dsh-vision-3090-fix: A Local Proxy Plugin to Limit the Number of Images per Request

Introduction¶

The DeepSeek Harness (DSH) plugin ecosystem allows developers to extend functionality through plugins. When configuring a local vLLM model (for example, Qwen 3.8 27B) as a vision model backend, a common issue arises: vLLM’s default parameter --limit-mm-per-prompt image=1 limits each request to include only one image. However, the LLM adapter built into DSH restricts images only by byte size, not by count. As a result, when a conversation contains multiple images across turns, or when a tool call returns multiple images, the backend directly returns a 400 error.

The dsh-vision-3090-fix plugin is designed specifically to solve this problem. It runs as a local HTTP reverse proxy and, before forwarding a request to the backend, replaces extra images with text placeholders, thereby bypassing the backend’s count limit.

Plugin Overview¶

This is a DeepSeek Harness plugin that sets the maximum number of images per single request to 1 for a self-hosted vLLM backend. It does not modify backend logic; instead, it intercepts requests and rewrites their content at the proxy layer.

Core Features¶

• Limit image count: Forces the number of images in each request to be limited to a specified maximum (default is 1).
• Local reverse proxy: Launches as a standard Cordis plugin and runs inside the DSH process.
• Placeholder replacement: Replaces older images beyond the limit with stable text placeholders.
• Supports streaming responses: Preserves the streaming characteristics of backend responses and forwards them byte by byte.
• Cross-message processing: Supports handling image_url content parts in the messages array and in tool calls.

Installation and Enabling¶

Use the following command to install the plugin into the specified DSH configuration profile:

dsh plugin --profile web add /path/to/dsh-vision-3090-fix

After installation, you need to configure the basic proxy parameters in cordis.patch.yml and change the provider’s baseURL to point to this proxy.

1. Configure upstreamOrigin, listenHost, listenPort, maxImagesPerRequest, and models in cordis.patch.yml:

    - id: vision-3090-fix
      name: dsh-vision-3090-fix
      config:
        upstreamOrigin: http://10.204.100.243:1234   # 真实后端地址，不含路径
        listenHost: 127.0.0.1
        listenPort: 8931
        maxImagesPerRequest: 1                        # 匹配你的服务器 --limit-mm-per-prompt 参数
        models: [qwen3.8-27b]                         # 可选，不填则对所有模型生效

2. Modify settings.yaml to point the provider’s baseURL to the proxy address:

    llm-pi-ai:
      providers:
        pure:
          displayName: pure
          apiKeyEnv: PURE_API_KEY
          api: openai-completions
          baseURL: http://127.0.0.1:8931/v1   # 指向本地代理
          models:
            - id: qwen3.8-27b

Use Cases and Notes¶

• Use cases: Self-hosted vision model environments where vLLM is deployed and --limit-mm-per-prompt image=1 is set.
• Notes:
  • The plugin fully buffers the request body before forwarding (to parse and rewrite the JSON), so it is not suitable for handling large file uploads.
  • There is no retry logic or request queueing mechanism; it is a lightweight pass-through proxy.
  • It assumes the backend uses OpenAI’s image_url line format; if the backend format is different, the plugin will not recognize it.
  • The plugin will not crash the proxy due to timeouts; timeouts are logged and the current request is terminated.
  • Check the source code and license (MIT) before installation.

Summary¶

dsh-vision-3090-fix provides a low-cost way to resolve the image-count conflict between DSH and restricted backends without modifying the vLLM server configuration.

• Plugin Directory
• GitHub Repository