dsh-qwen-mode
Manifest validQwen/GenieX providers, reasoning parser, warmup and per-turn TTFT/TPS for DeepSeek Harness.
DSH Qwen Mode / DSH Qwen 模式
中文
在 DSH 中使用 Qwen:连接本机 GenieX,或连接阿里云百炼 QwenCloud。Qwen 模式会自动发现并维护可用模型,把思考模式、采样、预热、MTP、用量与每轮统计集中到设置页和对话界面。
功能
- 自动维护 From GenieX / From QwenCloud 模型列表,模型来源变化后自动更新。
- 自动识别每个模型的上下文窗口、输出上限、图片输入、思考模式与 MTP 支持。
- 思考模式按模型支持的等级选择;QwenCloud 只发送模型支持的等级,GenieX 只在模型支持时开启思考。
- 每轮显示模型名、首字延迟、解码速度与耗时;包含工具调用的轮次显示本轮平均值。
- 会话内显示 GenieX 的累计 Token 用量,并区分输入与输出。
- 可选预热:开始输入消息后,将选中的 From GenieX 模型预加载到内存,模型列表用状态圆点显示加载情况。
- 采样参数可手动填写,也可读取模型推荐值,并支持固定随机种子。
- 可选 MTP 加速,仅适用于支持 MTP 的 From GenieX 本地模型。
- 破限 / 无审查模型名称高亮。
- 可选「仅回复」:隐藏并拒绝工具调用,只进行纯文本对话。
系统要求
- DSH
>=0.2.2-rc.1 <0.3.0-0(0.2.2 预发布线及以上、0.3 之前),并支持插件安装。 - 使用 GenieX 时,本机需要已安装并运行 GenieX,且模型已下载。
- 使用 QwenCloud 时,需要可用的阿里云百炼兼容端点与 API Key。
安装
- 在 DSH 中执行
dsh plugin --profile <profile> add github:Fuxfantx/dsh-qwen-mode,或通过 DSH 的插件管理安装离线包或插件目录。 - 刷新 DSH 页面。
- 打开 设置 → Qwen 模式。
使用
- 选择 关闭 或 自动。只有自动模式会维护模型列表。
- 选择来源并填写地址:
- GenieX:填写本机地址,默认为
http://127.0.0.1:18181。 - QwenCloud:填写完整的兼容端点,例如
https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1。
- GenieX:填写本机地址,默认为
- 使用 QwenCloud 时,填写保存 API Key 的环境变量名(不是 Key 本身)。GenieX 不需要 Key。
- 按需调整上下文、最大输出、思考、采样、预热、MTP 与推理后端。
- 点击 保存并应用。
- 在模型选择器中选择 From GenieX 或 From QwenCloud 下的模型。
设置说明
- 上下文窗口:留空表示暂不指定;勾选「优先使用模型上下文阈限」会使用模型自身的上下文上限。
- 最大输出:默认 16384,不会超过上下文窗口或模型自身的上限。
- 压缩提示:设置页会提示 DSH 压缩阈值与推荐的压缩门槛,并提示当前上下文是否足够。
- 采样参数:只补充请求中缺失的字段;留空表示不注入。部分参数仅对 GenieX 生效;可使用「读取推荐值」填入模型推荐参数;可固定随机种子。
- 预热:启用后,开始输入消息时会将选中的 From GenieX 模型预加载到内存。圆点状态:黄色加载中、绿色已加载、红色未加载、蓝色其它来源。
- MTP:仅适用于支持的 From GenieX 本地模型。启用后发送图片时,会要求选择「直接发送」或「本次关闭 MTP」。
- 推理后端:Compute 与 Vision Encoder 可留空,由 GenieX 决定;Vision Encoder 仅对图片模型生效。
- 仅回复:从权限选择器选择后,当前会话隐藏并拒绝工具调用;再次选择其它权限即可恢复。插件停用或卸载时该预设会随之移除,DSH 自带的其它权限预设不会被改写。
限制
- 预热只减少模型加载时间,不减少长提示词的处理时间。
- GenieX 同一时间只加载一个模型;不同会话使用不同模型时会互相卸载。
- MTP 与图片不能同时生效,发送前需要选择。
- 模型列表由模型来源提供;来源不可达时不会出现模型,插件会在后台重试。
- 预热、MTP 与推理后端仅适用于 From GenieX;QwenCloud 保留模型管理、思考、采样、用量与每轮统计、仅回复。
- 每轮统计与 GenieX 用量只统计 Qwen 模式下的回复。
卸载
从 Plugins 页停用或卸载插件即可。「仅回复」权限预设随插件移除。
English
Use Qwen in DSH: connect to models running on a local GenieX server, or to QwenCloud on Alibaba Cloud Model Studio. Qwen Mode discovers and maintains the available models automatically, and brings thinking mode, sampling, warmup, MTP, usage, and per-turn statistics into the settings page and chat UI.
Features
- Maintains the From GenieX / From QwenCloud model lists automatically and refreshes them when the model source changes.
- Recognizes each model's context window, output limit, image input, thinking modes, and MTP support.
- Thinking mode is chosen from the levels the model supports. QwenCloud receives only supported levels, and GenieX enables thinking only on models that support it.
- Shows the model name, first-response delay, decode speed, and elapsed time per turn; turns that use tools show a turn average.
- Shows cumulative GenieX token usage for the session, split into input and output.
- Optional warmup: once you start typing, preload the selected From GenieX model, with loading-status dots in the model list.
- Sampling parameters can be entered manually or read from the model's recommended values, including a fixed random seed.
- Optional MTP acceleration for supported local From GenieX models.
- Highlights abliterated / uncensored model names.
- Optional Reply Only: hides and denies tool calls, keeping replies plain text.
System requirements
- DSH
>=0.2.2-rc.1 <0.3.0-0(the 0.2.2 prerelease line and later 0.2.x releases, before 0.3), with plugin installation support. - For GenieX: GenieX installed and running locally, with the models already downloaded.
- For QwenCloud: a working Alibaba Cloud Model Studio compatible endpoint and an API key.
Install
- Run
dsh plugin --profile <profile> add github:Fuxfantx/dsh-qwen-modein DSH, or install the offline package or plugin directory through DSH plugin management. - Refresh the DSH page.
- Open Settings → Qwen Mode.
Usage
- Choose Off or Auto. Only Auto maintains the model lists.
- Choose a source and enter its address:
- GenieX: enter the local address,
http://127.0.0.1:18181by default. - QwenCloud: enter the full compatible endpoint, for example
https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1.
- GenieX: enter the local address,
- With QwenCloud, enter the environment variable that stores the API key (not the key itself). GenieX needs no key.
- Adjust context, output limit, thinking, sampling, warmup, MTP, and inference backend as needed.
- Click Save and apply.
- Select a model under From GenieX or From QwenCloud in the model picker.
Settings and behavior
- Context window: empty means not specified; Prefer model context limit uses the model's own context limit.
- Max output: defaults to 16384 and never exceeds the context window or the model's own limit.
- Compaction hints: the settings page shows DSH's compaction threshold, a recommended compaction gate, and whether the current context is sufficient.
- Sampling parameters: only fields missing from the request are filled; empty means not injected. Some parameters apply to GenieX only; Read recommended fills the model's values; a random seed can be fixed.
- Warmup: when enabled, start typing to preload the selected From GenieX model. Dots: yellow loading, green loaded, red not loaded, blue another source.
- MTP: for supported local From GenieX models only. While enabled, sending an image asks whether to send directly or disable MTP for that send.
- Inference backend: empty Compute and Vision Encoder values are decided by GenieX; Vision Encoder applies to image models only.
- Reply Only: selecting it in the permission picker hides and denies tool calls in the current session; choose another permission to restore. The preset is removed when the plugin is disabled or uninstalled, and DSH's own permission presets are never rewritten.
Limitations
- Warmup reduces model load time, not time spent processing a long prompt.
- GenieX keeps one model loaded at a time; using different models in different sessions unloads the previous one.
- MTP and images cannot be active together; the send flow asks first.
- The model lists come from the model source. When the source is unreachable no models appear, and the plugin retries in the background.
- Warmup, MTP, and inference backends are From GenieX-only; QwenCloud keeps model management, thinking, sampling, usage and per-turn statistics, and Reply Only.
- Per-turn statistics and GenieX usage cover Qwen-mode replies only.
Uninstall
Disable or uninstall the plugin from the Plugins page. The Reply Only permission preset is removed with the plugin.
License / 许可证
MIT-0. See LICENSE.
Comments
Loading…
Similar plugins
automatically set reasoning efforts for self-defined models in Deepseek Harness
★ 0
MIT
JavaScript
dsh plugin --profile web add dsh-reasoning-adapterby PaysNatal
Connects all seven Qwen Token Plan Harness-entitlement MCP services (enhanced search, web fetch, code interpreter, vertical search, image, TTS, ASR) plus the decision-model-preview model (qwen_decide)
★ 0
MIT
JavaScript
Nov 12, 2026
dsh plugin --profile web add dsh-qwen-mcpPer-model reasoning effort declarations for GPT, Gemini and OpenAI-compatible providers in DeepSeek Harness
★ 0
dsh plugin --profile web add dsh-more-models-thinking-levelby AkotaP
DSH web plugin: a Model Capabilities page in Settings — manage per-model Context Window, Max Output, Input Modalities and Reasoning Levels (Default = drop the field), plus global preset lists.
★ 0
MIT
JavaScript
Nov 10, 2026
dsh plugin --profile web add dsh-model-tweaksby vit-cerny
Import the models, providers, API keys and OAuth logins that other agent harnesses keep on your machine (opencode, Codex/ChatGPT, Claude Code, Gemini, Qwen, .env) into DeepSeek Harness, with a clean S
★ 0
MIT
JavaScript
Oct 28, 2026
dsh plugin --profile web add dsh-import-modelsby ZhouXuan-Yu
Multi-source model rotation plugin for DeepSeek Harness — group models by source, order them (manual/auto), preset model + reasoning effort, auto-failover when quota runs out, and see which model actually answered.
★ 0
MIT
JavaScript
Oct 31, 2026
dsh plugin --profile web add dsh-model-rotation