Update modeling_genbio_pathfm.py

To work with latest transformer version

modeling_genbio_pathfm.py

@@ -5,6 +5,8 @@ This file is intended to live in the HuggingFace repo at
 
     from transformers import AutoModel
     model = AutoModel.from_pretrained("genbio-ai/genbio-pathfm", trust_remote_code=True)
+
+Compatible with transformers v4.x and v5.0+.
 """
 
 import math
@@ -484,14 +486,21 @@ class GenBioPathFMModel(PreTrainedModel):
         model = AutoModel.from_pretrained("genbio-ai/genbio-pathfm", trust_remote_code=True)
 
         # CLS-only:  [B, embed_dim*3]
-        cls_features = model(rgb_tensor)
+        cls_features = model(pixel_values)
 
         # CLS + patch tokens:
-        cls_features, patch_features = model.forward_with_patches(rgb_tensor)
+        cls_features, patch_features = model.forward_with_patches(pixel_values)
     """
 
     config_class = GenBioPathFMConfig
 
+    # -- HuggingFace class attributes for compatibility (v4.x + v5.0+) --
+    base_model_prefix = "backbone"
+    main_input_name = "pixel_values"
+    supports_gradient_checkpointing = False
+    _no_split_modules = ["SelfAttentionBlock"]
+    _tied_weights_keys = []  # no tied weights (required by transformers v5.0+)
+
     def __init__(self, config: GenBioPathFMConfig):
         super().__init__(config)
         self.backbone = VisionTransformer(
@@ -519,6 +528,9 @@ class GenBioPathFMModel(PreTrainedModel):
             pos_embed_rope_rescale_coords=config.pos_embed_rope_rescale_coords,
             pos_embed_rope_dtype=config.pos_embed_rope_dtype,
         )
+        # Required: sets up all_tied_weights_keys, parallelism plans, etc.
+        # Safe on v4.x, required on v5.0+.
+        self.post_init()
 
     def _encode(self, x: Tensor) -> Dict[str, Tensor]:
         """Encode single-channel [B, 1, H, W] images."""