Run on CPU: replace GPU decorators and cuda providers

app.py

@@ -1,106 +1,106 @@
-import spaces
-import os
-import sys
-import site
-
-if "--use-sage-attention" not in sys.argv:
-    sys.argv.append("--use-sage-attention")
-    print("🚀 [SageAttention] Injected '--use-sage-attention' into sys.argv.")
-
-APP_DIR = os.path.dirname(os.path.abspath(__file__))
-if APP_DIR not in sys.path:
-    sys.path.insert(0, APP_DIR)
-    print(f"✅ Added project root '{APP_DIR}' to sys.path.")
-
-SAGE_PATCH_APPLIED = False
-
-def apply_sage_attention_patch():
-    global SAGE_PATCH_APPLIED
-    if SAGE_PATCH_APPLIED:
-        return "SageAttention patch already applied."
-
-    try:
-        from comfy import model_management
-        import sageattention
-        
-        print("--- [Runtime Patch] sageattention package found. Applying patch... ---")
-        model_management.sage_attention_enabled = lambda: True
-        model_management.pytorch_attention_enabled = lambda: False
-        
-        SAGE_PATCH_APPLIED = True
-        return "✅ Successfully enabled SageAttention."
-    except ImportError:
-        SAGE_PATCH_APPLIED = False
-        msg = "--- [Runtime Patch] ⚠️ sageattention package not found. Continuing with default attention. ---"
-        print(msg)
-        return msg
-    except Exception as e:
-        SAGE_PATCH_APPLIED = False
-        msg = f"--- [Runtime Patch] ❌ An error occurred while applying SageAttention patch: {e} ---"
-        print(msg)
-        return msg
-
-@spaces.GPU
-def dummy_gpu_for_startup():
-    try:
-        print("--- [GPU Startup] Dummy function for startup check initiated. ---")
-        patch_result = apply_sage_attention_patch()
-        print(f"--- [GPU Startup] {patch_result} ---")
-        print("--- [GPU Startup] Startup check passed. ---")
-        return "Startup check passed."
-    except BaseException as e:
-        err_msg = str(e)
-        if "uncorrectable ECC error" in err_msg or "cudaErrorECCUncorrectable" in err_msg:
-            print("\n" + "="*80)
-            print(f"🚨 [Fatal GPU Error] Captured uncorrectable ECC error during inference: {err_msg}")
-            print("🚨 Terminating process to trigger an automatic container restart...")
-            print("="*80 + "\n")
-            os._exit(1)
-        raise e
-
-
-def main():
-    from comfy_integration import setup as setup_comfyui
-    from utils.app_utils import load_ipadapter_presets
-    
-    print("--- [Setup] Starting ComfyUI initialization ---")
-    setup_comfyui.initialize_comfyui()
-    
-    print("--- [Setup] Applying SageAttention Runtime Patch ---")
-    patch_result = apply_sage_attention_patch()
-    print(f"--- [Setup] {patch_result} ---")
-
-    print("--- [Setup] Reloading site-packages to detect newly installed packages... ---")
-    try:
-        site.main()
-        print("--- [Setup] ✅ Site-packages reloaded. ---")
-    except Exception as e:
-        print(f"--- [Setup] ⚠️  Warning: Could not fully reload site-packages: {e} ---")
-
-    print("--- Initiating GPU Startup Check & SageAttention Patch Verification ---")
-    try:
-        dummy_gpu_for_startup()
-    except Exception as e:
-        print(f"--- [GPU Startup] ⚠️ Warning: Startup check failed: {e} ---")
-
-    print("--- Starting Application Setup ---")
-    
-    print("--- Loading IPAdapter presets ---")
-    load_ipadapter_presets()
-    print("--- ✅ IPAdapter setup complete. ---")
-
-
-    print("--- Environment configured. Proceeding with module imports. ---")
-    from ui.layout import build_ui
-    from ui.events import attach_event_handlers
-
-    print(f"✅ Working directory is stable: {os.getcwd()}")
-
-    demo = build_ui(attach_event_handlers)
-        
-    print("--- Launching Gradio Interface ---")
-    demo.queue().launch(server_name="0.0.0.0", server_port=7860)
-
-
-if __name__ == "__main__":
+import spaces
+import os
+import sys
+import site
+
+if "--use-sage-attention" not in sys.argv:
+    sys.argv.append("--use-sage-attention")
+    print("🚀 [SageAttention] Injected '--use-sage-attention' into sys.argv.")
