Registry / ai-ml / tensordict-nightly

tensordict-nightly

library2026.6.14pypypi✓ verified 84d ago

TensorDict is a PyTorch-dedicated tensor container that provides a dictionary-like class inheriting properties from `torch.Tensor`. It streamlines the organization and manipulation of collections of tensors, enabling efficient batch operations, shape transformations, and seamless device management. As a nightly build, `tensordict-nightly` offers the latest features and bug fixes, with frequent updates that may introduce breaking changes.

ai-ml data

pip install tensordict-nightly

github ↗package ↗

INSTALL

IMPORT

SIG · TENSORDICT-NIGHTLY

tensordict-nightly

ai-mlpythonv2026.6.14

Install

67.9s avg

Import

11566ms

Disk

4838MB

Pass rate

4/ 10

Env Coverage4 / 10

glibc

3.9–3.13

musl

3.9–3.13

Install & Compatibility

Where this runs

tested against v2026.6.7 · pip install

no network on importno background threads

Install × environment matrix

Each cell = how many times install + import succeeded across repeated harness runs. Partial = flaky.
glibc = Debian/Ubuntu slim · musl = Alpine Linux

musl

glibc

py 3.10

4/8 runs

✓ 79.2s

py 3.11

4/8 runs

✓ 73.08s

py 3.12

4/8 runs

✓ 61.83s

py 3.13

4/8 runs

✓ 57.65s

py 3.9

4/8 runs

4838MB installed

● package 4838MB

Code

Verified usage

Verified import paths — ran on the pinned version, not inferred.

TensorDict

✓ from tensordict import TensorDict

MemoryMappedTensor

✓ from tensordict import MemoryMappedTensor

✗ my_memmap_tensor._tensor

The `_tensor` property on `MemoryMappedTensor` was removed in v0.11.0 and now raises a RuntimeError. Access the instance directly.

TensorDictModule

✓ from tensordict.nn import TensorDictModule

tensorclass

✓ from tensordict import tensorclass

This quickstart demonstrates how to create a `TensorDict`, access and add elements, move it to a different device (if CUDA is available), and perform basic slicing operations.

import torch from tensordict import TensorDict # Create a TensorDict with a specified batch_size td = TensorDict( {"observations": torch.randn(128, 84), "actions": torch.randn(128, 4)}, batch_size=[128] ) print("Original TensorDict:\n", td) print("Batch size:", td.batch_size) # Accessing elements obs = td["observations"] print("\nObservations shape:", obs.shape) # Adding a new key td["rewards"] = torch.randn(128, 1) print("\nTensorDict after adding rewards:\n", td) # Moving to device if torch.cuda.is_available(): td_gpu = td.to("cuda") print(f"\nTensorDict moved to {td_gpu.device}:\n", td_gpu) # Slicing sub_td = td[:64] print("\nSliced TensorDict (first 64 elements):\n", sub_td) print("Sliced batch size:", sub_td.batch_size)

Debug

Known issues

breakingPython 3.9 support was dropped in TensorDict v0.11.0. Python 3.10 or newer is now required.

fix

Upgrade your Python environment to version 3.10 or higher.

affects: >=0.11.0

breakingDeprecated methods `lock`, `unlock`, and `rename_key` (without a trailing underscore) were removed in v0.11.0.

fix

Use `lock_`, `unlock_`, and `rename_key_` instead for in-place modifications, e.g., `td.lock_()` instead of `td.lock()`.

affects: >=0.11.0

breakingThe `MemoryMappedTensor._tensor` property now raises a `RuntimeError` since v0.11.0. Users should interact with the `MemoryMappedTensor` instance directly as it is a tensor subclass.

fix

Access the tensor content directly, e.g., `my_memmap_tensor` instead of `my_memmap_tensor._tensor`.

affects: >=0.11.0

gotchaFrom v0.10.0, lists assigned to a TensorDict will be automatically stacked by default, potentially raising a `FutureWarning`.

fix

Review code that assigns lists to TensorDicts. If list stacking is not the desired default behavior or to suppress the warning, explicitly use context managers like `td.set_` or specify the desired behavior.

affects: >=0.10.0

gotchaCalling `.to()` method on a `TensorDict` with a `dtype` argument will raise an error. The `to()` method is for device casting only.

fix

To change the dtype of tensors within a TensorDict, iterate through its items and apply `item.to(dtype=...)` or use `td.apply(lambda x: x.to(dtype=...))`.

affects: All versions

Errors

Common errors & fixes

IndexError: tuple index out of range

Passing a dictionary with non-string keys (e.g., integers, tuples) to the `TensorDict` constructor or `make_tensordict` function.

fix

Ensure all keys in the input dictionary are strings. If you need nested keys, use string keys or consider using `flatten_keys` later with a separator.

RuntimeError: Cannot modify locked TensorDict.

Attempting to modify a `TensorDict` instance that has been locked (e.g., by calling `td.lock_()` or being created with `lock=True`).

fix

Unlock the TensorDict using `td.unlock_()` before modification, or use in-place methods with a trailing underscore (e.g., `td.set_(key, value)`) if the key already exists.

Wrong values in TensorDict with device='cpu' specified

When creating or moving a `TensorDict` to CPU with `non_blocking=True` (which is often implicit or default), data transfer might not be synchronized, leading to incorrect values if accessed immediately. This is more prevalent with non-CUDA devices.

fix

Explicitly set `non_blocking=False` when moving to CPU if immediate and synchronized access is critical, or ensure a synchronization call (`torch.cuda.synchronize()` if applicable) is made before accessing the data.

_dist_sample hasattr error OR load_state_dict fails if checkpoint lacks entries for TensorDictParams

These are general issues often related to specific versions of PyTorch or `tensordict`, or complex model architectures and distributed setups where `TensorDict` or `tensorclass` objects are used in state dictionaries or sampling processes.

fix

Check the `pytorch/tensordict` GitHub issues for similar reports and potential workarounds or targeted bug fixes for your specific `tensordict` and PyTorch versions. Consider updating to the latest nightly builds for potential fixes.

Upgrade

Version history

2026.6.14latest on PyPI · released Jun 14, 2026

Audit

Dependencies

torchrequiredCore dependency, TensorDict is built for PyTorch tensors.

numpyoptionalRequired for certain operations or data handling.

cloudpickleoptionalUsed for serialization, especially in multiprocessing/distributed settings.

packagingoptionalUsed for version parsing and compatibility checks.

importlib_metadataoptionalBackward compatibility for Python versions prior to 3.8.

orjsonoptionalOptional dependency for faster JSON serialization, primarily for Python < 3.13.

Agent activity

14 hits · last 30 days

node

OpenAI (training)

Resources

githubhttps://github.com/pytorch/tensordict ↗

packagehttps://pypi.org/project/tensordict-nightly/ ↗