[Doc] Update DeepSeek-V3.2 document by Fridge003 · Pull Request #14321 · sgl-project/sglang

Fridge003 · 2025-12-02T23:28:17Z

Motivation

Following #14249

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.
Follow the SGLang code style guidance.
Work with maintainers to merge your PR. See the PR Merge Process

gemini-code-assist · 2025-12-02T23:28:32Z

Summary of Changes

Hello @Fridge003, 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 documentation for the DeepSeek-V3.2 model family, providing clearer guidance on its various versions, their respective tool parsing configurations, and detailed instructions for launching them with SGLang. It also expands the accuracy testing section with new results and setup procedures for different DeepSeek-V3.2 variants.

Highlights

DeepSeek-V3.2 Tool Parser Support: Added a new deepseekv32 tool parser entry to the documentation, specifically for the deepseek-ai/DeepSeek-V3.2 model, distinguishing it from deepseekv31 which now explicitly covers DeepSeek-V3.1 and DeepSeek-V3.2-Exp.
Expanded DeepSeek-V3.2 Documentation: The DeepSeek V3.2 usage document has been updated to clarify that it covers the entire DeepSeek-V3.2 model family (including -Exp, -Speciale, and the base V3.2). It also includes a note on tool call parser differences between these variants.
Detailed Launch Commands: New, specific launch commands are provided for DeepSeek-V3.2-Exp, DeepSeek-V3.2, and DeepSeek-V3.2-Speciale when using function calling and reasoning parsers, detailing the appropriate --tool-call-parser and --reasoning-parser arguments.
Updated AIME 2025 Accuracy Tests: The documentation now includes accuracy test results for DeepSeek-V3.2 and DeepSeek-V3.2-Speciale on the AIME 2025 benchmark. Instructions for hardcoding thinking mode for these models during evaluation have also been added, along with updated evaluation script parameters.

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 updates the documentation for DeepSeek-V3.2, including its variants. The changes are generally good and add valuable information. I've identified a few areas for improvement, including an incorrect model path in a command, some typos and grammatical errors, a broken markdown link, and some confusing or inconsistent formatting. My review includes specific suggestions to address these points and enhance the clarity and accuracy of the documentation.

gemini-code-assist · 2025-12-02T23:30:15Z

+To launch `DeepSeek-V3.2-Exp` with function calling and reasoning parser:
+```bash
+python3 -m sglang.launch_server \
+  --model-path deepseek-ai/DeepSeek-V3.2 \


The model path for DeepSeek-V3.2-Exp appears to be incorrect. It should point to the DeepSeek-V3.2-Exp model, not DeepSeek-V3.2.

Suggested change

--model-path deepseek-ai/DeepSeek-V3.2 \

--model-path deepseek-ai/DeepSeek-V3.2-Exp \

gemini-code-assist · 2025-12-02T23:30:15Z

+    "| `deepseekv31` | DeepSeek-V3.1 and DeepSeek-V3.2-Exp (e.g. `deepseek-ai/DeepSeek-V3.1`, `deepseek-ai/DeepSeek-V3.2-Exp`) | Recommend adding `--chat-template ./examples/chat_template/tool_chat_template_deepseekv31.jinja` (Or ..deepseekv32.jinja for DeepSeek-V3.2) to launch command. |\n",
+    "| `deepseekv32` | DeepSeek-V3.2 (`deepseek-ai/DeepSeek-V3.2`) | |\n",


With the addition of the deepseekv32 parser, the note for deepseekv31 has become confusing. It's better to simplify the note for deepseekv31 to only refer to its corresponding chat template and add a similar note for the new deepseekv32 parser.

Suggested change

"| `deepseekv31` | DeepSeek-V3.1 and DeepSeek-V3.2-Exp (e.g. `deepseek-ai/DeepSeek-V3.1`, `deepseek-ai/DeepSeek-V3.2-Exp`) | Recommend adding `--chat-template ./examples/chat_template/tool_chat_template_deepseekv31.jinja` (Or ..deepseekv32.jinja for DeepSeek-V3.2) to launch command. |\n",

"| `deepseekv32` | DeepSeek-V3.2 (`deepseek-ai/DeepSeek-V3.2`) | |\n",

"| `deepseekv31` | DeepSeek-V3.1 and DeepSeek-V3.2-Exp (e.g. `deepseek-ai/DeepSeek-V3.1`, `deepseek-ai/DeepSeek-V3.2-Exp`) | Recommend adding `--chat-template ./examples/chat_template/tool_chat_template_deepseekv31.jinja` to launch command. |\n",

"| `deepseekv32` | DeepSeek-V3.2 (`deepseek-ai/DeepSeek-V3.2`) | Recommend adding `--chat-template ./examples/chat_template/tool_chat_template_deepseekv32.jinja` to launch command. |\n",

gemini-code-assist · 2025-12-02T23:30:15Z

 # DeepSeek V3.2 Usage

-[DeepSeek-V3.2-Exp](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp) equips DeepSeek-V3.1-Terminus with DeepSeek Sparse Attention (DSA) through continued training. With DSA, a fine-grained sparse attention mechanism powered by a lightning indexer, DeepSeek-V3.2 achieves efficiency improvements in long-context scenarios.
+DeepSeek-V3.2 model families equips DeepSeek-V3.1-Terminus with DeepSeek Sparse Attention (DSA) through continued training. With DSA, a fine-grained sparse attention mechanism powered by a lightning indexer, DeepSeek-V3.2 achieves efficiency improvements in long-context scenarios.


There is a grammatical error here. Since "model families" is plural, the verb should be "equip", not "equips".

Suggested change

DeepSeek-V3.2 model families equips DeepSeek-V3.1-Terminus with DeepSeek Sparse Attention (DSA) through continued training. With DSA, a fine-grained sparse attention mechanism powered by a lightning indexer, DeepSeek-V3.2 achieves efficiency improvements in long-context scenarios.

DeepSeek-V3.2 model families equip DeepSeek-V3.1-Terminus with DeepSeek Sparse Attention (DSA) through continued training. With DSA, a fine-grained sparse attention mechanism powered by a lightning indexer, DeepSeek-V3.2 achieves efficiency improvements in long-context scenarios.

gemini-code-assist · 2025-12-02T23:30:16Z

+  pip install git+https://github.com/NVIDIA/NeMo-Skills.git --ignore-installed blinker
+  ```
+
+Nemo Skill can't enable thinking method from client side, so we need some hardcoding before launching server:


There's a typo here. For consistency with the library name, "Nemo Skill" should be "NeMo-Skills".

Suggested change

Nemo Skill can't enable thinking method from client side, so we need some hardcoding before launching server:

NeMo-Skills can't enable thinking method from client side, so we need some hardcoding before launching server:

gemini-code-assist · 2025-12-02T23:30:16Z

-Run the following script to evaluate AIME 2025:
+**For `DeepSeek-V3.2` and `DeepSeek-V3.2-Speciale`**:
+
+Hardcode the thinking mode to be `thinking` in (`_apply_jinja_template`)[https://github.com/sgl-project/sglang/blob/7c38eca1e4a704bf09fe6b52ea040a41d3cfc55d/python/sglang/srt/entrypoints/openai/serving_chat.py#L286`], then launch the server as usual:


The Markdown link syntax is incorrect, which breaks the link. It should be [text](url). Also, it's a good practice to link to the main branch instead of a specific commit hash to prevent the link from becoming outdated.

Suggested change

Hardcode the thinking mode to be `thinking` in (`_apply_jinja_template`)[https://github.com/sgl-project/sglang/blob/7c38eca1e4a704bf09fe6b52ea040a41d3cfc55d/python/sglang/srt/entrypoints/openai/serving_chat.py#L286`], then launch the server as usual:

Hardcode the thinking mode to be `thinking` in [`_apply_jinja_template`](https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/entrypoints/openai/serving_chat.py#L286), then launch the server as usual:

gemini-code-assist · 2025-12-02T23:30:16Z


 Test results:

+DeepSeek-V3.2-Exp：


There's a minor formatting issue. A full-width colon (：) is used here, which is inconsistent with the rest of the document. It should be a standard half-width colon (:).

Suggested change

DeepSeek-V3.2-Exp：

DeepSeek-V3.2-Exp:

Fridge003 added 4 commits December 2, 2025 12:25

upd

5e81d2d

upd

65c0ed2

upd

7b8c8ec

upd

7897067

github-actions Bot added documentation Improvements or additions to documentation deepseek labels Dec 2, 2025

gemini-code-assist Bot reviewed Dec 2, 2025

View reviewed changes

Fridge003 added 2 commits December 2, 2025 16:16

upd

4f683c7

upd

28db97b

Fridge003 merged commit 9220540 into main Dec 3, 2025
45 checks passed

Fridge003 deleted the baizhou/doc branch December 3, 2025 02:19

yingluosanqian pushed a commit to yingluosanqian/sglang that referenced this pull request Dec 4, 2025

[Doc] Update DeepSeek-V3.2 document (sgl-project#14321)

e9a63d2

tonyluj pushed a commit to openanolis/sglang that referenced this pull request Dec 5, 2025

[Doc] Update DeepSeek-V3.2 document (sgl-project#14321)

6b33de0

yuchengz816-bot pushed a commit to yuchengz816-bot/sglang that referenced this pull request Dec 8, 2025

[Doc] Update DeepSeek-V3.2 document (sgl-project#14321)

de5383e

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Doc] Update DeepSeek-V3.2 document#14321

[Doc] Update DeepSeek-V3.2 document#14321
Fridge003 merged 6 commits intomainfrom
baizhou/doc

Fridge003 commented Dec 2, 2025

Uh oh!

gemini-code-assist Bot commented Dec 2, 2025

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot Dec 2, 2025

Uh oh!

gemini-code-assist Bot Dec 2, 2025

Uh oh!

gemini-code-assist Bot Dec 2, 2025

Uh oh!

gemini-code-assist Bot Dec 2, 2025

Uh oh!

gemini-code-assist Bot Dec 2, 2025

Uh oh!

gemini-code-assist Bot Dec 2, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

	--model-path deepseek-ai/DeepSeek-V3.2 \
	--model-path deepseek-ai/DeepSeek-V3.2-Exp \

		"\| `deepseekv31` \| DeepSeek-V3.1 and DeepSeek-V3.2-Exp (e.g. `deepseek-ai/DeepSeek-V3.1`, `deepseek-ai/DeepSeek-V3.2-Exp`) \| Recommend adding `--chat-template ./examples/chat_template/tool_chat_template_deepseekv31.jinja` (Or ..deepseekv32.jinja for DeepSeek-V3.2) to launch command. \|\n",
		"\| `deepseekv32` \| DeepSeek-V3.2 (`deepseek-ai/DeepSeek-V3.2`) \| \|\n",

	DeepSeek-V3.2 model families equips DeepSeek-V3.1-Terminus with DeepSeek Sparse Attention (DSA) through continued training. With DSA, a fine-grained sparse attention mechanism powered by a lightning indexer, DeepSeek-V3.2 achieves efficiency improvements in long-context scenarios.
	DeepSeek-V3.2 model families equip DeepSeek-V3.1-Terminus with DeepSeek Sparse Attention (DSA) through continued training. With DSA, a fine-grained sparse attention mechanism powered by a lightning indexer, DeepSeek-V3.2 achieves efficiency improvements in long-context scenarios.

	Nemo Skill can't enable thinking method from client side, so we need some hardcoding before launching server:
	NeMo-Skills can't enable thinking method from client side, so we need some hardcoding before launching server:

	Hardcode the thinking mode to be `thinking` in (`_apply_jinja_template`)[https://github.com/sgl-project/sglang/blob/7c38eca1e4a704bf09fe6b52ea040a41d3cfc55d/python/sglang/srt/entrypoints/openai/serving_chat.py#L286`], then launch the server as usual:
	Hardcode the thinking mode to be `thinking` in [`_apply_jinja_template`](https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/entrypoints/openai/serving_chat.py#L286), then launch the server as usual:

Conversation

Fridge003 commented Dec 2, 2025

Motivation

Modifications

Accuracy Tests

Benchmarking and Profiling

Checklist

Uh oh!

gemini-code-assist Bot commented Dec 2, 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!

gemini-code-assist Bot Dec 2, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Dec 2, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Dec 2, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Dec 2, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Dec 2, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Dec 2, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant