[FIX] Don't reference pykeops.torch when torch is not installed

[flohorovicic]

Problem

gempy_engine/core/backend_tensor.py imports only pykeops.numpy at the top (guarded by is_pykeops_installed), but the wrapped pykeops functions referenced pykeops.torch.LazyTensor unconditionally in four places (_exp, _sum, _divide, _sqrt_fn).

pykeops only exposes its torch submodule when torch is importable, so running the numpy backend with pykeops enabled in an environment without torch crashed with:

AttributeError: module 'pykeops' has no attribute 'torch'

inside _sum (reached from _kernels_assembler._compute_distances_generic). Reproduced with gempy_engine 2026.0.3 running the Weisweiler model on macOS CPU (numpy backend + DEFAULT_PYKEOPS=True).

Fix

Build a module-level _LAZY_TENSOR_TYPES tuple at import time that always contains pykeops.numpy.LazyTensor and adds pykeops.torch.LazyTensor only when import pykeops.torch succeeds. All four type checks now use that tuple; the redundant import pykeops statements inside the wrapped functions were removed.

Verification

• In a venv with numpy + pykeops and no torch: on main, a script enabling the pykeops numpy backend and calling the four wrapped functions with a pykeops.numpy.LazyTensor-typed object reproduces the exact AttributeError; on this branch all four functions dispatch correctly and plain numpy arrays still take the numpy path.
• In an environment with torch + pykeops: _LAZY_TENSOR_TYPES contains both the numpy and torch LazyTensor classes, so behavior is unchanged.

🤖 Generated with Claude Code

[Leguark]

this was trying to use an environment Pykeops + numpy withouth torch or in an environment only with numpy did not work either?

[flohorovicic]

Only the first one. An environment with just numpy (no pykeops) works fine both before and after this PR — the pykeops import at module level is guarded by is_pykeops_installed, and _wrap_pykeops_functions() is only called when pykeops is enabled, so the bad pykeops.torch references were never reached. I re-verified this in a numpy-only venv against main: the numpy backend works, and requesting pykeops there raises the existing clear "library is not installed: pykeops" error.

The crash was specific to: pykeops installed + pykeops enabled (numpy backend) + torch not installed. In that case pykeops loads but never creates its torch submodule, so the unconditional pykeops.torch.LazyTensor checks raised AttributeError as soon as a wrapped function ran.

[Leguark]

okay good. So I do not have to hotfix the pip release. Are you able to run Keops on MacOS? I thought it is Linux only