update modernvbert modeling (#404)

* update modernvbert modeling

* lint

* handle legacy

* lint

* changelog and pyproject

Author: paultltc

CHANGELOG.md

@@ -11,6 +11,14 @@ and this project adheres to [Semantic Versioning](http://semver.org/).
 
 - Add ColQwen3.5 and BiQwen3.5 support (model + processor). Pretrained checkpoint: [athrael-soju/colqwen3.5-4.5B-v3](https://huggingface.co/athrael-soju/colqwen3.5-4.5B-v3).
 
+### Changed
+
+- Bump minimum supported `transformers` version to `5.3.0`.
+
+### Fixed
+
+- Fix ModernVBERT wrappers to rely on the upstream Hugging Face implementation and keep checkpoint key conversion mapping working with current Transformers v5 loading.
+
 ## [0.3.14] - 2026-02-24
 
 ### Added

colpali_engine/models/modernvbert/__init__.py

@@ -1,2 +1,2 @@
-from .bivbert import BiModernVBert, BiModernVBertProcessor
-from .colvbert import ColModernVBert, ColModernVBertProcessor
+from .bimodernvbert import BiModernVBert, BiModernVBertProcessor
+from .colmodernvbert import ColModernVBert, ColModernVBertProcessor

…e/models/modernvbert/bivbert/__init__.py …ls/modernvbert/bimodernvbert/__init__.pycolpali_engine/models/modernvbert/bivbert/__init__.py renamed to colpali_engine/models/modernvbert/bimodernvbert/__init__.py colpali_engine/models/modernvbert/bivbert/__init__.py renamed to colpali_engine/models/modernvbert/bimodernvbert/__init__.py

@@ -1,8 +1,7 @@
 from typing import Literal
 
 import torch
-
-from colpali_engine.models.modernvbert.modeling_modernvbert import ModernVBertModel, ModernVBertPreTrainedModel
+from transformers import ModernVBertModel, ModernVBertPreTrainedModel
 
 
 class BiModernVBert(ModernVBertPreTrainedModel):

…nvbert/bivbert/modeling_bimodernvbert.py …/bimodernvbert/modeling_bimodernvbert.pycolpali_engine/models/modernvbert/bivbert/modeling_bimodernvbert.py renamed to colpali_engine/models/modernvbert/bimodernvbert/modeling_bimodernvbert.py colpali_engine/models/modernvbert/bivbert/modeling_bimodernvbert.py renamed to colpali_engine/models/modernvbert/bimodernvbert/modeling_bimodernvbert.py

@@ -3,7 +3,7 @@
 import torch
 from transformers import BatchEncoding, BatchFeature
 
-from colpali_engine.models.modernvbert.colvbert import ColModernVBertProcessor  # noqa: N801
+from colpali_engine.models.modernvbert.colmodernvbert import ColModernVBertProcessor  # noqa: N801
 
 
 class BiModernVBertProcessor(ColModernVBertProcessor):  # noqa: N801

…bert/bivbert/processing_bimodernvbert.py …imodernvbert/processing_bimodernvbert.pycolpali_engine/models/modernvbert/bivbert/processing_bimodernvbert.py renamed to colpali_engine/models/modernvbert/bimodernvbert/processing_bimodernvbert.py colpali_engine/models/modernvbert/bivbert/processing_bimodernvbert.py renamed to colpali_engine/models/modernvbert/bimodernvbert/processing_bimodernvbert.py

@@ -1,12 +1,12 @@
 from torch import nn
+from transformers import ModernVBertModel, ModernVBertPreTrainedModel
 from transformers.conversion_mapping import get_checkpoint_conversion_mapping, register_checkpoint_conversion_mapping
 from transformers.core_model_loading import WeightRenaming
 
-from colpali_engine.models.modernvbert.modeling_modernvbert import ModernVBertModel, ModernVBertPreTrainedModel
-
 
 class ColModernVBert(ModernVBertPreTrainedModel):
     _checkpoint_conversion_mapping = {
+        r"^base_model\.model\.model\.text_model": "model.text_model",
         r"^base_model\.model\.custom_text_proj": "custom_text_proj",
     }
     """