llama3-llava-next-8b 全参微调报错

[liboaccn]

Reminder

• [x] I have read the above rules and searched the existing issues.

System Info

llamafactory 已经更新至最新版本

• llamafactory version: 0.9.2.dev0
• Platform: Linux-5.10.0-1.0.0.32-x86_64-with-glibc2.17
• Python version: 3.9.18
• PyTorch version: 2.1.0+cu118 (GPU)
• Transformers version: 4.45.0
• Datasets version: 2.19.2
• Accelerate version: 0.34.2
• PEFT version: 0.12.0
• TRL version: 0.9.6
• GPU type: NVIDIA A100-SXM4-40GB
• GPU number: 8
• GPU memory: 39.39GB
• DeepSpeed version: 0.15.4

Reproduction

全参数微调过程报错。 一共有2个报错，

第一个是image_sizes = iter(mm_inputs["image_sizes"].tolist()) 这里的mm_inputs 中没有image_sizes，image_sizes在后面的orig_height, orig_width = next(image_sizes) 会用到 第二个是 height, width = get_image_size(to_numpy_array(mm_inputs["pixel_values"][0][0])) 中的 mm_inputs["pixel_values"][0][0] 是不是应该是mm_inputs["pixel_values"][0]？？

raceback (most recent call last):
  File "/home/users/code/LLaMA-Factory/src/llamafactory/launcher.py", line 23, in <module>
    launch()
  File "/home/users/code/LLaMA-Factory/src/llamafactory/launcher.py", line 19, in launch
    run_exp()
  File "/home/users/code/LLaMA-Factory/src/llamafactory/train/tuner.py", line 93, in run_exp
    _training_function(config={"args": args, "callbacks": callbacks})
  File "/home/users/code/LLaMA-Factory/src/llamafactory/train/tuner.py", line 67, in _training_function
    run_sft(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
  File "/home/users/code/LLaMA-Factory/src/llamafactory/train/sft/workflow.py", line 51, in run_sft
    dataset_module = get_dataset(template, model_args, data_args, training_args, stage="sft", **tokenizer_module)
  File "/home/users/code/LLaMA-Factory/src/llamafactory/data/loader.py", line 325, in get_dataset
    dataset = _get_preprocessed_dataset(
  File "/home/users/code/LLaMA-Factory/src/llamafactory/data/loader.py", line 258, in _get_preprocessed_dataset
    dataset = dataset.map(
  File "/home/users/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 602, in wrapper
    out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
  File "/home/users/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 567, in wrapper
    out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
  File "/home/users/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 3156, in map
    for rank, done, content in Dataset._map_single(**dataset_kwargs):
  File "/home/users/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 3547, in _map_single
    batch = apply_function_on_filtered_inputs(
  File "/home/users/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 3416, in apply_function_on_filtered_inputs
    processed_inputs = function(*fn_args, *additional_args, **fn_kwargs)
  File "/home/users/code/LLaMA-Factory/src/llamafactory/data/processor/supervised.py", line 99, in preprocess_dataset
    input_ids, labels = self._encode_data_example(
  File "/home/users/code/LLaMA-Factory/src/llamafactory/data/processor/supervised.py", line 43, in _encode_data_example
    messages = self.template.mm_plugin.process_messages(prompt + response, images, videos, audios, self.processor)
  File "/home/users/code/LLaMA-Factory/src/llamafactory/data/mm_plugin.py", line 397, in process_messages
    height, width = get_image_size(to_numpy_array(mm_inputs["pixel_values"][0][0]))
  File "/home/users/miniconda3/lib/python3.9/site-packages/transformers/image_utils.py", line 295, in get_image_size
    channel_dim = infer_channel_dimension_format(image)
  File "/home/users/miniconda3/lib/python3.9/site-packages/transformers/image_utils.py", line 243, in infer_channel_dimension_format
    raise ValueError(f"Unsupported number of image dimensions: {image.ndim}")
ValueError: Unsupported number of image dimensions: 2
[2025-02-27 19:49:42,031] torch.distributed.elastic.mu

