[INFO] Hybrid work with RTX

[Smiggy]

Hi Selur,

I had a question.

Will Hybrid work on a RTX Pro 6000 (Blackwell)? Particularly, the upscale model RealESRGAN realsr-anime model and the SCUNet tool?

Thank you.

[Selur]

It should.

[Smiggy]

Hi,

Thank you for your answer. Given the power of the GPU.

Will the encoding be faster compared to a 4090 or 5090?

[Selur]

Not owning a 5090, I can only say that from the specs it should.
It has more VRAM (32 vs 24), more CUDA Cores (21760 vs 16384), newer Tensor Cores.

[Smiggy]

Hello,

We tried and got this error message. Please assist.

2025-07-22 14:52:58.031
C:\Users\user\Downloads\Hybrid\64bit\Vapoursynth\Lib\site-packages\torch\cuda\__init__.py:287: UserWarning:
NVIDIA RTX PRO 6000 Blackwell Workstation Edition with CUDA capability sm_120 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_61 sm_70 sm_75 sm_80 sm_86 sm_90.
If you want to use the NVIDIA RTX PRO 6000 Blackwell Workstation Edition GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

warnings.warn(

C:\Users\user\Downloads\Hybrid\64bit\Vapoursynth\Lib\site-packages\torch\cuda\__init__.py:287: UserWarning:
NVIDIA RTX PRO 6000 Blackwell Workstation Edition with CUDA capability sm_120 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_61 sm_70 sm_75 sm_80 sm_86 sm_90.
If you want to use the NVIDIA RTX PRO 6000 Blackwell Workstation Edition GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

warnings.warn(

2025-07-22 14:52:58.056
Failed to evaluate the script:
Python exception: CUDA error: invalid device ordinal
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.


Traceback (most recent call last):
File "src/cython/vapoursynth.pyx", line 3378, in vapoursynth._vpy_evaluate
File "src/cython/vapoursynth.pyx", line 3379, in vapoursynth._vpy_evaluate
File "C:\Users\user\AppData\Local\Temp\tempPreviewVapoursynthFile14_52_44_742.vpy", line 38, in
clip = SCUNet(clip=clip, model=3, device_index=1)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "contextlib.py", line 81, in inner
File "C:\Users\user\Downloads\Hybrid\64bit\Vapoursynth\Lib\site-packages\torch\utils\_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\user\Downloads\Hybrid\64bit\Vapoursynth\Lib\site-packages\vsscunet\__init__.py", line 134, in scunet
inf_streams = [torch.cuda.Stream(device) for _ in range(num_streams)]
^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\user\Downloads\Hybrid\64bit\Vapoursynth\Lib\site-packages\torch\cuda\streams.py", line 39, in new
with torch.cuda.device(device):
^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\user\Downloads\Hybrid\64bit\Vapoursynth\Lib\site-packages\torch\cuda\__init__.py", line 495, in enter
self.prev_idx = torch.cuda._exchange_device(self.idx)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: CUDA error: invalid device ordinal
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

[Selur]

Seems like the pytorch version used does not support the card.
One would probably have to update pytorch which then requires updates of some of the other stuff.

[Smiggy]

Hi,

Where can we update pytorch?

[Selur]

The portable python environment is under Hybrid/64bit/Vapoursynth, but be warned doing this is not easy.

once could try calling:

python -m   pip install -U torch torchvision torch_tensorrt --index-url https://download.pytorch.org/whl/cu126 --extra-index-url https://pypi.nvidia.com

but instead of using cu126, one would use the latest nightly.

But no clue how much this could break, one could end up having to setup a custom torch add-on.
I outlined the basic step here:

Here's how I build Hybrids torch-addon:

# Base setup
* Create an empty Vapoursynth folder.
* Download portable Python (3.12)
https://www.python.org/ftp/python/3.12.9/python-3.12.9-embed-amd64.zip
* Download portable Vapoursynth
https://github.com/vapoursynth/vapoursynth/releases/download/R72/VapourSynth64-Portable-R72.zip
* Extract Python and then Vapoursynth into the Vapoursynth folder
* Change the content of the python312._pth to
  ```
  Scripts
  Lib\site-packages
  python312.zip
  .
  
  # Uncomment to run site.main() automatically
  #import site
  ```
* Install pip
* Download the pip installer
   https://bootstrap.pypa.io/get-pip.py and move it into your Vapoursynth folder
* Open a terminal inside the Vapoursynth and call  'python get-pip.py'

* Integrate Vapoursynth to the environment by calling 'python -m pip install wheel/VapourSynth-72-cp312-abi3-win_amd64.whl'

* Install torch
  ```
  python -m pip install -U packaging setuptools Wheel
  python -m   pip install -U torch torchvision torch_tensorrt --index-url https://download.pytorch.org/whl/cu126 --extra-index-url https://pypi.nvidia.com

