Upload 2 files

FlowFacade.py

@@ -29,7 +29,7 @@ class FlowFacade:
     def _calculate_gpu_duration(self, image: Image.Image, duration_seconds: float,
                                 num_inference_steps: int, enable_prompt_expansion: bool, **kwargs) -> int:
         BASE_FRAMES_HEIGHT_WIDTH = 81 * 832 * 624
-        BASE_STEP_DURATION = 45  # Increased for CPU offload overhead
+        BASE_STEP_DURATION = 12  # Optimized for sequential CPU offload
 
         resized_image = self.video_engine.resize_image(image)
         width, height = resized_image.width, resized_image.height
@@ -37,13 +37,13 @@ class FlowFacade:
 
         factor = frames * width * height / BASE_FRAMES_HEIGHT_WIDTH
         step_duration = BASE_STEP_DURATION * factor ** 1.5
-        total_duration = 60 + int(num_inference_steps) * step_duration  # Increased base time
+        total_duration = 40 + int(num_inference_steps) * step_duration
 
         if enable_prompt_expansion:
-            total_duration += 50
+            total_duration += 35
 
-        # Minimum 300 seconds (5 minutes) for CPU offload mode
-        return max(int(total_duration), 300)
+        # Minimum 120 seconds (2 minutes) for sequential offload
+        return max(int(total_duration), 120)
 
     @spaces.GPU(duration=_calculate_gpu_duration)
     def generate_video_from_image(self, image: Image.Image, user_instruction: str,

VideoEngine.py

@@ -252,13 +252,13 @@ class VideoEngine:
             except Exception as e:
                 print(f"⚠ VAE optimization warning: {e}")
 
-            # Stage 4: Enable model CPU offload for ZeroGPU memory efficiency
-            print("→ [4/8] Enabling model CPU offload...")
-            self.pipeline.enable_model_cpu_offload()
+            # Stage 4: Enable sequential CPU offload (faster than model offload)
+            print("→ [4/8] Enabling sequential CPU offload...")
+            self.pipeline.enable_sequential_cpu_offload()
             gc.collect()
             if torch.cuda.is_available():
                 torch.cuda.empty_cache()
-            print("✓ Model CPU offload enabled (dynamic GPU/CPU memory management)")
+            print("✓ Sequential CPU offload enabled (only active component on GPU)")
 
             # Stage 5: Enable TF32 acceleration
             self._enable_tf32()