Source installation shows dependency conflicts inside official PhysicsNeMo containers (25.08 & 25.11)

[jak-sch]

Summary

Following the official installation instructions for PhysicsNeMo-Curator from source inside the recommended NVIDIA PhysicsNeMo containers (both 25.08 and 25.11), the installation consistently produces dependency conflicts in a fresh virtual environment. These conflicts appear even when installing without the [dev] extras and may indicate that pyproject.toml declares dependencies that are not compatible with the base container environment.

The installation completes, and physicsnemo_curator can be imported, but the environment remains in an inconsistent state.

Environment

• OS: NVIDIA PhysicsNeMo container
  • nvcr.io/nvidia/physicsnemo/physicsnemo:25.08 (from installation docs)
  • nvcr.io/nvidia/physicsnemo/physicsnemo:25.11 (latest)

Steps to reproduce

Inside either the 25.08 or 25.11 container:

git clone https://github.com/NVIDIA/physicsnemo-curator.git
cd physicsnemo-curator

python3 -m venv clean_env
source clean_env/bin/activate
pip install --upgrade pip

pip install -e ".[dev]"

(also reproducible with pip install -e .)

Actual behaviour / Errors

For container version 25.08 (indicated in README), pip reports the following dependency conflicts:

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
dgl 2.4.0+cu124 requires torch<=2.4.0, but you have torch 2.8.0a0+5228986c39.nv25.6 which is incompatible.
nvidia-dali-cuda120 1.50.0 requires packaging<=24.2, but you have packaging 25.0 which is incompatible.
nvidia-resiliency-ext 0.4.0 requires psutil>=6.0.0, but you have psutil 5.9.8 which is incompatible.

For container version 25.11, the following dependency conflicts are reported:

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
nvidia-resiliency-ext 0.4.1+cuda13 requires pynvml>=12.0.0, which is not installed.
nvidia-resiliency-ext 0.4.1+cuda13 requires psutil>=6.0.0, but you have psutil 5.9.8 which is incompatible.

The installation completes, but the environment is left in a state where:

• DGL depends on an older PyTorch version
• DALI requires an older packaging
• resiliency-ext requires a newer psutil

These versions are not aligned with the versions inside either PhysicsNeMo container.