支持packing算法，提高训练速度

[JasonCZH4]

No description provided.

[wxhcore]

@JasonCZH4 Thank you very much for submitting a PR to this project! I think adding the sequence packing feature is a great optimization that can significantly improve GPU utilization and training efficiency. After reviewing the newly added code, I have the following suggestions:

1. Several utility functions have been added in dataset.py; could we move these utility functions into a new file for unified management?

2. I notice that only SFTDataset has been modified. Considering potential future support for other dataset types, I suggest using inheritance or the strategy pattern for extensibility. Alternatively, if you have a better approach, you could also refer to the implementation in supervised.py from LLaMA-Factory.

3. Actually, your current code doesn’t work correctly: in base_trainer.py, the loss computation explicitly passes input_ids and attention_mask, but your code doesn’t pass position_ids to the model, so it may need to be changed to pass the batch using **batch.

4. I see you’ve added two new parameters, so the corresponding YAML config and argparse setup likely need to be updated to ensure the parameters are passed correctly.

5. Since I haven’t integrated CI on GitHub yet, I’ll need you to run the existing tests and add unit or functional tests for the new feature to ensure everything works as expected.

Finally, thank you again for contributing to this project!

[JasonCZH4]

Thanks for reviewing! It is my first time to submit PR. There are still many issues with the current code, and I am working hard to fix them in the coming time. In fact, I am currently conducting many tests. Once everything is OK, I will let you know. Thanks again!

[wxhcore]

Thanks for reviewing! It is my first time to submit PR. There are still many issues with the current code, and I am working hard to fix them in the coming time. In fact, I am currently conducting many tests. Once everything is OK, I will let you know. Thanks again!

Thank you very much for your efforts! This is also BumbleCore's first PR, so I attach great importance to it. Wish you all the best!