+
+APP_DIR = os.path.dirname(os.path.abspath(__file__))
+if APP_DIR not in sys.path:
+    sys.path.insert(0, APP_DIR)
+    print(f"✅ Added project root '{APP_DIR}' to sys.path.")
+
+SAGE_PATCH_APPLIED = False
+
+def apply_sage_attention_patch():
+    global SAGE_PATCH_APPLIED
+    if SAGE_PATCH_APPLIED:
+        return "SageAttention patch already applied."
+
+    try:
+        from comfy import model_management
+        import sageattention
+        
+        print("--- [Runtime Patch] sageattention package found. Applying patch... ---")
+        model_management.sage_attention_enabled = lambda: True
+        model_management.pytorch_attention_enabled = lambda: False
+        
+        SAGE_PATCH_APPLIED = True
+        return "✅ Successfully enabled SageAttention."
+    except ImportError:
+        SAGE_PATCH_APPLIED = False
+        msg = "--- [Runtime Patch] ⚠️ sageattention package not found. Continuing with default attention. ---"
+        print(msg)
+        return msg
+    except Exception as e:
+        SAGE_PATCH_APPLIED = False
+        msg = f"--- [Runtime Patch] ❌ An error occurred while applying SageAttention patch: {e} ---"
+        print(msg)
+        return msg
+
+# @spaces.cpu (disabled for CPU)
+def dummy_gpu_for_startup():
+    try:
+        print("--- [GPU Startup] Dummy function for startup check initiated. ---")
+        patch_result = apply_sage_attention_patch()
+        print(f"--- [GPU Startup] {patch_result} ---")
+        print("--- [GPU Startup] Startup check passed. ---")
+        return "Startup check passed."
+    except BaseException as e:
+        err_msg = str(e)
+        if "uncorrectable ECC error" in err_msg or "cudaErrorECCUncorrectable" in err_msg:
+            print("\n" + "="*80)
+            print(f"🚨 [Fatal GPU Error] Captured uncorrectable ECC error during inference: {err_msg}")
+            print("🚨 Terminating process to trigger an automatic container restart...")
+            print("="*80 + "\n")
+            os._exit(1)
+        raise e
+
+
+def main():
+    from comfy_integration import setup as setup_comfyui
+    from utils.app_utils import load_ipadapter_presets
+    
+    print("--- [Setup] Starting ComfyUI initialization ---")
+    setup_comfyui.initialize_comfyui()
+    
+    print("--- [Setup] Applying SageAttention Runtime Patch ---")
+    patch_result = apply_sage_attention_patch()
+    print(f"--- [Setup] {patch_result} ---")
+
+    print("--- [Setup] Reloading site-packages to detect newly installed packages... ---")
+    try:
+        site.main()
+        print("--- [Setup] ✅ Site-packages reloaded. ---")
+    except Exception as e:
+        print(f"--- [Setup] ⚠️  Warning: Could not fully reload site-packages: {e} ---")
+
+    print("--- Initiating GPU Startup Check & SageAttention Patch Verification ---")
+    try:
+        dummy_gpu_for_startup()
+    except Exception as e:
+        print(f"--- [GPU Startup] ⚠️ Warning: Startup check failed: {e} ---")
+
+    print("--- Starting Application Setup ---")
+    
+    print("--- Loading IPAdapter presets ---")
+    load_ipadapter_presets()
+    print("--- ✅ IPAdapter setup complete. ---")
+
+
+    print("--- Environment configured. Proceeding with module imports. ---")
+    from ui.layout import build_ui
+    from ui.events import attach_event_handlers
+
+    print(f"✅ Working directory is stable: {os.getcwd()}")
+
+    demo = build_ui(attach_event_handlers)
+        
+    print("--- Launching Gradio Interface ---")
+    demo.queue().launch(server_name="0.0.0.0", server_port=7860)
+
+
+if __name__ == "__main__":
     main()

chain_injectors/flux1_ipadapter_injector.py

@@ -1,46 +1,46 @@
-def inject(assembler, chain_definition, chain_items):
-    if not chain_items:
-        return
-
-    ksampler_name = chain_definition.get('ksampler_node', 'ksampler')
-    if ksampler_name not in assembler.node_map:
-        print(f"Warning: KSampler node '{ksampler_name}' not found for Flux1 IPAdapter chain. Skipping.")
