QCS6490 (Radxa Q6A) – QAIRT SDK Version Mismatch: AI Hub Artifacts (2.45.x) Incompatible with Runtime 2.42.0.251225

Dragon Series Q6A
1

We are currently working on AI model deployment on the Qualcomm QCS6490 platform (Radxa Dragon Q6A) using Qualcomm AI Runtime (QAIRT/QNN) targeting HTP/NPU execution. During deployment and validation, we identified a significant SDK compatibility issue that is preventing successful execution of Qualcomm AI Hub generated artifacts on the target hardware.

Target Platform Information

Hardware Platform: Qualcomm QCS6490 / Radxa Dragon Q6A
Target Runtime: Qualcomm AI Runtime (QAIRT/QNN) on HTP/NPU
Officially Supported SDK Version: QAIRT 2.42.0.251225
Host Environment: Ubuntu 24.04 x86_64, Python 3.10

Issue Summary

We successfully used Qualcomm AI Hub to export and compile AI models (for example, YOLOv8 and Real-ESRGAN) targeting:

qnn_context_binary
w8a8 quantization
qualcomm-qcs6490 chipset

The export, quantization, linking, profiling, and hosted inference validation on Qualcomm AI Hub completed successfully.

AI Hub profiling confirmed NPU execution on a hosted QCS6490 device:

Real-ESRGAN x4plus: npu (1096 ops), cpu (0), gpu (0)
YOLOv8 detection: npu (254 ops), cpu (0), gpu (0)

Generated artifacts were successfully downloaded, including .bin context binaries.

Primary Issue – QAIRT SDK Runtime Compatibility

The generated model metadata indicates that Qualcomm AI Hub currently compiles artifacts using:

QAIRT/QNN SDK version: 2.45.x

However, the target Radxa Dragon Q6A environment officially supports:

QAIRT SDK version: 2.42.0.251225

When attempting local deployment using qnn-net-run, execution fails with the following runtime error:

Initializing HtpProvider
Creating context from binary file: yolov8_det.bin

[ERROR] Using newer context binary on old SDK
[ERROR] Fail to get context blob with err 5000
[ERROR] Failed to create context from binary with err 0x1388
[ERROR] Could not create context from binary
Create From Binary failure
Technical Observation

This behavior suggests a runtime/compiler compatibility issue:

QNN/QAIRT serialized HTP context binaries appear to be version-sensitive.
Artifacts generated using QAIRT 2.45.x do not execute on QAIRT 2.42.0.251225 runtime.
Matching compiler/runtime versions may be required for successful deployment.

2

@Morgan do you know if we can download older QAIRT SDK?

3

hi, @vijay

[ERROR] Using newer context binary on old SDK

The new QNN context binary cannot run in older QAIRT runtime environments. You can download the same version of the QAIRT SDK with the version of context binary.

when you use the AI hub to export the model, you can specify QNN version,
please have a look this link:

best,
Morgan

4

Hi, my github repo with QAIRT Model Quantization Toolkit

5

here’s my repo. let me know when if you’ve successfully ran the inference or not. I have another project that for running the inference , written with c++ entirely for optimization. will push that one too.