The problem
I'm working on a GitHub Actions workflow to cog push a model to Replicate. Part of the release process includes running a script to download weights:
cog run script/download-weights
Unfortunately that script can't be run in the Actions workflow, because GitHub's hosted runners don't (yet?) have GPUs, and this model has gpu: true set in it's cog.yaml file. But there isn't anything in the download-weights script that specifically requires a GPU to run. The current workaround is to approximate the Cog environment by manually installing Python, installing Python dependencies, etc, but that's a pain.
A possible solution
It would be nice to be able to run one-off cog commands and somehow specify that the GPU is not needed:
cog run script/download-weights --no-gpu-needed-thanks
The problem
I'm working on a GitHub Actions workflow to
cog pusha model to Replicate. Part of the release process includes running a script to download weights:cog run script/download-weightsUnfortunately that script can't be run in the Actions workflow, because GitHub's hosted runners don't (yet?) have GPUs, and this model has
gpu: trueset in it'scog.yamlfile. But there isn't anything in thedownload-weightsscript that specifically requires a GPU to run. The current workaround is to approximate the Cog environment by manually installing Python, installing Python dependencies, etc, but that's a pain.A possible solution
It would be nice to be able to run one-off cog commands and somehow specify that the GPU is not needed:
cog run script/download-weights --no-gpu-needed-thanks