There are 2 issues at hand here. Firstly, your equality check is going to fail because you’re comparing an s64 tensor to an f32 tensor, and those will always be different Elixir-wise.
Nx unit tests use an assert_equal helper that you can yank, or better yet, assert all close.
Now for the precision issue. You’re dealing with floating point numbers, and the first difference I spot is that in Julia you used f64, while in Nx you used f32. That by itself is already a source of significant difference. Aside that, Julia might be doing some rounding in the results, especially in the underlying mechanisms of the linear system solution (e.g zero-rounding in PLU or QR decomposition before doing triangular solves)
