When combining an optimal control problem (OCP) with a PyTorch model, L4CasADi offers a useful solution. However, compiling the OCP can trigger an error:

RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

This occurs after converting and compiling a PyTorch model in an OCP using L4CasADi. The problematic code pattern is:

import l4casadi as l4c

l4casadi_model = l4c.L4CasADi(pytorch_model, device=device)

The solution

Create a separate PyTorch model for the OCP instead:

import l4casadi as l4c

# copy model weight and bias from the original model
pytorch_model.load_state_dict(original_model.state_dict())
l4casadi_model = l4c.L4CasADi(pytorch_model, device=device)

This approach preserves gradient tracking by keeping distinct model instances — one for optimization and one for training.

Building PyTorch from source becomes necessary when using Python 3.8 on JetPack 4.6.3, since Nvidia doesn’t provide pre-built wheels for this configuration. Downgrading to Python 3.6 was possible, but ROS Foxy requires Python 3.8, which made building from source the only viable option.

Setup specifications

• Python 3.8
• PyTorch 1.10.2
• Jetson Xavier with JetPack 4.6.3 and L4T 32.7.3
• CUDA 10.2

Build steps

Clone and checkout:

git clone --recursive https://github.com/pytorch/pytorch
cd pytorch
git checkout tags/v1.10.2
git submodule sync
git submodule update --init --recursive --jobs 0

Maximize performance:

sudo nvpmodel -m 0
sudo jetson_clocks

Configure the build environment:

export USE_NCCL=1
export USE_DISTRIBUTED=1
export USE_QNNPACK=0
export USE_PYTORCH_QNNPACK=0
export TORCH_CUDA_ARCH_LIST="7.2"
export PYTORCH_BUILD_VERSION=1.10.2
export PYTORCH_BUILD_NUMBER=1

Install dependencies:

sudo apt-get install cmake libopenblas-dev libopenmpi-dev
python3 -m pip install -r requirements.txt
python3 -m pip install scikit-build ninja
python3 -m pip install setuptools==59.5.0

Set the compiler version:

sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-8 1
sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-8 1

Build and install:

python3 setup.py bdist_wheel
cd dist
python3 -m pip install --no-cache <wheel_name>
python3 -c "import torch; torch.cuda.is_available()"

Checking the CUDA architecture

To verify CUDA capability using the deviceQuery utility:

1. Navigate to /usr/local/cuda/samples/1_Utilities/deviceQuery
2. Run sudo make to build
3. Run ./deviceQuery to display the specifications

Jetson Xavier should report compute capability 7.2.

My environment

• Windows 11
• PyTorch 2.5.1
• CUDA 12.4

Download CUDA

Download CUDA 12.4 from NVIDIA. There are two installer types: the local installer downloads the entire CUDA package (about 3.0 GB), while the network installer is a lightweight version that downloads CUDA as you install.

Note: attempting to install CUDA via Conda produced errors and was unsuccessful.

Install PyTorch

Using Conda is the easiest route on Windows. For CUDA 12.4 compatibility:

conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia

Testing

Verify the installation succeeded:

import torch
torch.cuda.is_available()
# True

Conclusion

Make sure the version matches between pytorch-cuda and your CUDA version. If you need a different CUDA version, make sure your pytorch-cuda version changes accordingly.