# 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
# mypy: ignore-errors

import contextlib
import functools
from collections.abc import Callable
from typing import Any

import torch
from vllm.triton_utils import triton

# Default chunk size used across FLA Triton kernels.
FLA_CHUNK_SIZE = 64


def tensor_cache(fn: Callable[..., torch.Tensor]) -> Callable[..., torch.Tensor]:
    """Cache recent tensor-derived metadata keyed by tensor object identity."""
    cache_entries: tuple[tuple | None, dict | None, Any] = []
    cache_size = 8

    @functools.wraps(fn)
    def wrapper(*args: Any, **kwargs: Any) -> Any:
        nonlocal cache_entries
        for i, entry in enumerate(cache_entries):
            last_args, last_kwargs, last_result = entry
            if (
                len(args) == len(last_args)
                and len(kwargs) == len(last_kwargs)
                and all(a is b for a, b in zip(args, last_args))
                and all(k in last_kwargs and v is last_kwargs[k] for k, v in kwargs.items())
            ):
                cache_entries = cache_entries[:i] + cache_entries[i + 1 :] + [(args, kwargs, last_result)]
                return last_result

        result = fn(*args, **kwargs)
        if len(cache_entries) >= cache_size:
            cache_entries = cache_entries[1:]
        cache_entries.append((args, kwargs, result))
        return result

    return wrapper


@tensor_cache
def prepare_lens(cu_seqlens: torch.LongTensor) -> torch.LongTensor:
    return cu_seqlens[1:] - cu_seqlens[:-1]


@tensor_cache
def prepare_chunk_indices(cu_seqlens: torch.LongTensor, chunk_size: int) -> torch.LongTensor:
    indices = torch.cat([torch.arange(n) for n in triton.cdiv(prepare_lens(cu_seqlens), chunk_size).tolist()])
    return torch.stack([indices.eq(0).cumsum(0) - 1, indices], 1).to(cu_seqlens)


@tensor_cache
def prepare_chunk_offsets(cu_seqlens: torch.LongTensor, chunk_size: int) -> torch.LongTensor:
    return torch.cat([cu_seqlens.new_tensor([0]), triton.cdiv(prepare_lens(cu_seqlens), chunk_size)]).cumsum(-1)


def input_guard(fn: Callable[..., torch.Tensor]) -> Callable[..., torch.Tensor]:
    """
    A decorator to make sure all input tensors are contiguous and set the device based on input tensors.
    """

    @functools.wraps(fn)
    def wrapper(*args, **kwargs):
        contiguous_args = (i if not isinstance(i, torch.Tensor) else i.contiguous() for i in args)
        contiguous_kwargs = {k: (v if not isinstance(v, torch.Tensor) else v.contiguous()) for k, v in kwargs.items()}

        tensor = None
        for arg in args:
            if isinstance(arg, torch.Tensor):
                tensor = arg
                break
        if tensor is None:
            for value in kwargs.values():
                if isinstance(value, torch.Tensor):
                    tensor = value
                    break

        if tensor is not None:
            ctx = torch.npu.device(tensor.device.index)
        else:
            ctx = contextlib.nullcontext()

        with ctx:
            return fn(*contiguous_args, **contiguous_kwargs)

    return wrapper