🐛 Bug Description

When using model.save_pretrained_torchao(), the function incorrectly uses AutoModel instead of AutoModelForCausalLM to reload the 16-bit model.

This causes the saved config.json in the final -torchao directory to have the base model architecture (e.g., Qwen3Model) instead of the language modeling head architecture (e.g., Qwen3ModelForCausalLM).

reproducing the bug

You can see this in the unsloth/save.py file, inside the unsloth_save_pretrained_torchao function.

The problematic lines are:

On line 2772:
from transformers import AutoModel, AutoTokenizer, TorchAoConfig

And around line 2791:
model = AutoModel.from_pretrained(...)

✅ The Fix

This bug is fixed by changing the function to use AutoModelForCausalLM:

1. Change the import to:
   from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig

2. Change the model loading line to:
   model = AutoModelForCausalLM.from_pretrained(...)