diffusers 0.27.1__py3-none-any.whl → 0.32.2__py3-none-any.whl

diffusers/utils/testing_utils.py

@@ -1,5 +1,6 @@
 import functools
 import importlib
+import importlib.metadata
 import inspect
 import io
 import logging
@@ -14,7 +15,6 @@ import time
 import unittest
 import urllib.parse
 from contextlib import contextmanager
-from distutils.util import strtobool
 from io import BytesIO, StringIO
 from pathlib import Path
 from typing import Callable, Dict, List, Optional, Union
@@ -28,14 +28,19 @@ from packaging import version
 
 from .import_utils import (
     BACKENDS_MAPPING,
+    is_accelerate_available,
+    is_bitsandbytes_available,
     is_compel_available,
     is_flax_available,
+    is_gguf_available,
     is_note_seq_available,
     is_onnx_available,
     is_opencv_available,
     is_peft_available,
+    is_timm_available,
     is_torch_available,
     is_torch_version,
+    is_torchao_available,
     is_torchsde_available,
     is_transformers_available,
 )
@@ -54,6 +59,7 @@ _required_transformers_version = is_transformers_available() and version.parse(
 ) > version.parse("4.33")
 
 USE_PEFT_BACKEND = _required_peft_version and _required_transformers_version
+BIG_GPU_MEMORY = int(os.getenv("BIG_GPU_MEMORY", 40))
 
 if is_torch_available():
     import torch
@@ -106,10 +112,21 @@ def numpy_cosine_similarity_distance(a, b):
     return distance
 
 
-def print_tensor_test(tensor, filename="test_corrections.txt", expected_tensor_name="expected_slice"):
+def print_tensor_test(
+    tensor,
+    limit_to_slices=None,
+    max_torch_print=None,
+    filename="test_corrections.txt",
+    expected_tensor_name="expected_slice",
+):
+    if max_torch_print:
+        torch.set_printoptions(threshold=10_000)
+
     test_name = os.environ.get("PYTEST_CURRENT_TEST")
     if not torch.is_tensor(tensor):
         tensor = torch.from_numpy(tensor)
+    if limit_to_slices:
+        tensor = tensor[0, -3:, -3:, -1]
 
     tensor_str = str(tensor.detach().cpu().flatten().to(torch.float32)).replace("\n", "")
     # format is usually:
@@ -118,7 +135,7 @@ def print_tensor_test(tensor, filename="test_corrections.txt", expected_tensor_n
     test_file, test_class, test_fn = test_name.split("::")
     test_fn = test_fn.split()[0]
     with open(filename, "a") as f:
-        print(";".join([test_file, test_class, test_fn, output_str]), file=f)
+        print("::".join([test_file, test_class, test_fn, output_str]), file=f)
 
 
 def get_tests_dir(append_path=None):
@@ -142,6 +159,22 @@ def get_tests_dir(append_path=None):
         return tests_dir
 
 
+# Taken from the following PR:
+# https://github.com/huggingface/accelerate/pull/1964
+def str_to_bool(value) -> int:
+    """
+    Converts a string representation of truth to `True` (1) or `False` (0). True values are `y`, `yes`, `t`, `true`,
+    `on`, and `1`; False value are `n`, `no`, `f`, `false`, `off`, and `0`;
+    """
+    value = value.lower()
+    if value in ("y", "yes", "t", "true", "on", "1"):
+        return 1
+    elif value in ("n", "no", "f", "false", "off", "0"):
+        return 0
+    else:
+        raise ValueError(f"invalid truth value {value}")
+
+
 def parse_flag_from_env(key, default=False):
     try:
         value = os.environ[key]
@@ -151,7 +184,7 @@ def parse_flag_from_env(key, default=False):
     else:
         # KEY is set, convert it to True or False.
         try:
-            _value = strtobool(value)
+            _value = str_to_bool(value)
         except ValueError:
             # More values are supported, but let's keep the message simple.
             raise ValueError(f"If set, {key} must be yes or no.")
@@ -160,6 +193,7 @@ def parse_flag_from_env(key, default=False):
 
 _run_slow_tests = parse_flag_from_env("RUN_SLOW", default=False)
 _run_nightly_tests = parse_flag_from_env("RUN_NIGHTLY", default=False)
+_run_compile_tests = parse_flag_from_env("RUN_COMPILE", default=False)
 
 
 def floats_tensor(shape, scale=1.0, rng=None, name=None):
@@ -198,6 +232,16 @@ def nightly(test_case):
     return unittest.skipUnless(_run_nightly_tests, "test is nightly")(test_case)
 
