[MODEL] Adding Support for Qwen3.5 Models by zju-stu-lizheng · Pull Request #18489 · sgl-project/sglang

zju-stu-lizheng · 2026-02-09T12:58:05Z

Purpose

This PR adds model support for the upcoming Qwen3.5 models, including both dense and MoE variants.

Special thanks to @cao1zhg, @yizhang2077, and @attack204 for their review, and to @hnyls2002 and @mickqian from the SGLang team for their valuable review and support.

Reference HF implementation - huggingface/transformers#43830

Main Changes

Model Support: Added support for two new model classes:
- Qwen3_5MoeForConditionalGeneration
- Qwen3_5ForConditionalGeneration
Attention Layer: Introduced a new linear attention layer Qwen3_5GatedDeltaNet to optimize the fix_qkvz reordering operation.

Launch Server

Hardware setup: 8 × H200 GPUs

Without MTP

TP_SIZE=8

python -m sglang.launch_server \
    --model $vlmoe \
    --tp-size ${TP_SIZE} \
    --enable-multimodal \
    --max-mamba-cache-size 128 \
    --max-running-requests 128 \
    --chunked-prefill-size 2048 \
    --mamba-ssm-dtype bfloat16

With MTP

Add the following arguments to the above command:

    --speculative-algo NEXTN    \
    --speculative-num-steps 3     \
    --speculative-eagle-topk 1     \
    --speculative-num-draft-tokens 4 \

Check our cookbook for detailed deployment instruction: https://cookbook.sglang.io/autoregressive/Qwen/Qwen3.5

use rope_parameters rather than rope_scaling

1. multimodal inputs for mtp 2. dense model tie embedding for mtp

…base

1. rename Qwen3_vl_next to Qwen3_5 in config 2. merge qwen3_5.py and qwen3_5_moe.py

2. rename Qwen3_5LLMModel to Qwen3_5ForCausalLM

use rope_parameters rather than rope_scaling

1. multimodal inputs for mtp 2. dense model tie embedding for mtp

…base

1. rename Qwen3_vl_next to Qwen3_5 in config 2. merge qwen3_5.py and qwen3_5_moe.py

2. rename Qwen3_5LLMModel to Qwen3_5ForCausalLM

gemini-code-assist · 2026-02-09T12:59:27Z

Summary of Changes

Hello @zju-stu-lizheng, 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 support for the Qwen3.5 series of models, encompassing both dense and MoE variants. It includes a new attention layer, configuration updates, and code modifications to ensure seamless integration and optimized performance within the existing framework. The changes enhance the system's ability to handle advanced language models and improve overall processing efficiency.

Highlights

Model Support: Adds support for Qwen3.5 models, including both dense and MoE variants, by introducing new model classes: Qwen3_5MoeForConditionalGeneration and Qwen3_5ForConditionalGeneration.
Attention Layer: Implements a new linear attention layer Qwen3_5GatedDeltaNet to optimize the fix_qkvz reordering operation, enhancing attention mechanisms.
Configuration Files: Introduces new configuration files (qwen3_5.py) and modifies existing ones (__init__.py, model_config.py) to integrate Qwen3.5 models into the system.
Code Modifications: Updates several files to ensure compatibility and proper handling of the new Qwen3.5 model architectures, including changes to logits processing, rotary embeddings, and multimodal utilities.

🧠 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

benchmark/kernels/fused_moe_triton/common_utils.py
- Added 'Qwen3_5MoeForConditionalGeneration' to the list of supported MoE models.
python/sglang/srt/configs/init.py
- Imported and exposed Qwen3_5Config and Qwen3_5MoeConfig.
python/sglang/srt/configs/model_config.py
- Modified draft model configuration to include Qwen3_5 architectures.
- Added Qwen3_5 architectures to the list of generation models.
python/sglang/srt/configs/qwen3_5.py
- Added new configuration files for Qwen3_5 and Qwen3_5Moe models, defining their architecture and parameters.
python/sglang/srt/layers/logits_processor.py
- Added mm_input_embeds to LogitsProcessorOutput and LogitsMetadata to handle multimodal inputs.
python/sglang/srt/layers/rotary_embedding.py
- Modified get_rope_index to support Qwen3_5 models when using video grid THW.
python/sglang/srt/managers/mm_utils.py
- Modified general_mm_embed_routine to handle mm_input_embeds in forward_batch.
python/sglang/srt/model_executor/forward_batch_info.py
- Added mm_input_embeds to ForwardBatch to support multimodal inputs.
python/sglang/srt/model_executor/model_runner.py
- Imported Qwen3_5Config and Qwen3_5MoeConfig.
- Modified hybrid_gdn_config to include Qwen3_5 configurations.
- Modified model_is_mrope to check for rope_parameters or rope_scaling.
python/sglang/srt/models/qwen3_5.py
- Added new model classes Qwen3_5ForCausalLM and Qwen3_5MoeForCausalLM, including attention and decoder layers.
python/sglang/srt/models/qwen3_5_mtp.py
- Added Qwen3_5ForCausalLMMTP to support Multi Token Prediction.
python/sglang/srt/models/qwen3_next.py
- Modified rope_scaling to check for rope_parameters or rope_scaling.
python/sglang/srt/multimodal/processors/qwen_vl.py
- Added support for Qwen3_5 models in QwenVLImageProcessor.
- Modified process_mm_data_async to handle video metadata for Qwen3_5 models.
python/sglang/srt/server_args.py
- Modified _handle_model_specific_adjustments to include Qwen3_5 models.
python/sglang/srt/speculative/eagle_worker.py
- Modified forward_draft_extend to include mm_input_embeds.
python/sglang/srt/utils/common.py
- Modified load_video to support video file as list, tuple, torch.Tensor, or np.ndarray.
python/sglang/srt/utils/hf_transformers_utils.py
- Imported Qwen3_5Config and Qwen3_5MoeConfig.

Activity

The pull request introduces support for Qwen3.5 models, including both dense and MoE variants.
A new linear attention layer Qwen3_5GatedDeltaNet is implemented to optimize the fix_qkvz reordering operation.
Configuration files are added and modified to integrate Qwen3.5 models into the system.
Code modifications are made to ensure compatibility and proper handling of the new Qwen3.5 model architectures.

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 adds support for the new Qwen3.5 models, including both dense and MoE variants. The changes are extensive, introducing new model and configuration files, and updating various parts of the system to accommodate these new models. The implementation appears to be well-integrated with the existing codebase. My review focuses on a few opportunities for code simplification and correctness improvements.

gemini-code-assist · 2026-02-09T13:01:23Z

python/sglang/srt/models/qwen3_5.py

+        if hasattr(config, "rope_parameters"):
+            self.rope_scaling = getattr(config, "rope_parameters", None)
+        else:
+            self.rope_scaling = getattr(config, "rope_scaling", None)


This logic for setting self.rope_scaling can be simplified to a single line. Using getattr with a default value is more concise and readable.

self.rope_scaling = getattr(config, "rope_parameters", getattr(config, "rope_scaling", None))

gemini-code-assist · 2026-02-09T13:01:23Z

python/sglang/srt/models/qwen3_5_mtp.py

+                forward_batch=forward_batch,
+            )
+        else:
+            raise ("not implementation for other mtp layers[self.num_mtp_layers > 1]")


Raising a string literal does not raise a proper exception. It's better to raise a specific exception type like NotImplementedError for unimplemented features.

