tensordict-nightly 2026.6.20

TensorDict is a pytorch dedicated tensor container.

TensorDict

TensorDict is a batched, nested dict[str, Tensor] that behaves like a tensor.

Move it, slice it, reshape it, stack it, save it, compile it, or do arithmetic on it: every tensor leaf follows the same operation, and one shared batch_size keeps the structure honest.

TensorDict(batch_size=[32])
|-- obs:      Tensor[32, 128]
|-- action:   Tensor[32]
|-- reward:   Tensor[32]
`-- next:
    `-- obs:  Tensor[32, 128]

30-second demo | Why TensorDict | What is new in 0.13 | Patterns | Installation | Ecosystem | Citation

30-second demo

import torch
from tensordict import TensorDict

batch = TensorDict(
    {
        "obs": torch.randn(32, 128),
        "action": torch.randint(0, 4, (32,)),
        "reward": torch.randn(32),
        "next": {"obs": torch.randn(32, 128)},
    },
    batch_size=[32],
)

mini = batch[:8]                 # slices every leaf
device = "cuda" if torch.cuda.is_available() else "cpu"
on_device = batch.to(device)       # moves every leaf; non-blocking internally
scaled = batch * 0.5             # arithmetic on the whole structure
merged = batch + batch           # leaf-wise TensorDict arithmetic
stacked = torch.stack([batch, batch], 0)

print(mini.shape)                # torch.Size([8])
print(stacked.shape)             # torch.Size([2, 32])

The object remains a mapping, but the batch acts like a tensor. That is the point: write the operation once, apply it to every tensor that belongs to the same example, rollout, batch, parameter set, or dataset shard.

Why TensorDict

Plain dictionaries are flexible. TensorDict keeps that flexibility and adds the parts tensor programs need once the code gets serious.

With a plain dict	With TensorDict
Manually keep leading dimensions aligned	One batch_size validates the structure
Repeat .to(device) for every tensor	td.to(device) moves the full batch
Hand-roll slicing, stacking, reshaping	td[:32], torch.stack, td.reshape
Manually recurse through nested state	Nested keys are first-class
Duplicate arithmetic over leaves	td + td, td * scalar, td.abs()
Invent checkpoint formats	td.save, td.memmap, load_memmap
Hope generic code keeps working	PyTorch-native APIs, torch.compile coverage

Use TensorDict when the unit of data is not one tensor anymore, but it should still move through your program like one tensor.

Performance is part of the API

TensorDict is not just syntax for recursive Python loops. Core paths are built for high-throughput PyTorch workloads:

• Arithmetic dispatch: operations such as td + td, td * 0.5, td.abs() and in-place variants apply directly to leaves and use PyTorch foreach kernels where available.
• Device and host transfers: D2H and H2D copies are dispatched across the full structure. TensorDict uses non-blocking leaf transfers internally when possible, so the common path is just td.to(device); pass non_blocking=False only when you need an explicitly synchronous transfer.
• Shape operations without boilerplate: indexing, view, reshape, permute, unsqueeze, squeeze, flatten, unflatten, stack and cat operate on the batch structure rather than on hand-maintained lists of leaves.
• Low-allocation workflows: lazy stacks, preallocation, memory mapping and inplace=True shape-changing operations help reduce peak memory in data-heavy pipelines.
• Compile-aware internals: TensorDict is used in compiled training and RL loops, and the codebase carries dedicated torch.compile coverage for hot paths.

For deeper numbers, see the benchmark notes.

What is new in 0.13

TensorDict 0.13 focuses on making structured tensor programs more practical in large training systems:

• Tabular import/export for pandas, CSV, Parquet and JSON workflows.
• More inplace=True shape operations, including gather, repeat, repeat_interleave, roll, reshape, flatten, unflatten and contiguous.
• Improved torch.compile behavior for TensorClass initialization, dynamic-shape export, locking paths and shallow clones.
• Safer memmap filenames by default through robust key encoding.
• A migration path for module state preservation with to_module(..., preserve_module_state=...).
• CPU-only release wheels for TensorDict, avoiding duplicate GPU wheel artifacts for a package whose compiled extension is device-independent.

