first stab at adding video

server/models.py

@@ -5,6 +5,7 @@ from server.vad import StreamingVAD
 from server.asr import ASREngine
 from server.llm import LLMEngine
 from server.tts import TTSEngine
+from server.video import VideoConfig, VideoEngine
 
 log = logging.getLogger(__name__)
 
@@ -31,6 +32,7 @@ class ModelManager:
         self.asr_engine: ASREngine | None = None
         self.llm_engine: LLMEngine | None = None
         self.tts_engine: TTSEngine | None = None
+        self.video_engine: VideoEngine | None = None
 
     def load_all(self):
         """Load all models sequentially. Call from the main process."""
@@ -38,6 +40,7 @@ class ModelManager:
         self._load_asr()
         self._load_llm()
         self._load_tts()
+        self._load_video()
         log.info("All models loaded successfully.")
 
     def _load_vad(self):
@@ -84,8 +87,8 @@ class ModelManager:
             log.info("Loading Qwen3-4B (GPTQ 4-bit)...")
             from transformers import AutoModelForCausalLM, AutoTokenizer
 
-            model_name = "Qwen/Qwen3.5-0.8B"
-            
+            # model_name = "Qwen/Qwen3.5-0.8B"
+            model_name = "dphn/Dolphin-X1-8B-FP8"
             tokenizer = AutoTokenizer.from_pretrained(model_name)
             device = get_device()
             model = AutoModelForCausalLM.from_pretrained(
@@ -101,6 +104,27 @@ class ModelManager:
         self.tts_engine = TTSEngine()
         log.info("Kokoro TTS loaded.")
 
+    def _load_video(self):
+        """Load the avatar video stack iff config.video.enabled is true.
+
+        Leaves ``video_engine`` as None when disabled so existing voice flow
+        is untouched. Later phases replace this stub with actual Wan2.2 +
+        MuseTalk loading inside ``VideoEngine``.
+        """
+        from server.config import config
+
+        video_cfg_raw = config.get("video", {}) or {}
+        if not video_cfg_raw.get("enabled", False):
+            log.info("Video engine disabled (config.video.enabled=false). Skipping load.")
+            return
+
+        log.info("Loading avatar video engine...")
+        cfg = VideoConfig.from_dict(video_cfg_raw)
+        self.video_engine = VideoEngine(cfg)
+        if cfg.loras:
+            self.video_engine.load_loras(cfg.loras)
+        log.info("Avatar video engine loaded (mode=%s).", cfg.mode)
+
     def create_vad(self) -> StreamingVAD:
         """Create a new StreamingVAD instance for a client session."""
         return StreamingVAD(self.vad_model)