Implement LoRA for MoE with support for LoRA injection for nn.parameters #9337

Ziheng-Zhang-AUS · 2025-10-23T07:12:27Z

What does this PR do?

Fixes #9251

Support PEFT 0.17.0 features of LoRA injection for nn.parameters.
Add corresponding GUI module
Add corresponding LoRA tests

Before submitting

Did you read the contributor guideline?
Did you write any new necessary tests?

add lora parameter webUI

…-config-logic 3 add formating parameters to lora config logic

gemini-code-assist · 2025-10-23T07:12:45Z

Summary of Changes

Hello @Ziheng-Zhang-AUS, 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 significantly enhances the LoRA fine-tuning capabilities by enabling users to target specific "nn.Parameters" for LoRA injection, rather than being limited to entire modules. This granular control is especially valuable for advanced model architectures like Mixture-of-Experts (MoE) models, allowing for more precise and efficient adaptation. The changes span across configuration, model adaptation logic, testing, and the graphical user interface, ensuring a complete and robust implementation.

Highlights

Direct LoRA Parameter Injection: Introduced a new "lora_parameters" argument to allow direct application of LoRA to specific "nn.Parameters", which is particularly useful for fine-tuning components like expert parameters in Mixture-of-Experts (MoE) models.
PEFT 0.17.0 Feature Support: Implemented support for the LoRA injection features available in PEFT version 0.17.0, enhancing compatibility and leveraging the latest capabilities.
Web UI Integration: Added a dedicated input field in the web-based training interface for users to easily specify "lora_parameters" through the GUI.
Comprehensive Testing: Included new test cases to validate the correct functionality of LoRA injection into specified parameters, ensuring robustness and reliability.

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.

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 valuable feature to support LoRA injection for nn.Parameters, which is particularly useful for Mixture-of-Experts (MoE) models. The implementation looks mostly correct, with new arguments added to the configuration, CLI, and web UI.

However, I've found several critical issues in the tests. The utility function check_lora_model has been refactored, but the existing tests that rely on it have not been updated correctly, causing them to fail or test the wrong thing. The new test for lora_parameters also has incorrect logic. Additionally, the refactored check_lora_model seems to have lost some functionality compared to the old version, specifically for checking modules_to_save.

I've also noticed a minor issue with a print statement that should be a logger call, and an unrelated .drawio file that might have been added by mistake.

Please address the critical issues in the tests to ensure the new functionality is properly verified and doesn't introduce regressions.

tests/model/test_lora.py

src/llamafactory/train/test_utils.py

src/llamafactory/model/adapter.py

src/llamafactory/train/test_utils.py

未命名绘图.drawio

hiyouga · 2025-10-23T08:12:49Z

Please resolve the gemini's review

…-config-logic changed the implementation of func check_lora_model

Ziheng-Zhang-AUS · 2025-10-25T04:16:35Z

I have modified the implementation according to Gemini's review :)

…-config-logic conflict lora parameters test

.DS_Store

hiyouga · 2025-10-26T08:28:18Z

src/llamafactory/model/adapter.py

+        target_parameters = []
        if len(finetuning_args.lora_target) == 1 and finetuning_args.lora_target[0] == "all":
-            target_modules = find_all_linear_modules(model, finetuning_args.freeze_vision_tower)
+            if finetuning_args.lora_parameters: # if specified the parameters to be adapted, use them


When we specify the target parameters, the target modules should not be affected

This if-else is strange

Sorry, I didn't understand the idea very clearly, because I noticed that Lora target has a default value of "all". I was thinking that if this default value is not changed and Lora parameters are used for injection without the user specifying a target, then this if-else should be judged here.

The current implementation will prevent target_parameters (lora_parameters) from being passed to peftconfig when target_modules (lora_target) is specified. There is no need to modify this peice of code to pass lora_parameters. Instead I would pass it immediatly to peft_kwargs:

peft_kwargs = { "r": finetuning_args.lora_rank, "target_modules": target_modules, "target_parameters": finetuning_args.lora_parameters, "lora_alpha": finetuning_args.lora_alpha, "lora_dropout": finetuning_args.lora_dropout, "use_rslora": finetuning_args.use_rslora, "use_dora": finetuning_args.use_dora, "modules_to_save": finetuning_args.additional_target, }

target_parameters is an optional argument with None as default.

hiyouga · 2025-10-26T08:30:48Z

src/llamafactory/model/adapter.py

                "use_rslora": finetuning_args.use_rslora,
                "use_dora": finetuning_args.use_dora,
                "modules_to_save": finetuning_args.additional_target,
+                "target_parameters": target_parameters,


Is target parameters always be defined?

zzm1102 and others added 8 commits October 12, 2025 17:15

add lora parameter webUI

e61ff00

Merge pull request #2 from suyabot/1-webui-integrate-lora

8c67986

add lora parameter webUI

change the finetuning args of lora

ce14b48

change the adapter of lora

d202ae9

change of adapters

96f15f0

lora_parameters_tests

8342bf8

aaa

282026b

Merge pull request #4 from suyabot/3-add-formating-parameters-to-lora…

69c5f15

…-config-logic 3 add formating parameters to lora config logic

gemini-code-assist bot reviewed Oct 23, 2025

View reviewed changes

Ziheng-Zhang-AUS added 2 commits October 25, 2025 15:11

changed the implementation of func check_lora_model

770bf8e

Merge pull request #5 from suyabot/3-add-formating-parameters-to-lora…

640755d

…-config-logic changed the implementation of func check_lora_model

Ziheng-Zhang-AUS added 2 commits October 25, 2025 21:54

conflict lora parameters test

03ba684

Merge pull request #6 from suyabot/3-add-formating-parameters-to-lora…

12ec8da

…-config-logic conflict lora parameters test

Kuangdd01 added the pending This problem is yet to be addressed label Oct 25, 2025

Ziheng-Zhang-AUS temporarily deployed to docker October 26, 2025 08:26 — with GitHub Actions Inactive

hiyouga reviewed Oct 26, 2025

View reviewed changes

Delete .DS_Store

18418ba

Implement LoRA for MoE with support for LoRA injection for nn.parameters #9337

Are you sure you want to change the base?

Implement LoRA for MoE with support for LoRA injection for nn.parameters #9337

Conversation

Ziheng-Zhang-AUS commented Oct 23, 2025

What does this PR do?

Before submitting

Uh oh!

gemini-code-assist bot commented Oct 23, 2025

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

hiyouga commented Oct 23, 2025

Uh oh!

Ziheng-Zhang-AUS commented Oct 25, 2025

Uh oh!

Uh oh!

hiyouga Oct 26, 2025

Choose a reason for hiding this comment

Uh oh!

hiyouga Oct 26, 2025

Choose a reason for hiding this comment

Uh oh!

Ziheng-Zhang-AUS Oct 26, 2025

Choose a reason for hiding this comment

Uh oh!

nomandormosh Nov 1, 2025

Choose a reason for hiding this comment

Uh oh!

hiyouga Oct 26, 2025

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants