v4.0.0

[HolyWu]

• Properly handle frame duration.
• Change default num_streams to 2.
• Put model files in a separate directory.
• Set CUDA_MODULE_LOADING environment variable to LAZY.
• Always clamp input tensor to 0.0-1.0 range.
• Remove nvFuser and CUDA Graphs.
• Change function name to lowercase.
• Add 4.7 model.
• Bump TensorRT version requirement to 8.6.1.
• Bump PyTorch version requirement to 2.0.1.
• Bump VapourSynth version requirement to R62.

---

This discussion was created from the release v4.0.0.

[Samhayne]

VERY probably it’s just a stupid coincidence, and I don’t want to make anyone feel unsafe, but I still want to drop a note in case someone else had the same problem and it’s NOT just a coincidence:

One day after updating to vs-rife 4.0 something seems to have killed the VRAM of my one week old and so far well working 4090 - all channels are reported as faulty by Nvidia’s diagnostics tool “mats”.
In the mid of switching between different rife models for several minutes I suddenly got artifacts on the screen - and they came to stay.

Again: This is NOT a warning of vs-rife 4.0. The card was still pretty new and I rather guess it’s a manufacturing/material problem. I myself plan to continue using vs-rife as soon as I get a replacement card.
Still: If someone else SHOULD run into the same issue, take this as a motivation to also drop a note. :)

edit after three weeks: I tortured my old GTX 1070 for more than two weeks now with vsrife3.0 - no issues so far.