DeepFaceLab
DeepFaceLab copied to clipboard
iperov
GeForce RTX 3090 Doesn't Response
THIS IS NOT TECH SUPPORT FOR NEWBIE FAKERS POST ONLY ISSUES RELATED TO BUGS OR CODE
Expected behavior
Faceset extraction from extracted images for data_src or data_dst on my RTX 3090 (GE FORCE).
Actual behavior
GPU shows up to select, but when selected it starts. Then it goes to the part that says "Extracting faces... Running on GeForce RTX 3090" and it hovers at 0% for a few minutes. Then I get an error message saying "GeForce RTX 3090 doesn't response. Terminating it."
Steps to reproduce
- Extract images (this works fine as it uses my CPU I think) from data_src/dst
- Click 4) data_src faceset extract
- Default settings
And thats' when the error happens.
Other relevant information
Using windows 10. I have done DDU and driver update, clean reinstallation of windows, and I've tested the GPU itself with 3d mark and several games. Whatever the issue is, it's unique to DFL.
have you tried turning Gpu Scheduling "ON" under start>graphic settings
(this setting shows if you have Windows 20h2 update ) ?
Such errors are not limited to your GTX model, nor are they limited to DFL. Unfortunately, Microsoft has undergone major changes in how it "verifies" memory access, much like MacOS running in rootless mode. If you can, run the whole thing in Administrator mode, then lock it to admin mode using properties and run it for Windows 7. I've found that some graphics code, even under premiere and after effects, will not run properly until admin mode and windows 7 are used, mainly because the way memory for the GPU is handled, but also for the code library support. Some code has been altered in WIN10, not just the directX support, but other code as well, whose changes have affected the way memory is passed into the card. Unfortunately, without being able to trace the memory location specifically, I couldn't begin to point you to where it should go. I ran this a while ago, in beta, and I had to run it in Win7 with admin. Not a pleasant thing but required.
I've been trying to work on this for almost a month now with no luck. Haven't tried this yet but good to hear someone else had the same problem and solved it. Will give this a shot thanks.
Such errors are not limited to your GTX model, nor are they limited to DFL. Unfortunately, Microsoft has undergone major changes in how it "verifies" memory access, much like MacOS running in rootless mode. If you can, run the whole thing in Administrator mode, then lock it to admin mode using properties and run it for Windows 7. I've found that some graphics code, even under premiere and after effects, will not run properly until admin mode and windows 7 are used, mainly because the way memory for the GPU is handled, but also for the code library support. Some code has been altered in WIN10, not just the directX support, but other code as well, whose changes have affected the way memory is passed into the card. Unfortunately, without being able to trace the memory location specifically, I couldn't begin to point you to where it should go. I ran this a while ago, in beta, and I had to run it in Win7 with admin. Not a pleasant thing but required.
I don't suppose you know where to get a working win 7 license key
Have someone fixed this problem? My cuda version is 11.x and in tensorflow, "tf.test.is_gpu_available()" gives me a 'True'.
Just increase this option self.no_response_time_sec in core/joblib/SubprocessorBase.py
