Troubleshooting

Debug from the lowest layer upward.

GPU Not Visible

Symptoms:

gpu target unavailable
ROCm device not found
container cannot see AMD GPU

Check:

rocminfo
rocm-smi
ls -l /dev/kfd /dev/dri
groups

In Docker, include:

--device=/dev/kfd
--device=/dev/dri
--group-add video
--ipc=host
--security-opt seccomp=unconfined

Run:

migraphx-driver compile model.onnx --gpu

Then check:

unsupported operator
bad dynamic shape
opset mismatch
invalid ONNX graph
NMS output shape behavior

Try:

python -m onnx.checker model.onnx

Print the program parameters and outputs from a Python smoke test. Compare them with your postprocess contract.

For object detection, confirm:

batch dimension
number of detections
field count per detection
coordinate format
class index type
score position

Separate the measurements:

MIGraphX driver perf
native backend inference only
preprocess + inference
full decode -> output pipeline

If migraphx-driver perf is good but the app is slow, focus on application memory movement, batching, preprocessing, or output stages.

Check:

Actions build succeeded
Pages source is GitHub Actions
base path matches repository name
out/ artifact uploaded
browser console asset URLs

For custom domains, set NEXT_PUBLIC_BASE_PATH to an empty value during build.