Does Hailo8L on RPI5 support Vulkan standard, such that libraries like ncnn can leverage higher inference speed automagically?
Short answer no.
To run a CNN on the Hailo-8 it needs to be converted in multiple steps into the HEF (Hailo Executable Format) using the Hailo Dataflow Compiler.
The quantization step is specific for each trained model even if the model architecture is the same because the weights are different. Depending the optimization algorithm this requires the model to be run so the algorithm can select the correct quantization for each layer.
There are other steps like the allocation that also require quite a lot of compute and time.
So, the overall process cannot be done on the fly.