The cuda:0 point hit close to home. Lost half a day on a production deploy once because GPU allocation was dynamic in our Kubernetes setup and nobody had thought to make that configurable. The research code had worked flawlessly on every lab machine for months. Such a small assumption, such a disproportionate headache.
Sadiqur Rahman
Building 100+ free tools for developers & everyday users at sadiqbd.com
The cuda:0 point hit close to home. Lost half a day on a production deploy once because GPU allocation was dynamic in our Kubernetes setup and nobody had thought to make that configurable. The research code had worked flawlessly on every lab machine for months. Such a small assumption, such a disproportionate headache.