diffusers 0.31.0__py3-none-any.whl → 0.32.1__py3-none-any.whl

diffusers/models/model_loading_utils.py

@@ -17,6 +17,7 @@
 import importlib
 import inspect
 import os
+from array import array
 from collections import OrderedDict
 from pathlib import Path
 from typing import List, Optional, Union
@@ -25,8 +26,8 @@ import safetensors
 import torch
 from huggingface_hub.utils import EntryNotFoundError
 
-from ..quantizers.quantization_config import QuantizationMethod
 from ..utils import (
+    GGUF_FILE_EXTENSION,
     SAFE_WEIGHTS_INDEX_NAME,
     SAFETENSORS_FILE_EXTENSION,
     WEIGHTS_INDEX_NAME,
@@ -34,6 +35,8 @@ from ..utils import (
     _get_model_file,
     deprecate,
     is_accelerate_available,
+    is_gguf_available,
+    is_torch_available,
     is_torch_version,
     logging,
 )
@@ -140,6 +143,8 @@ def load_state_dict(checkpoint_file: Union[str, os.PathLike], variant: Optional[
         file_extension = os.path.basename(checkpoint_file).split(".")[-1]
         if file_extension == SAFETENSORS_FILE_EXTENSION:
             return safetensors.torch.load_file(checkpoint_file, device="cpu")
+        elif file_extension == GGUF_FILE_EXTENSION:
+            return load_gguf_checkpoint(checkpoint_file)
         else:
             weights_only_kwarg = {"weights_only": True} if is_torch_version(">=", "1.13") else {}
             return torch.load(
@@ -176,11 +181,12 @@ def load_model_dict_into_meta(
     hf_quantizer=None,
     keep_in_fp32_modules=None,
 ) -> List[str]:
+    if device is not None and not isinstance(device, (str, torch.device)):
+        raise ValueError(f"Expected device to have type `str` or `torch.device`, but got {type(device)=}.")
     if hf_quantizer is None:
         device = device or torch.device("cpu")
     dtype = dtype or torch.float32
     is_quantized = hf_quantizer is not None
-    is_quant_method_bnb = getattr(model, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES
 
     accepts_dtype = "dtype" in set(inspect.signature(set_module_tensor_to_device).parameters.keys())
     empty_state_dict = model.state_dict()
@@ -211,17 +217,18 @@ def load_model_dict_into_meta(
                     set_module_kwargs["dtype"] = dtype
 
         # bnb params are flattened.
+        # gguf quants have a different shape based on the type of quantization applied
         if empty_state_dict[param_name].shape != param.shape:
             if (
-                is_quant_method_bnb
+                is_quantized
                 and hf_quantizer.pre_quantized
                 and hf_quantizer.check_if_quantized_param(model, param, param_name, state_dict, param_device=device)
             ):
-                hf_quantizer.check_quantized_param_shape(param_name, empty_state_dict[param_name].shape, param.shape)
-            elif not is_quant_method_bnb:
+                hf_quantizer.check_quantized_param_shape(param_name, empty_state_dict[param_name], param)
+            else:
                 model_name_or_path_str = f"{model_name_or_path} " if model_name_or_path is not None else ""
                 raise ValueError(
-                    f"Cannot load {model_name_or_path_str} because {param_name} expected shape {empty_state_dict[param_name]}, but got {param.shape}. If you want to instead overwrite randomly initialized weights, please make sure to pass both `low_cpu_mem_usage=False` and `ignore_mismatched_sizes=True`. For more information, see also: https://github.com/huggingface/diffusers/issues/1619#issuecomment-1345604389 as an example."
+                    f"Cannot load {model_name_or_path_str} because {param_name} expected shape {empty_state_dict[param_name].shape}, but got {param.shape}. If you want to instead overwrite randomly initialized weights, please make sure to pass both `low_cpu_mem_usage=False` and `ignore_mismatched_sizes=True`. For more information, see also: https://github.com/huggingface/diffusers/issues/1619#issuecomment-1345604389 as an example."
                 )
 
         if is_quantized and (
@@ -396,3 +403,78 @@ def _fetch_index_file_legacy(
                 index_file = None
 
     return index_file
+
+
+def _gguf_parse_value(_value, data_type):
+    if not isinstance(data_type, list):
+        data_type = [data_type]
+    if len(data_type) == 1:
+        data_type = data_type[0]
+        array_data_type = None
+    else:
+        if data_type[0] != 9:
+            raise ValueError("Received multiple types, therefore expected the first type to indicate an array.")
+        data_type, array_data_type = data_type
+
+    if data_type in [0, 1, 2, 3, 4, 5, 10, 11]:
+        _value = int(_value[0])
+    elif data_type in [6, 12]:
+        _value = float(_value[0])
+    elif data_type in [7]:
+        _value = bool(_value[0])
+    elif data_type in [8]:
+        _value = array("B", list(_value)).tobytes().decode()
+    elif data_type in [9]:
+        _value = _gguf_parse_value(_value, array_data_type)
+    return _value
+
+
+def load_gguf_checkpoint(gguf_checkpoint_path, return_tensors=False):
+    """
+    Load a GGUF file and return a dictionary of parsed parameters containing tensors, the parsed tokenizer and config
+    attributes.
+
+    Args:
+        gguf_checkpoint_path (`str`):
+            The path the to GGUF file to load
+        return_tensors (`bool`, defaults to `True`):
+            Whether to read the tensors from the file and return them. Not doing so is faster and only loads the
+            metadata in memory.
+    """
+
+    if is_gguf_available() and is_torch_available():
+        import gguf
+        from gguf import GGUFReader
+
+        from ..quantizers.gguf.utils import SUPPORTED_GGUF_QUANT_TYPES, GGUFParameter
+    else:
+        logger.error(
+            "Loading a GGUF checkpoint in PyTorch, requires both PyTorch and GGUF>=0.10.0 to be installed. Please see "
+            "https://pytorch.org/ and https://github.com/ggerganov/llama.cpp/tree/master/gguf-py for installation instructions."
+        )
+        raise ImportError("Please install torch and gguf>=0.10.0 to load a GGUF checkpoint in PyTorch.")
+
+    reader = GGUFReader(gguf_checkpoint_path)
+
+    parsed_parameters = {}
+    for tensor in reader.tensors:
+        name = tensor.name
+        quant_type = tensor.tensor_type
+
+        # if the tensor is a torch supported dtype do not use GGUFParameter
+        is_gguf_quant = quant_type not in [gguf.GGMLQuantizationType.F32, gguf.GGMLQuantizationType.F16]
+        if is_gguf_quant and quant_type not in SUPPORTED_GGUF_QUANT_TYPES:
+            _supported_quants_str = "\n".join([str(type) for type in SUPPORTED_GGUF_QUANT_TYPES])
+            raise ValueError(
+                (
+                    f"{name} has a quantization type: {str(quant_type)} which is unsupported."
+                    "\n\nCurrently the following quantization types are supported: \n\n"
+                    f"{_supported_quants_str}"
+                    "\n\nTo request support for this quantization type please open an issue here: https://github.com/huggingface/diffusers"
+                )
+            )
+
+        weights = torch.from_numpy(tensor.data.copy())
+        parsed_parameters[name] = GGUFParameter(weights, quant_type=quant_type) if is_gguf_quant else weights
+
+    return parsed_parameters

diffusers/models/modeling_flax_utils.py

@@ -530,7 +530,7 @@ class FlaxModelMixin(PushToHubMixin):
 
         if push_to_hub:
             commit_message = kwargs.pop("commit_message", None)
-            private = kwargs.pop("private", False)
+            private = kwargs.pop("private", None)
             create_pr = kwargs.pop("create_pr", False)
             token = kwargs.pop("token", None)
             repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])

diffusers/models/modeling_utils.py

@@ -99,21 +99,39 @@ def get_parameter_device(parameter: torch.nn.Module) -> torch.device:
 
 
 def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype:
-    try:
-        return next(parameter.parameters()).dtype
-    except StopIteration:
-        try:
-            return next(parameter.buffers()).dtype
-        except StopIteration:
-            # For torch.nn.DataParallel compatibility in PyTorch 1.5
-
-            def find_tensor_attributes(module: torch.nn.Module) -> List[Tuple[str, Tensor]]:
-                tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
-                return tuples
-
-            gen = parameter._named_members(get_members_fn=find_tensor_attributes)
-            first_tuple = next(gen)
-            return first_tuple[1].dtype
+    """
+    Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found.
+    """
+    last_dtype = None
+    for param in parameter.parameters():
+        last_dtype = param.dtype
+        if param.is_floating_point():
+            return param.dtype
+
+    for buffer in parameter.buffers():
+        last_dtype = buffer.dtype
+        if buffer.is_floating_point():
+            return buffer.dtype
+
+    if last_dtype is not None:
+        # if no floating dtype was found return whatever the first dtype is
+        return last_dtype
+
+    # For nn.DataParallel compatibility in PyTorch > 1.5
+    def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:
+        tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
+        return tuples
+
+    gen = parameter._named_members(get_members_fn=find_tensor_attributes)
+    last_tuple = None
+    for tuple in gen:
+        last_tuple = tuple
+        if tuple[1].is_floating_point():
+            return tuple[1].dtype
+
+    if last_tuple is not None:
+        # fallback to the last dtype
+        return last_tuple[1].dtype
 
 
 class ModelMixin(torch.nn.Module, PushToHubMixin):
@@ -208,6 +226,35 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         """
         self.set_use_npu_flash_attention(False)
 
+    def set_use_xla_flash_attention(
+        self, use_xla_flash_attention: bool, partition_spec: Optional[Callable] = None
+    ) -> None:
+        # Recursively walk through all the children.
+        # Any children which exposes the set_use_xla_flash_attention method
+        # gets the message
+        def fn_recursive_set_flash_attention(module: torch.nn.Module):
+            if hasattr(module, "set_use_xla_flash_attention"):
+                module.set_use_xla_flash_attention(use_xla_flash_attention, partition_spec)
+
+            for child in module.children():
+                fn_recursive_set_flash_attention(child)
+
+        for module in self.children():
+            if isinstance(module, torch.nn.Module):
+                fn_recursive_set_flash_attention(module)
+
+    def enable_xla_flash_attention(self, partition_spec: Optional[Callable] = None):
+        r"""
+        Enable the flash attention pallals kernel for torch_xla.
+        """
+        self.set_use_xla_flash_attention(True, partition_spec)
+
+    def disable_xla_flash_attention(self):
+        r"""
+        Disable the flash attention pallals kernel for torch_xla.
+        """
+        self.set_use_xla_flash_attention(False)
+
     def set_use_memory_efficient_attention_xformers(
         self, valid: bool, attention_op: Optional[Callable] = None
     ) -> None:
@@ -338,7 +385,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
 
         if push_to_hub:
             commit_message = kwargs.pop("commit_message", None)
-            private = kwargs.pop("private", False)
+            private = kwargs.pop("private", None)
             create_pr = kwargs.pop("create_pr", False)
             token = kwargs.pop("token", None)
             repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
@@ -673,8 +720,9 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         if hf_quantizer is not None:
             if device_map is not None:
                 raise NotImplementedError(
-                    "Currently, `device_map` is automatically inferred for quantized models. Support for providing `device_map` as an input will be added in the future."
+                    "Currently, providing `device_map` is not supported for quantized models. Providing `device_map` as an input will be added in the future."
                 )
+
             hf_quantizer.validate_environment(torch_dtype=torch_dtype, from_flax=from_flax, device_map=device_map)
             torch_dtype = hf_quantizer.update_torch_dtype(torch_dtype)
 
@@ -771,6 +819,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                     revision=revision,
                     subfolder=subfolder or "",
                 )
+                # TODO: https://github.com/huggingface/diffusers/issues/10013
                 if hf_quantizer is not None:
                     model_file = _merge_sharded_checkpoints(sharded_ckpt_cached_folder, sharded_metadata)
                     logger.info("Merged sharded checkpoints as `hf_quantizer` is not None.")
@@ -829,14 +878,11 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                 if device_map is None and not is_sharded:
                     # `torch.cuda.current_device()` is fine here when `hf_quantizer` is not None.
                     # It would error out during the `validate_environment()` call above in the absence of cuda.
-                    is_quant_method_bnb = (
-                        getattr(model, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES
-                    )
                     if hf_quantizer is None:
                         param_device = "cpu"
                     # TODO (sayakpaul,  SunMarc): remove this after model loading refactor
-                    elif is_quant_method_bnb:
-                        param_device = torch.cuda.current_device()
+                    else:
+                        param_device = torch.device(torch.cuda.current_device())
                     state_dict = load_state_dict(model_file, variant=variant)
                     model._convert_deprecated_attention_blocks(state_dict)
 
@@ -1010,14 +1056,14 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
                     dtype_present_in_args = True
                     break
 
-        # Checks if the model has been loaded in 4-bit or 8-bit with BNB
-        if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
+        if getattr(self, "is_quantized", False):
             if dtype_present_in_args:
                 raise ValueError(
-                    "You cannot cast a bitsandbytes model in a new `dtype`. Make sure to load the model using `from_pretrained` using the"
-                    " desired `dtype` by passing the correct `torch_dtype` argument."
+                    "Casting a quantized model to a new `dtype` is unsupported. To set the dtype of unquantized layers, please "
+                    "use the `torch_dtype` argument when loading the model using `from_pretrained` or `from_single_file`"
                 )
 
+        if getattr(self, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES:
             if getattr(self, "is_loaded_in_8bit", False):
                 raise ValueError(
                     "`.to` is not supported for `8-bit` bitsandbytes models. Please use the model as it is, since the"

diffusers/models/normalization.py

@@ -22,10 +22,7 @@ import torch.nn.functional as F
 
 from ..utils import is_torch_version
 from .activations import get_activation
-from .embeddings import (
-    CombinedTimestepLabelEmbeddings,
-    PixArtAlphaCombinedTimestepSizeEmbeddings,
-)
+from .embeddings import CombinedTimestepLabelEmbeddings, PixArtAlphaCombinedTimestepSizeEmbeddings
 
 
 class AdaLayerNorm(nn.Module):
@@ -266,6 +263,7 @@ class AdaLayerNormSingle(nn.Module):
         hidden_dtype: Optional[torch.dtype] = None,
     ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
         # No modulation happening here.
+        added_cond_kwargs = added_cond_kwargs or {"resolution": None, "aspect_ratio": None}
         embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype)
         return self.linear(self.silu(embedded_timestep)), embedded_timestep
 
@@ -358,20 +356,21 @@ class LuminaLayerNormContinuous(nn.Module):
         out_dim: Optional[int] = None,
     ):
         super().__init__()
+
         # AdaLN
         self.silu = nn.SiLU()
         self.linear_1 = nn.Linear(conditioning_embedding_dim, embedding_dim, bias=bias)
+
         if norm_type == "layer_norm":
             self.norm = LayerNorm(embedding_dim, eps, elementwise_affine, bias)
+        elif norm_type == "rms_norm":
+            self.norm = RMSNorm(embedding_dim, eps=eps, elementwise_affine=elementwise_affine)
         else:
             raise ValueError(f"unknown norm_type {norm_type}")
-        # linear_2
+
+        self.linear_2 = None
         if out_dim is not None:
-            self.linear_2 = nn.Linear(
-                embedding_dim,
-                out_dim,
-                bias=bias,
-            )
+            self.linear_2 = nn.Linear(embedding_dim, out_dim, bias=bias)
 
     def forward(
         self,
@@ -486,20 +485,24 @@ else:
 
 
 class RMSNorm(nn.Module):
-    def __init__(self, dim, eps: float, elementwise_affine: bool = True):
+    def __init__(self, dim, eps: float, elementwise_affine: bool = True, bias: bool = False):
         super().__init__()
 
         self.eps = eps
+        self.elementwise_affine = elementwise_affine
 
         if isinstance(dim, numbers.Integral):
             dim = (dim,)
 
         self.dim = torch.Size(dim)
 
