On Thu, Nov 25, 2021 at 02:45 AM, Nicolas Jeker wrote:
On Wed, 2021-11-24 at 09:55 -0800, Tim Orling wrote:Ah OK, I wasn't aware of the the python naming convention. That is likely my problem since I'm using Honister which uses Python 3.9. I pulled the wheel from NVIDIA's forums for pytorch, unfortunately they've not released one for Python 3.9 so I will likely have to create it myself using the build from scratch method described in the article I linked. Unfortunately this will likely open a new can of worms...
There's a section about building from source with a patch in the
I'm using Honister and the machine is 'jetson-tx2-dev-kit-tx2i', I'm making a custom distro based on the meta-tegra-demo from this:
A large part of the problem that I have is that many of these custom packages don't provide a nice .tar.gz pypi source distribution that I could use with the pypi class. My target install ends up on a spacecraft, so there is a strong desire to have a fully managed build that can just be flashed onto the target tx2i without the need to do any post installation or configuration. Sadly I'm finding that a lot of these third party dependencies do not lend themselves well to the yocto paradigm. I'm thinking that I may have to setup the target board, install the third party packages on the target using whatever is required to do that, then copy the build products back to the host and use a bitbake recipe to just do_install:append the built products into the rootfs during the yocto build. Does that sound feasible? I've not had to do this before, but it seems like it might be a reasonable approach given the complexity of the situation.