inital commit

vae/autoencoder_kl_causal_3d.py

@@ -0,0 +1,625 @@
+# 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.
+# ==============================================================================
+#
+# Modified from diffusers==0.29.2
+#
+# ==============================================================================
+from typing import Dict, Optional, Tuple, Union
+from dataclasses import dataclass
+
+import torch
+import torch.nn as nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+try:
+    # This diffusers is modified and packed in the mirror.
+    from diffusers.loaders import FromOriginalVAEMixin
+except ImportError:
+    # Use this to be compatible with the original diffusers.
+    from diffusers.loaders.single_file_model import FromOriginalModelMixin as FromOriginalVAEMixin
+from diffusers.utils.accelerate_utils import apply_forward_hook
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    Attention,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.modeling_outputs import AutoencoderKLOutput
+from diffusers.models.modeling_utils import ModelMixin
+from .vae import DecoderCausal3D, BaseOutput, DecoderOutput, DiagonalGaussianDistribution, EncoderCausal3D
+
+@dataclass
+class DecoderOutput2(BaseOutput):
+    sample: torch.FloatTensor
+    posterior: Optional[DiagonalGaussianDistribution] = None
+
+
+class AutoencoderKLCausal3D(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
+    r"""
+    A VAE model with KL loss for encoding images/videos into latents and decoding latent representations into images/videos.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str] = ("DownEncoderBlockCausal3D",),
+        up_block_types: Tuple[str] = ("UpDecoderBlockCausal3D",),
+        block_out_channels: Tuple[int] = (64,),
+        layers_per_block: int = 1,
+        act_fn: str = "silu",
+        latent_channels: int = 4,
+        norm_num_groups: int = 32,
+        tile_sample_min_size: int = 256,
+        sample_tsize: int = 64,
+        overlap_factor: float = 0.25,
+        scaling_factor: float = 0.18215,
+        force_upcast: float = True,
+        spatial_compression_ratio: int = 8,
+        time_compression_ratio: int = 4,
+        mid_block_add_attention: bool = True,
+    ):
+        super().__init__()
+
+        self.time_compression_ratio = time_compression_ratio
+
+        self.encoder = EncoderCausal3D(
+            in_channels=in_channels,
+            out_channels=latent_channels,
+            down_block_types=down_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            act_fn=act_fn,
+            norm_num_groups=norm_num_groups,
+            double_z=True,
+            time_compression_ratio=time_compression_ratio,
+            spatial_compression_ratio=spatial_compression_ratio,
+            mid_block_add_attention=mid_block_add_attention,
+        )
+
+        self.decoder = DecoderCausal3D(
+            in_channels=latent_channels,
+            out_channels=out_channels,
+            up_block_types=up_block_types,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            norm_num_groups=norm_num_groups,
+            act_fn=act_fn,
+            time_compression_ratio=time_compression_ratio,
+            spatial_compression_ratio=spatial_compression_ratio,
+            mid_block_add_attention=mid_block_add_attention,
+        )
+
+        self.quant_conv = nn.Conv3d(
+            2 * latent_channels, 2 * latent_channels, kernel_size=1)
+        self.post_quant_conv = nn.Conv3d(
+            latent_channels, latent_channels, kernel_size=1)
+
+        self.use_slicing = False
+        self.use_spatial_tiling = False
+        self.use_temporal_tiling = False
+
+        self.sample_tsize = sample_tsize
+
+        # only relevant if vae tiling is enabled
+        self.tile_sample_min_tsize = self.sample_tsize
+        self.tile_latent_min_tsize = self.sample_tsize // time_compression_ratio
+
+        self.tile_sample_min_size = tile_sample_min_size
+
+        self.tile_latent_min_size = int(
+            self.tile_sample_min_size / (2 ** (len(self.config.block_out_channels) - 1)))
+
+        self.tile_overlap_factor = overlap_factor
+        self.t_tile_overlap_factor = overlap_factor
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (EncoderCausal3D, DecoderCausal3D)):
+            module.gradient_checkpointing = value
+
+    def enable_temporal_tiling(self, use_tiling: bool = True):
+        self.use_temporal_tiling = use_tiling
+
+    def disable_temporal_tiling(self):
+        self.enable_temporal_tiling(False)
+
+    def enable_spatial_tiling(self, use_tiling: bool = True):
+        self.use_spatial_tiling = use_tiling
+
+    def disable_spatial_tiling(self):
+        self.enable_spatial_tiling(False)
+
+    def enable_tiling(self, use_tiling: bool = True):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger videos.