+        self.weight = None
+        self.bias = None
+
         if elementwise_affine:
             self.weight = nn.Parameter(torch.ones(dim))
-        else:
-            self.weight = None
+            if bias:
+                self.bias = nn.Parameter(torch.zeros(dim))
 
     def forward(self, hidden_states):
         input_dtype = hidden_states.dtype
@@ -511,12 +514,44 @@ class RMSNorm(nn.Module):
             if self.weight.dtype in [torch.float16, torch.bfloat16]:
                 hidden_states = hidden_states.to(self.weight.dtype)
             hidden_states = hidden_states * self.weight
+            if self.bias is not None:
+                hidden_states = hidden_states + self.bias
         else:
             hidden_states = hidden_states.to(input_dtype)
 
         return hidden_states
 
 
+# TODO: (Dhruv) This can be replaced with regular RMSNorm in Mochi once `_keep_in_fp32_modules` is supported
+# for sharded checkpoints, see: https://github.com/huggingface/diffusers/issues/10013
+class MochiRMSNorm(nn.Module):
+    def __init__(self, dim, eps: float, elementwise_affine: bool = True):
+        super().__init__()
+
+        self.eps = eps
+
+        if isinstance(dim, numbers.Integral):
+            dim = (dim,)
+
+        self.dim = torch.Size(dim)
+
+        if elementwise_affine:
+            self.weight = nn.Parameter(torch.ones(dim))
+        else:
+            self.weight = None
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
+
+        if self.weight is not None:
+            hidden_states = hidden_states * self.weight
+        hidden_states = hidden_states.to(input_dtype)
+
+        return hidden_states
+
+
 class GlobalResponseNorm(nn.Module):
     # Taken from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105
     def __init__(self, dim):
@@ -528,3 +563,33 @@ class GlobalResponseNorm(nn.Module):
         gx = torch.norm(x, p=2, dim=(1, 2), keepdim=True)
         nx = gx / (gx.mean(dim=-1, keepdim=True) + 1e-6)
         return self.gamma * (x * nx) + self.beta + x
+
+
+class LpNorm(nn.Module):
+    def __init__(self, p: int = 2, dim: int = -1, eps: float = 1e-12):
+        super().__init__()
+
+        self.p = p
+        self.dim = dim
+        self.eps = eps
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        return F.normalize(hidden_states, p=self.p, dim=self.dim, eps=self.eps)
+
+
+def get_normalization(
+    norm_type: str = "batch_norm",
+    num_features: Optional[int] = None,
+    eps: float = 1e-5,
+    elementwise_affine: bool = True,
+    bias: bool = True,
+) -> nn.Module:
+    if norm_type == "rms_norm":
+        norm = RMSNorm(num_features, eps=eps, elementwise_affine=elementwise_affine, bias=bias)
+    elif norm_type == "layer_norm":
+        norm = nn.LayerNorm(num_features, eps=eps, elementwise_affine=elementwise_affine, bias=bias)
+    elif norm_type == "batch_norm":
+        norm = nn.BatchNorm2d(num_features, eps=eps, affine=elementwise_affine)
+    else:
+        raise ValueError(f"{norm_type=} is not supported.")
+    return norm

