使用流式加载报错，传递了额外参数：TypeError: IterableDataset.map() got an unexpected keyword argument 'num_proc'

[CxsGhost]

Reminder

• I have read the README and searched the existing issues.

System Info

使用流式加载时，TypeError: IterableDataset.map() got an unexpected keyword argument 'num_proc'
问题出在：https://github.com/Qihoo360/360-LLaMA-Factory/blob/3bc07289eefcf8c8ea05f553e4ef0b82008419e4/src/llamafactory/data/loader.py#L224。
经检查Datasets库中IterableDataset map函数无法接收kwargs中的三个参数：

    kwargs = dict(
        num_proc=data_args.preprocessing_num_workers,
        load_from_cache_file=(not data_args.overwrite_cache) or (training_args.local_process_index != 0),
        desc="Running sequence parallel split on dataset",
    )

Reproduction

开启流式加载即可 --streaming True

Expected behavior

一般的Dataset map函数可以接收这些参数：

流式加载IterableDataset map：

修复方式：只需在 _get_sequence_parallel_dataset 中添加额外的判断逻辑即可，目前我本地运行良好

Others

No response

[chenhaoxing]

Hello. I found that when using stream reading, there is no stepped Loss. Have you ever encountered this? For example, 1000 steps can train all the data once. When the max step is set to 3000, no stepped loss occurs. However, when using non-streaming and without SP, there is a stepped loss.