vai_c_xir PyTorch compilation fails with XCOM_ITEM_UNDEFINED

[afzalxo]

I am using U50 with Vitis AI 1.4.1.978 and DPUCAHX8L_ISA0 to run and evaluate thousands of custom convolutional models to construct a large dataset of convnet architectures along with their performance on board.

For some models, when I try to compile the quantized model using vai_c_xir, the compilation fails with

[UNILOG][INFO] Begin to compile...
[UNILOG][FATAL][XCOM_ITEM_UNDEFINED][The requested item is not finded or defined!]
*** Check failure stack trace: ***
Aborted (core dumped)

A sample quantized xmodel file is here: custom_convnet.xmodel.zip.

Help would be appreciated. Thanks

[afzalxo]

I have identified many different models that are valid, can be trained and quantized, but do not compile using vai_c_xir due to XCOM_ITEM_UNDEFINED error. I cannot however make out a pattern; different expansion factors in MobileNet-like models, and different number of layers per block also generate this result. For some expansion factors/number of layers, things work fine, but fail for others without a proper justification of failures. The compiler should at least throw a decipherable error message, which is not the case here.

[afzalxo]

A few updates:
I compiled the same quantized model using ZCU102 as the target for DPUCZDX8G and Vitis AI 2.5 and the compilation goes smoothly. The error only appears when compiling for DPU DPUCAHX8L utilized by U50 (Can't use DPUCAHX8H since I'm using depthwise conv). The operations contained in the model are all supported by both DPUs. Hence the problem is likely on the compilers end vai_c_xir.

[SuZhan0322]

Hi @afzalxo , Is your model a pytorch model? And the vai_c_xir only supports the compilation of pytorch models. Thanks!