Merge branch 'main' into improve-failure-handling-peft

Author: sayakpaul

examples/research_projects/sd3_lora_colab/train_dreambooth_lora_sd3_miniature.py

@@ -765,7 +765,7 @@ def load_model_hook(models, input_dir):
         lora_state_dict = StableDiffusion3Pipeline.lora_state_dict(input_dir)
 
         transformer_state_dict = {
-            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")
+            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
         }
         transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
         incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")

setup.py

@@ -135,6 +135,7 @@
     "transformers>=4.41.2",
     "urllib3<=2.0.0",
     "black",
+    "phonemizer",
 ]
 
 # this is a lookup table with items like:
@@ -227,6 +228,7 @@ def run(self):
     "scipy",
     "torchvision",
     "transformers",
+    "phonemizer",
 )
 extras["torch"] = deps_list("torch", "accelerate")
 

src/diffusers/dependency_versions_table.py

@@ -43,4 +43,5 @@
     "transformers": "transformers>=4.41.2",
     "urllib3": "urllib3<=2.0.0",
     "black": "black",
+    "phonemizer": "phonemizer",
 }

src/diffusers/loaders/single_file_utils.py

@@ -186,6 +186,7 @@
     "inpainting": 512,
     "inpainting_v2": 512,
     "controlnet": 512,
+    "instruct-pix2pix": 512,
     "v2": 768,
     "v1": 512,
 }

src/diffusers/pipelines/audioldm2/pipeline_audioldm2.py

@@ -237,7 +237,7 @@ def disable_vae_slicing(self):
         """
         self.vae.disable_slicing()
 
-    def enable_model_cpu_offload(self, gpu_id=0):
+    def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
         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`
@@ -249,11 +249,23 @@ def enable_model_cpu_offload(self, gpu_id=0):
         else:
             raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
 
-        device = torch.device(f"cuda:{gpu_id}")
+        torch_device = torch.device(device)
+        device_index = torch_device.index
+
+        if gpu_id is not None and device_index is not None:
+            raise ValueError(
+                f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}"
+                f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of the device: `device`={torch_device.type}"
+            )
+
+        device_type = torch_device.type
+        device = torch.device(f"{device_type}:{gpu_id or torch_device.index}")
 
         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)
+            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)
 
         model_sequence = [
             self.text_encoder.text_model,

src/diffusers/pipelines/latte/pipeline_latte.py

@@ -30,6 +30,7 @@
 from ...utils import (
     BACKENDS_MAPPING,
     BaseOutput,
+    deprecate,
     is_bs4_available,
     is_ftfy_available,
     is_torch_xla_available,
@@ -848,7 +849,14 @@ def __call__(
                 if XLA_AVAILABLE:
                     xm.mark_step()
 
-        if not output_type == "latents":
+        if output_type == "latents":
+            deprecation_message = (
+                "Passing `output_type='latents'` is deprecated. Please pass `output_type='latent'` instead."
+            )
+            deprecate("output_type_latents", "1.0.0", deprecation_message, standard_warn=False)
+            output_type = "latent"
+
+        if not output_type == "latent":
             video = self.decode_latents(latents, video_length, decode_chunk_size=14)
             video = self.video_processor.postprocess_video(video=video, output_type=output_type)
         else:

tests/pipelines/audioldm2/test_audioldm2.py

@@ -469,8 +469,8 @@ def test_xformers_attention_forwardGenerator_pass(self):
         pass
 
     def test_dict_tuple_outputs_equivalent(self):
-        # increase tolerance from 1e-4 -> 2e-4 to account for large composite model
-        super().test_dict_tuple_outputs_equivalent(expected_max_difference=2e-4)
+        # increase tolerance from 1e-4 -> 3e-4 to account for large composite model
+        super().test_dict_tuple_outputs_equivalent(expected_max_difference=3e-4)
 
     def test_inference_batch_single_identical(self):
         # increase tolerance from 1e-4 -> 2e-4 to account for large composite model

tests/single_file/single_file_testing_utils.py

@@ -47,6 +47,8 @@ def download_diffusers_config(repo_id, tmpdir):
 
 
 class SDSingleFileTesterMixin:
+    single_file_kwargs = {}
+
     def _compare_component_configs(self, pipe, single_file_pipe):
         for param_name, param_value in single_file_pipe.text_encoder.config.to_dict().items():
             if param_name in ["torch_dtype", "architectures", "_name_or_path"]:
@@ -154,7 +156,7 @@ def test_single_file_components_with_original_config_local_files_only(
         self._compare_component_configs(pipe, single_file_pipe)
 
     def test_single_file_format_inference_is_same_as_pretrained(self, expected_max_diff=1e-4):
-        sf_pipe = self.pipeline_class.from_single_file(self.ckpt_path, safety_checker=None)
+        sf_pipe = self.pipeline_class.from_single_file(self.ckpt_path, safety_checker=None, **self.single_file_kwargs)
         sf_pipe.unet.set_attn_processor(AttnProcessor())
         sf_pipe.enable_model_cpu_offload(device=torch_device)
 
@@ -170,7 +172,7 @@ def test_single_file_format_inference_is_same_as_pretrained(self, expected_max_d
 
         max_diff = numpy_cosine_similarity_distance(image.flatten(), image_single_file.flatten())
 
-        assert max_diff < expected_max_diff
+        assert max_diff < expected_max_diff, f"{image.flatten()} != {image_single_file.flatten()}"
 
     def test_single_file_components_with_diffusers_config(
         self,

tests/single_file/test_stable_diffusion_single_file.py

@@ -132,6 +132,7 @@ class StableDiffusionInstructPix2PixPipelineSingleFileSlowTests(unittest.TestCas
         "https://raw.githubusercontent.com/timothybrooks/instruct-pix2pix/refs/heads/main/configs/generate.yaml"
     )
     repo_id = "timbrooks/instruct-pix2pix"
+    single_file_kwargs = {"extract_ema": True}
 
     def setUp(self):
         super().setUp()