fps stuck on 0.2

[crybaby212]

i have a 6GB rtx 3060 with windows 10 pro, 16GB ram and intel i7-2600 and I have tried installing CUDA and CUDANN but I have no idea what path to install it on for it to work and its only working on my CPU. when I try to run it with cuda, I have tried multiple versions of cuda to make sure I'm using the correct one, it shows me this.

someone please tell me how to do this step by step because I have spent my whole day on it and I cant get it to work.

[galoisgroupcn]

Let's fix your CUDA/GPU issue step by step! Your RTX 3060 should be able to run CUDA workloads, and the problem is likely with CUDA/cuDNN installation, driver, or Python environment mismatches. Here's what you should do:

1. Verify Your NVIDIA Drivers

Open the NVIDIA Control Panel or run nvidia-smi in Command Prompt.
If you get a version and see your GPU, your driver is installed.
If not, download the latest driver for 3060 from NVIDIA's site.

2. Uninstall ALL CUDA Toolkits and cuDNNs

Go to "Add or Remove Programs" and uninstall every entry named "NVIDIA CUDA Toolkit" and any cuDNN you installed.

Why? Multiple versions often conflict.

3. (Re)Install CUDA & cuDNN the Right Way

a) Find the Right Version for Your PyTorch

Go to PyTorch Get Started.
Select your settings (Stable, Windows, Pip/Conda, Python, CUDA 12.1 for 3060 is recommended).
Copy the provided install command. For example, for pip:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

Do NOT install CUDA Toolkit separately for PyTorch unless you plan to compile CUDA code yourself. PyTorch wheels come with their own CUDA runtime.

b) Install cuDNN (Rarely Needed for PyTorch)

For most users, you do not need to install cuDNN manually when using PyTorch wheels.

4. Check Your Python Environment

Open Command Prompt, run

python -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0) if 
torch.cuda.is_available() else 'No device')"

Output should be:

True
NVIDIA GeForce RTX 3060

If "False" or "No device", your install is broken. Go back to step 3.

5. Check run.py for Device Selection

In your project, make sure you're not forcing CPU.
Look in run.py for any lines like:

device = torch.device('cpu')

or

device = "cpu"

If so, change to:

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

or

device = "cuda" if torch.cuda.is_available() else "cpu"

If you want, I can review your run.py for you—just let me know!

6. Run Your Script With the Right Environment

Always open the Anaconda Prompt or Command Prompt where you installed PyTorch.
Activate your environment (if using conda or venv).
Run your script using:
sh
python run.py

If you have a run-cuda.bat, open it with Notepad and confirm it just calls python run.py (and not something special that disables CUDA).

7. Monitor GPU Usage

Open Task Manager > Performance > GPU.
Or run nvidia-smi while your script is running. You should see a process using the GPU.

8. If You Still Get 0.2 FPS
   It's likely your model is running on CPU, or there's another bottleneck (like disk I/O or data loading).
   Post the output of:

sh

python -c "import torch; print(torch.__version__); print(torch.cuda.is_available()); 
print(torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'No device')"

Or share your run.py so I can check for issues!

Summary

1. You don't need to install CUDA Toolkit manually for PyTorch—just install the CUDA-enabled PyTorch wheel.
2. Make sure your script is using cuda and not forced to cpu.
3. Use Task Manager or nvidia-smi to confirm actual GPU usage.

If you want, upload your run.py contents here, and I'll double-check the device selection code for you.

[crybaby212]

i have followed the steps and it says a syntax error for python -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0) if
torch.cuda.is_available() else 'No device')"

also, when I run it now there is a black square along with the low fps which wasn't there before. i have run pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121 but it says an error for one of the things it installs. is there any way someone could just help me install the program from the beginning and show me exactly which path to install cuda and how to use the GPU because this is just very complicated.

[prgarg007]

@crybaby212 please run 'nvidia-smi' on your command prompt & share output here. Let's try to debug the issue with your drivers

[manumaan]

Thanks @galoisgroupcn Following your steps I was able to make it run on RTX 2050. FPS went fro 0.4 to 1.5 now. Is there any way to bring it up further?

[Duranntl]

Using a RTX 3080 10GB il getting around 2 fps. (1.7-19 fps exactly) but don't know how can improve this...

im trying to use another model instead of "buffalo_l" to improve performance a little, but when trying to use "antelopev2" when i hit live, i have this errors...

[prgarg007]

@Duranntl seems cuda is not configured properly in your setup & its running on CPU instead.
please run 'nvidia-smi' on your command prompt & share output here.

[Duranntl]

@Duranntl seems cuda is not configured properly in your setup & its running on CPU instead. please run 'nvidia-smi' on your command prompt & share output here.

Yeah, i confirmed that when i opened task manager, the cpu utilization goes higher when i use the app

[prgarg007]

@Duranntl oka, please check if for torch and related libraries you are using the Cuda 12.9 version for them (if using any other version it won't work)
If you followed the original requirements, then you need to upgrade your torch packages for cuda 12.9, as in the original requirements it's using cuda 12.8 (which in your case would not be compatible)