reduce peak VRAM memory usage of IP adapter

invokeai/app/invocations/latent.py

@@ -50,7 +50,6 @@ from invokeai.app.invocations.primitives import DenoiseMaskOutput, ImageOutput,
 from invokeai.app.invocations.t2i_adapter import T2IAdapterField
 from invokeai.app.services.shared.invocation_context import InvocationContext
 from invokeai.app.util.controlnet_utils import prepare_control_image
-from invokeai.backend.ip_adapter.ip_adapter import IPAdapter, IPAdapterPlus
 from invokeai.backend.lora import LoRAModelRaw
 from invokeai.backend.model_manager import BaseModelType, LoadedModel
 from invokeai.backend.model_manager.config import MainConfigBase, ModelVariantType
@@ -672,6 +671,39 @@ class DenoiseLatentsInvocation(BaseInvocation):
 
         return controlnet_data
 
+    def prep_ip_adapter_image_prompts(
+        self,
+        context: InvocationContext,
+        ip_adapter: Optional[Union[IPAdapterField, list[IPAdapterField]]],
+    ) -> List[Tuple[torch.Tensor, torch.Tensor]]:
+        """Run the IPAdapter CLIPVisionModel, returning image prompt embeddings."""
+        # ip_adapter could be a list or a single IPAdapterField. Normalize to a list here.
+        if not isinstance(ip_adapter, list):
+            ip_adapter = [ip_adapter]
+
+        if len(ip_adapter) == 0:
+            return []
+
+        image_prompts = []
+        for single_ip_adapter in ip_adapter:
+            with context.models.load(single_ip_adapter.ip_adapter_model) as ip_adapter_model:
+                image_encoder_model_info = context.models.load(single_ip_adapter.image_encoder_model)
+                # `single_ip_adapter.image` could be a list or a single ImageField. Normalize to a list here.
+                single_ipa_image_fields = single_ip_adapter.image
+                if not isinstance(single_ipa_image_fields, list):
+                    single_ipa_image_fields = [single_ipa_image_fields]
+
+                single_ipa_images = [context.images.get_pil(image.image_name) for image in single_ipa_image_fields]
+                with image_encoder_model_info as image_encoder_model:
+                    assert isinstance(image_encoder_model, CLIPVisionModelWithProjection)
+                    # Get image embeddings from CLIP and ImageProjModel.
+                    image_prompt_embeds, uncond_image_prompt_embeds = ip_adapter_model.get_image_embeds(
+                        single_ipa_images, image_encoder_model
+                    )
+                    image_prompts.append((image_prompt_embeds, uncond_image_prompt_embeds))
+
+        return image_prompts
+
     def prep_ip_adapter_data(
         self,
         context: InvocationContext,
@@ -680,6 +712,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
         latent_height: int,
         latent_width: int,
         dtype: torch.dtype,
+        image_prompts: List[Tuple[torch.Tensor, torch.Tensor]],
     ) -> Optional[list[IPAdapterData]]:
         """If IP-Adapter is enabled, then this function loads the requisite models, and adds the image prompt embeddings
         to the `conditioning_data` (in-place).
@@ -696,26 +729,9 @@ class DenoiseLatentsInvocation(BaseInvocation):
 
         ip_adapter_data_list = []
         for single_ip_adapter in ip_adapter:
-            ip_adapter_model: Union[IPAdapter, IPAdapterPlus] = exit_stack.enter_context(
-                context.models.load(single_ip_adapter.ip_adapter_model)
-            )
+            ip_adapter_model = exit_stack.enter_context(context.models.load(single_ip_adapter.ip_adapter_model))
 
-            image_encoder_model_info = context.models.load(single_ip_adapter.image_encoder_model)
-            # `single_ip_adapter.image` could be a list or a single ImageField. Normalize to a list here.
-            single_ipa_image_fields = single_ip_adapter.image
-            if not isinstance(single_ipa_image_fields, list):
-                single_ipa_image_fields = [single_ipa_image_fields]
-
-            single_ipa_images = [context.images.get_pil(image.image_name) for image in single_ipa_image_fields]
-
-            # TODO(ryand): With some effort, the step of running the CLIP Vision encoder could be done before any other
-            # models are needed in memory. This would help to reduce peak memory utilization in low-memory environments.
-            with image_encoder_model_info as image_encoder_model:
-                assert isinstance(image_encoder_model, CLIPVisionModelWithProjection)
-                # Get image embeddings from CLIP and ImageProjModel.
-                image_prompt_embeds, uncond_image_prompt_embeds = ip_adapter_model.get_image_embeds(
-                    single_ipa_images, image_encoder_model
-                )
+            image_prompt_embeds, uncond_image_prompt_embeds = image_prompts.pop(0)
 
             mask = single_ip_adapter.mask
             if mask is not None:
@@ -912,6 +928,8 @@ class DenoiseLatentsInvocation(BaseInvocation):
                 do_classifier_free_guidance=True,
             )
 
+            image_prompts = self.prep_ip_adapter_image_prompts(context=context, ip_adapter=self.ip_adapter)
+
             # get the unet's config so that we can pass the base to dispatch_progress()
             unet_config = context.models.get_config(self.unet.unet.key)
 
@@ -975,6 +993,7 @@ class DenoiseLatentsInvocation(BaseInvocation):
                     latent_height=latent_height,
                     latent_width=latent_width,
                     dtype=unet.dtype,
+                    image_prompts=image_prompts,
                 )
 
                 num_inference_steps, timesteps, init_timestep, scheduler_step_kwargs = self.init_scheduler(