+        """
+        self.enable_spatial_tiling(use_tiling)
+        self.enable_temporal_tiling(use_tiling)
+
+    def disable_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.disable_spatial_tiling()
+        self.disable_temporal_tiling()
+
+    def enable_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.use_slicing = True
+
+    def disable_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+        decoding in one step.
+        """
+        self.use_slicing = False
+
+    @property
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(
+                    return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(
+                    f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(
+        self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
+    ):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor, _remove_lora=_remove_lora)
+                else:
+                    module.set_processor(processor.pop(
+                        f"{name}.processor"), _remove_lora=_remove_lora)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(
+                    f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor, _remove_lora=True)
+
+    @apply_forward_hook
+    def encode(
+        self, x: torch.FloatTensor, return_dict: bool = True
+    ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
+        """
+        Encode a batch of images/videos into latents.
+
+        Args:
+            x (`torch.FloatTensor`): Input batch of images/videos.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+                The latent representations of the encoded images/videos. If `return_dict` is True, a
+                [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+        """
+        assert len(x.shape) == 5, "The input tensor should have 5 dimensions"
+
+        if self.use_temporal_tiling and x.shape[2] > self.tile_sample_min_tsize:
+            return self.temporal_tiled_encode(x, return_dict=return_dict)
+
+        if self.use_spatial_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size):
+            return self.spatial_tiled_encode(x, return_dict=return_dict)
+
+        if self.use_slicing and x.shape[0] > 1:
+            encoded_slices = [self.encoder(x_slice) for x_slice in x.split(1)]
+            h = torch.cat(encoded_slices)
+        else:
+            h = self.encoder(x)
+
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def _decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
+        assert len(z.shape) == 5, "The input tensor should have 5 dimensions"
+
+        if self.use_temporal_tiling and z.shape[2] > self.tile_latent_min_tsize:
+            return self.temporal_tiled_decode(z, return_dict=return_dict)
+
+        if self.use_spatial_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size):
+            return self.spatial_tiled_decode(z, return_dict=return_dict)
+
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z)
+
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    @apply_forward_hook
+    def decode(
+        self, z: torch.FloatTensor, return_dict: bool = True, generator=None
+    ) -> Union[DecoderOutput, torch.FloatTensor]:
+        """
+        Decode a batch of images/videos.
+
+        Args:
+            z (`torch.FloatTensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+
+        """        
+        if self.use_slicing and z.shape[0] > 1:
+            decoded_slices = [self._decode(
+                z_slice).sample for z_slice in z.split(1)]
+            decoded = torch.cat(decoded_slices)
+        else:
+            decoded = self._decode(z).sample
+
+        if not return_dict:
+            return (decoded,)
+
+        return DecoderOutput(sample=decoded)
+
+    def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[-2], b.shape[-2], blend_extent)
+        for y in range(blend_extent):
+            b[:, :, :, y, :] = a[:, :, :, -blend_extent + y, :] * \
+                (1 - y / blend_extent) + b[:, :, :, y, :] * (y / blend_extent)
+        return b
+
+    def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[-1], b.shape[-1], blend_extent)
+        for x in range(blend_extent):
+            b[:, :, :, :, x] = a[:, :, :, :, -blend_extent + x] * \
+                (1 - x / blend_extent) + b[:, :, :, :, x] * (x / blend_extent)
+        return b
+
+    def blend_t(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+        blend_extent = min(a.shape[-3], b.shape[-3], blend_extent)
+        for x in range(blend_extent):
+            b[:, :, x, :, :] = a[:, :, -blend_extent + x, :, :] * \
+                (1 - x / blend_extent) + b[:, :, x, :, :] * (x / blend_extent)
+        return b
+
+    def spatial_tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True, return_moments: bool = False) -> AutoencoderKLOutput:
+        r"""Encode a batch of images/videos using a tiled encoder.
+
+        When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+        steps. This is useful to keep memory use constant regardless of image/videos size. The end result of tiled encoding is
+        different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
+        tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
+        output, but they should be much less noticeable.
+
+        Args:
+            x (`torch.FloatTensor`): Input batch of images/videos.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`:
+                If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain
+                `tuple` is returned.
+        """
+        overlap_size = int(self.tile_sample_min_size *
+                           (1 - self.tile_overlap_factor))
+        blend_extent = int(self.tile_latent_min_size *
+                           self.tile_overlap_factor)
+        row_limit = self.tile_latent_min_size - blend_extent
+
+        # Split video into tiles and encode them separately.
+        rows = []
+        for i in range(0, x.shape[-2], overlap_size):
+            row = []
+            for j in range(0, x.shape[-1], overlap_size):
+                tile = x[:, :, :, i: i + self.tile_sample_min_size,
+                         j: j + self.tile_sample_min_size]
+                tile = self.encoder(tile)
+                tile = self.quant_conv(tile)
+                row.append(tile)
+            rows.append(row)
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                # blend the above tile and the left tile
+                # to the current tile and add the current tile to the result row
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :, :row_limit, :row_limit])
+            result_rows.append(torch.cat(result_row, dim=-1))
+
+        moments = torch.cat(result_rows, dim=-2)
+        if return_moments:
+            return moments
+
+        posterior = DiagonalGaussianDistribution(moments)
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def spatial_tiled_decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
+        r"""
+        Decode a batch of images/videos using a tiled decoder.
+
+        Args:
+            z (`torch.FloatTensor`): Input batch of latent vectors.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.vae.DecoderOutput`] or `tuple`:
+                If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+                returned.
+        """
+        overlap_size = int(self.tile_latent_min_size *
+                           (1 - self.tile_overlap_factor))
+        blend_extent = int(self.tile_sample_min_size *
+                           self.tile_overlap_factor)
+        row_limit = self.tile_sample_min_size - blend_extent
+
+        total_rows = (z.shape[-2] + overlap_size - 1) // overlap_size
+
+        # Split z into overlapping tiles with progress bar
+        rows = []
+        for i in range(0, z.shape[-2], overlap_size):
+            row = []
+            for j in range(0, z.shape[-1], overlap_size):
+                tile = z[:, :, :, i:i + self.tile_latent_min_size, j:j + self.tile_latent_min_size]
+                tile = self.post_quant_conv(tile)
+                decoded = self.decoder(tile)
+                row.append(decoded)
+            rows.append(row)
+
+        # Process results with progress bar
+        result_rows = []
+        for i, row in enumerate(rows):
+            result_row = []
+            for j, tile in enumerate(row):
+                if i > 0:
+                    tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+                if j > 0:
+                    tile = self.blend_h(row[j - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :, :row_limit, :row_limit])
+            result_rows.append(torch.cat(result_row, dim=-1))
+
+        dec = torch.cat(result_rows, dim=-2)
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    def temporal_tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True) -> AutoencoderKLOutput:
+
+        B, C, T, H, W = x.shape
+        overlap_size = int(self.tile_sample_min_tsize *
+                           (1 - self.t_tile_overlap_factor))
+        blend_extent = int(self.tile_latent_min_tsize *
+                           self.t_tile_overlap_factor)
+        t_limit = self.tile_latent_min_tsize - blend_extent
+
+        # Split the video into tiles and encode them separately.
+        row = []
+        for i in range(0, T, overlap_size):
+            tile = x[:, :, i: i + self.tile_sample_min_tsize + 1, :, :]
+            if self.use_spatial_tiling and (tile.shape[-1] > self.tile_sample_min_size or tile.shape[-2] > self.tile_sample_min_size):
+                tile = self.spatial_tiled_encode(tile, return_moments=True)
+            else:
+                tile = self.encoder(tile)
+                tile = self.quant_conv(tile)
+            if i > 0:
+                tile = tile[:, :, 1:, :, :]
+            row.append(tile)
+        result_row = []
+        for i, tile in enumerate(row):
+            if i > 0:
+                tile = self.blend_t(row[i - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :t_limit, :, :])
+            else:
+                result_row.append(tile[:, :, :t_limit+1, :, :])
+
+        moments = torch.cat(result_row, dim=2)
+        posterior = DiagonalGaussianDistribution(moments)
+
+        if not return_dict:
+            return (posterior,)
+
+        return AutoencoderKLOutput(latent_dist=posterior)
+
+    def temporal_tiled_decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
+        # Split z into overlapping tiles and decode them separately.
