Closed blooop closed 2 weeks ago
Here is how we create a pip package. Use either a docker with pytorch or a python environment with torch already installed and run the below commands:
python3 -m pip install build
cd curobo && python3 -m build --no-isolation
This will create the .whl
file that you can host in your pip registry.
Let me know if this works. You might have to install venv
if it's not already installed.
We are looking at putting on pypi.
Thanks. Those steps produce the wheel, but the jit compile steps don't go away. Its not really that big of a deal though, I can wait for an official pip as this temporary solution works well enough.
You can also reduce the compilation time by setting this environment variable export TORCH_CUDA_ARCH_LIST="7.0+PTX"
to only compile for one architecture with forward compatability.
We are deferring this to a later time.
I made curobo into a pip package using:
and uploaded it to a local pip registry so that I'm now able to:
on different repos without needing to compile it each time.
However, when I use the pip installed curobo wheel I need to jit compile the cuda kernels every time. The output looks like:
Thanks