Setting tensor's backend

[rgerganov]

After recent CUDA refactoring, we no longer set GGML_BACKEND_TYPE_GPU in ggml_backend_cuda_buffer_init_tensor() and tensors end up with GGML_BACKEND_TYPE_CPU backend even when using CUDA.

Is this a bug or part of a bigger plan which removes the backend property of the tensors?

[slaren]

The goal is to remove backend from the tensor entirely. Backends can use the buffer type to identify the storage location of the tensors instead. I think that at this point, only the OpenCL backend depends on the value of the backend property.

[rgerganov]

Thanks. That confirms my assumption.