diffusers 0.33.1__py3-none-any.whl → 0.35.0__py3-none-any.whl

diffusers/quantizers/pipe_quant_config.py

@@ -0,0 +1,202 @@
+# Copyright 2025 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 inspect
+from typing import Dict, List, Optional, Union
+
+from ..utils import is_transformers_available, logging
+from .quantization_config import QuantizationConfigMixin as DiffQuantConfigMixin
+
+
+try:
+    from transformers.utils.quantization_config import QuantizationConfigMixin as TransformersQuantConfigMixin
+except ImportError:
+
+    class TransformersQuantConfigMixin:
+        pass
+
+
+logger = logging.get_logger(__name__)
+
+
+class PipelineQuantizationConfig:
+    """
+    Configuration class to be used when applying quantization on-the-fly to [`~DiffusionPipeline.from_pretrained`].
+
+    Args:
+        quant_backend (`str`): Quantization backend to be used. When using this option, we assume that the backend
+            is available to both `diffusers` and `transformers`.
+        quant_kwargs (`dict`): Params to initialize the quantization backend class.
+        components_to_quantize (`list`): Components of a pipeline to be quantized.
+        quant_mapping (`dict`): Mapping defining the quantization specs to be used for the pipeline
+            components. When using this argument, users are not expected to provide `quant_backend`, `quant_kawargs`,
+            and `components_to_quantize`.
+    """
+
+    def __init__(
+        self,
+        quant_backend: str = None,
+        quant_kwargs: Dict[str, Union[str, float, int, dict]] = None,
+        components_to_quantize: Optional[List[str]] = None,
+        quant_mapping: Dict[str, Union[DiffQuantConfigMixin, "TransformersQuantConfigMixin"]] = None,
+    ):
+        self.quant_backend = quant_backend
+        # Initialize kwargs to be {} to set to the defaults.
+        self.quant_kwargs = quant_kwargs or {}
+        self.components_to_quantize = components_to_quantize
+        self.quant_mapping = quant_mapping
+        self.config_mapping = {}  # book-keeping Example: `{module_name: quant_config}`
+        self.post_init()
+
+    def post_init(self):
+        quant_mapping = self.quant_mapping
+        self.is_granular = True if quant_mapping is not None else False
+
+        self._validate_init_args()
+
+    def _validate_init_args(self):
+        if self.quant_backend and self.quant_mapping:
+            raise ValueError("Both `quant_backend` and `quant_mapping` cannot be specified at the same time.")
+
+        if not self.quant_mapping and not self.quant_backend:
+            raise ValueError("Must provide a `quant_backend` when not providing a `quant_mapping`.")
+
+        if not self.quant_kwargs and not self.quant_mapping:
+            raise ValueError("Both `quant_kwargs` and `quant_mapping` cannot be None.")
+
+        if self.quant_backend is not None:
+            self._validate_init_kwargs_in_backends()
+
+        if self.quant_mapping is not None:
+            self._validate_quant_mapping_args()
+
+    def _validate_init_kwargs_in_backends(self):
+        quant_backend = self.quant_backend
+
+        self._check_backend_availability(quant_backend)
+
+        quant_config_mapping_transformers, quant_config_mapping_diffusers = self._get_quant_config_list()
+
+        if quant_config_mapping_transformers is not None:
+            init_kwargs_transformers = inspect.signature(quant_config_mapping_transformers[quant_backend].__init__)
+            init_kwargs_transformers = {name for name in init_kwargs_transformers.parameters if name != "self"}
+        else:
+            init_kwargs_transformers = None
+
+        init_kwargs_diffusers = inspect.signature(quant_config_mapping_diffusers[quant_backend].__init__)
+        init_kwargs_diffusers = {name for name in init_kwargs_diffusers.parameters if name != "self"}
+
+        if init_kwargs_transformers != init_kwargs_diffusers:
+            raise ValueError(
+                "The signatures of the __init__ methods of the quantization config classes in `diffusers` and `transformers` don't match. "
+                f"Please provide a `quant_mapping` instead, in the {self.__class__.__name__} class. Refer to [the docs](https://huggingface.co/docs/diffusers/main/en/quantization/overview#pipeline-level-quantization) to learn more about how "
+                "this mapping would look like."
+            )
+
+    def _validate_quant_mapping_args(self):
+        quant_mapping = self.quant_mapping
