Chat template: return vectorized output in processors (#34275)

* update chat template * style * fix tests * Update src/transformers/image_utils.py Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com> * typehints + docs * fix tests * remove unnecessary warnings * forgot code style :( * allow users to pass backend and num frames * Update docs/source/en/chat_templating.md Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com> * Update src/transformers/image_utils.py Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com> * Update src/transformers/image_utils.py Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com> * Update src/transformers/image_utils.py Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com> * Update src/transformers/image_utils.py Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com> * Update src/transformers/image_utils.py Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com> * Update src/transformers/image_utils.py Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com> * Update src/transformers/processing_utils.py Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com> * typo fix * style * address comments * align with "pipeline" template * update docs * update docs * unpack for all kwargs? * wrong conflict resolution while rebasing * tmp * update docs * Update docs/source/en/chat_templating.md Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com> * Update docs/source/en/chat_templating.md Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com> * Update docs/source/en/chat_templating.md Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com> * Update docs/source/en/chat_templating.md Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com> --------- Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com> Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
2025-01-10 11:05:29 +01:00
parent 5f087d1335
commit e0646f3dce
12 changed files with 880 additions and 46 deletions
--- a/tests/models/llava_onevision/test_processor_llava_onevision.py
+++ b/tests/models/llava_onevision/test_processor_llava_onevision.py
@@ -16,8 +16,8 @@ import shutil
 import tempfile
 import unittest

-from transformers.testing_utils import require_vision
-from transformers.utils import is_vision_available
+from transformers.testing_utils import require_av, require_torch, require_vision
+from transformers.utils import is_torch_available, is_vision_available

 from ...test_processing_common import ProcessorTesterMixin

@@ -31,6 +31,9 @@ if is_vision_available():
        Qwen2TokenizerFast,
    )

+if is_torch_available:
+    import torch
+

@require_vision
 class LlavaOnevisionProcessorTest(ProcessorTesterMixin, unittest.TestCase):
@@ -100,3 +103,60 @@ class LlavaOnevisionProcessorTest(ProcessorTesterMixin, unittest.TestCase):

        formatted_prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
        self.assertEqual(expected_prompt, formatted_prompt)
+
+    @require_av
+    def test_chat_template_dict(self):
+        processor = AutoProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf")
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "video"},
+                    {"type": "text", "text": "What is shown in this video?"},
+                ],
+            },
+        ]
+
+        formatted_prompt_tokenized = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True)
+        expected_output = [[151644, 872, 220, 151647, 198, 3838, 374, 6839, 304, 419, 2766, 30, 151645, 151644, 77091, 198]]  # fmt: skip
+        self.assertListEqual(expected_output, formatted_prompt_tokenized)
+
+        out_dict = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True)
+        self.assertListEqual(list(out_dict.keys()), ["input_ids", "attention_mask"])
+
+        # add image URL for return dict
+        messages[0]["content"][0] = {
+            "type": "video",
+            "url": "https://test-videos.co.uk/vids/bigbuckbunny/mp4/h264/720/Big_Buck_Bunny_720_10s_10MB.mp4",
+        }
+        out_dict_with_video = processor.apply_chat_template(
+            messages, add_generation_prompt=True, tokenize=True, return_dict=True
+        )
+        self.assertListEqual(list(out_dict_with_video.keys()), ["input_ids", "attention_mask", "pixel_values_videos"])
+
+    @require_torch
+    @require_av
+    def test_chat_template_dict_torch(self):
+        processor = AutoProcessor.from_pretrained("llava-hf/llava-onevision-qwen2-7b-ov-hf")
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "video",
+                        "url": "https://test-videos.co.uk/vids/bigbuckbunny/mp4/h264/720/Big_Buck_Bunny_720_10s_10MB.mp4",
+                    },
+                    {"type": "text", "text": "What is shown in this video?"},
+                ],
+            },
+        ]
+
+        out_dict_tensors = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+        )
+        self.assertListEqual(list(out_dict_tensors.keys()), ["input_ids", "attention_mask", "pixel_values_videos"])
+        self.assertTrue(isinstance(out_dict_tensors["input_ids"], torch.Tensor))