Quantized ONNX models on IREE with Vulkan backend

I am interested in quantizing a YOLO-NAS model to run on an embedded GPU with a vulkan backend. To achieve better performance, I've performed Post-Training Quantization on the model, using a small calibration dataset.

I am starting with an ONNX format, which I import with _iree-import-onnx_ into a *.mlir format.
I notice that, on the MLIR file, most of the operands have been kept on a float 32 bits data type (Possibly ignored on quantization pass).

Then, I use the following to compile:
` iree-compile YOLO_NAS.mlir --iree-hal-target-device=vulkan --iree-vulkan-target=architecture_name --output-format=vm-bytecode -o YOLO_NAS.vmfb`

I saw there are some flags that can be used for compiling considering quantization, such as:
--iree-global-opt-enable-quantized-matmul-reassociation
--iree-input-demote-f32-to-f16
--iree-input-demote-i64-to-i32
--Iree-global-opt-enable-demote-contraction-inputs-to-bf16=all

However, even with those flags, the model still performs poorly on the GPU, taking too long to realize inferences.
Is there a preferred way on IREE to deal with quantized models that should increase performance? Especially for Vulkan backends.

Thanks for the support.