This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

TDA4VM: The trouble at "TVM/Neo-AI-DLR + TIDL Heterogeneous Execution"

Part Number: TDA4VM

Hi

I'm facing the trouble at "TVM/Neo-AI-DLR + TIDL Heterogeneous Execution".

I' m referring at the "TIDL specific lines in compilation script" in

software-dl.ti.com/.../md_tidl_osr_tvm_tidl.html

and also referring "est_tidl_j7.py".

The problem is that I cannot compile the Pytorch model directly (i.e. without Onnx) for EVM.

But, it's ended with error as below.

It's happened only when I try to compile  with a large number of images (up to 100) used for quantization calibration.

The model I use is same with one on e2e.ti.com/.../977560.

Thanks and Best regards,

HaTa

  • Hi HaTa,

        I do not have the model, compilation script and calibration images to reproduce the problem.

        Could this be a memory problem?  i.e. the calibration process ran out of memory?  When you say "It's happened only when I try to compile  with a large number of images (up to 100) used for quantization calibration."  Does it mean that it succeeds with smaller number of calibration images, say 50?

        More information would help trouble shoot the issue.  Thanks!

    -Yuan

  • Hi Yuan

    I send the file via BOX. Thanks!

    Thanks and Best regards,

    HaTa

  • Hi HaTa,

        I tried 100 calibration images.  The compilation finished calibration on my Linux box without any problem.  My Linux box has 16GB memory.  I suspect it is a memory problem: you are running out of memory on your Linux box.

        I did see that the compilation failed in the perfsim tool, which you reported in another e2e thread.  Anshu will be looking into that issue.

    -Yuan

  • Hi Yuan

    Thanks a lot!

    I'll confirm to customer if their Linux machine have enough memory.