VRAM choking when in VRChat

[Kaylee-West]

Hi all!

Firstly let me say that the v.15.3.15 onnxdirectML - cuda real time voice changer is working beautifully on my laptop most of the time. I use it in Second Life (absolutely no issues) and VRChat (occasional issues).

I am having a bit of an issue in some VRChat worlds where I run out of VRAM and VRChat freezes, sometimes temporarily, sometimes requiring a reboot. I do reduce the graphics settings right down and often this solves the problem, but occasionally I have to close the voice changer completely and then re-start it later in less graphically demanding worlds.

I was just wondering if there are any particular settings I can adjust or any other real time voice changer versions that I can use that will help avoid this problem? I actually only use one voice model, but there are 6 (other than the one I use) showing on the UI. Would removing / deleting the other voice models lessen the demand on VRAM?

My computer specs are as follows:

Processor 12th Gen Intel(R) Core(TM) i7-12700H (2.30 GHz)
Installed RAM 16.0 GB (15.7 GB usable)
System type 64-bit operating system, x64-based processor

My GPU specs are as follows:

NVIDIA GeForce RTX 3070 Ti Laptop GPU, NVIDIA compatible

Any guidance on how I might lessen the real time voice changer's burden on VRAM would be warmly welcomed.

Kaylee

[HomidWay]

My first guess would be to use CUDA version and not DirectML one, as DirectML is an additional layer between GPU and computations.

In cuda version basic models would take around 1.2 gig VRAM, since 3070 has 8 gb total you would have to lower your graphics, primarily textures and resolution to fit into your VRAM.

There isn't much you can do to make RVC models take less VRAM, at least I'm not aware of anything except for using smaller models, but even that would be difference of megabytes and basically insignificant

[Kaylee-West]

Hi HomidWay!

Thank you so much for getting back to me. I actually thought I was using the CUDA version, but clearly I was mistaken.

Is the CUDA version you are suggesting this one: https://huggingface.co/wok000/vcclient000/blob/main/MMVCServerSIO_win_onnxgpu-cuda_v.1.5.3.18a.zip

Once again, ty very much for your help.

Kaylee

[Kuuko-fokkusugaru]

Nice to meet another fellow SL user!
You definitely need to switch to the CUDA version and, while on it, use the latest one. You can either use the latest v1, which is the one that you point out, or you can use the latest v2 which has some improvements in CPU usage (2.1.4 at time of writing this I believe). Both of them works fine still.

As you can guess, as soon as a game performance takes over the CPU or GPU, if there is nothing spare for RVC, you may experience stutter or big delays. About which settings you may be able to tweak, the lower the chunk size, the more performance it requires. So you could try to sacrifice a bit of delay in exchange of less GPU usage. I personally use 0.3 seconds but you could also try 0.5 seconds which isn't so different from 0.3 in terms of delay. Also, using bigger chunk also translates into better audio quality so you can try different settings and see if they actually help. Keep in mind that RVC is meant to be used from a different computer connected to the same network since you can access to RVC through any web browser as long as both are in the same network but I understand that having two computers isn't a cheap option. I am just mentioning it because the other "fix" is to get a high end GPU like a 4090 or 5090 and, at that point, it's way cheaper to just get a laptop with a 3060 and the likes solely for RVC. This all depends on how much you actually like the software and how much you need it to enjoy your virtual experiences.

Anyway, try the mentioned settings and let's see if they can work out for you. Also, keep in mind that you can also reduce other settings like limiting the framerate. For example, SL isn't really a game where high fps are needed. And if you have been around long enough, playing at like 15 fps was pretty much the norm even for high end computers lol. So you can always consider lowering the framerate. If you play on high refresh rates monitors, try to limit it to 60 or even 30. And in crowded areas you could set it to 20 or 15. For vrchat is a bit more complicated since the graphics settings are very limited but, if you are used to vr, you may be able to set your headset to 72 hz instead of like 90 or 120 and get some spare room for RVC. You can also resize the desktop window of vrchat and make it small if you aren't like streaming or anything reducing a bit the cpu and gpu usage. And never ever open something like a browser or even browser based apps like discord as those tend to use quite a big chunk of gpu unless you turn off hardware acceleration.