+
+        B, C, T, H, W = z.shape
+        overlap_size = int(self.tile_latent_min_tsize *
+                           (1 - self.t_tile_overlap_factor))
+        blend_extent = int(self.tile_sample_min_tsize *
+                           self.t_tile_overlap_factor)
+        t_limit = self.tile_sample_min_tsize - blend_extent
+
+        row = []
+        for i in range(0, T, overlap_size):
+            tile = z[:, :, i: i + self.tile_latent_min_tsize + 1, :, :]
+            if self.use_spatial_tiling and (tile.shape[-1] > self.tile_latent_min_size or tile.shape[-2] > self.tile_latent_min_size):
+                decoded = self.spatial_tiled_decode(
+                    tile, return_dict=True).sample
+            else:
+                tile = self.post_quant_conv(tile)
+                decoded = self.decoder(tile)
+            if i > 0:
+                decoded = decoded[:, :, 1:, :, :]
+            row.append(decoded)
+        result_row = []
+        for i, tile in enumerate(row):
+            if i > 0:
+                tile = self.blend_t(row[i - 1], tile, blend_extent)
+                result_row.append(tile[:, :, :t_limit, :, :])
+            else:
+                result_row.append(tile[:, :, :t_limit+1, :, :])
+
+        dec = torch.cat(result_row, dim=2)
+        if not return_dict:
+            return (dec,)
+
+        return DecoderOutput(sample=dec)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        sample_posterior: bool = False,
+        return_dict: bool = True,
+        return_posterior: bool = False,
+        generator: Optional[torch.Generator] = None,
+    ) -> Union[DecoderOutput2, torch.FloatTensor]:
+        r"""
+        Args:
+            sample (`torch.FloatTensor`): Input sample.
+            sample_posterior (`bool`, *optional*, defaults to `False`):
+                Whether to sample from the posterior.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+        """
+        x = sample
+        posterior = self.encode(x).latent_dist
+        if sample_posterior:
+            z = posterior.sample(generator=generator)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z).sample
+
+        if not return_dict:
+            if return_posterior:
+                return (dec, posterior)
+            else:
+                return (dec,)
+        if return_posterior:
+            return DecoderOutput2(sample=dec, posterior=posterior)
+        else:
+            return DecoderOutput2(sample=dec)
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError(
+                    "`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+    # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)

vae/unet_causal_3d_blocks.py

@@ -0,0 +1,797 @@
+# 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.
+# ==============================================================================
+#
+# Modified from diffusers==0.29.2
+#
+# ==============================================================================
+
+from typing import  Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from einops import rearrange
+
+from diffusers.utils import logging
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import SpatialNorm
+from diffusers.models.attention_processor import Attention
+from diffusers.models.normalization import AdaGroupNorm
+from diffusers.models.normalization import RMSNorm
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def prepare_causal_attention_mask(
+    n_frame: int, n_hw: int, dtype, device, batch_size: int | None = None
+):
+    indices = torch.arange(1, n_frame + 1, dtype=torch.int32, device=device)
+    indices_blocks = indices.repeat_interleave(n_hw)
+    x, y = torch.meshgrid(indices_blocks, indices_blocks, indexing="xy")
+    mask = torch.where(x <= y, 0, -float("inf")).to(dtype=dtype)
+    if batch_size is not None:
+        mask = mask.unsqueeze(0).expand(batch_size, -1, -1)
+    return mask
+
+
+class CausalConv3d(nn.Module):
+    """
+    Implements a causal 3D convolution layer where each position only depends on previous timesteps and current spatial locations.
+    This maintains temporal causality in video generation tasks.
+    """
+
+    def __init__(
+        self,
+        chan_in,
+        chan_out,
+        kernel_size: Union[int, Tuple[int, int, int]],
+        stride: Union[int, Tuple[int, int, int]] = 1,
+        dilation: Union[int, Tuple[int, int, int]] = 1,
+        pad_mode='replicate',
+        **kwargs
+    ):
+        super().__init__()
+
+        self.pad_mode = pad_mode
+        padding = (kernel_size // 2, kernel_size // 2, kernel_size //
+                   2, kernel_size // 2, kernel_size - 1, 0)  # W, H, T
+        self.time_causal_padding = padding
+
+        self.conv = nn.Conv3d(chan_in, chan_out, kernel_size,
+                              stride=stride, dilation=dilation, **kwargs)
+
+    def forward(self, x):
+        x = F.pad(x, self.time_causal_padding, mode=self.pad_mode)
+        return self.conv(x)
+
+
+class UpsampleCausal3D(nn.Module):
+    """
+    A 3D upsampling layer with an optional convolution.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        use_conv: bool = False,
+        use_conv_transpose: bool = False,
+        out_channels: Optional[int] = None,
+        name: str = "conv",
+        kernel_size: Optional[int] = None,
+        padding=1,
