[Bugfix][MiniCPM-o] Align Daily-Omni bench with MiniCPM interleaved A… (#5625)

Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

vllm_omni/entrypoints/cli/benchmark/cli_args.py

@@ -95,6 +95,23 @@ def add_daily_omni_cli_args(parser: argparse.ArgumentParser) -> None:
         "'all': video + WAV together. Sets mm_processor_kwargs.use_audio_in_video=false "
         "and matches official separate video/audio streams.",
     )
+    group.add_argument(
+        "--daily-omni-pack-mode",
+        type=str,
+        choices=["qwen", "minicpm-interleave"],
+        default="qwen",
+        help="How to pack multimodal parts into OpenAI chat messages. "
+        "'qwen' (default): one video_url + one audio_url (Daily-Omni/Qwen protocol). "
+        "'minicpm-interleave': MiniCPM-o official recipe — 1fps frames interleaved with "
+        "matching 1s audio segments as image_url/audio_url pairs (needed to approach "
+        "OpenBMB ~80%% Daily-Omni; requires local Videos extract). By default this mode "
+        "writes JPEG/WAV segments under "
+        "<daily-omni-video-dir>/.minicpm_daily_omni_interleave and sends file:// URLs, "
+        "so start the server with --allowed-local-media-path <daily-omni-video-dir> "
+        "(or a parent). Alternatively pass --daily-omni-inline-local-video to embed "
+        "segments as data URLs. For MiniCPM-o 4.5 string chat templates also start the "
+        "server with --interleave-mm-strings.",
+    )
     group.add_argument(
         "--daily-omni-save-eval-items",
         action="store_true",