[model] support LLaVA-OneVision-1.5

docs/source/Instruction/支持的模型和数据集.md

@@ -929,6 +929,10 @@
 |[AI-ModelScope/llava-next-72b](https://modelscope.cn/models/AI-ModelScope/llava-next-72b)|llava_next_qwen|llava_next_qwen|transformers>=4.42, av|✘|vision|[lmms-lab/llava-next-72b](https://huggingface.co/lmms-lab/llava-next-72b)|
 |[AI-ModelScope/llava-next-110b](https://modelscope.cn/models/AI-ModelScope/llava-next-110b)|llava_next_qwen|llava_next_qwen|transformers>=4.42, av|✘|vision|[lmms-lab/llava-next-110b](https://huggingface.co/lmms-lab/llava-next-110b)|
 |[AI-ModelScope/llama3-llava-next-8b](https://modelscope.cn/models/AI-ModelScope/llama3-llava-next-8b)|llama3_llava_next|llama3_llava_next|transformers>=4.42, av|✘|vision|[lmms-lab/llama3-llava-next-8b](https://huggingface.co/lmms-lab/llama3-llava-next-8b)|
+|[lmms-lab/LLaVA-OneVision-1.5-4B-Instruct](https://modelscope.cn/models/lmms-lab/LLaVA-OneVision-1.5-4B-Instruct)|llava_onevision1_5|llava_onevision1_5|transformers>=4.53, qwen_vl_utils>=0.0.6, decord|✘|vision, video|[lmms-lab/LLaVA-OneVision-1.5-4B-Instruct](https://huggingface.co/lmms-lab/LLaVA-OneVision-1.5-4B-Instruct)|
+|[lmms-lab/LLaVA-OneVision-1.5-8B-Instruct](https://modelscope.cn/models/lmms-lab/LLaVA-OneVision-1.5-8B-Instruct)|llava_onevision1_5|llava_onevision1_5|transformers>=4.53, qwen_vl_utils>=0.0.6, decord|✘|vision, video|[lmms-lab/LLaVA-OneVision-1.5-8B-Instruct](https://huggingface.co/lmms-lab/LLaVA-OneVision-1.5-8B-Instruct)|
+|[lmms-lab/LLaVA-OneVision-1.5-4B-Base](https://modelscope.cn/models/lmms-lab/LLaVA-OneVision-1.5-4B-Base)|llava_onevision1_5|llava_onevision1_5|transformers>=4.53, qwen_vl_utils>=0.0.6, decord|✘|vision, video|[lmms-lab/LLaVA-OneVision-1.5-4B-Base](https://huggingface.co/lmms-lab/LLaVA-OneVision-1.5-4B-Base)|
+|[lmms-lab/LLaVA-OneVision-1.5-8B-Base](https://modelscope.cn/models/lmms-lab/LLaVA-OneVision-1.5-8B-Base)|llava_onevision1_5|llava_onevision1_5|transformers>=4.53, qwen_vl_utils>=0.0.6, decord|✘|vision, video|[lmms-lab/LLaVA-OneVision-1.5-8B-Base](https://huggingface.co/lmms-lab/LLaVA-OneVision-1.5-8B-Base)|
 |[deepseek-ai/deepseek-vl-1.3b-chat](https://modelscope.cn/models/deepseek-ai/deepseek-vl-1.3b-chat)|deepseek_vl|deepseek_vl|-|✘|vision|[deepseek-ai/deepseek-vl-1.3b-chat](https://huggingface.co/deepseek-ai/deepseek-vl-1.3b-chat)|
 |[deepseek-ai/deepseek-vl-7b-chat](https://modelscope.cn/models/deepseek-ai/deepseek-vl-7b-chat)|deepseek_vl|deepseek_vl|-|✘|vision|[deepseek-ai/deepseek-vl-7b-chat](https://huggingface.co/deepseek-ai/deepseek-vl-7b-chat)|
 |[deepseek-ai/deepseek-vl2-tiny](https://modelscope.cn/models/deepseek-ai/deepseek-vl2-tiny)|deepseek_vl2|deepseek_vl2|transformers<4.42|&#x2718;|vision|[deepseek-ai/deepseek-vl2-tiny](https://huggingface.co/deepseek-ai/deepseek-vl2-tiny)|

docs/source_en/Instruction/Supported-models-and-datasets.md

@@ -911,6 +911,10 @@ The table below introduces the models integrated with ms-swift:
 |[llava-hf/llava-v1.6-vicuna-13b-hf](https://modelscope.cn/models/llava-hf/llava-v1.6-vicuna-13b-hf)|llava1_6_vicuna_hf|llava1_6_vicuna_hf|transformers>=4.39|✘|vision|[llava-hf/llava-v1.6-vicuna-13b-hf](https://huggingface.co/llava-hf/llava-v1.6-vicuna-13b-hf)|
 |[llava-hf/llava-v1.6-34b-hf](https://modelscope.cn/models/llava-hf/llava-v1.6-34b-hf)|llava1_6_yi_hf|llava1_6_yi_hf|transformers>=4.39|✘|vision|[llava-hf/llava-v1.6-34b-hf](https://huggingface.co/llava-hf/llava-v1.6-34b-hf)|
 |[llava-hf/llama3-llava-next-8b-hf](https://modelscope.cn/models/llava-hf/llama3-llava-next-8b-hf)|llama3_llava_next_hf|llama3_llava_next_hf|transformers>=4.39|✘|vision|[llava-hf/llama3-llava-next-8b-hf](https://huggingface.co/llava-hf/llama3-llava-next-8b-hf)|
+|[lmms-lab/LLaVA-OneVision-1.5-4B-Instruct](https://modelscope.cn/models/lmms-lab/LLaVA-OneVision-1.5-4B-Instruct)|llava_onevision1_5|llava_onevision1_5|transformers>=4.53, qwen_vl_utils>=0.0.6, decord|✘|vision, video|[lmms-lab/LLaVA-OneVision-1.5-4B-Instruct](https://huggingface.co/lmms-lab/LLaVA-OneVision-1.5-4B-Instruct)|
+|[lmms-lab/LLaVA-OneVision-1.5-8B-Instruct](https://modelscope.cn/models/lmms-lab/LLaVA-OneVision-1.5-8B-Instruct)|llava_onevision1_5|llava_onevision1_5|transformers>=4.53, qwen_vl_utils>=0.0.6, decord|✘|vision, video|[lmms-lab/LLaVA-OneVision-1.5-8B-Instruct](https://huggingface.co/lmms-lab/LLaVA-OneVision-1.5-8B-Instruct)|
+|[lmms-lab/LLaVA-OneVision-1.5-4B-Base](https://modelscope.cn/models/lmms-lab/LLaVA-OneVision-1.5-4B-Base)|llava_onevision1_5|llava_onevision1_5|transformers>=4.53, qwen_vl_utils>=0.0.6, decord|✘|vision, video|[lmms-lab/LLaVA-OneVision-1.5-4B-Base](https://huggingface.co/lmms-lab/LLaVA-OneVision-1.5-4B-Base)|
+|[lmms-lab/LLaVA-OneVision-1.5-8B-Base](https://modelscope.cn/models/lmms-lab/LLaVA-OneVision-1.5-8B-Base)|llava_onevision1_5|llava_onevision1_5|transformers>=4.53, qwen_vl_utils>=0.0.6, decord|✘|vision, video|[lmms-lab/LLaVA-OneVision-1.5-8B-Base](https://huggingface.co/lmms-lab/LLaVA-OneVision-1.5-8B-Base)|
 |[llava-hf/llava-next-72b-hf](https://modelscope.cn/models/llava-hf/llava-next-72b-hf)|llava_next_qwen_hf|llava_next_qwen_hf|transformers>=4.39|✘|vision|[llava-hf/llava-next-72b-hf](https://huggingface.co/llava-hf/llava-next-72b-hf)|
 |[llava-hf/llava-next-110b-hf](https://modelscope.cn/models/llava-hf/llava-next-110b-hf)|llava_next_qwen_hf|llava_next_qwen_hf|transformers>=4.39|✘|vision|[llava-hf/llava-next-110b-hf](https://huggingface.co/llava-hf/llava-next-110b-hf)|
 |[llava-hf/LLaVA-NeXT-Video-7B-DPO-hf](https://modelscope.cn/models/llava-hf/LLaVA-NeXT-Video-7B-DPO-hf)|llava_next_video_hf|llava_next_video_hf|transformers>=4.42, av|✘|video|[llava-hf/LLaVA-NeXT-Video-7B-DPO-hf](https://huggingface.co/llava-hf/LLaVA-NeXT-Video-7B-DPO-hf)|

swift/llm/model/constant.py

@@ -222,6 +222,7 @@ class MLLMModelType:
     llava1_6_yi = 'llava1_6_yi'
     llava_next_qwen = 'llava_next_qwen'
     llama3_llava_next = 'llama3_llava_next'
+    llava_onevision1_5 = 'llava_onevision1_5'
 
     deepseek_vl = 'deepseek_vl'
     deepseek_vl2 = 'deepseek_vl2'

swift/llm/model/model/llava.py

@@ -5,6 +5,7 @@
 from typing import Any, Dict
 
 from transformers import AutoConfig
+from transformers.dynamic_module_utils import get_class_from_dynamic_module
 
 from swift.llm import TemplateType
 from ..constant import MLLMModelType
@@ -389,3 +390,32 @@ def _new_forward(*args, **kwargs):
         requires=['transformers>=4.42', 'av'],
         tags=['vision'],
         model_arch=None))
+
+
+def get_model_tokenizer_llava_onevision1_5(model_dir, *args, **kwargs):
+    model_cls = get_class_from_dynamic_module('modeling_llavaonevision1_5.LLaVAOneVision1_5_ForConditionalGeneration',
+                                              model_dir)
+    model_cls._no_split_modules = ['LLaVAOneVision1_5_DecoderLayer', 'RiceBlock']
+    model, processor = get_model_tokenizer_multimodal(model_dir, *args, **kwargs)
+    model.config.vision_start_token_id = 151652
+    return model, processor
+
+
+register_model(
+    ModelMeta(
+        MLLMModelType.llava_onevision1_5,
+        [
+            ModelGroup([
+                Model('lmms-lab/LLaVA-OneVision-1.5-4B-Instruct', 'lmms-lab/LLaVA-OneVision-1.5-4B-Instruct'),
+                Model('lmms-lab/LLaVA-OneVision-1.5-8B-Instruct', 'lmms-lab/LLaVA-OneVision-1.5-8B-Instruct'),
+                Model('lmms-lab/LLaVA-OneVision-1.5-4B-Base', 'lmms-lab/LLaVA-OneVision-1.5-4B-Base'),
+                Model('lmms-lab/LLaVA-OneVision-1.5-8B-Base', 'lmms-lab/LLaVA-OneVision-1.5-8B-Base'),
+            ], ),
+        ],
+        TemplateType.llava_onevision1_5,
+        get_model_tokenizer_llava_onevision1_5,
+        architectures=['LLaVAOneVision1_5_ForConditionalGeneration'],
+        model_arch=ModelArch.llava_onevision1_5,
+        requires=['transformers>=4.53.0', 'qwen_vl_utils'],
+        tags=['vision'],
+    ))

swift/llm/model/model_arch.py

@@ -46,6 +46,7 @@ class MLLMModelArch:
     llava_hf = 'llava_hf'
     llava_hf_legacy = 'llava_hf_legacy'  # transformers<4.52
     llava_next_video_hf = 'llava_next_video_hf'
+    llava_onevision1_5 = 'llava_onevision1_5'
 
     llava_llama = 'llava_llama'
     llava_mistral = 'llava_mistral'
@@ -705,6 +706,14 @@ def register_model_arch(model_arch: ModelKeys, *, exist_ok: bool = False) -> Non
     language_model='model',
 ))
 
+register_model_arch(
+    MultiModelKeys(
+        MLLMModelArch.llava_onevision1_5,
+        language_model='model.language_model',
+        aligner='model.visual.merger',
+        vision_tower='model.visual',
+    ))
+
 
 def get_model_arch(arch_name: Optional[str]) -> Optional[MultiModelKeys]:
     return MODEL_ARCH_MAPPING.get(arch_name)

swift/llm/template/constant.py

@@ -152,6 +152,7 @@ class MLLMTemplateType:
     llava1_6_yi = 'llava1_6_yi'
     llava_next_qwen = 'llava_next_qwen'
     llama3_llava_next = 'llama3_llava_next'
+    llava_onevision1_5 = 'llava_onevision1_5'
 
     yi_vl = 'yi_vl'
 

swift/llm/template/template/llava.py

@@ -6,6 +6,7 @@
 import transformers
 from packaging import version
 
+from swift.utils import get_env_args
 from ..base import Template
 from ..constant import MLLMTemplateType
 from ..register import TemplateMeta, register_template
@@ -307,3 +308,101 @@ def _data_collator(self, batch: List[Dict[str, Any]], *, padding_to: Optional[in
     ))
 
 register_template(QwenTemplateMeta(MLLMTemplateType.llava_next_qwen, template_cls=LLavaTemplate))
+
+
+class LLavaOneVision1_5Template(Template):
+    image_token_id = 151655
+    video_token_id = 151656
+    placeholder_tokens = ['<|image_pad|>', '<|video_pad|>']
+    use_model = True
+    support_padding_free = True
+
+    def init_env_args(self):
+        super().init_env_args()
+        self.bbox_format = get_env_args('QWENVL_BBOX_FORMAT', str, 'legacy')
+
+    def replace_tag(self, media_type: Literal['image', 'video', 'audio'], index: int,
+                    inputs: StdTemplateInputs) -> List[Context]:
+        from qwen_vl_utils import fetch_image, fetch_video
+        assert media_type in {'image', 'video'}
+        if media_type == 'image':
+            inputs.images[index] = fetch_image({'image': inputs.images[index]})
+            if self.mode == 'lmdeploy':
+                return ['<|vision_start|>', [-100], '<|vision_end|>']
+            else:
+                return ['<|vision_start|><|image_pad|><|vision_end|>']
+        else:
+            video = inputs.videos[index]
+            video, video_kwargs = fetch_video({'video': video}, return_video_sample_fps=True)
+            inputs.mm_processor_kwargs.setdefault('fps', []).append(video_kwargs)
+            tokens = ['<|vision_start|><|video_pad|><|vision_end|>']
+            if isinstance(video, torch.Tensor):
+                video = video.to(torch.uint8)
+            inputs.videos[index] = video
+            return tokens
+
+    def replace_ref(self, ref: str, index: int, inputs: StdTemplateInputs) -> List[Context]:
+        if self.bbox_format == 'legacy':
+            return [f'<|object_ref_start|>{ref}<|object_ref_end|>']
+        else:
+            return [ref]
+
+    def replace_bbox(self, bbox: List[int], index: int, inputs: StdTemplateInputs) -> List[Context]:
+        if self.bbox_format == 'legacy':
+            return [f'<|box_start|>{self._get_bbox_str(bbox)}<|box_end|>']
+        else:
+            return [str(bbox)]
+
+    def _encode(self, inputs: StdTemplateInputs) -> Dict[str, Any]:
+        encoded = super()._encode(inputs)
+        processor = self.processor
+        input_ids = encoded['input_ids']
+        labels = encoded['labels']
+        loss_scale = encoded.get('loss_scale', None)
+        for media_type in ['images', 'videos']:
+            mm_data = getattr(inputs, media_type)
+            if mm_data:
+                if media_type == 'images':
+                    media_token = self.image_token_id
+                    media_inputs = processor.image_processor(images=mm_data, return_tensors='pt', do_resize=False)
+                    media_grid_thw = media_inputs['image_grid_thw']
+                else:
+                    kwargs = {}
+                    if hasattr(processor, 'video_processor'):
+                        processor_func = processor.video_processor
+                    else:
+                        processor_func = processor.image_processor
+                        kwargs['images'] = None
+                    media_inputs = processor_func(videos=mm_data, return_tensors='pt', do_resize=False, **kwargs)
+                    media_grid_thw = media_inputs['video_grid_thw']
+                    media_token = self.video_token_id
+                idx_list = findall(input_ids, media_token)
+                merge_length = processor.image_processor.merge_size**2
+
+                def _get_new_tokens(i):
+                    token_len = (media_grid_thw[i].prod() // merge_length)
+                    return [media_token] * token_len
+
+                input_ids, labels, loss_scale = self._extend_tokens(input_ids, labels, loss_scale, idx_list,
+                                                                    _get_new_tokens)
+                encoded.update(media_inputs)
+
+        encoded['input_ids'] = input_ids
+        encoded['labels'] = labels
+        encoded['loss_scale'] = loss_scale
+        return encoded
+
+    def _post_encode(self, model, inputs: Dict[str, Any]) -> Dict[str, Any]:
+        if not self.is_training:
+            return inputs
+        input_ids = inputs['input_ids']
+        base_model = self.get_base_model(model)
+        if hasattr(base_model.model, 'embed_tokens'):
+            inputs_embeds = base_model.model.embed_tokens(input_ids)
+        else:
+            inputs_embeds = base_model.model.language_model.embed_tokens(input_ids)
+        inputs_embeds = self._get_inputs_embeds_hf(inputs_embeds, inputs, model.visual, self.processor, model.config)
+        return {'inputs_embeds': inputs_embeds}
+
+
+register_template(QwenTemplateMeta(MLLMTemplateType.llava_onevision1_5, template_cls=LLavaOneVision1_5Template))