Yep, everything’s installed correctly.
In the previous version, there was an issue if it detected only one song, but I fixed that in the latest version… Anyway, just to be sure that’s not the problem, try this:
Press Shift+Click on “AI Analysis,” which will run the analysis again
Tell it there’s more than one song, even if there’s only one
Wait for the analysis to finish to see if it found sections this time.
Finding the sections is what matters.
The BPMs it shows you are per song, not per section; it’s a general average of the song.
I tried that just now and it produced the same result. I should have mentioned, any song I run through it, regardless of the tempo, always comes back saying it’s 120bpm. If it helps, here’s the terminal output when I run the AI analysis.
The error appears in the log: “[ERROR] [Sections] Section-based detection failed: CUDA error: no kernel image is available for execution on the device…”
Okay, the architecture is correct.
It’s important to know about the second message you mentioned that doesn’t return anything. When you say it doesn’t return anything, do you mean:
It tells you that flash attn isn’t installed
It gives you an error
What does it show you? Because some kind of message should appear.
Okay, it looks like another dependency might be using a more modern architecture that your GPU doesn’t support.
You’ll need to open a Command Prompt again and run these commands:
python\python.exe -m pip list | findstr -i "xformers bitsandbytes triton"
Let me know the result.
Then try uninstalling xformers and testing the app:
python\python.exe -m pip uninstall xformers -y
If it keeps failing, you could try uninstalling bitsandbytes, although it’s not very likely that it’s causing the problem:
Let me know if any of these solutions helped to solve your problem.
Don’t worry about uninstalling these things. Two of them are accelerators, so you might lose a bit of speed, but the app should still work fine. The last one is used for quantizing LLM models (in JOI moods, which is something you probably won’t even use).
Damn it! Well, at least we’ve ruled out a few things.
Sorry for the trouble, but it would be great to find the error because it would help to prepare the application for other users who have a Nvidia from the same series.
If it’s not too much trouble, I’d need you to do the following:
Open a CMD again and run these commands:
set CUDA_LAUNCH_BLOCKING=1
python\python.exe -c "import torch; t = torch.zeros(1).cuda(); print('CUDA OK')"
If that returns “CUDA OK,” then the basic Torch functionality is working properly, and the problem lies in a specific package that is loaded during section analysis.
Next, to generate a log with more specific error details, you’ll need to run the application as follows:
Open a CMD window and:
set CUDA_LAUNCH_BLOCKING=1
python\python.exe app.py
Then attach the most recent .log file you have in the \logs folder.
Thanks!
Sounds like troubleshooting is mutually beneficial! It returns CUDA OK when I run those commands. When I run the second set, I get the following:
Traceback (most recent call last):
File "F:\Programs\FAPS_v0.8.1\app.py", line 64, in <module>
from core import (
ModuleNotFoundError: No module named 'core'
It didn’t generate a new log file after running that command.
Thanks!
It seems the problem lies with all-in-one, which is precisely what’s responsible for finding the sections, even though Demucs works fine for you.
The problem is likely “flash-attn,” which, even though you uninstalled it, PyTorch 2.7 has its own flash attention backend built in.
To force it not to use it, you should replace your “main.py” file with this one:
Let me know if this works for you.
If it does, you can try reinstalling the things that were uninstalled for testing purposes (except for Flash Attention, of course):
Thanks!
Yep, setups that use the latest AI models are always a bit tricky (dealing with PyTorch, CUDA, dependencies, etc.).
You can find documentation in \docs, though it’s not up to date because, honestly, until it’s fully polished and all the changes are done, it’s not worth maintaining documentation that keeps changing, but I’m sure it’ll help you with some things.
Since you mention “PRIMARY,” I assume you’re using the “Fap Mixer” mood. Just wanted to let you know that I’ve actually been tweaking it a bit today. For example, something you probably don’t know because it’s not very intuitive is that if you open “Configure AI,” you can choose which stem to use for each “block”.
Now, with the changes I’ve made, there’s a drop-down menu right next to each block. Also, for “Fap Mixer,” I’ve created another automatic method for assigning stems to blocks that should work better.
Oh, and by the way, in the new version v0.8.1, you no longer need to install Python 3.12, it comes with an embedded version of Python, although you still need to run “install.bat”
