Fix cluster client communication by starting local backend process

- Modified cluster client to start a local backend process alongside workers
- Backend process handles communication between cluster client and local workers
- Fixed process cleanup to properly terminate backend and worker processes
- This resolves the timeout issue when cluster client forwards jobs to local backend

vidai/cluster_client.py

@@ -129,7 +129,7 @@ class ClusterClient:
                 self.connected = False
 
     async def start_local_processes(self) -> None:
-        """Start local worker processes based on available backends (GPU and CPU)."""
+        """Start local backend and worker processes based on available backends (GPU and CPU)."""
         from .compat import detect_gpu_backends, get_available_backends
 
         gpu_info = detect_gpu_backends()
@@ -140,6 +140,14 @@ class ClusterClient:
 
         # Only start processes if not already started
         if not self.local_processes:
+            # Start local backend process first
+            backend_cmd = [sys.executable, '-m', 'vidai.backend']
+            self.local_processes['backend'] = subprocess.Popen(backend_cmd)
+            print("Started local backend process")
+
+            # Give backend time to start up
+            await asyncio.sleep(2)
+
             # Start analysis workers for available backends (including CPU)
             for backend in available_backends:
                 proc_name = f'analysis_{backend}'
@@ -481,10 +489,13 @@ class ClusterClient:
             except:
                 pass
 
-        # Clear the processes dict
+        # Clear the processes dict (but keep backend)
+        backend_proc = self.local_processes.get('backend')
         self.local_processes.clear()
         self.process_weights.clear()
         self.process_models.clear()
+        if backend_proc:
+            self.local_processes['backend'] = backend_proc
 
         # Check if backend is available
         from .compat import get_available_backends
@@ -709,9 +720,15 @@ class ClusterClient:
 
         # Final cleanup
         for proc in self.local_processes.values():
-            proc.terminate()
+            try:
+                proc.terminate()
+            except:
+                pass
         for proc in self.local_processes.values():
-            proc.wait()
+            try:
+                proc.wait()
+            except:
+                pass
 
 
 def start_cluster_client(host: str, port: int, token: str, optimize: bool = False, flash: bool = False, weight: int = 100, shared_dir: str = None) -> None: