History¶

append(history: History, *, inplace: bool = True, dim: int = - 1) → History[source]¶

Appends a new history to the current one.

Parameters:

history (History) – The new history to append.
inplace (bool, optional) – Whether to perform the operation in-place. Defaults to True.
dim (int, optional) – The dimension to append along. Defaults to -1.

Returns:

The appended History object.

Return type:

History

apply_chat_template(*, tokenizer: transformers.AutoTokenizer | transformers.AutoProcessor, add_generation_prompt: bool = True, chat_template: str | None = None, chat_template_name: str | None = None, continue_final_message: bool = False, tokenize: bool | None = None, padding: bool | str = False, truncation: bool | str = False, return_tensors: str | None = None, return_dict: bool | None = None, return_assistant_tokens_mask: bool = False, **kwargs) → str | list[str] | TensorDict[source]¶

Applies a chat template to the history.

Keyword Arguments:

tokenizer (transformers.PreTrainedTokenizer | transformers.AutoProcessor) – The tokenizer to use.
add_generation_prompt (bool, optional) – Whether to add a generation prompt (e.g. “<|im_start|>assistant”). Defaults to True.
chat_template (str, optional) – The chat template to use. Defaults to the tokenizer’s default template.
chat_template_name (str, optional) – The name of the chat template to use. Prevalent over tokenizer.chat_template. If None, the method will automatically detect the model family and use the appropriate template. Defaults to None.
continue_final_message (bool, optional) – Whether to continue the final message. Defaults to False.
tokenize (bool, optional) – Whether to tokenize the output. Defaults to False.
padding (bool | str, optional) – The padding strategy to use. Defaults to False.
truncation (bool | str, optional) – The truncation strategy to use. Defaults to False.
return_tensors (str | None, optional) – The type of tensors to return. Defaults to “pt”.
return_dict (bool, optional) – Whether to return a dictionary. Defaults to False.
return_assistant_tokens_mask (bool, optional) –
Whether to return a mask of the assistant generated tokens. If True, the mask will be written to the assistant_masks key. For tokens generated by the assistant, the mask will contain 1. For user and system tokens, the mask will contain 0. This functionality is only available for chat templates that support it via the {% generation %} keyword. Defaults to False.

Note

Assistant token masking is supported across multiple model families: - Qwen family: Uses custom template with full tool calling support - DialoGPT family: Uses custom template for conversation format - Falcon family: Uses custom template for instruction format - DeepSeek family: Uses custom template with native format - Other models: Use the default chatml_format template

The method automatically detects the model family and selects the appropriate template.
**kwargs – Additional keyword arguments to pass to the tokenizer apply_chat_template method.

Returns:

The formatted history.

cat(dim: int = 0, *, out=None)¶

Concatenates tensordicts into a single tensordict along the given dimension.

This call is equivalent to calling torch.cat() but is compatible with torch.compile.

classmethod default_spec(shape=(- 1,))[source]¶

A default spec to use in transforms / envs that return History objects.

Parameters:: shape (torch.Size, optional) – The shape of the returned History spec. Defaults to (-1) (variable length along the time dimension).

Example

>>> import tensordict
>>> from torchrl.data import History
>>> tensordict.set_list_to_stack(True).set()
>>>
>>> history = History(role=["system", "user"], content=["a message", "another message"], batch_size=(2,))
>>> spec = history.default_spec()
>>> print(spec)
Composite(
    role: NonTensor(
        shape=torch.Size([-1]),
        space=None,
        device=None,
        dtype=None,
        domain=None,
        example_data=foo),
    content: NonTensor(
        shape=torch.Size([-1]),
        space=None,
        device=None,
        dtype=None,
        domain=None,
        example_data=foo),
    device=None,
    shape=torch.Size([-1]))
>>> print(spec.zero())
History(
    content=NonTensorData(data=foo, batch_size=torch.Size([1]), device=None),
    role=NonTensorData(data=foo, batch_size=torch.Size([1]), device=None),
    batch_size=torch.Size([1]),
    device=None,
    is_shared=False)

property device: device¶: Retrieves the device type of tensor class.

dumps(prefix: str | None = None, copy_existing: bool = False, *, num_threads: int = 0, return_early: bool = False, share_non_tensor: bool = False) → Self¶

Saves the tensordict to disk.

This function is a proxy to memmap().

classmethod fields()¶

Return a tuple describing the fields of this dataclass.

Accepts a dataclass or an instance of one. Tuple elements are of type Field.

from_any(*, auto_batch_size: bool = False, batch_dims: int | None = None, device: torch.device | None = None, batch_size: torch.Size | None = None)¶

Recursively converts any object to a TensorDict.

Note

from_any is less restrictive than the regular TensorDict constructor. It can cast data structures like dataclasses or tuples to a tensordict using custom heuristics. This approach may incur some extra overhead and involves more opinionated choices in terms of mapping strategies.

Note

This method recursively converts the input object to a TensorDict. If the object is already a TensorDict (or any similar tensor collection object), it will be returned as is.

Parameters:

obj – The object to be converted.

Keyword Arguments:

auto_batch_size (bool, optional) – if True, the batch size will be computed automatically. Defaults to False.
batch_dims (int, optional) – If auto_batch_size is True, defines how many dimensions the output tensordict should have. Defaults to None (full batch-size at each level).
device (torch.device, optional) – The device on which the TensorDict will be created.
batch_size (torch.Size, optional) – The batch size of the TensorDict. Exclusive with auto_batch_size.

Returns:

A TensorDict representation of the input object.

Supported objects:

Dataclasses through from_dataclass() (dataclasses will be converted to TensorDict instances, not tensorclasses).
Namedtuples through from_namedtuple().
Dictionaries through from_dict().
Tuples through from_tuple().
NumPy’s structured arrays through from_struct_array().
HDF5 objects through from_h5().

classmethod from_chats(chats: list[list[dict]]) → History[source]¶

Create a History object from a list of chats.

Parameters:: chats (list[list[dict]]) – A list of chats, where each chat is a list of dictionaries.

from_dataclass(*, dest_cls: Type | None = None, auto_batch_size: bool = False, batch_dims: int | None = None, as_tensorclass: bool = False, device: torch.device | None = None, batch_size: torch.Size | None = None)¶

Converts a dataclass into a TensorDict instance.

Parameters:

dataclass – The dataclass instance to be converted.

Keyword Arguments:

dest_cls (tensorclass, optional) – A tensorclass type to be used to map the data. If not provided, a new class is created. Without effect if obj is a type or as_tensorclass is False.
auto_batch_size (bool, optional) – If True, automatically determines and applies batch size to the resulting TensorDict. Defaults to False.
batch_dims (int, optional) – If auto_batch_size is True, defines how many dimensions the output tensordict should have. Defaults to None (full batch-size at each level).
as_tensorclass (bool, optional) – If True, delegates the conversion to the free function from_dataclass() and returns a tensor-compatible class (tensorclass()) or instance instead of a TensorDict. Defaults to False.
device (torch.device, optional) – The device on which the TensorDict will be created. Defaults to None.
batch_size (torch.Size, optional) – The batch size of the TensorDict. Defaults to None.

Returns:

A TensorDict instance derived from the provided dataclass, unless as_tensorclass is True, in which case a tensor-compatible class or instance is returned.

Raises:

TypeError – If the provided input is not a dataclass instance.

Warning

This method is distinct from the free function from_dataclass and serves a different purpose. While the free function returns a tensor-compatible class or instance, this method returns a TensorDict instance.

Note

This method creates a new TensorDict instance with keys corresponding to the fields of the input dataclass.
Each key in the resulting TensorDict is initialized using the cls.from_any method.
The auto_batch_size option allows for automatic batch size determination and application to the resulting TensorDict.

from_h5(*, mode: str = 'r', auto_batch_size: bool = False, batch_dims: int | None = None, batch_size: torch.Size | None = None)¶

Creates a PersistentTensorDict from a h5 file.

Parameters:

filename (str) – The path to the h5 file.

Keyword Arguments:

mode (str, optional) – Reading mode. Defaults to "r".
auto_batch_size (bool, optional) – If True, the batch size will be computed automatically. Defaults to False.
batch_dims (int, optional) – If auto_batch_size is True, defines how many dimensions the output tensordict should have. Defaults to None (full batch-size at each level).
batch_size (torch.Size, optional) – The batch size of the TensorDict. Defaults to None.

Returns:

A PersistentTensorDict representation of the input h5 file.

Examples

>>> td = TensorDict.from_h5("path/to/file.h5")
>>> print(td)
PersistentTensorDict(
    fields={
        key1: Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, is_shared=False),
        key2: Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([]),
    device=None,
    is_shared=False)

from_modules(*, as_module: bool = False, lock: bool = True, use_state_dict: bool = False, lazy_stack: bool = False, expand_identical: bool = False)¶

Retrieves the parameters of several modules for ensebmle learning/feature of expects applications through vmap.

Parameters:

modules (sequence of nn.Module) – the modules to get the parameters from. If the modules differ in their structure, a lazy stack is needed (see the lazy_stack argument below).

Keyword Arguments:

as_module (bool, optional) – if True, a TensorDictParams instance will be returned which can be used to store parameters within a torch.nn.Module. Defaults to False.
lock (bool, optional) – if True, the resulting tensordict will be locked. Defaults to True.
use_state_dict (bool, optional) –
if True, the state-dict from the module will be used and unflattened into a TensorDict with the tree structure of the model. Defaults to False.

Note

This is particularly useful when state-dict hooks have to be used.
lazy_stack (bool, optional) –
whether parameters should be densly or lazily stacked. Defaults to False (dense stack).

Note

lazy_stack and as_module are exclusive features.

Warning

There is a crucial difference between lazy and non-lazy outputs in that non-lazy output will reinstantiate parameters with the desired batch-size, while lazy_stack will just represent the parameters as lazily stacked. This means that whilst the original parameters can safely be passed to an optimizer when lazy_stack=True, the new parameters need to be passed when it is set to True.

Warning

Whilst it can be tempting to use a lazy stack to keep the orignal parameter references, remember that lazy stack perform a stack each time get() is called. This will require memory (N times the size of the parameters, more if a graph is built) and time to be computed. It also means that the optimizer(s) will contain more parameters, and operations like step() or zero_grad() will take longer to be executed. In general, lazy_stack should be reserved to very few use cases.
expand_identical (bool, optional) – if True and the same parameter (same identity) is being stacked to itself, an expanded version of this parameter will be returned instead. This argument is ignored when lazy_stack=True.

Examples

>>> from torch import nn
>>> from tensordict import TensorDict
>>> torch.manual_seed(0)
>>> empty_module = nn.Linear(3, 4, device="meta")
>>> n_models = 2
>>> modules = [nn.Linear(3, 4) for _ in range(n_models)]
>>> params = TensorDict.from_modules(*modules)
>>> print(params)
TensorDict(
    fields={
        bias: Parameter(shape=torch.Size([2, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        weight: Parameter(shape=torch.Size([2, 4, 3]), device=cpu, dtype=torch.float32, is_shared=False)},
    batch_size=torch.Size([2]),
    device=None,
    is_shared=False)
>>> # example of batch execution
>>> def exec_module(params, x):
...     with params.to_module(empty_module):
...         return empty_module(x)
>>> x = torch.randn(3)
>>> y = torch.vmap(exec_module, (0, None))(params, x)
>>> assert y.shape == (n_models, 4)
>>> # since lazy_stack = False, backprop leaves the original params untouched
>>> y.sum().backward()
>>> assert params["weight"].grad.norm() > 0
>>> assert modules[0].weight.grad is None

With lazy_stack=True, things are slightly different:

>>> params = TensorDict.from_modules(*modules, lazy_stack=True)
>>> print(params)
LazyStackedTensorDict(
    fields={
        bias: Tensor(shape=torch.Size([2, 4]), device=cpu, dtype=torch.float32, is_shared=False),
        weight: Tensor(shape=torch.Size([2, 4, 3]), device=cpu, dtype=torch.float32, is_shared=False)},
    exclusive_fields={
    },
    batch_size=torch.Size([2]),
    device=None,
    is_shared=False,
    stack_dim=0)
>>> # example of batch execution
>>> y = torch.vmap(exec_module, (0, None))(params, x)
>>> assert y.shape == (n_models, 4)
>>> y.sum().backward()
>>> assert modules[0].weight.grad is not None

from_namedtuple(*, auto_batch_size: bool = False, batch_dims: int | None = None, device: torch.device | None = None, batch_size: torch.Size | None = None)¶

Converts a namedtuple to a TensorDict recursively.

Parameters:

named_tuple – The namedtuple instance to be converted.

Keyword Arguments:

auto_batch_size (bool, optional) – if True, the batch size will be computed automatically. Defaults to False.
batch_dims (int, optional) – If auto_batch_size is True, defines how many dimensions the output tensordict should have. Defaults to None (full batch-size at each level).
device (torch.device, optional) – The device on which the TensorDict will be created. Defaults to None.
batch_size (torch.Size, optional) – The batch size of the TensorDict. Defaults to None.

Returns:

A TensorDict representation of the input namedtuple.

Examples

>>> from tensordict import TensorDict
>>> import torch
>>> data = TensorDict({
...     "a_tensor": torch.zeros((3)),
...     "nested": {"a_tensor": torch.zeros((3)), "a_string": "zero!"}}, [3])
>>> nt = data.to_namedtuple()
>>> print(nt)
GenericDict(a_tensor=tensor([0., 0., 0.]), nested=GenericDict(a_tensor=tensor([0., 0., 0.]), a_string='zero!'))
>>> TensorDict.from_namedtuple(nt, auto_batch_size=True)
TensorDict(
    fields={
        a_tensor: Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, is_shared=False),
        nested: TensorDict(
            fields={
                a_string: NonTensorData(data=zero!, batch_size=torch.Size([3]), device=None),
                a_tensor: Tensor(shape=torch.Size([3]), device=cpu, dtype=torch.float32, is_shared=False)},
            batch_size=torch.Size([3]),
            device=None,
            is_shared=False)},
    batch_size=torch.Size([3]),
    device=None,
    is_shared=False)

from_pytree(*, batch_size: torch.Size | None = None, auto_batch_size: bool = False, batch_dims: int | None = None)¶

Converts a pytree to a TensorDict instance.

This method is designed to keep the pytree nested structure as much as possible.

Additional non-tensor keys are added to keep track of each level’s identity, providing a built-in pytree-to-tensordict bijective transform API.

Accepted classes currently include lists, tuples, named tuples and dict.

Note

For dictionaries, non-NestedKey keys are registered separately as NonTensorData instances.

Note

Tensor-castable types (such as int, float or np.ndarray) will be converted to torch.Tensor instances. Note that this transformation is surjective: transforming back the tensordict to a pytree will not recover the original types.

Examples

>>> # Create a pytree with tensor leaves, and one "weird"-looking dict key
>>> class WeirdLookingClass:
...     pass
...
>>> weird_key = WeirdLookingClass()
>>> # Make a pytree with tuple, lists, dict and namedtuple
>>> pytree = (
...     [torch.randint(10, (3,)), torch.zeros(2)],
...     {
...         "tensor": torch.randn(
...             2,
...         ),
...         "td": TensorDict({"one": 1}),
...         weird_key: torch.randint(10, (2,)),
...         "list": [1, 2, 3],
...     },
...     {"named_tuple": TensorDict({"two": torch.ones(1) * 2}).to_namedtuple()},
... )
>>> # Build a TensorDict from that pytree
>>> td = TensorDict.from_pytree(pytree)
>>> # Recover the pytree
>>> pytree_recon = td.to_pytree()
>>> # Check that the leaves match
>>> def check(v1, v2):
>>>     assert (v1 == v2).all()
>>>
>>> torch.utils._pytree.tree_map(check, pytree, pytree_recon)
>>> assert weird_key in pytree_recon[1]

from_remote_init(group: 'ProcessGroup' | None = None, device: torch.device | None = None) → Self¶

Creates a new tensordict instance initialized from remotely sent metadata.

This class method receives the metadata sent by init_remote, creates a new tensordict with matching shape and dtype, and then asynchronously receives the actual tensordict content.

Parameters:

src (int) – The rank of the source process that sent the metadata.
group ("ProcessGroup", optional) – The process group to use for communication. Defaults to None.
device (torch.device, optional) – The device to use for tensor operations. Defaults to None.

Returns:

A new tensordict instance initialized with the received metadata and content.

Return type:

TensorDict

See also

The sending process should have called ~.init_remote to send the metadata and content.

from_struct_array(*, auto_batch_size: bool = False, batch_dims: int | None = None, device: torch.device | None = None, batch_size: torch.Size | None = None) → Self¶

Converts a structured numpy array to a TensorDict.

The resulting TensorDict will share the same memory content as the numpy array (it is a zero-copy operation). Changing values of the structured numpy array in-place will affect the content of the TensorDict.

Note

This method performs a zero-copy operation, meaning that the resulting TensorDict will share the same memory content as the input numpy array. Therefore, changing values of the numpy array in-place will affect the content of the TensorDict.

Parameters:

struct_array (np.ndarray) – The structured numpy array to be converted.

Keyword Arguments:

auto_batch_size (bool, optional) – If True, the batch size will be computed automatically. Defaults to False.
batch_dims (int, optional) – If auto_batch_size is True, defines how many dimensions the output tensordict should have. Defaults to None (full batch-size at each level).
device (torch.device, optional) –
The device on which the TensorDict will be created. Defaults to None.

Note

Changing the device (i.e., specifying any device other than None or "cpu") will transfer the data, resulting in a change to the memory location of the returned data.
batch_size (torch.Size, optional) – The batch size of the TensorDict. Defaults to None.

Returns:

A TensorDict representation of the input structured numpy array.

Examples

>>> x = np.array(
...     [("Rex", 9, 81.0), ("Fido", 3, 27.0)],
...     dtype=[("name", "U10"), ("age", "i4"), ("weight", "f4")],
... )
>>> td = TensorDict.from_struct_array(x)
>>> x_recon = td.to_struct_array()
>>> assert (x_recon == x).all()
>>> assert x_recon.shape == x.shape
>>> # Try modifying x age field and check effect on td
>>> x["age"] += 1
>>> assert (td["age"] == np.array([10, 4])).all()

classmethod from_tensordict(tensordict: TensorDictBase, non_tensordict: dict | None = None, safe: bool = True) → Self¶

Tensor class wrapper to instantiate a new tensor class object.

Parameters:

tensordict (TensorDictBase) – Dictionary of tensor types
non_tensordict (dict) – Dictionary with non-tensor and nested tensor class objects
safe (bool) – Whether to raise an error if the tensordict is not a TensorDictBase instance

Inverts a chat template into a History object.

Parameters:

text (str | list[str]) – The chat template to invert.
chat_template_name (str, optional) – The name of the chat template to use.
tokenizer (transformers.AutoTokenizer | transformers.AutoProcessor, optional) – The tokenizer to use.

Returns:

The inverted History object.

Return type:

History

Examples

>>> from torchrl.data.llm.history import History
>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
>>> text = "<|im_start|>system\nYou are a helpful assistant.\n<|im_end|>\n<|im_start|>user\nWrite a python script that gives the capital of France or Germany.\n<|im_end|>\n<|im_start|>assistant\n<think>The capital of France is Paris, the capital of Germany is Berlin.</think>\n<answer><python>\n"
>>> history = History.from_text(text, tokenizer=tokenizer)
>>> print(history)
History(
    content=NonTensorStack(
        ['You are a helpful assistant.', 'Write a python s...,
        batch_size=torch.Size([3]),
        device=None),
    is_complete=NonTensorStack(
        [True, True, False],
        batch_size=torch.Size([3]),
        device=None),
    role=NonTensorStack(
        ['system', 'user', 'assistant'],
        batch_size=torch.Size([3]),
        device=None),
    tool_calls=None,
    tool_responses=None,
    batch_size=torch.Size([3]),
    device=None,
    is_shared=False)

from_tuple(*, auto_batch_size: bool = False, batch_dims: int | None = None, device: torch.device | None = None, batch_size: torch.Size | None = None)¶

Converts a tuple to a TensorDict.