+        norm_type=None,
+        eps=None,
+        elementwise_affine=None,
+        bias=True,
+        interpolate=True,
+        upsample_factor=(2, 2, 2),
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+        self.interpolate = interpolate
+        self.upsample_factor = upsample_factor
+
+        if norm_type == "ln_norm":
+            self.norm = nn.LayerNorm(channels, eps, elementwise_affine)
+        elif norm_type == "rms_norm":
+            self.norm = RMSNorm(channels, eps, elementwise_affine)
+        elif norm_type is None:
+            self.norm = None
+        else:
+            raise ValueError(f"unknown norm_type: {norm_type}")
+
+        conv = None
+        if use_conv_transpose:
+            assert False, "Not Implement yet"
+            if kernel_size is None:
+                kernel_size = 4
+            conv = nn.ConvTranspose2d(
+                channels, self.out_channels, kernel_size=kernel_size, stride=2, padding=padding, bias=bias
+            )
+        elif use_conv:
+            if kernel_size is None:
+                kernel_size = 3
+            conv = CausalConv3d(self.channels, self.out_channels,
+                                kernel_size=kernel_size, bias=bias)
+
+        if name == "conv":
+            self.conv = conv
+        else:
+            self.Conv2d_0 = conv
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        output_size: Optional[int] = None,
+        scale: float = 1.0,
+    ) -> torch.FloatTensor:
+        assert hidden_states.shape[1] == self.channels
+
+        if self.norm is not None:
+            assert False, "Not Implement yet"
+            hidden_states = self.norm(
+                hidden_states.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
+
+        if self.use_conv_transpose:
+            return self.conv(hidden_states)
+
+        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
+        dtype = hidden_states.dtype
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.float32)
+
+        # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+        if hidden_states.shape[0] >= 64:
+            hidden_states = hidden_states.contiguous()
+
+        # if `output_size` is passed we force the interpolation output
+        # size and do not make use of `scale_factor=2`
+        if self.interpolate:
+            B, C, T, H, W = hidden_states.shape
+            first_h, other_h = hidden_states.split((1, T-1), dim=2)
+            if output_size is None:
+                if T > 1:
+                    other_h = F.interpolate(
+                        other_h, scale_factor=self.upsample_factor, mode="nearest")
+
+                first_h = first_h.squeeze(2)
+                first_h = F.interpolate(
+                    first_h, scale_factor=self.upsample_factor[1:], mode="nearest")
+                first_h = first_h.unsqueeze(2)
+            else:
+                assert False, "Not Implement yet"
+                other_h = F.interpolate(
+                    other_h, size=output_size, mode="nearest")
+
+            if T > 1:
+                hidden_states = torch.cat((first_h, other_h), dim=2)
+            else:
+                hidden_states = first_h
+
+        # If the input is bfloat16, we cast back to bfloat16
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(dtype)
+
+        if self.use_conv:
+            if self.name == "conv":
+                hidden_states = self.conv(hidden_states)
+            else:
+                hidden_states = self.Conv2d_0(hidden_states)
+
+        return hidden_states
+
+
+class DownsampleCausal3D(nn.Module):
+    """
+    A 3D downsampling layer with an optional convolution.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        use_conv: bool = False,
+        out_channels: Optional[int] = None,
+        padding: int = 1,
+        name: str = "conv",
+        kernel_size=3,
+        norm_type=None,
+        eps=None,
+        elementwise_affine=None,
+        bias=True,
+        stride=2,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.padding = padding
+        stride = stride
+        self.name = name
+
+        if norm_type == "ln_norm":
+            self.norm = nn.LayerNorm(channels, eps, elementwise_affine)
+        elif norm_type == "rms_norm":
+            self.norm = RMSNorm(channels, eps, elementwise_affine)
+        elif norm_type is None:
+            self.norm = None
+        else:
+            raise ValueError(f"unknown norm_type: {norm_type}")
+
+        if use_conv:
+            conv = CausalConv3d(
+                self.channels, self.out_channels, kernel_size=kernel_size, stride=stride, bias=bias
+            )
+        else:
+            raise NotImplementedError
+
+        if name == "conv":
+            self.Conv2d_0 = conv
+            self.conv = conv
+        elif name == "Conv2d_0":
+            self.conv = conv
+        else:
+            self.conv = conv
+
+    def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
+        assert hidden_states.shape[1] == self.channels
+
+        if self.norm is not None:
+            hidden_states = self.norm(
+                hidden_states.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
+
+        assert hidden_states.shape[1] == self.channels
+
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class ResnetBlockCausal3D(nn.Module):
+    r"""
+    A Resnet block.
