Deploy autotuned network as bundle on ARM fails

[MaxS1996]

I created a simple MNIST test case to evaluate the performance of different DL Compilers on an Raspberry Pi 4. I was able to build a standalone bundle by modifying the bundle example at /apps/bundle.
But this fails, if I generated the model.o, params.bin and graph.json with autoTVM.
The binary still gets created, but fails when I try to run it on the device

The error message is (6 different versions):
function handle for fused_... not found

• transpose_multiply_round_clip_cast_layout_transform
• nn_contrib_conv2d_NCHWc_add_right_shift_clip_cast_1
• cast_multiply_add_nn_relu_1
• nn_contrib_conv2d_NCHWc_add_right_shift_clip_cast
• nn_max_pool2d_multiply_round_clip_cast
• cast_multiply_add_nn_relu

I guess, these functions have been created by the autontuning step, but somehow do not end up in the executable. How can I add them?

[tqchen]

Thanks for reporting the problem, the community uses https://discuss.tvm.ai/ for related trouble shooting questions, please open a new thread there and provide a few mode steps.

Please also consider to look at https://github.com/apache/incubator-tvm/tree/master/apps/benchmark for reference examples.

[MaxS1996]

https://discuss.tvm.ai/t/autotvm-or-quantization-for-raspberry-pi-fails/7319

I created a topic there to afterwards, as I am not sure if its a bug or just a user error