Codecov Report

❌ Patch coverage is 86.66667% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.9%. Comparing base (95829c9) to head (5563cba).
⚠️ Report is 37 commits behind head on main.

@@           Coverage Diff           @@
##            main    #918     +/-   ##
=======================================
+ Coverage   85.6%   85.9%   +0.3%     
=======================================
  Files        151     151             
  Lines      16310   16524    +214     
=======================================
+ Hits       13967   14192    +225     
+ Misses      2343    2332     -11     

We'll need an example that reproduces the issue and which can be used in tests. If this is fixing some behaviour, a test should ensure that this is not reverted back in future.

We could, instead, refactor the upstream kernel functions to take either a scalar or array and do the right thing. This way you could keep the original vectorized implementation.

that also looks jittable?

We could, instead, refactor the upstream kernel functions to take either a scalar or array and do the right thing. This way you could keep the original vectorized implementation.

+1 on that.

Which kernel causes trouble? I am still waiting for the reproducer :)

@martinfleis , here is the explanation of the issue and how we can address it:

i searched and thought about possible failure and the failure occurs when a user passes a custom user-defined callable as the kernel parameter that contains scalar control flow (such as bounds checking for co-located points): (look at the following code below):

def safe_custom_kernel(distances, bw):
    # Prevents division by zero for co-located points
    if bw < 0.001: 
        bw = 0.001
    return np.exp(-distances / bw)

a test for this, could be, for example:

on testing with the vectorized apporach:

with iterative approach:

We could, instead, refactor the upstream kernel functions to take either a scalar or array and do the right thing. This way you could keep the original vectorized implementation.

was thinking the exact same..
(and, maybe i'm wrong too with scenario i have for user having a custom user-defined callable..)

If it is only the case of custom callable, I think we should document that it should expect array input and behave correctly.