Running dlc models? Invalid ELF header on InceptionV3 on 6490

[janjongboom]

According to the docs inception_v3 should run on 6490... I grabbed it from AI Hub (https://aihub.qualcomm.com/models/inception_v3?searchTerm=inception) in w8a8 and float format; then try to use it like this:

import os
from qai_appbuilder import (QNNContext, Runtime, LogLevel, ProfilingLevel, PerfProfile, QNNConfig)

curr_dir = os.path.dirname(os.path.realpath(__file__))
# Download w8a8 Qualcomm AI Runtime binary from https://aihub.qualcomm.com/models/inception_v3?searchTerm=inception
MODEL_FILE = os.path.join(curr_dir, 'inception_v3-inception-v3-w8a8.dlc')

# InceptionV3 class which inherited from the class QNNContext.
class InceptionV3(QNNContext):
    def Inference(self, input_data):
        input_datas=[input_data]
        output_data = super().Inference(input_datas)[0]
        return output_data

# Config AppBuilder environment (/usr/lib contains all my .so files; and I run other models (eg TFLite w/ QNN delegates; or LLMs w/ the QAI AppBuilder Genie bindings) fine)
QNNConfig.Config('/usr/lib', Runtime.HTP, LogLevel.WARN, ProfilingLevel.BASIC)

# Instance for InceptionV3 objects.
inceptionV3 = InceptionV3("inceptionV3", MODEL_FILE)

This yields:

     0.4ms [ ERROR ] Unable to load model. pal::dynamicloading::dlError(): /home/ubuntu/context-binary-demo/inception_v3-inception-v3-w8a8.dlc: invalid ELF header
     0.4ms [ ERROR ] Error initializing QNN Function Pointers: could not load model: /home/ubuntu/context-binary-demo/inception_v3-inception-v3-w8a8.dlc
Error initializing QNN Function Pointers: could not load model: /home/ubuntu/context-binary-demo/inception_v3-inception-v3-w8a8.dlc

Any idea?

EfficientNetB0 has the same issue (also with qnn-net-run):

(.venv) ubuntu@ubuntu:~/context-binary-demo$ qnn-net-run --backend libQnnHtp.so --model efficientnet_b0-efficientnet-b0-w8a16.dlc
qnn-net-run pid:7300
qnn-net-run build version: v2.35.0.250530123435_121478
     0.6ms [ ERROR ] Unable to load model. dlerror(): efficientnet_b0-efficientnet-b0-w8a16.dlc: invalid ELF header
     0.6ms [ ERROR ] Failed to load ModelLib!
     0.6ms [ ERROR ] Could not load model : efficientnet_b0-efficientnet-b0-w8a16.dlc!
Initialization failure

dlc looks like a zip file:

(.venv) ubuntu@ubuntu:~/context-binary-demo$ file efficientnet_b0-efficientnet-b0-w8a16.dlc
efficientnet_b0-efficientnet-b0-w8a16.dlc: Zip archive data, at least v4.5 to extract, compression method=store

And it extracts (but throws some CRC errors).

(.venv) ubuntu@ubuntu:~/context-binary-demo$ unzip -d inception inception_v3-inception-v3-w8a8.dlc
Archive:  inception_v3-inception-v3-w8a8.dlc
 extracting: inception/dlc.metadata2.1.0   bad CRC 05688358  (should be 00000000)
 extracting: inception/model          bad CRC 0f32a6ff  (should be 00000000)
 extracting: inception/model.params   bad CRC c8758672  (should be 00000000)
 extracting: inception/model.params.bin   bad CRC b2f93d75  (should be 00000000)

I guess I need a bin file? From reading the example, but how can I get this bin file? The model.params.bin file at least does not work.

ChatGPT suggested this:

(.venv) ubuntu@ubuntu:~/context-binary-demo$ qnn-net-run --backend /usr/lib/libQnnHtp.so --model /usr/lib/libQnnModelDlc.so --dlc_path $PWD/inception_v3-inception-v3-w8a8.dlc
qnn-net-run pid:8095
qnn-net-run build version: v2.35.0.250530123435_121478
 <W> Initializing HtpProvider
qnn-net-run log level is : QNN_LOG_LEVEL_ERROR
/prj/qct/webtech_scratch20/mlg_user_admin/qaisw_source_repo/rel/qairt-2.35.0/release/snpe_src/avante-tools/prebuilt/dsp/hexagon-sdk-5.4.0/ipc/fastrpc/rpcmem/src/rpcmem_android.c:38:dummy call to rpcmem_init, rpcmem APIs will be used from libxdsprpc
Composing Graphs
   162.6ms [ ERROR ] Failed to create dlc handle with code 1002 for dlc file at /home/ubuntu/context-binary-demo/inception_v3-inception-v3-w8a8.dlc.
   162.6ms [ ERROR ] ComposeGraphs Failed with error = 1
Graph Prepare failure
 <W> Backend 1 free cleanup called during process exit
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
/prj/qct/webtech_scratch20/mlg_user_admin/qaisw_source_repo/rel/qairt-2.35.0/release/snpe_src/avante-tools/prebuilt/dsp/hexagon-sdk-5.4.0/ipc/fastrpc/rpcmem/src/rpcmem_android.c:42:dummy call to rpcmem_deinit, rpcmem APIs will be used from libxdsprpc

^ Also fails.

[janjongboom]

Guess i need https://github.com/quic/ai-engine-direct-helper/tree/main/tools/convert/dlc2bin - but what are the configs for 6490?

Tried to guess one, but same error as before:

(.venv) ubuntu@ubuntu:~/context-binary-demo$ qnn-context-binary-generator --backend /usr/lib/libQnnHtp.so --model /usr/lib/libQnnModelDlc.so --dlc_path  ./inception_v3-inception-v3-w8a8.dlc --binary_file "inception_v3-inception-v3-w8a8" --output_dir output --config_file config_local.json
qnn-context-binary-generator pid:9056
 <W> Initializing HtpProvider
qnn-net-run log level is : QNN_LOG_LEVEL_ERROR
/prj/qct/webtech_scratch20/mlg_user_admin/qaisw_source_repo/rel/qairt-2.35.0/release/snpe_src/avante-tools/prebuilt/dsp/hexagon-sdk-5.4.0/ipc/fastrpc/rpcmem/src/rpcmem_android.c:38:dummy call to rpcmem_init, rpcmem APIs will be used from libxdsprpc
   165.5ms [ ERROR ] Failed to create dlc handle with code 1002 for dlc file at ./inception_v3-inception-v3-w8a8.dlc.
   165.6ms [ ERROR ] ComposeGraphs Failed with error = 1
Graph Compose failure
 <W> Backend 1 free cleanup called during process exit
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
 <E> exits with 2147483647, undefined m_mutex handle object
/prj/qct/webtech_scratch20/mlg_user_admin/qaisw_source_repo/rel/qairt-2.35.0/release/snpe_src/avante-tools/prebuilt/dsp/hexagon-sdk-5.4.0/ipc/fastrpc/rpcmem/src/rpcmem_android.c:42:dummy call to rpcmem_deinit, rpcmem APIs will be used from libxdsprpc

edit: I guess this was created by QAIRT 2.37; but we run QNN 2.35 (which is why this won't work I guess):

(.venv) ubuntu@ubuntu:~/context-binary-demo$ strings ./inception_v3-inception-v3-w8a8.dlc | egrep -i 'qnn|qairt|export|version' | head
        "header_version": {
        "version": {
                "tool": "qairt-converter",
                "converterVersion": "2.37.0.250724175447_124859",
                    "disable_qnn_op_config_validation": "False",
                    "dump_exported_onnx": "False",
                    "export_format": "DLC_DEFAULT",
                    "model_version": "None",
                "tool": "qairt-converter",
                "converterVersion": "2.37.0.250724175447_124859",

[janjongboom]

According to AI Hub team the dlc->bin can be done on device; it would be great if this would just happen transparently in QAI AppBuilder when you pass in a dlc file w/o having to figure out any local conversion.

[quic-zhanweiw]

Thanks @janjongboom for your question.
QAI AppBuilder can't load DLC model directly currently. We'll try to support this.

We have some doc on how to convert DLC to bin:
https://github.com/quic/ai-engine-direct-helper/tree/main/tools/convert/dlc2bin