TorchRL

Documentation | TensorDict | Features | Examples, tutorials and demos | Citation | Installation | Asking a question | Contributing

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

🚀 What's New

LLM API - Complete Framework for Language Model Fine-tuning

TorchRL now includes a comprehensive LLM API for post-training and fine-tuning of language models! This new framework provides everything you need for RLHF, supervised fine-tuning, and tool-augmented training:

• 🤖 Unified LLM Wrappers: Seamless integration with Hugging Face models and vLLM inference engines - more to come!
• 💬 Conversation Management: Advanced History class for multi-turn dialogue with automatic chat template detection
• 🛠️ Tool Integration: Built-in support for Python code execution, function calling, and custom tool transforms
• 🎯 Specialized Objectives: GRPO (Group Relative Policy Optimization) and SFT loss functions optimized for language models
• ⚡ High-Performance Collectors: Async data collection with distributed training support
• 🔄 Flexible Environments: Transform-based architecture for reward computation, data loading, and conversation augmentation

The LLM API follows TorchRL's modular design principles, allowing you to mix and match components for your specific use case. Check out the complete documentation and GRPO implementation example to get started!

python
1from torchrl.envs.llm import ChatEnv
2from torchrl.modules.llm import TransformersWrapper
3from torchrl.objectives.llm import GRPOLoss
4from torchrl.collectors.llm import LLMCollector
5
6# Create environment with Python tool execution
7env = ChatEnv(
8    tokenizer=tokenizer,
9    system_prompt="You are an assistant that can execute Python code.",
10    batch_size=[1]
11).append_transform(PythonInterpreter())
12
13# Wrap your language model
14llm = TransformersWrapper(
15    model=model,
16    tokenizer=tokenizer,
17    input_mode="history"
18)
19
20# Set up GRPO training
21loss_fn = GRPOLoss(llm, critic, gamma=0.99)
22collector = LLMCollector(env, llm, frames_per_batch=100)
23
24# Training loop
25for data in collector:
26    loss = loss_fn(data)
27    loss.backward()
28    optimizer.step()

Key features

• 🐍 Python-first: Designed with Python as the primary language for ease of use and flexibility
• ⏱️ Efficient: Optimized for performance to support demanding RL research applications
• 🧮 Modular, customizable, extensible: Highly modular architecture allows for easy swapping, transformation, or creation of new components
• 📚 Documented: Thorough documentation ensures that users can quickly understand and utilize the library
• ✅ Tested: Rigorously tested to ensure reliability and stability
• ⚙️ Reusable functionals: Provides a set of highly reusable functions for cost functions, returns, and data processing

Design Principles

• 🔥 Aligns with PyTorch ecosystem: Follows the structure and conventions of popular PyTorch libraries (e.g., dataset pillar, transforms, models, data utilities)
• ➖ Minimal dependencies: Only requires Python standard library, NumPy, and PyTorch; optional dependencies for common environment libraries (e.g., OpenAI Gym) and datasets (D4RL, OpenX...)

Read the full paper for a more curated description of the library.

Getting started

Check our Getting Started tutorials for quickly ramp up with the basic features of the library!

Documentation and knowledge base

The TorchRL documentation can be found here. It contains tutorials and the API reference.

TorchRL also provides a RL knowledge base to help you debug your code, or simply learn the basics of RL. Check it out here.

We have some introductory videos for you to get to know the library better, check them out:

• TalkRL podcast
• TorchRL intro at PyTorch day 2022
• PyTorch 2.0 Q&A: TorchRL

Spotlight publications

TorchRL being domain-agnostic, you can use it across many different fields. Here are a few examples:

• ACEGEN: Reinforcement Learning of Generative Chemical Agents for Drug Discovery
• BenchMARL: Benchmarking Multi-Agent Reinforcement Learning
• BricksRL: A Platform for Democratizing Robotics and Reinforcement Learning Research and Education with LEGO
• OmniDrones: An Efficient and Flexible Platform for Reinforcement Learning in Drone Control
• RL4CO: an Extensive Reinforcement Learning for Combinatorial Optimization Benchmark
• Robohive: A unified framework for robot learning

