Multistage builds help separate heavier development dependencies from lighter production dependencies without compromising on all the things that make containers nice.
The classic example is separating compilation from deployment. If training is model compilation, why not borrow this best practice?
Less ambitiously, it might be possible to use them for splitting frontend/backend(s). Right now, we've got gradio and pandas in our Lambda, which is costing us whole seconds each inference.
Multistage builds help separate heavier development dependencies from lighter production dependencies without compromising on all the things that make containers nice.
The classic example is separating compilation from deployment. If training is model compilation, why not borrow this best practice?
Less ambitiously, it might be possible to use them for splitting frontend/backend(s). Right now, we've got gradio and pandas in our Lambda, which is costing us whole seconds each inference.