Releases Note

• 1: R36 - Add more telemetry events, service health checks, and UX improvements
• 2: R34 - SmartCode chat supports #variables, Devi public preview
• 3: R32 - Optimized admin features, SmartCode chat variable support, and more
• 4: R30 - Mermaid Diagram, ARM deployment, UOS certification, and UX improvements
• 5: R28 - 体验更好/更稳定的SmartCode
• 6: R26 - DeepSeek私有化部署完整支持
• 7: R24 - DeepSeek初步支持以及SmartCode v2

1 - R36 - Add more telemetry events, service health checks, and UX improvements

Key features or improvements

1. AI Base Service - Telemetry Data Improvements

Added the telemetry data reporting of the following scenarios to improve the accuracy of plug-in usage statistics.

• User Login Logout
• Plugin update operation
• Right-click menu and codelens function trigger
• Chat actions

After adding these telemetry data reports, the following plug-in statistics can provide more accurate plug-in usage statistics, not only counting the scenarios where the user completes the code in the plug-in, but also including all the above scenarios, which will be recorded as plug-in usage.

2. AI Base Service - Added service health checks

Implement service health check function to detect the running status of all services. Perform regular health checks on all system services through background independent processes to ensure that all services are in a healthy state, and record abnormal states when problems occur. At the same time, provide health check API that can be called by the operation and maintenance system to report the system health status regularly for event monitoring and abnormal alarms.

3. SmartCode - Multiple UX improvements

• Enhancement：VSCode Plugin - Add a shortcut in Chat Input control for model switching to facilitate users to quickly switch models.

• Enhancement：JetBrains Plugin - When configuring an invalid address in the JetBrains plugin, the plugin attempts to connect to the server and reports an error. A friendly error message is required to prompt the user to check the configuration.
• Enhancement：JetBrains Plugin - Supplementary telemetry data for specific functions such as login, logout, update, etc
• Fix：JetBrains Plugin - Code completion without logging in to smartcode will prompt not logged in once. Code completion after restarting idea will not prompt not logged in.
• Enhancement：JetBrains Plugin - Treesitter parsing optimization
• Enhancement：JetBrains Plugin - Update menu “Quick Start” to make terminology more uniform and easier to understand.
• Enhancement：JetBrains Plugin - Upgrade method Enhancement, when a new version appears, the user clicks Upgrade to no longer open the IDE built-in plug-in upgrade page, but automatically completes the installation package download and installation actions.
• Enhancement：VScode Plugin - Supplementary telemetry data for specific functions such as login, logout, update, etc
• Fix：VScode Plugin - Fixed an issue in variable display to ensure that variable names and content are displayed correctly.
• Fix：VScode Plugin - Fixed variable file file name multiple selections to ensure that the file name is selected correctly and uniquely.
• Fix：VScode Plugin - Fixed the issue of deleting after selecting content to ensure that the deletion operation proceeds normally.
• Fix：VScode Plugin - Fixed an issue where content became rich text after pasting, ensuring that pasted content remains in its original format.
• Enhancement：VScode Plugin - Optimize the use of variable #files to reduce the number of unnecessary interactions.
• Enhancement：VScode Plugin - Update menu “Quick Start” to make terminology more uniform and easier to understand.
• Enhancement：VScode Plugin - When modifying the model for the current session, modify the global model settings to ensure consistency.

4. Devi Configuration UI and MacOS Support

Devi for GitHub Copilot (DB Agent) provides developers with the ability to use @devi calls in GitHub Copilot Chat to implement database conversations, Text2SQL, database document generation, etc. In this release, we have Enhancement of MySQL data connection configuration, providing the following visual configuration interface. For specific configuration methods, please refer to: [Devi Operation Manual] (https://github.com/devi-run/copilot-extension/blob/main/docs/content/en/_index.md)

This release also provides support for the MacOS operating system, starting with version 0.35.2126 8, developers can install and Devi plug-ins in Visual Studio Code on Windows and MacOS.

You can install Devi for GitHub Copilot (DB Agent) from

• Devi for GitHub Copilot
• Or you can search for’Devi’ installation through the Visual Studio Code extension tool.