+        transformers_map, diffusers_map = self._get_quant_config_list()
+
+        available_transformers = list(transformers_map.values()) if transformers_map else None
+        available_diffusers = list(diffusers_map.values())
+
+        for module_name, config in quant_mapping.items():
+            if any(isinstance(config, cfg) for cfg in available_diffusers):
+                continue
+
+            if available_transformers and any(isinstance(config, cfg) for cfg in available_transformers):
+                continue
+
+            if available_transformers:
+                raise ValueError(
+                    f"Provided config for module_name={module_name} could not be found. "
+                    f"Available diffusers configs: {available_diffusers}; "
+                    f"Available transformers configs: {available_transformers}."
+                )
+            else:
+                raise ValueError(
+                    f"Provided config for module_name={module_name} could not be found. "
+                    f"Available diffusers configs: {available_diffusers}."
+                )
+
+    def _check_backend_availability(self, quant_backend: str):
+        quant_config_mapping_transformers, quant_config_mapping_diffusers = self._get_quant_config_list()
+
+        available_backends_transformers = (
+            list(quant_config_mapping_transformers.keys()) if quant_config_mapping_transformers else None
+        )
+        available_backends_diffusers = list(quant_config_mapping_diffusers.keys())
+
+        if (
+            available_backends_transformers and quant_backend not in available_backends_transformers
+        ) or quant_backend not in quant_config_mapping_diffusers:
+            error_message = f"Provided quant_backend={quant_backend} was not found."
+            if available_backends_transformers:
+                error_message += f"\nAvailable ones (transformers): {available_backends_transformers}."
+            error_message += f"\nAvailable ones (diffusers): {available_backends_diffusers}."
+            raise ValueError(error_message)
+
+    def _resolve_quant_config(self, is_diffusers: bool = True, module_name: str = None):
+        quant_config_mapping_transformers, quant_config_mapping_diffusers = self._get_quant_config_list()
+
+        quant_mapping = self.quant_mapping
+        components_to_quantize = self.components_to_quantize
+
+        # Granular case
+        if self.is_granular and module_name in quant_mapping:
+            logger.debug(f"Initializing quantization config class for {module_name}.")
+            config = quant_mapping[module_name]
+            self.config_mapping.update({module_name: config})
+            return config
+
+        # Global config case
+        else:
+            should_quantize = False
+            # Only quantize the modules requested for.
+            if components_to_quantize and module_name in components_to_quantize:
+                should_quantize = True
+            # No specification for `components_to_quantize` means all modules should be quantized.
+            elif not self.is_granular and not components_to_quantize:
+                should_quantize = True
+
+            if should_quantize:
+                logger.debug(f"Initializing quantization config class for {module_name}.")
+                mapping_to_use = quant_config_mapping_diffusers if is_diffusers else quant_config_mapping_transformers
+                quant_config_cls = mapping_to_use[self.quant_backend]
+                quant_kwargs = self.quant_kwargs
+                quant_obj = quant_config_cls(**quant_kwargs)
+                self.config_mapping.update({module_name: quant_obj})
+                return quant_obj
+
+        # Fallback: no applicable configuration found.
+        return None
+
+    def _get_quant_config_list(self):
+        if is_transformers_available():
+            from transformers.quantizers.auto import (
+                AUTO_QUANTIZATION_CONFIG_MAPPING as quant_config_mapping_transformers,
+            )
+        else:
+            quant_config_mapping_transformers = None
+
+        from ..quantizers.auto import AUTO_QUANTIZATION_CONFIG_MAPPING as quant_config_mapping_diffusers
+
+        return quant_config_mapping_transformers, quant_config_mapping_diffusers
+
+    def __repr__(self):
+        out = ""
+        config_mapping = dict(sorted(self.config_mapping.copy().items()))
+        for module_name, config in config_mapping.items():
+            out += f"{module_name} {config}"
+        return out

