support deepseek-v2-lite online train and support yarn rope

[jiapingW]

Motivation

Since only the config for deepseek-v2-lite training is provided, we found after running LlamaForEagle3 that it does not support yarn position encoding. Therefore, I added adaptation for yarn position encoding and trained Deepseek-v2-lite using the test data online.

Modifications

Related Issues

Fix this issue: #219
We will also add support for YARN in sglang later.

Accuracy Test

Benchmark & Profiling

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://sgl-fru7574.slack.com/archives/C09784E3EN6 to discuss your PR.

[gemini-code-assist]

Summary of Changes

Hello @jiapingW, 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 introduces comprehensive support for online training of the DeepSeek-V2-Lite model with the Eagle3 architecture. It addresses the previous lack of YARN (Yet another RoPE N-dimensional) position encoding support by integrating the YARN rotary embedding method and providing a dedicated script to facilitate this training process.

Highlights

• YARN RoPE Implementation: Implemented the LlamaYarnRotaryEmbedding class and associated helper functions in specforge/modeling/draft/llama3_eagle.py to enable YARN (Yet another RoPE N-dimensional) position encoding.
• YARN Integration: Modified the LlamaAttention module to dynamically select and apply the newly implemented YARN rotary embedding based on configuration, ensuring compatibility with DeepSeek-V2-Lite.
• DeepSeek-V2-Lite Online Training Script: Introduced a new shell script (examples/run_deepseek_v2_lite_eagle3_online.sh) to streamline the online training process for DeepSeek-V2-Lite using the Eagle3 architecture and the updated YARN configuration.
• Configuration Update: Updated the configs/deepseek-v2-lite-eagle3.json file to correctly specify "rope_type": "yarn", aligning with the new YARN implementation.

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 in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

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 issue 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 adds support for YaRN rope scaling to enable online training for deepseek-v2-lite. The changes include a new LlamaYarnRotaryEmbedding class with its helper functions, updates to the model configuration to use YaRN, and a new example script for training. The implementation of YaRN appears correct. I have a couple of minor suggestions to improve code style and adherence to conventions.

[gemini-code-assist]

It's a good practice for text files, including shell scripts, to end with a newline character. This prevents potential issues with file concatenation and some command-line tools. Please add a newline at the end of this file.

[gemini-code-assist]

The function parameters min and max shadow Python's built-in functions. While it doesn't cause an issue in this specific function, it's a bad practice that can lead to subtle bugs if the built-in functions were needed. It's recommended to use alternative names like min_val and max_val to avoid shadowing.¹

Suggested change

	def yarn_linear_ramp_mask(min, max, dim):
	if min == max:
	max += 0.001 # Prevent singularity
	linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)
	ramp_func = torch.clamp(linear_func, 0, 1)
	return ramp_func
	def yarn_linear_ramp_mask(min_val, max_val, dim):
	if min_val == max_val:
	max_val += 0.001 # Prevent singularity
	linear_func = (torch.arange(dim, dtype=torch.float32) - min_val) / (max_val - min_val)
	ramp_func = torch.clamp(linear_func, 0, 1)
	return ramp_func

Style Guide References

Footnotes

1. As per PEP 8, it is recommended to avoid using names that shadow built-in functions. ↩

[dongyibo]

Hello, does the use of yarn rotation position encoding improve the accuracy of eagle3 compared to not using any rotation position encoding?
@jiapingW

[jiapingW]

Hello, does the use of yarn rotation position encoding improve the accuracy of eagle3 compared to not using any rotation position encoding?

I trained DeepSeek-v2-Lite on ShareGPT for one epoch using YARN. The accuracy for position 1 was 0.36. I didn't use the standard rope for training. Due to limited resources, I couldn't train with a long context. I think using rope and YARN shouldn't significantly impact training performance unless you're training with very long inputs.