Patterns

One batch through the whole training step

TensorDict lets datasets, models and losses agree on one container instead of a long argument list.

for batch in dataloader:
    batch = batch.to(device)
    batch = model(batch)
    loss = loss_module(batch)

    loss.backward()
    optimizer.step()
    optimizer.zero_grad()

That loop can stay stable while the schema changes from classification to segmentation, RL rollouts, model-based prediction or LLM post-training batches.

Nested data without custom plumbing

td = TensorDict(
    {
        "agents": {
            "policy": torch.randn(64, 8),
            "value": torch.randn(64, 1),
        },
        "env": {
            "reward": torch.randn(64),
            "done": torch.zeros(64, dtype=torch.bool),
        },
    },
    batch_size=[64],
)

policy = td["agents", "policy"]
td["env", "reward"] = td["env", "reward"].clip(-1, 1)

Nested keys are part of the API, not an afterthought.

Functional modules and parameter sets

TensorDict can hold module parameters, swap them into modules, vectorize over ensembles and make model state explicit.

from tensordict import TensorDict

params = TensorDict.from_module(module)

with params.to_module(module, preserve_module_state=True):
    out = module(inputs)

This is the same foundation used by TorchRL modules and functional training utilities.

Checkpoint and share large tensor batches

td = TensorDict({"tokens": tokens, "scores": scores}, batch_size=[n])
td.memmap("/tmp/batch")          # memory-map every leaf
reloaded = TensorDict.load_memmap("/tmp/batch")

Memory-mapped TensorDicts are useful for large offline datasets, replay buffers, inter-process handoff and checkpointed intermediate state.

Key features

• Tensor-like collection ops: indexing, slicing, device casting, dtype casting, reshaping, stacking and concatenation. [tutorial]
• Nested structures with tuple keys and predictable batch semantics. [tutorial]
• Fast memory workflows: asynchronous transfers, memmap, consolidated tensors, lazy stacks and preallocation. [tutorial]
• Functional programming with parameter TensorDicts, to_module and compatibility with torch.vmap. [tutorial]
• @tensorclass: a tensor-aware dataclass for structured tensor objects. [tutorial]
• Distributed and multiprocessed pipelines across workers, devices and machines. [doc]
• Serialization and memory mapping for efficient checkpointing and dataset storage. [doc]

For a longer tour, start with GETTING_STARTED.md or the online documentation.

Installation

With pip:

pip install tensordict

With conda:

conda install -c conda-forge tensordict

Nightly builds:

pip install tensordict-nightly

From source with an existing PyTorch install:

pip install -e . --no-deps

If you use uv with PyTorch nightlies, keep torch pinned to the PyTorch wheel index or install TensorDict with --no-deps so the resolver does not replace your existing PyTorch build:

uv pip install -e . --no-deps
uv pip install -e . --prerelease=allow -f "https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html"

Ecosystem

TensorDict started in reinforcement learning, where batches quickly become nested trajectories. It is now used anywhere tensor batches are structured data: RL rollouts, LLM post-training samples, robotics trajectories, simulation state, model parameters, checkpointed datasets and scientific pipelines.

Domain	Projects
Reinforcement Learning	TorchRL (PyTorch), DreamerV3-torch, Dreamer4, SkyRL
LLM Post-Training	verl, ROLL (Alibaba), LMFlow, LoongFlow (Baidu)
Robotics and Simulation	MuJoCo Playground (Google DeepMind), ProtoMotions (NVIDIA), holosoma (Amazon)
Physics and Scientific ML	PhysicsNeMo (NVIDIA)
Genomics	Medaka (Oxford Nanopore)

Citation

If you use TensorDict, please cite the TorchRL paper:

@misc{bou2023torchrl,
      title={TorchRL: A data-driven decision-making library for PyTorch},
      author={Albert Bou and Matteo Bettini and Sebastian Dittert and Vikash Kumar and Shagun Sodhani and Xiaomeng Yang and Gianni De Fabritiis and Vincent Moens},
      year={2023},
      eprint={2306.00577},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

License

TensorDict is licensed under the MIT License. See LICENSE for details.