torchRL version compatibility: Failed to import torchrl C++ binaries

[MarioAsti8]

After installing ACEGEN and its dependencies, I tried to run scripts/reinforce/reinforce.py. I had an error that seems to be related to torchRL version and compatibility with other dependencies. I tried both the instructions given in the README file and creating my conda environment through the provided .yml file, only changing the cuda version (mine is 11.8), with the same result in both cases. Below is the full error:

python scripts/reinforce/reinforce.py --config-name config_denovo
/home/mario/miniconda3/envs/acegen/lib/python3.10/site-packages/torchrl/data/replay_buffers/samplers.py:37: UserWarning: Failed to import torchrl C++ binaries. Some modules (eg, prioritized replay buffers) may not work with your installation. You seem to be using the nightly version of TorchRL. If this is a local install, there might be an issue with the local installation. Here are some tips to debug this:

• make sure ninja and cmake were installed
• make sure you ran python setup.py clean && python setup.py develop and that no error was raised
• make sure the version of PyTorch you are using matches the one that was present in your virtual env during setup.
  warnings.warn(EXTENSION_WARNING)
  WARNING:molscore.scoring_functions:OpenEye functions: currently unavailable due to the following: No module named 'openeye'
  WARNING:moleval.metrics.posecheck:PoseCheck metrics: currently unavailable due to the following: No module named 'prolif'
  Error executing job with overrides: []
  Traceback (most recent call last):
  File "/home/mario/software/acegen-open/scripts/reinforce/reinforce.py", line 62, in main
  run_task(cfg, run_reinforce, file)
  File "/home/mario/miniconda3/envs/acegen/lib/python3.10/site-packages/acegen/script_helpers.py", line 88, in run_task
  algorithm(cfg, task)
  File "/home/mario/software/acegen-open/scripts/reinforce/reinforce.py", line 129, in run_reinforce
  sampler=PrioritizedSampler(storage.max_size, alpha=1.0, beta=1.0),
  File "/home/mario/miniconda3/envs/acegen/lib/python3.10/site-packages/torchrl/data/replay_buffers/samplers.py", line 384, in init
  self._init()
  File "/home/mario/miniconda3/envs/acegen/lib/python3.10/site-packages/torchrl/data/replay_buffers/samplers.py", line 418, in _init
  self._sum_tree = SumSegmentTreeFp32(self._max_capacity)
  NameError: name 'SumSegmentTreeFp32' is not defined

Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

[zhf3564859793]

I also meet this problem.
For de novo generation:
python scripts/a2c/a2c.py --config-name config_denovo
/root/miniconda3/lib/python3.10/site-packages/torchrl/data/replay_buffers/samplers.py:37: UserWarning: Failed to import torchrl C++ binaries. Some modules (eg, prioritized replay buffers) may not work with your installation. You seem to be using the nightly version of TorchRL. If this is a local install, there might be an issue with the local installation. Here are some tips to debug this:

• make sure ninja and cmake were installed
• make sure you ran python setup.py clean && python setup.py develop and that no error was raised
• make sure the version of PyTorch you are using matches the one that was present in your virtual env during setup.
  warnings.warn(EXTENSION_WARNING)
  SKLearnModel: currently unavailable due to the following: No module named 'joblib'
  OpenEye functions: currently unavailable due to the following: No module named 'openeye'
  PoseCheck metrics: currently unavailable due to the following: No module named 'prolif'
  /root/miniconda3/lib/python3.10/site-packages/torch/_utils.py:831: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly. To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
  return self.fget.get(instance, owner)()
  100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 10000/10000 [05:23<00:00, 30.89it/s]

Although it has warning, but it still works, and the result seems good.

But for scaffold decoration:
python scripts/a2c/a2c.py --config-name config_scaffold
/root/miniconda3/lib/python3.10/site-packages/torchrl/data/replay_buffers/samplers.py:37: UserWarning: Failed to import torchrl C++ binaries. Some modules (eg, prioritized replay buffers) may not work with your installation. You seem to be using the nightly version of TorchRL. If this is a local install, there might be an issue with the local installation. Here are some tips to debug this:

• make sure ninja and cmake were installed
• make sure you ran python setup.py clean && python setup.py develop and that no error was raised
• make sure the version of PyTorch you are using matches the one that was present in your virtual env during setup.
  warnings.warn(EXTENSION_WARNING)
  SKLearnModel: currently unavailable due to the following: No module named 'joblib'
  OpenEye functions: currently unavailable due to the following: No module named 'openeye'
  PoseCheck metrics: currently unavailable due to the following: No module named 'prolif'
  Failed to identify lib-invent, attempting to create it automatically (this may take several minutes)
  [2025-12-21 23:08:25,437][base][WARNING] - Failed to identify lib-invent, attempting to create it automatically (this may take several minutes)
  Collecting package metadata (repodata.json): done
  Solving environment: /
  Then it doesn't work.

Thank you very much if you could help me with this problem.

[MorganCThomas]

Yes resolving the TorchRL / Tensordict / Torch dependencies needs improving as TorchRL and Tensordict were still under very active development. I'll look at that this week to try and come up with a more robust solution - apologies for you running into these issues.

@MarioAsti8 @zhf3564859793 could you please confirm the torch versions in your environment with something like
pip list | grep torch

[zhf3564859793]

Dear MorganCThomas,

Thanks for your nice reply!

These are my versions for your reference.
PyTorch 2.1.0
Python 3.10(ubuntu22.04)
CUDA 12.1

Thanks for your nice help!