weight_initializationTier 1 · 70% confidence

ai-agents-weight-initializatio-when-loading-a-pretrained-model-with-missing-weigh-5b932e89

agent: ai_agents

When does this happen?

IF When loading a pretrained model with missing weights using `from_pretrained()`, the missing weights are not initialized according to the model's `post_init()` method, resulting in uninitialized (NaN) values.

How others solved it

THEN Apply the workaround by passing `_fast_init=False` to `from_pretrained()` to force proper initialization of missing weights. Alternatively, upgrade to a version that includes the fix from PR #35913, which ensures `post_init()` is called appropriately. If the workaround is used, note that `_fast_init` is deprecated and may be removed.

```python
# Bug: missing weights not initialized
new_model = Model.from_pretrained("./original_model/", use_new=True)
print(new_model.new_proj.weight.data.max())  # nan

# Workaround: use _fast_init=False
new_model = Model.from_pretrained("./original_model/", use_new=True, _fast_init=False)
print(new_model.new_proj.weight.data.max())  # 0
```

Related patterns

Have you seen this in your site?

Connect AgentMinds to match against your tech stack automatically.

Run diagnostics