Other updates

• Enhancement：AI Base Service - Enhancement of user batch import, modify the import template, user gender, account status, role is modified to drop-down selection, making it more concise and easy to fill in, which is convenient for users to import in batches according to the content of the template. Add space check to import data more conveniently.
• Enhancement：AI Base Service - Chat content management Added code hover display

2 - R34 - SmartCode chat supports #variables, Devi public preview

Key features and improvements

1. SmartCode VSCode chat supports #variables

This update to SmartCode adds support for chat variables, which are used to reference code snippet in IDE, providing users with a more flexible way to organize prompts. Three variables are provided in this release:

• #selection: reference code selected in the current active editor
• #editor：reference any code block in the current active editor
• #file：selecting files as reference

By using chat variables, developers can implement some previously difficult-to-implement prompts, such as: Developers can refer to a file content with the following prompts, and ask the AI to generate code based on the reference. The example is as follows:

• Refer to #file:BankTransactionController.java generate API documentation and output using standard markdown format.
• Refer to #file:Dockerfile write a docker-compose.yaml and set the external port to 8090, and map the logs to the local path
• Analysis #selection ，focusing on possible code security issues and providing solutions and sample code

Users can also combine the above variables to create more complex prompts, such as:

• Refer to #file:api_reference.yaml，fix the logic in the following code #selection
• Refer to #editor:model.py77-89，generate 10 unit tests to cover as many scenarios as possible

The following video demonstrates the process of code explaination in Japanese using the #selection variable

The following video demonstrates the process of using the #file variable to reference the entire file and guide the AI to summarize the file content

The following video demonstrates the process of using the #file variable to reference an API entry file and generate documentation.

This update only implements the above capabilities in VSCode. In future versions, we will implement the above capabilities in the JetBrains plugin.

2. Devi for GitHub Copilot Public Preview

Devi is your AI agent building blocks. Building agentic workflow with generative AI is a new and innovative area, developers need to build many low level basic components before they can readlly focus on the actual valuable workflow. We want to simplify the complexity of this process by building a set of common AI agent components.

GitHub Copilot 插件

Devi for GitHub Copilot is a series of AI agents built on the GitHub Copilot ecosystem. Devi provides the atomic capabilities of these agents and provides services to developers through GitHub Copilot Chat. Developers can install the Devi extension through the Visual Studio Marketplace, and then use the @devi in GitHub Copilot Chat to implement out-of-the-box capabilities of the DB agent, such as: database conversations, Text2SQL, database document generation, etc.

• Devi for GitHub Copilot
• Or you can search for’Devi’ installation through the Visual Studio Code extension tool.

For more information on how to use Devi for GitHub Copilot, please refer to Devi for GitHub Copilot 操作手册

Demo Video

Devi NPM Package

Devi NPM Package provids the atomic capabilities of an AI agent, developers can install the Devi Package through NPM and then use the capabilities provided by Devi in your own applications.

• Please refer to Devi NPM Package for details of how to use it.

Devi for GitHub Copilot 特性说明

开发者在GitHub Copilot Chat中使用 @devi 调用 代维DB智能体实现如下能力：

Chat with your Database

Automatically identify the database structure and allow users to ask questions about the database in natural language, e.g.:

• What is this database used for?
• Which database objects are related to each other?

DB Query with Natural Language (Text2SQL)

Generate database queries with natural language (Text2SQL), users can use natural language to ask questions, and @devi will generate SQL statements based on the database schema，e.g.:

• Help me generate a summary query of all apps and the corresponding number of reviews
• Help me generate a list of all users with the most reviews, with descending order

Database Document Generation

Generate standard database documents via built-in @devi /docs <table_name> shortcuts

Other updates

1. Enhancement AI Base Service - Avoiding multiple post requests on the same UI in dashboard, significantly improving the response speed of the dashboard and improving the user experience
2. Fix AI Base Service - The problem that the list is not displayed because the picture is not uploaded in the OAuth2 configuration page
3. Enhancement Handbook/Website - Allows control of video sources through whitelisting, saving public server traffic
4. Enhancement SmartCode - Show friendly message when the user does not have permission to access the current code completion model
5. Enhancement SmartCode Jetbrains - When the user opens multiple JetBrains windows at the same time, the automatic update cannot be completed. Now the plug-in side will give a friendly prompt, prompting the user to close the window before upgrading.
6. Fix SmartCode JetBrains - Fix the error caused by the IDE service not starting during the plug-in loading process, improve the stability and reliability of the plug-in loading
7. Fix SmartCode JetBrains - Optimize and solve the memory overflow problem in the process of code parsing using tree-sitter, and improve the stability and performance of plug-in dynamic parsing of code.