-        return
-        
-    ksampler_id = assembler.node_map[ksampler_name]
-
-    if 'model' not in assembler.workflow[ksampler_id]['inputs']:
-        print(f"Warning: KSampler node '{ksampler_name}' is missing 'model' input. Skipping Flux1 IPAdapter chain.")
-        return
-        
-    current_model_connection = assembler.workflow[ksampler_id]['inputs']['model']
-    
-    for item_data in chain_items:
-        image_loader_id = assembler._get_unique_id()
-        image_loader_node = assembler._get_node_template("LoadImage")
-        image_loader_node['inputs']['image'] = item_data['image']
-        assembler.workflow[image_loader_id] = image_loader_node
-
-        ipadapter_loader_id = assembler._get_unique_id()
-        ipadapter_loader_node = assembler._get_node_template("IPAdapterFluxLoader")
-        ipadapter_loader_node['inputs']['ipadapter'] = "ip-adapter.bin"
-        ipadapter_loader_node['inputs']['clip_vision'] = "google/siglip-so400m-patch14-384"
-        ipadapter_loader_node['inputs']['provider'] = "cuda"
-        assembler.workflow[ipadapter_loader_id] = ipadapter_loader_node
-        
-        apply_ipa_id = assembler._get_unique_id()
-        apply_ipa_node = assembler._get_node_template("ApplyIPAdapterFlux")
-        
-        apply_ipa_node['inputs']['weight'] = item_data['weight']
-        apply_ipa_node['inputs']['start_percent'] = item_data.get('start_percent', 0.0)
-        apply_ipa_node['inputs']['end_percent'] = item_data.get('end_percent', 0.6)
-        
-        apply_ipa_node['inputs']['model'] = current_model_connection
-        apply_ipa_node['inputs']['ipadapter_flux'] = [ipadapter_loader_id, 0]
-        apply_ipa_node['inputs']['image'] = [image_loader_id, 0]
-        
-        assembler.workflow[apply_ipa_id] = apply_ipa_node
-        current_model_connection = [apply_ipa_id, 0]
-
-    assembler.workflow[ksampler_id]['inputs']['model'] = current_model_connection
+def inject(assembler, chain_definition, chain_items):
+    if not chain_items:
+        return
+
+    ksampler_name = chain_definition.get('ksampler_node', 'ksampler')
+    if ksampler_name not in assembler.node_map:
+        print(f"Warning: KSampler node '{ksampler_name}' not found for Flux1 IPAdapter chain. Skipping.")
+        return
+        
+    ksampler_id = assembler.node_map[ksampler_name]
+
+    if 'model' not in assembler.workflow[ksampler_id]['inputs']:
+        print(f"Warning: KSampler node '{ksampler_name}' is missing 'model' input. Skipping Flux1 IPAdapter chain.")
+        return
+        
+    current_model_connection = assembler.workflow[ksampler_id]['inputs']['model']
+    
+    for item_data in chain_items:
+        image_loader_id = assembler._get_unique_id()
+        image_loader_node = assembler._get_node_template("LoadImage")
+        image_loader_node['inputs']['image'] = item_data['image']
+        assembler.workflow[image_loader_id] = image_loader_node
+
+        ipadapter_loader_id = assembler._get_unique_id()
+        ipadapter_loader_node = assembler._get_node_template("IPAdapterFluxLoader")
+        ipadapter_loader_node['inputs']['ipadapter'] = "ip-adapter.bin"
+        ipadapter_loader_node['inputs']['clip_vision'] = "google/siglip-so400m-patch14-384"
+        ipadapter_loader_node['inputs']['provider'] = "cpu"
+        assembler.workflow[ipadapter_loader_id] = ipadapter_loader_node
+        
+        apply_ipa_id = assembler._get_unique_id()
+        apply_ipa_node = assembler._get_node_template("ApplyIPAdapterFlux")
+        
+        apply_ipa_node['inputs']['weight'] = item_data['weight']
+        apply_ipa_node['inputs']['start_percent'] = item_data.get('start_percent', 0.0)
+        apply_ipa_node['inputs']['end_percent'] = item_data.get('end_percent', 0.6)
+        
+        apply_ipa_node['inputs']['model'] = current_model_connection
+        apply_ipa_node['inputs']['ipadapter_flux'] = [ipadapter_loader_id, 0]
+        apply_ipa_node['inputs']['image'] = [image_loader_id, 0]
+        
+        assembler.workflow[apply_ipa_id] = apply_ipa_node
+        current_model_connection = [apply_ipa_id, 0]
+
+    assembler.workflow[ksampler_id]['inputs']['model'] = current_model_connection
     print(f"Flux1 IPAdapter injector applied. KSampler model input re-routed through {len(chain_items)} IPAdapter(s).")

chain_injectors/sd3_ipadapter_injector.py

@@ -1,66 +1,66 @@
-def inject(assembler, chain_definition, chain_items):
-    if not chain_items:
-        return
-
-    ksampler_name = chain_definition.get('ksampler_node', 'ksampler')
-    if ksampler_name not in assembler.node_map:
-        print(f"Warning: KSampler node '{ksampler_name}' not found for SD3 IPAdapter chain. Skipping.")
-        return
-        
-    ksampler_id = assembler.node_map[ksampler_name]
-
-    if 'model' not in assembler.workflow[ksampler_id]['inputs']:
-        print(f"Warning: KSampler node '{ksampler_name}' is missing 'model' input. Skipping SD3 IPAdapter chain.")
-        return
-        
-    current_model_connection = assembler.workflow[ksampler_id]['inputs']['model']
-    
-    clip_vision_loader_id = assembler._get_unique_id()
-    clip_vision_loader_node = assembler._get_node_template("CLIPVisionLoader")
-    clip_vision_loader_node['inputs']['clip_name'] = "sigclip_vision_patch14_384.safetensors"
-    assembler.workflow[clip_vision_loader_id] = clip_vision_loader_node
-
-    ipadapter_loader_id = assembler._get_unique_id()
-    ipadapter_loader_node = assembler._get_node_template("IPAdapterSD3Loader")
-    ipadapter_loader_node['inputs']['ipadapter'] = "ip-adapter_sd35l_instantx.bin"
-    ipadapter_loader_node['inputs']['provider'] = "cuda"
-    assembler.workflow[ipadapter_loader_id] = ipadapter_loader_node
-
-    for item_data in chain_items:
-        image_loader_id = assembler._get_unique_id()
-        image_loader_node = assembler._get_node_template("LoadImage")
-        image_loader_node['inputs']['image'] = item_data['image']
-        assembler.workflow[image_loader_id] = image_loader_node
-
-        image_scaler_id = assembler._get_unique_id()
-        image_scaler_node = assembler._get_node_template("ImageScaleToTotalPixels")
-        image_scaler_node['inputs']['image'] = [image_loader_id, 0]
-        image_scaler_node['inputs']['upscale_method'] = 'nearest-exact'
-        image_scaler_node['inputs']['megapixels'] = 1.0
-        assembler.workflow[image_scaler_id] = image_scaler_node
-
-        clip_vision_encode_id = assembler._get_unique_id()
-        clip_vision_encode_node = assembler._get_node_template("CLIPVisionEncode")
-        clip_vision_encode_node['inputs']['crop'] = "center"
-        clip_vision_encode_node['inputs']['clip_vision'] = [clip_vision_loader_id, 0]
-        clip_vision_encode_node['inputs']['image'] = [image_scaler_id, 0]
-        assembler.workflow[clip_vision_encode_id] = clip_vision_encode_node
-        
-        apply_ipa_id = assembler._get_unique_id()
-        apply_ipa_node = assembler._get_node_template("ApplyIPAdapterSD3")
-        
-        apply_ipa_node['inputs']['weight'] = item_data.get('weight', 1.0)
-        apply_ipa_node['inputs']['start_percent'] = item_data.get('start_percent', 0.0)
-        apply_ipa_node['inputs']['end_percent'] = item_data.get('end_percent', 1.0)
-        
-        apply_ipa_node['inputs']['model'] = current_model_connection
-        apply_ipa_node['inputs']['ipadapter'] = [ipadapter_loader_id, 0]
-        apply_ipa_node['inputs']['image_embed'] = [clip_vision_encode_id, 0]
-        
-        assembler.workflow[apply_ipa_id] = apply_ipa_node
-        
-        current_model_connection = [apply_ipa_id, 0]
-
-    assembler.workflow[ksampler_id]['inputs']['model'] = current_model_connection
-    
+def inject(assembler, chain_definition, chain_items):
+    if not chain_items:
+        return
+
+    ksampler_name = chain_definition.get('ksampler_node', 'ksampler')
+    if ksampler_name not in assembler.node_map:
+        print(f"Warning: KSampler node '{ksampler_name}' not found for SD3 IPAdapter chain. Skipping.")
