2,101 Reply by Drakko01

  • Thanks: 49

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

How can I test R70 vapoursynth instead of R64 provided for svp, and you have  plan to updated ? I dunno if there a real benefit but a think so many things fix/changed on it , that a like to try it. So if someone get me pointers , it will be very appreciated.

2,102 Reply by Chainik

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

> I dunno if there a real benefit

nope

> How can I test R70 vapoursynth

put portable release into SVP 4/mpv64

2,103 Reply by Drakko01

  • Thanks: 49

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

Chainik wrote:

> I dunno if there a real benefit

nope

> How can I test R70 vapoursynth

put portable release into SVP 4/mpv64

Thanks for respond, which file do a put into mpv64 folder, if copy the entire folder i get unable to initialize vapoursynth api 4.0 error

2,104 Reply by Chainik

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

all files
dunno which Python version it wants, probably you have to replace portable Python too

2,105 Reply by Drakko01

  • Thanks: 49

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

Chainik wrote:

all files
dunno which Python version it wants, probably you have to replace portable Python too

Thanks, I think better leave like it is, wait until you decided update on the future.

2,106 Reply by Drakko01

  • Thanks: 49
  • Thanks for the post: 2

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

For those interesting Model 4.26_heavy its out , for me at least don't work as well as 4.25, maybe for those with 4090 will.

  • Thanks: 99
  • Thanks for the post: 1

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

Drakko01 wrote:

4.26_heavy its out

These files should allow SVP to find files ending with "_large.onnx".

Post's attachments

helpers.py 855 b, 24 downloads since 2024-11-18 

vsmlrt.py 102.46 kb, 27 downloads since 2024-11-18 

2,108 Reply by narkohol (edited by narkohol 19-11-2024 11:42:54)

  • Thanks: 8

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

Does changing these as suggested in vsmlrt.py make sense for RTX 40XX gpus?

https://i.postimg.cc/505Pf7zr/image.png

If yes...    which numbers are recommended for 'Streams' and 'Workspace'?

any other recommendations for the other parameters?

2,109 Reply by dawkinscm (edited by dawkinscm 19-11-2024 12:38:35)

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

narkohol wrote:

Does changing these as suggested in vsmlrt.py make sense for RTX 40XX gpus?

https://i.postimg.cc/505Pf7zr/image.png

If yes...    which numbers are recommended for 'Streams' and 'Workspace'?

any other recommendations for the other parameters?

What changes?

2,110 Reply by narkohol (edited by narkohol 19-11-2024 18:50:17)

  • Thanks: 8

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

dawkinscm wrote:
narkohol wrote:

Does changing these as suggested in vsmlrt.py make sense for RTX 40XX gpus?

https://i.postimg.cc/505Pf7zr/image.png

If yes...    which numbers are recommended for 'Streams' and 'Workspace'?

any other recommendations for the other parameters?

What changes?

Those specified in that header as 'basic performance tuning':

        set fp16 = True (on RTX GPUs)
        increase num_streams
        increase workspace
        set use_cuda_graph = True
       

That by default are:

        fp16: bool = False
        num_streams: int = 1
        workspace: typing.Optional[int] = None
        use_cuda_graph: bool = False

2,111 Reply by Chainik

  • Thanks: 1622
  • Thanks for the post: 1

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

num_streams comes from the RIFE profile
force_fp16 = True, tf32 = True, use_cuda_graph = True
come from helpers.py

2,112 Reply by narkohol

  • Thanks: 8

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

Chainik wrote:

num_streams comes from the RIFE profile
force_fp16 = True, tf32 = True, use_cuda_graph = True
come from helpers.py

What about 'Workspace'?

vsmlrt.py 'performance tuning' header says "increase workspace", but in helpers.py it is set to backend.workspace = None

https://i.postimg.cc/Dy3PQKX9/image.png

2,113 Reply by aloola

  • Thanks: 86
  • Thanks for the post: 2

Re: New RIFE filter - 3x faster AI interpolation possible in SVP!!!

narkohol wrote:
Chainik wrote:

num_streams comes from the RIFE profile
force_fp16 = True, tf32 = True, use_cuda_graph = True
come from helpers.py

What about 'Workspace'?

vsmlrt.py 'performance tuning' header says "increase workspace", but in helpers.py it is set to backend.workspace = None

https://i.postimg.cc/Dy3PQKX9/image.png

"The maximum workspace is set to None for the total memory size of the GPU."

none is the best