Fix GroupNorm.forward with IPEX

modules/devices.py

@@ -145,8 +145,15 @@ def set_cuda_params():
 
 args = cmd_args.parser.parse_args()
 if args.use_ipex:
+    #Fix broken function in ipex 1.13.120+xpu
+    from modules.sd_hijack_utils import CondFunc
+    CondFunc('torch.nn.modules.GroupNorm.forward',
+        lambda orig_func, *args, **kwargs: orig_func(args[0], args[1].to(args[0].weight.data.dtype)),
+        lambda *args, **kwargs: args[2].dtype != args[1].weight.data.dtype)
+    
+    #Use XPU instead of CPU. %20 Perf improvement on weak CPUs.
     if args.device_id is not None:
-        cpu = torch.device(f"xpu:{args.device_id}") #Use XPU instead of CPU. %20 Perf improvement on weak CPUs.
+        cpu = torch.device(f"xpu:{args.device_id}")
     else:
         cpu = torch.device("xpu")
 else:

modules/shared.py

@@ -304,7 +304,7 @@ options_templates.update(options_section(('cuda', "Compute Settings"), {
     "cuda_dtype": OptionInfo("FP32" if sys.platform == "darwin" else "FP16", "Device precision type", gr.Radio, lambda: {"choices": ["FP32", "FP16", "BF16"]}),
     "no_half": OptionInfo(False, "Use full precision for model (--no-half)", None, None, None),
     "no_half_vae": OptionInfo(False, "Use full precision for VAE (--no-half-vae)"),
-    "upcast_sampling": OptionInfo(True if sys.platform == "darwin" or cmd_opts.use_ipex else False, "Enable upcast sampling"),
+    "upcast_sampling": OptionInfo(True if sys.platform == "darwin" else False, "Enable upcast sampling"),
     "upcast_attn": OptionInfo(False, "Enable upcast cross attention layer"),
     "disable_nan_check": OptionInfo(True, "Disable NaN check in produced images/latent spaces"),
     "rollback_vae": OptionInfo(False, "Attempt to roll back VAE when produced NaN values, requires NaN check (experimental)"),