huggingface
diff --git a/‎tests/models/audioflamingo3/test_processing_audioflamingo3.py‎
Lines changed: 29 additions & 11 deletions b/‎tests/models/audioflamingo3/test_processing_audioflamingo3.py‎
Lines changed: 29 additions & 11 deletions
diff --git a/‎tests/models/colpali/test_processing_colpali.py‎
Lines changed: 5 additions & 4 deletions b/‎tests/models/colpali/test_processing_colpali.py‎
Lines changed: 5 additions & 4 deletions
diff --git a/‎tests/models/deepseek_ocr2/test_processing_deepseek_ocr2.py‎
Lines changed: 6 additions & 8 deletions b/‎tests/models/deepseek_ocr2/test_processing_deepseek_ocr2.py‎
Lines changed: 6 additions & 8 deletions
diff --git a/‎tests/models/florence2/test_processing_florence2.py‎
Lines changed: 6 additions & 9 deletions b/‎tests/models/florence2/test_processing_florence2.py‎
Lines changed: 6 additions & 9 deletions
diff --git a/‎tests/models/glmasr/test_processing_glmasr.py‎
Lines changed: 7 additions & 7 deletions b/‎tests/models/glmasr/test_processing_glmasr.py‎
Lines changed: 7 additions & 7 deletions
diff --git a/‎tests/models/got_ocr2/test_processing_got_ocr2.py‎
Lines changed: 12 additions & 10 deletions b/‎tests/models/got_ocr2/test_processing_got_ocr2.py‎
Lines changed: 12 additions & 10 deletions
diff --git a/‎tests/models/granite4_vision/test_processing_granite4_vision.py‎
Lines changed: 18 additions & 14 deletions b/‎tests/models/granite4_vision/test_processing_granite4_vision.py‎
Lines changed: 18 additions & 14 deletions
@@ -25,21 +25,22 @@
     AutoTokenizer,
     WhisperFeatureExtractor,
 )
-from transformers.testing_utils import require_librosa, require_torch
+from transformers.testing_utils import require_librosa, require_torch, slow
 
 from ...test_processing_common import MODALITY_INPUT_DATA, ProcessorTesterMixin
 
 
 class AudioFlamingo3ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
     processor_class = AudioFlamingo3Processor
+    # Tiny processor created with make_tiny_processor.py from "nvidia/audio-flamingo-3-hf"
+    tiny_model_id = "hf-internal-testing/tiny-processor-audioflamingo3"
+    checkpoint = "nvidia/audio-flamingo-3-hf"
 
     @classmethod
     @require_torch
     def setUpClass(cls):
-        cls.checkpoint = "nvidia/audio-flamingo-3-hf"
         cls.tmpdirname = tempfile.mkdtemp()
-
-        processor = AudioFlamingo3Processor.from_pretrained(cls.checkpoint)
+        processor = AudioFlamingo3Processor.from_pretrained(cls.tiny_model_id)
         processor.save_pretrained(cls.tmpdirname)
 
     @require_torch
@@ -60,14 +61,14 @@ def tearDownClass(cls):
 
     @require_torch
     def test_can_load_various_tokenizers(self):
-        processor = AudioFlamingo3Processor.from_pretrained(self.checkpoint)
-        tokenizer = AutoTokenizer.from_pretrained(self.checkpoint)
+        processor = AudioFlamingo3Processor.from_pretrained(self.tiny_model_id)
+        tokenizer = AutoTokenizer.from_pretrained(self.tiny_model_id)
         self.assertEqual(processor.tokenizer.__class__, tokenizer.__class__)
 
     @require_torch
     def test_save_load_pretrained_default(self):
-        tokenizer = AutoTokenizer.from_pretrained(self.checkpoint)
-        processor = AudioFlamingo3Processor.from_pretrained(self.checkpoint)
+        tokenizer = AutoTokenizer.from_pretrained(self.tiny_model_id)
+        processor = AudioFlamingo3Processor.from_pretrained(self.tiny_model_id)
         feature_extractor = processor.feature_extractor
 
         processor = AudioFlamingo3Processor(tokenizer=tokenizer, feature_extractor=feature_extractor)
@@ -82,6 +83,21 @@ def test_save_load_pretrained_default(self):
 