Hope that helps 💛

[wajahatalikhan751-bot]

Hi all!

Firstly let me say that the v.15.3.15 onnxdirectML - cuda real time voice changer is working beautifully on my laptop most of the time. I use it in Second Life (absolutely no issues) and VRChat (occasional issues).

I am having a bit of an issue in some VRChat worlds where I run out of VRAM and VRChat freezes, sometimes temporarily, sometimes requiring a reboot. I do reduce the graphics settings right down and often this solves the problem, but occasionally I have to close the voice changer completely and then re-start it later in less graphically demanding worlds.

I was just wondering if there are any particular settings I can adjust or any other real time voice changer versions that I can use that will help avoid this problem? I actually only use one voice model, but there are 6 (other than the one I use) showing on the UI. Would removing / deleting the other voice models lessen the demand on VRAM?

My computer specs are as follows:

Processor 12th Gen Intel(R) Core(TM) i7-12700H (2.30 GHz) Installed RAM 16.0 GB (15.7 GB usable) System type 64-bit operating system, x64-based processor

My GPU specs are as follows:

NVIDIA GeForce RTX 3070 Ti Laptop GPU, NVIDIA compatible

Any guidance on how I might lessen the real time voice changer's burden on VRAM would be warmly welcomed.

Kaylee

Hi , I also use it in 2nd life , but my voice distorts alot when I use and in have almost exact same specs as you but I have normal rtx 3070 not ti version and I have 32 gigs ram , I want to ask you , how do you make it work in 2nd life

[Kuuko-fokkusugaru]

You should update to the v1.5.3.18a but, instead of using the ML version, use the CUDA one so you use your GPU for RVC and get better performance. That said, demanding games like SL and vrchat may still cause issues. Besides the vram usage, which you can't just use less even if you delete unused models, you may easily hit a roof because those games are very intensive for the GPU and CPU. There is not much that you can do beyond reducing graphics, draw distance and limiting texture size to 1024 or lower for SL. Maybe the CUDA version performs better but I always found issues with vrchat even when I had the 3080Ti with 12GB VRAM.

You can also try Vonovox and test if it performs better for you. It's a new voice changer with more active development. It also works with RVC pth voice models.

[wajahatalikhan751-bot]

You should update to the v1.5.3.18a but, instead of using the ML version, use the CUDA one so you use your GPU for RVC and get better performance. That said, demanding games like SL and vrchat may still cause issues. Besides the vram usage, which you can't just use less even if you delete unused models, you may easily hit a roof because those games are very intensive for the GPU and CPU. There is not much that you can do beyond reducing graphics, draw distance and limiting texture size to 1024 or lower for SL. Maybe the CUDA version performs better but I always found issues with vrchat even when I had the 3080Ti with 12GB VRAM.

You can also try Vonovox and test if it performs better for you. It's a new voice changer with more active development. It also works with RVC pth voice models.

I am using the cuda version and its v1.5 still i have the problem also when I don't play any game and try it, still the voice sometime distorts I use 748 chunks and the extra is all the way to full and idex is 0.7 , maybe its the voice problem and I do have amd cpu can it cause problem? Also how can I run it better and where can I get good quality voice models

[Kuuko-fokkusugaru]

Using index will always skyrocket the CPU usage. Isn't recommended for real time conversion.
Your CPU is fine as long as it have enough juice to carry them. You can run it better by setting index to 0. From there, you can't do much more than setting games to minimum settings and locking fps below your average ones so you make room for RVC to work properly. The only way of running it better is by getting a high end GPU like an nvidia 4090 or 5090.
About good quality models, you may need to search that by yourself. Is hard to know which ones are good quality until you try them. The distortions are tied to low quality models.