Writing simplified and portable RL codebase with TensorDict

RL algorithms are very heterogeneous, and it can be hard to recycle a codebase across settings (e.g. from online to offline, from state-based to pixel-based learning). TorchRL solves this problem through TensorDict, a convenient data structure⁽¹⁾ that can be used to streamline one's RL codebase. With this tool, one can write a complete PPO training script in less than 100 lines of code!

python
1import torch
2from tensordict.nn import TensorDictModule
3from tensordict.nn.distributions import NormalParamExtractor
4from torch import nn
5
6from torchrl.collectors import SyncDataCollector
7from torchrl.data.replay_buffers import TensorDictReplayBuffer, \
8  LazyTensorStorage, SamplerWithoutReplacement
9from torchrl.envs.libs.gym import GymEnv
10from torchrl.modules import ProbabilisticActor, ValueOperator, TanhNormal
11from torchrl.objectives import ClipPPOLoss
12from torchrl.objectives.value import GAE
13
14env = GymEnv("Pendulum-v1") 
15model = TensorDictModule(
16  nn.Sequential(
17      nn.Linear(3, 128), nn.Tanh(),
18      nn.Linear(128, 128), nn.Tanh(),
19      nn.Linear(128, 128), nn.Tanh(),
20      nn.Linear(128, 2),
21      NormalParamExtractor()
22  ),
23  in_keys=["observation"],
24  out_keys=["loc", "scale"]
25)
26critic = ValueOperator(
27  nn.Sequential(
28      nn.Linear(3, 128), nn.Tanh(),
29      nn.Linear(128, 128), nn.Tanh(),
30      nn.Linear(128, 128), nn.Tanh(),
31      nn.Linear(128, 1),
32  ),
33  in_keys=["observation"],
34)
35actor = ProbabilisticActor(
36  model,
37  in_keys=["loc", "scale"],
38  distribution_class=TanhNormal,
39  distribution_kwargs={"low": -1.0, "high": 1.0},
40  return_log_prob=True
41  )
42buffer = TensorDictReplayBuffer(
43  storage=LazyTensorStorage(1000),
44  sampler=SamplerWithoutReplacement(),
45  batch_size=50,
46  )
47collector = SyncDataCollector(
48  env,
49  actor,
50  frames_per_batch=1000,
51  total_frames=1_000_000,
52)
53loss_fn = ClipPPOLoss(actor, critic)
54adv_fn = GAE(value_network=critic, average_gae=True, gamma=0.99, lmbda=0.95)
55optim = torch.optim.Adam(loss_fn.parameters(), lr=2e-4)
56
57for data in collector:  # collect data
58  for epoch in range(10):
59      adv_fn(data)  # compute advantage
60      buffer.extend(data)
61      for sample in buffer:  # consume data
62          loss_vals = loss_fn(sample)
63          loss_val = sum(
64              value for key, value in loss_vals.items() if
65              key.startswith("loss")
66              )
67          loss_val.backward()
68          optim.step()
69          optim.zero_grad()
70  print(f"avg reward: {data['next', 'reward'].mean().item(): 4.4f}")

Here is an example of how the environment API relies on tensordict to carry data from one function to another during a rollout execution:

TensorDict makes it easy to re-use pieces of code across environments, models and algorithms.

diff
1- obs, done = env.reset()
2+ tensordict = env.reset()
3policy = SafeModule(
4    model,
5    in_keys=["observation_pixels", "observation_vector"],
6    out_keys=["action"],
7)
8out = []
9for i in range(n_steps):
10-     action, log_prob = policy(obs)
11-     next_obs, reward, done, info = env.step(action)
12-     out.append((obs, next_obs, action, log_prob, reward, done))
13-     obs = next_obs
14+     tensordict = policy(tensordict)
15+     tensordict = env.step(tensordict)
16+     out.append(tensordict)
17+     tensordict = step_mdp(tensordict)  # renames next_observation_* keys to observation_*
18- obs, next_obs, action, log_prob, reward, done = [torch.stack(vals, 0) for vals in zip(*out)]
19+ out = torch.stack(out, 0)  # TensorDict supports multiple tensor operations