3 - R32 - Optimized admin features, SmartCode chat variable support, and more

Key features and improvements

1. SmartChat - Mermaid charts support zooming in and exporting

Mermaid Markdown is a tool that uses Markdown syntax to create diagrams and flowcharts. Mermaid is a JavaScript-based diagramming tool that allows users to generate complex diagrams through simple text descriptions. In the R30 update of AISE, we began to support dynamic parsing of the markdown content generated by the model that conforms to the Mermaid syntax and displayed as diagrams. This R30 version has improved this function by adding enlargement and export functions.

2. AI Base Service - Marketplace - Support uploading application installation packages

The app store page of the AISE base service supports users to upload application installation packages, simplifying the operation complexity of administrators during the application update process. Administrators can directly upload application installation packages through the application page in AISE Management | app store, and the installation packages will be automatically saved in the file server. After the application installation package is uploaded, users can receive a prompt for the application update, and click Update to complete the application update.

3. AI Base Service - Dashboard - Added dashboard administrators and dashboard user groups to refine dashboard access control

Access to the dashboard is independent of AISE administrator privileges. Dashboard administrators can add dashboard users. Users in the dashboard users group can only access the dashboard and cannot access other AISE functions. It is convenient for users to provide different data analytics permissions to users in different roles within the organization.

4. SmartCode AI Coding Assistant - Chat Variable Support and Experience Optimization

4.1 SmartCode VSCode Chat Variable Support

#variables can help users choose more flexible code contexts and combine more complex prompt words. In this update, we have added support for #variables, including:

• #selection refers to the selected code in the code editor
• #file reference file

4.2 Other UX optimization

• JetBrains code completion, after the document changes, the IDE sends the request, from the last time, the delay is changed to 400 milliseconds
• JetBrains settings interface added: code parsing configuration, default: off Enable: Prompt, and model return code will be parsed; off: Prompt (default cursor before 2500, cursor after 2000 characters), and model return code will not be parsed
• Android Studio has not set JCEF. After installing the plug-in, the startup project will not flash back, which will not affect code completion, but Chat related functions are not available.

5. AISE has obtained the certification of Zhaoxin, Dongfang, Kirin and other Xinchuang localized manufacturers

AISE has been committed to providing complete support for the localized Xinchuang environment, and actively promotes cooperation and certification with various localized chip, operating system, and middleware manufacturers. In this R32 version, we have added certifications for Zhaoxin, Dongfang, Kirin and other manufacturers to provide users with a more stable and secure localized environment.

Kirin software certification

Zhaoxin certification

Dongtong certification

Other updates

• AI Base Services - Add a document link at the top of the page to make it easier for users to access documents.
• AI Base Services - Dashboard selection department and select the root department inconsistency, this R32 version fixes the selection department and select the root department logic to make them consistent.
• AI Base Services - Fix the problem that “normal user” cannot be found in the role list when the administrator adds new users after deleting all ordinary users, and fix the role list query logic to ensure that even after deleting all ordinary users, the administrator can still find the “normal user” role when adding new users.

4 - R30 - Mermaid Diagram, ARM deployment, UOS certification, and UX improvements

Major Feature and Updates

1. SmartChat - Initial support for generating mermaid charts

Mermaid Markdown is a tool that uses Markdown syntax to create diagrams and flowcharts. Mermaid is a JavaScript-based charting tool that allows users to generate complex diagrams through simple text descriptions. In this AISE update, SmartChat browser has started to support dynamic parsing of model-generated markdown content that conforms to Mermaid syntax and displays it as diagrams. As shown in the following video demonstration:

2. SmartCode AI Coding Assistant - UX Updates

• Codelens triggering method optimization: allows users to directly trigger code intelligence tasks without opening chat window
• Allows users to archive all session history with one click
• Optimize the plug-in version information display interface and unify the information display format
• Optimize the error prompt when the user triggers the code intelligence task in the unlogged state
• Optimize the way codelens select code blocks, ensuring that the indentation in front of the method body is fully selected to optimize the model response
• The codelens of the JetBrains plugin allow users to temporarily turn off the display of the current tab page, and automatically display it when reopened
• Improve the content loading of the chat window after the user exits the IDE, ensuring that the last conversation history is loaded according to the workspace.
• Code review prompt word optimization: ensure that the model responds reasonably to problem-free code, 1) optimize the output message format, use more explicit number and bold display; 2) fix the problem of repeatedly outputting the same error message in an infinite loop.
• JetBrains dynamically analyzes the code generated by the model, only displays the content of the complete code block, and truncates the subsequent content.

3. Support for deployment on ARM-based CPU and operating systems

The AISE base service has been able to run fully on the ARM operating system, and has passed the compatibility test of the UOS operating system Unixin UOS, meeting the compatibility requirements of the Domestically Developed Information Technology Innovation in terms of operating system and the domestic CPU, and has been awarded an official certification certificate; supported CPU platforms include: Haiguang 5000, Haiguang 7000, Zhaoxin KH-30000, Zhaoxin KH-20000, Kunpeng 916, Kunpeng 920, Feiteng FT-2000 +/64 and Tengyun S250.

4. AI base services - UX updates

• User import and export optimization: Support for specifying user roles during import
• For Azure OpenAI service connections that are dynamically authenticated using Azure Service Principle, an SPN authentication feature is provided to ensure that users have correct configuration items when configuring the SPN, and a friendly prompt message is provided if the SPN expires during use.
• Client side feedback optimization for model service exceptions: The AISE AI Base service supports parallel connections to multiple LLM providers, including Azure OpenAI and DeepSeek online services, and private deployment of LLMs running by enterprises through local GPU computing power. It is very common for model services to have intermittent networks or unstable operation. AISE specifically optimizes the user experience for these scenarios, ensuring user-friendly prompts when the model is invalid, and uses a dedicated error status code 12017401 to record background logs, which is convenient for system admins to troubleshoot problems.

Others

• AI base service - optimize the personal message management page, including message data statistics and filtering logic to ensure the normal operation of the number of messages and filtering conditions
• AI pedestal service - app store page (optimized) to avoid unnecessary re-login operations
• AI Dock Service - Fix the internal display of the release notes page

5 - R28 - 体验更好/更稳定的SmartCode

主要特性或改进

1. SmartCode用户体验优化

• 优化SmartCode 修复代码补全的一些细节问题：包括可能出现无关代码，token过期时没有给出友好提示等
• 优化代码评审更改提示词模版,使模型返回的答案效果更好，符合用户预期
• 优化单元测试提示词模版，根据新的模型模版返回正确的单元测试用例，符合用户预期
• 优化模型被禁用后的用户体验，优化提示信息
• 对话UX优化，针对长文本流式更新过程中用户查看信息场景进行优化，允许用户混动视图到已经输出的内容，同时停止自动滚动
• 增加关闭 codelens 按钮，允许用户禁用悬浮菜单（codelens）入口；同时在配置页面中允许用户重新启动。
• 增加对不同JDK版本的识别，并在代码补全中引导模型使用对应JDK版本的编码风格

2. 对话信息持久化特性改进

• 修复因用户消息内包含特殊字符造成的数据保存失败问题：当用户在对话中引入类似URL的内容，有可能引发后台数据报错失败。本次更新优化了这个场景，确保任何类型的用户输入数据很可以正常保存。
• 修复插件端报 Invalid SessionId 错误时造成用户无法继续对话的问题

其他更新

• 修复当用户token过期时，插件smartcode vscode一直处于登录状态，没有给出正确登录状态提示
• 修复补码补全功能 用户java开发环境JDK8，但AISE补全的java代码不能被jdk8所识别并报java语法错误。
• 修复smartcode jetbrains插件 ，代码补全代码中出现无关的代码现象。
• 修复重新滚动到底部后无法再次触发自动滚动
• 修复smartcode-vscode chat功能，用户在回答过程中，切换会话，导致会话卡住
• 修复smartcode vscode 增加codelens关闭功能，实现对上下文悬浮菜单关闭。
• SmartCode ChatApp 自动滚动更新，应该参考ChatGPT的方式，在消息超出一屏自动滚动过程中允许用户手工混动到正在输出的消息的中间位置，此时应该停止自动混动
• SmartCode Chat 模型被禁用后的用户体验改进，提示信息优化
• 修复smartcode jetbrain的报错：提示 Invalid json… …
• 完善smartCode JetbBrains完善升级提示
• 修复插件端补全代码非模型服务器错误未有友好提示，模型服务器返回状态码为200,orch返回状态码500
• 修复smartcode jetbrain(2.25.20104)存在错误信息的错误编码不规范，存在重复
• 优化 Orch日志文件大小： 15G，一个文件。建议增加Orch日志生成分片策略 。按天分片。
• 修复【Core】16k模型传入长token会导致历史问题丢失 - 原因是chat-data-v2 数据字段类型需要调整为text以便支持长文本
• 修复smartcode jetbrains插件提示 “1102911029401 Invalid SessionId！”

