Your documentation on custom inference using voxl-tflite-server is quite outdated. Yolo Ultralytics has stopped exporting frozen graph model format to tflite fp16 quantization. Ultralytics says there's no quantize=16 option — an FP32 .tflite model automatically runs in FP16 at runtime when using a GPU delegate (WebGPU/OpenCL/Metal), so a separate FP16 export file isn't needed. That's why half=True/quantize=16 throws the assertion.

Can voxl-tflite-server automatically execute FP32 models internally in FP16?
Hardware: VOXL2
Source code: Latest voxl-tflite-server code on modalai gitlab page.