To summarize the discussion from
https://discuss.scientific-python.org/t/should-scipy-tests-assume-float64-as-the-default-floating-point-dtype/1606/8 :

• in our testing with multiple array backends, we currently configure jax to default to float64 but keep torch defaulting to float32;
• the ability to run scipy tests with either f32 or f64 default dtype is important for some of us;
• we broadly agree that inconsistent defaults are not by design;
• there is a body of tests with handcrafted tolerances depending on whether the array backend is torch or something else.

The path forward---summarizing from and slightly extending the discourse thread above:

• parametrize the array API CI job into a pair: one job sets torch.set_default_dtype(torch.float64), the other does not
• add a helper (_lib feels like a right place?) to query the default, roughly _lib.default_fp_type(xp) which returns the xp-specific default dtype; do not add a publicly accessible backend-agnostic way to set the default.
• audit existing tests for backend-specific tolerances, atol=1e-7 if is_torch(xp) else 1e-15, change to if default_fp_type(xp) == xp.float32 where appropriate;
• once done, make float64 the default in conftest.py