Understanding update() and compute() functions for building custom metric

[kkarrancsu]

I'd like to implement a custom TorchMetric in order to use it as a loss function. A couple of questions:

1 - I noticed there is a property: is_differentiable What specifically needs to be done w/ the metric implementation to ensure that the implemented metric is actually differentiable vs. simply just setting the property?
The next 2 questions apply to the DDP setting:
a - Is it OK to understand the update() method as the computation that happens on each GPU?
b - Is it OK to understand that the compute() method is what happens when the outputs of update are aggregated across all GPUs?