Merge branch 'main' into pyramid-attention-broadcast

Author: a-r-r-o-w

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")

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/latte/pipeline_latte.py

@@ -31,6 +31,7 @@
 from ...utils import (
     BACKENDS_MAPPING,
     BaseOutput,
+    deprecate,
     is_bs4_available,
     is_ftfy_available,
     is_torch_xla_available,
@@ -853,6 +854,13 @@ def __call__(
 
         self._current_timestep = None
 
+        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=decode_chunk_size)
             video = self.video_processor.postprocess_video(video=video, output_type=output_type)

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()