diffusers/models/transformers/__init__.py

@@ -11,10 +11,15 @@ if is_torch_available():
     from .lumina_nextdit2d import LuminaNextDiT2DModel
     from .pixart_transformer_2d import PixArtTransformer2DModel
     from .prior_transformer import PriorTransformer
+    from .sana_transformer import SanaTransformer2DModel
     from .stable_audio_transformer import StableAudioDiTModel
     from .t5_film_transformer import T5FilmDecoder
     from .transformer_2d import Transformer2DModel
+    from .transformer_allegro import AllegroTransformer3DModel
     from .transformer_cogview3plus import CogView3PlusTransformer2DModel
     from .transformer_flux import FluxTransformer2DModel
+    from .transformer_hunyuan_video import HunyuanVideoTransformer3DModel
+    from .transformer_ltx import LTXVideoTransformer3DModel
+    from .transformer_mochi import MochiTransformer3DModel
     from .transformer_sd3 import SD3Transformer2DModel
     from .transformer_temporal import TransformerTemporalModel

diffusers/models/transformers/auraflow_transformer_2d.py

@@ -466,7 +466,7 @@ class AuraFlowTransformer2DModel(ModelMixin, ConfigMixin):
 
         # MMDiT blocks.
         for index_block, block in enumerate(self.joint_transformer_blocks):
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module, return_dict=None):
                     def custom_forward(*inputs):
@@ -497,7 +497,7 @@ class AuraFlowTransformer2DModel(ModelMixin, ConfigMixin):
             combined_hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
 
             for index_block, block in enumerate(self.single_transformer_blocks):
-                if self.training and self.gradient_checkpointing:
+                if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                     def create_custom_forward(module, return_dict=None):
                         def custom_forward(*inputs):

diffusers/models/transformers/cogvideox_transformer_3d.py

@@ -170,6 +170,8 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
             Whether to flip the sin to cos in the time embedding.
         time_embed_dim (`int`, defaults to `512`):
             Output dimension of timestep embeddings.
+        ofs_embed_dim (`int`, defaults to `512`):
+            Output dimension of "ofs" embeddings used in CogVideoX-5b-I2B in version 1.5
         text_embed_dim (`int`, defaults to `4096`):
             Input dimension of text embeddings from the text encoder.
         num_layers (`int`, defaults to `30`):
@@ -177,7 +179,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         dropout (`float`, defaults to `0.0`):
             The dropout probability to use.
         attention_bias (`bool`, defaults to `True`):
-            Whether or not to use bias in the attention projection layers.
+            Whether to use bias in the attention projection layers.
         sample_width (`int`, defaults to `90`):
             The width of the input latents.
         sample_height (`int`, defaults to `60`):
@@ -198,7 +200,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         timestep_activation_fn (`str`, defaults to `"silu"`):
             Activation function to use when generating the timestep embeddings.
         norm_elementwise_affine (`bool`, defaults to `True`):
-            Whether or not to use elementwise affine in normalization layers.
+            Whether to use elementwise affine in normalization layers.
         norm_eps (`float`, defaults to `1e-5`):
             The epsilon value to use in normalization layers.
         spatial_interpolation_scale (`float`, defaults to `1.875`):
@@ -219,6 +221,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         flip_sin_to_cos: bool = True,
         freq_shift: int = 0,
         time_embed_dim: int = 512,
+        ofs_embed_dim: Optional[int] = None,
         text_embed_dim: int = 4096,
         num_layers: int = 30,
         dropout: float = 0.0,
@@ -227,6 +230,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         sample_height: int = 60,
         sample_frames: int = 49,
         patch_size: int = 2,
+        patch_size_t: Optional[int] = None,
         temporal_compression_ratio: int = 4,
         max_text_seq_length: int = 226,
         activation_fn: str = "gelu-approximate",
@@ -237,6 +241,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         temporal_interpolation_scale: float = 1.0,
         use_rotary_positional_embeddings: bool = False,
         use_learned_positional_embeddings: bool = False,
+        patch_bias: bool = True,
     ):
         super().__init__()
         inner_dim = num_attention_heads * attention_head_dim
@@ -251,10 +256,11 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         # 1. Patch embedding
         self.patch_embed = CogVideoXPatchEmbed(
             patch_size=patch_size,
+            patch_size_t=patch_size_t,
             in_channels=in_channels,
             embed_dim=inner_dim,
             text_embed_dim=text_embed_dim,
-            bias=True,
+            bias=patch_bias,
             sample_width=sample_width,
             sample_height=sample_height,
             sample_frames=sample_frames,
@@ -267,10 +273,19 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         )
         self.embedding_dropout = nn.Dropout(dropout)
 
-        # 2. Time embeddings
+        # 2. Time embeddings and ofs embedding(Only CogVideoX1.5-5B I2V have)
+
         self.time_proj = Timesteps(inner_dim, flip_sin_to_cos, freq_shift)
         self.time_embedding = TimestepEmbedding(inner_dim, time_embed_dim, timestep_activation_fn)
 
+        self.ofs_proj = None
+        self.ofs_embedding = None
+        if ofs_embed_dim:
+            self.ofs_proj = Timesteps(ofs_embed_dim, flip_sin_to_cos, freq_shift)
+            self.ofs_embedding = TimestepEmbedding(
+                ofs_embed_dim, ofs_embed_dim, timestep_activation_fn
+            )  # same as time embeddings, for ofs
+
         # 3. Define spatio-temporal transformers blocks
         self.transformer_blocks = nn.ModuleList(
             [
@@ -298,7 +313,15 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
             norm_eps=norm_eps,
             chunk_dim=1,
         )
-        self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * out_channels)
+
+        if patch_size_t is None:
+            # For CogVideox 1.0
+            output_dim = patch_size * patch_size * out_channels
+        else:
+            # For CogVideoX 1.5
+            output_dim = patch_size * patch_size * patch_size_t * out_channels
+
+        self.proj_out = nn.Linear(inner_dim, output_dim)
 
         self.gradient_checkpointing = False
 
@@ -411,6 +434,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         encoder_hidden_states: torch.Tensor,
         timestep: Union[int, float, torch.LongTensor],
         timestep_cond: Optional[torch.Tensor] = None,
+        ofs: Optional[Union[int, float, torch.LongTensor]] = None,
         image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
         attention_kwargs: Optional[Dict[str, Any]] = None,
         return_dict: bool = True,
@@ -442,6 +466,12 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         t_emb = t_emb.to(dtype=hidden_states.dtype)
         emb = self.time_embedding(t_emb, timestep_cond)
 
+        if self.ofs_embedding is not None:
+            ofs_emb = self.ofs_proj(ofs)
+            ofs_emb = ofs_emb.to(dtype=hidden_states.dtype)
+            ofs_emb = self.ofs_embedding(ofs_emb)
+            emb = emb + ofs_emb
+
         # 2. Patch embedding
         hidden_states = self.patch_embed(encoder_hidden_states, hidden_states)
         hidden_states = self.embedding_dropout(hidden_states)
@@ -452,7 +482,7 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
 
         # 3. Transformer blocks
         for i, block in enumerate(self.transformer_blocks):
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):
@@ -491,12 +521,17 @@ class CogVideoXTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
         hidden_states = self.proj_out(hidden_states)
 
         # 5. Unpatchify
-        # Note: we use `-1` instead of `channels`:
-        #   - It is okay to `channels` use for CogVideoX-2b and CogVideoX-5b (number of input channels is equal to output channels)
-        #   - However, for CogVideoX-5b-I2V also takes concatenated input image latents (number of input channels is twice the output channels)
         p = self.config.patch_size
-        output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
-        output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
+        p_t = self.config.patch_size_t
+
+        if p_t is None:
+            output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
+            output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
+        else:
+            output = hidden_states.reshape(
+                batch_size, (num_frames + p_t - 1) // p_t, height // p, width // p, -1, p_t, p, p
+            )
+            output = output.permute(0, 1, 5, 4, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(1, 2)
 
         if USE_PEFT_BACKEND:
             # remove `lora_scale` from each PEFT layer

diffusers/models/transformers/dit_transformer_2d.py

@@ -184,7 +184,7 @@ class DiTTransformer2DModel(ModelMixin, ConfigMixin):
 
         # 2. Blocks
         for block in self.transformer_blocks:
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module, return_dict=None):
                     def custom_forward(*inputs):