 
+def is_torch_compile(test_case):
+    """
+    Decorator marking a test that runs compile tests in the diffusers CI.
+
+    Compile tests are skipped by default. Set the RUN_COMPILE environment variable to a truthy value to run them.
+
+    """
+    return unittest.skipUnless(_run_compile_tests, "test is torch compile")(test_case)
+
+
 def require_torch(test_case):
     """
     Decorator marking a test that requires PyTorch. These tests are skipped when PyTorch isn't installed.
@@ -214,6 +258,18 @@ def require_torch_2(test_case):
     )
 
 
+def require_torch_version_greater_equal(torch_version):
+    """Decorator marking a test that requires torch with a specific version or greater."""
+
+    def decorator(test_case):
+        correct_torch_version = is_torch_available() and is_torch_version(">=", torch_version)
+        return unittest.skipUnless(
+            correct_torch_version, f"test requires torch with the version greater than or equal to {torch_version}"
+        )(test_case)
+
+    return decorator
+
+
 def require_torch_gpu(test_case):
     """Decorator marking a test that requires CUDA and PyTorch."""
     return unittest.skipUnless(is_torch_available() and torch_device == "cuda", "test requires PyTorch+CUDA")(
@@ -229,6 +285,20 @@ def require_torch_accelerator(test_case):
     )
 
 
+def require_torch_multi_gpu(test_case):
+    """
+    Decorator marking a test that requires a multi-GPU setup (in PyTorch). These tests are skipped on a machine without
+    multiple GPUs. To run *only* the multi_gpu tests, assuming all test names contain multi_gpu: $ pytest -sv ./tests
+    -k "multi_gpu"
+    """
+    if not is_torch_available():
+        return unittest.skip("test requires PyTorch")(test_case)
+
+    import torch
+
+    return unittest.skipUnless(torch.cuda.device_count() > 1, "test requires multiple GPUs")(test_case)
+
+
 def require_torch_accelerator_with_fp16(test_case):
     """Decorator marking a test that requires an accelerator with support for the FP16 data type."""
     return unittest.skipUnless(_is_torch_fp16_available(torch_device), "test requires accelerator with fp16 support")(
@@ -243,6 +313,26 @@ def require_torch_accelerator_with_fp64(test_case):
     )
 
 
+def require_big_gpu_with_torch_cuda(test_case):
+    """
+    Decorator marking a test that requires a bigger GPU (24GB) for execution. Some example pipelines: Flux, SD3, Cog,
+    etc.
+    """
+    if not is_torch_available():
+        return unittest.skip("test requires PyTorch")(test_case)
+
+    import torch
+
+    if not torch.cuda.is_available():
+        return unittest.skip("test requires PyTorch CUDA")(test_case)
+
+    device_properties = torch.cuda.get_device_properties(0)
+    total_memory = device_properties.total_memory / (1024**3)
+    return unittest.skipUnless(
+        total_memory >= BIG_GPU_MEMORY, f"test requires a GPU with at least {BIG_GPU_MEMORY} GB memory"
+    )(test_case)
+
+
 def require_torch_accelerator_with_training(test_case):
     """Decorator marking a test that requires an accelerator with support for training."""
     return unittest.skipUnless(
@@ -285,6 +375,14 @@ def require_note_seq(test_case):
     return unittest.skipUnless(is_note_seq_available(), "test requires note_seq")(test_case)
 
 
+def require_accelerator(test_case):
+    """
+    Decorator marking a test that requires a hardware accelerator backend. These tests are skipped when there are no
+    hardware accelerator available.
+    """
+    return unittest.skipUnless(torch_device != "cpu", "test requires a hardware accelerator")(test_case)
+
+
 def require_torchsde(test_case):
     """
     Decorator marking a test that requires torchsde. These tests are skipped when torchsde isn't installed.
@@ -300,6 +398,27 @@ def require_peft_backend(test_case):
     return unittest.skipUnless(USE_PEFT_BACKEND, "test requires PEFT backend")(test_case)
 
 
+def require_timm(test_case):
+    """
+    Decorator marking a test that requires timm. These tests are skipped when timm isn't installed.
+    """
+    return unittest.skipUnless(is_timm_available(), "test requires timm")(test_case)
+
+
+def require_bitsandbytes(test_case):
+    """
+    Decorator marking a test that requires bitsandbytes. These tests are skipped when bitsandbytes isn't installed.
+    """
+    return unittest.skipUnless(is_bitsandbytes_available(), "test requires bitsandbytes")(test_case)
+
+
+def require_accelerate(test_case):
+    """
+    Decorator marking a test that requires accelerate. These tests are skipped when accelerate isn't installed.
+    """
+    return unittest.skipUnless(is_accelerate_available(), "test requires accelerate")(test_case)
+
+
 def require_peft_version_greater(peft_version):
     """
     Decorator marking a test that requires PEFT backend with a specific version, this would require some specific
@@ -317,6 +436,72 @@ def require_peft_version_greater(peft_version):
     return decorator
 
 
+def require_transformers_version_greater(transformers_version):
+    """
+    Decorator marking a test that requires transformers with a specific version, this would require some specific
+    versions of PEFT and transformers.
+    """
+
+    def decorator(test_case):
+        correct_transformers_version = is_transformers_available() and version.parse(
+            version.parse(importlib.metadata.version("transformers")).base_version
+        ) > version.parse(transformers_version)
+        return unittest.skipUnless(
+            correct_transformers_version,
+            f"test requires transformers with the version greater than {transformers_version}",
+        )(test_case)
+
+    return decorator
+
+
+def require_accelerate_version_greater(accelerate_version):
+    def decorator(test_case):
+        correct_accelerate_version = is_accelerate_available() and version.parse(
+            version.parse(importlib.metadata.version("accelerate")).base_version
+        ) > version.parse(accelerate_version)
+        return unittest.skipUnless(
+            correct_accelerate_version, f"Test requires accelerate with the version greater than {accelerate_version}."
+        )(test_case)
+
+    return decorator
+
+
+def require_bitsandbytes_version_greater(bnb_version):
+    def decorator(test_case):
+        correct_bnb_version = is_bitsandbytes_available() and version.parse(
+            version.parse(importlib.metadata.version("bitsandbytes")).base_version
+        ) > version.parse(bnb_version)
+        return unittest.skipUnless(
+            correct_bnb_version, f"Test requires bitsandbytes with the version greater than {bnb_version}."
+        )(test_case)
+
+    return decorator
+
+
+def require_gguf_version_greater_or_equal(gguf_version):
+    def decorator(test_case):
+        correct_gguf_version = is_gguf_available() and version.parse(
+            version.parse(importlib.metadata.version("gguf")).base_version
+        ) >= version.parse(gguf_version)
+        return unittest.skipUnless(
+            correct_gguf_version, f"Test requires gguf with the version greater than {gguf_version}."
+        )(test_case)
+
+    return decorator
+
+
+def require_torchao_version_greater_or_equal(torchao_version):
+    def decorator(test_case):
+        correct_torchao_version = is_torchao_available() and version.parse(
+            version.parse(importlib.metadata.version("torchao")).base_version
+        ) >= version.parse(torchao_version)
+        return unittest.skipUnless(
+            correct_torchao_version, f"Test requires torchao with version greater than {torchao_version}."
+        )(test_case)
+
+    return decorator
+
+
 def deprecate_after_peft_backend(test_case):
     """
     Decorator marking a test that will be skipped after PEFT backend
@@ -324,13 +509,10 @@ def deprecate_after_peft_backend(test_case):
     return unittest.skipUnless(not USE_PEFT_BACKEND, "test skipped in favor of PEFT backend")(test_case)
 
 
-def require_python39_or_higher(test_case):
-    def python39_available():
-        sys_info = sys.version_info
-        major, minor = sys_info.major, sys_info.minor
-        return major == 3 and minor >= 9
-
-    return unittest.skipUnless(python39_available(), "test requires Python 3.9 or higher")(test_case)
+def get_python_version():
+    sys_info = sys.version_info
+    major, minor = sys_info.major, sys_info.minor
+    return major, minor
 
 
 def load_numpy(arry: Union[str, np.ndarray], local_path: Optional[str] = None) -> np.ndarray:

diffusers/utils/torch_utils.py

@@ -14,6 +14,7 @@
 """
 PyTorch utilities: Utilities related to PyTorch
 """
+
 from typing import List, Optional, Tuple, Union
 
 from . import logging
@@ -101,6 +102,9 @@ def fourier_filter(x_in: "torch.Tensor", threshold: int, scale: int) -> "torch.T
     # Non-power of 2 images must be float32
     if (W & (W - 1)) != 0 or (H & (H - 1)) != 0:
         x = x.to(dtype=torch.float32)
+    # fftn does not support bfloat16
+    elif x.dtype == torch.bfloat16:
+        x = x.to(dtype=torch.float32)
 
