OpenVINO + FP16 + Waifu2xModel.cunet == ERROR

[WhitePr]

Arch linux
kernel: linux 6.18.2-arch2-1
CPU ryzen 7900x
intel arc b580
openvino 2025.4.0-2 from AUR

This is vapoursynth script

import vapoursynth as vs
from vapoursynth import core
import mvsfunc as mvf
import vsmlrt

vsmlrt.Waifu2x(clip = core.std.BlankClip(format=vs.RGBS),
               scale = 1, 
               noise = 3, 
               model = vsmlrt.Waifu2xModel.cunet, 
               backend = vsmlrt.Backend.OV_GPU(fp16=1, num_streams=2)).set_output()

error log from Terminal:

/usr/include/c++/15.2.1/bits/stl_vector.h:1282: std::vector<_Tp, _Alloc>::const_reference std::vector<_Tp, _Alloc>::operator[](size_type) const [with _Tp = kernel_selector::Tensor::DataTensor; _Alloc = std::allocator<kernel_selector::Tensor::DataTensor>; const_reference = const kernel_selector::Tensor::DataTensor&; size_type = long unsigned int]: Assertion '__n < this->size()' failed.

I tested other models (except swin) without problems. Also tried different noise and scale parameters, but the problem remained. Once I set OPENVINO to FP32, the problem disappeared. Checking the WIKI documentation, I found that the A770 is able to run the CUNET model in FP16 mode, but reports errors on my B580.

[WolframRhodium]

It could be caused by running out of GPU memory. A770 has 4GB more memory than B580.

[WhitePr]

It could be caused by running out of GPU memory. A770 has 4GB more memory than B580.

Only 3GB of VRAM was utilised in FP32 mode; switching to FP16 should require even less VRAM. Therefore, I don't believe the issue lies with insufficient graphics memory.
(1080x1920 used 3G VRAM)