Output AST of code run in cuda?

(Charles Durham) #1

Is there a way to output some form of the compiled cuda code that gets run in pytorch or some representation of the graph in human readable form? Preferably without a rebuild of the pytorch.

Thanks!

(Simon Wang) #2

Do you mean from JIT? PyTorch w/o JIT runs everything as a dynamic graph so there isn’t compilation.

(Charles Durham) #3

Gotcha, is there a way to log Cuda kernel launches or something like that?

(Simon Wang) #4

The profiler does something similar to that, but at op level not at kernel level