CUDA Mod kernel has no zero divisor check, unlike CPU after #27833

@tianleiwu is already working on this.

Since Sep 10, 2026.

• #32546 by @copilot-swe-agent — open

### Describe the issue

The CUDA `Mod` kernel has no zero divisor check, so an integer `Mod` with a zero in the divisor is undefined behaviour there. The CPU kernel got a check in #27833 and the two now disagree.

`_Mod` in `onnxruntime/core/providers/cuda/cu_inc/common.cuh`, line 503:

```cpp
template
__device__ __inline__ T _Mod(T a, T b) {
T r = a % b;
T zero = T(0);
if ((r > zero && b < zero) || (r < zero && b > zero)) {
r += b;
}
return r;
}
```

`a % b` runs with no guard on `b`. `_Fmod` seven lines below has the same shape.

On CPU the same input is rejected. `CheckZeroDivisorImpl` in `onnxruntime/core/providers/cpu/math/element_wise_ops.cc` does `ORT_RETURN_IF(b_data[i] == T{0}, "Integer modulo by zero")` before the kernel runs, which came from #27833.

So the same model errors cleanly on CPU and hits undefined behaviour on CUDA.

### To reproduce

Run an integer `Mod` node where the divisor tensor contains a zero, on the CUDA EP. For example `Mod(int64[3] {-3, 4, 7}, int64[3] {0, 2, 3})` with `fmod=0`.

I have not run this. I do not have NVIDIA hardware and cannot build the CUDA EP, so this is a read of the source rather than a measured crash, and I would not want it treated as a verified repro. Filing it because the asymmetry with #27833 looked worth recording. Happy to close if it turns out the divisor is already validated somewhere upstream of the kernel.

### Urgency

Low.

### Platform

Linux

### OS Version

n/a, reported from source

### ONNX Runtime Installation

Built from Source

### ONNX Runtime Version or Commit ID

main, `common.cuh` line 503 as of 2026-09-09

### ONNX Runtime API

Python

### Architecture

X64

### Execution Provider

CUDA