     @require_torch
     def test_tokenizer_integration(self):
+        slow_tokenizer = AutoTokenizer.from_pretrained(self.tiny_model_id, use_fast=False)
+        fast_tokenizer = AutoTokenizer.from_pretrained(self.tiny_model_id, from_slow=True, legacy=False)
+
+        prompt = (
+            "<|im_start|>system\nAnswer the questions.<|im_end|>"
+            "<|im_start|>user\n<sound>What is it?<|im_end|>"
+            "<|im_start|>assistant\n"
+        )
+
+        # Verify slow and fast tokenizers produce the same output (parity test)
+        self.assertEqual(slow_tokenizer.tokenize(prompt), fast_tokenizer.tokenize(prompt))
+
+    @slow
+    @require_torch
+    def test_tokenizer_full_integration(self):
         slow_tokenizer = AutoTokenizer.from_pretrained(self.checkpoint, use_fast=False)
         fast_tokenizer = AutoTokenizer.from_pretrained(self.checkpoint, from_slow=True, legacy=False)
 
@@ -118,7 +134,7 @@ def test_tokenizer_integration(self):
 
     @require_torch
     def test_chat_template(self):
-        processor = AutoProcessor.from_pretrained(self.checkpoint)
+        processor = self.get_processor()
         expected_prompt = (
             "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
             "<|im_start|>user\n<sound>What is surprising about the relationship between the barking and the music?<|im_end|>\n"
@@ -146,9 +162,11 @@ def test_chat_template(self):
 
     @require_torch
     def test_apply_transcription_request_single(self):
-        processor = AutoProcessor.from_pretrained(self.checkpoint)
+        processor = self.get_processor()
 
-        audio_url = "https://huggingface.co/datasets/nvidia/AudioSkills/resolve/main/assets/t_837b89f2-26aa-4ee2-bdf6-f73f0dd59b26.wav"
+        audio_url = (
+            "https://huggingface.co/datasets/raushan-testing-hf/audio-test/resolve/main/f2641_0_throatclearing.wav"
+        )
         helper_outputs = processor.apply_transcription_request(audio=audio_url)
 
         conversation = [
 
@@ -25,7 +25,7 @@
 
 
 if is_vision_available():
-    from transformers import ColPaliProcessor, GemmaTokenizer
+    from transformers import ColPaliProcessor, GemmaTokenizer, SiglipImageProcessor
 
 SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece.model")
 
@@ -40,8 +40,9 @@ def _setup_tokenizer(cls):
 
     @classmethod
     def _setup_image_processor(cls):
-        image_processor_class = cls._get_component_class_from_processor("image_processor")
-        image_processor = image_processor_class.from_pretrained("google/siglip-so400m-patch14-384")
+        # Use 64×64 instead of the default 384×384 from google/siglip-so400m-patch14-384 to avoid
+        # large tensors. image_seq_length=0 matches the processor attribute so token-count tests pass.
+        image_processor = SiglipImageProcessor(size={"height": 64, "width": 64})
         image_processor.image_seq_length = 0
         return image_processor
 
@@ -85,7 +86,7 @@ def test_process_images(self):
 
         # Assertions
         self.assertIn("pixel_values", batch_feature)
-        self.assertEqual(batch_feature["pixel_values"].shape, torch.Size([1, 3, 384, 384]))
+        self.assertEqual(batch_feature["pixel_values"].shape, torch.Size([1, 3, 64, 64]))
 
     @require_torch
     @require_vision
 
@@ -25,21 +25,19 @@
 @require_vision
 class DeepseekOcr2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
     processor_class = DeepseekOcr2Processor
+    # Tiny processor created with make_tiny_processor.py from "deepseek-community/DeepSeek-OCR-2"
+    tiny_model_id = "hf-internal-testing/tiny-processor-deepseek_ocr2"
 
