-
Notifications
You must be signed in to change notification settings - Fork 854
Description
🐛 Describe the bug
As title - example failing job
The root issue appears to be that the nncf dependency is using torch.ao.quantization.pt2e, but I believe this was recently migrated to torchao.quantization.pt2e
Copilot Explanation
The main cause of this failing job is a missing Python module: torch.ao.quantization.pt2e. The error appears during execution of nncf/experimental/torch/fx/quantization/quantize_pt2e.py, specifically:
ModuleNotFoundError: No module named 'torch.ao.quantization.pt2e'
Solution Steps
Ensure Correct Torch Version:
torch.ao.quantization.pt2e` is available only in PyTorch 2.0 and later.
Upgrade PyTorch in your environment with:
Code
pip install --upgrade torch
If using a Docker image, check the Dockerfile to confirm installation of PyTorch 2.x or later.
Verify nncf Version Compatibility
The NNCF package should be compatible with PyTorch 2.x.
Upgrade nncf package if needed:
Code
pip install --upgrade nncf
Review requirements.txt or environment.yml to ensure compatible versions.
Check PYTHONPATH and Installation
Ensure the Python environment running the job is using the correct installation and that torch.ao.quantization.pt2e is available (check with pip show torch and try importing in a shell).
Code Suggestion:
If your Dockerfile or install scripts use a specific torch version, update them. Example Dockerfile snippet:
Dockerfile
RUN pip install torch==2.1.0 nncf
Summary:
Upgrade your PyTorch installation to at least version 2.0 and verify compatibility of nncf. This will resolve the ModuleNotFoundError and unblock the job.
Versions
N/A