pytorch/tensordict

Health v2 · maintenance only

Adoption not part of the score

1036 stars · 117 forks observed · 2026-09-03

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

TensorDict is a PyTorch library providing a batched, nested dict-like tensor container where operations like slicing, stacking, and arithmetic propagate to every tensor leaf. It keeps a shared batch_size across nested structures, making it easy to handle heterogeneous tensor data such as RL rollouts or model batches.

Use cases

• manage batches of heterogeneous tensors in pytorch
• store reinforcement learning rollouts with nested observations
• slice and move nested tensor structures to gpu in one call
• stack and reshape batches of tensor dictionaries
• serialize and save collections of tensors with a shared batch size

When to choose

• you work in PyTorch with nested or heterogeneous tensor data like RL transitions
• you want dictionary flexibility with tensor-like batch operations
• you need consistent batch_size enforcement across nested structures

When to avoid

• your project does not use PyTorch
• you only need plain tensors or simple homogeneous batches
• you need a framework-agnostic container

Facets

library · maturity active

machine-learning deep-learning data-science serialization machine-learning deep-learning data-science reinforcement-learning python cross-platform pytorch tensor-container batched-dictionary rl

Member repositories

For agents

markdown · JSON · MCP: product_card(name="pytorch/tensordict")