Parameters:

obj – The tuple instance to be converted.

Keyword Arguments:

auto_batch_size (bool, optional) – If True, the batch size will be computed automatically. Defaults to False.
batch_dims (int, optional) – If auto_batch_size is True, defines how many dimensions the output tensordict should have. Defaults to None (full batch-size at each level).
device (torch.device, optional) – The device on which the TensorDict will be created. Defaults to None.
batch_size (torch.Size, optional) – The batch size of the TensorDict. Defaults to None.

Returns:

A TensorDict representation of the input tuple.

Examples

>>> my_tuple = (1, 2, 3)
>>> td = TensorDict.from_tuple(my_tuple)
>>> print(td)
TensorDict(
    fields={
        0: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.int64, is_shared=False),
        1: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.int64, is_shared=False),
        2: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.int64, is_shared=False)},
    batch_size=torch.Size([]),
    device=None,
    is_shared=False)

fromkeys(value: Any = 0)¶

Creates a tensordict from a list of keys and a single value.

Parameters:

keys (list of NestedKey) – An iterable specifying the keys of the new dictionary.
value (compatible type, optional) – The value for all keys. Defaults to 0.

get(key: NestedKey, *args, **kwargs)¶

Gets the value stored with the input key.

Parameters:

key (str, tuple of str) – key to be queried. If tuple of str it is equivalent to chained calls of getattr.
default – default value if the key is not found in the tensorclass.

Returns:

value stored with the input key

lazy_stack(dim: int = 0, *, out=None, **kwargs)¶

Creates a lazy stack of tensordicts.

See lazy_stack() for details.

load(*args, **kwargs) → Self¶

Loads a tensordict from disk.

This class method is a proxy to load_memmap().

load_(prefix: str | Path, *args, **kwargs)¶

Loads a tensordict from disk within the current tensordict.

This class method is a proxy to load_memmap_().

load_memmap(device: torch.device | None = None, non_blocking: bool = False, *, out: TensorDictBase | None = None) → Self¶

Loads a memory-mapped tensordict from disk.

Parameters:

prefix (str or Path to folder) – the path to the folder where the saved tensordict should be fetched.
device (torch.device or equivalent, optional) – if provided, the data will be asynchronously cast to that device. Supports “meta” device, in which case the data isn’t loaded but a set of empty “meta” tensors are created. This is useful to get a sense of the total model size and structure without actually opening any file.
non_blocking (bool, optional) – if True, synchronize won’t be called after loading tensors on device. Defaults to False.
out (TensorDictBase, optional) – optional tensordict where the data should be written.

Examples

>>> from tensordict import TensorDict
>>> td = TensorDict.fromkeys(["a", "b", "c", ("nested", "e")], 0)
>>> td.memmap("./saved_td")
>>> td_load = TensorDict.load_memmap("./saved_td")
>>> assert (td == td_load).all()

This method also allows loading nested tensordicts.

Examples

>>> nested = TensorDict.load_memmap("./saved_td/nested")
>>> assert nested["e"] == 0

A tensordict can also be loaded on “meta” device or, alternatively, as a fake tensor.

Examples

>>> import tempfile
>>> td = TensorDict({"a": torch.zeros(()), "b": {"c": torch.zeros(())}})
>>> with tempfile.TemporaryDirectory() as path:
...     td.save(path)
...     td_load = TensorDict.load_memmap(path, device="meta")
...     print("meta:", td_load)
...     from torch._subclasses import FakeTensorMode
...     with FakeTensorMode():
...         td_load = TensorDict.load_memmap(path)
...         print("fake:", td_load)
meta: TensorDict(
    fields={
        a: Tensor(shape=torch.Size([]), device=meta, dtype=torch.float32, is_shared=False),
        b: TensorDict(
            fields={
                c: Tensor(shape=torch.Size([]), device=meta, dtype=torch.float32, is_shared=False)},
            batch_size=torch.Size([]),
            device=meta,
            is_shared=False)},
    batch_size=torch.Size([]),
    device=meta,
    is_shared=False)
fake: TensorDict(
    fields={
        a: FakeTensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),
        b: TensorDict(
            fields={
                c: FakeTensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False)},
            batch_size=torch.Size([]),
            device=cpu,
            is_shared=False)},
    batch_size=torch.Size([]),
    device=cpu,
    is_shared=False)

