Closed przemoc closed 1 month ago
Since the height is above 1080 ( 1280p in your case ) it has defaulted to FP32 in order to avoid certain issues with Pixelation seen in resolutions above 1920x1080. I am not sure if it applies to 1280x1280 but TAS was hardcoded to work that way. ( If width or height are > 1920 or 1080, switch to UHD Mode / FP32, 1280 > 1080 therefore UDH mode on ).
FP32 uses roughly 2x more VRAM than FP16 therefore it has ended up VRAM OOM-ing, I am not sure why your nvidia-smi
didn't show it but it's clearly a VRAM issue coupled with some concat op that isn't supported apparently?.
Sadly I can not provide a fix for now other than either lowering the base resolution to 1080x1080 and / or using CUDA.
Thank you for the details. Next time I will try with lower resolution. CUDA and NCNN variants were working fine from my short testing (NCNN was slower, though).
My concat.mp4
file was the output of ffmpeg concat, basically combining multiple images into video.
#!/bin/sh
ls -1rt *.png | sed "s,^,file ',;s,$,'," >ffmpeg-concat.txt
ls -1rt *.png | sed "s,^,file ',;s,$,',;1q" >>ffmpeg-concat.txt
ffmpeg -f concat -safe 0 -i ffmpeg-concat.txt -c:v libx264 -crf 0 -pix_fmt yuv444p -vf "settb=AVTB,setpts=N/3/TB,fps=3" -movflags +faststart concat.mp4
I was using concat, as in the past I was also playing with ffmpeg minterpolate as a secondary step.
I was referring to this part of the logs: "ONNXTRT_Broadcast_470.../Concat_35"
If I remember correctly from my own testing a rule of thumb is.
Tensorrt > Cuda >= DirectML ( if it works ) > NCNN.
So your results are to be expected.
I haven't played with TensorRT so far at all, so maybe I'm doing something wrong. TAS v1.9.0. I tried both RIFE4.17 and RIFE4.18.
I checked in
nvidia-smi
and my RTX 2060 w/ 6 GB VRAM should have 5069 MB free, as OS + rest of apps occupy around 964 MB, which gives 6144 - 964 = 5180 MB.