ai-agents-model-weight-initial-when-loading-a-pretrained-hugging-face-model-with--7b4e9613

THEN As a temporary workaround, pass `_fast_init=False` to `from_pretrained()` to ensure `post_init()` is called for missing weights. For a permanent fix, update to a version of `transformers` that includes PR #35913, which properly initializes new layers after loading. Alternatively, manually initialize the missing weights using PyTorch's `nn.init` after loading the model.

```python
# Workaround: use _fast_init=False
model = Model.from_pretrained("./checkpoint/", use_new=True, _fast_init=False)
# Or manually initialize new weights after loading
model = Model.from_pretrained("./checkpoint/", use_new=True)
model.new_proj.weight.data.fill_(0.0)
```