Using this, TorchRL abstracts away the input / output signatures of the modules, env, collectors, replay buffers and losses of the library, allowing all primitives to be easily recycled across settings.

diff
1- for i, (obs, next_obs, action, hidden_state, reward, done) in enumerate(collector):
2+ for i, tensordict in enumerate(collector):
3-     replay_buffer.add((obs, next_obs, action, log_prob, reward, done))
4+     replay_buffer.add(tensordict)
5    for j in range(num_optim_steps):
6-         obs, next_obs, action, hidden_state, reward, done = replay_buffer.sample(batch_size)
7-         loss = loss_fn(obs, next_obs, action, hidden_state, reward, done)
8+         tensordict = replay_buffer.sample(batch_size)
9+         loss = loss_fn(tensordict)
10        loss.backward()
11        optim.step()
12        optim.zero_grad()

TensorDict supports multiple tensor operations on its device and shape (the shape of TensorDict, or its batch size, is the common arbitrary N first dimensions of all its contained tensors):

python
1# stack and cat
2tensordict = torch.stack(list_of_tensordicts, 0)
3tensordict = torch.cat(list_of_tensordicts, 0)
4# reshape
5tensordict = tensordict.view(-1)
6tensordict = tensordict.permute(0, 2, 1)
7tensordict = tensordict.unsqueeze(-1)
8tensordict = tensordict.squeeze(-1)
9# indexing
10tensordict = tensordict[:2]
11tensordict[:, 2] = sub_tensordict
12# device and memory location
13tensordict.cuda()
14tensordict.to("cuda:1")
15tensordict.share_memory_()

TensorDict comes with a dedicated tensordict.nn module that contains everything you might need to write your model with it. And it is functorch and torch.compile compatible!

diff
1transformer_model = nn.Transformer(nhead=16, num_encoder_layers=12)
2+ td_module = SafeModule(transformer_model, in_keys=["src", "tgt"], out_keys=["out"])
3src = torch.rand((10, 32, 512))
4tgt = torch.rand((20, 32, 512))
5+ tensordict = TensorDict({"src": src, "tgt": tgt}, batch_size=[20, 32])
6- out = transformer_model(src, tgt)
7+ td_module(tensordict)
8+ out = tensordict["out"]

python
1encoder_module = TransformerEncoder(...)
2encoder = TensorDictSequential(encoder_module, in_keys=["src", "src_mask"], out_keys=["memory"])
3decoder_module = TransformerDecoder(...)
4decoder = TensorDictModule(decoder_module, in_keys=["tgt", "memory"], out_keys=["output"])
5transformer = TensorDictSequential(encoder, decoder)
6assert transformer.in_keys == ["src", "src_mask", "tgt"]
7assert transformer.out_keys == ["memory", "output"]

python
1transformer.select_subsequence(out_keys=["memory"])  # returns the encoder
2transformer.select_subsequence(in_keys=["tgt", "memory"])  # returns the decoder

Check TensorDict tutorials to learn more!

Features

• A common interface for environments which supports common libraries (OpenAI gym, deepmind control lab, etc.)⁽¹⁾ and state-less execution (e.g. Model-based environments). The batched environments containers allow parallel execution⁽²⁾. A common PyTorch-first class of tensor-specification class is also provided. TorchRL's environments API is simple but stringent and specific. Check the documentation and tutorial to learn more!

  Code

  python
  1env_make = lambda: GymEnv("Pendulum-v1", from_pixels=True)
  2env_parallel = ParallelEnv(4, env_make)  # creates 4 envs in parallel
  3tensordict = env_parallel.rollout(max_steps=20, policy=None)  # random rollout (no policy given)
  4assert tensordict.shape == [4, 20]  # 4 envs, 20 steps rollout
  5env_parallel.action_spec.is_in(tensordict["action"])  # spec check returns True

• multiprocess and distributed data collectors⁽²⁾ that work synchronously or asynchronously. Through the use of TensorDict, TorchRL's training loops are made very similar to regular training loops in supervised learning (although the "dataloader" -- read data collector -- is modified on-the-fly):

  Code

  python
  1env_make = lambda: GymEnv("Pendulum-v1", from_pixels=True)
  2collector = MultiaSyncDataCollector(
  3    [env_make, env_make],
  4    policy=policy,
  5    devices=["cuda:0", "cuda:0"],
  6    total_frames=10000,
  7    frames_per_batch=50,
  8    ...
  9)
  10for i, tensordict_data in enumerate(collector):
  11    loss = loss_module(tensordict_data)
  12    loss.backward()
  13    optim.step()
  14    optim.zero_grad()
  15    collector.update_policy_weights_()

  Check our distributed collector examples to learn more about ultra-fast data collection with TorchRL.

• efficient⁽²⁾ and generic⁽¹⁾ replay buffers with modularized storage:

  Code

  python
  1storage = LazyMemmapStorage(  # memory-mapped (physical) storage
  2    cfg.buffer_size,
  3    scratch_dir="/tmp/"
  4)
  5buffer = TensorDictPrioritizedReplayBuffer(
  6    alpha=0.7,
  7    beta=0.5,
  8    collate_fn=lambda x: x,
  9    pin_memory=device != torch.device("cpu"),
  10    prefetch=10,  # multi-threaded sampling
  11    storage=storage
  12)

  Replay buffers are also offered as wrappers around common datasets for offline RL:

  Code

  python
  1from torchrl.data.replay_buffers import SamplerWithoutReplacement
  2from torchrl.data.datasets.d4rl import D4RLExperienceReplay
  3data = D4RLExperienceReplay(
  4    "maze2d-open-v0",
  5    split_trajs=True,
  6    batch_size=128,
  7    sampler=SamplerWithoutReplacement(drop_last=True),
  8)
  9for sample in data:  # or alternatively sample = data.sample()
  10    fun(sample)

• cross-library environment transforms⁽¹⁾, executed on device and in a vectorized fashion⁽²⁾, which process and prepare the data coming out of the environments to be used by the agent:

  Code

  python
  1env_make = lambda: GymEnv("Pendulum-v1", from_pixels=True)
  2env_base = ParallelEnv(4, env_make, device="cuda:0")  # creates 4 envs in parallel
  3env = TransformedEnv(
  4    env_base,
  5    Compose(
  6        ToTensorImage(),
  7        ObservationNorm(loc=0.5, scale=1.0)),  # executes the transforms once and on device
  8)
  9tensordict = env.reset()
  10assert tensordict.device == torch.device("cuda:0")

  Other transforms include: reward scaling (RewardScaling), shape operations (concatenation of tensors, unsqueezing etc.), concatenation of successive operations (CatFrames), resizing (Resize) and many more.

  Unlike other libraries, the transforms are stacked as a list (and not wrapped in each other), which makes it easy to add and remove them at will:

  python
  1env.insert_transform(0, NoopResetEnv())  # inserts the NoopResetEnv transform at the index 0

  Nevertheless, transforms can access and execute operations on the parent environment:

  python
  1transform = env.transform[1]  # gathers the second transform of the list
  2parent_env = transform.parent  # returns the base environment of the second transform, i.e. the base env + the first transform

• various tools for distributed learning (e.g. memory mapped tensors)⁽²⁾;

• various architectures and models (e.g. actor-critic)⁽¹⁾:

  Code

  python
  1# create an nn.Module
  2common_module = ConvNet(
  3    bias_last_layer=True,
  4    depth=None,
  5    num_cells=[32, 64, 64],
  6    kernel_sizes=[8, 4, 3],
  7    strides=[4, 2, 1],
  8)
  9# Wrap it in a SafeModule, indicating what key to read in and where to
  10# write out the output
  11common_module = SafeModule(
  12    common_module,
  13    in_keys=["pixels"],
  14    out_keys=["hidden"],
  15)
  16# Wrap the policy module in NormalParamsWrapper, such that the output
  17# tensor is split in loc and scale, and scale is mapped onto a positive space
  18policy_module = SafeModule(
  19    NormalParamsWrapper(
  20        MLP(num_cells=[64, 64], out_features=32, activation=nn.ELU)
  21    ),
  22    in_keys=["hidden"],
  23    out_keys=["loc", "scale"],
  24)
  25# Use a SafeProbabilisticTensorDictSequential to combine the SafeModule with a
  26# SafeProbabilisticModule, indicating how to build the
  27# torch.distribution.Distribution object and what to do with it
  28policy_module = SafeProbabilisticTensorDictSequential(  # stochastic policy
  29    policy_module,
  30    SafeProbabilisticModule(
  31        in_keys=["loc", "scale"],
  32        out_keys="action",
  33        distribution_class=TanhNormal,
  34    ),
  35)
  36value_module = MLP(
  37    num_cells=[64, 64],
  38    out_features=1,
  39    activation=nn.ELU,
  40)
  41# Wrap the policy and value funciton in a common module
  42actor_value = ActorValueOperator(common_module, policy_module, value_module)
  43# standalone policy from this
  44standalone_policy = actor_value.get_policy_operator()

• exploration wrappers and modules to easily swap between exploration and exploitation⁽¹⁾:

  Code

  python
  1policy_explore = EGreedyWrapper(policy)
  2with set_exploration_type(ExplorationType.RANDOM):
  3    tensordict = policy_explore(tensordict)  # will use eps-greedy
  4with set_exploration_type(ExplorationType.DETERMINISTIC):
  5    tensordict = policy_explore(tensordict)  # will not use eps-greedy

• A series of efficient loss modules and highly vectorized functional return and advantage computation.

  Code

  Loss modules

  python
  1from torchrl.objectives import DQNLoss
  2loss_module = DQNLoss(value_network=value_network, gamma=0.99)
  3tensordict = replay_buffer.sample(batch_size)
  4loss = loss_module(tensordict)

  Advantage computation

  python
  1from torchrl.objectives.value.functional import vec_td_lambda_return_estimate
  2advantage = vec_td_lambda_return_estimate(gamma, lmbda, next_state_value, reward, done, terminated)

• a generic trainer class⁽¹⁾ that executes the aforementioned training loop. Through a hooking mechanism, it also supports any logging or data transformation operation at any given time.

• various recipes to build models that correspond to the environment being deployed.

• LLM API: Complete framework for language model fine-tuning with unified wrappers for Hugging Face and vLLM backends, conversation management with automatic chat template detection, tool integration (Python execution, function calling), specialized objectives (GRPO, SFT), and high-performance async collectors. Perfect for RLHF, supervised fine-tuning, and tool-augmented training scenarios.

  Code

  python
  1from torchrl.envs.llm import ChatEnv
  2from torchrl.modules.llm import TransformersWrapper
  3from torchrl.envs.llm.transforms import PythonInterpreter
  4
  5# Create environment with tool execution
  6env = ChatEnv(
  7    tokenizer=tokenizer,
  8    system_prompt="You can execute Python code.",
  9    batch_size=[1]
  10).append_transform(PythonInterpreter())
  11
  12# Wrap language model for training
  13llm = TransformersWrapper(
  14    model=model,
  15    tokenizer=tokenizer,
  16    input_mode="history"
  17)
  18
  19# Multi-turn conversation with tool use
  20obs = env.reset(TensorDict({"query": "Calculate 2+2"}, batch_size=[1]))
  21llm_output = llm(obs)  # Generates response
  22obs = env.step(llm_output)  # Environment processes response

If you feel a feature is missing from the library, please submit an issue! If you would like to contribute to new features, check our call for contributions and our contribution page.

Examples, tutorials and demos

A series of State-of-the-Art implementations are provided with an illustrative purpose:

** The number indicates expected speed-up compared to eager mode when executed on CPU. Numbers may vary depending on architecture and device.

and many more to come!

Code examples displaying toy code snippets and training scripts are also available

• LLM API & GRPO - Complete language model fine-tuning pipeline
• RLHF
• Memory-mapped replay buffers

Check the examples directory for more details about handling the various configuration settings.

We also provide tutorials and demos that give a sense of what the library can do.

