[Refactor] Split out deepseek v2 weight loader function into mixin

[xyjixyjixyji]

Motivation

DeepseekV2 code has been developed fast and a lot of historical code and be more orgranized, including the weight loading part.

Issue related: #16291

Modifications

This PR just moves the weight loader function into a mixin with some documentations.

The further refactors of splitting the weight loading internal will come after this PR get merged.

Accuracy Tests

Server start with CI model: python3 -m sglang.launch_server --model lmzheng/sglang-ci-dsv3-test --tp 1 --trust-remote-code --base-gpu-id 5 --port 12345

Client accuracy test:

root@innomatrix-us-adc-ndh100-0004:/sgl-workspace/sglang# python3 benchmark/gsm8k/bench_sglang.py --num-shots 8 --num-questions 1319 --parallel 1319 --data-path /tmp/test.jsonl --port 12345
100%|███████████████████████████████████████████████████████████████████████████████████| 1319/1319 [00:19<00:00, 69.38it/s]
Accuracy: 0.665
Invalid: 0.001
Latency: 19.946 s
Output throughput: 6475.981 token/s

Benchmarking and Profiling

N/A

On H100:

root@innomatrix-us-adc-ndh100-0004:/sgl-workspace/sglang# python3 -m sglang.test.send_one --port 12345
......
+-------------+--------+------------+-----------------+
| Latency (s) | Tokens | Acc Length | Speed (token/s) |
+-------------+--------+------------+-----------------+
|    1.572    |  252   |   1.000    |     160.35      |
+-------------+--------+------------+-----------------+

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.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments (/tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci) or contact authorized users to do so.
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Summary of Changes

Hello @xyjixyjixyji, 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 focuses on enhancing the modularity and organization of the DeepseekV2 model's weight loading mechanism. By moving the intricate weight loading and post-processing logic into a reusable mixin, the main model class becomes cleaner and easier to manage. This change sets the stage for future refactors and ensures a more structured approach to handling model weights, especially concerning various quantization and parallelism configurations.

Highlights

• Code Refactoring: The core weight loading logic for Deepseek V2/V3 models has been extracted into a dedicated mixin class, DeepseekV2WeightLoaderMixin, to improve code organization and maintainability.
• Modularization: The DeepseekV2WeightLoaderMixin now encapsulates complex weight loading functionalities, including handling tensor/pipeline parallelism, various quantization formats (FP8, INT8, AWQ), MoE expert weights, NextN speculative decoding weights, and shared expert fusion optimizations.
• Dependency Management: A new utility file, deepseek_common/utils.py, has been introduced to house common functions like awq_dequantize_func and enable_nextn_moe_bf16_cast_to_fp8, centralizing Deepseek-specific utilities.
• Simplified Model Class: The DeepseekV2ForCausalLM class has been significantly simplified by inheriting from the new mixin, reducing its internal complexity and making its load_weights method a direct call to the mixin's implementation.

🧠 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 is a good refactoring that moves the Deepseek V2 weight loading logic into a dedicated mixin, DeepseekV2WeightLoaderMixin. This significantly cleans up the DeepseekV2ForCausalLM class and improves code organization by centralizing weight loading logic. The introduction of awq_dequantize_func in a new utility file is also a nice improvement for handling device-specific implementations. However, I've identified a critical bug in the newly added mixin that would cause a runtime error during weight loading due to incorrect argument passing.

[Qiaolin-Yu]

/tag-and-rerun-ci

[Fridge003]

We can do low-level refactor in following PRs

[Qiaolin-Yu]

/tag-and-rerun-ci

[Fridge003]

Failed tests unrelated
https://github.com/sgl-project/sglang/actions/runs/21101509440/job/60716014440?pr=16649#logs