[Doc][Model] Add DeepSeek-V4 Flash and Pro documentation by GDzhu01 · Pull Request #9995 · vllm-project/vllm-ascend

GDzhu01 · 2026-06-04T06:47:09Z

What this PR does / why we need it?

Add tutorial documentation for DeepSeek-V4 Flash and DeepSeek-V4 Pro models. Also includes fixes to the supported models matrix. This provides users with comprehensive guides on running DeepSeek-V4 models on Ascend hardware.

Changes:

New tutorial pages: DeepSeek-V4-Flash.md and DeepSeek-V4-Pro.md
Update models index to include both new pages
Update supported models matrix with DeepSeek V4-Flash and V4-Pro entries

Does this PR introduce any user-facing change?

No, this is a documentation-only update.

How was this patch tested?

Documentation-only change, CI verification is sufficient.

vLLM version: v0.20.2
vLLM main: vllm-project/vllm@9090368

Signed-off-by: GDzhu01 <809721801@qq.com>

gemini-code-assist · 2026-06-04T06:47:18Z

Summary of Changes

Hello, 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 new documentation for the DeepSeek-V4-Flash and DeepSeek-V4-Pro models to support users running these models on Ascend hardware. It also ensures the project's model support matrix and index are kept up to date with these additions.

Highlights

New Documentation: Added comprehensive tutorial pages for DeepSeek-V4-Flash and DeepSeek-V4-Pro models, detailing environment setup, deployment, and evaluation.
Matrix Updates: Updated the supported models index and the supported models matrix to include DeepSeek-V4-Flash and DeepSeek-V4-Pro.

New Features

🧠 You can now enable Memory (public preview) 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.

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 the 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 counterproductive. 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.

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. ↩

github-actions · 2026-06-04T06:47:41Z

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

A PR should do only one thing, smaller PRs enable faster reviews.
Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

gemini-code-assist

Code Review

This pull request introduces comprehensive deployment and verification documentation for the DeepSeek-V4-Flash and DeepSeek-V4-Pro models on Ascend hardware, updating both the models index and the supported models matrix. The review feedback suggests minor documentation improvements, including correcting a numbering inconsistency, fixing a typo, resolving a redundancy in the Pro model's description, and updating hardcoded link text in the support matrix to match specific model variants. Additionally, in accordance with the repository style guide, the suggested PR Title and PR Summary are provided below:

Suggested PR Title:

[Doc][Feature] Add DeepSeek-V4 Flash and Pro documentation

Suggested PR Summary:

### What this PR does / why we need it?

This PR adds comprehensive tutorial documentation for deploying and running the DeepSeek-V4 Flash and DeepSeek-V4 Pro models on Ascend hardware (Atlas 800 A2/A3). It includes guides for single-node deployment, multi-node deployment, and Prefill-Decode (P-D) disaggregation. Additionally, it updates the supported models matrix and the models index.

### Does this PR introduce _any_ user-facing change?

No, this is a documentation-only update.

### How was this patch tested?

Documentation-only change, verified via local rendering/CI.

gemini-code-assist · 2026-06-04T06:48:27Z

 Qwen3.5-397B-A17B.md
 DeepSeek-V3.1.md
 DeepSeek-V3.2.md
+DeepSeek-V4-Flash.md


To comply with the repository style guide, here are the suggested PR Title and PR Summary formatted in markdown code blocks for easy copying.\n\nSuggested PR Title:\n\nmarkdown\n[Doc][Feature] Add DeepSeek-V4 Flash and Pro documentation\n\n\nSuggested PR Summary:\n\nmarkdown\n### What this PR does / why we need it?\n\nThis PR adds comprehensive tutorial documentation for deploying and running the DeepSeek-V4 Flash and DeepSeek-V4 Pro models on Ascend hardware (Atlas 800 A2/A3). It includes guides for single-node deployment, multi-node deployment, and Prefill-Decode (P-D) disaggregation. Additionally, it updates the supported models matrix and the models index.\n\n### Does this PR introduce _any_ user-facing change?\n\nNo, this is a documentation-only update.\n\n### How was this patch tested?\n\nDocumentation-only change, verified via local rendering/CI.\n

References

The repository style guide requires providing a suggested PR Title and PR Summary in a specific format using markdown code blocks during PR reviews. ^(link)

gemini-code-assist · 2026-06-04T06:48:27Z

+| DeepSeek V4-Flash               | ✅        |                                                                      | ✅ | A2/A3 | ✅ | ✅ |✅|| ✅ |✅| ✅ || ✅ | ✅ | ✅ || ✅ | 1M || [DeepSeek-V4](../../tutorials/models/DeepSeek-V4-Flash.md) |
+| DeepSeek V4-Pro               | ✅        |                                                                      | ✅ | A2/A3 | ✅ | ✅ |✅|| ✅ |✅| ✅ || ✅ | ✅ | ✅ || ✅ | 1M || [DeepSeek-V4](../../tutorials/models/DeepSeek-V4-Pro.md) |


The link text for both DeepSeek V4-Flash and DeepSeek V4-Pro is currently hardcoded as [DeepSeek-V4]. To maintain consistency with other models in the table (such as [DeepSeek-V3.1] and [DeepSeek-V3.2]) and to avoid confusion, the link text should be updated to reflect the specific model variant.

Suggested change

| DeepSeek V4-Flash | ✅ | | ✅ | A2/A3 | ✅ | ✅ |✅|| ✅ |✅| ✅ || ✅ | ✅ | ✅ || ✅ | 1M || [DeepSeek-V4](../../tutorials/models/DeepSeek-V4-Flash.md) |

| DeepSeek V4-Pro | ✅ | | ✅ | A2/A3 | ✅ | ✅ |✅|| ✅ |✅| ✅ || ✅ | ✅ | ✅ || ✅ | 1M || [DeepSeek-V4](../../tutorials/models/DeepSeek-V4-Pro.md) |

| DeepSeek V4-Flash | ✅ | | ✅ | A2/A3 | ✅ | ✅ |✅|| ✅ |✅| ✅ || ✅ | ✅ | ✅ || ✅ | 1M || [DeepSeek-V4-Flash](../../tutorials/models/DeepSeek-V4-Flash.md) |\n| DeepSeek V4-Pro | ✅ | | ✅ | A2/A3 | ✅ | ✅ |✅|| ✅ |✅| ✅ || ✅ | ✅ | ✅ || ✅ | 1M || [DeepSeek-V4-Pro](../../tutorials/models/DeepSeek-V4-Pro.md) |

gemini-code-assist · 2026-06-04T06:48:27Z

+python launch_online_dp.py --dp-size 4 --tp-size 4 --dp-size-local 4 --dp-rank-start 0 --dp-address xx.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 7100
+```
+
+3. Decode node 0


There is a numbering inconsistency in the deployment steps. The list goes from 1. Prefill node directly to 3. Decode node 0. It should be updated to 2. Decode node 0.

Suggested change

3. Decode node 0

2. Decode node 0

gemini-code-assist · 2026-06-04T06:48:27Z

+
+Finally, Refer to [Prefill-Decode Disaggregation (Deepseek)](../features/pd_disaggregation_mooncake_multi_node.md) to deploy the P-D disaggregation proxy.
+
+For Atlas 800 A2 series machines, we can configure the deployment(4\*1P 1\*4D) as follows:.


There is a minor typo at the end of the sentence (as follows:. has a trailing period after the colon). Also, adding a space before the parenthesis improves readability.

Suggested change

For Atlas 800 A2 series machines, we can configure the deployment(4\*1P 1\*4D) as follows:.

For Atlas 800 A2 series machines, we can configure the deployment (4\*1P 1\*4D) as follows:

gemini-code-assist · 2026-06-04T06:48:27Z

+- The Manifold-Constrained Hyper-Connections (mHC) to strengthen conventional residual connections;
+- A hybrid attention architecture, which greatly improves long-context efficiency through Compress-4-Attention and Compress-128-Attention. For the Mixture-of Experts (MoE) components, it still adopt the DeepSeekMoE architecture, with only minor adjustments.
+
+DeepSeek-V4-Pro, the maximum reasoning effort mode of DeepSeek-V4-Pro, significantly advances the knowledge capabilities of open-source models, firmly establishing itself as the best open-source model available today. It achieves top-tier performance in coding benchmarks and significantly bridges the gap with leading closed-source models on reasoning and agentic tasks.


The phrase 'the maximum reasoning effort mode of DeepSeek-V4-Pro' is redundant. It should be corrected to 'the maximum reasoning effort mode of DeepSeek-V4' to accurately describe the relationship between the Pro variant and the base model.

Suggested change

DeepSeek-V4-Pro, the maximum reasoning effort mode of DeepSeek-V4-Pro, significantly advances the knowledge capabilities of open-source models, firmly establishing itself as the best open-source model available today. It achieves top-tier performance in coding benchmarks and significantly bridges the gap with leading closed-source models on reasoning and agentic tasks.