Citation

If you're using TorchRL, please refer to this BibTeX entry to cite this work:

1@misc{bou2023torchrl,
2      title={TorchRL: A data-driven decision-making library for PyTorch}, 
3      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},
4      year={2023},
5      eprint={2306.00577},
6      archivePrefix={arXiv},
7      primaryClass={cs.LG}
8}

Installation

Create a new virtual environment:

bash
1python -m venv torchrl
2source torchrl/bin/activate  # On Windows use: venv\Scripts\activate

Or create a conda environment where the packages will be installed.

1conda create --name torchrl python=3.9
2conda activate torchrl

Install dependencies:

PyTorch

Depending on the use of torchrl that you want to make, you may want to install the latest (nightly) PyTorch release or the latest stable version of PyTorch. See here for a detailed list of commands, including pip3 or other special installation instructions.

TorchRL offers a few pre-defined dependencies such as "torchrl[tests]", "torchrl[atari]" etc.

Torchrl

You can install the latest stable release by using

bash
1pip3 install torchrl

This should work on linux (including AArch64 machines), Windows 10 and OsX (Metal chips only). On certain Windows machines (Windows 11), one should build the library locally. This can be done in two ways:

bash
1# Install and build locally v0.8.1 of the library without cloning
2pip3 install git+https://github.com/pytorch/[email protected]
3# Clone the library and build it locally
4git clone https://github.com/pytorch/tensordict
5git clone https://github.com/pytorch/rl
6pip install -e tensordict
7pip install -e rl

Note that tensordict local build requires cmake to be installed via homebrew (MacOS) or another package manager such as apt, apt-get, conda or yum but NOT pip, as well as pip install "pybind11[global]".

One can also build the wheels to distribute to co-workers using

bash
1python setup.py bdist_wheel

Your wheels will be stored there ./dist/torchrl<name>.whl and installable via

bash
1pip install torchrl<name>.whl

The nightly build can be installed via

bash
1pip3 install tensordict-nightly torchrl-nightly

which we currently only ship for Linux machines. Importantly, the nightly builds require the nightly builds of PyTorch too. Also, a local build of torchrl with the nightly build of tensordict may fail - install both nightlies or both local builds but do not mix them.

Disclaimer: As of today, TorchRL is roughly compatible with any pytorch version >= 2.1 and installing it will not directly require a newer version of pytorch to be installed. Indirectly though, tensordict still requires the latest PyTorch to be installed and we are working hard to loosen that requirement. The C++ binaries of TorchRL (mainly for prioritized replay buffers) will only work with PyTorch 2.7.0 and above. Some features (e.g., working with nested jagged tensors) may also be limited with older versions of pytorch. It is recommended to use the latest TorchRL with the latest PyTorch version unless there is a strong reason not to do so.

Optional dependencies

The following libraries can be installed depending on the usage one wants to make of torchrl:

1# diverse
2pip3 install tqdm tensorboard "hydra-core>=1.1" hydra-submitit-launcher
3
4# rendering
5pip3 install "moviepy<2.0.0"
6
7# deepmind control suite
8pip3 install dm_control
9
10# gym, atari games
11pip3 install "gym[atari]" "gym[accept-rom-license]" pygame
12
13# tests
14pip3 install pytest pyyaml pytest-instafail
15
16# tensorboard
17pip3 install tensorboard
18
19# wandb
20pip3 install wandb

Versioning issues can cause error message of the type undefined symbol and such. For these, refer to the versioning issues document for a complete explanation and proposed workarounds.

Asking a question

If you spot a bug in the library, please raise an issue in this repo.

If you have a more generic question regarding RL in PyTorch, post it on the PyTorch forum.

Contributing

Internal collaborations to torchrl are welcome! Feel free to fork, submit issues and PRs. You can checkout the detailed contribution guide here. As mentioned above, a list of open contributions can be found in here.

Contributors are recommended to install pre-commit hooks (using pre-commit install). pre-commit will check for linting related issues when the code is committed locally. You can disable th check by appending -n to your commit command: git commit -m <commit message> -n

Disclaimer

This library is released as a PyTorch beta feature. BC-breaking changes are likely to happen but they will be introduced with a deprecation warranty after a few release cycles.

License

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