moved inherit_stack_trace inside inplace_allocempty decorator

theano/gpuarray/dnn.py

@@ -3383,19 +3383,16 @@ def local_abstractconv_gi_cudnn(node):
 
 @inplace_allocempty(GpuDnnConv, 2)
 def local_dnn_conv_inplace(node, inputs):
-    with inherit_stack_trace(node.outputs):
     return [GpuDnnConv(algo=node.op.algo, inplace=True, num_groups=node.op.num_groups)(*inputs)]
 
 
 @inplace_allocempty(GpuDnnConvGradW, 2)
 def local_dnn_convgw_inplace(node, inputs):
-    with inherit_stack_trace(node.outputs):
     return [GpuDnnConvGradW(algo=node.op.algo, inplace=True, num_groups=node.op.num_groups)(*inputs)]
 
 
 @inplace_allocempty(GpuDnnConvGradI, 2)
 def local_dnn_convgi_inplace(node, inputs):
-    with inherit_stack_trace(node.outputs):
     return [GpuDnnConvGradI(algo=node.op.algo, inplace=True, num_groups=node.op.num_groups)(*inputs)]
 
 optdb.register('local_dnna_conv_inplace',

theano/gpuarray/opt_util.py

@@ -5,6 +5,7 @@ import numpy as np
 
 from theano import tensor, scalar as scal, Constant
 from theano.gof import local_optimizer
+from theano.gof.opt import inherit_stack_trace
 from theano.tensor import (DimShuffle, get_scalar_constant_value,
                            NotScalarConstantError)
 
@@ -326,6 +327,7 @@ def inplace_allocempty(op, idx):
                     len(alloc.clients) > 1):
                 alloc_op = GpuAllocEmpty(alloc.owner.op.dtype, alloc.owner.op.context_name)
                 inputs[idx] = alloc_op(*alloc.owner.inputs)
+            with inherit_stack_trace(node.outputs):
                 return maker(node, inputs)
         return opt
     return wrapper