NeMo-Speech.cpp/docs/troubleshooting.md at main · NVIDIA/NeMo-Speech.cpp

GitHub

CLI exit codes

All commands use the same stable process status contract. With --json, an error is written to stderr as one JSON object containing the same exit_code and a machine-readable type; stdout remains reserved for successful results.

CodeMeaningJSON error type0success-1runtime, inference, network, or partial batch failureruntime_error2invalid command, option, input, or configurationinvalid_argument3required model or companion artifact is missingmissing_model4capability is not compiled or supported by the selected modelunsupported_featureStart with:

nemo-speech doctor nemo-speech model info ./models/model.ggufdoctor --json is suitable for support bundles and automation. It reports the runtime version, compiled capabilities, backend devices, and driver status.

Common failures

SymptomCheckResolutionCommand is absentnemo-speech --help; doctor featuresInstall/build an archive containing that component.No suitable modelmodel list; model info FILEUse an indexed model name, run model pull NAME, or pass a local GGUF path.Automatic model download is unavailabledoctor --jsonmodel_download field; curl --versionInstall curl and ensure it is on PATH, or use an existing local model.Missing companion modelcommand errorDownload the reported component and pass it explicitly or set it in YAML.GPU requested but unavailabledoctor --json devicesInstall the matching backend build/driver or use --device cpu.Server exits before listeningserver stderrCorrect the reported model/component path or compatibility error.HTTP 401request Authorization headerSend Authorization: Bearer $NEMO_SPEECH_HTTP_API_KEY.HTTP 413upload sizeRaise --max-upload-mb intentionally or split the input.gRPC UNAVAILABLEriva_server listener and portStart riva_server --bind 0.0.0.0:50051, verify binding/firewall, and use plaintext unless TLS is terminated externally.Unsupported WAVfile formatUse PCM16 or float32 WAV; the CLI handles mono/stereo and 8-96 kHz.Out of memorydevice inventory and compositionUse a smaller quantization, reduce concurrency/context, or load fewer capabilities.Keep diagnostics on stderr and command results on stdout when collecting logs. Never include API keys, registry tokens, or signed model URLs in a report.