diffusers/quantizers/quantization_config.py

@@ -75,7 +75,7 @@ class QuantizationConfigMixin:
         Args:
             config_dict (`Dict[str, Any]`):
                 Dictionary that will be used to instantiate the configuration object.
-            return_unused_kwargs (`bool`,*optional*, defaults to `False`):
+            return_unused_kwargs (`bool`, *optional*, defaults to `False`):
                 Whether or not to return a list of unused keyword arguments. Used for `from_pretrained` method in
                 `PreTrainedModel`.
             kwargs (`Dict[str, Any]`):
@@ -179,7 +179,7 @@ class BitsAndBytesConfig(QuantizationConfigMixin):
     This is a wrapper class about all possible attributes and features that you can play with a model that has been
     loaded using `bitsandbytes`.
 
-    This replaces `load_in_8bit` or `load_in_4bit`therefore both options are mutually exclusive.
+    This replaces `load_in_8bit` or `load_in_4bit` therefore both options are mutually exclusive.
 
     Currently only supports `LLM.int8()`, `FP4`, and `NF4` quantization. If more methods are added to `bitsandbytes`,
     then more arguments will be added to this class.
@@ -192,10 +192,10 @@ class BitsAndBytesConfig(QuantizationConfigMixin):
             `bitsandbytes`.
         llm_int8_threshold (`float`, *optional*, defaults to 6.0):
             This corresponds to the outlier threshold for outlier detection as described in `LLM.int8() : 8-bit Matrix
-            Multiplication for Transformers at Scale` paper: https://arxiv.org/abs/2208.07339 Any hidden states value
-            that is above this threshold will be considered an outlier and the operation on those values will be done
-            in fp16. Values are usually normally distributed, that is, most values are in the range [-3.5, 3.5], but
-            there are some exceptional systematic outliers that are very differently distributed for large models.
+            Multiplication for Transformers at Scale` paper: https://huggingface.co/papers/2208.07339 Any hidden states
+            value that is above this threshold will be considered an outlier and the operation on those values will be
+            done in fp16. Values are usually normally distributed, that is, most values are in the range [-3.5, 3.5],
+            but there are some exceptional systematic outliers that are very differently distributed for large models.
             These outliers are often in the interval [-60, -6] or [6, 60]. Int8 quantization works well for values of
             magnitude ~5, but beyond that, there is a significant performance penalty. A good default threshold is 6,
             but a lower threshold might be needed for more unstable models (small models, fine-tuning).
@@ -493,7 +493,7 @@ class TorchAoConfig(QuantizationConfigMixin):
         TORCHAO_QUANT_TYPE_METHODS = self._get_torchao_quant_type_to_method()
         if self.quant_type not in TORCHAO_QUANT_TYPE_METHODS.keys():
             is_floating_quant_type = self.quant_type.startswith("float") or self.quant_type.startswith("fp")
-            if is_floating_quant_type and not self._is_cuda_capability_atleast_8_9():
+            if is_floating_quant_type and not self._is_xpu_or_cuda_capability_atleast_8_9():
                 raise ValueError(
                     f"Requested quantization type: {self.quant_type} is not supported on GPUs with CUDA capability <= 8.9. You "
                     f"can check the CUDA capability of your GPU using `torch.cuda.get_device_capability()`."
@@ -645,7 +645,7 @@ class TorchAoConfig(QuantizationConfigMixin):
             QUANTIZATION_TYPES.update(INT8_QUANTIZATION_TYPES)
             QUANTIZATION_TYPES.update(UINTX_QUANTIZATION_DTYPES)
 
-            if cls._is_cuda_capability_atleast_8_9():
+            if cls._is_xpu_or_cuda_capability_atleast_8_9():
                 QUANTIZATION_TYPES.update(FLOATX_QUANTIZATION_TYPES)
 
             return QUANTIZATION_TYPES
@@ -655,14 +655,16 @@ class TorchAoConfig(QuantizationConfigMixin):
             )
 
     @staticmethod
-    def _is_cuda_capability_atleast_8_9() -> bool:
-        if not torch.cuda.is_available():
-            raise RuntimeError("TorchAO requires a CUDA compatible GPU and installation of PyTorch.")
-
-        major, minor = torch.cuda.get_device_capability()
-        if major == 8:
-            return minor >= 9
-        return major >= 9
+    def _is_xpu_or_cuda_capability_atleast_8_9() -> bool:
+        if torch.cuda.is_available():
+            major, minor = torch.cuda.get_device_capability()
+            if major == 8:
+                return minor >= 9
+            return major >= 9
+        elif torch.xpu.is_available():
+            return True
+        else:
+            raise RuntimeError("TorchAO requires a CUDA compatible GPU or Intel XPU and installation of PyTorch.")
 
     def get_apply_tensor_subclass(self):
         TORCHAO_QUANT_TYPE_METHODS = self._get_torchao_quant_type_to_method()

diffusers/quantizers/quanto/quanto_quantizer.py

@@ -175,3 +175,7 @@ class QuantoQuantizer(DiffusersQuantizer):
     @property
     def is_serializable(self):
         return True
+
+    @property
+    def is_compileable(self) -> bool:
+        return True

diffusers/quantizers/torchao/torchao_quantizer.py

@@ -19,6 +19,7 @@ https://github.com/huggingface/transformers/blob/3a8eb74668e9c2cc563b2f5c62fac17
 
 import importlib
 import types
+from fnmatch import fnmatch
 from typing import TYPE_CHECKING, Any, Dict, List, Union
 
 from packaging import version
@@ -262,7 +263,7 @@ class TorchAoHfQuantizer(DiffusersQuantizer):
         **kwargs,
     ):
         r"""
-        Each nn.Linear layer that needs to be quantized is processsed here. First, we set the value the weight tensor,
+        Each nn.Linear layer that needs to be quantized is processed here. First, we set the value the weight tensor,
         then we move it to the target device. Finally, we quantize the module.
         """
         module, tensor_name = get_module_from_name(model, param_name)
@@ -278,6 +279,31 @@ class TorchAoHfQuantizer(DiffusersQuantizer):
             module._parameters[tensor_name] = torch.nn.Parameter(param_value).to(device=target_device)
             quantize_(module, self.quantization_config.get_apply_tensor_subclass())
 
+    def get_cuda_warm_up_factor(self):
+        """
+        This factor is used in caching_allocator_warmup to determine how many bytes to pre-allocate for CUDA warmup.
+        - A factor of 2 means we pre-allocate the full memory footprint of the model.
+        - A factor of 4 means we pre-allocate half of that, and so on
+
+        However, when using TorchAO, calculating memory usage with param.numel() * param.element_size() doesn't give
+        the correct size for quantized weights (like int4 or int8) That's because TorchAO internally represents
+        quantized tensors using subtensors and metadata, and the reported element_size() still corresponds to the
+        torch_dtype not the actual bit-width of the quantized data.
+
+        To correct for this:
+        - Use a division factor of 8 for int4 weights
+        - Use a division factor of 4 for int8 weights
+        """
+        # Original mapping for non-AOBaseConfig types
+        # For the uint types, this is a best guess. Once these types become more used
+        # we can look into their nuances.
+        map_to_target_dtype = {"int4_*": 8, "int8_*": 4, "uint*": 8, "float8*": 4}
+        quant_type = self.quantization_config.quant_type
+        for pattern, target_dtype in map_to_target_dtype.items():
+            if fnmatch(quant_type, pattern):
+                return target_dtype
+        raise ValueError(f"Unsupported quant_type: {quant_type!r}")
+
     def _process_model_before_weight_loading(
         self,
         model: "ModelMixin",
@@ -335,3 +361,7 @@ class TorchAoHfQuantizer(DiffusersQuantizer):
     @property
     def is_trainable(self):
         return self.quantization_config.quant_type.startswith("int8")
+
+    @property
+    def is_compileable(self) -> bool:
+        return True

diffusers/schedulers/__init__.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.
@@ -60,6 +60,7 @@ else:
     _import_structure["scheduling_euler_discrete"] = ["EulerDiscreteScheduler"]
     _import_structure["scheduling_flow_match_euler_discrete"] = ["FlowMatchEulerDiscreteScheduler"]
     _import_structure["scheduling_flow_match_heun_discrete"] = ["FlowMatchHeunDiscreteScheduler"]
+    _import_structure["scheduling_flow_match_lcm"] = ["FlowMatchLCMScheduler"]
     _import_structure["scheduling_heun_discrete"] = ["HeunDiscreteScheduler"]
     _import_structure["scheduling_ipndm"] = ["IPNDMScheduler"]
     _import_structure["scheduling_k_dpm_2_ancestral_discrete"] = ["KDPM2AncestralDiscreteScheduler"]
@@ -161,6 +162,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .scheduling_euler_discrete import EulerDiscreteScheduler
         from .scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler
         from .scheduling_flow_match_heun_discrete import FlowMatchHeunDiscreteScheduler
+        from .scheduling_flow_match_lcm import FlowMatchLCMScheduler
         from .scheduling_heun_discrete import HeunDiscreteScheduler
         from .scheduling_ipndm import IPNDMScheduler
         from .scheduling_k_dpm_2_ancestral_discrete import KDPM2AncestralDiscreteScheduler

diffusers/schedulers/deprecated/scheduling_karras_ve.py

@@ -1,4 +1,4 @@
-# Copyright 2024 NVIDIA and The HuggingFace Team. All rights reserved.
+# Copyright 2025 NVIDIA and 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.
@@ -55,8 +55,9 @@ class KarrasVeScheduler(SchedulerMixin, ConfigMixin):
 
     <Tip>
 
-    For more details on the parameters, see [Appendix E](https://arxiv.org/abs/2206.00364). The grid search values used
-    to find the optimal `{s_noise, s_churn, s_min, s_max}` for a specific model are described in Table 5 of the paper.
+    For more details on the parameters, see [Appendix E](https://huggingface.co/papers/2206.00364). The grid search
+    values used to find the optimal `{s_noise, s_churn, s_min, s_max}` for a specific model are described in Table 5 of
+    the paper.
 
     </Tip>
 

diffusers/schedulers/deprecated/scheduling_sde_vp.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Google Brain and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Google Brain and 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.

diffusers/schedulers/scheduling_consistency_models.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.

diffusers/schedulers/scheduling_cosine_dpmsolver_multistep.py

@@ -1,4 +1,4 @@
-# Copyright 2024 TSAIL Team and The HuggingFace Team. All rights reserved.
+# Copyright 2025 TSAIL Team and 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.
@@ -30,7 +30,7 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
     Implements a variant of `DPMSolverMultistepScheduler` with cosine schedule, proposed by Nichol and Dhariwal (2021).
     This scheduler was used in Stable Audio Open [1].
 
-    [1] Evans, Parker, et al. "Stable Audio Open" https://arxiv.org/abs/2407.14358
+    [1] Evans, Parker, et al. "Stable Audio Open" https://huggingface.co/papers/2407.14358
 
     This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
     methods the library implements for all schedulers such as loading and saving.
@@ -44,8 +44,8 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
             The standard deviation of the data distribution. This is set to 1.0 in Stable Audio Open [1].
         sigma_schedule (`str`, *optional*, defaults to `exponential`):
             Sigma schedule to compute the `sigmas`. By default, we the schedule introduced in the EDM paper
-            (https://arxiv.org/abs/2206.00364). Other acceptable value is "exponential". The exponential schedule was
-            incorporated in this model: https://huggingface.co/stabilityai/cosxl.
+            (https://huggingface.co/papers/2206.00364). Other acceptable value is "exponential". The exponential
+            schedule was incorporated in this model: https://huggingface.co/stabilityai/cosxl.
         num_train_timesteps (`int`, defaults to 1000):
             The number of diffusion steps to train the model.
         solver_order (`int`, defaults to 2):
@@ -144,7 +144,7 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
 
     # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_inputs
     def precondition_inputs(self, sample, sigma):
-        c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
+        c_in = self._get_conditioning_c_in(sigma)
         scaled_sample = sample * c_in
         return scaled_sample
 
@@ -568,5 +568,10 @@ class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         noisy_samples = original_samples + noise * sigma
         return noisy_samples
 
+    # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._get_conditioning_c_in
+    def _get_conditioning_c_in(self, sigma):
+        c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
+        return c_in
+
     def __len__(self):
         return self.config.num_train_timesteps

diffusers/schedulers/scheduling_ddim.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Stanford University Team and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Stanford University Team and 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.
@@ -94,7 +94,7 @@ def betas_for_alpha_bar(
 
 def rescale_zero_terminal_snr(betas):
     """
-    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+    Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1)
 
 
     Args:
@@ -269,7 +269,7 @@ class DDIMScheduler(SchedulerMixin, ConfigMixin):
         pixels from saturation at each step. We find that dynamic thresholding results in significantly better
         photorealism as well as better image-text alignment, especially when using very large guidance weights."
 
