[Feature Request] Report error codes that represent critical errors, such as device-removal

[mingmingtasd]

Describe the feature request

Is there any plan to report more specific error code values for critical errors such as “device-removal” , “device-reset” and “out of memory”?

I looked into the ORT repo, I found that the OrtErrorCode values returned by ORT API like ORT_FAILand ORT_RUNTIME_EXCEPTION are somewhat general.

1. OrtModelEditorAPI::CreateSessionFromModel method tries to catch all exceptions and converts them to ORT_FAILat 1
2. InferenceSession::Run method tries to catch all exceptions and converts them to ORT_FAILat 2 or ORT_RUNTIME_EXCEPTIONat 3

Relevant issue: #26606

/cc @fdwr @RafaelCintron @huningxin @skottmckay

Describe scenario use case

For our WebNN implementation in Chrome browser, we handle the OrtErrorCode values from ORT API, but we don’t know which values should be treated as critical errors, such as
“device-removal” , “device-reset” and “out of memory”, as defined at DXGI_ERROR. For these errors, WebNN needs to destroy the context and may even decide to kill the Chrome GPU process.

[fdwr]

If they're all being funneled into ORT_FAIL and ORT_RUNTIME_EXCEPTION, then that's certainly not helpful. Is this error originating from the OpenVINO EP and/or DML EP?

which values should be treated as critical errors

OOM is critical, but I wouldn't consider device removal a critical error, as one could experience device removal simply from sleeping or closing the lid of their laptop (which switches power profiles), and so device removal should be recoverable.

Hmm, I see some other existing error codes that might be better mappings (ORT_EP_FAIL, ORT_MODEL_REQUIRES_COMPILATION...), but I'm guessing you want a distinct error code, yes?

typedef enum OrtErrorCode {
  ORT_OK,
  ORT_FAIL,
  ORT_INVALID_ARGUMENT,
  ORT_NO_SUCHFILE,
  ORT_NO_MODEL,
  ORT_ENGINE_ERROR,
  ORT_RUNTIME_EXCEPTION,
  ORT_INVALID_PROTOBUF,
  ORT_MODEL_LOADED,
  ORT_NOT_IMPLEMENTED,
  ORT_INVALID_GRAPH,
  ORT_EP_FAIL,
  ORT_MODEL_LOAD_CANCELED,
  ORT_MODEL_REQUIRES_COMPILATION,
  ORT_NOT_FOUND,
} OrtErrorCode;