Allow GPU acceleration with AMD GPUs with ROCm

[CodingAndCodingAgain]

I was able to run piper with an AMD GPU by installing the "onnxruntime-rocm" and the appropriate "torch" package based on my GPU.

Without taking into account the installation of the packages, it didn't require to much setup to get it to work so I thought more people could benefit of it.

Importing the "torch" package before the "onnxruntime" package it's actually necessary since "onnxruntime-rocm" will crash without it.

[malev-v]

NOTE ROCm Execution Provider has been removed since 1.23 release. Please Migrate your applications to use the MIGraphX Execution Provider
ROCm 7.0 is the last offiicaly AMD supported distribution of this provider and all builds going forward (ROCm 7.1+) will have ROCm EP removed.
ROCm-ExecutionProvider.html

But even if you install onnxruntime_rocm (Show)

Use AMD official for rocm-rel-7.0.2 instead pip:

onnxruntime_rocm-1.22.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl

It goes with libhipblas-5aa6e563.so.3.0.70002
link /opt/rocm/lib/librocm_smi64.so.1.0.70203 -> librocm_smi64.so.7
export LD_LIBRARY_PATH= ...

The situation won't change - Python still throws an error.

---

When attempting to use MIGraphXExecutionProvider, the following message appears:
2026-05-26 21:23:10.N [W:onnxruntime:Default, migraphx_execution_provider.cc:180 get_flags_from_env] [MIGraphX EP] MIGraphX ENV Override Variables Set:

But it still uses the CPU!

After --force-reinstall onnxruntime_migraphx ... Оnce again, it falls back to the CPU. Only this appear:

DT [W:onnxruntime:Default, migraphx_execution_provider_utils.h:155 canEvalNodeArgument] Node:/dp/flows.7/NonZero_5 Input:/dp/flows.7/Expand_28_output_0 Can't eval shape
DT [W:onnxruntime:Default, migraphx_execution_provider_utils.h:155 canEvalNodeArgument] Node:/dp/flows.7/NonZero Input:/dp/flows.7/Not_output_0 Can't eval shape
...
DT [W:onnxruntime:Default, migraphx_execution_provider_utils.h:155 canEvalNodeArgument] Node:/Range Input:/ReduceMax_output_0 Can't eval shape

Any ideas?