  ```
* Install VSGAN
  ```
  python -m pip install -U vsgan
  ```
* Install vs-rife by calling:
  ```
  python -m pip install -U vsrife
  ```
* Install BasicVSR++
  ```
  python -m pip install -U vsbasicvsrpp
  ```
* Install DPIR
  ```
  python -m pip install -U vsdpir
  ```
* Install SCUNet
  ```
  python -m pip install -U vsscunet
  ```
* Install REALEsrgan
  ```
  python -m pip install -U vsrealesrgan
  ```
* Install HINet
  ```
  python -m pip install -U vshinet
  ```
* Install AnimeSR
  ```
  python -m pip install -U vsanimesr
  ```
* Install FeMaSR
  ```
  python -m pip install -U vsfemasr
  ```
* Install CodeFormer
  ```
  python -m pip install -U vscodeformer
  python -m pip install https://github.com/eddiehe99/dlib-whl/releases/download/v20.0.0/dlib-20.0.0-cp312-cp312-win_amd64.whl
  ```
* Install GRLIR
  ```
  python -m pip install -U vsgrlir
  ```  
* Install MFDIN
  ```
  python -m pip install -U vsmfdin
  ```  
* Install ProPainter
  ```
  python -m pip install https://github.com/dan64/vs-propainter/releases/download/v1.2.1/vspropainter-1.2.1-py3-none-any.whl
  ```  
* Install VSDDColor
  ```
  python -m pip install https://github.com/dan64/vs-deoldify/releases/download/v4.0.0/vsddcolor-1.0.1-py3-none-any.whl
  ```
* Install VSDDColor
  ```
  python -m pip install -U vsswinir
  ```
* Install DeOdify
  ```
  python -m pip install -U scikit-image
  python -m pip install -U numba
  python -m pip install https://github.com/dan64/vs-deoldify/releases/download/v5.0.4/vsdeoldify-5.0.4-py3-none-any.whl
  ```
  * Download:
    * https://colorizers.s3.us-east-2.amazonaws.com/colorization_release_v2-9b330a0b.pth
    * https://colorizers.s3.us-east-2.amazonaws.com/siggraph17-df00044c.pth
    * https://data.deepai.org/deoldify/ColorizeArtistic_gen.pth
    * https://www.dropbox.com/s/axsd2g85uyixaho/ColorizeStable_gen.pth?dl=0
    * https://data.deepai.org/deoldify/ColorizeVideo_gen.pth
    and place them in the Vapoursynth/vsdeoldify/models-folder
    * https://github.com/zhangmozhe/Deep-Exemplar-based-Video-Colorization/releases/download/v1.0/colorization_checkpoint.zip
      and extract it into the Vapoursynth\Lib\site-packages\vsdeoldify\deepex
    * https://download.pytorch.org/models/resnet101-63fe2227.pth
    * https://download.pytorch.org/models/resnet50-19c8e357.pth
    * https://dl.fbaipublicfiles.com/dinov2/dinov2_vits14/dinov2_vits14_pretrain.pth
    and place them in the Vapoursynth\Lib\site-packages\vsdeoldify\deepex\checkpoints-folder
    * https://github.com/yyang181/colormnet/releases/download/v0.1/DINOv2FeatureV6_LocalAtten_s2_154000.pth
    and place it into the Vapoursynth/Lib/site-packages/vsdeoldify/colormnet/weights-folder
  * install spartial_correlation_sampler (needs new build, see: https://forum.selur.net/thread-3595-post-26036.html#pid26036)
    extract content of spatial_correlation_sampler (open as an archive) into Vapoursynth\Lib\site-packages

* Download Models for the installed tools
  ```
  python -m vsrife
  python -m vsdpir
  python -m vsscunet
  python -m vsrealesrgan
  python -m vshinet
  python -m vsfemasr
  python -m vscodeformer
  python -m vsgrlir
  python -m vsddcolor
  python -m vsswinir
  python -m vsbasicvsrpp
  ```
* Install DeepDeinterlace  (**fails due to openmim dependency**)
  Download https://github.com/pifroggi/vs_deepdeinterlace/archive/refs/heads/main.zip and place the vs_deepdeinterlace-folder (under vs_deepdeinterlace-main)  inside your Vapoursynth/Lib/site-packages folder
  ```
  python -m pip install positional_encodings
  ```
  not supporting DfConvEkSA atm. due to openmin issues.

I'll try finding some time, to create such a torch add-on in the next few days.
(best remind me if I didn't get around to this on Friday)

Cu Selur

[tailland]

 premier support, selur.

[Selur]

Won't probably get around to upload a new test torch, but I'm uploading a VapoursynthR72_torch_2025.06.06_torch2.7dev which I didn't build new, but which was a test setup I used in June.
Should be up in ~1hour, please test and let me know whether that one works.

Cu Selur

Ps.: when I remember correctly the problem with torch dev2.7 was that it caused problems with HAVC due to excessive log output,...