first commit

tests/models/llama4/test_image_processing_llama4.py

@@ -0,0 +1,116 @@
+# Copyright 2022 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import unittest
+
+from transformers.testing_utils import require_torch, require_vision
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+class Llama4ImageProcessingTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        max_patches=1,
+        do_resize=True,
+        size=None,
+        do_normalize=True,
+        do_pad=False,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+        do_convert_rgb=True,
+    ):
+        super().__init__()
+        size = size if size is not None else {"height": 20, "width": 20}
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.max_patches = max_patches
+        self.do_resize = do_resize
+        self.size = size
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_pad = do_pad
+        self.do_convert_rgb = do_convert_rgb
+
+    def prepare_image_processor_dict(self):
+        return {
+            "max_patches": self.max_patches,
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_convert_rgb": self.do_convert_rgb,
+            "do_pad": self.do_pad,
+        }
+
+    def expected_output_image_shape(self, images):
+        return self.num_channels, self.size["height"], self.size["width"]
+
+    def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        return prepare_image_inputs(
+            batch_size=self.batch_size,
+            num_channels=self.num_channels,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=equal_resolution,
+            numpify=numpify,
+            torchify=torchify,
+        )
+
+
+@require_torch
+@require_vision
+class Llama4ImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = Llama4ImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processor = image_processing_class(**self.image_processor_dict)
+            self.assertTrue(hasattr(image_processor, "do_resize"))
+            self.assertTrue(hasattr(image_processor, "size"))
+            self.assertTrue(hasattr(image_processor, "do_normalize"))
+            self.assertTrue(hasattr(image_processor, "image_mean"))
+            self.assertTrue(hasattr(image_processor, "image_std"))
+            self.assertTrue(hasattr(image_processor, "do_convert_rgb"))
+
+    def test_split_tiles(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processor = image_processing_class(**self.image_processor_dict)
+            image = self.image_processor_tester.prepare_image_inputs(equal_resolution=True)[0]
+            processed_images = image_processor(
+                image,
+                max_patches=16,
+            )
+            self.assertEqual(len(processed_images.pixel_values), 1)
+            self.assertEqual(processed_images.pixel_values[0].shape[0], 17)
+            self.assertEqual(processed_images.pixel_values[0].shape[-2:], (20, 20))

tests/models/llama4/test_modeling_llama4.py

@@ -0,0 +1,122 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Testing suite for the PyTorch Llama4 model."""
+
+import unittest
+
+from transformers import is_torch_available
+from transformers.testing_utils import (
+    Expectations,
+    cleanup,
+    require_torch_large_accelerator,
+    slow,
+    torch_device,
+)
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        Llama4ForConditionalGeneration,
+        Llama4Processor,
+    )
+
+
+@slow
+@require_torch_large_accelerator
+class Llama4IntegrationTest(unittest.TestCase):
+    model_id = "meta-llama/Llama-4-Scout-17B-16E"
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = Llama4ForConditionalGeneration.from_pretrained(
+            "meta-llama/Llama-4-Scout-17B-16E",
+            device_map="auto",
+            dtype=torch.float32,
+            attn_implementation="eager",
+        )
+
+    def setUp(self):
+        self.processor = Llama4Processor.from_pretrained("meta-llama/Llama-4-Scout-17B-16E", padding_side="left")
+
+        url = "https://huggingface.co/datasets/hf-internal-testing/fixtures-captioning/resolve/main/cow_beach_1.png"
+        self.messages_1 = [
+            {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image", "url": url},
+                    {"type": "text", "text": "What is shown in this image?"},
+                ],
+            },
+        ]
+
+        self.messages_2 = [
+            {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "image",
+                        "url": "https://huggingface.co/datasets/hf-internal-testing/fixtures-captioning/resolve/main/cow_beach_1.png",
+                    },
+                    {
+                        "type": "image",
+                        "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/australia.jpg",
+                    },
+                    {"type": "text", "text": "Are these images identical?"},
+                ],
+            },
+        ]
+
+    def tearDown(self):
+        cleanup(torch_device, gc_collect=True)
+
+    def test_model_17b_16e_fp32(self):
+        EXPECTED_TEXTS = Expectations(
+            {
+                ("xpu", 3): ['system\n\nYou are a helpful assistant.user\n\nWhat is shown in this image?assistant\n\nThe image shows a cow standing on a beach with a blue sky and a body of water in the background. The cow is brown with a white face'],
+                ("cuda", None): ['system\n\nYou are a helpful assistant.user\n\nWhat is shown in this image?assistant\n\nThe image shows a cow standing on a beach, with a blue sky and a body of water in the background. The cow is brown with a white'],
+            }
+        )  # fmt: skip
+        EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation()
+
+        inputs = self.processor.apply_chat_template(
+            self.messages_1, tokenize=True, add_generation_prompt=True, return_tensors="pt", return_dict=True
+        ).to(device=torch_device, dtype=self.model.dtype)
+        output = self.model.generate(**inputs, max_new_tokens=30, do_sample=False)
+        output_text = self.processor.batch_decode(output, skip_special_tokens=True)
+
+        print(output_text)
+        self.assertEqual(output_text, EXPECTED_TEXT)
+
+    def test_model_17b_16e_batch(self):
+        inputs = self.processor.apply_chat_template(
+            [self.messages_1, self.messages_2],
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+            padding=True,
+            add_generation_prompt=True,
+        ).to(device=torch_device, dtype=torch.float32)
+
+        output = self.model.generate(**inputs, max_new_tokens=30, do_sample=False)
+        output_text = self.processor.batch_decode(output, skip_special_tokens=True)
+
+        EXPECTED_TEXTS = [
+            'system\n\nYou are a helpful assistant.user\n\nWhat is shown in this image?assistant\n\nThe image shows a cow standing on a beach, with a blue sky and a body of water in the background. The cow is brown with a white',
+            'system\n\nYou are a helpful assistant.user\n\nAre these images identical?assistant\n\nNo, these images are not identical. The first image shows a cow standing on a beach with a blue sky and a white cloud in the background.'
+        ]  # fmt: skip
+        self.assertEqual(output_text, EXPECTED_TEXTS)

tests/models/llama4/test_processing_llama4.py

@@ -0,0 +1,39 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+from transformers import Llama4Processor
+from transformers.testing_utils import require_vision
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+@require_vision
+class Llama4ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = Llama4Processor
+
+    @classmethod
+    def _setup_image_processor(cls):
+        image_processor_class = cls._get_component_class_from_processor("image_processor")
+        return image_processor_class(max_patches=1, size={"height": 20, "width": 20})
+
+    @classmethod
+    def _setup_tokenizer(cls):
+        tokenizer_class = cls._get_component_class_from_processor("tokenizer")
+        return tokenizer_class.from_pretrained("unsloth/Llama-3.2-11B-Vision-Instruct-unsloth-bnb-4bit")
+
+    @classmethod
+    def _setup_test_attributes(cls, processor):
+        cls.image_token = processor.image_token