3 Releases published by 1 person

• v0.9.0 TorchRL 0.9.0 Release Notes

  published Jul 10, 2025

• v0.9.1 v0.9.1: fix for history-based vLLM and Transformers wrappers

  published Jul 11, 2025

• v0.9.2 v0.9.2: Bug fixes and perf improvements

  published Jul 17, 2025

57 Pull requests merged by 12 people

• [BE] Strong limit on warning appearance for failing C++ binaries

  #3115 merged Aug 6, 2025

• [BugFix] Fix parameter conflict resolution and validation tests in Wrappers

  #3114 merged Aug 6, 2025

• [Doc] fix typo in dqn module

  #3111 merged Aug 5, 2025

• [BE] Catching common errors in env.rollout and rb.add

  #3102 merged Aug 1, 2025

• [Algorithm] SOTA discrete offline CQL

  #3098 merged Aug 1, 2025

• [Formatting] Update pre-commit

  #3108 merged Aug 1, 2025

• [Feature,BugFix] Fix composite entropy for nested keys

  #3101 merged Aug 1, 2025

• [Feature] Common kwargs for generate across vLLM and Transformers

  #3107 merged Aug 1, 2025

• [Feature] Better batching in LLM Wrappers

  #3104 merged Jul 31, 2025

• [Feature] batching for vllm and transformers wrappers

  #3103 merged Jul 30, 2025

• [Versioning] Better usage of pyproject.toml

  #3089 merged Jul 29, 2025

• [Feature] Packing

  #3060 merged Jul 29, 2025

• [Versioning] Bump tensordict version

  #3100 merged Jul 28, 2025

• [BE] Avoid warning users who don't care about PRB

  #3099 merged Jul 28, 2025

• [BugFix] Patching not applied to NonTensorData observations, like Atari's

  #3091 merged Jul 28, 2025

• [BE] Avoid duplicated file names in test/ to make pytest happy

  #3096 merged Jul 28, 2025

• [Examples, Docs] Compressed rb nvcomp examples

  #3095 merged Jul 26, 2025

• [Versioning] Bump 0.10.0

  #3094 merged Jul 25, 2025

• Fix benchmark CI upload with conditional PR testing

  #3093 merged Jul 25, 2025

• [Feature] Compressed storage gpu

  #3062 merged Jul 25, 2025

• [Feature] Add support for loading datasets from local Minari cache

  #3068 merged Jul 24, 2025

• [CI] Fix benchmark runs

  #3092 merged Jul 24, 2025

• [BugFix] remove storage from ray init params

  #3090 merged Jul 24, 2025

• Re-enable CPU Pytest benchmark job

  #3048 merged Jul 23, 2025

• [BugFix] Minor fixes to wandb logger

  #2999 merged Jul 23, 2025

• [BugFix] Fix Qwen parsing for tools

  #3087 merged Jul 23, 2025

• [Quality] Readability improvements in value estimation

  #3088 merged Jul 22, 2025

• [Feature] public methods for SAC

  #3085 merged Jul 21, 2025

• [BE] Include PyTorch version in message for PRB import error

  #3086 merged Jul 20, 2025

• [BugFix] Fix Binary reshaping

  #3084 merged Jul 18, 2025

• [Doc] Add missing docstrings in History

  #3083 merged Jul 17, 2025

• [BugFix] Add reference to policy with state dict

  #3043 merged Jul 17, 2025

• [Performance] Avoid computing log-probs when retrieving dist

  #3081 merged Jul 16, 2025

• [BugFix] More robust handling of unbatched data in llm wrappers

  #3070 merged Jul 16, 2025

• [BugFix] Fix cross-entropy log-prob for batched input

  #3080 merged Jul 16, 2025

• [Quality] Uniform coeff suffix in A2C and PPO

  #3079 merged Jul 16, 2025

• [BugFix] Fix Binary clone() ops

  #3077 merged Jul 16, 2025

• [BugFix] Fix in-place modification of specs in TransformedEnv

  #3076 merged Jul 16, 2025

• Enable LineariseRewards to work with negative weights

  #3064 merged Jul 14, 2025

• [BugFix] Wrappers stack fn

  #3061 merged Jul 11, 2025

• [BugFix] Load on empty PRB

  #3057 merged Jul 11, 2025

• [CI] Fix win CI

  #3028 merged Jul 10, 2025

• [BugFix] Fix minari dataloading

  #3054 merged Jul 10, 2025

• [BugFix] Brax memory leak fix

  #3052 merged Jul 10, 2025

• [BugFix] remove update from add when transform is not in-place to keep new metadata

