Found it, was a very dummy mistake. I did not install pykeops before running the code. Weird enough, if pykeops is not installed then keops kernels in gpytorch.kernels.keops fallback to non-keops kernels. The fallback in my case was happening silently, with no warning (somehow the warning generated here was suppressed).

I figured this out by inspecting source code. IMHO, I think that gpytorch.kernels.keops should raise some exception when it's used whitout pykeops installed.