Add Mistral3 (#36790)

* initial start

* style and dummies

* Create convert_mistral3_weights_to_hf.py

* update

* typo

* typo

* Update convert_mistral3_weights_to_hf.py

* Update convert_mistral3_weights_to_hf.py

* Update convert_mistral3_weights_to_hf.py

* Update convert_mistral3_weights_to_hf.py

* up

* Update convert_mistral3_weights_to_hf.py

* Update convert_mistral3_weights_to_hf.py

* update

* update

* Update image_processing_mistral3.py

* Update convert_mistral3_weights_to_hf.py

* fix patch merger

* Update convert_mistral3_weights_to_hf.py

* Update convert_mistral3_weights_to_hf.py

* up

* update modular to fit

* style

* Update convert_mistral3_weights_to_hf.py

* typo

* Update modular_mistral3.py

* simplify a lot all shape shenanigans

* simplify

* add working test processor

* Add partially working common modeling tests

* All tests working and remove mistral3 image processors

* add docs and fixup

* fix inference with image size >1540

* 🚨fix test image proc pixtral

* Remove vision_feature_select_strategy

* Update convert_mistral3_weights_to_hf.py

* Update convert_mistral3_weights_to_hf.py

* Update convert_mistral3_weights_to_hf.py

* Update convert_mistral3_weights_to_hf.py

* clean

* fix test checkpoints

* Update test_modeling_mistral3.py

* Update test_modeling_mistral3.py

* style

* Use Pixtral processor

* up

* finish cleaning processor to use pixtral directly

* Update __init__.py

* Update processing_pixtral.py

* doc

* Update __init__.py

* Update mistral3.md

* Update _toctree.yml

---------

Co-authored-by: yonigozlan <yoni.gozlan@huggingface.co>
Co-authored-by: yonigozlan <yoni.gozlan10@gmail.com>

tests/models/pixtral/test_image_processing_pixtral.py

@@ -109,8 +109,8 @@ class PixtralImageProcessingTester:
 
             ratio = max(height / max_height, width / max_width)
             if ratio > 1:
-                height = int(np.ceil(height / ratio))
-                width = int(np.ceil(width / ratio))
+                height = int(np.floor(height / ratio))
+                width = int(np.floor(width / ratio))
 
             patch_height, patch_width = self.patch_size["height"], self.patch_size["width"]
             num_height_tokens = (height - 1) // patch_height + 1