fix: correct scale parameter remapping logic in Llama4ForConditionalGeneration

[JustinTong0323]

Motivation

Fix the issue below, which is caused by #10042

It looks like #10611 should be included in the latest release lmsysorg/sglang:v0.5.3-cu129 and lmsysorg/sglang:v0.5.3rc2-cu129 . but when i try python3 -m sglang.launch_server --model-path=nvidia/Llama-4-Scout-17B-16E-Instruct-FP8 --tp=8 --trust-remote-code --mem-fraction-static 0.7 --context-length=131072 --attention-backend=fa3 --enable-multimodal --tool-call-parser=pythonic --chat-template=llama-4 --cuda-graph-max-bs=48 --host=0.0.0.0 --speculative-algorithm=EAGLE3 --speculative-num-steps=3 --speculative-eagle-topk=4 --speculative-num-draft-tokens=8 --speculative-draft-model-path=/models/llama-4-scout-eagle/state_7 --port=8000
i now see the error

2025-10-06 17:09:50 TP6] Scheduler hit an exception: Traceback (most recent call last):
  File "/sgl-workspace/sglang/python/sglang/srt/managers/scheduler.py", line 2847, in run_scheduler_process
    scheduler = Scheduler(
                ^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/managers/scheduler.py", line 396, in __init__
    self.tp_worker = TpWorkerClass(
                     ^^^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/managers/tp_worker.py", line 100, in __init__
    self.model_runner = ModelRunner(
                        ^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/model_executor/model_runner.py", line 298, in __init__
    self.initialize(min_per_gpu_memory)
  File "/sgl-workspace/sglang/python/sglang/srt/model_executor/model_runner.py", line 345, in initialize
    self.load_model()
  File "/sgl-workspace/sglang/python/sglang/srt/model_executor/model_runner.py", line 789, in load_model
    self.model = get_model(
                 ^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/model_loader/__init__.py", line 28, in get_model
    return loader.load_model(
           ^^^^^^^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/model_loader/loader.py", line 494, in load_model
    self.load_weights_and_postprocess(
  File "/sgl-workspace/sglang/python/sglang/srt/model_loader/loader.py", line 502, in load_weights_and_postprocess
    model.load_weights(weights)
  File "/sgl-workspace/sglang/python/sglang/srt/models/mllama4.py", line 660, in load_weights
    if self._handle_stacked_params(
       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/sgl-workspace/sglang/python/sglang/srt/models/mllama4.py", line 719, in _handle_stacked_params
    param = params_dict[transformed_name]
            ~~~~~~~~~~~^^^^^^^^^^^^^^^^^^
KeyError: 'language_model.model.layers.29.self_attn.qkv_proj.k_scale'

Modifications

Accuracy Tests

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.

[gemini-code-assist]

Summary of Changes

Hello @JustinTong0323, 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 a logical flaw in the Llama4ForConditionalGeneration model's parameter remapping mechanism. It refines the condition within the _handle_scale_remapping function to ensure that scale parameters are correctly identified as 'handled,' encompassing scenarios where a parameter's name is changed as well as those where no remapping is necessary but the parameter still falls under the handling scope. This fix enhances the robustness and accuracy of parameter processing for Llama4 models.

Highlights

• Logic Correction: Corrected the conditional logic within the _handle_scale_remapping method in Llama4ForConditionalGeneration to accurately determine when a scale parameter has been handled.
• Parameter Remapping: The remapping logic now correctly accounts for cases where maybe_remap_kv_scale_name returns None (indicating no remapping occurred but the parameter is still considered handled) or when a remapping results in a different name.

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 corrects the logic for handling scale parameter remapping during model weight loading. The previous logic failed to correctly skip scale parameters that could not be remapped, potentially leading to errors. The new logic ensures that both unmappable and successfully remapped scale parameters are considered 'handled' and skipped by the loading loop. I've suggested a minor simplification to the return statement to improve readability.

[JustinTong0323]

cc @BowenBao @HaiShaw @kkHuang-amd for review.

[HaiShaw]

cc @BowenBao

[BowenBao]

LGTM, thanks!