[Performance 4/6] Precompute is_sdxl_inpaint flag

modules/processing.py

@@ -115,20 +115,17 @@ def txt2img_image_conditioning(sd_model, x, width, height):
         return x.new_zeros(x.shape[0], 2*sd_model.noise_augmentor.time_embed.dim, dtype=x.dtype, device=x.device)
 
     else:
-        sd = sd_model.model.state_dict()
-        diffusion_model_input = sd.get('diffusion_model.input_blocks.0.0.weight', None)
-        if diffusion_model_input is not None:
-            if diffusion_model_input.shape[1] == 9:
-                # The "masked-image" in this case will just be all 0.5 since the entire image is masked.
-                image_conditioning = torch.ones(x.shape[0], 3, height, width, device=x.device) * 0.5
-                image_conditioning = images_tensor_to_samples(image_conditioning,
-                                                              approximation_indexes.get(opts.sd_vae_encode_method))
+        if getattr(sd_model.model, "is_sdxl_inpaint", False):
+            # The "masked-image" in this case will just be all 0.5 since the entire image is masked.
+            image_conditioning = torch.ones(x.shape[0], 3, height, width, device=x.device) * 0.5
+            image_conditioning = images_tensor_to_samples(image_conditioning,
+                                                            approximation_indexes.get(opts.sd_vae_encode_method))
 
-                # Add the fake full 1s mask to the first dimension.
-                image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0)
-                image_conditioning = image_conditioning.to(x.dtype)
+            # Add the fake full 1s mask to the first dimension.
+            image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0)
+            image_conditioning = image_conditioning.to(x.dtype)
 
-                return image_conditioning
+            return image_conditioning
 
         # Dummy zero conditioning if we're not using inpainting or unclip models.
         # Still takes up a bit of memory, but no encoder call.
@@ -390,11 +387,8 @@ def img2img_image_conditioning(self, source_image, latent_image, image_mask=None
         if self.sampler.conditioning_key == "crossattn-adm":
             return self.unclip_image_conditioning(source_image)
 
-        sd = self.sampler.model_wrap.inner_model.model.state_dict()
-        diffusion_model_input = sd.get('diffusion_model.input_blocks.0.0.weight', None)
-        if diffusion_model_input is not None:
-            if diffusion_model_input.shape[1] == 9:
-                return self.inpainting_image_conditioning(source_image, latent_image, image_mask=image_mask)
+        if getattr(self.sampler.model_wrap.inner_model.model, "is_sdxl_inpaint", False):
+            return self.inpainting_image_conditioning(source_image, latent_image, image_mask=image_mask)
 
         # Dummy zero conditioning if we're not using inpainting or depth model.
         return latent_image.new_zeros(latent_image.shape[0], 5, 1, 1)

modules/sd_models.py

@@ -380,6 +380,13 @@ def load_model_weights(model, checkpoint_info: CheckpointInfo, state_dict, timer
     model.is_sd2 = not model.is_sdxl and hasattr(model.cond_stage_model, 'model')
     model.is_sd1 = not model.is_sdxl and not model.is_sd2
     model.is_ssd = model.is_sdxl and 'model.diffusion_model.middle_block.1.transformer_blocks.0.attn1.to_q.weight' not in state_dict.keys()
+    # Set is_sdxl_inpaint flag.
+    diffusion_model_input = state_dict.get('diffusion_model.input_blocks.0.0.weight', None)
+    model.is_sdxl_inpaint = (
+        model.is_sdxl and
+        diffusion_model_input is not None and
+        diffusion_model_input.shape[1] == 9
+    )
     if model.is_sdxl:
         sd_models_xl.extend_sdxl(model)
 

modules/sd_models_xl.py

@@ -35,11 +35,10 @@ def get_learned_conditioning(self: sgm.models.diffusion.DiffusionEngine, batch:
 
 
 def apply_model(self: sgm.models.diffusion.DiffusionEngine, x, t, cond):
-    sd = self.model.state_dict()
-    diffusion_model_input = sd.get('diffusion_model.input_blocks.0.0.weight', None)
-    if diffusion_model_input is not None:
-        if diffusion_model_input.shape[1] == 9:
-            x = torch.cat([x] + cond['c_concat'], dim=1)
+    """WARNING: This function is called once per denoising iteration. DO NOT add
+    expensive functionc calls such as `model.state_dict`. """
+    if self.is_sdxl_inpaint:
+        x = torch.cat([x] + cond['c_concat'], dim=1)
 
     return self.model(x, t, cond)