Also register in fast_compile on the CPU to only compute the forward when the grad isn't needed.

theano/tensor/nnet/ctc.py

@@ -224,7 +224,7 @@ def ctc(activations, labels, input_lengths):
 
 
 # Disable gradient computation if not needed
-@register_canonicalize
+@register_canonicalize('fast_compile')
 @local_optimizer([ConnectionistTemporalClassification])
 def local_ctc_no_grad(node):
     if isinstance(node.op, ConnectionistTemporalClassification):