Suggested change

raise ("not implementation for other mtp layers[self.num_mtp_layers > 1]")

raise NotImplementedError("not implementation for other mtp layers[self.num_mtp_layers > 1]")

gemini-code-assist · 2026-02-09T13:01:23Z

python/sglang/srt/models/qwen3_next.py

+        if "rope_parameters" in config:
+            self.rope_scaling = getattr(config, "rope_parameters", None)
+        else:
+            self.rope_scaling = getattr(config, "rope_scaling", None)


This logic for setting self.rope_scaling can be simplified. Using getattr with a default value is more concise. Also, using hasattr is generally preferred over the in operator for checking attributes on an object.

self.rope_scaling = getattr(config, "rope_parameters", getattr(config, "rope_scaling", None))

mickqian · 2026-02-09T13:01:35Z

/tag-and-rerun-ci

mickqian · 2026-02-09T15:04:26Z

/rerun-failed-ci

mickqian · 2026-02-09T15:34:28Z

/rerun-failed-ci

mickqian · 2026-02-09T16:04:30Z

/rerun-failed-ci

mickqian · 2026-02-09T16:34:33Z

/rerun-failed-ci

mickqian · 2026-02-09T17:04:35Z

/rerun-failed-ci

mickqian · 2026-02-09T18:41:10Z

/rerun-failed-ci

mickqian · 2026-02-09T21:37:46Z

/rerun-failed-ci

mickqian · 2026-02-09T22:07:48Z

/rerun-failed-ci

mickqian · 2026-02-10T01:07:05Z

/rerun-failed-ci

mickqian · 2026-02-10T01:37:08Z

/rerun-failed-ci

mickqian · 2026-02-10T02:07:10Z

/rerun-failed-ci

mickqian · 2026-02-10T02:37:11Z

/rerun-failed-ci

Co-authored-by: 瑀澈 <yuche.lz@alibaba-inc.com>

Huixxi · 2026-02-25T09:01:02Z

Emm, may I ask why qwen3.5's moe sparse block reuse qwen2's not qwen3's? @zju-stu-lizheng

cao1zhg · 2026-02-26T07:40:29Z

Emm, may I ask why qwen3.5's moe sparse block reuse qwen2's not qwen3's? @zju-stu-lizheng

for shared_expert

Cppowboy · 2026-03-03T06:47:25Z

Which tool call paser should be used for qwen3.5 series model?

Huixxi · 2026-03-04T08:03:20Z

Which tool call paser should be used for qwen3.5 series model?

Seems qwen3_coder from https://www.modelscope.cn/models/Qwen/Qwen3.5-35B-A3B

riZZZhik · 2026-03-04T10:40:22Z

Hello,
Why does half of the docs suggest setting --speculative-algorithm to EAGLE (e.g., cookbook) and the other half to NEXTN (e.g., this and another PR descriptions)

Which one is correct?

riZZZhik · 2026-03-04T10:41:22Z

And why is expert parallelism not suggested for this model?
It's not supported yet or has worse performance?

Co-authored-by: 瑀澈 <yuche.lz@alibaba-inc.com>

瑀澈 added 30 commits January 22, 2026 00:15

[qwen3_next vl] first push draft model file

2e20be6

[qwen3_next vl] load moe, need to fix some bugs

fd9bdb7

[mrope] fix for transformers v5:

f1cec48

use rope_parameters rather than rope_scaling

[qwen3_next vl] support moe version

6f2dfe1

[qwen3_next vl] support dense model

89f10c9

[lint] delete debug info

c6a6eaa

[qwen3_next vl] support mtp model

ab2f31b

1. multimodal inputs for mtp 2. dense model tie embedding for mtp

fix : add multimodel config for qwen3.5

c046573

refactor(srt): Renamed modules and simplified the Qwen 3.5 model code…

afc5a35

…base

fix: rope_theta and partial_rotary_factor factor

39f3499

[qwen3_5] add qwen3_5 dense and moe model

6c04e0b

1. rename Qwen3_vl_next to Qwen3_5 in config 2. merge qwen3_5.py and qwen3_5_moe.py

fix(srt): import FusedMoE

f5eeb96

fix: offline preprocessed video

20ed51f

1. fallback qwen3vl

ed4bb6d

2. rename Qwen3_5LLMModel to Qwen3_5ForCausalLM

fix: remove visual parameter for Qwen3_5ForCausalLM

f004a2c

fix import

7a8906f

[qwen3_next vl] first push draft model file

14f5fa3

[qwen3_next vl] load moe, need to fix some bugs

69c5b3a

[mrope] fix for transformers v5:

