Cuda.is_available returns True in consel but False in program

deployment
2

Often these issues are caused by (accidentally) using different environments and binaries.
Check the torch.__path__ in both setups as these could be different.