Why Does Ncc-TFLite(TFLite-to-DLA) Conversion Fail Without Error Log on Genio (Neuron6) Platforms?

NeuroPilot
, , ,
1

When attempting to convert a TensorFlow Lite model to DLA format on a Genio-510/700/1200 (Neuron6) platform using the following command,


ncc-tflite xxxx.tflite -arch mdla3.0 -d xxxx.dla --relax-fp32 --verbose

the tool fails with only:


ERROR: Fail to create tflite model

No additional error details or debug logs are shown, making it unclear how to troubleshoot or identify unsupported operations.

How can developers address this lack of error log, and what is the recommended workflow for debugging or solving TFLite-to-DLA conversion failures on these platforms?

2

Yocto Platform (NP6SDK)

  • The NP6 SDK used for DLA model conversion on Genio-510/700/1200 (Neuron6) platforms is no longer maintained. As a result, the ncc-tflite tool often provides minimal or no error logs when failures occur.
  • As of now, developers must continue using the NP6 SDK to generate DLA models for these platforms, since the compatible Neuron library upgrade package is still under development by MediaTek and not yet released.
  • If the ncc-tflite tool fails without detailed output, it is recommended to validate model compatibility and identify unsupported operators by executing:
    benchmark_model --stable_delegate_settings_file=/usr/share/label_image/stable_delegate_settings.json --use_nnapi=false --use_xnnpack=false --use_gpu=false --min_secs=20 --graph=your/model/path.tflite
    • The resulting log will indicate unsupported operator types (e.g., “ERROR: OP RELU_0_TO_1 (v1) is not supported”).
  • MediaTek recommends monitoring official announcements for future library releases that will enable migration to Neuron8 tools and improved debugging support.

Android Platform (NP8SDK)

  • For Genio-510/700/1200 (Neuron6) platforms running Android, MediaTek recommends upgrading device runtime libraries to Neuron8 and using the NP8 SDK for DLA compilation.

  • The NP8 toolchain provides detailed error logging and improved compatibility when compiling models for these devices.

  • Note: DLA models compiled with NP8 SDK require the device’s runtime to be upgraded to Neuron8; otherwise, runtime version conflicts may occur.

  • On-device model compilation is no longer supported for these platforms; all conversions and builds should be done with the host-side toolchain.