  #3050 merged Jul 9, 2025

• [BugFix] Variable length vllm wrapper answer stacking

  #3049 merged Jul 9, 2025

• [Feature] Multi-node Ray support for GRPO sota-implementation

  #3040 merged Jul 9, 2025

• [BugFix] RB.add unsqueezes tds when applying the transform

  #3047 merged Jul 9, 2025

• [Feature] Reconsider prompts for GRPO

  #3030 merged Jul 9, 2025

• [BugFix,Test] Update CliffWalking version

  #3045 merged Jul 9, 2025

• Disable unrunnable benchmarks cpu test

  #3046 merged Jul 8, 2025

• [Feature] Migrate CQL from D4RLExperienceReplay to MinariExperienceReplay + fix W&B logging and SLURM usage

  #3035 merged Jul 8, 2025

• [Minor] Fix sota runs

  #3042 merged Jul 7, 2025

• [Feature] LLMMaskedCategorical

  #3041 merged Jul 7, 2025

• [Feature] Make dialog_turns_per_batch optional when yield_completed_trajectories=True

  #3039 merged Jul 7, 2025

• Minor fix: remove unused variables in GRPO scripts

  #3038 merged Jul 7, 2025

• [Feature] Allow non-peft models in vLLMUpdater

  #3037 merged Jul 7, 2025

4 Pull requests opened by 2 people

• Fix Habitat

  #3065 opened Jul 14, 2025

• [Refactor] refactor noisy linear

  #3082 opened Jul 17, 2025

• [Bugfix] Wrong minari download first element

  #3106 opened Jul 31, 2025

• [Doc] Better doc on multi-head entropy

  #3109 opened Aug 1, 2025

20 Issues closed by 6 people

• [BUG] UserWarning: Failed to import torchrl C++ binaries.

  #3112 closed Aug 5, 2025

• [BUG] bad markdown display in dqn readme file

  #3110 closed Aug 5, 2025

• [Feature Request] Discrete offline CQL

  #3097 closed Aug 1, 2025

• [Feature Request] MinariExperienceReplay now can handle text fields like "mission"

  #3071 closed Jul 31, 2025

• [Feature Request] Add support for loading datasets from local Minari cache

  #3067 closed Jul 28, 2025

• [Feature Request] Compressing data stored in the Replay Buffer

  #2983 closed Jul 25, 2025

• [BUG] Unexpected use of step in WandbLogger

  #2998 closed Jul 23, 2025

• [BUG] `MaskedCategorical`'s `use_cross_entropy` is not compatible with previously working shapes

  #3078 closed Jul 16, 2025

• [BUG] Inconsistent naming of coefficients of `PPOLoss`

  #3072 closed Jul 16, 2025

• [BUG] Binary tensor spec is not compatible with `shape=()`

  #3074 closed Jul 16, 2025

• [BUG] `TransformedEnv` modifies its`base_env` input_spec

  #3073 closed Jul 16, 2025

• [Feature Request] Enable `LineariseRewards` to work with negative weights

  #3063 closed Jul 14, 2025

• [Feature Request] Explained Variance

  #3009 closed Jul 11, 2025

• [BUG] Cannot load serialized replay buffer with prioritized sampler from disk

  #3056 closed Jul 11, 2025

• [BUG] Replay buffer sampler serialization error

  #3055 closed Jul 10, 2025

• [BUG] Unable to write replay buffer contents to disk

  #3053 closed Jul 10, 2025

• [BUG] `RuntimeError` when creating specific dataset through `MinariExperienceReplay`

  #3051 closed Jul 10, 2025

• [BUG] Memory Leak in BraxEnv with requires_grad=True

  #2837 closed Jul 10, 2025

• [BUG] MultistepTransform is returning emtpy slice first valid datapoint

  #3036 closed Jul 9, 2025

• [Feature Request] Migrate D4RLExperienceReplay to MinariExperienceReplay following official deprecation

  #3034 closed Jul 7, 2025

6 Issues opened by 5 people

• [BUG] AtariDQNExperienceReplay shows AccessDeniedException

  #3113 opened Aug 6, 2025

• [BUG] Download of categorical data in MinariExperienceReplay is completely broken

  #3105 opened Jul 31, 2025

• Bad performance with NoisyLinear

  #3069 opened Jul 14, 2025

• [BUG] Data collection error on M1 Macs

  #3066 opened Jul 14, 2025

• [BUG] QValueActor does not respect action_spec

  #3059 opened Jul 11, 2025

• [Feature Request] Documentation explaining the recommended way to save and load agents

  #3044 opened Jul 8, 2025

5 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• [Feature, Example] A3C Atari Implementation for TorchRL

  #3001 commented on Jul 27, 2025 • 9 new comments

• [Feature Request] Documentation about weight update propagation

  #3032 commented on Jul 11, 2025 • 0 new comments

• [Feature Request] Missing ActionScaling and FlattenAction

  #1209 commented on Jul 15, 2025 • 0 new comments

• [Feature] PPOTrainer

  #2550 commented on Jul 18, 2025 • 0 new comments

• [Algorithm] DPO

  #3025 commented on Jul 16, 2025 • 0 new comments