对应代码部分

class LlavaNextPlugin(BasePlugin):
    @override
    def process_messages(
        self,
        messages: Sequence[Dict[str, str]],
        images: Sequence["ImageInput"],
        videos: Sequence["VideoInput"],
        audios: Sequence["AudioInput"],
        processor: Optional["ProcessorMixin"],
    ) -> List[Dict[str, str]]:
        self._validate_input(processor, images, videos, audios)
        num_image_tokens = 0
        messages = deepcopy(messages)
        mm_inputs = self._get_mm_inputs(images, videos, audios, processor)
        if "pixel_values" in mm_inputs:
            image_sizes = iter(mm_inputs["image_sizes"].tolist())   <-----------------mm_inputs 中 image_sizes  应该不存在，所以这一行有报错
            height, width = get_image_size(to_numpy_array(mm_inputs["pixel_values"][0][0]))<---------------- 这是是不是应该是mm_inputs["pixel_values"][0]  而不是 mm_inputs["pixel_values"][0][0] 否则有上面的报错 ValueError: Unsupported number of image dimensions: 2




### Others

_No response_

[hiyouga]

应该是存在的https://github.com/huggingface/transformers/blob/v4.49.0/src/transformers/models/llava_next/processing_llava_next.py#L152-L153

[liboaccn]

报错的不是 transformers/models/llava_next/processing_llava_next.py 而是llamafactory 的代码

而是这 src/llamafactory/data/mm_plugin.py", line 396,

Traceback (most recent call last):
  File "/home/users/code/LLaMA-Factory/src/llamafactory/launcher.py", line 23, in <module>
    launch()
  File "/home/users/code/LLaMA-Factory/src/llamafactory/launcher.py", line 19, in launch
    run_exp()
  File "/home/users/code/LLaMA-Factory/src/llamafactory/train/tuner.py", line 93, in run_exp
    _training_function(config={"args": args, "callbacks": callbacks})
  File "/home/users/code/LLaMA-Factory/src/llamafactory/train/tuner.py", line 67, in _training_function
    run_sft(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
  File "/home/users/code/LLaMA-Factory/src/llamafactory/train/sft/workflow.py", line 51, in run_sft
    dataset_module = get_dataset(template, model_args, data_args, training_args, stage="sft", **tokenizer_module)
  File "/home/users/code/LLaMA-Factory/src/llamafactory/data/loader.py", line 325, in get_dataset
    dataset = _get_preprocessed_dataset(
  File "/home/users/code/LLaMA-Factory/src/llamafactory/data/loader.py", line 258, in _get_preprocessed_dataset
    dataset = dataset.map(
  File "/home/users/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 602, in wrapper
    out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
  File "/home/users/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 567, in wrapper
    out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
  File "/home/users/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 3156, in map
    for rank, done, content in Dataset._map_single(**dataset_kwargs):
  File "/home/users/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 3547, in _map_single
    batch = apply_function_on_filtered_inputs(
  File "/home/users/miniconda3/lib/python3.9/site-packages/datasets/arrow_dataset.py", line 3416, in apply_function_on_filtered_inputs
    processed_inputs = function(*fn_args, *additional_args, **fn_kwargs)
  File "/home/users/code/LLaMA-Factory/src/llamafactory/data/processor/supervised.py", line 99, in preprocess_dataset
    input_ids, labels = self._encode_data_example(
  File "/home/users/code/LLaMA-Factory/src/llamafactory/data/processor/supervised.py", line 43, in _encode_data_example
    messages = self.template.mm_plugin.process_messages(prompt + response, images, videos, audios, self.processor)
  File "/home/users/code/LLaMA-Factory/src/llamafactory/data/mm_plugin.py", line 396, in process_messages
    image_sizes = iter(mm_inputs["image_sizes"].tolist())
KeyError: 'image_sizes'

[Kuangdd01]

https://github.com/huggingface/transformers/blob/51083d1bac7905aa8316b75f7897bdd4e5302044/src/transformers/models/llava_next/image_processing_llava_next.py#L726C9-L728C10

  return BatchFeature(
      data={"pixel_values": processed_images, "image_sizes": image_sizes}, tensor_type=return_tensors
  )

经过了llava-next的image_processor之后应该会存在这个image_sizes key的，图片输入正确吗

[namingsohard]

I'm facing the same issue as you. How did you solve it? My training data runs normally when using llama-factory to train qwen2.5vl, and I'm not sure where the problem exactly lies.