[Demirrr]

Found while testing #658: on a machine with a GPU, 6 of the 7 tests/test_nero_simple.py tests failed on unmodified develop with RuntimeError: Expected all tensors to be on the same device, but got index is on cpu, different from other tensors on cuda:0. CI has no GPU, so it never showed up.

• The model is moved to self.device for training, but 8 places built the index tensors as torch.LongTensor([[...]]) on the CPU (search, fit, fit_from_iterable, the two embedding getters). They now use a new NERO._index_tensor that creates them on self.device.
• The public forward moves its inputs to self.device.
• sort_idxs.detach().numpy() (3 places) would have raised TypeError on a CUDA tensor right after the first fix; now .detach().cpu().numpy().
• Regression test that runs on CPU-only CI: it gives the learner a meta device and checks that the index tensors follow it. It fails if the CPU is hard-coded again.

Local: test_nero_simple 8 passed with CUDA available and 8 passed with CUDA_VISIBLE_DEVICES=""; ruff clean.

🤖 Generated with Claude Code