Call post_init() in PhyloGPNModel for transformers v5 compatibility
#1
by jburtoft - opened
Loading PhyloGPN with transformers v5 (tested on 5.22.2) fails inside from_pretrained:
File ".../transformers/modeling_utils.py", in _move_missing_keys_from_meta_to_device
for key in missing_keys - self.all_tied_weights_keys.keys():
AttributeError: 'PhyloGPNModel' object has no attribute 'all_tied_weights_keys'. Did you mean: '_tied_weights_keys'?
In transformers v5, PreTrainedModel.post_init() sets attributes the loader depends on (e.g. all_tied_weights_keys). PhyloGPNModel.__init__ never calls it. v4 didn't require the call; v5 does.
This PR adds a single self.post_init() call at the end of PhyloGPNModel.__init__.
Testing (CPU, torch 2.15.0):
- transformers 5.22.2: the model-card snippet now runs. All 485 tensors load, with no missing or unexpected keys.
- transformers 4.57.6: still works.
- Outputs (embeddings and F81 log-rates, on the model-card example and a random 2048-bp sequence) are bit-identical between 4.57.6 without the patch and 5.22.2 with it (max abs diff 0.0).
post_init()does not re-initialize loaded weights.
calbors changed pull request status to merged