update vllm miltimodal for api calls convenience

nemo_skills/inference/generate.py

@@ -207,7 +207,7 @@ class GenerationTaskConfig:
     enable_litellm_cache: bool = False
 
     # List of content types to drop from messages (e.g., base64 audio) to keep output files smaller
-    drop_content_types: list[str] = field(default_factory=lambda: ["audio_url"])
+    drop_content_types: list[str] = field(default_factory=lambda: ["audio_url", "input_audio"])
 
     # Audio configuration - set by benchmarks that need audio processing (mmau-pro, audiobench, etc.)
     enable_audio: bool = False  # Enable audio preprocessing (set by benchmark configs)

nemo_skills/inference/model/audio_utils.py

@@ -145,6 +145,8 @@ def make_audio_content_block(base64_audio: str, audio_format: str = "audio_url")
     if audio_format == "input_audio":
         # OpenAI native format (works with NVIDIA API / Gemini / Azure)
         return {"type": "input_audio", "input_audio": {"data": base64_audio, "format": "wav"}}
-    else:
+    elif audio_format == "audio_url":
         # Data URI format (works with vLLM / Qwen)
         return {"type": "audio_url", "audio_url": {"url": f"data:audio/wav;base64,{base64_audio}"}}
+    else:
+        raise ValueError(f"Unsupported audio_format '{audio_format}'. Use 'audio_url' or 'input_audio'.")