DML cannot use device_id = 1 , run_with_iobinding failed.

[Jay19751103]

Describe the issue

The system have two graphic cards (Hybrid Graphics,Tested on A+A and I+N)
Use the olive tag 6ebb02a969077d57f5bb14420d85fcff3fa537f4 to run llama_v2
Model is 7B , FP16 (size around 12G~13G)
command
python run_llama_v2_io_binding.py --prompt="Hi how are you" --max_seq_len=100 --device_id=0 => system can run
python run_llama_v2_io_binding.py --prompt="Hi how are you" --max_seq_len=100 --device_id=1 => system cannot run

To reproduce

Copy FP16 llama2 onnx model to olive and run following will unexpected exit.
python run_llama_v2_io_binding.py --prompt="Hi how are you" --max_seq_len=100 --device_id=1

If we change the iobinding from dml to cpu , it can run, the result is wrong.
If we make a test patch to force the device id to 1 in onnxruntime. it can pass and result is okay. (please check the attached patch file)
0002-test_dev_id1.patch

Urgency

It's Urgent. This is for multi-gpu support general case.

#13545 last comment looks similar issues

Platform

Windows

OS Version

Windows 11 Pro OS build 22631.3593

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

4573740

ONNX Runtime API

Python

Architecture

X64

Execution Provider

DirectML

Execution Provider Library Version

1.18.0

[Jay19751103]

Hello
Any body can help on this issue ?

[Jay19751103]

Any one can help on this issue ?

[Jay19751103]

Hi
Anyone have suggestion for the io binding issue ?
I cannot use different GPU device if the application is not bind with a specified GPU.