+        return
+        
+    ksampler_id = assembler.node_map[ksampler_name]
+
+    if 'model' not in assembler.workflow[ksampler_id]['inputs']:
+        print(f"Warning: KSampler node '{ksampler_name}' is missing 'model' input. Skipping SD3 IPAdapter chain.")
+        return
+        
+    current_model_connection = assembler.workflow[ksampler_id]['inputs']['model']
+    
+    clip_vision_loader_id = assembler._get_unique_id()
+    clip_vision_loader_node = assembler._get_node_template("CLIPVisionLoader")
+    clip_vision_loader_node['inputs']['clip_name'] = "sigclip_vision_patch14_384.safetensors"
+    assembler.workflow[clip_vision_loader_id] = clip_vision_loader_node
+
+    ipadapter_loader_id = assembler._get_unique_id()
+    ipadapter_loader_node = assembler._get_node_template("IPAdapterSD3Loader")
+    ipadapter_loader_node['inputs']['ipadapter'] = "ip-adapter_sd35l_instantx.bin"
+    ipadapter_loader_node['inputs']['provider'] = "cpu"
+    assembler.workflow[ipadapter_loader_id] = ipadapter_loader_node
+
+    for item_data in chain_items:
+        image_loader_id = assembler._get_unique_id()
+        image_loader_node = assembler._get_node_template("LoadImage")
+        image_loader_node['inputs']['image'] = item_data['image']
+        assembler.workflow[image_loader_id] = image_loader_node
+
+        image_scaler_id = assembler._get_unique_id()
+        image_scaler_node = assembler._get_node_template("ImageScaleToTotalPixels")
+        image_scaler_node['inputs']['image'] = [image_loader_id, 0]
+        image_scaler_node['inputs']['upscale_method'] = 'nearest-exact'
+        image_scaler_node['inputs']['megapixels'] = 1.0
+        assembler.workflow[image_scaler_id] = image_scaler_node
+
+        clip_vision_encode_id = assembler._get_unique_id()
+        clip_vision_encode_node = assembler._get_node_template("CLIPVisionEncode")
+        clip_vision_encode_node['inputs']['crop'] = "center"
+        clip_vision_encode_node['inputs']['clip_vision'] = [clip_vision_loader_id, 0]
+        clip_vision_encode_node['inputs']['image'] = [image_scaler_id, 0]
+        assembler.workflow[clip_vision_encode_id] = clip_vision_encode_node
+        
+        apply_ipa_id = assembler._get_unique_id()
+        apply_ipa_node = assembler._get_node_template("ApplyIPAdapterSD3")
+        
+        apply_ipa_node['inputs']['weight'] = item_data.get('weight', 1.0)
+        apply_ipa_node['inputs']['start_percent'] = item_data.get('start_percent', 0.0)
+        apply_ipa_node['inputs']['end_percent'] = item_data.get('end_percent', 1.0)
+        
+        apply_ipa_node['inputs']['model'] = current_model_connection
+        apply_ipa_node['inputs']['ipadapter'] = [ipadapter_loader_id, 0]
+        apply_ipa_node['inputs']['image_embed'] = [clip_vision_encode_id, 0]
+        
+        assembler.workflow[apply_ipa_id] = apply_ipa_node
+        
+        current_model_connection = [apply_ipa_id, 0]
+
+    assembler.workflow[ksampler_id]['inputs']['model'] = current_model_connection
+    
     print(f"SD3 IPAdapter injector applied. KSampler model input re-routed through {len(chain_items)} IPAdapter(s).")

core/pipelines/base_pipeline.py

@@ -1,65 +1,66 @@
-from abc import ABC, abstractmethod
-from typing import List, Any, Dict
-import gradio as gr
-import spaces
-import tempfile
-import imageio
-import numpy as np
-import sys
-import os
-
-class BasePipeline(ABC):
-    def __init__(self):
-        from core.model_manager import model_manager
-        self.model_manager = model_manager
-
-    @abstractmethod
-    def get_required_models(self, **kwargs) -> List[str]:
-        pass
-
-    @abstractmethod
-    def run(self, *args, progress: gr.Progress, **kwargs) -> Any:
-        pass
-
-    def _ensure_models_downloaded(self, progress: gr.Progress, **kwargs):
-        """Ensures model files are downloaded before requesting GPU."""
