[GH-ISSUE #1669] Is It Possible to Use CPU & RAM for Training?

bsakaguchi commented

2026-10-20 10:06:01 +00:00

Owner

Originally created by @TechVillain on GitHub (Jan 1, 2024).
Original GitHub issue: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/1669

For inference, I run RVC on CPU and RAM, it is a little slow, but works.

CPU Training

For Training, I only have 2GB of Nvidia VRAM, it won't work even with batch size of 1. It's a laptop, so it's impossible to replace the GPU without abandoning the whole computer with already installed programs.

Google Colab is not a good option either, as it's based on Linux and being complicated to setup for noobs like me.

The questions is, is it possible to modify the source code, and run training with CPU and RAM?

It will be slow, but that's ok. As long as it allows the training to run.

I have 12GB of RAM, Intel Quadcore i7 CPU, and I only need to train one model of Jay Chou's voice.

Please let me know what source code I can modify to make it happen.

Originally created by @TechVillain on GitHub (Jan 1, 2024). Original GitHub issue: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/1669 For inference, I run RVC on CPU and RAM, it is a little slow, but works. ### CPU Training For Training, I only have 2GB of Nvidia VRAM, it won't work even with batch size of 1. It's a laptop, so it's impossible to replace the GPU without abandoning the whole computer with already installed programs. Google Colab is not a good option either, as it's based on Linux and being complicated to setup for noobs like me. **The questions is, is it possible to modify the source code, and run training with CPU and RAM?** It will be slow, but that's ok. As long as it allows the training to run. I have 12GB of RAM, Intel Quadcore i7 CPU, and I only need to train one model of Jay Chou's voice. Please let me know what source code I can modify to make it happen.

bsakaguchi

2026-10-20 10:06:01 +00:00

closed this issue
added the
stale
label

bsakaguchi commented

2026-10-20 10:06:01 +00:00

Author

Owner

@kubinka0505 commented on GitHub (Jan 11, 2024):

not in colab, as FileNotFoundError: [Errno 2] No such file or directory: '/content/arsvc/logs/modelname/3_feature256'

@kubinka0505 commented on GitHub (Jan 11, 2024): not in colab, as `FileNotFoundError: [Errno 2] No such file or directory: '/content/arsvc/logs/modelname/3_feature256'`

bsakaguchi commented

2026-10-20 10:06:01 +00:00

Author

Owner

@bleriot14 commented on GitHub (Jan 19, 2024):

I am also fallowing this issue. Any chance ?

@bleriot14 commented on GitHub (Jan 19, 2024): I am also fallowing this issue. Any chance ?

bsakaguchi commented

2026-10-20 10:06:02 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Apr 28, 2024):

This issue was closed because it has been inactive for 15 days since being marked as stale.

@github-actions[bot] commented on GitHub (Apr 28, 2024): This issue was closed because it has been inactive for 15 days since being marked as stale.

bsakaguchi commented

2026-10-20 10:06:02 +00:00

Author

Owner

@thefellaguy commented on GitHub (May 14, 2024):

I train locally on a laptop with 6 cores and 32 gigs ram. Harvest works well.

@thefellaguy commented on GitHub (May 14, 2024): I train locally on a laptop with 6 cores and 32 gigs ram. Harvest works well.

bsakaguchi commented

2026-10-20 10:06:02 +00:00

Author

Owner

@TechVillain commented on GitHub (May 14, 2024):

That is amazing! Is there any special setup you did in order to run the training on your CPU and RAM?

What is your benchmark? How many minutes does it take for going through one epoch?

From: thefellaguy @.>
Sent: Monday, May 13, 2024 9:30:12 PM
To: RVC-Project/Retrieval-based-Voice-Conversion-WebUI @.>
Cc: TechVillain @.>; Author @.>
Subject: Re: [RVC-Project/Retrieval-based-Voice-Conversion-WebUI] Is It Possible to Use CPU & RAM for Training? (Issue #1669)

I train locally on a laptop with 6 cores and 32 gigs ram. Harvest works well.

