4-channel Fine-tuned model BSOD clock_watchdog_timeout between 45% and 80% completion when run on CPU

[NerdyKyogre]

Kind of at my wit's end with this one. Any time I attempt to split a song using the 4-channel fine-tuned model on CPU, I get a guaranteed CLOCK_WATCHDOG_TIMEOUT BSOD between 45% and 80% completion. The fast 4-channel model works perfectly, as does the 6-channel, and all models work using CUDA. If it matters, I'm using local FLACs as this is the only file type I'm interested in splitting.

What I've tried:

• Updating Stemroller (2.1.2 and 2.2.1 both affected)
• Moving the source file from my NAS to a local SSD
• Changing output type from FLAC to WAV (it seems to last slightly longer before crashing this way)
• Moving the output path to a secondary NVMe
• Reverting my overclock back to stock settings and running DISM health check + SFC
• Killing all background processes
• Using another computer entirely, with the same results

System 1 specs: i5-12600K, 32 GB DDR4, RX 6800, Windows 11 Pro
System 2 specs: i7-7700HK, 32 GB DDR4, GTX 1070 (CUDA works on this), Windows 10 Pro

[iffyloop]

This is a very strange issue and I'm not sure what could be causing it. Is there a reason to not use CUDA in your situation? That should be a lot faster anyways; the CPU mode is just a fallback for if CUDA isn't supported.

[NerdyKyogre]

The main reason I'm not using CUDA is because I don't have a supported GPU on my main desktop PC and would prefer not to have to break out my gaming laptop just for this. The only reason I mention that machine is that I wanted to test if the issue was specific to my desktop.

[iffyloop]

I see - that makes a lot of sense. Don't know of anything I can do about this, but hopefully someday another user will find this issue and have a more helpful comment. For now it may be worth just using the regular non-finetuned 4 stem model. Although the finetuned one is supposed to provide better quality, I haven't noticed a significant difference myself yet (but also haven't done thorough A/B comparisons). The finetuned one is a lot slower too (as you've probably seen).