As discussed, the API currently prevents any scenario that would require this support. We hence will just make it a clean error. We'll likely need to come back to it when we expand the API e.g. to integrate with other GPU-accelerated packages.

I might be missing something, but why does the API prevent this scenario? I thought this was possible today, and if so we need an error or an implementation

thrust::host_vector<thrust::complex<double>> hostVector(4);
hostVector[0] = M_SQRT1_2;
hostVector[3] = M_SQRT1_2;
thrust::device_vector<thrust::complex<double>> devState = hostVector;
auto *devPtr = thrust::raw_pointer_cast(&devState[0]);
auto deviceState = cudaq::state::from_data(
      std::make_pair(reinterpret_cast<std::complex<double> *>(devPtr), 4));
auto kernel = [](cudaq::state s) __qpu__ { cudaq::qvector q (s); };
cudaq::get_state(kernel, deviceState).dump()

There's a test for the from_data part in unittests/gpu/get_state_tester.cu.

I might be missing something, but why does the API prevent this scenario? I thought this was possible today, and if so we need an error or an implementation

thrust::host_vector<thrust::complex<double>> hostVector(4);
hostVector[0] = M_SQRT1_2;
hostVector[3] = M_SQRT1_2;
thrust::device_vector<thrust::complex<double>> devState = hostVector;
auto *devPtr = thrust::raw_pointer_cast(&devState[0]);
auto deviceState = cudaq::state::from_data(
      std::make_pair(reinterpret_cast<std::complex<double> *>(devPtr), 4));
auto kernel = [](cudaq::state s) __qpu__ { cudaq::qvector q (s); };
cudaq::get_state(kernel, deviceState).dump()

There's a test for the from_data part in unittests/gpu/get_state_tester.cu.

This is a somewhat 'stale' comment in this addQubitsToState(std::size_t count, const void *stateIn) code path whereby we expect a complex<T>* pointer.

From the current set of qvector constructors, this can only be a host pointer.

The other one below, addQubitsToState(const cudaq::SimulationState &in_state) will handle the above example: handling the qvector(state&) constructor.

Hence, I think we just need to clean up those obsolete comments and add a verification check about this assumption.