#pragma once #include namespace torch { namespace jit { // propagate autograd zero information through a gradient graph and // remove grad_of blocks if present. // Note: this is a very limited pass. It only propagates autograd zeros for // operations generated by the symbolic autodiff code and cleans up // AutogradAdds when possible. Outputs of other nodes are conservatively // marked Unknown and not optimized. TORCH_API void specializeAutogradZero(Graph &g); } // namespace jit } // namespace torch