transformers/tests/models/llava_onevision/test_modeling_llava_onevision.py

# coding=utf-8
# Copyright 2024 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 Llava-NeXT model."""

import unittest

import numpy as np
import requests
from huggingface_hub import hf_hub_download

from transformers import (
    AutoProcessor,
    LlavaOnevisionConfig,
    LlavaOnevisionForConditionalGeneration,
    is_torch_available,
    is_vision_available,
)
from transformers.testing_utils import (
    cleanup,
    require_bitsandbytes,
    require_torch,
    slow,
    torch_device,
)

from ...generation.test_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import (
    ModelTesterMixin,
    _config_zero_init,
    floats_tensor,
    ids_tensor,
)


if is_torch_available():
    import torch

else:
    is_torch_greater_or_equal_than_2_0 = False

if is_vision_available():
    from PIL import Image


class LlavaOnevisionVisionText2TextModelTester:
    def __init__(
        self,
        parent,
        ignore_index=-100,
        image_token_index=1,
        projector_hidden_act="gelu",
        seq_length=7,
        vision_feature_select_strategy="full",
        vision_feature_layer=-1,
        text_config={
            "model_type": "qwen2",
            "seq_length": 7,
            "is_training": True,
            "use_input_mask": True,
            "use_token_type_ids": False,
            "use_labels": True,
            "vocab_size": 99,
            "hidden_size": 32,
            "num_hidden_layers": 2,
            "num_attention_heads": 4,
            "num_key_value_heads": 4,
            "intermediate_size": 37,
            "hidden_act": "gelu",
            "hidden_dropout_prob": 0.1,
            "attention_probs_dropout_prob": 0.1,
            "max_position_embeddings": 580,
            "type_vocab_size": 16,
            "type_sequence_label_size": 2,
            "initializer_range": 0.02,
            "num_labels": 3,
            "num_choices": 4,
            "pad_token_id": 0,
        },
        is_training=True,
        vision_config={
            "image_size": 16,
            "patch_size": 8,
            "num_channels": 3,
            "is_training": True,
            "hidden_size": 32,
            "projection_dim": 32,
            "num_hidden_layers": 2,
            "num_attention_heads": 4,
            "intermediate_size": 37,
            "dropout": 0.1,
            "attention_dropout": 0.1,
            "initializer_range": 0.02,
        },
    ):
        self.parent = parent
        self.ignore_index = ignore_index
        self.image_token_index = image_token_index
        self.projector_hidden_act = projector_hidden_act
        self.vision_feature_select_strategy = vision_feature_select_strategy
        self.vision_feature_layer = vision_feature_layer
        self.text_config = text_config
        self.vision_config = vision_config
        self.pad_token_id = text_config["pad_token_id"]
        self.num_image_tokens = 10
        self.seq_length = seq_length + self.num_image_tokens

        self.num_hidden_layers = text_config["num_hidden_layers"]
        self.vocab_size = text_config["vocab_size"]
        self.hidden_size = text_config["hidden_size"]
        self.num_attention_heads = text_config["num_attention_heads"]
        self.is_training = is_training

        self.batch_size = 3
        self.num_channels = 3
        self.image_size = 30
        self.image_grid_pinpoints = [[16, 16]]

    def get_config(self):
        return LlavaOnevisionConfig(
            text_config=self.text_config,
            vision_config=self.vision_config,
            ignore_index=self.ignore_index,
            image_token_index=self.image_token_index,
            projector_hidden_act=self.projector_hidden_act,
            vision_feature_select_strategy=self.vision_feature_select_strategy,
            vision_feature_layer=self.vision_feature_layer,
            image_grid_pinpoints=self.image_grid_pinpoints,
        )

    def prepare_config_and_inputs(self):
        pixel_values = floats_tensor(
            [
                self.batch_size,
                3,
                self.vision_config["num_channels"],
                self.vision_config["image_size"],
                self.vision_config["image_size"],
            ]
        )
        config = self.get_config()

        return config, pixel_values

    def prepare_config_and_inputs_for_common(self):
        config_and_inputs = self.prepare_config_and_inputs()
        config, pixel_values = config_and_inputs
        input_ids = ids_tensor([self.batch_size, self.seq_length], config.text_config.vocab_size - 2) + 2
        attention_mask = torch.ones(input_ids.shape, dtype=torch.long).to(torch_device)

        input_ids[input_ids == config.image_token_index] = self.pad_token_id
        input_ids[:, : self.num_image_tokens] = config.image_token_index

        labels = torch.zeros((self.batch_size, self.seq_length), dtype=torch.long, device=torch_device)
        labels[:, : self.num_image_tokens] == self.ignore_index

