Fix whisper-server: check for whisper-server FIRST in transcription endpoint

- Move whisper-server check before audio_model check
- Now whisper-server will be used if available, regardless of audio_model setting
- Also update multi_model_manager.audio_models with cached path

coderai

@@ -2693,26 +2693,27 @@ async def create_transcription(
     - Local faster-whisper models (when --audio-model is specified)
     - whisper.cpp server (when --whisper-server is specified)
     """
+    # Check if whisper-server is available FIRST (before checking audio_model)
+    if multi_model_manager.whisper_server and multi_model_manager.whisper_server.is_running():
+        # Use whisper-server - read file and send to server
+        file_content = await file.read()
+        result = multi_model_manager.whisper_server.transcribe(
+            file_content,
+            language=language,
+            prompt=prompt
+        )
+        if "error" in result:
+            raise HTTPException(status_code=500, detail=result["error"])
+        # Convert whisper-server response to OpenAI format
+        text = result.get("text", "")
+        return {
+            "text": text
+        }
+    
     audio_model = multi_model_manager.audio_model
     
     # If no audio model configured, return an error
     if not audio_model:
-        # Check if whisper-server is available
-        if multi_model_manager.whisper_server and multi_model_manager.whisper_server.is_running():
-            # Use whisper-server - read file and send to server
-            file_content = await file.read()
-            result = multi_model_manager.whisper_server.transcribe(
-                file_content,
-                language=language,
-                prompt=prompt
-            )
-            if "error" in result:
-                raise HTTPException(status_code=500, detail=result["error"])
-            # Convert whisper-server response to OpenAI format
-            text = result.get("text", "")
-            return {
-                "text": text
-            }
         raise HTTPException(
             status_code=400,
             detail="Audio transcription not configured. Use --audio-model or --whisper-server to specify a model."
@@ -4459,13 +4460,11 @@ def main():
                 gpu_device = getattr(args, 'audio_vulkan_device', 0) or 0
                 actual_model_path = whisper_server_mgr.start(model_path=model_to_use, gpu_device=gpu_device)
                 if actual_model_path:
-                    # Update audio_models to store the actual path (not the URL)
+                    # Update audio_models in multi_model_manager to store the actual path (not the URL)
                     if model_to_use != actual_model_path:
-                        audio_models[0] = actual_model_path
-                        multi_model_manager.set_audio_model(actual_model_path, {
-                            'ctx': args.audio_ctx,
-                            'offload': args.audio_offload,
-                        })
+                        # Update the manager's audio_models list
+                        if multi_model_manager.audio_models and multi_model_manager.audio_models[0] == model_to_use:
+                            multi_model_manager.audio_models[0] = actual_model_path
                     print(f"Whisper server started with model: {actual_model_path}")
                 else:
                     print("Warning: Failed to start whisper-server, falling back to other backends")