llm.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from __future__ import annotations

from collections import deque
from collections.abc import Mapping
from copy import copy, deepcopy
from typing import Any, Callable, Iterable, Literal

import torch
from tensordict import (
    lazy_stack,
    NestedKey,
    TensorDict,
    TensorDictBase,
    unravel_key,
)
from tensordict.nn import ProbabilisticTensorDictModule, TensorDictParams
from tensordict.utils import _zip_strict, is_seq_of_nested_key
from torch import nn

from torchrl.data.tensor_specs import Composite, NonTensor, TensorSpec, Unbounded
from torchrl.envs.transforms.transforms import TensorDictPrimer, Transform
from torchrl.envs.transforms.utils import _set_missing_tolerance, _stateless_param
from torchrl.envs.utils import make_composite_from_td


def as_nested_tensor(list_of_tensordicts: list[TensorDictBase]) -> TensorDictBase:
    """Stacks a list of tensordicts into a single tensordict with nested tensors.

    Args:
        list_of_tensordicts (list[TensorDictBase]): A list of tensordicts to stack.

    Returns:
        TensorDictBase: A tensordict with nested tensors.

    """

    def _as_nested_tensor(*list_of_tensors):
        return torch.nested.as_nested_tensor(list_of_tensors, layout=torch.jagged)

    batch_size = list(list_of_tensordicts[0].shape)
    batch_size.insert(0, len(list_of_tensordicts))
    return list_of_tensordicts[0].apply(
        _as_nested_tensor, *list_of_tensordicts[1:], batch_size=batch_size
    )


def as_padded_tensor(
    list_of_tensordicts: list[[TensorDictBase]], dim=0, stack_dim: int = 0
) -> TensorDictBase:
    """Stacks a list of tensordicts into a single tensordict with padded tensors.

    Args:
        list_of_tensordicts (list[[TensorDictBase]]): A list of tensordicts to stack.
        dim (int, optional): The dimension along which to pad. Defaults to 0.
        stack_dim (int, optional): The dimension along which to stack. Defaults to 0.

    Returns:
        TensorDictBase: A tensordict with padded tensors.
    """

    def _stack_tensors(*list_of_tensors):
        if dim < 0:
            raise ValueError("dim must be >= 0")
        max_length = max([t.size(dim) for t in list_of_tensors])

        def pad_tensor(tensor):
            padding_length = max_length - tensor.size(dim)
            shape = [
                s if i != dim else padding_length for i, s in enumerate(tensor.shape)
            ]
            return torch.cat((tensor.new_zeros(shape), tensor), dim=dim)

        return torch.stack([pad_tensor(t) for t in list_of_tensors], dim=stack_dim)

    batch_size = list(list_of_tensordicts[0].shape)
    batch_size.insert(dim, len(list_of_tensordicts))
    result = list_of_tensordicts[0].apply(
        _stack_tensors, *list_of_tensordicts[1:], batch_size=batch_size
    )
    return result


class DataLoadingPrimer(TensorDictPrimer):
    """A primer that loads data from a dataloader and converts it into a tensordict using ``stack_method``.

    Args:
        dataloader (Iterable[Any]): The dataloader to load data from.

    Keyword Args:
        primers (Composite | None, optional): The primers to use for each key in the dataloader. Defaults to None.
        data_keys (List[NestedKey] | None, optional): The keys to use for each item in the dataloader. Defaults to None.
        data_specs (List[TensorSpec] | None, optional): The specs to use for each item in the dataloader. Defaults to None.
        example_data (Any, optional): Example data to use for initializing the primer. Defaults to None.
        stack_method (Callable[[Any], Any] | Literal["as_nested_tensor", "as_padded_tensor"], optional): The method to use for stacking the data. Defaults to ``maybe_dense_stack``.
        use_buffer (bool, optional): Whether to use a buffer to load the batches. When an environment has a batch-size
            that differs from the dataloader's, or when partial resets are to be expected, using a buffer to store data
            ensures that `next()` is called on the dataloader only when necessary, and that elements of the dataset
            are loaded in order.
            Defaults to ``True`` whenever the batch-size of the dataloader is greater than 1.
        auto_batch_size (bool, optional): If ``True`` (default if `dataloader.batch_size > 0`), the batch size of the
            tensordict returned by the transform will be automatically determined assuming that there is a single batch
            dimension.
        repeats (int, optional): How many times the same sample needs to appear successively. This can be useful in
            situations like GRPO where a single prompt is used multiple times to estimate the advantage using Monte-Carlo
            samples (rather than an advantage module).

    Attributes:
        dataloader (Iterable[Any]): The dataloader to load data from.
        endless_dataloader (Iterable[Any]): An endless iterator over the dataloader.
        data_keys (List[NestedKey]): The keys to use for each item in the dataloader.
        stack_method (Callable[[Any], Any]): The method to use for stacking the data.

    .. seealso:: :class:`~torchrl.envs.LLMEnv` and :class:`~torchrl.envs.LLMEnv.from_dataloader`.

    Example of a dataloader yielding strings:
        >>> import random
        >>> import string
        >>> import tensordict as td
        >>> import torch
        >>> from tensordict import TensorDict
        >>> from torchrl.data import Unbounded
        >>> from torchrl.envs import DataLoadingPrimer, LLMEnv
        >>> td.set_capture_non_tensor_stack(False).set()
        >>> class DummyDataLoader:
        ...     '''A dummy dataloader that generates random strings.'''
        ...     def __init__(self, batch_size: int = 0):
        ...         self.batch_size = batch_size
        ...     def generate_random_string(self, length: int = 10) -. str:
        ...         '''Generate a random string of a given length.'''
        ...         return ''.join(random.choice(string.ascii_lowercase) for _ in range(length))
        ...     def __iter__(self):
        ...         return self
        ...     def __next__(self):
        ...         if self.batch_size == 0:
        ...             return self.generate_random_string()
        ...         else:
        ...             return [self.generate_random_string() for _ in range(self.batch_size)]
        >>> # Create an LLM environment with string-to-string input/output.
        >>> env = LLMEnv(str2str=True)
        >>> # Append a DataLoadingPrimer to the environment.
        >>> env = env.append_transform(
        >>>     DataLoadingPrimer(
        >>>         dataloader=DummyDataLoader(),
        >>>         data_keys=["observation"],
        >>>         example_data="a string!",
        >>>     )
        >>> )
        >>> # Test the environment.
        >>> print(env.rand_action(TensorDict()))
        TensorDict(
            fields={
                action: NonTensorData(data=a string, batch_size=torch.Size([]), device=None)},
            batch_size=torch.Size([]),
            device=None,
            is_shared=False)
        >>> print(env.rollout(3))
        TensorDict(
            fields={
                action: NonTensorStack(
                    ['a string', 'a string', 'a string'],
                    batch_size=torch.Size([3]),
                    device=None),
                done: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                next: TensorDict(
                    fields={
                        done: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                        observation: NonTensorStack(
                            ['zxwvupirska string', 'zxwvupirska stringa string...,
                            batch_size=torch.Size([3]),
                            device=None),
                        terminated: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                        truncated: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
                    batch_size=torch.Size([3]),
                    device=None,
                    is_shared=False),
                observation: NonTensorStack(
                    ['zxwvupirsk', 'zxwvupirska string', 'zxwvupirska ...,
                    batch_size=torch.Size([3]),
                    device=None),
                terminated: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([3]),
            device=None,
            is_shared=False)
        >>> # Roll out the environment with a specific initial state.
        >>> init_state = env.reset(TensorDict(batch_size=[3]))
        >>> print(env.rollout(3, auto_reset=False, tensordict=init_state))
        TensorDict(
            fields={
                action: NonTensorStack(
                    [['a string', 'a string', 'a string'], ['a string'...,
                    batch_size=torch.Size([3, 3]),
                    device=None),
                done: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                next: TensorDict(
                    fields={
                        done: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                        observation: NonTensorStack(
                            [[array(['nngcmflsana string', 'vrrbnhzpmga string...,
                            batch_size=torch.Size([3, 3]),
                            device=None),
                        terminated: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                        truncated: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
                    batch_size=torch.Size([3, 3]),
                    device=None,
                    is_shared=False),
                observation: NonTensorStack(
                    [['nngcmflsan', array(['nngcmflsana string', 'vrrb...,
                    batch_size=torch.Size([3, 3]),
                    device=None),
                terminated: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            batch_size=torch.Size([3, 3]),
            device=None,
            is_shared=False)

    Example of dataloader yielding tensors:
        >>> import random
        >>> import string
        >>>
        >>> import tensordict as td
        >>> import torch
        >>> from tensordict import TensorDict
        >>> from torchrl.data import Unbounded
        >>> from torchrl.envs import DataLoadingPrimer, LLMEnv
        >>>
        >>> td.set_capture_non_tensor_stack(False).set()
        >>>
        >>>
        >>> class DummyTensorDataLoader:
        ...     '''A dummy dataloader that generates tensors of random int64 values.'''
        ...
        ...     def __init__(self, batch_size: int = 0, max_length: int = 10, padding: bool = False):
        ...         '''
        ...         Args:
        ...             batch_size (int, optional): The batch size of the generated tensors. Defaults to 0.
        ...             max_length (int, optional): The maximum length of the generated tensors. Defaults to 10.
        ...             padding (bool, optional): Whether to pad the tensors to the maximum length. Defaults to False.
        ...         '''
        ...         self.batch_size = batch_size
        ...         self.max_length = max_length
        ...         self.padding = padding
        ...
        ...     def generate_random_tensor(self) -. torch.Tensor:
        ...         '''Generate a tensor of random int64 values.'''
        ...         length = random.randint(1, self.max_length)
        ...         return torch.tensor([random.randint(0, 100) for _ in range(length)], dtype=torch.int64)
        ...
        ...     def pad_tensor(self, tensor: torch.Tensor) -. torch.Tensor:
        ...         '''Pad a tensor to the maximum length.'''
        ...         padding_length = self.max_length - len(tensor)
        ...         return torch.cat((torch.zeros(padding_length, dtype=torch.int64), tensor))
        ...
        ...     def __iter__(self):
        ...         return self
        ...
        ...     def __next__(self):
        ...         if self.batch_size == 0:
        ...             tensor = self.generate_random_tensor()
        ...             return self.pad_tensor(tensor) if self.padding else tensor
        ...         else:
        ...             tensors = [self.generate_random_tensor() for _ in range(self.batch_size)]
        ...             if self.padding:
        ...                 tensors = [self.pad_tensor(tensor) for tensor in tensors]
        ...                 return torch.stack(tensors)
        ...             else:
        ...                 return tensors
        >>>
        >>> # Create an LLM environment with non-string input/output and append a DataLoadingPrimer.
        >>> env = LLMEnv(str2str=False)
        >>> env = env.append_transform(
        >>>     DataLoadingPrimer(
        >>>         dataloader=DummyTensorDataLoader(),
        >>>         data_keys=["observation"],
        >>>         data_specs=[Unbounded(shape=(-1,), dtype=torch.int64)],
        >>>     )
        >>> )
        >>> print(env.rand_action(TensorDict()))
        TensorDict(
            fields={
                action: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.int64, is_shared=False)},
            batch_size=torch.Size([]),
            device=None,
            is_shared=False)
        >>> print(env.rollout(3))
        LazyStackedTensorDict(
            fields={
                action: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.int64, is_shared=False),
                done: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                next: LazyStackedTensorDict(
                    fields={
                        done: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                        observation: Tensor(shape=torch.Size([3, -1]), device=cpu, dtype=torch.int64, is_shared=False),
                        terminated: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                        truncated: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
                    exclusive_fields={
                    },
                    batch_size=torch.Size([3]),
                    device=None,
                    is_shared=False,
                    stack_dim=0),
                observation: Tensor(shape=torch.Size([3, -1]), device=cpu, dtype=torch.int64, is_shared=False),
                terminated: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([3, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            exclusive_fields={
            },
            batch_size=torch.Size([3]),
            device=None,
            is_shared=False,
            stack_dim=0)
        >>> # Create an LLM environment with padded tensor input/output and append a DataLoadingPrimer.
        >>> env = LLMEnv(str2str=False)
        >>> env = env.append_transform(
        >>>     DataLoadingPrimer(
        >>>         dataloader=DummyTensorDataLoader(padding=True),
        >>>         data_keys=["observation"],
        >>>         data_specs=[Unbounded(shape=(-1,), dtype=torch.int64)],
        >>>         stack_method="as_padded_tensor",
        >>>     )
        >>> )
        >>> print(env.rollout(3, auto_reset=False, tensordict=env.reset(TensorDict(batch_size=[3]))))
        LazyStackedTensorDict(
            fields={
                action: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.int64, is_shared=False),
                done: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                next: LazyStackedTensorDict(
                    fields={
                        done: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                        observation: Tensor(shape=torch.Size([3, 3, -1]), device=cpu, dtype=torch.int64, is_shared=False),
                        terminated: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                        truncated: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
                    exclusive_fields={
                    },
                    batch_size=torch.Size([3, 3]),
                    device=None,
                    is_shared=False,
                    stack_dim=1),
                observation: Tensor(shape=torch.Size([3, 3, -1]), device=cpu, dtype=torch.int64, is_shared=False),
                terminated: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.bool, is_shared=False),
                truncated: Tensor(shape=torch.Size([3, 3, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
            exclusive_fields={
            },
            batch_size=torch.Size([3, 3]),
            device=None,
            is_shared=False,
            stack_dim=1)

    """

    def __init__(
        self,
        dataloader: Iterable[Any],
        *,
        primers: Composite | None = None,
        data_keys: list[NestedKey] | None = None,
        data_specs: list[TensorSpec] | None = None,
        example_data: Any = None,
        stack_method: Callable[[Any], Any]
        | Literal["as_nested_tensor", "as_padded_tensor"] = None,
        use_buffer: bool | None = None,
        auto_batch_size: bool = True,
        repeats: int | None = None,
        device: torch.device | None = None,
    ):
        self.dataloader = dataloader
        if repeats is None:
            repeats = 0
        self.repeats = repeats
        if (
            getattr(dataloader, "batch_size", 1) > 1 and use_buffer is None
        ) or repeats > 0:
            use_buffer = True

        self.use_buffer = use_buffer
        if self.use_buffer:
            self._queue = deque()

        # No auto_batch_size if we know we have a single element
        self.auto_batch_size = auto_batch_size and (
            getattr(dataloader, "batch_size", 1) > 0
        )
        self.endless_dataloader = self._endless_iter(self.dataloader)

        if stack_method is None:
            stack_method = lazy_stack
        elif stack_method == "as_nested_tensor":
            stack_method = as_nested_tensor
        elif stack_method == "as_padded_tensor":
            stack_method = as_padded_tensor
        elif not callable(stack_method):
            raise ValueError(f"Unknown stack_method={stack_method}")
        self.stack_method = stack_method

        if primers is None and not self.use_buffer:
            if data_keys is None:
                data_keys = ["data"]
            if data_specs is None:
                data_specs = [NonTensor(example_data=example_data, shape=())]
            primers = Composite(
                {
                    data_key: data_spec
                    for data_key, data_spec in _zip_strict(data_keys, data_specs)
                }
            )
            self.data_keys = data_keys
        elif primers is None:
            self.data_keys = data_keys
            # We can get the primer from the dataloader itself
            data = self._load_from_dataloader()
            primers = make_composite_from_td(data, dynamic_shape=True)
            self._queue.insert(0, data)
            if data_keys is None:
                self.data_keys = list(primers.keys(True, True))
        else:
            self.data_keys = list(primers.keys(True, True))

        super().__init__(
            primers=primers,
            default_value=self._load_from_dataloader,
            reset_key=None,
            expand_specs=None,
            single_default_value=True,
            call_before_env_reset=True,
            device=device,
        )
        self._reset_key = "_reset"

    @classmethod
    def _endless_iter(self, obj):
        while True:
            yield from obj

    # def _reset_env_preprocess(self, tensordict: TensorDictBase) -> TensorDictBase:
    #     td = super()._reset_env_preprocess(tensordict)
    #     return lazy_stack(list(td.unbind(0)))
    #
    def _load_from_dataloader(self, reset: torch.Tensor | None = None):
        """Loads a single element from the dataloader, or alternatively from the buffer.

        If `reset` is passed, then one element per reset will be loaded.
        """
        if reset is not None:
            if not reset.any():
                raise RuntimeError("reset must have at least one True value.")
            if reset.ndim > 0:
                loaded = [self._load_from_dataloader() for i in range(reset.sum())]
                return self.stack_method(loaded)

        primers = getattr(self, "primers", None)
        if primers is not None:
            device = self.primers.device
        else:
            device = None

        if self.use_buffer and len(self._queue) > 0:
            result = self._queue.popleft()
            if result.device != device:
                result = result.to(device)
            return result

        data = next(self.endless_dataloader)
        # Some heuristic here:
        # if data is a map, assume its keys match the keys in spec
        # TODO: one could rename the keys too
        if isinstance(data, Mapping):
            out = TensorDict.from_dict(
                data,
                auto_batch_size=self.auto_batch_size,
                batch_dims=1,
                device=device,
            )
        elif self.data_keys is None:
            raise RuntimeError(
                f"Cannot lazily instantiate the {type(self).__name__} as the data_keys was "
                f"not passed but the data is not a Mapping, therefore the keys cannot be retrieved "
                f"automatically. Please pass the data_keys to the constructor."
            )
        elif len(self.data_keys) > 1 and isinstance(data, (list, tuple)):
            out = TensorDict.from_dict(
                {k: val for k, val in _zip_strict(self.data_keys, data)},
                auto_batch_size=self.auto_batch_size,
                batch_dims=1,
                device=device,
            )
        elif len(self.data_keys) == 1:
            out = TensorDict.from_dict(
                {self.data_keys[0]: data},
                auto_batch_size=self.auto_batch_size,
                batch_dims=1,
                device=device,
            )
        else:
            raise ValueError(
                f"Unrecognized data type: {type(data)} with keys {self.data_keys}."
            )
        if self.use_buffer:
            if not out.ndim:
                out = out.unsqueeze(0)
            self._queue.extend(
                [d for d in out.unbind(0) for _ in range(max(1, self.repeats))]
            )
            return self._queue.popleft()
        return out


class KLRewardTransform(Transform):
    """A transform to add a KL[pi_current||pi_0] correction term to the reward.

    This transform is used to constrain the policy to remain close to its original
    configuration which limits overfitting when fine-tuning using RLHF.

    Args:
        actor (ProbabilisticTensorDictModule): a probabilistic actor. It must
            have the following features: it must have a set of input (``in_keys``)
            and output keys (``out_keys``). It must have a ``get_dist`` method
            that outputs the distribution of the action.
        coef (:obj:`float`): the coefficient of the KL term. Defaults to ``1.0``.
        in_keys (str or list of str/tuples of str): the input key where the
            reward should be fetched. Defaults to ``"reward"``.
        out_keys (str or list of str/tuples of str): the output key where the
            reward should be written. Defaults to ``"reward"``.
        requires_grad (bool, optional): if ``True``, the frozen parameters will
            consist of differentiable clones of the original params.
            Defaults to ``False``.

    .. note:: If the parameters are not differentiable (default), they will *not*
        follow the module when dtype or device casting operations will be called
        (such as :meth:`cuda`, :meth:`to` etc.). When ``requires_grad=True``,
        casting operations will work as expected.

    Examples:
        >>> from torchrl.envs.libs.gym import GymEnv
        >>> from torchrl.envs import TransformedEnv
        >>> from tensordict.nn import TensorDictModule as Mod, NormalParamExtractor
        >>> from torchrl.modules import ProbabilisticActor
        >>> from tensordict import TensorDict
        >>> from torchrl.modules.distributions import TanhNormal
        >>> from torch import nn
        >>> base_env = GymEnv("Pendulum-v1")
        >>> n_obs = base_env.observation_spec["observation"].shape[-1]
        >>> n_act = base_env.action_spec.shape[-1]
        >>> module = Mod(
        ...     nn.Sequential(nn.Linear(n_obs, n_act * 2), NormalParamExtractor()),
        ...     in_keys=["observation"],
        ...     out_keys=["loc", "scale"],
        ... )
        >>> actor = ProbabilisticActor(
        ...     module,
        ...     in_keys=["loc", "scale"],
        ...     distribution_class=TanhNormal,
        ...     return_log_prob=True,
        ... )
        >>> transform = KLRewardTransform(actor, out_keys="reward_kl")
        >>> env = TransformedEnv(base_env, transform)
        >>> with torch.no_grad():
        ...     # modify the actor parameters
        ...     _ = TensorDict(dict(actor.named_parameters()), []).apply_(lambda x: x.data.copy_(x.data + 1))
        ...     td = env.rollout(3, actor)
        >>> # check that rewards have been modified
        >>> assert (td.get(("next", "reward")) != td.get(("next", "reward_kl"))).all()

    .. note:: Because the KL formula is not always available and the parameters of the
      original distribution may not have been recorded, we use a stochastic estimate
      of the KL divergence.

    """

    DEFAULT_IN_KEYS = ["reward"]

    def __init__(
        self,
        actor: ProbabilisticTensorDictModule,
        coef=1.0,
        in_keys=None,
        out_keys=None,
        requires_grad=False,
    ):
        if in_keys is None:
            in_keys = self.DEFAULT_IN_KEYS
        if out_keys is None:
            out_keys = copy(in_keys)
        super().__init__(in_keys=in_keys, out_keys=out_keys)
        if not is_seq_of_nested_key(self.in_keys) or not is_seq_of_nested_key(
            self.out_keys
        ):
            raise ValueError(
                f"invalid in_keys / out_keys:\nin_keys={self.in_keys} \nout_keys={self.out_keys}"
            )
        if len(self.in_keys) != 1 or len(self.out_keys) != 1:
            raise ValueError(
                f"Only one in_key/out_key is allowed, got in_keys={self.in_keys}, out_keys={self.out_keys}."
            )
        # for convenience, convert out_keys to tuples
        self._out_keys = [
            out_key if isinstance(out_key, tuple) else (out_key,)
            for out_key in self._out_keys
        ]

        # update the in_keys for dispatch etc
        self.in_keys = self.in_keys + actor.in_keys

        # check that the model has parameters
        params = TensorDict.from_module(actor)
        with params.apply(
            _stateless_param, device="meta", filter_empty=False
        ).to_module(actor):
            # copy a stateless actor
            self.__dict__["functional_actor"] = deepcopy(actor)
        # we need to register these params as buffer to have `to` and similar
        # methods work properly

        def _make_detached_param(x):

            if isinstance(x, nn.Parameter):
                # we need an nn.Parameter since some modules (RNN) require nn.Parameters
                return nn.Parameter(x.data.clone(), requires_grad=requires_grad)
            elif x.requires_grad:
                raise ValueError(
                    "Encountered a value that requires gradients but is not an nn.Parameter instance."
                )
            return x.clone()

        self.frozen_params = params.apply(_make_detached_param, filter_empty=False)
        if requires_grad:
            # includes the frozen params/buffers in the module parameters/buffers
            self.frozen_params = TensorDictParams(self.frozen_params, no_convert=True)

        # self._buffers["actor_params"] = params.clone().detach()

        # find the sample log-prob key
        self.sample_log_prob_key = "sample_log_prob"

        def find_sample_log_prob(module):
            if hasattr(module, "log_prob_key"):
                self.sample_log_prob_key = module.log_prob_key

        self.functional_actor.apply(find_sample_log_prob)

        if not isinstance(coef, torch.Tensor):
            coef = torch.as_tensor(coef)
        self.register_buffer("coef", coef)

    def _reset(
        self, tensordict: TensorDictBase, tensordict_reset: TensorDictBase
    ) -> TensorDictBase:
        with _set_missing_tolerance(self, True):
            tensordict_reset = self._call(tensordict_reset)
        return tensordict_reset

    def _call(self, next_tensordict: TensorDictBase) -> TensorDictBase:
        # run the actor on the tensordict
        action = next_tensordict.get("action", None)
        if action is None:
            # being called after reset or without action, skipping
            if self.out_keys[0] != ("reward",) and self.parent is not None:
                next_tensordict.set(self.out_keys[0], self.parent.reward_spec.zero())
            return next_tensordict
        with self.frozen_params.to_module(self.functional_actor):
            dist = self.functional_actor.get_dist(next_tensordict.clone(False))
        # get the log_prob given the original model
        log_prob = dist.log_prob(action)
        reward_key = self.in_keys[0]
        reward = next_tensordict.get("next").get(reward_key)
        curr_log_prob = next_tensordict.get(self.sample_log_prob_key)
        # we use the unbiased consistent estimator of the KL: log_p(x) - log_q(x) when x ~ p(x)
        kl = (curr_log_prob - log_prob).view_as(reward)
        next_tensordict.set(("next", *self.out_keys[0]), reward + self.coef * kl)
        return next_tensordict

    def _step(
        self, tensordict: TensorDictBase, next_tensordict: TensorDictBase
    ) -> TensorDictBase:
        with tensordict.unlock_():
            return self._call(tensordict.set("next", next_tensordict)).pop("next")

    forward = _call

    def transform_output_spec(self, output_spec: Composite) -> Composite:
        in_key = unravel_key(self.in_keys[0])
        out_key = unravel_key(self.out_keys[0])

        if in_key == "reward" and out_key == "reward":
            parent = self.parent
            reward_spec = Unbounded(
                device=output_spec.device,
                shape=output_spec["full_reward_spec"][parent.reward_key].shape,
            )
            output_spec["full_reward_spec"] = Composite(
                {parent.reward_key: reward_spec},
                shape=output_spec["full_reward_spec"].shape,
            )
        elif in_key == "reward":
            parent = self.parent
            reward_spec = output_spec["full_reward_spec"][parent.reward_key].clone()
            # then we need to populate the output keys
            observation_spec = output_spec["full_observation_spec"]
            observation_spec[out_key] = reward_spec
        else:
            observation_spec = output_spec["full_observation_spec"]
            reward_spec = observation_spec[in_key].clone()
            # then we need to populate the output keys
            observation_spec[out_key] = reward_spec
        return output_spec