add lm studio option

server/models.py

@@ -46,7 +46,7 @@ class ModelManager:
         from server.vad import SileroVADOnnx
 
         model_path = hf_hub_download(
-            repo_id="onnx-community/silero-vad", filename="silero_vad.onnx"
+            repo_id="onnx-community/silero-vad", filename="onnx/model.onnx"
         )
         self.vad_model = SileroVADOnnx(model_path)
         log.info("Silero VAD loaded (ONNX, CPU).")
@@ -66,19 +66,32 @@ class ModelManager:
         log.info("Qwen3-ASR-0.6B loaded.")
 
     def _load_llm(self):
-        log.info("Loading Qwen3-4B (GPTQ 4-bit)...")
-        from transformers import AutoModelForCausalLM, AutoTokenizer
+        from server.config import config
 
-        model_name = "Qwen/Qwen3.5-0.8B"
-        
-        tokenizer = AutoTokenizer.from_pretrained(model_name)
-        device = get_device()
-        model = AutoModelForCausalLM.from_pretrained(
-            model_name,
-            device_map=device,
-        )
-        self.llm_engine = LLMEngine(model, tokenizer)
-        log.info("Qwen3-4B-GPTQ-Int4 loaded (~2.5GB VRAM).")
+        backend = config.get("llm", {}).get("backend", "local")
+
+        if backend == "lmstudio":
+            from server.llm import LMStudioEngine
+
+            lms = config.get("llm", {}).get("lmstudio", {})
+            url = lms.get("url", "http://host.docker.internal:1234")
+            model = lms.get("model", "") or ""
+            log.info(f"Using LM Studio backend at {url} (model={model or 'server default'})")
+            self.llm_engine = LMStudioEngine(url, model)
+        else:
+            log.info("Loading Qwen3-4B (GPTQ 4-bit)...")
+            from transformers import AutoModelForCausalLM, AutoTokenizer
+
+            model_name = "Qwen/Qwen3.5-0.8B"
+            
+            tokenizer = AutoTokenizer.from_pretrained(model_name)
+            device = get_device()
+            model = AutoModelForCausalLM.from_pretrained(
+                model_name,
+                device_map=device,
+            )
+            self.llm_engine = LLMEngine(model, tokenizer)
+            log.info("Qwen3-4B-GPTQ-Int4 loaded (~2.5GB VRAM).")
 
     def _load_tts(self):
         log.info("Loading Kokoro TTS...")