Model loading fails silently on Python 3.12 with PyTorch Nightly

[Liux100720]

Hello, the node is failing to load the joycaption-beta-one model on my system.

My Environment:

• OS: Linux
• GPU: NVIDIA GeForce RTX 5070 Ti
• Python Version: 3.12.3
• PyTorch Version: Latest Nightly Build (torch-2.9.0+cu128 or similar)
• Transformers Version: Latest (4.57.1 or similar)

The Problem: When I queue a prompt, the terminal shows "Loading checkpoint shards: 0%" and then the prompt immediately finishes in about 1.5 seconds. The node in the UI shows the error: "Error loading model: 'JC_Models' object has no attribute 'model'".

Important Details:

• This happens in both normal GPU mode and in CPU-only mode (--cpu).
• All my model files and paths are correct. The model shards are fully downloaded.
• My Python environment has been fully updated to resolve all other issues (setuptools, etc.).

This seems to be a code-level incompatibility between the node and the latest versions of the core AI libraries required for modern hardware.

[1038lab]

Thanks for the detailed report! This seems related to Python 3.12 and the latest PyTorch nightly build. I haven’t tested JoyCaption with that combination yet.

Could you please try with a stable PyTorch release (e.g. 2.8.0) and see if the issue still happens? If it works there, I’ll add compatibility updates for 3.12 + nightly in the next patch.

[MannySJH]

Same thing happens to me on pytorch 2.9 stable, cu130

[1038lab]

Testing occurs on CUDA 12.8 and PyTorch 2.8 python 3.12 execution is smooth.