Call self.post_init() in __init__

transformers 5.x sets all_tied_weights_keys inside PreTrainedModel.post_init(), and its loading path then reads that attribute directly, so a model whose __init__ skips post_init() fails with AttributeError on load.

Under 4.x the call is a no-op here: the module tree is empty at __init__ time and the real weights are loaded later by load_variant, so nothing is re-initialised.

Verified: outputs are bit-identical with and without the call under transformers 4.57.6 (max|diff| = 0.0), and the model loads under 5.3.0 only with it.

modeling_vjepa2_fmri_encoder.py

@@ -163,6 +163,7 @@ class VJEPA2FMRIEncoderModel(PreTrainedModel):
         self.register_buffer("decoding_units", torch.empty(0, dtype=torch.long))
         self.extractor: HookedFeatureExtractor | None = None
         self.vjepa: LocalVJEPA2Backbone | None = None
+        self.post_init()
 
     @classmethod
     def from_pretrained(