[Kimi-K2.5] Fix NVFP4 Kimi-K2.5 weight mapping and exclude list

[mmangkad]

Motivation

Make nvidia/Kimi-K2.5-NVFP4 load cleanly in SGLang. The checkpoint’s naming and hf_quant_config.json exclude list use language_model.layers.*, which doesn’t match SGLang’s current Kimi/Deepseek module prefixes. This mismatch caused weight shape errors and unintended quantization of self‑attn.

Modifications

• Map language_model.layers.* to language_model.model.layers.* in KimiK25ForConditionalGeneration before loading weights.
• Apply the same mapping to ModelOpt NVFP4 exclude_modules, and also add stripped variants (without language_model.) so the inner Deepseek prefixes (model.layers.*) are excluded as intended.

Accuracy Tests

python -m sglang.launch_server --model-path nvidia/Kimi-K2.5-NVFP4 --tensor-parallel-size 8 --trust-remote-code --reasoning-parser kimi_k2 --tool-call-parser kimi_k2 --moe-runner-backend flashinfer_trtllm --quantization modelopt_fp4 --model-loader-extra-config '{"enable_multithread_load": true,"num_threads": 119}'

GSM8K 20 shots

Accuracy: 0.936
Invalid: 0.001
Latency: 90.726 s
Output throughput: 1491.511 token/s

Benchmarking and Profiling

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 or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Summary of Changes

Hello @mmangkad, 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 addresses compatibility issues when loading the nvidia/Kimi-K2.5-NVFP4 model within SGLang. By introducing a mechanism to correctly map module prefixes and adjust quantization exclusion lists, it resolves discrepancies in checkpoint naming that previously led to weight shape errors and unintended quantization of model components, enabling the model to load cleanly.

Highlights

• Kimi-K2.5 Weight Mapping Fix: Implemented a WeightsMapper in KimiK25ForConditionalGeneration to correctly map language_model.layers.* to language_model.model.layers.* when loading weights for the nvidia/Kimi-K2.5-NVFP4 model, resolving module prefix mismatches.
• ModelOpt NVFP4 Exclude List Adjustment: Modified ModelOptQuantConfig to apply the same weight name mapping to exclude_modules and added stripped variants (without language_model.) to ensure correct exclusion of inner Deepseek prefixes during NVFP4 quantization, preventing unintended quantization of self-attention layers.

🧠 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.

Changelog

• python/sglang/srt/layers/quantization/modelopt_quant.py
  • Imported WeightsMapper from sglang.srt.models.utils.
  • Added a new method apply_weight_name_mapper to ModelOptQuantConfig to dynamically adjust the exclude_modules list based on a provided WeightsMapper, ensuring proper handling of module name discrepancies for quantization.
  • Expanded the exclude_modules list to include both mapped and stripped variants of module names (e.g., language_model.layers.0.attn and layers.0.attn) to cover different naming conventions.
• python/sglang/srt/models/kimi_k25.py
  • Imported WeightsMapper from sglang.srt.models.utils.
  • Defined a hf_to_sglang_mapper within KimiK25ForConditionalGeneration to specifically handle the nvidia/Kimi-K2.5-NVFP4 model's weight naming convention, mapping language_model.layers. to language_model.model.layers..
  • Integrated the hf_to_sglang_mapper into the load_weights method of KimiK25ForConditionalGeneration to automatically apply the necessary weight name transformations during model loading.

Activity

• The author has provided accuracy test results for GSM8K 20 shots, showing an accuracy of 0.936, an invalid rate of 0.001, a latency of 90.726 s, and an output throughput of 1491.511 token/s.
• The pull request includes a detailed checklist for formatting, unit tests, documentation, and benchmarking, indicating readiness for review.
• Instructions for the review process, including pinging Merge Oncalls and triggering CI tests, are outlined.

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 a fix for loading nvidia/Kimi-K2.5-NVFP4 weights by correctly mapping weight names and quantization exclusion lists. The changes are well-implemented and address the issue described. A WeightsMapper is introduced in KimiK25ForConditionalGeneration to handle the prefix difference in layer names. This mapper is also applied to the exclude_modules in ModelOptQuantConfig to ensure quantization is correctly skipped for the intended modules. The implementation is clean and robust. I've made one suggestion to improve memory efficiency during weight loading.

[mickqian]

/tag-and-rerun-ci

[mickqian]

/rerun-failed-ci