Spaces:
Running
Running
Using neuron cache instead of creating a pull-request in the target repo
#1
by
dacorvo
- opened
Our experience with the neuron SDK showed us that a model exported to Neuron for a specific version might not work as expected for the next SDK versions.
This is why storing exported models in dedicated repositories is not the most efficient way to share them: it is better to just store the cached compilation artifacts in the aws-neuron/optimum-neuron_compile-cache directory instead. This can be achieved transparently after having exported a model by calling the optimum neuron cache synchronization helper, either programatically or through the CLI.
Got it, makes perfect sense. I'll modify the export workflow to use the compilation cache approach instead of storing exported models in repos.