h *= exp(b_g)
        # [BV]
        b_v -= tl.sum(b_h * b_k[:, None], 0)
        if IS_BETA_HEADWISE:
            b_beta = tl.load(p_beta, mask=mask_v, other=0).to(tl.float32)
        else:
            b_beta = tl.load(p_beta).to(tl.float32)
        b_v *= b_beta
        # [BK, BV]
        b_h += b_k[:, None] * b_v[None, :]
        # [BV]
        if not DISABLE_OUTPUT_CALCULATION:
            b_o = tl.sum(b_h * b_q[:, None], 0)
            # core attn output
            tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=mask_v)

        # store intermediate states if enabled
        if CACHE_INTERMEDIATE_STATES:
            if cache_idx >= 0:
                # Compute cache pointer for this step
                step_offset = step_idx * HV * K * V
                cache_ptr = (
                    intermediate_states_buffer
                    + cache_idx * cache_steps * HV * K * V
                    + step_offset
                    + i_hv * K * V
                    + o_k[:, None] * V
                    + o_v[None, :]
                )
                tl.store(cache_ptr, b_h.to(cache_ptr.dtype.element_ty), mask=mask_h)

        step_idx += 1

        p_q += H * K
        p_k += H * K
        p_o += HV * V
        p_v += HV * V
        p_g += HV
        p_beta += HV * (V if IS_BETA_HEADWISE else 1)

    # Store final state back to h0_source with bounds checking
    # ssm states
    if not DISABLE_STATE_UPDATE:
        idx = tl.load(h0_indices + i_n)
        if idx >= 0:  # Add bounds checking
            p_h0 = (
                h0_source
                + idx * HV * K * V
                + i_hv * K * V
                + o_k[:, None] * V
                + o_v[None, :]
            )
            tl.store(p_h0, b_h.to(p_h0.dtype.element_ty), mask=mask_h)


def fused_recurrent_gated_delta_rule_update_fwd(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    g: torch.Tensor,
    beta: torch.Tensor,
    scale: float,
    initial_state_source: torch.Tensor,
    initial_state_indices: torch.Tensor,
    use_qk_l2norm_in_kernel: bool = False,
    cu_seqlens: Optional[torch.LongTensor] = None,
    disable_state_update: bool = False,
    disable_output_calculation: bool = False,
    intermediate_states_buffer: Optional[torch.Tensor] = None,
    intermediate_state_indices: Optional[torch.Tensor] = None,
    cache_steps: Optional[int] = None,
    retrieve_parent_token: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    B, T, H, K, V = *k.shape, v.shape[-1]
    HV = v.shape[2]
    N = B if cu_seqlens is None else len(cu_seqlens) - 1
    BK, BV = triton.next_power_of_2(K), min(triton.next_power_of_2(V), 8)
    NK, NV = triton.cdiv(K, BK), triton.cdiv(V, BV)
    assert NK == 1, "NK > 1 is not supported yet"
    num_stages = 3
    num_warps = 1

    if disable_output_calculation:
        # When output calculation is disabled, allocate minimal tensor
        o = q.new_empty(NK, 1, 1, 1, 1)  # minimal allocation
    else:
        o = q.new_empty(NK, *v.shape)

    grid = (NK, NV, N * HV)

    # prepare retrieve next token buffer strides if provided
    if retrieve_parent_token is not None:
        stride_retrieve_parent_token_seq, stride_retrieve_parent_token_token = (
            retrieve_parent_token.stride(0),
            retrieve_parent_token.stride(1),
        )
    else:
        stride_retrieve_parent_token_seq = stride_retrieve_parent_token_token = 0

