[GH-ISSUE #1584] Full Precision Option? #1055

New issue

Closed

opened 2026-10-20 10:05:32 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-20 10:05:32 +00:00

Owner

Originally created by @0RadicalLarry0 on GitHub (Nov 29, 2023).
Original GitHub issue: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/1584

I have an RX 5700XT attempting to train with rocm on linux and in order to get training to work I need to run this with full precision (float32) instead of half (float16).

I did disable fp16_run in the config, however, I'm unsure if everything actually uses full precision.

Setting fp16_run to false did fix it not training at all, but when comparing a model trained on the same dataset locally vs google colab the locally trained model sounds really strange. It has a thick accent that messes up every word and overall sounds like a really bad impression of the character. Yet the model trained on google colab comes out sounding great. Both trained to the same epochs too.

My only conclusion is that something somewhere is still using half precision and screwing up the training process. Unless there is something I'm missing it would be fantastic if there were an option to enable full precision pitch extraction and training.

Originally created by @0RadicalLarry0 on GitHub (Nov 29, 2023). Original GitHub issue: https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/1584 I have an RX 5700XT attempting to train with rocm on linux and in order to get training to work I need to run this with full precision (float32) instead of half (float16). I did disable fp16_run in the config, however, I'm unsure if everything actually uses full precision. Setting fp16_run to false did fix it not training at all, but when comparing a model trained on the same dataset locally vs google colab the locally trained model sounds really strange. It has a thick accent that messes up every word and overall sounds like a really bad impression of the character. Yet the model trained on google colab comes out sounding great. Both trained to the same epochs too. My only conclusion is that something somewhere is still using half precision and screwing up the training process. Unless there is something I'm missing it would be fantastic if there were an option to enable full precision pitch extraction and training.

bsakaguchi

2026-10-20 10:05:32 +00:00

closed this issue
added the
stale
label

bsakaguchi commented

2026-10-20 10:05:33 +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.