6 - R26 - DeepSeek私有化部署完整支持

主要特性或改进

1. 对DeepSeek模型提供完整支持

针对DeepSeek 6.7 Base 和 DeepSeek 6.7 Instruct模型的支持已经完整，已经可以支持以下场景

• 使用 DeepSeek 6.7 Base 模型进行代码补全和提示词（完成模式）模板的支持，包括：explain, review, unit test 和 code check 四个提示词均可以正常使用。
• 使用 DeepSeek 6.7 Instruct 模型进行对话，在SmartChat Web （ChatUI）中已经支持

部署环境支持

• 使用 Nvidia A10 部署 DeepSeek 6.7 Base 和 Instruct 模型并支持以上场景
• 使用 华为晟腾910B/310部署 DeepSeek 6.7 Base和Instruct模型
  • 提供初始版本的部署手册 华为910/310系列mindie离线部署方案和 FT038 - DeepSeek 模型适配。手册内容仍然需要打磨和整理
  • 对关键模型参数组合进行验证，主要包括：maxSeqLen, maxIterTimes 和 input/prompt入参长度问题的探索 模型参数设置合理值

2. SmartChat 多模态 - 支持GPT4o模型，上传图片进行识别

在SmartChat中上传图像，针对图像内容进行理解，解析和内容生成,默认是关闭状态，可按照如下顺序打开图片识别开关

1. 首先打开 aise.feature.chat_image 这个开关
2. 在 aise.model.config.default 里面正确配置 image_model_type
3. 设置用户对于image_model_type 所对应的模型权限访问 备注：图片上传识别功能的使用与用户当前对话模型无关，是一个独立任务；只要以上条件满足，用户可以在与任何模型对话过程中使用此功能。 后续还需要调用通过私有化模型提供类似多模态能力，

3. 针对华为晟腾NPU+mindIE部署的优化

针对使用华为晟腾910/310 NPU作为算力，使用 mindIE 作为推理服务的环境，优化了模型，推理服务端的部署参数以及对应的Deepseek模型的个性化参数，确保硬件可以稳定运行。

4. SmartCode 用户体验改进

• 针对deepseek模型在华为晟腾mindIE推理服务下流式返回数据进行适配，修复信息截断问题
• 插件端动态获取模型适配器上的默认模型行为参数，针对不同模型进行最优化的默认行为控制
• 插件端增加模型行为控制界面，允许对模型行为进行动态调整
• 插件信息增加服务器地址信息，方便问题定位
• VSCode 菜单增加 “重新加载”选项，方便用户在插件异常情况下回复插件状态
• JetBrains 简化配置菜单，去除与用户无关的设置
• JetBrains 增加日志级别选项：信息/调试
• JetBrains 修复设置私有插件市场链接中可能出现的重复设置问题
• JetBrains codelens新增X关闭按钮，设置界面新增启用、禁用设置；用户可以通过配置界面再次开启 codelens。

其他更新

• 【插件】SmartCode Chat 适配模型适配器参数和个性化参数支持
• smartcode vscode 增加codelens关闭功能，实现对上下文悬浮菜单关闭。
• 修复【插件】SmartCode Chat 历史对话一直加载无法打开
• 修复【插件】SmartCode Chat 在 晟腾910/310 + MindIE 环境运行 DeepSeek 6.7B Instruct模型下对话信息被截断问题
• 修复【插件】【代码补全】SmartCode JetBrains 代码上下文结构化分析 - 多行补全模式下，实现对生成内容的按方法体截断
• 修复smartcode-jetbrains《上下文悬浮》菜单位置出现不对。
• 优化 smartcode-vscode chat功能，用户在回答过程中，切换会话，导致会话卡住。
• 修复博时 客户smartcode-vscode(2.24.19606) 插件端，提示警告信息：SmartCode is encountering temporary issues 。在什么情况下提示上述信息
• 修复 smartcode-vscode 对话task，过程中会输出到一半停止，停止后不输出内容，点停止按钮也不管用
• 【插件】SmartCode Chat 适配模型适配器参数和个性化参数支持
• 修复【插件】SmartCode Chat 历史对话一直加载无法打开
• 完善增加完模型端异常排查提示处理机制
• 实现单机部署模式
• 修复当AISE升级后 ，出现用户头像丢失情况
• 修复菜单有问题，《会话明细信息》不应该显示，将菜单变更增加到脚本变更中。
• 实现【多模态支持】SmartChat for Web 支持上传图片，调用 GPT4o 模型进行图片内容提取
• 【RAG】Workspace 增加deepseek API支持，方便个人开发者试用
• 【RAG】Workspace 加入 AISE 服务器，见说明详情

