pub fn auto_detect_weight_prefix(store: &WeightStore, config: &mut ModelConfig)Expand description
Resolve the weight-key prefix a nested (multimodal) checkpoint uses.
Lives here rather than in the server because it depends only on the store
and the config, and BOTH the serve path and the integration harness need it:
a nested checkpoint stores model.language_model.layers.0.…, and a caller
that skips this step fails with “Weight ‘model.layers.0.input_layernorm.
weight’ not found in store” after a full weight load. Keeping one copy is
what stops the test from accepting a different set of checkpoints than
production does.