from typing import Any
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
# Inference-only adaptation; see LICENSES and NOTICE.
from typing import Optional
import torch
from .chunk_delta_h import chunk_gated_delta_rule_fwd_h
from .chunk_fwd import chunk_gated_delta_rule_fwd_intra
from .chunk_o import chunk_fwd_o
from .cumsum import chunk_local_cumsum_scalar as chunk_local_cumsum
CHUNK_SIZE = 64
SUPPRESS_LEVEL = 0

def chunk_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,
    initial_state_indices: torch.Tensor,
    cu_seqlens: Optional[torch.LongTensor] = None,
    chunk_indices: torch.LongTensor | None = None,
) -> Any:
    g = chunk_local_cumsum(
        g, chunk_size=CHUNK_SIZE, cu_seqlens=cu_seqlens, chunk_indices=chunk_indices
    )

    # fused kkt + solve_tril + recompute_w_u
    w, u, A = chunk_gated_delta_rule_fwd_intra(
        k=k,
        v=v,
        g=g,
        beta=beta,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
    )

    h, v_new = chunk_gated_delta_rule_fwd_h(
        k=k,
        w=w,
        u=u,
        g=g,
        initial_state=initial_state,
        initial_state_indices=initial_state_indices,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
    )
    o = chunk_fwd_o(
        q=q,
        k=k,
        v=v_new,
        h=h,
        g=g,
        scale=scale,
        cu_seqlens=cu_seqlens,
    )
    if SUPPRESS_LEVEL < 3:
        return g, o, A, None, h, None
    elif SUPPRESS_LEVEL >= 3:
        return g, o, A, w, h, v_new