        inputs_dict = {
            "pixel_values": pixel_values,
            "image_sizes": torch.tensor([[45, 45]] * self.batch_size),
            "input_ids": input_ids,
            "attention_mask": attention_mask,
            "labels": labels,
        }
        return config, inputs_dict

    def create_and_check_llava_onevision_model_fp16_forward(
        self, config, input_ids, pixel_values, attention_mask, image_sizes
    ):
        model = LlavaOnevisionForConditionalGeneration(config=config)
        model.to(torch_device)
        model.half()
        model.eval()
        logits = model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            image_sizes=image_sizes,
            pixel_values=pixel_values.to(torch.bfloat16),
            return_dict=True,
        )["logits"]
        self.parent.assertFalse(torch.isnan(logits).any().item())

    def create_and_check_llava_onevision_model_fp16_autocast_forward(
        self, config, input_ids, pixel_values, attention_mask, image_sizes
    ):
        config.torch_dtype = torch.float16
        model = LlavaOnevisionForConditionalGeneration(config=config)
        model.to(torch_device)
        model.eval()
        with torch.autocast(device_type="cuda", dtype=torch.float16):
            logits = model(
                input_ids=input_ids,
                attention_mask=attention_mask,
                image_sizes=image_sizes,
                pixel_values=pixel_values.to(torch.bfloat16),
                return_dict=True,
            )["logits"]
        self.parent.assertFalse(torch.isnan(logits).any().item())


@require_torch
class LlavaOnevisionForConditionalGenerationModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
    """
    Model tester for `LlavaOnevisionForConditionalGeneration`.
    """

    all_model_classes = (LlavaOnevisionForConditionalGeneration,) if is_torch_available() else ()
    all_generative_model_classes = (LlavaOnevisionForConditionalGeneration,) if is_torch_available() else ()
    pipeline_model_mapping = (
        {"image-text-to-text": LlavaOnevisionForConditionalGeneration} if is_torch_available() else {}
    )
    test_pruning = False
    test_head_masking = False
    _is_composite = True

    def setUp(self):
        self.model_tester = LlavaOnevisionVisionText2TextModelTester(self)
        self.config_tester = ConfigTester(self, config_class=LlavaOnevisionConfig, has_text_modality=False)

    def test_initialization(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        configs_no_init = _config_zero_init(config)
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            for name, param in model.named_parameters():
                # LLaVa Onevision has SigLIP backbone which init weights differently from CLIP
                if "image_newline" in name or "vision_tower" in name:
                    continue
                elif param.requires_grad:
                    self.assertIn(
                        ((param.data.mean() * 1e9).round() / 1e9).item(),
                        [0.0, 1.0],
                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
                    )

    # overwrite inputs_embeds tests because we need to delete "pixel values" for LVLMs
    def test_inputs_embeds(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            inputs = self._prepare_for_class(inputs_dict, model_class)

            input_ids = inputs["input_ids"]
            del inputs["input_ids"]
            del inputs["pixel_values"]

            wte = model.get_input_embeddings()
            inputs["inputs_embeds"] = wte(input_ids)

            with torch.no_grad():
                model(**inputs)

    # overwrite inputs_embeds tests because we need to delete "pixel values" for LVLMs
    # while some other models require pixel_values to be present
    def test_inputs_embeds_matches_input_ids(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            inputs = self._prepare_for_class(inputs_dict, model_class)
            input_ids = inputs["input_ids"]
            del inputs["input_ids"]
            del inputs["pixel_values"]

            inputs_embeds = model.get_input_embeddings()(input_ids)

            with torch.no_grad():
                out_ids = model(input_ids=input_ids, **inputs)[0]
                out_embeds = model(inputs_embeds=inputs_embeds, **inputs)[0]
            self.assertTrue(torch.allclose(out_embeds, out_ids))

    @unittest.skip(
        reason="This architecure seem to not compute gradients properly when using GC, SiglipVisionModel does not support standalone training"
    )
    def test_training_gradient_checkpointing(self):
        pass

    @unittest.skip(
        reason="This architecure seem to not compute gradients properly when using GC, SiglipVisionModel does not support standalone training"
    )
    def test_training_gradient_checkpointing_use_reentrant(self):
        pass

    @unittest.skip(
        reason="This architecure seem to not compute gradients properly when using GC, SiglipVisionModel does not support standalone training"
    )
    def test_training_gradient_checkpointing_use_reentrant_false(self):
        pass

    @unittest.skip("VLMs can't do assisted decoding yet!")
    def test_assisted_decoding_with_num_logits_to_keep(self):
        pass

    @unittest.skip("FlashAttention only support fp16 and bf16 data type")
    def test_flash_attn_2_fp32_ln(self):
        pass

    @unittest.skip(
        "VLMs need lots of steps to prepare images/mask correctly to get pad-free inputs. Can be tested as part of LLM test"
    )
    def test_flash_attention_2_padding_matches_padding_free_with_position_ids(self):
        pass


@require_torch
class LlavaOnevisionForConditionalGenerationIntegrationTest(unittest.TestCase):
    def setUp(self):
        self.processor = AutoProcessor.from_pretrained(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf", padding_side="left"
        )
        image_file = hf_hub_download(
            repo_id="raushan-testing-hf/images_test", filename="llava_v1_5_radar.jpg", repo_type="dataset"
        )
        video_file = hf_hub_download(
            repo_id="raushan-testing-hf/videos-test", filename="video_demo.npy", repo_type="dataset"
        )
        self.image = Image.open(image_file)
        self.video = np.load(video_file)
        self.prompt_image = "user\n<image>\nWhat do you see in this image?<|im_end|>\n<|im_start|>assistant\n"
        self.prompt_video = "user\n<video>\nWhat do you see in this video?<|im_end|>\n<|im_start|>assistant\n"

    def tearDown(self):
        cleanup(torch_device, gc_collect=True)

    @slow
    @require_bitsandbytes
    def test_small_model_integration_test(self):
        model = LlavaOnevisionForConditionalGeneration.from_pretrained(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf", torch_dtype="float16", device_map=torch_device
        )

        inputs = self.processor(images=self.image, text=self.prompt_image, return_tensors="pt").to(
            torch_device, torch.float16
        )
        self.assertTrue(inputs.input_ids.shape[1] == 6567)  # should expand num-image-tokens times
        self.assertTrue(inputs.pixel_values.shape == torch.Size([1, 10, 3, 384, 384]))
        self.assertTrue(inputs.image_sizes.tolist() == [[899, 1024]])

        # verify single forward pass
        inputs = inputs.to(torch_device)
        with torch.no_grad():
            output = model(**inputs)

        expected_slice = torch.tensor(
            [[-12.3125, -14.5625, -12.8750], [3.4023, 5.0508, 9.5469], [3.5762, 4.4922, 7.8906]],
            dtype=torch.float32,
            device=torch_device,
        )
        self.assertTrue(torch.allclose(output.logits[0, :3, :3], expected_slice, atol=1e-3))

        # verify generation
        output = model.generate(**inputs, max_new_tokens=100)
        EXPECTED_DECODED_TEXT = 'user\n\nWhat do you see in this image?\nassistant\nThe image is a radar chart that compares the performance of different models in a specific task, likely related to natural language processing or machine learning. The chart is divided into several axes, each representing a different model or method. The models are color-coded and labeled with their respective names. The axes are labeled with terms such as "VQA," "GQA," "MQA," "VIZ," "TextVQA," "SQA-IMG," and "MQE." The radar chart shows'  # fmt: skip

        self.assertEqual(
            self.processor.decode(output[0], skip_special_tokens=True),
            EXPECTED_DECODED_TEXT,
        )

    @slow
    @require_bitsandbytes
    def test_small_model_integration_test_batch(self):
        model = LlavaOnevisionForConditionalGeneration.from_pretrained(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf", torch_dtype="float16", device_map=torch_device
        )

        inputs = self.processor(
            text=[self.prompt_image, self.prompt_video],
            images=self.image,
            videos=self.video,
            return_tensors="pt",
            padding=True,
        ).to(torch_device, torch.float16)

        output = model.generate(**inputs, max_new_tokens=20)

        EXPECTED_DECODED_TEXT = ['user\n\nWhat do you see in this image?\nassistant\nThe image is a radar chart that compares the performance of different models in a specific task, likely related', 'user\n\nWhat do you see in this video?\nassistant\nA child wearing a light blue sleeveless top and pink pants is seen sitting on a bed, eng']  # fmt: skip

        self.assertEqual(
            self.processor.batch_decode(output, skip_special_tokens=True),
            EXPECTED_DECODED_TEXT,
        )

    @slow
    @require_bitsandbytes
    def test_small_model_integration_test_video(self):
        # related to (#29835)
        model = LlavaOnevisionForConditionalGeneration.from_pretrained(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf",
            torch_dtype="float16",
            device_map=torch_device,
        )

        inputs = self.processor(text=self.prompt_video, videos=self.video, return_tensors="pt").to(
            torch_device, torch.float16
        )

        # verify generation
        output = model.generate(**inputs, max_new_tokens=40)
        EXPECTED_DECODED_TEXT = 'user\n\nWhat do you see in this video?\nassistant\nA child wearing a light blue sleeveless top and pink pants is seen sitting on a bed, engrossed in reading a book.'  # fmt: skip

        self.assertEqual(
            self.processor.decode(output[0], skip_special_tokens=True),
            EXPECTED_DECODED_TEXT,
        )