+    """
+
+    def __init__(
+        self,
+        *,
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        conv_shortcut: bool = False,
+        dropout: float = 0.0,
+        temb_channels: int = 512,
+        groups: int = 32,
+        groups_out: Optional[int] = None,
+        pre_norm: bool = True,
+        eps: float = 1e-6,
+        non_linearity: str = "swish",
+        skip_time_act: bool = False,
+        # default, scale_shift, ada_group, spatial
+        time_embedding_norm: str = "default",
+        kernel: Optional[torch.FloatTensor] = None,
+        output_scale_factor: float = 1.0,
+        use_in_shortcut: Optional[bool] = None,
+        up: bool = False,
+        down: bool = False,
+        conv_shortcut_bias: bool = True,
+        conv_3d_out_channels: Optional[int] = None,
+    ):
+        super().__init__()
+        self.pre_norm = pre_norm
+        self.pre_norm = True
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+        self.up = up
+        self.down = down
+        self.output_scale_factor = output_scale_factor
+        self.time_embedding_norm = time_embedding_norm
+        self.skip_time_act = skip_time_act
+
+        linear_cls = nn.Linear
+
+        if groups_out is None:
+            groups_out = groups
+
+        if self.time_embedding_norm == "ada_group":
+            self.norm1 = AdaGroupNorm(
+                temb_channels, in_channels, groups, eps=eps)
+        elif self.time_embedding_norm == "spatial":
+            self.norm1 = SpatialNorm(in_channels, temb_channels)
+        else:
+            self.norm1 = torch.nn.GroupNorm(
+                num_groups=groups, num_channels=in_channels, eps=eps, affine=True)
+
+        self.conv1 = CausalConv3d(
+            in_channels, out_channels, kernel_size=3, stride=1)
+
+        if temb_channels is not None:
+            if self.time_embedding_norm == "default":
+                self.time_emb_proj = linear_cls(temb_channels, out_channels)
+            elif self.time_embedding_norm == "scale_shift":
+                self.time_emb_proj = linear_cls(
+                    temb_channels, 2 * out_channels)
+            elif self.time_embedding_norm == "ada_group" or self.time_embedding_norm == "spatial":
+                self.time_emb_proj = None
+            else:
+                raise ValueError(
+                    f"unknown time_embedding_norm : {self.time_embedding_norm} ")
+        else:
+            self.time_emb_proj = None
+
+        if self.time_embedding_norm == "ada_group":
+            self.norm2 = AdaGroupNorm(
+                temb_channels, out_channels, groups_out, eps=eps)
+        elif self.time_embedding_norm == "spatial":
+            self.norm2 = SpatialNorm(out_channels, temb_channels)
+        else:
+            self.norm2 = torch.nn.GroupNorm(
+                num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True)
+
+        self.dropout = torch.nn.Dropout(dropout)
+        conv_3d_out_channels = conv_3d_out_channels or out_channels
+        self.conv2 = CausalConv3d(
+            out_channels, conv_3d_out_channels, kernel_size=3, stride=1)
+
+        self.nonlinearity = get_activation(non_linearity)
+
+        self.upsample = self.downsample = None
+        if self.up:
+            self.upsample = UpsampleCausal3D(in_channels, use_conv=False)
+        elif self.down:
+            self.downsample = DownsampleCausal3D(
+                in_channels, use_conv=False, name="op")
+
+        self.use_in_shortcut = self.in_channels != conv_3d_out_channels if use_in_shortcut is None else use_in_shortcut
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = CausalConv3d(
+                in_channels,
+                conv_3d_out_channels,
+                kernel_size=1,
+                stride=1,
+                bias=conv_shortcut_bias,
+            )
+
+    def forward(
+        self,
+        input_tensor: torch.FloatTensor,
+        temb: torch.FloatTensor,
+        scale: float = 1.0,
+    ) -> torch.FloatTensor:
+        hidden_states = input_tensor
+
+        if self.time_embedding_norm == "ada_group" or self.time_embedding_norm == "spatial":
+            hidden_states = self.norm1(hidden_states, temb)
+        else:
+            hidden_states = self.norm1(hidden_states)
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        if self.upsample is not None:
+            # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+            if hidden_states.shape[0] >= 64:
+                input_tensor = input_tensor.contiguous()
+                hidden_states = hidden_states.contiguous()
+            input_tensor = (
+                self.upsample(input_tensor, scale=scale)
+            )
+            hidden_states = (
+                self.upsample(hidden_states, scale=scale)
+            )
+        elif self.downsample is not None:
+            input_tensor = (
+                self.downsample(input_tensor, scale=scale)
+            )
+            hidden_states = (
+                self.downsample(hidden_states, scale=scale)
+            )
+
+        hidden_states = self.conv1(hidden_states)
+
+        if self.time_emb_proj is not None:
+            if not self.skip_time_act:
+                temb = self.nonlinearity(temb)
+            temb = (
+                self.time_emb_proj(temb, scale)[:, :, None, None]
+            )
+
+        if temb is not None and self.time_embedding_norm == "default":
+            hidden_states = hidden_states + temb
+
+        if self.time_embedding_norm == "ada_group" or self.time_embedding_norm == "spatial":
