torchrl

TorchRL is the open-source Reinforcement Learning (RL) library for PyTorch.

[torchrl @torchrl1]

📢 Introducing TorchRL Data Hub, the most comprehensive single-format data repository for offline RL. With its unified format, swapping or merging datasets is a breeze. It includes: OpenX Embodiment, Roboset, Minari, (V-)D4RL, GenDGRL, AtariDQN, OpenML. https://t.co/XdLBLFVzvN

[vmoens @VincentMoens]

torchrl 0.13 is out!
It provides a lot of deep optimizations including scan/triton-based RNNs with support for intermediate resets (=> padding free) + cuda-based prioritized RBs

[vmoens @VincentMoens]

ultra-fast on-device prioritized sampling over millions of replay items on CUDA, brought to you by @torchrl1

[torchrl @torchrl1]

RNs: https://t.co/cl5C7UFEGP
Doc: https://t.co/qJbbeyLv5H

[torchrl @torchrl1]

🚀 TorchRL 0.9 is here and it's LLM ready!
This major release brings comprehensive LLM fine-tuning capabilities to our RL ecosystem. We're enabling LLM post-training with your favorite RL tools like replay buffers and data collectors.
🧵 Thread below 👇

[torchrl @torchrl1]

📚 What's included in 0.9:
->Unified wrappers for Hugging Face and vLLM models
->Complete GRPO implementation with multi-GPU support
->Specialized objectives: GRPO and SFT loss functions optimized for LLMs
->Comprehensive docs and tutorials

[torchrl @torchrl1]

Check it out here https://t.co/TRZRVcaTGD

[torchrl @torchrl1]

Just released TorchRL 0.8.0, which now incorporates async envs and simplifies extending collectors weight update strategies

[torchrl @torchrl1]

Diff: https://t.co/RcToEnf7rd

[torchrl @torchrl1]

torchrl 🤝 gymnasium happy ever after

With the help of the @FaramaFound team, we managed to make TorchRL compatible with gymnasium v1.1 onward!

[Kosta Derpanis (sabbatical in Zurich) @CSProfKGD]

Collegues in Europe are running this poll about #NeurIPS2025 participation. If in Europe, highly recommended to participate.

[RanW @_RanW_]

Hi imitation learning friends, I am excited to introduce CleanIL (https://t.co/jacC3COKXZ), a repo of high quality single-file implementations of imitation learning and inverse RL algos inspired by CleanRL and built on @torchrl1.

[torchrl @torchrl1]

We're looking for brave OSS contributors to help us scale this up! We need:

- an asyncio api
- ray support
- support for transformed async environments

Please reach out if you're willing to help!

[torchrl @torchrl1]

If anyone is interested in Async environments in torchrl, we just landed that feature in the lib!

[Soumith Chintala @soumithchintala]

A few months ago we quietly open-sourced a @PyTorch video decoding library called torchcodec -- small, nimble, fast.
The @LeRobotHF folks had some nice things to say about it.

[torchrl @torchrl1]

@deependu__ We’re very much alive! No need to worry

[torchrl @torchrl1]

Check out the release notes here: https://t.co/u7DM5zfNQe

[torchrl @torchrl1]

TorchRL 0.7.0 is out!

We brought in torch.compile and with it some considerable speedups!