—
Reply to this email directly, view it on GitHubhttps://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/1669#issuecomment-2109208433, or unsubscribehttps://github.com/notifications/unsubscribe-auth/A7ON6X4JUALQTEGV7MMRZDDZCGAMJAVCNFSM6AAAAABBI3RINKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMBZGIYDQNBTGM.
You are receiving this because you authored the thread.Message ID: @.***>

@TechVillain commented on GitHub (May 14, 2024): That is amazing! Is there any special setup you did in order to run the training on your CPU and RAM? What is your benchmark? How many minutes does it take for going through one epoch? ________________________________ From: thefellaguy ***@***.***> Sent: Monday, May 13, 2024 9:30:12 PM To: RVC-Project/Retrieval-based-Voice-Conversion-WebUI ***@***.***> Cc: TechVillain ***@***.***>; Author ***@***.***> Subject: Re: [RVC-Project/Retrieval-based-Voice-Conversion-WebUI] Is It Possible to Use CPU & RAM for Training? (Issue #1669) I train locally on a laptop with 6 cores and 32 gigs ram. Harvest works well. — Reply to this email directly, view it on GitHub<https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/1669#issuecomment-2109208433>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/A7ON6X4JUALQTEGV7MMRZDDZCGAMJAVCNFSM6AAAAABBI3RINKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMBZGIYDQNBTGM>. You are receiving this because you authored the thread.Message ID: ***@***.***>

bsakaguchi commented

2026-10-20 10:06:02 +00:00

Author

Owner

@thefellaguy commented on GitHub (May 14, 2024):

no special setup. just make sure the files are in the right place. git clone the repo, then pulled the files from hugging face that go in pretrained_v2 folder. im on a linux laptop with no gpu. 6 cores 32 gigs ram. epochs were between 2:30 and 3:15. total time for 20 epochs was about 55 mins. it did want python 3.8, so i had to install than and i modified the run script to use a python 3.8 venv. (default python on this laptop is 3.12) it was pretty cut and dry. im gonna move it to my server later. that has more hp, should be able to go faster.

@thefellaguy commented on GitHub (May 14, 2024): no special setup. just make sure the files are in the right place. git clone the repo, then pulled the files from hugging face that go in pretrained_v2 folder. im on a linux laptop with no gpu. 6 cores 32 gigs ram. epochs were between 2:30 and 3:15. total time for 20 epochs was about 55 mins. it did want python 3.8, so i had to install than and i modified the run script to use a python 3.8 venv. (default python on this laptop is 3.12) it was pretty cut and dry. im gonna move it to my server later. that has more hp, should be able to go faster.

bsakaguchi commented

2026-10-20 10:06:02 +00:00

Author

Owner

@TechVillain commented on GitHub (May 15, 2024):

Wow, you are getting around 22 epochs per hour, on a CPU, that's impressive! If you leave it cooking during the night and go to bed sleep, the next morning you wake up you get about 150 epochs. Normally a 200 to 300 epoch model is considered a high quality model.

How many minutes of audio do you have for the training dataset? Is it sing or speech, or both?

With your 32GB RAM, what batch size number did you set? I assume the training didn't use up the full 32GB memory, am I right?

Also, I assume, in your Python 3.8 venv, you installed the torch-cpu version, not the torch-cuda version, in order to enable CPU training, right?

My Discord account is the same name as my GitHub account. Add me as a friend. I'm interested in hearing your CPU training experience, and how accurate is your model compare to the source speaker.

From: thefellaguy @.>
Sent: Tuesday, May 14, 2024 4:13:18 PM
To: RVC-Project/Retrieval-based-Voice-Conversion-WebUI @.>
Cc: TechVillain @.>; Author @.>
Subject: Re: [RVC-Project/Retrieval-based-Voice-Conversion-WebUI] Is It Possible to Use CPU & RAM for Training? (Issue #1669)

no special setup. just make sure the files are in the right place. git clone the repo, then pulled the filed from hugging face that go in pretrained_v2 folder. im on a linux laptop with no gpu. 6 cores 32 gigs ram. epochs were between 2:30 and 3:15. total time for 20 epochs was about 55 mins. it did want python 3.8, so i had to install than and i modified the run script to use a python 3.8 venv. (default python on this laptop is 3.12) it was pretty cut and dry. im gonna move it to my server later. that has more hp, should be able to go faster.

—
Reply to this email directly, view it on GitHubhttps://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/1669#issuecomment-2111223501, or unsubscribehttps://github.com/notifications/unsubscribe-auth/A7ON6X3OWWZ3FZRXUWXPVY3ZCKD75AVCNFSM6AAAAABBI3RINKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMJRGIZDGNJQGE.
You are receiving this because you authored the thread.Message ID: @.***>

