Feat/roboflow workflow dsp

[klemen1999]

Purpose

Summary

• Migrate the Roboflow Workflow standalone app to the released OAK4 QNN runtime stack:

  • Use luxonis/oakapp-base:1.4.0-onnxruntime.
  • Pin depthai==3.9.0, depthai-nodes==0.6.1, and inference==1.3.7.
  • Run the app on the base image’s Python 3.12 runtime.
  • Pass through the Hexagon DSP/HTP devices and NPU runtime mount.

• Route Roboflow-created ONNX Runtime sessions through depthai_nodes.runtime.onnx_qnn_session:

  • Fix dynamic batch dimensions to static batch size 1 before QNN compilation.
  • Run QNN on HTP with fp16 and reuse cached compiled graphs.
  • Fall back to CPU by default; expose EP=cpu and STRICT=1 for debugging.

• Replace the custom OpenCV generator-capture/probe path with Roboflow’s VideoFrameProducer integration backed by a DepthAI output queue.

• Discover workflow outputs and parameter definitions from the Roboflow API rather than performing dummy inference.

• Support runtime workflow controls:

  • Switch API key, workspace, workflow ID, and workflow parameters.
  • Rebuild DepthAI visualizer topics when output schemas change.
  • Refresh the active workflow definition from Roboflow without redeploying.
  • Surface inference pipeline health/errors in the frontend.

• Prevent workflow switching from aborting the app:

  • Quiesce AnnotationNode before DepthAI pipeline teardown.
  • Handle closed output queues inside the host-node callback.

• Update the custom frontend:

  • Render controls from the discovered workflow interface.
  • Add error/retry and refresh UI states.
  • Align Luxonis frontend dependencies and enforce one React/ReactDOM version.
  • Limit Rollup parallel file operations and raise the frontend build file-descriptor limit to avoid EMFILE.

• Simplify and update documentation:

  • Update README/agent guidance for released QNN support, runtime workflow controls, and visualization rules.

Specification

None / not applicable

Dependencies & Potential Impact

None / not applicable

Deployment Plan

None / not applicable

Testing & Validation

None / not applicable

AI Usage

Assisted-by: AGENT_NAME:MODEL_VERSION [TOOL1] [TOOL2]

Submitted code was reviewed by a human: YES/NO

The author is taking the responsibility for the contribution: YES/NO