[feat] Add Support Cut-Cross-Entropy (CCE)

[w1ida]

This PR adds optional Cut-Cross-Entropy (CCE) loss into the ms-swift training pipeline.
CCE is a memory-efficient alternative to standard CE, from Apple’s project:

➡️ https://github.com/apple/ml-cross-entropy

---

✨ Features

• Add use_cce=True in TrainArguments to enable CCE.
• Plug-and-play replacement; no model modification required.
• Add test script: tests/train/test_cce.py.

Usage:

TrainArguments(
    model='Qwen/Qwen2.5-0.5B-Instruct',
    dataset=['gsm8k#1024'],
    split_dataset_ratio=0.01,
    use_cce=True,
    **kwargs
)

Install dependency:

pip install "cut-cross-entropy @ git+https://github.com/w1ida/ml-cross-entropy.git"

---

📊 Memory Result (A10, BS=64)

• Hardware: NVIDIA A10
• transformers: 4.56.1
• Model: Qwen/Qwen2.5-0.5B-Instruct
• Dataset: gsm8k#1024
• Batch size: 64

Environment

• Hardware: NVIDIA A10
• transformers: 4.56.1
• Model: Qwen/Qwen2.5-0.5B-Instruct
• Dataset: gsm8k#1024
• Batch size: 64

Memory Usage

Loss	Memory (GiB)
transformers CE	16.24
CCE	2.62

≈ 84% memory reduction.

---

📄 Reference

Cut Your Losses in Large-Vocabulary Language Models
https://arxiv.org/abs/2411.09009

[gemini-code-assist]

Summary of Changes

Hello @w1ida, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request integrates Cut-Cross-Entropy (CCE) loss into the training pipeline, offering a significant memory optimization for large language models. By providing a memory-efficient alternative to traditional cross-entropy, it enables training of larger models or larger batch sizes on existing hardware, making the training process more accessible and scalable.

Highlights

• Cut-Cross-Entropy (CCE) Integration: Introduced optional CCE loss into the training pipeline, a memory-efficient alternative to standard cross-entropy, based on Apple's ml-cross-entropy project.
• Memory Efficiency: Demonstrated approximately 84% memory reduction on an NVIDIA A10 GPU for a Qwen model with a batch size of 64, significantly optimizing resource usage.
• Ease of Use: CCE can be enabled by setting "use_cce=True" in "TrainArguments" and is designed as a plug-and-play replacement without requiring model modifications.
• Model Support: The CCE patching mechanism supports various LLM architectures including Llama, Mistral, Phi3, Gemma2, and Qwen2.
• Test Coverage: A new test script ("tests/train/test_cce.py") has been added to validate CCE functionality during SFT training, ensuring proper integration and operation.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces support for Cut-Cross-Entropy (CCE) loss, offering a memory-efficient alternative for training. The changes integrate CCE into the training pipeline, add UI elements for activation, and include a dedicated test script. The implementation generally follows existing patterns for kernel integrations, such as the Liger kernel, ensuring consistency. The UI updates are well-localized and clearly explain the new feature. Overall, this is a valuable addition for optimizing memory usage during training.

[Jintao-Huang]

哈咯，这是我这测试的情况，看起来liger_kernel的显存占用更少

使用的是你给的例子

"""
liger_kernel:
{'loss': 0.72876334, 'grad_norm': 5.18441868, 'learning_rate': 9.989e-05, 'epoch': 0.06, 'global_step/max_steps': '1/48', 'percentage': '2.08%', 'elapsed_time': '3s', 'remaining_time': '2m 27s', 'memory(GiB)': 6.26, 'train_speed(iter/s)': 0.318979}
{'loss': 0.55229092, 'grad_norm': 1.04675198, 'learning_rate': 9.735e-05, 'epoch': 0.31, 'global_step/max_steps': '5/48', 'percentage': '10.42%', 'elapsed_time': '8s', 'remaining_time': '1m 9s', 'memory(GiB)': 6.27, 'train_speed(iter/s)': 0.622059}
cce:
{'loss': 0.72674572, 'grad_norm': 5.16368008, 'learning_rate': 9.989e-05, 'epoch': 0.06, 'global_step/max_steps': '1/48', 'percentage': '2.08%', 'elapsed_time': '5s', 'remaining_time': '4m 3s', 'memory(GiB)': 6.74, 'train_speed(iter/s)': 0.193078}
{'loss': 0.55218643, 'grad_norm': 1.03230286, 'learning_rate': 9.735e-05, 'epoch': 0.31, 'global_step/max_steps': '5/48', 'percentage': '10.42%', 'elapsed_time': '11s', 'remaining_time': '1m 39s', 'memory(GiB)': 6.74, 'train_speed(iter/s)': 0.430464}
原始
{'loss': 0.72737986, 'grad_norm': 5.16535616, 'learning_rate': 9.989e-05, 'token_acc': 0.83662044, 'epoch': 0.06, 'global_step/max_steps': '1/48', 'percentage': '2.08%', 'elapsed_time': '2s', 'remaining_time': '1m 44s', 'memory(GiB)': 61.58, 'train_speed(iter/s)': 0.450271}
{'loss': 0.55272961, 'grad_norm': 1.02546477, 'learning_rate': 9.735e-05, 'token_acc': 0.85563783, 'epoch': 0.31, 'global_step/max_steps': '5/48', 'percentage': '10.42%', 'elapsed_time': '7s', 'remaining_time': '1m 0s', 'memory(GiB)': 61.58, 'train_speed(iter/s)': 0.714248}
"""

[w1ida]

哈咯，这是我这测试的情况，看起来liger_kernel的显存占用更少

使用的是你给的例子

""" liger_kernel: {'loss': 0.72876334, 'grad_norm': 5.18441868, 'learning_rate': 9.989e-05, 'epoch': 0.06, 'global_step/max_steps': '1/48', 'percentage': '2.08%', 'elapsed_time': '3s', 'remaining_time': '2m 27s', 'memory(GiB)': 6.26, 'train_speed(iter/s)': 0.318979} {'loss': 0.55229092, 'grad_norm': 1.04675198, 'learning_rate': 9.735e-05, 'epoch': 0.31, 'global_step/max_steps': '5/48', 'percentage': '10.42%', 'elapsed_time': '8s', 'remaining_time': '1m 9s', 'memory(GiB)': 6.27, 'train_speed(iter/s)': 0.622059} cce: {'loss': 0.72674572, 'grad_norm': 5.16368008, 'learning_rate': 9.989e-05, 'epoch': 0.06, 'global_step/max_steps': '1/48', 'percentage': '2.08%', 'elapsed_time': '5s', 'remaining_time': '4m 3s', 'memory(GiB)': 6.74, 'train_speed(iter/s)': 0.193078} {'loss': 0.55218643, 'grad_norm': 1.03230286, 'learning_rate': 9.735e-05, 'epoch': 0.31, 'global_step/max_steps': '5/48', 'percentage': '10.42%', 'elapsed_time': '11s', 'remaining_time': '1m 39s', 'memory(GiB)': 6.74, 'train_speed(iter/s)': 0.430464} 原始 {'loss': 0.72737986, 'grad_norm': 5.16535616, 'learning_rate': 9.989e-05, 'token_acc': 0.83662044, 'epoch': 0.06, 'global_step/max_steps': '1/48', 'percentage': '2.08%', 'elapsed_time': '2s', 'remaining_time': '1m 44s', 'memory(GiB)': 61.58, 'train_speed(iter/s)': 0.450271} {'loss': 0.55272961, 'grad_norm': 1.02546477, 'learning_rate': 9.735e-05, 'token_acc': 0.85563783, 'epoch': 0.31, 'global_step/max_steps': '5/48', 'percentage': '10.42%', 'elapsed_time': '7s', 'remaining_time': '1m 0s', 'memory(GiB)': 61.58, 'train_speed(iter/s)': 0.714248} """

@Jintao-Huang 嗯嗯，liger_kernel 这边还额外优化了 RMSNorm、RoPE、SwiGLU 等算子，从整体显存占用来看确实更低。 paper里只对 logits + CE 部分的显存开销做了对比。
另外，axolotl 这个框架已经支持了 CCE 和 liger-kernel，感觉可以考虑在 ms-swift 里也集成一下CCE的支持~

[Jintao-Huang]

支持CCE会额外引入维护成本，我和其他开发者讨论一下