-        required_models = self.get_required_models(**kwargs)
-        self.model_manager.ensure_models_downloaded(required_models, progress=progress)
-
-    def _execute_gpu_logic(self, gpu_function: callable, duration: int, default_duration: int, task_name: str, *args, **kwargs):
-        final_duration = default_duration
-        try:
-            if duration is not None and int(duration) > 0:
-                final_duration = int(duration)
-        except (ValueError, TypeError):
-            print(f"Invalid ZeroGPU duration input for {task_name}. Using default {default_duration}s.")
-            pass
-        
-        print(f"Requesting ZeroGPU for {task_name} with duration: {final_duration} seconds.")
-        gpu_runner = spaces.GPU(duration=final_duration)(gpu_function)
-        
-        try:
-            return gpu_runner(*args, **kwargs)
-        except BaseException as e:
-            err_msg = str(e)
-            if "uncorrectable ECC error" in err_msg or "cudaErrorECCUncorrectable" in err_msg:
-                print("\n" + "="*80)
-                print(f"🚨 [Fatal GPU Error] Captured uncorrectable ECC error during inference: {err_msg}")
-                print("🚨 Terminating process to trigger an automatic container restart...")
-                print("="*80 + "\n")
-                os._exit(1)
-            raise e
-
-    def _encode_video_from_frames(self, frames_tensor_cpu: 'torch.Tensor', fps: int, progress: gr.Progress) -> str:
-        progress(0.9, desc="Encoding video on CPU...")
-        frames_np = (frames_tensor_cpu.numpy() * 255.0).astype(np.uint8)
-        
-        with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as temp_video_file:
-            video_path = temp_video_file.name
-            writer = imageio.get_writer(video_path, fps=fps, codec='libx264', quality=8)
-            for frame in frames_np:
-                writer.append_data(frame)
-            writer.close()
-            
-        progress(1.0, desc="Done!")
+from abc import ABC, abstractmethod
+from typing import List, Any, Dict
+import gradio as gr
+import spaces
+import tempfile
+import imageio
+import numpy as np
+import sys
+import os
+
+class BasePipeline(ABC):
+    def __init__(self):
+        from core.model_manager import model_manager
+        self.model_manager = model_manager
+
+    @abstractmethod
+    def get_required_models(self, **kwargs) -> List[str]:
+        pass
+
+    @abstractmethod
+    def run(self, *args, progress: gr.Progress, **kwargs) -> Any:
+        pass
+
+    def _ensure_models_downloaded(self, progress: gr.Progress, **kwargs):
+        """Ensures model files are downloaded before requesting GPU."""
+        required_models = self.get_required_models(**kwargs)
+        self.model_manager.ensure_models_downloaded(required_models, progress=progress)
+
+    def _execute_gpu_logic(self, gpu_function: callable, duration: int, default_duration: int, task_name: str, *args, **kwargs):
+        final_duration = default_duration
+        try:
+            if duration is not None and int(duration) > 0:
+                final_duration = int(duration)
+        except (ValueError, TypeError):
+            print(f"Invalid ZeroGPU duration input for {task_name}. Using default {default_duration}s.")
+            pass
+        
+        print(f"Requesting ZeroGPU for {task_name} with duration: {final_duration} seconds.")
+        # Direct call without GPU allocation for CPU execution
+        gpu_runner = gpu_function
+        
+        try:
+            return gpu_runner(*args, **kwargs)
+        except BaseException as e:
+            err_msg = str(e)
+            if "uncorrectable ECC error" in err_msg or "cudaErrorECCUncorrectable" in err_msg:
+                print("\n" + "="*80)
+                print(f"🚨 [Fatal GPU Error] Captured uncorrectable ECC error during inference: {err_msg}")
+                print("🚨 Terminating process to trigger an automatic container restart...")
+                print("="*80 + "\n")
+                os._exit(1)
+            raise e
+
+    def _encode_video_from_frames(self, frames_tensor_cpu: 'torch.Tensor', fps: int, progress: gr.Progress) -> str:
+        progress(0.9, desc="Encoding video on CPU...")
+        frames_np = (frames_tensor_cpu.numpy() * 255.0).astype(np.uint8)
+        
+        with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as temp_video_file:
+            video_path = temp_video_file.name
+            writer = imageio.get_writer(video_path, fps=fps, codec='libx264', quality=8)
+            for frame in frames_np:
+                writer.append_data(frame)
+            writer.close()
+            
+        progress(1.0, desc="Done!")
         return video_path