As mentioned in#29315 (see#29315 (comment)),numpy.quantile currently can overflow for integer inputs in spite of returning a float:

>>>np.quantile(np.array([32767,-1],dtype=np.int16),0.5)np.float64(49151.0)

This can be fixed by casting integers to floats, before interpolating.
This yields the correct result:

>>>np.quantile(np.array([32767,-1],dtype=np.int16),0.5)np.float64(16383.0)

We ensure a consistent behavior by adding a property test:

quantile=np.quantile(a,q)assertnp.min(a)<=quantileassertquantile<=np.max(a)

So far, this is just a quick fix, I haven't check if all paths are correct.

We probably need to parameterize by the additional parameters ofnp.quantile.
I just wanted to first cross-check if this is the right approach to fix the issue.

At this point, we still incorrectly get:

>>>np.quantile([1],0)np.int64(1)

While the documentation ensures us that we should getnp.float64(0.0).
Currently, the return type depends on the value ofq:

>>>np.quantile([1],0.5)np.float64(1.0)

which is undesired behavior.
I can add the return type to the property test.
We also haven't looked into the second part of the documentation: "If the input contains ... floats smaller than float64, the output data-type is float64."

Another point worth discussing is usingnp.int64 with values bigger than can be exactly represented bynp.float64 (i.e. integers larger than 2⁵³).
I am not quite sure, if we might lose accuracy in this case.
On the other hand, it is clearly documented that we catnp.float64, so we should expect a loss of accuracy for integers outside the representable range.