====================================================
DirectML PReLU Native Crash Report
====================================================

1. Environment
----------------
Python: 3.11.9
torch: 2.4.1+cpu
torchvision: 0.19.1+cpu
torch-directml: 0.2.5.dev240914

Windows:
(Build)

GPU:
(AMD...)

Driver:
(xx.xx.xx)

====================================================
Problem
====================================================

Expected:

PReLU(num_parameters=C) should execute correctly
for arbitrary channel counts.

Observed:

num_parameters=1  -> PASS

num_parameters=2  -> Native crash

====================================================
Crash
====================================================

[F810 ...]
Check failed:
rank <= DML_TENSOR_DIMENSION_COUNT_MAX

====================================================
Reproducer
====================================================

Short code

...

====================================================
Isolation Process
====================================================

The issue was reproduced

✔ fresh Python 3.11 installation

✔ fresh virtual environment

✔ independent of Real-ESRGAN

✔ independent of instrumentation framework

✔ CPU passes all tests

✔ ReLU passes

✔ LeakyReLU passes

✔ SiLU passes

✔ Sigmoid passes

✔ Tanh passes

✔ PReLU(num_parameters=1) passes

✔ PReLU(num_parameters>1) crashes

====================================================
Question
====================================================

Is this a known limitation or bug in the current
torch-directml implementation of PReLU?