DeepSeek-V4-Pro, the maximum reasoning effort mode of DeepSeek-V4, significantly advances the knowledge capabilities of open-source models, firmly establishing itself as the best open-source model available today. It achieves top-tier performance in coding benchmarks and significantly bridges the gap with leading closed-source models on reasoning and agentic tasks.

…t#9995) ## What this PR does / why we need it? Add tutorial documentation for DeepSeek-V4 Flash and DeepSeek-V4 Pro models. Also includes fixes to the supported models matrix. This provides users with comprehensive guides on running DeepSeek-V4 models on Ascend hardware. Changes: - New tutorial pages: `DeepSeek-V4-Flash.md` and `DeepSeek-V4-Pro.md` - Update models index to include both new pages - Update supported models matrix with DeepSeek V4-Flash and V4-Pro entries ## Does this PR introduce _any_ user-facing change? No, this is a documentation-only update. ## How was this patch tested? Documentation-only change, CI verification is sufficient. - vLLM version: v0.20.2 - vLLM main: vllm-project/vllm@9090368 --------- Signed-off-by: GDzhu01 <809721801@qq.com>

… into main-dsv4 * 'main-dsv4' of https://github.com/chen-commits/vllm-ascend: (35 commits) 测试qwen3.5-122B [BugFix][UT] Fix platform eager mode log assertion (vllm-project#10075) [Misc] Use logger.info_once for enable_npugraph_ex log (vllm-project#9699) [CI] delete DeepSeek-R1-W8A8.yaml update DeepSeek-R1-0528-W8A8-EPLB.yaml (vllm-project#10055) [CI] fix ci error when rebase on latest main (vllm-project#10057) [CI] Upgrade the image version of pr_e2e_comment (vllm-project#10052) [BugFix] Include zero cached prompt token details (vllm-project#9911) [Doc][Misc] Update release notes known issues (vllm-project#10039) [Doc]Update MiniMax-M2.5.md change A2 to single node (vllm-project#9481) [Doc]Add InternVL3.5 documentation (vllm-project#9980) [BugFix]Fix the question that cant find num_kv_head from some model config (vllm-project#10008) [Ascend950][BugFix] Fix MoonViT3dPretrainedModel.to overriding quantized ViT weight dtype (vllm-project#9929) [CI][Test] Optimize selective test routing and CI test stability (vllm-project#9994) [BugFix] fix dsv4 piecewise scenario (vllm-project#10003) [Test][CI] final part for selected test (vllm-project#10017) [Feature] Enable prefix caching with PCP/DCP (vllm-project#9638) [Doc]Update GLM-5 documentation and support matrix (vllm-project#9889) [CI] add MiniMax A2 acc case (vllm-project#9855) [Bugfix] [kimi] upgrade transformers from 5.5.3 to 5.5.4 (vllm-project#9996) [Doc][Model] Add DeepSeek-V4 Flash and Pro documentation (vllm-project#9995) ...

…t#9995) ## What this PR does / why we need it? Add tutorial documentation for DeepSeek-V4 Flash and DeepSeek-V4 Pro models. Also includes fixes to the supported models matrix. This provides users with comprehensive guides on running DeepSeek-V4 models on Ascend hardware. Changes: - New tutorial pages: `DeepSeek-V4-Flash.md` and `DeepSeek-V4-Pro.md` - Update models index to include both new pages - Update supported models matrix with DeepSeek V4-Flash and V4-Pro entries ## Does this PR introduce _any_ user-facing change? No, this is a documentation-only update. ## How was this patch tested? Documentation-only change, CI verification is sufficient. - vLLM version: v0.20.2 - vLLM main: vllm-project/vllm@9090368 --------- Signed-off-by: GDzhu01 <809721801@qq.com>

nwpu-zxr · 2026-06-17T11:50:06Z

+    --quantization ascend \
+    --port 8900 \
+    --block-size 128 \
+    --speculative-config '{"num_speculative_tokens": 1,"method": "mtp","enforce_eager": true}' \


We should inform the user that the Prefill node is best used only with mtp1, while the decode node can use mtp1 or 3, etc.

…t#9995) ## What this PR does / why we need it? Add tutorial documentation for DeepSeek-V4 Flash and DeepSeek-V4 Pro models. Also includes fixes to the supported models matrix. This provides users with comprehensive guides on running DeepSeek-V4 models on Ascend hardware. Changes: - New tutorial pages: `DeepSeek-V4-Flash.md` and `DeepSeek-V4-Pro.md` - Update models index to include both new pages - Update supported models matrix with DeepSeek V4-Flash and V4-Pro entries ## Does this PR introduce _any_ user-facing change? No, this is a documentation-only update. ## How was this patch tested? Documentation-only change, CI verification is sufficient. - vLLM version: v0.20.2 - vLLM main: vllm-project/vllm@9090368 --------- Signed-off-by: GDzhu01 <809721801@qq.com>

GDzhu01 added 2 commits June 4, 2026 14:45

support deepseekv4 docs

d8fc07d

Signed-off-by: GDzhu01 <809721801@qq.com>

fix lint

2baf2ac

Signed-off-by: GDzhu01 <809721801@qq.com>

GDzhu01 requested review from LCAIZJ, Yikun and wangxiyuan as code owners June 4, 2026 06:47

github-actions Bot added the documentation Improvements or additions to documentation label Jun 4, 2026

gemini-code-assist Bot reviewed Jun 4, 2026

View reviewed changes

weijinqian0 approved these changes Jun 4, 2026

View reviewed changes

weijinqian0 merged commit 3dc73a6 into vllm-project:main Jun 4, 2026
7 checks passed

nwpu-zxr reviewed Jun 17, 2026

View reviewed changes

Navigation Menu

Uh oh!

[Doc][Model] Add DeepSeek-V4 Flash and Pro documentation - #9995

[Doc][Model] Add DeepSeek-V4 Flash and Pro documentation#9995
weijinqian0 merged 2 commits into
vllm-project:mainfrom
GDzhu01:main_v4_docs

GDzhu01 commented Jun 4, 2026 •

edited by github-actions Bot

Loading

Uh oh!

gemini-code-assist Bot commented Jun 4, 2026

Uh oh!

github-actions Bot commented Jun 4, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot Jun 4, 2026

Uh oh!

gemini-code-assist Bot Jun 4, 2026

Uh oh!

gemini-code-assist Bot Jun 4, 2026

Uh oh!

gemini-code-assist Bot Jun 4, 2026

Uh oh!

gemini-code-assist Bot Jun 4, 2026

Uh oh!

Uh oh!

nwpu-zxr Jun 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

		\| DeepSeek V4-Flash \| ✅ \| \| ✅ \| A2/A3 \| ✅ \| ✅ \|✅\|\| ✅ \|✅\| ✅ \|\| ✅ \| ✅ \| ✅ \|\| ✅ \| 1M \|\| [DeepSeek-V4](../../tutorials/models/DeepSeek-V4-Flash.md) \|
		\| DeepSeek V4-Pro \| ✅ \| \| ✅ \| A2/A3 \| ✅ \| ✅ \|✅\|\| ✅ \|✅\| ✅ \|\| ✅ \| ✅ \| ✅ \|\| ✅ \| 1M \|\| [DeepSeek-V4](../../tutorials/models/DeepSeek-V4-Pro.md) \|


		Finally, Refer to [Prefill-Decode Disaggregation (Deepseek)](../features/pd_disaggregation_mooncake_multi_node.md) to deploy the P-D disaggregation proxy.

		For Atlas 800 A2 series machines, we can configure the deployment(4\1P 1\4D) as follows:.

	DeepSeek-V4-Pro, the maximum reasoning effort mode of DeepSeek-V4-Pro, significantly advances the knowledge capabilities of open-source models, firmly establishing itself as the best open-source model available today. It achieves top-tier performance in coding benchmarks and significantly bridges the gap with leading closed-source models on reasoning and agentic tasks.
	DeepSeek-V4-Pro, the maximum reasoning effort mode of DeepSeek-V4, significantly advances the knowledge capabilities of open-source models, firmly establishing itself as the best open-source model available today. It achieves top-tier performance in coding benchmarks and significantly bridges the gap with leading closed-source models on reasoning and agentic tasks.

Uh oh!

Conversation

GDzhu01 commented Jun 4, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What this PR does / why we need it?

Does this PR introduce any user-facing change?

How was this patch tested?

Uh oh!

gemini-code-assist Bot commented Jun 4, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

github-actions Bot commented Jun 4, 2026

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Jun 4, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jun 4, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jun 4, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jun 4, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Jun 4, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

nwpu-zxr Jun 17, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

GDzhu01 commented Jun 4, 2026 •

edited by github-actions Bot

Loading