You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Let's unify interfaces(genTensors, genTrainingTensors, and genKernels) of TrainableBackendContext.
Why
Currently those three functions(genTensors, genTrainingTensors, and genKernels) are called sequentially because the creation of tensors is the prerequisite for generating kernels in most backends and generating kernels has no prerequisite related to generating tensors.
However, we are going to introduce a feature in which generating kernels is the prerequisite for generating tensors in the train backend. That means generating tensors and generating kernels are interdependent on each other. Therefore, the calling priority of those three functions cannot be determined outside the train backend. We need to pass the determination to each training backend instead of onert core.
What
Let's unify interfaces(
genTensors
,genTrainingTensors
, andgenKernels
) ofTrainableBackendContext
.Why
Currently those three functions(
genTensors
,genTrainingTensors
, andgenKernels
) are called sequentially because the creation of tensors is the prerequisite for generating kernels in most backends and generating kernels has no prerequisite related to generating tensors.However, we are going to introduce a feature in which generating kernels is the prerequisite for generating tensors in the
train
backend. That means generating tensors and generating kernels are interdependent on each other. Therefore, the calling priority of those three functions cannot be determined outside thetrain
backend. We need to pass the determination to each training backend instead of onert core.draft : #13629
The text was updated successfully, but these errors were encountered: