[Issue]: `concat_op.axis >= 0` assertion violates onnx spec

[shssoichiro]

Problem Description

When compiling certain onnx models via migraphx, an assertion is thrown from this line. I reproduced this on multiple of the SwinIR models from the vs-mlrt repo. However, the onnx spec states that valid axis values are "[-r, r-1] where r = rank(inputs)". Therefore, this assert violates the onnx spec, and the code needs to be updated to support negative axis values.

Operating System

CachyOS Linux 6.18.6

CPU

AMD Ryzen 9800x3d

GPU

AMD Radeon RX 7900 XTX

Other

No response

ROCm Version

ROCm 6.0.0

Steps to Reproduce

• Download the onnx-formatted SwinIR models from vs-mlrt
• Run the migraphx compilation command, e.g. migraphx-driver compile --onnx /usr/lib/vapoursynth/models/swinir/003_realSR_BSRGAN_DFO_s64w8_SwinIR-M_x2_GAN.onnx --gpu --optimize --binary --output /usr/lib/vapoursynth/models/swinir/003_realSR_BSRGAN_DFO_s64w8_SwinIR-M_x2_GAN.onnx.1920x1080_fast_migx-2.14.0_AMD-Radeon-RX-7900-XTX_79bab741.mxr --input-dim @input 1 3 1080 1920

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

The dropdown for ROCm version only goes as high as 6.0, but I encountered this running ROCm 7.1.1. The dropdown needs updated in the issue template.

[pfultz2]

Very likely the problem is the --optimize flag as that doesnt run normalization. Does it work if you remove the --optimize flag?

The model is still optimized when doing a compile. The --optimize flag is a flag for developers to apply some optimziations without compiling.

[shssoichiro]

Thanks that did seem to fix the issue! I can advise the vs-mlrt devs to remove the flag from their script.

Unfortunately this model still seems to be affected by #4164, as compilation was ongoing for over an hour (didn't finish because earlyoom killed it 💀) and 46 GB of RAM (source file is 49MB), although that is unrelated to this specific issue.

Is there work that should be done internally for the --optimize flag to handle this, or is it safe to close this issue?