GPU selection for Keras 3 with PyTorch DirectML

[TaiXeflar]

I'm here for asking GPU selection problems since I'm still new for Keras 3 when I jump into Keras 3 from PyTorch a while.

I have a Windows 11 Machine contains 2 cards with NVIDIA GeForce RTX 4070ti Super and AMD Radeon RX 7800XT respectively.
Now I have a independent Python 3.11.9 env installed torch-directml. In PyTorch we can manually use .to(device) to set GPU device, and can do check like this:

import torch_directml

# List all available DML devices
device_count = torch_directml.device_count()
print(f"Available DirectML devices: {device_count}")

# Loop over devices and print info
for i in range(device_count):
    dml_device = torch_directml.device(i)
    print(f"Device {i}: {dml_device}")

# Select device 1 (example)
dml = torch_directml.device(1)

Available DirectML devices: 2
Device 0: privateuseone:0
Device 1: privateuseone:1

From the above codes and the result, we can know that device 1 is the AMD card.
How can I manage/add preprocess "on DirectML device select" in Keras? If we just do nothing, the model will be run on CPU only.

Hardware and Software/Environ:

Intel Xeon W-3175X
NVIDIA GeForce RTX 4070ti Super
AMD Radeon RX 7800XT

Windows 11 24H2
Python 3.13.2 ---> PyTorch 2.6.0+cu126, Keras 3            # The NVIDIA CUDA one
Python 3.11.9 ---> PyTorch-DirectML, Keras 3                 # The DirectML for AMD card environment

[innat]

#21190

[TaiXeflar]

I just reviewed the post then doing test and have to say sorry where I find nothing can helps. Here's what happends:

First of all, I'm using Pytorch. keras.distribution not work where it returns None type under torch backend whether is torch+cuda or torch-directml. So keras.distribution.list_devices() not work on anywhere(My Pytorch deployment). The issue #21190 discovered for is TF2/JAX backends.

Second is I used like with keras.device("privateuseone:1"), it still can't run successfully. I uses Task Manager to watch GPU engines and just find out the model have no message hints and running on cpu.

[innat]

Its better to open a ticket. Keras team doesn't monitor the discussion much.