    @slow
    @require_bitsandbytes
    def test_small_model_integration_test_multi_image(self):
        # related to (#29835)
        model = LlavaOnevisionForConditionalGeneration.from_pretrained(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf",
            torch_dtype="float16",
            device_map=torch_device,
        )

        url = "https://www.ilankelman.org/stopsigns/australia.jpg"
        image = Image.open(requests.get(url, stream=True).raw)
        prompt = (
            "user\n<image><image>\nWhat is the difference between these images?<|im_end|>\n<|im_start|>assistant\n"
        )
        inputs = self.processor(text=prompt, images=[self.image, image], return_tensors="pt").to(
            torch_device, torch.float16
        )

        # verify generation
        output = model.generate(**inputs, max_new_tokens=40)
        EXPECTED_DECODED_TEXT = "user\n\nWhat is the difference between these images?\nassistant\nThe images you've provided appear to be related to a graphical representation of a radar chart, which is a type of data visualization used to show the distribution of a particular variable across a geographic area. The"  # fmt: skip

        self.assertEqual(
            self.processor.decode(output[0], skip_special_tokens=True),
            EXPECTED_DECODED_TEXT,
        )

    @slow
    @require_bitsandbytes
    def test_small_model_integration_test_multi_video(self):
        # related to (#29835)
        model = LlavaOnevisionForConditionalGeneration.from_pretrained(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf",
            torch_dtype="float16",
            device_map=torch_device,
        )

        prompt = "user\n<video><video>\nAre these videos identical?<|im_end|>\n<|im_start|>assistant\n"
        inputs = self.processor(text=prompt, videos=[self.video, self.video], return_tensors="pt").to(
            torch_device, torch.float16
        )

        # verify generation
        output = model.generate(**inputs, max_new_tokens=40)
        EXPECTED_DECODED_TEXT = "user\n\nAre these videos identical?\nassistant\nNo, the video is not identical; it shows slight variations in the child's actions and the background."  # fmt: skip

        self.assertEqual(
            self.processor.decode(output[0], skip_special_tokens=True),
            EXPECTED_DECODED_TEXT,
        )

    @slow
    @require_bitsandbytes
    def test_small_model_integration_test_batch_different_resolutions(self):
        model = LlavaOnevisionForConditionalGeneration.from_pretrained(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf", torch_dtype="float16", device_map=torch_device
        )

        url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        lowres_url = "https://4.img-dpreview.com/files/p/TS560x560~forums/56876524/03975b28741443319e9a94615e35667e"
        cats_image = Image.open(requests.get(url, stream=True).raw)
        lowres_img = Image.open(requests.get(lowres_url, stream=True).raw)

        inputs = self.processor(
            text=[self.prompt_image, self.prompt_image],
            images=[lowres_img, cats_image],
            return_tensors="pt",
            padding=True,
        ).to(torch_device, torch.float16)

        # verify generation
        output = model.generate(**inputs, max_new_tokens=50)
        EXPECTED_DECODED_TEXT = ['user\n\nWhat do you see in this image?\nassistant\nThe image shows a scene from a wildlife camera, likely a security camera, capturing a moment in a natural setting. It features two deer, one larger and one smaller, grazing on the grass. The environment is foggy, suggesting early morning or late', 'user\n\nWhat do you see in this image?\nassistant\nIn the tranquil setting of this image, two cats are enjoying a peaceful nap on a vibrant pink blanket. The cat on the left, with its gray and black striped fur, is lying on its side, its head comfortably resting on the blanket. Its']  # fmt: skip
        self.assertEqual(
            self.processor.batch_decode(output, skip_special_tokens=True),
            EXPECTED_DECODED_TEXT,
        )

    @slow
    @require_bitsandbytes
    def test_small_model_integration_test_batch_matches_single(self):
        model = LlavaOnevisionForConditionalGeneration.from_pretrained(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf",
            torch_dtype="float16",
            device_map=torch_device,
        )

        url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        lowres_url = "https://4.img-dpreview.com/files/p/TS560x560~forums/56876524/03975b28741443319e9a94615e35667e"
        cats_image = Image.open(requests.get(url, stream=True).raw)
        lowres_img = Image.open(requests.get(lowres_url, stream=True).raw)

        inputs_batched = self.processor(
            text=[self.prompt_image, self.prompt_image],
            images=[lowres_img, cats_image],
            return_tensors="pt",
            padding=True,
        ).to(torch_device, torch.float16)

        inputs_single = self.processor(
            text=self.prompt_image, images=lowres_img, return_tensors="pt", padding=True
        ).to(torch_device, torch.float16)

        # verify generation
        output_batched = model.generate(**inputs_batched, max_new_tokens=50)
        output_single = model.generate(**inputs_single, max_new_tokens=50)

        self.assertEqual(
            self.processor.decode(output_batched[0], skip_special_tokens=True),
            self.processor.decode(output_single[0], skip_special_tokens=True),
        )