0ad9dc1

use rope_parameters rather than rope_scaling

[qwen3_next vl] support moe version

e263cac

[qwen3_next vl] support dense model

e71b10c

[lint] delete debug info

11ef96c

[qwen3_next vl] support mtp model

36ddc67

1. multimodal inputs for mtp 2. dense model tie embedding for mtp

fix : add multimodel config for qwen3.5

f3e43e8

refactor(srt): Renamed modules and simplified the Qwen 3.5 model code…

87e60df

…base

fix: rope_theta and partial_rotary_factor factor

542670c

[qwen3_5] add qwen3_5 dense and moe model

2bbfe83

1. rename Qwen3_vl_next to Qwen3_5 in config 2. merge qwen3_5.py and qwen3_5_moe.py

fix(srt): import FusedMoE

12922e6

fix: offline preprocessed video

0cc1cba

1. fallback qwen3vl

e7b8130

2. rename Qwen3_5LLMModel to Qwen3_5ForCausalLM

zju-stu-lizheng requested review from BBuf, Edwardf0t1, HaiShaw and ch-wan as code owners February 9, 2026 12:58

gemini-code-assist bot reviewed Feb 9, 2026

View reviewed changes

github-actions bot added the run-ci label Feb 9, 2026

ispobock mentioned this pull request Feb 9, 2026

[Feature] Support Qwen3.5 #18465

Open

2 tasks

mickqian merged commit 27c4476 into sgl-project:main Feb 9, 2026
260 of 287 checks passed

Johnsonms pushed a commit to Johnsonms/sglang that referenced this pull request Feb 14, 2026

model: support Qwen3.5 (sgl-project#18489)

989bf30

Co-authored-by: 瑀澈 <yuche.lz@alibaba-inc.com>

magicYang1573 pushed a commit to magicYang1573/sglang that referenced this pull request Mar 9, 2026

model: support Qwen3.5 (sgl-project#18489)

f753488

Co-authored-by: 瑀澈 <yuche.lz@alibaba-inc.com>

	raise ("not implementation for other mtp layers[self.num_mtp_layers > 1]")
	raise NotImplementedError("not implementation for other mtp layers[self.num_mtp_layers > 1]")

Conversation

zju-stu-lizheng commented Feb 9, 2026 • edited by JustinTong0323 Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Purpose

Main Changes

Launch Server

Without MTP

With MTP

Uh oh!

gemini-code-assist bot commented Feb 9, 2026

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!

gemini-code-assist bot Feb 9, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Feb 9, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Feb 9, 2026

Choose a reason for hiding this comment

Uh oh!

mickqian commented Feb 9, 2026

Uh oh!

mickqian commented Feb 9, 2026

Uh oh!

mickqian commented Feb 9, 2026

Uh oh!

mickqian commented Feb 9, 2026

Uh oh!

Uh oh!

mickqian commented Feb 9, 2026

Uh oh!

mickqian commented Feb 9, 2026

Uh oh!

mickqian commented Feb 9, 2026

Uh oh!

mickqian commented Feb 9, 2026

Uh oh!

mickqian commented Feb 9, 2026

Uh oh!

mickqian commented Feb 10, 2026

Uh oh!

mickqian commented Feb 10, 2026

Uh oh!

mickqian commented Feb 10, 2026

Uh oh!

mickqian commented Feb 10, 2026

Uh oh!

Huixxi commented Feb 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cao1zhg commented Feb 26, 2026

Uh oh!

Cppowboy commented Mar 3, 2026

Uh oh!

Huixxi commented Mar 4, 2026

Uh oh!

riZZZhik commented Mar 4, 2026

Uh oh!

riZZZhik commented Mar 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

zju-stu-lizheng commented Feb 9, 2026 •

edited by JustinTong0323

Loading

Huixxi commented Feb 25, 2026 •

edited

Loading