     # FFT
     x_freq = fftn(x, dim=(-2, -1))

diffusers/video_processor.py

@@ -0,0 +1,113 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from typing import List, Optional, Union
+
+import numpy as np
+import PIL
+import torch
+
+from .image_processor import VaeImageProcessor, is_valid_image, is_valid_image_imagelist
+
+
+class VideoProcessor(VaeImageProcessor):
+    r"""Simple video processor."""
+
+    def preprocess_video(self, video, height: Optional[int] = None, width: Optional[int] = None) -> torch.Tensor:
+        r"""
+        Preprocesses input video(s).
+
+        Args:
+            video (`List[PIL.Image]`, `List[List[PIL.Image]]`, `torch.Tensor`, `np.array`, `List[torch.Tensor]`, `List[np.array]`):
+                The input video. It can be one of the following:
+                * List of the PIL images.
+                * List of list of PIL images.
+                * 4D Torch tensors (expected shape for each tensor `(num_frames, num_channels, height, width)`).
+                * 4D NumPy arrays (expected shape for each array `(num_frames, height, width, num_channels)`).
+                * List of 4D Torch tensors (expected shape for each tensor `(num_frames, num_channels, height,
+                  width)`).
+                * List of 4D NumPy arrays (expected shape for each array `(num_frames, height, width, num_channels)`).
+                * 5D NumPy arrays: expected shape for each array `(batch_size, num_frames, height, width,
+                  num_channels)`.
+                * 5D Torch tensors: expected shape for each array `(batch_size, num_frames, num_channels, height,
+                  width)`.
+            height (`int`, *optional*, defaults to `None`):
+                The height in preprocessed frames of the video. If `None`, will use the `get_default_height_width()` to
+                get default height.
+            width (`int`, *optional*`, defaults to `None`):
+                The width in preprocessed frames of the video. If `None`, will use get_default_height_width()` to get
+                the default width.
+        """
+        if isinstance(video, list) and isinstance(video[0], np.ndarray) and video[0].ndim == 5:
+            warnings.warn(
+                "Passing `video` as a list of 5d np.ndarray is deprecated."
+                "Please concatenate the list along the batch dimension and pass it as a single 5d np.ndarray",
+                FutureWarning,
+            )
+            video = np.concatenate(video, axis=0)
+        if isinstance(video, list) and isinstance(video[0], torch.Tensor) and video[0].ndim == 5:
+            warnings.warn(
+                "Passing `video` as a list of 5d torch.Tensor is deprecated."
+                "Please concatenate the list along the batch dimension and pass it as a single 5d torch.Tensor",
+                FutureWarning,
+            )
+            video = torch.cat(video, axis=0)
+
+        # ensure the input is a list of videos:
+        # - if it is a batch of videos (5d torch.Tensor or np.ndarray), it is converted to a list of videos (a list of 4d torch.Tensor or np.ndarray)
+        # - if it is is a single video, it is convereted to a list of one video.
+        if isinstance(video, (np.ndarray, torch.Tensor)) and video.ndim == 5:
+            video = list(video)
+        elif isinstance(video, list) and is_valid_image(video[0]) or is_valid_image_imagelist(video):
+            video = [video]
+        elif isinstance(video, list) and is_valid_image_imagelist(video[0]):
+            video = video
+        else:
+            raise ValueError(
+                "Input is in incorrect format. Currently, we only support numpy.ndarray, torch.Tensor, PIL.Image.Image"
+            )
+
+        video = torch.stack([self.preprocess(img, height=height, width=width) for img in video], dim=0)
+
+        # move the number of channels before the number of frames.
+        video = video.permute(0, 2, 1, 3, 4)
+
+        return video
+
+    def postprocess_video(
+        self, video: torch.Tensor, output_type: str = "np"
+    ) -> Union[np.ndarray, torch.Tensor, List[PIL.Image.Image]]:
+        r"""
+        Converts a video tensor to a list of frames for export.
+
+        Args:
+            video (`torch.Tensor`): The video as a tensor.
+            output_type (`str`, defaults to `"np"`): Output type of the postprocessed `video` tensor.
+        """
+        batch_size = video.shape[0]
+        outputs = []
+        for batch_idx in range(batch_size):
+            batch_vid = video[batch_idx].permute(1, 0, 2, 3)
+            batch_output = self.postprocess(batch_vid, output_type)
+            outputs.append(batch_output)
+
+        if output_type == "np":
+            outputs = np.stack(outputs)
+        elif output_type == "pt":
+            outputs = torch.stack(outputs)
+        elif not output_type == "pil":
+            raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']")
+
+        return outputs