[BUG] mx.divmod for floats truncates the quotient — disagrees with mx.floor_divide and breaks q·b+r == a

Describe the bug

This bug is found from property based testing.

For floating-point inputs, mx.divmod(a, b) returns a quotient that truncates
toward zero, while mx.floor_divide(a, b) correctly floors toward −∞
(matching NumPy). So for any a/b < 0 the two operations return different
quotients, and divmod's (q, r) no longer reconstruct the dividend:

q * b + r == a      # the universal divmod invariant — BROKEN for floats

To Reproduce

import mlx.core as mx
import numpy as np

a, b = np.float32(-7.0), np.float32(2.0)
np.array(mx.floor_divide(mx.array(a), mx.array(b)))   # -4.0  (correct, floors)
q, r = mx.divmod(mx.array(a), mx.array(b))
np.array(q), np.array(r)                                # (-3.0, 1.0)  WRONG
# reconstruction:
(-3.0) * 2.0 + 1.0                                      # -5.0  != -7.0

Expected behavior

>>> a, b = np.float32(-7.0), np.float32(2.0)
>>> q, r = mx.divmod(mx.array(a), mx.array(b))
>>> np.array(q).item(), np.array(r).item()
(-3.0, 1.0)            # actual — wrong
>>> # expected (NumPy parity, and consistency with mx.floor_divide):
>>> np.floor_divide(a, b), np.remainder(a, b)
(-4.0, 1.0)            # q=-4, r=1; check: -4*2 + 1 = -7 ✓
>>> np.array(mx.floor_divide(mx.array(a), mx.array(b))).item()
-4.0                   # mx.floor_divide is already correct for floats

FYI

      a/b  floor_div  divmod_q  divmod_r    q*b+r   want  np_floor
-7.0/ 2.0      -4.00     -3.00      1.00    -5.00   -7.0     -4.00  <- BUG
 7.0/-2.0      -4.00     -3.00     -1.00     5.00    7.0     -4.00  <- BUG
-5.5/ 2.0      -3.00     -2.00      0.50    -3.50   -5.5     -3.00  <- BUG
 5.5/-2.0      -3.00     -2.00     -0.50     3.50    5.5     -3.00  <- BUG
-9.0/ 4.0      -3.00     -2.00      3.00    -5.00   -9.0     -3.00  <- BUG

floor_divide(-7.0,2.0) = -4.0   divmod(-7.0,2.0).q = -3.0   agree? False

fix: make divmod floor the quotient and match numpy

Reproduced on 0.32.0 and on main (a1e0e0b).

Two things not in the report:

Integers are affected too. mx.divmod(-7, 2) returns (-3, 1) on GPU and (-3, -1) on CPU. numpy gives (-4, 1).
So CPU and GPU return different remainders for the same input.

Cause: DivMod pairs a truncating quotient with a floored remainder. In binary_ops.h, FloorDivide does trunc(x/y) while Remainder floors. The CPU path uses std::trunc and std::fmod, which is self consistent but does not match numpy.

test_divmod misses this because it only uses uniform(1, 100) and only checks out[0].

I have a fix locally: q = floor(x/y), r = x - q*y, plus tests with negative cases. FloorDivide is only used by DivMod, so nothing else changes.

I do not have a Metal compiler on this machine, so I verified the CPU backend only. Want me to open a PR?

Navigation Menu

Repository navigation

[BUG] `mx.divmod` for floats truncates the quotient — disagrees with `mx.floor_divide` and breaks `q·b+r == a` #4119

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[BUG] mx.divmod for floats truncates the quotient — disagrees with mx.floor_divide and breaks q·b+r == a #4119

Description

Activity

krrishapatel commented on Aug 16, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

[BUG] `mx.divmod` for floats truncates the quotient — disagrees with `mx.floor_divide` and breaks `q·b+r == a` #4119