PyTorch Setup Cheat Sheet II

In the previous post, we checked out PyTorch Setup as an open-source machine learning framework widely used for building + training neural networks. Here we installed and configured our nVidia driver and CUDA Toolkit to build some simple PyTorch programs. However, we built PyTorch3d + TorchSparse packages from source each time. Now we'd like to pre-build custom wheels housed on localhost for future examples reuse.
Let's check it out!

Example I
All examples listed are Python 3.10.19 compatible. Launch PyCharm | New Project. Enter the following info:

Location:	~/HelloPyTorch3dWheels
Interpreter type:	uv
Python version:	3.10
Path to uv:	~/.local/bin/uv

PyCharm should setup UV virtual environment and configure Python interpreter if not then enter commands:

uv venv --python 3.10   source .venv/bin/activate # OR .\.venv\Scripts\activate   which python   `which python` --version # Python 3.10.19

In the PyCharm Terminal | Enter the following commands for UV to install packages to build custom wheels:

uv pip install twine   uv pip install numpy==1.26.4   uv pip install --index-url https://download.pytorch.org/whl/cu121 \    "torch==2.2.0+cu121" "torchvision==0.17.0" "torchaudio==2.2.0"

Create directory to house custom wheels mkdir -p wheelhouse-cu121. Write shell scripts to code wheel logic:

PyTorch3D	chmod +x steveprobuild_pytorch3d_wheel.sh
TorchSparse	chmod +x steveprobuild_torchsparse_wheel.sh

Execute shell scripts to build custom wheels. These will be the custom wheels used locally on next examples:

COMMAND	ARTIFACT
bash steveprobuild_pytorch3d_wheel.sh	stevepropytorch3d-0.7.7-cp310-cp310-linux_x86_64.whl
bash steveprobuild_torchsparse_wheel.sh	steveprotorchsparse-2.0.0b0-cp310-cp310-linux_x86_64.whl

Example II
Repeat previous exercise using locally built wheels. Launch PyCharm | New Project. Enter the following info:

Location:	~/HelloPyTorch3dWheels
Interpreter type:	uv
Python version:	3.10
Path to uv:	~/.local/bin/uv

Copy custom wheels built from previous example: mkdir wheelhouse-cu121. Copy files into new directory:

mkdir -p wheelhouse-cu121   cp -r ../01-Example/wheelhouse-cu121 .

In the PyCharm Terminal | Enter the following commands for UV to install and sync package dependencies:

export MAX_JOBS=1   export NVCC_THREADS=1

uv lock   uv sync

Create the following file: main.py. Enter the command uv run main.py. Verify PyTorch packages installed!

Example II

Hello PyTorch3d   torch 2.2.0+cu121   pytorch3d 0.7.7   cuda 12.1   cuda True   torch_geometric 2.7.0   torch_scatter 2.1.2+pt22cu121   torch_sparse 0.6.18+pt22cu121   torch_cluster 1.6.3+pt22cu121   torch_spline_conv 1.2.2+pt22cu121   torchsparse 2.0.0b

Example III
Repeat previous exercise but upload custom wheels. Launch PyCharm | New Project. Enter the following info

Location:	~/HelloPyTorch3dWheels
Interpreter type:	uv
Python version:	3.10
Path to uv:	~/.local/bin/uv

Copy custom wheels built from previous example: mkdir wheelhouse-cu121. Copy files into new directory:

mkdir -p wheelhouse-cu121   cp -r ../01-Example/wheelhouse-cu121 .

In the PyCharm Terminal | Enter the following commands to install devpi packages to upload wheels locally:

uv pip install devpi-server   uv pip install devpi-client

uv pip install devpi-web   uv pip install twine

Launch Terminal #1. Initialize devpi-server on localhost port 3141 and start. Navigate to the homepage URL

devpi-init # Launch browser   devpi-server --host 127.0.0.1 --port 3141 # Navigate http://localhost:3141

Launch Terminal #2. Connect to the devpi-client and login. Create custom cuda-wheels index and activate

devpi use http://localhost:3141   devpi login root --password=''

devpi index -c cuda-wheels bases=root/pypi     devpi use root/cuda-wheels

Finally upload 2x custom wheel files to devpi-web available at http://localhost:3141/root/cuda-wheels

twine upload --repository-url http://localhost:3141/root/cuda-wheels/ wheelhouse-cu121/*

Example IV
Repeat previous exercise using uploaded wheels. Launch PyCharm | New Project. Enter the following info:

Location:	~/HelloPyTorch3dWheels
Interpreter type:	uv
Python version:	3.10
Path to uv:	~/.local/bin/uv

In the PyCharm Terminal | Enter the following commands to install devpi packages to consume local wheels:

uv pip install devpi-server   uv pip install devpi-client

uv pip install devpi-web   uv pip install twine

Launch Terminal #1. Initialize devpi-server on localhost port 3141 and start. Navigate to the homepage URL

devpi-init # Launch browser   devpi-server --host 127.0.0.1 --port 3141 # Navigate http://localhost:3141

In the PyCharm Terminal | Enter the following commands for UV to install and sync package dependencies:

export MAX_JOBS=1   export NVCC_THREADS=1

uv lock   uv sync

Create the following file: main.py. Enter the command uv run main.py. Verify PyTorch packages installed!

Example IV

Hello PyTorch3d   torch 2.2.0+cu121   pytorch3d 0.7.7   cuda 12.1   cuda True   torch_geometric 2.7.0   torch_scatter 2.1.2+pt22cu121   torch_sparse 0.6.18+pt22cu121   torch_cluster 1.6.3+pt22cu121   torch_spline_conv 1.2.2+pt22cu121   torchsparse 2.0.0b

Example V
Repeat previous exercise but wrap logic as Azure ML endpoint. Launch PyCharm | New Project. Enter info:

Location:	~/HelloAzureML
Interpreter type:	uv
Python version:	3.10
Path to uv:	~/.local/bin/uv

PyCharm should setup UV virtual environment and configure Python interpreter if not then enter commands:

uv venv --python 3.10   source .venv/bin/activate # OR .\.venv\Scripts\activate   which python   `which python` --version # Python 3.10.19

In the PyCharm Terminal | Enter the following commands for UV to install and sync package dependencies:

export MAX_JOBS=1   export NVCC_THREADS=1

uv lock   uv sync

Create the following directories app and tests. Enter all code for app/scoring.py. Test all code using pytest.
Launch terminals. Enter following commands: build Docker image run Docker container and submit request:

Terminal #1	docker build -t azml-gpu-local:latest .   docker run --gpus all -p 5001:5001 azml-gpu-local:latest
Terminal #2	curl --location --request POST 'http://localhost:5001/score' \    --header 'Content-Type: application/json' \    --data-raw '{    "points": [    [0.0, 0.0, 1.0],    [0.1, 0.0, 0.99],    [-0.1, 0.0, 0.99],    [0.0, 0.1, 0.99],    [0.0, -0.1, 0.99],    [0.7, 0.7, 0.0],    [-0.7, 0.7, 0.0],    [0.7, -0.7, 0.0],    [-0.7, -0.7, 0.0],    [0.0, 0.0, -1.0]    ]   }'

Example V

{    "num_vertices": 1524,    "num_faces": 2892,    "mean_pixel": 0.9926620125770569,    "image_shape": [    128,    128,    3    ],    "device": "cuda:0"   }

Cleanup as necessary: stop and remove the docker process esp. if that process is continued to be in use:

docker stop $(docker ps -q)   docker rm -f $(docker ps -q)

Summary
To summarize, we have built PyTorch3d and TorchSparse packages from source, pre-built the custom wheels housed on localhost for examples reuse. We are now in an excellent position: drive Deep Learning education!