    NP2_T = triton.next_power_of_2(T)
    fused_recurrent_gated_delta_rule_update_fwd_kernel[grid](
        q=q,
        k=k,
        v=v,
        g=g,
        beta=beta,
        o=o,
        h0_source=initial_state_source,
        h0_indices=initial_state_indices,
        cu_seqlens=cu_seqlens,
        scale=scale,
        intermediate_states_buffer=intermediate_states_buffer,
        intermediate_state_indices=intermediate_state_indices,
        cache_steps=0 if cache_steps is None else cache_steps,
        retrieve_parent_token_ptr=retrieve_parent_token,
        stride_retrieve_parent_token_seq=stride_retrieve_parent_token_seq,
        stride_retrieve_parent_token_token=stride_retrieve_parent_token_token,
        T=T,
        NP2_T=NP2_T,
        B=B,
        H=H,
        HV=HV,
        K=K,
        V=V,
        BK=BK,
        BV=BV,
        USE_INITIAL_STATE=initial_state_source is not None,
        IS_VARLEN=cu_seqlens is not None,
        CACHE_INTERMEDIATE_STATES=intermediate_states_buffer is not None,
        HAS_EAGLE_TREE_CUSTOM_ATTN_MASK=retrieve_parent_token is not None,
        IS_BETA_HEADWISE=beta.ndim == v.ndim,
        USE_QK_L2NORM_IN_KERNEL=use_qk_l2norm_in_kernel,
        DISABLE_STATE_UPDATE=disable_state_update,
        DISABLE_OUTPUT_CALCULATION=disable_output_calculation,
        num_warps=num_warps,
        num_stages=num_stages,
    )
    o = o.squeeze(0)
    return o


class FusedRecurrentUpdateFunction(torch.autograd.Function):

    @staticmethod
    @input_guard
    def forward(
        ctx,
        q: torch.Tensor,
        k: torch.Tensor,
        v: torch.Tensor,
        g: torch.Tensor,
        beta: torch.Tensor,
        scale: float,
        initial_state_source: torch.Tensor,
        initial_state_indices: torch.Tensor,
        cu_seqlens: Optional[torch.LongTensor] = None,
        use_qk_l2norm_in_kernel: bool = False,
        disable_state_update: bool = False,
        disable_output_calculation: bool = False,
        intermediate_states_buffer: Optional[torch.Tensor] = None,
        intermediate_state_indices: Optional[torch.Tensor] = None,
        cache_steps: Optional[int] = None,
        retrieve_parent_token: Optional[torch.Tensor] = None,
    ):
        o = fused_recurrent_gated_delta_rule_update_fwd(
            q=q,
            k=k,
            v=v,
            g=g,
            beta=beta,
            scale=scale,
            initial_state_source=initial_state_source,
            initial_state_indices=initial_state_indices,
            use_qk_l2norm_in_kernel=use_qk_l2norm_in_kernel,
            cu_seqlens=cu_seqlens,
            disable_state_update=disable_state_update,
            disable_output_calculation=disable_output_calculation,
            intermediate_states_buffer=intermediate_states_buffer,
            intermediate_state_indices=intermediate_state_indices,
            cache_steps=cache_steps,
            retrieve_parent_token=retrieve_parent_token,
        )

        return o

    @staticmethod
    @input_guard
    def backward(ctx, do, dht):
        raise NotImplementedError(
            "Backward pass is not implemented yet and we do not have plans to implement it "
            "because we haven't figured out how to compute dg without materializing the full "
            "hidden states for all time steps."
        )


def fused_recurrent_gated_delta_rule_update(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    g: torch.Tensor,
    beta: torch.Tensor = None,
    scale: float = None,
    initial_state_source: torch.Tensor = None,
    initial_state_indices: torch.Tensor = None,
    cu_seqlens: Optional[torch.LongTensor] = None,
    use_qk_l2norm_in_kernel: bool = False,
    disable_state_update: bool = False,
    disable_output_calculation: bool = False,
    intermediate_states_buffer: Optional[torch.Tensor] = None,
    intermediate_state_indices: Optional[torch.Tensor] = None,
    cache_steps: Optional[int] = None,
    retrieve_parent_token: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    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_source is not None:
            if initial_state_indices.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_indices.shape[0]}."
                )
            if initial_state_indices.shape[0] != intermediate_state_indices.shape[0]:
                raise ValueError(
                    f"The number of intermediate state indices is expected to be equal to the number of input sequences, "
                    f"i.e., {initial_state_indices.shape[0]} != {intermediate_state_indices.shape[0]}."
                )
    if scale is None:
        scale = k.shape[-1] ** -0.5
    else:
        assert scale > 0, "scale must be positive"
    if beta is None:
        beta = torch.ones_like(q[..., 0])
    o = FusedRecurrentUpdateFunction.apply(
        q,
        k,
        v,
        g,
        beta,
        scale,
        initial_state_source,
        initial_state_indices,
        cu_seqlens,
        use_qk_l2norm_in_kernel,
        disable_state_update,
        disable_output_calculation,
        intermediate_states_buffer,
        intermediate_state_indices,
        cache_steps,
        retrieve_parent_token,
    )
    return o