In exception handlers, clear the torch CUDA cache (if we're using CUD…

invokeai/backend/invoke_ai_web_server.py

@@ -1208,19 +1208,31 @@ def diffusers_step_callback_adapter(*cb_args, **kwargs):
             )
 
         except KeyboardInterrupt:
+            # Clear the CUDA cache on an exception
+            self.empty_cuda_cache()
             self.socketio.emit("processingCanceled")
             raise
         except CanceledException:
+            # Clear the CUDA cache on an exception
+            self.empty_cuda_cache()
             self.socketio.emit("processingCanceled")
             pass
         except Exception as e:
+            # Clear the CUDA cache on an exception
+            self.empty_cuda_cache()
             print(e)
             self.socketio.emit("error", {"message": (str(e))})
             print("\n")
 
             traceback.print_exc()
             print("\n")
 
+    def empty_cuda_cache(self):
+        if self.generate.device.type == "cuda":
+            import torch.cuda
+
+            torch.cuda.empty_cache()
+
     def parameters_to_generated_image_metadata(self, parameters):
         try:
             # top-level metadata minus `image` or `images`

ldm/generate.py

@@ -211,7 +211,7 @@ def __init__(
                 print('>> xformers memory-efficient attention is available but disabled')
         else:
             print('>> xformers not installed')
-            
+
         # model caching system for fast switching
         self.model_manager = ModelManager(mconfig,self.device,self.precision,max_loaded_models=max_loaded_models)
         # don't accept invalid models
@@ -565,11 +565,19 @@ def process_image(image,seed):
                                              image_callback = image_callback)
 
         except KeyboardInterrupt:
+            # Clear the CUDA cache on an exception
+            if self._has_cuda():
+                torch.cuda.empty_cache()
+
             if catch_interrupts:
                 print('**Interrupted** Partial results will be returned.')
             else:
                 raise KeyboardInterrupt
         except RuntimeError:
+            # Clear the CUDA cache on an exception
+            if self._has_cuda():
+                torch.cuda.empty_cache()
+
             print(traceback.format_exc(), file=sys.stderr)
             print('>> Could not generate image.')
 

ldm/invoke/generator/txt2img2img.py

@@ -65,6 +65,10 @@ def make_image(x_T):
                 mode="bilinear"
             )
 
+            # Free up memory from the last generation.
+            if self.model.device.type == 'cuda':
+                torch.cuda.empty_cache()
+
             second_pass_noise = self.get_noise_like(resized_latents)
 
             verbosity = get_verbosity()