-        https://arxiv.org/abs/2205.11487
+        https://huggingface.co/papers/2205.11487
         """
         dtype = sample.dtype
         batch_size, channels, *remaining_dims = sample.shape
@@ -312,7 +312,7 @@ class DDIMScheduler(SchedulerMixin, ConfigMixin):
 
         self.num_inference_steps = num_inference_steps
 
-        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891
         if self.config.timestep_spacing == "linspace":
             timesteps = (
                 np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps)
@@ -387,7 +387,7 @@ class DDIMScheduler(SchedulerMixin, ConfigMixin):
                 "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
             )
 
-        # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+        # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502
         # Ideally, read DDIM paper in-detail understanding
 
         # Notation (<variable name> -> <name in paper>
@@ -408,7 +408,7 @@ class DDIMScheduler(SchedulerMixin, ConfigMixin):
         beta_prod_t = 1 - alpha_prod_t
 
         # 3. compute predicted original sample from predicted noise also called
-        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502
         if self.config.prediction_type == "epsilon":
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
             pred_epsilon = model_output
@@ -441,10 +441,10 @@ class DDIMScheduler(SchedulerMixin, ConfigMixin):
             # the pred_epsilon is always re-derived from the clipped x_0 in Glide
             pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
 
-        # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502
         pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon
 
-        # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 7. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502
         prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
 
         if eta > 0:

diffusers/schedulers/scheduling_ddim_cogvideox.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# Copyright 2025 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
 # All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -94,7 +94,7 @@ def betas_for_alpha_bar(
 
 def rescale_zero_terminal_snr(alphas_cumprod):
     """
-    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+    Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1)
 
 
     Args:
@@ -275,7 +275,7 @@ class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin):
 
         self.num_inference_steps = num_inference_steps
 
-        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891
         if self.config.timestep_spacing == "linspace":
             timesteps = (
                 np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps)
@@ -350,7 +350,7 @@ class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin):
                 "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
             )
 
-        # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+        # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502
         # Ideally, read DDIM paper in-detail understanding
 
         # Notation (<variable name> -> <name in paper>
@@ -371,7 +371,7 @@ class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin):
         beta_prod_t = 1 - alpha_prod_t
 
         # 3. compute predicted original sample from predicted noise also called
-        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502
         # To make style tests pass, commented out `pred_epsilon` as it is an unused variable
         if self.config.prediction_type == "epsilon":
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)

diffusers/schedulers/scheduling_ddim_flax.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Stanford University Team and The HuggingFace Team. All rights reserved.
+# Copyright 2025 Stanford University Team and 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.
@@ -73,7 +73,7 @@ class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
     [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
     [`~SchedulerMixin.from_pretrained`] functions.
 
-    For more details, see the original paper: https://arxiv.org/abs/2010.02502
+    For more details, see the original paper: https://huggingface.co/papers/2010.02502
 
     Args:
         num_train_timesteps (`int`): number of diffusion steps used to train the model.
@@ -230,7 +230,7 @@ class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
                 "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
             )
 
-        # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+        # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502
         # Ideally, read DDIM paper in-detail understanding
 
         # Notation (<variable name> -> <name in paper>
@@ -254,7 +254,7 @@ class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
         beta_prod_t = 1 - alpha_prod_t
 
         # 3. compute predicted original sample from predicted noise also called
-        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502
         if self.config.prediction_type == "epsilon":
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
             pred_epsilon = model_output
@@ -281,10 +281,10 @@ class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
         variance = self._get_variance(state, timestep, prev_timestep)
         std_dev_t = eta * variance ** (0.5)
 
-        # 5. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 5. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502
         pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon
 
-        # 6. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 6. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502
         prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
 
         if not return_dict:

diffusers/schedulers/scheduling_ddim_inverse.py

@@ -1,4 +1,4 @@
-# Copyright 2024 The HuggingFace Team. All rights reserved.
+# Copyright 2025 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.
@@ -93,7 +93,7 @@ def betas_for_alpha_bar(
 # Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
 def rescale_zero_terminal_snr(betas):
     """
-    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+    Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1)
 
 
     Args:
@@ -266,7 +266,7 @@ class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
 
         self.num_inference_steps = num_inference_steps
 
-        # "leading" and "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+        # "leading" and "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891
         if self.config.timestep_spacing == "leading":
             step_ratio = self.config.num_train_timesteps // self.num_inference_steps
             # creates integer timesteps by multiplying by ratio
@@ -338,7 +338,7 @@ class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
         beta_prod_t = 1 - alpha_prod_t
 
         # 3. compute predicted original sample from predicted noise also called
-        # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502
         if self.config.prediction_type == "epsilon":
             pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
             pred_epsilon = model_output
@@ -360,10 +360,10 @@ class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
                 -self.config.clip_sample_range, self.config.clip_sample_range
             )
 
-        # 5. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 5. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502
         pred_sample_direction = (1 - alpha_prod_t_prev) ** (0.5) * pred_epsilon
 
-        # 6. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+        # 6. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502
         prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
 
         if not return_dict: