Some overrides to make it work

invokeai/backend/stable_diffusion/diffusers_pipeline.py

@@ -339,6 +339,12 @@ class StableDiffusionGeneratorPipeline(StableDiffusionPipeline):
         """
         if xformers is available, use it, otherwise use sliced attention.
         """
+        # TODO: use DML_DEVICE?
+        if self.device.type == "privateuseone":
+            self.enable_attention_slicing(slice_size="max")
+            #self.unet.set_default_attn_processor()
+            return
+
         config = InvokeAIAppConfig.get_config()
         if torch.cuda.is_available() and is_xformers_available() and not config.disable_xformers:
             self.enable_xformers_memory_efficient_attention()

invokeai/backend/util/hotfixes.py

@@ -18,6 +18,43 @@ from diffusers.models.unet_2d_condition import UNet2DConditionModel
 import diffusers
 from diffusers.models.controlnet import ControlNetConditioningEmbedding, ControlNetOutput, zero_module
 
+try:
+    import torch_directml
+    torch_directml_installed = True
+    torch_directml_device_type = torch_directml.device().type # privateuseone
+except:
+    torch_directml_installed = False
+
+if torch_directml_installed:
+    from contextlib import contextmanager
+    #torch.inference_mode = torch.no_grad
+
+    def empty_enter(self):
+        pass
+
+    torch.inference_mode.__enter__ = empty_enter
+    """
+    orig_inference_mode = torch.inference_mode
+    @contextmanager
+    def fake_inference_mode(mode):
+        yield
+        #try:
+        #    with torch.inference_mode(False):
+        #        yield
+        #finally:
+        #    pass
+
+    torch.inference_mode = fake_inference_mode
+    """
+
+
+    orig_torch_generator = torch.Generator
+    def new_torch_Generator(device="cpu"):
+        if torch_directml_installed and type(device) is torch.device and device.type == torch_directml_device_type:
+            device = "cpu"
+        return orig_torch_generator(device)
+    torch.Generator = new_torch_Generator
+
 # Modified ControlNetModel with encoder_attention_mask argument added