load_state_dict(state_dict: dict[str, Any], strict=True, assign=False, from_flatten=False)¶: Loads a state_dict attemptedly in-place on the destination tensorclass.

maybe_dense_stack(dim: int = 0, *, out=None, **kwargs)¶

Attempts to make a dense stack of tensordicts, and falls back on lazy stack when required..

See maybe_dense_stack() for details.

memmap(prefix: str | None = None, copy_existing: bool = False, *, num_threads: int = 0, return_early: bool = False, share_non_tensor: bool = False, existsok: bool = True) → Self¶

Writes all tensors onto a corresponding memory-mapped Tensor in a new tensordict.

Parameters:

prefix (str) – directory prefix where the memory-mapped tensors will be stored. The directory tree structure will mimic the tensordict’s.
copy_existing (bool) – If False (default), an exception will be raised if an entry in the tensordict is already a tensor stored on disk with an associated file, but is not saved in the correct location according to prefix. If True, any existing Tensor will be copied to the new location.

Keyword Arguments:

num_threads (int, optional) – the number of threads used to write the memmap tensors. Defaults to 0.
return_early (bool, optional) – if True and num_threads>0, the method will return a future of the tensordict.
share_non_tensor (bool, optional) – if True, the non-tensor data will be shared between the processes and writing operation (such as inplace update or set) on any of the workers within a single node will update the value on all other workers. If the number of non-tensor leaves is high (e.g., sharing large stacks of non-tensor data) this may result in OOM or similar errors. Defaults to False.
existsok (bool, optional) – if False, an exception will be raised if a tensor already exists in the same path. Defaults to True.

The TensorDict is then locked, meaning that any writing operations that isn’t in-place will throw an exception (eg, rename, set or remove an entry). Once the tensordict is unlocked, the memory-mapped attribute is turned to False, because cross-process identity is not guaranteed anymore.

Returns:: A new tensordict with the tensors stored on disk if return_early=False, otherwise a TensorDictFuture instance.

Note

Serialising in this fashion might be slow with deeply nested tensordicts, so it is not recommended to call this method inside a training loop.

memmap_(prefix: str | None = None, copy_existing: bool = False, *, num_threads: int = 0, return_early: bool = False, share_non_tensor: bool = False, existsok: bool = True) → Self¶

Writes all tensors onto a corresponding memory-mapped Tensor, in-place.

Parameters:

prefix (str) – directory prefix where the memory-mapped tensors will be stored. The directory tree structure will mimic the tensordict’s.
copy_existing (bool) – If False (default), an exception will be raised if an entry in the tensordict is already a tensor stored on disk with an associated file, but is not saved in the correct location according to prefix. If True, any existing Tensor will be copied to the new location.

Keyword Arguments:

num_threads (int, optional) – the number of threads used to write the memmap tensors. Defaults to 0.
return_early (bool, optional) – if True and num_threads>0, the method will return a future of the tensordict. The resulting tensordict can be queried using future.result().
share_non_tensor (bool, optional) – if True, the non-tensor data will be shared between the processes and writing operation (such as inplace update or set) on any of the workers within a single node will update the value on all other workers. If the number of non-tensor leaves is high (e.g., sharing large stacks of non-tensor data) this may result in OOM or similar errors. Defaults to False.
existsok (bool, optional) – if False, an exception will be raised if a tensor already exists in the same path. Defaults to True.

The TensorDict is then locked, meaning that any writing operations that isn’t in-place will throw an exception (eg, rename, set or remove an entry). Once the tensordict is unlocked, the memory-mapped attribute is turned to False, because cross-process identity is not guaranteed anymore.

