Failed to quantize ResNet50 model to int8

[gyulaz-htec]

I'm trying to load a resnet50 model with quantize_int8 using calibration data, but getting the following error: LLVM ERROR: Expected to find GEMM, convolution, or attention op, and didn't
The error comes from MLIR:
https://github.com/ROCm/rocMLIR/blob/45d830dbbc15fe84c41d95585f526a50719020eb/mlir/lib/Dialect/Rock/Tuning/RockTuningImpl.cpp#L493

Description:

• I'm trying to use the https://github.com/ROCm/migraphx-mlperf repo to run MLPerf inference for ResNet50 with int8 quantization.
• The model loading is in C++ and the related code part is here: https://github.com/ROCm/migraphx-mlperf/blob/mlperf_resurrection/prototypes/inference_v2/harness/program.hpp#L202-L232
• We're loading the calibration data from a numpy file
• The model is modified with a graph surgeon script before we load it. The graph surgeon script is in the migraphx-mlperf repo here
  There is a description in the script about what it changes in the model: https://github.com/ROCm/migraphx-mlperf/blob/mlperf_resurrection/prototypes/inference_v2/resnet50/scripts/rn50_graphsurgeon.py#L322-L366

cc @TedThemistokleous @pfultz2

I can compile the list of reproduce steps or a pre-built docker image if needed

[gyulaz-htec]

This turned out to be an user error from my side. The input data was in the wrong format for the int8 quantization (I wanted to load int8 data but float was the expected). I'm closing this issue.