deepseek-ocr ft failed when batch_size > 1

[happened]

根据 #6238
Describe the bug
What the bug is, and how to reproduce, better with screenshots(描述bug以及复现过程，最好有截图)

自定义数据集的情况下batch-size超过1的时候报错
File "/app/ms-swift/swift/llm/template/template/deepseek.py", line 373, in _data_collator_mm_data
images_seq_mask = self.concat_tensor(batch, 'images_seq_mask', 0)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/app/ms-swift/swift/llm/template/base.py", line 1461, in concat_tensor
return torch.concat(res, dim=dim) if res else None
^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: Sizes of tensors must match except in dimension 0. Expected size 299 but got size 292 for tensor number 1 in the list.

Your hardware and system info
Write your system info like CUDA version/system/GPU/torch version here(在这里给出硬件信息和系统信息，如CUDA版本，系统，GPU型号和torch版本等)

Additional context
Add any other context about the problem here(在这里补充其他信息)

[qinb]

因为images_seq_mask = self.concat_tensor(batch, 'images_seq_mask', 0)，这块是每个图片的valid tokens，所以会报错。这行注释掉，换成list保存

[happened]

使用gemini大法结合回复可以跑下去，但最终实验结果怪怪的

def _data_collator_mm_data(self, batch: List[Dict[str, Any]]) -> Dict[str, Any]:
        res = super()._data_collator_mm_data(batch)
        images = self.gather_list(batch, 'images')
        if images:
            res['images'] = images
        #images_seq_mask = self.concat_tensor(batch, 'images_seq_mask', 0)
        masks_2d = [
            b['images_seq_mask'] for b in batch if 'images_seq_mask' in b and b['images_seq_mask'] is not None
        ]

        if masks_2d:
            # 2. 将它们 squeeze 成一维张量的列表，形状变为 [seq_len]
            masks_1d = [m.squeeze(0) for m in masks_2d]
            
            # 3. 使用 pad_sequence 将这些不同长度的1D掩码填充成一个规整的2D批次张量
            images_seq_mask_padded = pad_sequence(masks_1d, batch_first=True, padding_value=False)
            res['images_seq_mask'] = images_seq_mask_padded
        
        # --- 整理结束 ---

        images_spatial_crop = self.concat_tensor(batch, 'images_spatial_crop', 0)
        if images_spatial_crop is not None:
            res['images_spatial_crop'] = images_spatial_crop
        return res

[TidalWish]

我自己修改了_data_collator_mm_data函数，将同一batch中的images_seq_mask按最长的用False进行padding，在我完成之后发现可以使用一些已有的函数来简化代码，最终我的_data_collator_mm_data函数如下：

    def _data_collator_mm_data(self, batch: List[Dict[str, Any]]) -> Dict[str, Any]:
        res = super()._data_collator_mm_data(batch)
        images = self.gather_list(batch, 'images')
        if images:
            res['images'] = images
        images_seq_mask_list = self.gather_list(batch, 'images_seq_mask')
        images_seq_mask = self._pad_sequence(images_seq_mask_list, False)
        images_spatial_crop = self.concat_tensor(batch, 'images_spatial_crop', 0)
        if images_seq_mask is not None:
            res['images_seq_mask'] = images_seq_mask
        if images_spatial_crop is not None:
            res['images_spatial_crop'] = images_spatial_crop
        return res

目前训练loss正在稳步下降。但修改的正确性还有待验证。 @Jintao-Huang