[torch.compile] fix graph break problems partially

src/diffusers/models/controlnet.py

@@ -817,7 +817,6 @@ def forward(
         # 6. scaling
         if guess_mode and not self.config.global_pool_conditions:
             scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device)  # 0.1 to 1.0
-
             scales = scales * conditioning_scale
             down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)]
             mid_block_res_sample = mid_block_res_sample * scales[-1]  # last one

src/diffusers/models/unet_2d_condition.py

@@ -874,9 +874,11 @@ def forward(
         forward_upsample_size = False
         upsample_size = None
 
-        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
-            # Forward upsample size to force interpolation output size.
-            forward_upsample_size = True
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
 
         # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
         # expects mask of shape:

src/diffusers/pipelines/controlnet/pipeline_controlnet.py

@@ -20,6 +20,7 @@
 import PIL.Image
 import torch
 import torch.nn.functional as F
+from packaging import version
 from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from ...image_processor import PipelineImageInput, VaeImageProcessor
@@ -976,8 +977,17 @@ def __call__(
 
         # 8. Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        is_unet_compiled = is_compiled_module(self.unet)
+        is_controlnet_compiled = is_compiled_module(self.controlnet)
+        is_torch_higher_equal_than_2_1 = version.parse(version.parse(torch.__version__).base_version) >= version.parse(
+            "2.1"
+        )
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
+                # Relevant thread:
+                # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
+                if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_than_2_1:
+                    torch._inductor.cudagraph_mark_step_begin()
                 # expand the latents if we are doing classifier free guidance
                 latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py

@@ -20,6 +20,7 @@
 import PIL.Image
 import torch
 import torch.nn.functional as F
+from packaging import version
 from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
 
 from diffusers.utils.import_utils import is_invisible_watermark_available
@@ -1144,8 +1145,17 @@ def __call__(
 
         # 8. Denoising loop
         num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        is_unet_compiled = is_compiled_module(self.unet)
+        is_controlnet_compiled = is_compiled_module(self.controlnet)
+        is_torch_higher_equal_than_2_1 = version.parse(version.parse(torch.__version__).base_version) >= version.parse(
+            "2.1"
+        )
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
+                # Relevant thread:
+                # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
+                if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_than_2_1:
+                    torch._inductor.cudagraph_mark_step_begin()
                 # expand the latents if we are doing classifier free guidance
                 latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
                 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

src/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py

@@ -814,7 +814,8 @@ def __call__(
                     encoder_hidden_states=prompt_embeds,
                     cross_attention_kwargs=cross_attention_kwargs,
                     down_intrablock_additional_residuals=[state.clone() for state in adapter_state],
-                ).sample
+                    return_dict=False,
+                )[0]
 
                 # perform guidance
                 if do_classifier_free_guidance:

src/diffusers/pipelines/versatile_diffusion/modeling_text_unet.py

@@ -1084,9 +1084,11 @@ def forward(
         forward_upsample_size = False
         upsample_size = None
 
-        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
-            # Forward upsample size to force interpolation output size.
-            forward_upsample_size = True
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
 
         # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
         # expects mask of shape: