RIFE + MIGX backend causes segmentation fault on Arch Linux (AMD RX 7700 XT)

[AlmeasqViolet]

[AlmeasqViolet@ArchLinux ~]$ mpv '/run/media/AlmeasqViolet/Anime/Anime/[Snow-Raws] 悪魔のリドル riddle story of devil/映像特典/[Snow-Raws] 悪魔のリドル riddle story of devil OP01 (BD 1920x1080 HEVC-YUV420P10 FLAC).mkv' [playlistmanager] auto showamount: 19 ● Video --vid=1 --vlang=jpn (hevc 1920x1080 23.976 fps) [default] ● Audio --aid=1 --alang=jpn (flac 2ch 48000 Hz) [default] [2025-12-11 16:45:04.214] [MANGOHUD] [info] [gpu.cpp:89] Set renderD128 as active GPU (driver=amdgpu id=1002:747e pci_dev=0000:0d:00.0) Using hardware decoding (vaapi). [autoconvert] HW-downloading from vaapi [autoconvert] Converting p010 -> yuv420p10 Running [ MIGraphX Version: 2.14.0.20250912-17-113-g89196378a ]: /usr/lib/vapoursynth/vsmlrt-hip/migraphx-driver compile --onnx /usr/lib/vapoursynth/models/rife/rife_v4.6.onnx --gpu --optimize --binary --output /usr/lib/vapoursynth/models/rife/e46d21ca.mxr --input-dim @input 1 11 1024 1920 --fp16 [2025-12-11 16:45:04] Reading: /usr/lib/vapoursynth/models/rife/rife_v4.6.onnx Quantizing to fp16 ... Compiling ... [2025-12-11 16:47:49] [ MIGraphX Version: 2.14.0.20250912-17-113-g89196378a ] Complete(164.524s): /usr/lib/vapoursynth/vsmlrt-hip/migraphx-driver compile --onnx /usr/lib/vapoursynth/models/rife/rife_v4.6.onnx --gpu --optimize --binary --output /usr/lib/vapoursynth/models/rife/e46d21ca.mxr --input-dim @input 1 11 1024 1920 --fp16 Segmentation fault (コアダンプ) mpv '/run/media/AlmeasqViolet/Anime/Anime/[Snow-Raws] 悪魔のリドル riddle story of devil/映像特典/[Snow-Raws] 悪魔のリドル riddle story of devil OP01 (BD 1920x1080 HEVC-YUV420P10 FLAC).mkv'
`import vapoursynth as vs
from vapoursynth import core
import vsmlrt

clip = video_in

#向上取整：19201080>===>19201088
#w, h = clip.width, clip.height
#w = ((w + 63) // 64) * 64
#h = ((h + 63) // 64) * 64
#向上取整：19201080>===>19201088
#w, h = clip.width, clip.height
#w = ((w + 127) // 128) * 128
#h = ((h + 127) // 128) * 128
#向下取整：19201080>===>19201024
w, h = clip.width, clip.height
w = (w // 64) * 64
h = (h // 64) * 64
#向下取整：19201080>===>19201024
#w, h = clip.width, clip.height
#w = (w // 64) * 64
#h = (h // 64) * 64

clip = core.resize.Bicubic(
clip,
width=w,
height=h,
format=vs.RGBS,
matrix_in_s="709"
)

clip = vsmlrt.RIFE(
clip,
model=46,
multi=2,
backend=vsmlrt.Backend.MIGX(
device_id=0,
fp16=True,
num_streams=2,
opt_shapes=None,
fast_math=True,
exhaustive_tune=False,
short_path=True,
supports_onnx_serialization=False,
)
)

clip = core.resize.Bicubic(
clip,
format=vs.YUV420P10,
matrix_s="709"
)

clip.set_output()`

My configuration:
archlinux
linux-zen
R9 3900
RX 7700XT
DDR4 3200 44G
16G+500G+1T SSD
2T HDD

Note:
The 16GB one is an Intel M20, which I use as disk cache. My animation files are stored on the 2TB HDD.
I have tried adjusting many MIGX parameters, including image cropping, output formats, and others, but the issue remains unresolved. I don't have a deep technical background, so if anyone is willing, I would appreciate some help with this.

[WolframRhodium]

This is a known limitation of migx, I have never been able to run rife on it.

[AlmeasqViolet]

This is a known limitation of migx, I have never been able to run rife on it.

Thank you, I'd like to ask, will MIGX be able to run RIFE successfully in the future?

[WolframRhodium]

Eventually maybe, by migx's developers. Practically, I am not convinced that it will happen soon. In the meantime, you could use VapourSynth-RIFE-ncnn-Vulkan.

[AlmeasqViolet]

Eventually maybe, by migx's developers. Practically, I am not convinced that it will happen soon. In the meantime, you could use VapourSynth-RIFE-ncnn-Vulkan.

vapoursynth-plugin-mlrt-ort-runtime-git's ROCm version won't work either? I'm planning to try this.

[WolframRhodium]

It won't. The ort interface in vs-mlrt does not support the ROCM provider of onnxruntime, and the provider has also been removed in upstream.

[AlmeasqViolet]

It won't. The ort interface in vs-mlrt does not support the ROCM provider of onnxruntime, and the provider has also been removed in upstream.

Okay, thank you.

[AlmeasqViolet]

It won't. The ort interface in vs-mlrt does not support the ROCM provider of onnxruntime, and the provider has also been removed in upstream.

I tried switching to NCNN_VK, but encountered "GridSample not supported yet!" How should I resolve this?

[AlmeasqViolet@ArchLinux ~]$ mpv '/run/media/AlmeasqViolet/Anime/Anime/[Snow-Raws] 悪魔のリドル riddle story of devil/映像特典/[Snow-Raws] 悪魔のリドル riddle story of devil OP01 (BD 1920x1080 HEVC-YUV420P10 FLAC).mkv'
[playlistmanager] auto showamount: 19
 ● Video  --vid=1  --vlang=jpn  (hevc 1920x1080 23.976 fps) [default]
 ● Audio  --aid=1  --alang=jpn  (flac 2ch 48000 Hz) [default]
[2025-12-11 18:14:51.175] [MANGOHUD] [info] [gpu.cpp:89] Set renderD128 as active GPU (driver=amdgpu id=1002:747e pci_dev=0000:0d:00.0)
Using hardware decoding (vaapi).
[autoconvert] HW-downloading from vaapi
[autoconvert] Converting p010 -> yuv420p10
GridSample not supported yet!
  # align_corners=1
  # mode=bilinear
  # padding_mode=border
GridSample not supported yet!
  # align_corners=1
  # mode=bilinear
  # padding_mode=border
GridSample not supported yet!
  # align_corners=1
  # mode=bilinear
  # padding_mode=border
GridSample not supported yet!
  # align_corners=1
  # mode=bilinear
  # padding_mode=border
Unsupported split axis !
[0 AMD Radeon RX 7700 XT (RADV NAVI32)]  queueC=1[4]  queueG=0[1]  queueT=0[1]
[0 AMD Radeon RX 7700 XT (RADV NAVI32)]  bugsbn1=0  bugbilz=0  bugcopc=0  bugihfa=0
[0 AMD Radeon RX 7700 XT (RADV NAVI32)]  fp16-p/s/u/a=1/1/1/1  int8-p/s/u/a=1/1/1/1
[0 AMD Radeon RX 7700 XT (RADV NAVI32)]  subgroup=64(32~64)  ops=1/1/1/1/1/1/1/1/1/1
[0 AMD Radeon RX 7700 XT (RADV NAVI32)]  fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0
AO: [pipewire] 48000Hz stereo 2ch s32
Segmentation fault         (コアダンプ) mpv '/run/media/AlmeasqViolet/Anime/Anime/[Snow-Raws] 悪魔のリドル riddle story of devil/映像特典/[Snow-Raws] 悪魔のリドル riddle story of devil OP01 (BD 1920x1080 HEVC-YUV420P10 FLAC).mkv'
[AlmeasqViolet@ArchLinux ~]$ 

import vapoursynth as vs
from vapoursynth import core
import vsmlrt

clip = video_in

#向上取整：1920*1080>===>1920*1088
#w, h = clip.width, clip.height
#w = ((w + 63) // 64) * 64
#h = ((h + 63) // 64) * 64
#向上取整：1920*1080>===>1920*1088
#w, h = clip.width, clip.height
#w = ((w + 127) // 128) * 128
#h = ((h + 127) // 128) * 128
#向下取整：1920*1080>===>1920*1024
w, h = clip.width, clip.height
w = (w // 64) * 64
h = (h // 64) * 64
#向下取整：1920*1080>===>1920*1024
#w, h = clip.width, clip.height
#w = (w // 64) * 64
#h = (h // 64) * 64

clip = core.resize.Bicubic(
    clip, 
    width=w, 
    height=h, 
    format=vs.RGBS,
    matrix_in_s="709"
)

clip = vsmlrt.RIFE(
    clip,
    model=46,
    multi=2,
    backend=vsmlrt.Backend.NCNN_VK(
        device_id=0,
        fp16=True,
        num_streams=2,
    )
)

clip = core.resize.Bicubic(
    clip, 
    format=vs.YUV420P10,
    matrix_s="709"
)

clip.set_output()

[WolframRhodium]

NCNN_VK is different from VapourSynth-RIFE-ncnn-Vulkan. The former does not support GridSample natively.