"""Isolated hybrid cache branches, including LFM2 convolution state.

The deepcopy + reorder approach follows notnotsamuel/LFM2.5-350M-RLCD
(MIT; see THIRD_PARTY_NOTICES.md). Never broadcast writable state views.
"""

import copy

import torch


def state_tensors(cache):
    for layer in cache.layers:
        for name in ("keys", "values"):
            value = getattr(layer, name, None)
            if isinstance(value, torch.Tensor):
                yield value
        for name in ("conv_states", "recurrent_states"):
            states = getattr(layer, name, {})
            values = states.values() if isinstance(states, dict) else states
            for value in values:
                if isinstance(value, torch.Tensor) and value.numel():
                    yield value


def fork_cache(cache, count):
    if type(count) is not int or count <= 0:
        raise ValueError("branch count must be positive")
    tensors = list(state_tensors(cache))
    if not tensors or any(t.shape[0] != 1 for t in tensors):
        raise ValueError("expected an initialized batch-one prefix cache")
    branch = copy.deepcopy(cache)
    branch.reorder_cache(torch.zeros(count, dtype=torch.long, device=tensors[0].device))
    return branch