Replace empty_cache with torch_gc

extensions-builtin/LDSR/ldsr_model_arch.py

@@ -111,10 +111,7 @@ class LDSR:
         eta = 1.0
 
         gc.collect()
-        if torch.cuda.is_available:
-            torch.cuda.empty_cache()
-        if devices.backend == 'ipex':
-            torch.xpu.empty_cache()
+        devices.torch_gc()
 
         im_og = image
         width_og, height_og = im_og.size
@@ -150,8 +147,7 @@ class LDSR:
 
         del model
         gc.collect()
-        if torch.cuda.is_available:
-            torch.cuda.empty_cache()
+        devices.torch_gc()
 
         return a
 

extensions-builtin/ScuNET/scripts/scunet_model.py

@@ -88,7 +88,7 @@ class UpscalerScuNET(modules.upscaler.Upscaler):
 
     def do_upscale(self, img: PIL.Image.Image, selected_file):
 
-        torch.cuda.empty_cache()
+        devices.torch_gc()
 
         model = self.load_model(selected_file)
         if model is None:
@@ -112,7 +112,7 @@ class UpscalerScuNET(modules.upscaler.Upscaler):
         torch_output = torch_output[:, :h * 1, :w * 1] # remove padding, if any
         np_output: np.ndarray = torch_output.float().cpu().clamp_(0, 1).numpy()
         del torch_img, torch_output
-        torch.cuda.empty_cache()
+        devices.torch_gc()
 
         output = np_output.transpose((1, 2, 0))  # CHW to HWC
         output = output[:, :, ::-1]  # BGR to RGB

extensions-builtin/SwinIR/scripts/swinir_model.py

@@ -40,13 +40,7 @@ class UpscalerSwinIR(Upscaler):
             return img
         model = model.to(device_swinir, dtype=devices.dtype)
         img = upscale(img, model)
-        try:
-            if devices.backend == 'ipex':
-                torch.xpu.empty_cache()
-            else:
-                torch.cuda.empty_cache()
-        except Exception:
-            pass
+        devices.torch_gc()
         return img
 
     def load_model(self, path, scale=4):

modules/codeformer_model.py

@@ -101,10 +101,7 @@ def setup_model(dirname):
                             output = self.net(cropped_face_t, w=w if w is not None else shared.opts.code_former_weight, adain=True)[0]
                             restored_face = tensor2img(output, rgb2bgr=True, min_max=(-1, 1))
                         del output
-                        if devices.backend == 'ipex':
-                            torch.xpu.empty_cache()
-                        else:
-                            torch.cuda.empty_cache()
+                        devices.torch_gc()
                     except Exception as error:
                         print(f'\tFailed inference for CodeFormer: {error}', file=sys.stderr)
                         restored_face = tensor2img(cropped_face_t, rgb2bgr=True, min_max=(-1, 1))

modules/devices.py

@@ -193,7 +193,7 @@ else:
 if backend == 'ipex':
     #Fix broken functions with ipex
     from modules.sd_hijack_utils import CondFunc
-    torch.cuda.empty_cache = torch.xpu.empty_cache
+    torch.cuda.empty_cache = torch_gc
 
     #Functions with dtype errors:
     CondFunc('torch.nn.modules.GroupNorm.forward',