[new-model] Add support for Cohere2ForCausalLM behind Command-A and Command-R Models

[ljw-mc]

Motivation

This PR addresses #4570 and provides support for the Cohere2ForCausalLM architecture behind Cohere's Command-A and Command-R7B model variants for SGLang.

Background

Cohere2ForCausalLM features a hybrid attention approach, compared to CohereForCausalLM:

• Alternates between 3 SWA layers and 1 global attention layer.
• RoPE applied only to SWA layers, NoPE on global attention layer.

For more details refer to the original Command-A Paper by Cohere

Benchmarking and Profiling

Benchmarked through MMLU using the following commands:

python3 -m sglang.launch_server \
    --model-path CohereLabs/c4ai-command-r7b-12-2024 \
    --trust-remote-code

python3 bench_sglang.py 

SGLang model accuracy (66.1) vs Cohere Paper reported accuracy (65.2) for Command R7b

mmlu:
Total latency: 133.511   (on L40 GPU)
Average accuracy: 66.1     (Cohere official paper: 65.2)

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

[gemini-code-assist]

Summary of Changes

Hello @ljw-mc, 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 Cohere's second-generation models, specifically the Cohere2ForCausalLM architecture, which encompasses models like Command-A and Command-R. The core changes involve adapting the attention mechanism to correctly utilize Cohere2's sliding window attention (SWA) by integrating a sliding_window_size parameter. Additionally, the application of rotary embeddings has been refined to ensure proper behavior and compatibility across both Cohere v1 and v2 models, particularly in the context of SWA layers.

Highlights

• New Model Support: Added support for the Cohere2ForCausalLM architecture, enabling compatibility with Cohere's Command-A and Command-R model variants.
• Sliding Window Attention (SWA) Integration: Implemented the sliding_window_size attribute within CohereAttention to properly handle and support Sliding Window Attention (SWA) in Cohere2 models.
• Conditional Rotary Embeddings: Modified the application of rotary embeddings to be conditional, applying them only for Cohere v1 models or Cohere2 layers that utilize a sliding window attention mechanism.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

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 adds support for the Cohere2ForCausalLM architecture, which is used by Cohere's Command-A and Command-R models. The changes correctly identify the model version and configure the sliding window attention for Cohere2 models. However, I've found a critical issue in how the layer type is determined, which could lead to an IndexError at runtime. My review includes a suggested fix for this issue.

[JustinTong0323]

/tag-and-rerun-ci

[JustinTong0323]

Need to cherry pick this PR #17236 (dfc03db) to launch the server, and the test result seems reasonable.

mmlu:
Total latency: 35.161
Average accuracy: 0.660

[JustinTong0323]

/tag-and-rerun-ci