[BUG] Inconsistent behaviour of the floot_divide integer arrays truncates towards 0

[aaishwarymishra]

☑️ I understand it is strictly prohibited to use AI to write issues.

Describe the bug
The mx.floor_divide is inconsistent , it behaves different based on integer and float dtypes

To Reproduce

Include code snippet

x = mx.array(1, dtype=mx.float32)
y = mx.array(-2, dtype=mx.float32)

z = mx.floor_divide(x, y)

np_x = np.array(1, dtype=np.float32)
np_y = np.array(-2, dtype=np.float32)
np_z = np.floor_divide(np_x, np_y)

print(z)
print(np_z)

outputs:

array(-1, dtype=float32)
-1.0

when changed to int32

x = mx.array(1, dtype=mx.int32)
y = mx.array(-2, dtype=mx.int32)

z = mx.floor_divide(x, y)

np_x = np.array(1, dtype=np.int32)
np_y = np.array(-2, dtype=np.int32)
np_z = np.floor_divide(np_x, np_y)

print(z)
print(np_z)

outputs:

array(0, dtype=int32)
-1

Expected behavior
A clear and concise description of what you expected to happen.

The expected behaviour is that the integer arrays should behave like float arrays instead of truncating towards 0.

Desktop (please complete the following information):

• OS Version: macos tahoe 26.6.1
• Version: 0.32.3.dev20260902+117188cd

Additional context
Discovered while working on data-apis/array-api-compat#451

[Ved235]

This is happening becuase in floor_divide(), inputs of subdtype inexact are routed to the correct floor(divide()) operation combo. Whereas, integers go through the following logic:

  auto inputs = broadcast_arrays({astype(a, dtype, s), astype(b, dtype, s)}, s);
  auto shape = inputs[0].shape();
  return array(
      shape, dtype, std::make_shared<Divide>(to_stream(s)), std::move(inputs));

This is just normal division. To fix it I believe we can do:

  auto q = array(
      shape, dtype, std::make_shared<Divide>(to_stream(s)), std::move(inputs));
  if (!issubdtype(dtype, signedinteger)) {
    return q;
  }

  auto r = remainder(a, b, s);
  auto zero = array(0, dtype);
  auto round_down = logical_and(
      not_equal(r, zero, s),
      not_equal(less(a, zero, s), less(b, zero, s), s),
      s);
  return subtract(q, astype(round_down, dtype, s), s);

But this might have a performance impact (we could use divmod() also... it should be more performant). I can check later how pytorch implements this.

[zcbenz]

There is a fix in #4346.

[aaishwarymishra]

@ayaangazali Hi are you still working on it ? :)

[aaishwarymishra]

@zcbenz will it be fine if I make a PR? Author of the previous PR has not responded for some time.

[zcbenz]

I'm good with a new PR.