QNN EP removes needed names from graph - leading to crash in QnnHtpPrepare

[john-dance]

Describe the issue

When running the model, there are initializers and NodeArgs that are removed because they are no longer used. For example:
Removing initializer 'onnx_Mul_818'. It is no longer used by any node.
Removing NodeArg 'Transpose_token_428_out0'. It is no longer used by any node.

However, the QNN EP seems to remove some needed inputs. This results in some error messages like this:
Input name not exist: network_0_network_0_0_mlp_fc1_Conv.weight_dq at location qnn_model_wrapper.cc:138 onnxruntime::qnn::QnnModelWrapper::CreateQnnInputOutputTensors
nput name not exist: network_0_network_0_1_mlp_fc1_Conv.weight_dq at location qnn_model_wrapper.cc:138 onnxruntime::qnn::QnnModelWrapper::CreateQnnInputOutputTensors

And that leads to a crash in QNNHtpPrepare.

If I run the same model on the CPU, it works.
If I run the same model with DirectML EP it works. I also noticed the following messages when using DML EP:
Add MemcpyFromHost after network_0_network_0_0_mlp_fc1_Conv.weight_dq for DmlExecutionProvider
Notice that this is one of the names that does not exist when running with the QNN EP.

To reproduce

Run the model found in this AI Hub job: https://app.aihub.qualcomm.com/jobs/jgdxeoy6p with the QNN EP.
(Notice that this will only be available to members of the MSFT QNN EP team.)

Urgency

No response

Platform

Windows

OS Version

Windows 11

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

1.19.2

ONNX Runtime API

C++

Architecture

ARM64

Execution Provider

Other / Unknown

Execution Provider Library Version

QNN EP

[ashumish-QCOM]

Hi @john-dance

Could you please provide the following details:

The specific model you are using (if different from the one linked).
Any modifications made to the model before running it with the QNN EP.
The complete log output when the error occurs.