PyTorch Version Compatibility

This page documents known issues with specific PyTorch versions when using NequIP framework models.

Warning

We recommend testing your workflow with your target PyTorch version before deploying to production, especially when using compilation features.

Known Issues by PyTorch Version

PyTorch 2.10.0

Issue 1: CPU + AOTInductor compilation failure

Affected Feature: nequip-compile --mode aotinductor --device cpu

Status: Known bug, may be fixed in future versions

Workaround: Use PyTorch 2.9.1, or use --mode torchscript, or compile for CUDA

Issue 2: AOTInductor runtime load failure in Python scripts

Affected Feature: Running AOTInductor-compiled models from Python, including ASE and torch-sim integrations

Error:

AttributeError: module 'torch._inductor' has no attribute 'codecache'

Status: Fixed in NequIP >=0.17.0

Workaround: Upgrade to NequIP >=0.17.0. If this error appears in your Python script, initialize NequIP global state near the top of the script:

from nequip.utils.global_state import set_global_state
set_global_state()

---

PyTorch 2.9.1

Issue: AOTInductor compilation accuracy failure on A100 GPUs

Affected Feature: nequip-compile --mode aotinductor --device cuda (with OpenEquivariance)

Reproducer: (#574)

nequip-compile \
  nequip.net:mir-group/NequIP-OAM-L:0.1 \
  mir-group__NequIP-OAM-L__0.1.nequip.pt2 \
  --mode aotinductor \
  --device cuda \
  --target ase \
  --modifiers enable_OpenEquivariance

Status: Known bug in PyTorch 2.9.1, compilation check fails with MaxAbsError exceeding tolerance

Workaround: Use PyTorch 2.9.0

---

PyTorch 2.6.0+

Issue: CPU train-time compilation issues

Affected Feature: compile_mode: compile on CPU

Status: Known limitation

Workaround: Avoid CPU train-time compilation; use GPU or eager mode

General CPU Compilation Advisory

Warning

CPU + Compilation has known problems across multiple PyTorch versions.

Be aware that there are known issues when using PyTorch compilation features on CPU devices:

• nequip-compile with --mode aotinductor

• Train-time compilation with compile_mode: compile

If you plan to use compilation on CPU, carefully test your specific workflow and PyTorch version before deployment.

Reporting Issues

If you encounter PyTorch version-specific issues not listed here, please open a GitHub issue with your PyTorch version, workflow details, and error messages.