Update classic.py

classic.py

@@ -208,8 +208,6 @@ model_ledger = ModelLedger(
 )
 
 canny_processor = CannyDetector()
-
-# 2. Set these to None globally. Do NOT load them here.
 depth_processor = None
 text_encoder = None
 
@@ -217,15 +215,33 @@ def ensure_models_loaded():
     """Helper function to load CUDA models safely inside the ZeroGPU environment."""
     global depth_processor, text_encoder
     
+    # 1. Ensure the pipeline itself is switched to CUDA
+    if pipeline.device.type == "cpu":
+        print("Moving base pipeline to GPU...")
+        pipeline.device = torch.device("cuda")
+    
+    # 2. Load Depth Processor
     if depth_processor is None:
         print("Loading Depth Processor to GPU...")
         depth_processor = MidasDetector.from_pretrained("lllyasviel/Annotators").to("cuda")
         
+    # 3. Load Text Encoder
     if text_encoder is None:
         print("Loading Text Encoder to GPU...")
         text_encoder = model_ledger.text_encoder()
-        print("=" * 80)
-        print("Text encoder loaded and ready!")
+        print("✓ Text encoder loaded!")
+        
+    # 4. Load Video Encoder
+    if not hasattr(pipeline, '_video_encoder') or pipeline._video_encoder is None:
+        print("Loading Video Encoder to GPU...")
+        pipeline._video_encoder = pipeline.model_ledger.video_encoder()
+        print("✓ Video encoder loaded!")
+
+    # 5. Load the Transformer (this was your hidden culprit!)
+    if not hasattr(pipeline, '_transformer') or pipeline._transformer is None:
+        print("Loading Transformer to GPU...")
+        pipeline._transformer = pipeline.model_ledger.transformer()
+        print("✓ Transformer loaded!")
         print("=" * 80)
 
 def on_lora_change(selected: str):
@@ -730,11 +746,12 @@ pipeline = DistilledPipeline(
     local_files_only=False,
 )
 
-pipeline._video_encoder = pipeline.model_ledger.video_encoder()
-pipeline._transformer = pipeline.model_ledger.transformer()
+# DELETE or COMMENT OUT these two lines from the global scope:
+# pipeline._video_encoder = pipeline.model_ledger.video_encoder()
+# pipeline._transformer = pipeline.model_ledger.transformer()
 
 print("=" * 80)
-print("Pipeline fully loaded and ready!")
+print("Pipeline framework initialized on CPU (waiting for ZeroGPU allocation)...")
 print("=" * 80)
 
 class RadioAnimated(gr.HTML):