`jnp.floor_divide()` returns incorrect results when used with `scan`

import jax
from jax import lax
import jax.numpy as jnp

x = jnp.array([2, 2, -3])

def reduce_for_loop(func, arr):
  val = arr[0]
  for i in range(1, len(arr)):
    val = func(val, arr[i])
  return val

def reduce_scan(func, arr):
  val = arr[0]
  body_func = lambda i, val: func(val, arr[i])
  return lax.fori_loop(1, len(arr), body_func, val)

print(reduce_for_loop(jnp.floor_divide, x)) # prints -1
print(reduce_scan(jnp.floor_divide, x))     # prints 0
with jax.disable_jit():
  print(reduce_scan(jnp.floor_divide, x))   # prints -1

Came up when working on #9529

[chsigg]

Fixed with https://github.com/tensorflow/tensorflow/commit/6d6d7fece1752523699993d7a30ee7790e84adba. I will leave it to Jake to close this issue.

[jakevdp]

Looks like this is fixed in the jaxlib 0.3.14 release