Merge branch 'main' into skip-better-2

Author: sayakpaul

src/diffusers/hooks/group_offloading.py

@@ -96,9 +96,6 @@ def __init__(
         else:
             self.cpu_param_dict = self._init_cpu_param_dict()
 
-        if self.stream is None and self.record_stream:
-            raise ValueError("`record_stream` cannot be True when `stream` is None.")
-
     def _init_cpu_param_dict(self):
         cpu_param_dict = {}
         if self.stream is None:
@@ -513,6 +510,9 @@ def apply_group_offloading(
         else:
             raise ValueError("Using streams for data transfer requires a CUDA device, or an Intel XPU device.")
 
+    if not use_stream and record_stream:
+        raise ValueError("`record_stream` cannot be True when `use_stream=False`.")
+
     _raise_error_if_accelerate_model_or_sequential_hook_present(module)
 
     if offload_type == "block_level":

src/diffusers/models/autoencoders/autoencoder_kl_cosmos.py

@@ -110,8 +110,11 @@ def __init__(self, patch_size: int = 1, patch_method: str = "haar") -> None:
         self.patch_size = patch_size
         self.patch_method = patch_method
 
-        self.register_buffer("wavelets", _WAVELETS[patch_method], persistent=False)
-        self.register_buffer("_arange", torch.arange(_WAVELETS[patch_method].shape[0]), persistent=False)
+        wavelets = _WAVELETS.get(patch_method).clone()
+        arange = torch.arange(wavelets.shape[0])
+
+        self.register_buffer("wavelets", wavelets, persistent=False)
+        self.register_buffer("_arange", arange, persistent=False)
 
     def _dwt(self, hidden_states: torch.Tensor, mode: str = "reflect", rescale=False) -> torch.Tensor:
         dtype = hidden_states.dtype
@@ -185,12 +188,11 @@ def __init__(self, patch_size: int = 1, patch_method: str = "haar"):
         self.patch_size = patch_size
         self.patch_method = patch_method
 
-        self.register_buffer("wavelets", _WAVELETS[patch_method], persistent=False)
-        self.register_buffer(
-            "_arange",
-            torch.arange(_WAVELETS[patch_method].shape[0]),
-            persistent=False,
-        )
+        wavelets = _WAVELETS.get(patch_method).clone()
+        arange = torch.arange(wavelets.shape[0])
+
+        self.register_buffer("wavelets", wavelets, persistent=False)
+        self.register_buffer("_arange", arange, persistent=False)
 
     def _idwt(self, hidden_states: torch.Tensor, rescale: bool = False) -> torch.Tensor:
         device = hidden_states.device

src/diffusers/pipelines/omnigen/pipeline_omnigen.py

@@ -23,12 +23,14 @@
 from ...models.autoencoders import AutoencoderKL
 from ...models.transformers import OmniGenTransformer2DModel
 from ...schedulers import FlowMatchEulerDiscreteScheduler
-from ...utils import is_torch_xla_available, logging, replace_example_docstring
+from ...utils import is_torch_xla_available, is_torchvision_available, logging, replace_example_docstring
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
-from .processor_omnigen import OmniGenMultiModalProcessor
 
 
+if is_torchvision_available():
+    from .processor_omnigen import OmniGenMultiModalProcessor
+
 if is_torch_xla_available():
     XLA_AVAILABLE = True
 else:

src/diffusers/pipelines/omnigen/processor_omnigen.py

@@ -18,7 +18,12 @@
 import numpy as np
 import torch
 from PIL import Image
-from torchvision import transforms
+
+from ...utils import is_torchvision_available
+
+
+if is_torchvision_available():
+    from torchvision import transforms
 
 
 def crop_image(pil_image, max_image_size):

tests/models/test_modeling_common.py

@@ -1528,14 +1528,16 @@ def test_fn(storage_dtype, compute_dtype):
         test_fn(torch.float8_e5m2, torch.float32)
         test_fn(torch.float8_e4m3fn, torch.bfloat16)
 
+    @torch.no_grad()
     def test_layerwise_casting_inference(self):
         from diffusers.hooks.layerwise_casting import DEFAULT_SKIP_MODULES_PATTERN, SUPPORTED_PYTORCH_LAYERS
 
         torch.manual_seed(0)
         config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
-        model = self.model_class(**config).eval()
-        model = model.to(torch_device)
-        base_slice = model(**inputs_dict)[0].flatten().detach().cpu().numpy()
+        model = self.model_class(**config)
+        model.eval()
+        model.to(torch_device)
+        base_slice = model(**inputs_dict)[0].detach().flatten().cpu().numpy()
 
         def check_linear_dtype(module, storage_dtype, compute_dtype):
             patterns_to_check = DEFAULT_SKIP_MODULES_PATTERN
@@ -1573,6 +1575,7 @@ def test_layerwise_casting(storage_dtype, compute_dtype):
         test_layerwise_casting(torch.float8_e4m3fn, torch.bfloat16)
 
     @require_torch_accelerator
+    @torch.no_grad()
     def test_layerwise_casting_memory(self):
         MB_TOLERANCE = 0.2
         LEAST_COMPUTE_CAPABILITY = 8.0
@@ -1706,10 +1709,6 @@ def test_group_offloading_with_disk(self, record_stream, offload_type):
         if not self.model_class._supports_group_offloading:
             pytest.skip("Model does not support group offloading.")
 
-        torch.manual_seed(0)
-        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
-        model = self.model_class(**init_dict)
-
         torch.manual_seed(0)
         init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
         model = self.model_class(**init_dict)
@@ -1725,7 +1724,7 @@ def test_group_offloading_with_disk(self, record_stream, offload_type):
                 **additional_kwargs,
             )
             has_safetensors = glob.glob(f"{tmpdir}/*.safetensors")
-            assert has_safetensors, "No safetensors found in the directory."
+            self.assertTrue(len(has_safetensors) > 0, "No safetensors found in the offload directory.")
             _ = model(**inputs_dict)[0]
 
     def test_auto_model(self, expected_max_diff=5e-5):