     @classmethod
     def _setup_image_processor(cls):
+        # Small size (64×64) reduces the number of tiles produced by the tiling logic,
+        # keeping token counts low. tile_size=512 is a safe sentinel above the image size.
         image_processor_class = cls._get_component_class_from_processor("image_processor")
         image_processor = image_processor_class()
         image_processor.size = {"height": 64, "width": 64}
         image_processor.tile_size = 512
         return image_processor
 
-    @classmethod
-    def _setup_tokenizer(cls):
-        tokenizer_class = cls._get_component_class_from_processor("tokenizer")
-        tokenizer = tokenizer_class.from_pretrained("deepseek-community/DeepSeek-OCR-2")
-        return tokenizer
-
     @classmethod
     def _setup_test_attributes(cls, processor):
         cls.image_token = processor.image_token
@@ -73,8 +71,8 @@ def test_image_token_expansion_large_image(self):
         processor.image_processor.size = {"height": 1024, "width": 1024}
         processor.image_processor.tile_size = 768
 
-        # Large image: max(2448, 3264) > 768 → local patches
-        image = torch.randint(0, 256, (3, 3264, 2448), dtype=torch.uint8)
+        # Large image: max(769, 577) > 768 → local patches; same 2×3 grid as 3264×2448 (ar≈0.75)
+        image = torch.randint(0, 256, (3, 769, 577), dtype=torch.uint8)
         prompt = "<image>\nFree OCR."
 
         inputs = processor(images=image, text=prompt, return_tensors="pt")
 
@@ -30,22 +30,19 @@
 @require_vision
 class Florence2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
     processor_class = Florence2Processor
+    # Tiny processor created with make_tiny_processor.py from "microsoft/Florence-2-base"
+    tiny_model_id = "hf-internal-testing/tiny-processor-florence2"
 
     @classmethod
     def _setup_image_processor(cls):
+        # Florence2Processor reads image_processor.image_seq_length at construction time
+        # (processing_florence2.py line 99) to set num_image_tokens. Use a small value (2)
+        # to avoid large token sequences in tests.
         image_processor_class = cls._get_component_class_from_processor("image_processor")
-        image_processor = image_processor_class.from_pretrained("florence-community/Florence-2-base")
+        image_processor = image_processor_class.from_pretrained(cls.tiny_model_id)
         image_processor.image_seq_length = 2
         return image_processor
 
-    @classmethod
-    def _setup_tokenizer(cls):
-        tokenizer_class = cls._get_component_class_from_processor("tokenizer")
-        tokenizer = tokenizer_class.from_pretrained("florence-community/Florence-2-base")
-        tokenizer.image_token = "<image>"
-        tokenizer.image_token_id = tokenizer.encode(tokenizer.image_token, add_special_tokens=False)[0]
-        return tokenizer
-
     @classmethod
     def _setup_test_attributes(cls, processor):
         # override: Florence shouldn't have any image-token in input text
 
@@ -31,14 +31,14 @@
 
 class GlmAsrProcessorTest(ProcessorTesterMixin, unittest.TestCase):
     processor_class = GlmAsrProcessor
+    # Tiny processor created with make_tiny_processor.py from "zai-org/GLM-ASR-Nano-2512"
+    tiny_model_id = "hf-internal-testing/tiny-processor-glmasr"
 
     @classmethod
     @require_torch
     def setUpClass(cls):
-        cls.checkpoint = "zai-org/GLM-ASR-Nano-2512"
         cls.tmpdirname = tempfile.mkdtemp()
-
-        processor = GlmAsrProcessor.from_pretrained(cls.checkpoint)
+        processor = GlmAsrProcessor.from_pretrained(cls.tiny_model_id)
         processor.save_pretrained(cls.tmpdirname)
 
     @require_torch
@@ -59,14 +59,14 @@ def tearDownClass(cls):
 
     @require_torch
     def test_can_load_various_tokenizers(self):
-        processor = GlmAsrProcessor.from_pretrained(self.checkpoint)
-        tokenizer = AutoTokenizer.from_pretrained(self.checkpoint)
+        processor = GlmAsrProcessor.from_pretrained(self.tiny_model_id)
+        tokenizer = AutoTokenizer.from_pretrained(self.tiny_model_id)
         self.assertEqual(processor.tokenizer.__class__, tokenizer.__class__)
 
     @require_torch
     def test_save_load_pretrained_default(self):
-        tokenizer = AutoTokenizer.from_pretrained(self.checkpoint)
-        processor = GlmAsrProcessor.from_pretrained(self.checkpoint)
+        tokenizer = AutoTokenizer.from_pretrained(self.tiny_model_id)
+        processor = GlmAsrProcessor.from_pretrained(self.tiny_model_id)
         feature_extractor = processor.feature_extractor
 
         processor = GlmAsrProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
 
@@ -23,12 +23,14 @@
 @require_vision
 class GotOcr2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
     processor_class = GotOcr2Processor
+    # Tiny processor created with make_tiny_processor.py from "stepfun-ai/GOT-OCR-2.0-hf"
+    tiny_model_id = "hf-internal-testing/tiny-processor-got_ocr2"
 
     @classmethod
-    def _setup_tokenizer(cls):
-        tokenizer_class = cls._get_component_class_from_processor("tokenizer")
-        tokenizer = tokenizer_class.from_pretrained("stepfun-ai/GOT-OCR-2.0-hf")
-        return tokenizer
+    def _setup_image_processor(cls):
+        # Instantiate directly to avoid loading the full 384×384 image processor from Hub.
+        image_processor_class = cls._get_component_class_from_processor("image_processor")
+        return image_processor_class()
 
     @unittest.skip("GotOcr2Processor pop the image processor output 'num_patches'")
     def test_image_processor_defaults(self):
@@ -38,27 +40,27 @@ def test_ocr_queries(self):
         processor = self.get_processor()
         image_input = self.prepare_image_inputs()
         inputs = processor(image_input, return_tensors="pt")
-        self.assertEqual(inputs["input_ids"].shape, (1, 286))
+        self.assertEqual(inputs["input_ids"].shape, (1, 324))
         self.assertEqual(inputs["pixel_values"].shape, (1, 3, 384, 384))
 
         inputs = processor(image_input, return_tensors="pt", format=True)
-        self.assertEqual(inputs["input_ids"].shape, (1, 288))
+        self.assertEqual(inputs["input_ids"].shape, (1, 328))
         self.assertEqual(inputs["pixel_values"].shape, (1, 3, 384, 384))
 
         inputs = processor(image_input, return_tensors="pt", color="red")
-        self.assertEqual(inputs["input_ids"].shape, (1, 290))
+        self.assertEqual(inputs["input_ids"].shape, (1, 329))
         self.assertEqual(inputs["pixel_values"].shape, (1, 3, 384, 384))
 
         inputs = processor(image_input, return_tensors="pt", box=[0, 0, 100, 100])
-        self.assertEqual(inputs["input_ids"].shape, (1, 303))
+        self.assertEqual(inputs["input_ids"].shape, (1, 341))
         self.assertEqual(inputs["pixel_values"].shape, (1, 3, 384, 384))
 
         inputs = processor([image_input, image_input], return_tensors="pt", multi_page=True, format=True)
-        self.assertEqual(inputs["input_ids"].shape, (1, 547))
+        self.assertEqual(inputs["input_ids"].shape, (1, 595))
         self.assertEqual(inputs["pixel_values"].shape, (2, 3, 384, 384))
 
         inputs = processor(image_input, return_tensors="pt", crop_to_patches=True, max_patches=6)
-        self.assertEqual(inputs["input_ids"].shape, (1, 1826))
+        self.assertEqual(inputs["input_ids"].shape, (1, 1872))
         self.assertEqual(inputs["pixel_values"].shape, (7, 3, 384, 384))
 
     def test_processor_text_has_no_visual(self):
 
@@ -17,7 +17,7 @@
 
 import torch
 
-from transformers import Granite4VisionProcessor
+from transformers import Granite4VisionProcessor, LlavaNextImageProcessor
 from transformers.testing_utils import require_vision
 
 from ...test_processing_common import ProcessorTesterMixin
@@ -26,21 +26,24 @@
 @require_vision
 class Granite4VisionProcessorTest(ProcessorTesterMixin, unittest.TestCase):
     processor_class = Granite4VisionProcessor
+    # Tiny processor created with make_tiny_processor.py from "ibm-granite/granite-vision-4.1-4b"
+    tiny_model_id = "hf-internal-testing/tiny-processor-granite4_vision"
     # Image token expansion with downsample_rate="1/2" produces more tokens than the defaults
     image_text_kwargs_max_length = 300
     image_text_kwargs_override_max_length = 280
     image_unstructured_max_length = 260
 
     @classmethod
-    def _setup_tokenizer(cls):
-        tokenizer_class = cls._get_component_class_from_processor("tokenizer")
-        tokenizer = tokenizer_class.from_pretrained("huggyllama/llama-7b")
-        tokenizer.add_special_tokens({"additional_special_tokens": ["<image>"]})
-        if not tokenizer.pad_token:
-            tokenizer.pad_token = "[PAD]"
-            if tokenizer.pad_token_id is None:
-                tokenizer.pad_token_id = 0
-        return tokenizer
+    def _setup_image_processor(cls):
+        # Must use LlavaNextImageProcessor (not CLIPImageProcessor from the tiny repo) because
+        # processing_granite4_vision.py calls iter(image_inputs["image_sizes"]), which requires
+        # the image_sizes key that only LlavaNextImageProcessor produces. Small sizes keep
+        # tensor allocations minimal.
+        return LlavaNextImageProcessor(
+            size={"shortest_edge": 64},
+            crop_size={"height": 64, "width": 64},
+            image_grid_pinpoints=[[64, 64]],
+        )
 
     @classmethod
     def _setup_test_attributes(cls, processor):
@@ -52,6 +55,7 @@ def prepare_processor_dict():
             "chat_template": "{% for message in messages %}{% if message['role'] != 'system' %}{{ message['role'].upper() + ': '}}{% endif %}{# Render all images first #}{% for content in message['content'] | selectattr('type', 'equalto', 'image') %}{{ '<image>\n' }}{% endfor %}{# Render all text next #}{% if message['role'] != 'assistant' %}{% for content in message['content'] | selectattr('type', 'equalto', 'text') %}{{ content['text'] + ' '}}{% endfor %}{% else %}{% for content in message['content'] | selectattr('type', 'equalto', 'text') %}{% generation %}{{ content['text'] + ' '}}{% endgeneration %}{% endfor %}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ 'ASSISTANT:' }}{% endif %}",
             "patch_size": 14,
             "vision_feature_select_strategy": "default",
+            "num_additional_image_tokens": 1,
             "downsample_rate": "1/2",
         }  # fmt: skip
 
@@ -97,10 +101,10 @@ def test_image_token_filling(self):
         #   aspect: 316/503 = 0.628, 12/24 = 0.5 -> orig > current -> new_height = round(503*(12/316)) = 19
         #   padding = (24-19)//2 = 2, current_height = 24 - 4 = 20
         #   unpadded = 20*12 = 240, newline = 20
-        #   base = 12*12 + 0 = 144
-        #   total = 240 + 20 + 144 = 404
-        #   with "default" strategy: 404 - 1 = 403
-        expected_image_tokens = 403
+        #   base = 12*12 + num_additional_image_tokens(1) = 145
+        #   total = 240 + 20 + 145 = 405
+        #   with "default" strategy: 405 - 1 = 404
+        expected_image_tokens = 404
 
         messages = [
             {