Native DirectML crash with torch.nn.PReLU(num_parameters > 1) (dml_tensor_desc.cc:80)

Hi everyone,

I've isolated what appears to be a reproducible DirectML backend issue involving channel-wise PReLU.

Environment

Minimal behaviour

CPU:

• ReLU ✔
• LeakyReLU ✔
• SiLU ✔
• Sigmoid ✔
• Tanh ✔
• PReLU(num_parameters=1) ✔
• PReLU(num_parameters=2) ✔

DirectML:

• ReLU ✔
• LeakyReLU ✔
• SiLU ✔
• Sigmoid ✔
• Tanh ✔
• PReLU(num_parameters=1) ✔
• PReLU(num_parameters=2) ❌ Native crash

The process aborts with

The crash happens before Python can raise an exception.

Isolation steps

I reproduced this

• in a completely fresh Python 3.11 virtual environment,
• without Real-ESRGAN,
• without my instrumentation framework,
• using a standalone minimal reproducer.

The crash consistently occurs when switching from a single PReLU parameter to channel-wise PReLU (num_parameters > 1).

Has anyone seen this before?

Is this a known limitation or a regression in torch-directml?

Any insight would be greatly appreciated.Hi everyone,

I've isolated what appears to be a reproducible DirectML backend issue involving channel-wise PReLU.

Environment

Minimal behaviour

CPU:

• ReLU ✔
• LeakyReLU ✔
• SiLU ✔
• Sigmoid ✔
• Tanh ✔
• PReLU(num_parameters=1) ✔
• PReLU(num_parameters=2) ✔

DirectML:

• ReLU ✔
• LeakyReLU ✔
• SiLU ✔
• Sigmoid ✔
• Tanh ✔
• PReLU(num_parameters=1) ✔
• PReLU(num_parameters=2) ❌ Native crash

The process aborts with

The crash happens before Python can raise an exception.

Isolation steps

I reproduced this

• in a completely fresh Python 3.11 virtual environment,
• without Real-ESRGAN,
• without my instrumentation framework,
• using a standalone minimal reproducer.

The crash consistently occurs when switching from a single PReLU parameter to channel-wise PReLU (num_parameters > 1).

Has anyone seen this before?

Is this a known limitation or a regression in torch-directml?

Any insight would be greatly appreciated.DirectML_PReLU_Bug_Report.txt