# GPU usage zero with Pennylane and PyTorch **URL:** \ **Category:** PennyLane Help\ **Created:** [July 21, 2023, 12:41pm UTC](https://discuss.pennylane.ai/t/gpu-usage-zero-with-pennylane-and-pytorch/3204 "2023-07-21T12:41:32Z")\ **Posts on this page:** 5\ **Page:** 1
**Author:** ![imakash](https://yyz2.discourse-cdn.com/flex012/user_avatar/discuss.pennylane.ai/imakash/32/1428_2.png) [@imakash](https://discuss.pennylane.ai/u/imakash)\ **Post date:** [July 21, 2023, 12:41pm UTC](https://discuss.pennylane.ai/t/gpu-usage-zero-with-pennylane-and-pytorch/3204/1 "2023-07-21T12:41:32Z")
Hello, I am trying to train two models on a GPU cluster: 1. Hybrid quantum-classical reinforcement learning (RL) model and 2. Classical deep RL model. The model is pretty complex and therefore it’s difficult to share the complete code on this forum. When I train the classical model, the training happens on GPU but, when I train the quantum-classical model, I realized that the GPU usage is 0, thus making each iteration extremely slow. I tried debugging the code to figure out the issue and I realized that the input tensor as well as the parameters of the model (classical and quantum) are on GPU. Following are further details of my implementation. I am using V100-16GB GPU for the ```auto dev = qml.device("default.qubit", wires=4) # The qnode is created using the following annotation: # @qml.qnode(dev, interface='torch') ``` Please let me know if there are any flags that I should change or any tensor that I should evaluate to figure out the reason why GPU is not being used for quantum simulation. Apologies that I couldn’t share a working code snippet to reproduce the issue. **Output of `qml.about()`** Name: PennyLane Version: 0.28.0 Summary: PennyLane is a Python quantum machine learning library by Xanadu Inc. Home-page: [GitHub - PennyLaneAI/pennylane: PennyLane is a cross-platform Python library for differentiable programming of quantum computers. Train a quantum computer the same way as a neural network.](https://github.com/XanaduAI/pennylane) Author: Author-email: License: Apache License 2.0 Location: /home/sinha/miniconda3/envs/qns/lib/python3.8/site-packages Requires: appdirs, autograd, autoray, cachetools, networkx, numpy, pennylane-lightning, requests, retworkx, scipy, semantic-version, toml Required-by: PennyLane-Lightning, PennyLane-Lightning-GPU Platform info: Linux-5.15.0-1028-nvidia-x86\_64-with-glibc2.17 Python version: 3.8.16 Numpy version: 1.22.3 Scipy version: 1.7.3 Installed devices: - lightning.gpu (PennyLane-Lightning-GPU-0.30.0) - lightning.qubit (PennyLane-Lightning-0.28.1) - default.gaussian (PennyLane-0.28.0) - default.mixed (PennyLane-0.28.0) - default.qubit (PennyLane-0.28.0) - default.qubit.autograd (PennyLane-0.28.0) - default.qubit.jax (PennyLane-0.28.0) - default.qubit.tf (PennyLane-0.28.0) - default.qubit.torch (PennyLane-0.28.0) - default.qutrit (PennyLane-0.28.0) - null.qubit (PennyLane-0.28.0) ---
**Author:** ![Alvaro\_Ballon](https://yyz2.discourse-cdn.com/flex012/user_avatar/discuss.pennylane.ai/alvaro_ballon/32/874_2.png) [@Alvaro\_Ballon](https://discuss.pennylane.ai/u/Alvaro_Ballon)\ **Post date:** [July 21, 2023, 3:04pm UTC](https://discuss.pennylane.ai/t/gpu-usage-zero-with-pennylane-and-pytorch/3204/2 "2023-07-21T15:04:29Z")
Hi @imakash Sorry, but we seem to be lacking any information to figure out what exactly is going on here. However, [this part](https://docs.pennylane.ai/en/stable/introduction/interfaces/torch.html#gpu-and-cuda-support) of the documentation may be of use to you. If the wire count is small, GPUs are usually much slower. You might be better off using a CPU at these scales, but it might also depend on the code. Hope this helps! Alvaro ---
**Author:** ![imakash](https://yyz2.discourse-cdn.com/flex012/user_avatar/discuss.pennylane.ai/imakash/32/1428_2.png) [@imakash](https://discuss.pennylane.ai/u/imakash)\ **Post date:** [July 21, 2023, 3:16pm UTC](https://discuss.pennylane.ai/t/gpu-usage-zero-with-pennylane-and-pytorch/3204/3 "2023-07-21T15:16:26Z")
Hello @Alvaro_Ballon Thanks for your response. My implementation is actually inspired by the documentation you shared in your answer. I understand that it would be difficult for you to come up with a solution for my problem without the code to reproduce the error. However, if you could suggest some debugging steps to help me out here? Best Regards, Akash ---
**Author:** ![Alvaro\_Ballon](https://yyz2.discourse-cdn.com/flex012/user_avatar/discuss.pennylane.ai/alvaro_ballon/32/874_2.png) [@Alvaro\_Ballon](https://discuss.pennylane.ai/u/Alvaro_Ballon)\ **Post date:** [July 21, 2023, 7:45pm UTC](https://discuss.pennylane.ai/t/gpu-usage-zero-with-pennylane-and-pytorch/3204/4 "2023-07-21T19:45:40Z")
Hello @imakash Maybe you could try using one of the lightning devices? If that doesn’t work, the Torch/JAX backprop GPU pipeline could be faster than lightning at those problem sizes. Even so, a CPU is probably your best bet here. I’m sorry, but without more information, these are all the suggestions we can come up with ☹ Cheers! Alvaro ---
**Author:** ![kevinkawchak](https://yyz2.discourse-cdn.com/flex012/user_avatar/discuss.pennylane.ai/kevinkawchak/32/1801_2.png) [@kevinkawchak](https://discuss.pennylane.ai/u/kevinkawchak)\ **Post date:** [July 23, 2023, 8:48am UTC](https://discuss.pennylane.ai/t/gpu-usage-zero-with-pennylane-and-pytorch/3204/5 "2023-07-23T08:48:21Z")
You can also try 90 core qsim, Best regards. - Kevin