7 - R24 - DeepSeek初步支持以及SmartCode v2

主要特性或改进

1. 对DeepSeek模型的基本支持

针对DeepSeek 6.7 Base 和 DeepSeek 6.7 Instruct模型的支持基本完整，24_2 版本已经可以支持以下场景

• 使用 DeepSeek 6.7 Base 模型进行代码补全和提示词（完成模式）模板的支持，包括：explain, review, unit test 和 code check 四个提示词均可以正常使用。
• 使用 DeepSeek 6.7 Instruct 模型进行对话，在SmartChat Web （ChatUI）中已经支持，但是 SmartCode v2中的内置Chat组件仍然有兼容问题（nvidia卡部署工作正常，晟腾卡部署有问题），预计在 sprint 25 修复。 部署环境支持
• 使用 Nvidia A10 部署 DeepSeek 6.7 Base 和 Instruct 模型并支持以上场景
• 使用 华为晟腾部署 DeepSeek 6.7 Base和Instruct模型并支持以上场景仍然有问题，预计在Sprint25修复。

2. SmartCode V2 功能基本完整

SmartCode V2 主要增加内置chat能力，在24版本中主要针对提示词库提供了动态加载支持，并且修复了不少客户现场问题。 当前已经支持的场景有

• GPT模型的完整支持：使用GPT 3.5 Instruct模型进行代码补全和提示词（完成模式）的生成，使用 GPT 3.5, GPT 3.5 16K, GPT 4, GPT 4 32k 和 GPT-4o模型进行普通对话。
• DeepSeek模型在NV环境下的完整支持：使用 deepseek 6.7 base模型进行代码补全和提示词（完成模式）的生成，使用 deepseek 6.7 instruct 和 deepseek API Coder/Chat 模型进行普通对话

3. 使用 GPT-4o 进行普通对话

已经支持添加 GPT-4o模型支持普通对话，暂不支持多模态能力。

其他更新

• SmatCode V2 支持提示词仓库，确保提示词数据动态获取
• 优化SmartCode Chat功能 登录体验；
• SmartChat for Web 支持上传图片，调用 GPT4o 模型进行图片内容提取
• 修复Core各调用redis服务端容器日志都显示有重新连接redis日志，Reconnecting to aise-redis
• 修复【RAG】Workspace 修改打包方式为 dir 方式，优化启动速度。当前的onefile方式启动速度太慢，主要是因为需要动态解压文件造成。尝试修改成dir的方式，看看是否能有较大提升。
• 实现Workspace 加入 AISE 服务器
• 修复博时 UAT测试 优化 smartcode-vscode(2.24.19435-rel2) 右键 代码检查 功能已正常。 但该提示词基于不同方法（完成 \对话 ）生成反馈结果体验有差异，用户更接受采用对话方法产生的结果
• 修复博时 UAT测试 smartcode-vscode(2.24.19435-rel2) core(spring23-1) 测试右键smartcode->代码检查 vscode插件报错
• 优化DeepSeek 模型适配，验证适配效果，模型配置和提示词版本匹配情况。包括：API 模式，VLLM模式（N卡使用vllm部署），TGI模式（晟腾910/310使用mindie部署）。整理DeepSeek部署标准化配置手册，包括代码补全、代码任务和对话场景。
• DeepSeek在晟腾环境部署（mindIE）时使用 generate_steam 进行chat操作的 chat 模板改造
  • 晟腾910B3/310部署方案，交付完整端到端部署文档
  • 模型优化】DeepSeek 在 N卡上的部署文档