vllm_omni.entrypoints.cli.benchmark.omni_duplex_eval ¶

CLI for Omni-DuplexEval generate / evaluate / summarize.

generate still writes per-sample timed sentences and *.meta.json. Duplex performance metrics are written to <response-root>/duplex_metrics.json. Skipped samples keep previously recorded rows; incoming (split, sample_id) rows replace the matching ones, then duplex_stream_* is recomputed.

help = "Generate, evaluate, or summarize Omni-DuplexEval artifacts."

name = 'omni-duplex-eval'

add_cli_args(parser: ArgumentParser) -> None

cmd(args: Namespace) -> None

add_cli_args(parser: ArgumentParser) -> None

run(args: Namespace) -> int