[G720 / Yocto] Self-compiled DLA fails on NeuronRT with "Cannot create device for EDPA_1_2" (official DLA works)

NeuroPilot NeuroPilot - Analytical AI
, , ,
1

I’m trying to use the official SDK neuropilot-sdk-basic-8.0.11-build20260211.tar.gz to deploy YOLOv5 offline. After completing the export, convert, quantize, and ncc-tflite compilation processes for YOLOv5, I finally obtained a model with the .dla file extension. However, I encountered errors when testing this newly generated DLA model on my development board using NeuroRT. The errors are concentrated in libcmdl.so, EDPA_1_2, and APUSYS_2_0. To troubleshoot whether it’s a problem with the board’s environment, I tested the official YOLOv5 reference DLA file on the development board. However, the official DLA can infer normally, which may indicate that the board’s inference environment is working correctly. Then I suspected the problem might be due to the model being too complex, so I created a very simple TensorFlow model consisting only of Conv2D + ReLU. I first verified locally that TensorFlow and TFLite inference worked correctly. Then I exported TFLite, quantized it, and compiled it into a DLA. Running this extremely simple model on the development board still resulted in the same failure message. Based on these experiments, I suspect that the DLA generator I’m using is incompatible with the runtime or target-specific data on the G720 Yocto system’s development board. I want to confirm the officially recommended offline DLA process and whether ncc-tflite requires additional platform/arch/debug parameters for board compatibility.

The attachment for this question is here：np8_g720_issue_bundle_20260607_110922.zip

3

Hi,

This is an international forum, so please post questions in English. More people will understand your question and you’re more likely to get helpful answers.

tico

4

Thank you for pointing it out. Now, I have expressed it in English. Could you please help answer this question? Thank you!

5

Hi Huang_Hui,

Thanks for reaching out!

This issue is caused by using the wrong target during DLA compilation. On G720, you should compile against the EDMA target (not EDPA), which is why the runtime reports Cannot create device for EDPA_1_2.

Please re-compile your model with the following command:

ncc-tflite --arch=mdla5.3,edma3.6 <path/to/model.tflite> -o <path/to/model.dla> --relax-fp32 --opt=3

After recompiling with the correct target, NeuronRT on the board should be able to load and execute the .dla without the EDPA_1_2 / APUSYS_2_0 errors.

For more background and discussion, refer to this thread: Found an unsupported target: EDPA_1_2 when executing DLA on G720

Best,
Jun