Returns:: self if return_early=False, otherwise a TensorDictFuture instance.

Note

Serialising in this fashion might be slow with deeply nested tensordicts, so it is not recommended to call this method inside a training loop.

memmap_like(prefix: str | None = None, copy_existing: bool = False, *, existsok: bool = True, num_threads: int = 0, return_early: bool = False, share_non_tensor: bool = False) → Self¶

Creates a contentless Memory-mapped tensordict with the same shapes as the original one.

Parameters:

prefix (str) – directory prefix where the memory-mapped tensors will be stored. The directory tree structure will mimic the tensordict’s.
copy_existing (bool) – If False (default), an exception will be raised if an entry in the tensordict is already a tensor stored on disk with an associated file, but is not saved in the correct location according to prefix. If True, any existing Tensor will be copied to the new location.

Keyword Arguments:

num_threads (int, optional) – the number of threads used to write the memmap tensors. Defaults to 0.
return_early (bool, optional) – if True and num_threads>0, the method will return a future of the tensordict.
share_non_tensor (bool, optional) – if True, the non-tensor data will be shared between the processes and writing operation (such as inplace update or set) on any of the workers within a single node will update the value on all other workers. If the number of non-tensor leaves is high (e.g., sharing large stacks of non-tensor data) this may result in OOM or similar errors. Defaults to False.
existsok (bool, optional) – if False, an exception will be raised if a tensor already exists in the same path. Defaults to True.

The TensorDict is then locked, meaning that any writing operations that isn’t in-place will throw an exception (eg, rename, set or remove an entry). Once the tensordict is unlocked, the memory-mapped attribute is turned to False, because cross-process identity is not guaranteed anymore.

Returns:: A new TensorDict instance with data stored as memory-mapped tensors if return_early=False, otherwise a TensorDictFuture instance.

Note

This is the recommended method to write a set of large buffers on disk, as memmap_() will copy the information, which can be slow for large content.

Examples

>>> td = TensorDict({
...     "a": torch.zeros((3, 64, 64), dtype=torch.uint8),
...     "b": torch.zeros(1, dtype=torch.int64),
... }, batch_size=[]).expand(1_000_000)  # expand does not allocate new memory
>>> buffer = td.memmap_like("/path/to/dataset")

memmap_refresh_()¶

Refreshes the content of the memory-mapped tensordict if it has a saved_path.

This method will raise an exception if no path is associated with it.

save(prefix: str | None = None, copy_existing: bool = False, *, num_threads: int = 0, return_early: bool = False, share_non_tensor: bool = False) → Self¶

Saves the tensordict to disk.

This function is a proxy to memmap().

set(key: NestedKey, value: Any, inplace: bool = False, non_blocking: bool = False)¶

Sets a new key-value pair.

Parameters:

key (str, tuple of str) – name of the key to be set. If tuple of str it is equivalent to chained calls of getattr followed by a final setattr.
value (Any) – value to be stored in the tensorclass
inplace (bool, optional) – if True, set will tentatively try to update the value in-place. If False or if the key isn’t present, the value will be simply written at its destination.

Returns:

self

stack(dim: int = 0, *, out=None)¶

Stacks tensordicts into a single tensordict along the given dimension.

This call is equivalent to calling torch.stack() but is compatible with torch.compile.

state_dict(destination=None, prefix='', keep_vars=False, flatten=False) → dict[str, Any]¶: Returns a state_dict dictionary that can be used to save and load data from a tensorclass.

to_tensordict(*, retain_none: bool | None = None) → TensorDict¶

Convert the tensorclass into a regular TensorDict.

Makes a copy of all entries. Memmap and shared memory tensors are converted to regular tensors.

Parameters:: retain_none (bool) – if True, the None values will be written in the tensordict. Otherwise they will be discrarded. Default: True.
Returns:: A new TensorDict object containing the same values as the tensorclass.

unbind(dim: int)¶

Returns a tuple of indexed tensorclass instances unbound along the indicated dimension.

Resulting tensorclass instances will share the storage of the initial tensorclass instance.

History¶

Docs

Tutorials

Resources