@TechVillain commented on GitHub (May 15, 2024): Wow, you are getting around 22 epochs per hour, on a CPU, that's impressive! If you leave it cooking during the night and go to bed sleep, the next morning you wake up you get about 150 epochs. Normally a 200 to 300 epoch model is considered a high quality model. How many minutes of audio do you have for the training dataset? Is it sing or speech, or both? With your 32GB RAM, what batch size number did you set? I assume the training didn't use up the full 32GB memory, am I right? Also, I assume, in your Python 3.8 venv, you installed the torch-cpu version, not the torch-cuda version, in order to enable CPU training, right? My Discord account is the same name as my GitHub account. Add me as a friend. I'm interested in hearing your CPU training experience, and how accurate is your model compare to the source speaker. ________________________________ From: thefellaguy ***@***.***> Sent: Tuesday, May 14, 2024 4:13:18 PM To: RVC-Project/Retrieval-based-Voice-Conversion-WebUI ***@***.***> Cc: TechVillain ***@***.***>; Author ***@***.***> Subject: Re: [RVC-Project/Retrieval-based-Voice-Conversion-WebUI] Is It Possible to Use CPU & RAM for Training? (Issue #1669) no special setup. just make sure the files are in the right place. git clone the repo, then pulled the filed from hugging face that go in pretrained_v2 folder. im on a linux laptop with no gpu. 6 cores 32 gigs ram. epochs were between 2:30 and 3:15. total time for 20 epochs was about 55 mins. it did want python 3.8, so i had to install than and i modified the run script to use a python 3.8 venv. (default python on this laptop is 3.12) it was pretty cut and dry. im gonna move it to my server later. that has more hp, should be able to go faster. — Reply to this email directly, view it on GitHub<https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/1669#issuecomment-2111223501>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/A7ON6X3OWWZ3FZRXUWXPVY3ZCKD75AVCNFSM6AAAAABBI3RINKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMJRGIZDGNJQGE>. You are receiving this because you authored the thread.Message ID: ***@***.***>

bsakaguchi commented

2026-10-20 10:06:02 +00:00

Author

Owner

@thefellaguy commented on GitHub (May 15, 2024):

sent you a friend request, but I'm not a fan of putting it online :P

pytorch came down with running the run.sh file that came down with the git repo. the last lines of the log file before training are:

Retrieval-based-Voice-Conversion-WebUI/.venv/lib64/python3.8/site-packages/torch/nn/utils/weight_norm.py:28: UserWarning: torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm.
  warnings.warn("torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm.")
move model to cpu

launching sytem default python can't import torch. launcing the rvc venv python shows this:

>>> print(torch.__version__)
2.3.0+cu121
>>>

quick look on the filesystem for torch maxes out my terminal frame buffer. So no idea. I can say the algo's that require a gpu don't work. fails with a missing gpu error. So I'm not sure that only one or the other gets installed. First time messing with pytorch though, so pure speculation here.

I used the recommended 10 mins of audio.

My plan is to move this to my server, its a 16 core with 64 gigs of ram. Its an older server, practically e-waste now, but I can let this stuff run all night. It has 16TB of disk as well so i can keep more logs and chart progress better. Other threads just dismiss cpu training as taking weeks, but I dont think they actually tried. Getting this up and running on linux the first time too only took about 30 mins. I dunno if linux is just easier or what. The only thing I had to do manually was install python 3.8 and populate the pretrained_v2 folder from the hugging face repo. the reason i had to manually install python3.8 was because the run.sh setup script assumed ubuntu of you were on linux and tried apt get. it only tries brew and apt get then tells you to do it yourself. And I sympathize with that!

@thefellaguy commented on GitHub (May 15, 2024): sent you a friend request, but I'm not a fan of putting it online :P pytorch came down with running the `run.sh` file that came down with the git repo. the last lines of the log file before training are: ``` Retrieval-based-Voice-Conversion-WebUI/.venv/lib64/python3.8/site-packages/torch/nn/utils/weight_norm.py:28: UserWarning: torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm. warnings.warn("torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm.") move model to cpu ``` launching sytem default python can't import torch. launcing the rvc venv python shows this: ``` >>> print(torch.__version__) 2.3.0+cu121 >>> ``` quick look on the filesystem for torch maxes out my terminal frame buffer. So no idea. I can say the algo's that require a gpu don't work. fails with a `missing gpu` error. So I'm not sure that only one or the other gets installed. First time messing with pytorch though, so pure speculation here. I used the recommended 10 mins of audio. My plan is to move this to my server, its a 16 core with 64 gigs of ram. Its an older server, practically e-waste now, but I can let this stuff run all night. It has 16TB of disk as well so i can keep more logs and chart progress better. Other threads just dismiss cpu training as taking weeks, but I dont think they actually tried. Getting this up and running on linux the first time too only took about 30 mins. I dunno if linux is just easier or what. The only thing I had to do manually was install python 3.8 and populate the `pretrained_v2` folder from the hugging face repo. the reason i had to manually install python3.8 was because the `run.sh` setup script assumed ubuntu of you were on linux and tried `apt get`. it only tries brew and apt get then tells you to do it yourself. And I sympathize with that!

Rows
Columns

[GH-ISSUE #1669] Is It Possible to Use CPU & RAM for Training? #1108

CPU Training