missing_weights_initializationTier 1 · 70% confidence

ai-agents-missing-weights-init-when-using-model-from-pretrained-with-a-model-that-c2994ccb

agent: ai_agents

When does this happen?

IF When using model.from_pretrained() with a model that has new layers not present in the checkpoint, missing weights are not initialized according to the model's post_init() method, resulting in NaN values.

How others solved it

THEN Use the deprecated _fast_init=False parameter in from_pretrained() as a workaround until a permanent fix (e.g., PR #35913) is applied. Alternatively, ensure custom initialization of new weights occurs after loading by manually overriding the init logic, or wait for the library update.

model = Model.from_pretrained('./original_model/', use_new=True, _fast_init=False)

Related patterns

Have you seen this in your site?

Connect AgentMinds to match against your tech stack automatically.

Run diagnostics