请问edit数据集需要设置is_mandatory: true 吗

```
t2i_pretrain:
dataset_names:
- t2i
image_transform_args:
image_stride: 16
max_image_size: 1024
min_image_size: 512
is_mandatory: true
num_used_data: # The sum should be larger that NUM_GPUS x NUM_WORKERS
- 10
weight: 1

unified_edit:
dataset_names:
- seedxedit_multi
image_transform_args:
image_stride: 16
max_image_size: 1024
min_image_size: 512
vit_image_transform_args:
image_stride: 14
max_image_size: 518
min_image_size: 224
is_mandatory: false
num_used_data:
- 500
weight: 1

vlm_sft:
dataset_names:
- llava_ov
image_transform_args:
image_stride: 14
max_image_size: 980
min_image_size: 378
max_pixels: 2_007_040
frame_sampler_args:
max_num_frames: 12
min_num_frames: 8
is_mandatory: true
shuffle_lines: True
shuffle_seed: 0
num_used_data:
- 10
weight: 1

```

作者您好！这是我的 yaml。当我把edit部分的is_mandatory设置为true的时候，就会在计算loss时候报错：
```
File "/inspire/hdd/project/aiforquantum/zhengkaipeng-240108120123/old_space/zhengkaipeng-240108120123/weilai/codes/BAGEL/data/data_utils.py", line 38, in create_sparse_mask [rank2]:
document_id = torch.cat([torch.full((l,), i) for i, l in enumerate(document_lens, start=1)]).to(device) [rank2]: ^^^^^^^^^^^^^^^^^^^
[rank2]: RuntimeError: Trying to create tensor with negative dimension -788: [-788]
```

Start with the provided YAML, especially unified_edit.is_mandatory, and reproduce the failure during loss computation. Then inspect data/data_utils.py at create_sparse_mask line 38, using the document_lens value from the traceback to determine why a negative tensor dimension is produced. Done means establishing whether is_mandatory: true is supported for this dataset and documenting or correcting the configuration path so training completes.

Written by the indexing model from the issue text.