Merge branch 'main' into compile_utils

Author: sayakpaul

src/diffusers/models/modeling_utils.py

@@ -1421,8 +1421,10 @@ def compile_repeated_blocks(self, *args, **kwargs):
         class name matches will be compiled.
 
         2. **`_no_split_modules`** – Fallback.  If the preferred attribute is
-        missing or empty, we fall back to the legacy Diffusers attribute
-        `_no_split_modules`.
+        missing or empty, we fall back to the Diffusers attribute
+        `_no_split_modules`. The original purpose of this attribute is to
+        prevent splitting when device_map is present, but they can also act as a
+        good proxy for repeated blocks.
 
         Once discovered, each matching sub-module is compiled by calling
         `submodule.compile(*args, **kwargs)`.  Any positional or keyword
@@ -1445,8 +1447,8 @@ class name matches will be compiled.
         has_compiled_region = False
         for submod in self.modules():
             if submod.__class__.__name__ in repeated_blocks:
-                has_compiled_region = True
                 submod.compile(*args, **kwargs)
+                has_compiled_region = True
 
         if not has_compiled_region:
             raise ValueError(

src/diffusers/models/transformers/transformer_ltx.py

@@ -328,6 +328,9 @@ class LTXVideoTransformer3DModel(ModelMixin, ConfigMixin, FromOriginalModelMixin
 
     _supports_gradient_checkpointing = True
     _skip_layerwise_casting_patterns = ["norm"]
+    _repeated_blocks = [
+        "LTXVideoTransformerBlock",
+    ]
 
     @register_to_config
     def __init__(

src/diffusers/models/unets/unet_2d_condition.py

@@ -167,6 +167,9 @@ class conditioning with `class_embed_type` equal to `None`.
     _supports_gradient_checkpointing = True
     _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D", "CrossAttnUpBlock2D"]
     _skip_layerwise_casting_patterns = ["norm"]
+    _repeated_blocks = [
+        "BasicTransformerBlock",
+    ]
 
     @register_to_config
     def __init__(

src/diffusers/pipelines/audioldm2/pipeline_audioldm2.py

@@ -41,7 +41,7 @@
     replace_example_docstring,
 )
 from ...utils.import_utils import is_transformers_version
-from ...utils.torch_utils import randn_tensor
+from ...utils.torch_utils import empty_device_cache, randn_tensor
 from ..pipeline_utils import AudioPipelineOutput, DiffusionPipeline
 from .modeling_audioldm2 import AudioLDM2ProjectionModel, AudioLDM2UNet2DConditionModel
 
@@ -267,9 +267,7 @@ def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[t
 
         if self.device.type != "cpu":
             self.to("cpu", silence_dtype_warnings=True)
-            device_mod = getattr(torch, device.type, None)
-            if hasattr(device_mod, "empty_cache") and device_mod.is_available():
-                device_mod.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+            empty_device_cache(device.type)
 
         model_sequence = [
             self.text_encoder.text_model,

src/diffusers/pipelines/consisid/consisid_utils.py

@@ -294,7 +294,7 @@ def prepare_face_models(model_path, device, dtype):
 
     Parameters:
     - model_path: Path to the directory containing model files.
-    - device: The device (e.g., 'cuda', 'cpu') where models will be loaded.
+    - device: The device (e.g., 'cuda', 'xpu', 'cpu') where models will be loaded.
     - dtype: Data type (e.g., torch.float32) for model inference.
 
     Returns:

src/diffusers/pipelines/controlnet/pipeline_controlnet.py

@@ -37,7 +37,7 @@
     scale_lora_layers,
     unscale_lora_layers,
 )
-from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
+from ...utils.torch_utils import empty_device_cache, is_compiled_module, is_torch_version, randn_tensor
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
@@ -1339,7 +1339,7 @@ def __call__(
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.unet.to("cpu")
             self.controlnet.to("cpu")
-            torch.cuda.empty_cache()
+            empty_device_cache()
 
         if not output_type == "latent":
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[

src/diffusers/pipelines/controlnet/pipeline_controlnet_img2img.py

@@ -36,7 +36,7 @@
     scale_lora_layers,
     unscale_lora_layers,
 )
-from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ...utils.torch_utils import empty_device_cache, is_compiled_module, randn_tensor
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionPipelineOutput
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
@@ -1311,7 +1311,7 @@ def __call__(
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.unet.to("cpu")
             self.controlnet.to("cpu")
-            torch.cuda.empty_cache()
+            empty_device_cache()
 
         if not output_type == "latent":
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[

src/diffusers/pipelines/controlnet/pipeline_controlnet_inpaint.py

@@ -38,7 +38,7 @@
     scale_lora_layers,
     unscale_lora_layers,
 )
-from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ...utils.torch_utils import empty_device_cache, is_compiled_module, randn_tensor
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion import StableDiffusionPipelineOutput
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
@@ -1500,7 +1500,7 @@ def __call__(
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.unet.to("cpu")
             self.controlnet.to("cpu")
-            torch.cuda.empty_cache()
+            empty_device_cache()
 
         if not output_type == "latent":
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[

src/diffusers/pipelines/controlnet/pipeline_controlnet_inpaint_sd_xl.py

@@ -51,7 +51,7 @@
     scale_lora_layers,
     unscale_lora_layers,
 )
-from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ...utils.torch_utils import empty_device_cache, is_compiled_module, randn_tensor
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 
@@ -1858,7 +1858,7 @@ def denoising_value_valid(dnv):
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.unet.to("cpu")
             self.controlnet.to("cpu")
-            torch.cuda.empty_cache()
+            empty_device_cache()
 
         if not output_type == "latent":
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]

src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py

@@ -1465,7 +1465,11 @@ def __call__(
 
                 # 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_2_1:
+                if (
+                    torch.cuda.is_available()
+                    and (is_unet_compiled and is_controlnet_compiled)
+                    and is_torch_higher_equal_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 self.do_classifier_free_guidance else latents

src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py

@@ -53,7 +53,7 @@
     scale_lora_layers,
     unscale_lora_layers,
 )
-from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ...utils.torch_utils import empty_device_cache, is_compiled_module, randn_tensor
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 
@@ -921,7 +921,7 @@ def prepare_latents(
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
-            torch.cuda.empty_cache()
+            empty_device_cache()
 
         image = image.to(device=device, dtype=dtype)
 
@@ -1632,7 +1632,7 @@ def __call__(
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.unet.to("cpu")
             self.controlnet.to("cpu")
-            torch.cuda.empty_cache()
+            empty_device_cache()
 
         if not output_type == "latent":
             # make sure the VAE is in float32 mode, as it overflows in float16

src/diffusers/pipelines/controlnet/pipeline_controlnet_union_inpaint_sd_xl.py

@@ -51,7 +51,7 @@
     scale_lora_layers,
     unscale_lora_layers,
 )
-from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ...utils.torch_utils import empty_device_cache, is_compiled_module, randn_tensor
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 
@@ -1766,7 +1766,7 @@ def denoising_value_valid(dnv):
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.unet.to("cpu")
             self.controlnet.to("cpu")
-            torch.cuda.empty_cache()
+            empty_device_cache()
 
         if not output_type == "latent":
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]

src/diffusers/pipelines/controlnet/pipeline_controlnet_union_sd_xl_img2img.py

@@ -53,7 +53,7 @@
     scale_lora_layers,
     unscale_lora_layers,
 )
-from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ...utils.torch_utils import empty_device_cache, is_compiled_module, randn_tensor
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
 
@@ -876,7 +876,7 @@ def prepare_latents(
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
-            torch.cuda.empty_cache()
+            empty_device_cache()
 
         image = image.to(device=device, dtype=dtype)
 
@@ -1574,7 +1574,7 @@ def __call__(
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.unet.to("cpu")
             self.controlnet.to("cpu")
-            torch.cuda.empty_cache()
+            empty_device_cache()
 
         if not output_type == "latent":
             # make sure the VAE is in float32 mode, as it overflows in float16

src/diffusers/pipelines/controlnet_xs/pipeline_controlnet_xs.py

@@ -36,7 +36,7 @@
     scale_lora_layers,
     unscale_lora_layers,
 )
-from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
+from ...utils.torch_utils import empty_device_cache, is_compiled_module, is_torch_version, randn_tensor
 from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
@@ -853,7 +853,7 @@ def __call__(
             for i, t in enumerate(timesteps):
                 # Relevant thread:
                 # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
-                if is_controlnet_compiled and is_torch_higher_equal_2_1:
+                if torch.cuda.is_available() and is_controlnet_compiled and is_torch_higher_equal_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
@@ -902,7 +902,7 @@ def __call__(
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.unet.to("cpu")
             self.controlnet.to("cpu")
-            torch.cuda.empty_cache()
+            empty_device_cache()
 
         if not output_type == "latent":
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[

src/diffusers/pipelines/kandinsky/pipeline_kandinsky_combined.py

@@ -193,7 +193,7 @@ def __init__(
     def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
         self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
 
-    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None):
         r"""
         Offloads all models (`unet`, `text_encoder`, `vae`, and `safety checker` state dicts) to CPU using 🤗
         Accelerate, significantly reducing memory usage. Models are moved to a `torch.device('meta')` and loaded on a
@@ -411,7 +411,7 @@ def __init__(
     def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
         self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
 
-    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None):
         r"""
         Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
         text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
@@ -652,7 +652,7 @@ def __init__(
     def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
         self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
 
-    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None):
         r"""
         Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
         text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a

src/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py

@@ -179,7 +179,7 @@ def __init__(
     def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
         self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
 
-    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None):
         r"""
         Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
         text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
@@ -407,7 +407,7 @@ def __init__(
     def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
         self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
 
-    def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+    def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None):
         r"""
         Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
         to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
@@ -417,7 +417,7 @@ def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[t
         self.prior_pipe.enable_model_cpu_offload(gpu_id=gpu_id, device=device)
         self.decoder_pipe.enable_model_cpu_offload(gpu_id=gpu_id, device=device)
 
-    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None):
         r"""
         Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
         text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
@@ -656,7 +656,7 @@ def __init__(
     def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
         self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
 
-    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+    def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = None):
         r"""
         Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
         text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a

src/diffusers/pipelines/kolors/pipeline_kolors_img2img.py

@@ -25,7 +25,7 @@
 from ...models.attention_processor import AttnProcessor2_0, FusedAttnProcessor2_0, XFormersAttnProcessor
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import is_torch_xla_available, logging, replace_example_docstring
-from ...utils.torch_utils import randn_tensor
+from ...utils.torch_utils import empty_device_cache, randn_tensor
 from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
 from .pipeline_output import KolorsPipelineOutput
 from .text_encoder import ChatGLMModel
@@ -618,7 +618,7 @@ def prepare_latents(
         # Offload text encoder if `enable_model_cpu_offload` was enabled
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
             self.text_encoder_2.to("cpu")
-            torch.cuda.empty_cache()
+            empty_device_cache()
 
         image = image.to(device=device, dtype=dtype)
 

src/diffusers/pipelines/musicldm/pipeline_musicldm.py

@@ -35,7 +35,7 @@
     logging,
     replace_example_docstring,
 )
-from ...utils.torch_utils import randn_tensor
+from ...utils.torch_utils import empty_device_cache, get_device, randn_tensor
 from ..pipeline_utils import AudioPipelineOutput, DeprecatedPipelineMixin, DiffusionPipeline, StableDiffusionMixin
 
 
@@ -397,20 +397,22 @@ def prepare_latents(self, batch_size, num_channels_latents, height, dtype, devic
     def enable_model_cpu_offload(self, gpu_id=0):
         r"""
         Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
-        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
-        method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
-        `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+        to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the accelerator when its
+        `forward` method is called, and the model remains in accelerator until the next model runs. Memory savings are
+        lower than with `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution
+        of the `unet`.
         """
         if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
             from accelerate import cpu_offload_with_hook
         else:
             raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
 
-        device = torch.device(f"cuda:{gpu_id}")
+        device_type = get_device()
+        device = torch.device(f"{device_type}:{gpu_id}")
 
         if self.device.type != "cpu":
             self.to("cpu", silence_dtype_warnings=True)
-            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+            empty_device_cache()  # otherwise we don't see the memory savings (but they probably exist)
 
         model_sequence = [
             self.text_encoder.text_model,