+            hidden_states = self.norm2(hidden_states, temb)
+        else:
+            hidden_states = self.norm2(hidden_states)
+
+        if temb is not None and self.time_embedding_norm == "scale_shift":
+            scale, shift = torch.chunk(temb, 2, dim=1)
+            hidden_states = hidden_states * (1 + scale) + shift
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = (
+                self.conv_shortcut(input_tensor)
+            )
+
+        output_tensor = (input_tensor + hidden_states) / \
+            self.output_scale_factor
+
+        return output_tensor
+
+
+def get_down_block3d(
+    down_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    temb_channels: int,
+    add_downsample: bool,
+    downsample_stride: int,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    downsample_padding: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    resnet_out_scale_factor: float = 1.0,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = None,
+    downsample_type: Optional[str] = None,
+    dropout: float = 0.0,
+):
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    down_block_type = down_block_type[7:] if down_block_type.startswith(
+        "UNetRes") else down_block_type
+    if down_block_type == "DownEncoderBlockCausal3D":
+        return DownEncoderBlockCausal3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            dropout=dropout,
+            add_downsample=add_downsample,
+            downsample_stride=downsample_stride,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block3d(
+    up_block_type: str,
+    num_layers: int,
+    in_channels: int,
+    out_channels: int,
+    prev_output_channel: int,
+    temb_channels: int,
+    add_upsample: bool,
+    upsample_scale_factor: Tuple,
+    resnet_eps: float,
+    resnet_act_fn: str,
+    resolution_idx: Optional[int] = None,
+    transformer_layers_per_block: int = 1,
+    num_attention_heads: Optional[int] = None,
+    resnet_groups: Optional[int] = None,
+    cross_attention_dim: Optional[int] = None,
+    dual_cross_attention: bool = False,
+    use_linear_projection: bool = False,
+    only_cross_attention: bool = False,
+    upcast_attention: bool = False,
+    resnet_time_scale_shift: str = "default",
+    attention_type: str = "default",
+    resnet_skip_time_act: bool = False,
+    resnet_out_scale_factor: float = 1.0,
+    cross_attention_norm: Optional[str] = None,
+    attention_head_dim: Optional[int] = None,
+    upsample_type: Optional[str] = None,
+    dropout: float = 0.0,
+) -> nn.Module:
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    up_block_type = up_block_type[7:] if up_block_type.startswith(
+        "UNetRes") else up_block_type
+    if up_block_type == "UpDecoderBlockCausal3D":
+        return UpDecoderBlockCausal3D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            resolution_idx=resolution_idx,
+            dropout=dropout,
+            add_upsample=add_upsample,
+            upsample_scale_factor=upsample_scale_factor,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            temb_channels=temb_channels,
+        )
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class UNetMidBlockCausal3D(nn.Module):
+    """
+    A 3D UNet mid-block [`UNetMidBlockCausal3D`] with multiple residual blocks and optional attention blocks.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",  # default, spatial
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        attn_groups: Optional[int] = None,
+        resnet_pre_norm: bool = True,
+        add_attention: bool = True,
+        attention_head_dim: int = 1,
+        output_scale_factor: float = 1.0,
+    ):
+        super().__init__()
+        resnet_groups = resnet_groups if resnet_groups is not None else min(
+            in_channels // 4, 32)
+        self.add_attention = add_attention
+
+        if attn_groups is None:
+            attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None
+
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlockCausal3D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        if attention_head_dim is None:
+            logger.warn(
+                f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
+            )
+            attention_head_dim = in_channels
+
+        for _ in range(num_layers):
+            if self.add_attention:
+                # assert False, "Not implemented yet"
+                attentions.append(
+                    Attention(
+                        in_channels,
+                        heads=in_channels // attention_head_dim,
+                        dim_head=attention_head_dim,
+                        rescale_output_factor=output_scale_factor,
+                        eps=resnet_eps,
+                        norm_num_groups=attn_groups,
+                        spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
+                        residual_connection=True,
+                        bias=True,
+                        upcast_softmax=True,
+                        _from_deprecated_attn_block=True,
+                    )
+                )
+            else:
+                attentions.append(None)
+
+            resnets.append(
+                ResnetBlockCausal3D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            if attn is not None:
+                B, C, T, H, W = hidden_states.shape
+                hidden_states = rearrange(
+                    hidden_states, "b c f h w -> b (f h w) c")
+                attention_mask = prepare_causal_attention_mask(
+                    T, H * W, hidden_states.dtype, hidden_states.device, batch_size=B)
+                hidden_states = attn(
+                    hidden_states, temb=temb, attention_mask=attention_mask)
+                hidden_states = rearrange(
+                    hidden_states, "b (f h w) c -> b c f h w", f=T, h=H, w=W)
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class DownEncoderBlockCausal3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_downsample: bool = True,
+        downsample_stride: int = 2,
+        downsample_padding: int = 1,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlockCausal3D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=None,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    DownsampleCausal3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        padding=downsample_padding,
+                        name="op",
+                        stride=downsample_stride,
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+    def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
+        for resnet in self.resnets:
+            hidden_states = resnet(hidden_states, temb=None, scale=scale)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states, scale)
+
+        return hidden_states
+
+
+class UpDecoderBlockCausal3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        resolution_idx: Optional[int] = None,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",  # default, spatial
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        output_scale_factor: float = 1.0,
+        add_upsample: bool = True,
+        upsample_scale_factor=(2, 2, 2),
+        temb_channels: Optional[int] = None,
+    ):
+        super().__init__()
+        resnets = []
+
+        for i in range(num_layers):
+            input_channels = in_channels if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlockCausal3D(
+                    in_channels=input_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList(
+                [
+                    UpsampleCausal3D(
+                        out_channels,
+                        use_conv=True,
+                        out_channels=out_channels,
+                        upsample_factor=upsample_scale_factor,
+                    )
+                ]
+            )
+        else:
+            self.upsamplers = None
+
+        self.resolution_idx = resolution_idx
+
+    def forward(
+        self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
+    ) -> torch.FloatTensor:
+        for resnet in self.resnets:
+            hidden_states = resnet(hidden_states, temb=temb, scale=scale)
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states)
+
+        return hidden_states

vae/vae.py

@@ -0,0 +1,373 @@
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+from diffusers.utils import BaseOutput, is_torch_version
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.models.attention_processor import SpatialNorm
+from .unet_causal_3d_blocks import (
+    CausalConv3d,
+    UNetMidBlockCausal3D,
+    get_down_block3d,
+    get_up_block3d,
+)
+
+
+@dataclass
+class DecoderOutput(BaseOutput):
+    r"""
+    Output of decoding method.
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The decoded output sample from the last layer of the model.
+    """
+
+    sample: torch.FloatTensor
+
+
+class EncoderCausal3D(nn.Module):
+    r"""
+    The `EncoderCausal3D` layer of a variational autoencoder that encodes its input into a latent representation.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        down_block_types: Tuple[str, ...] = ("DownEncoderBlockCausal3D",),
+        block_out_channels: Tuple[int, ...] = (64,),
+        layers_per_block: int = 2,
+        norm_num_groups: int = 32,
+        act_fn: str = "silu",
+        double_z: bool = True,
+        mid_block_add_attention=True,
+        time_compression_ratio: int = 4,
+        spatial_compression_ratio: int = 8,
+    ):
+        super().__init__()
+        self.layers_per_block = layers_per_block
+
+        self.conv_in = CausalConv3d(
+            in_channels, block_out_channels[0], kernel_size=3, stride=1)
+        self.mid_block = None
+        self.down_blocks = nn.ModuleList([])
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+            num_spatial_downsample_layers = int(
+                np.log2(spatial_compression_ratio))
+            num_time_downsample_layers = int(np.log2(time_compression_ratio))
+
+            if time_compression_ratio == 4:
+                add_spatial_downsample = bool(
+                    i < num_spatial_downsample_layers)
+                add_time_downsample = bool(i >= (
+                    len(block_out_channels) - 1 - num_time_downsample_layers) and not is_final_block)
+            elif time_compression_ratio == 8:
+                add_spatial_downsample = bool(
+                    i < num_spatial_downsample_layers)
+                add_time_downsample = bool(i < num_time_downsample_layers)
+            else:
+                raise ValueError(
+                    f"Unsupported time_compression_ratio: {time_compression_ratio}")
+
+            downsample_stride_HW = (2, 2) if add_spatial_downsample else (1, 1)
+            downsample_stride_T = (2, ) if add_time_downsample else (1, )
+            downsample_stride = tuple(
+                downsample_stride_T + downsample_stride_HW)
+            down_block = get_down_block3d(
+                down_block_type,
+                num_layers=self.layers_per_block,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                add_downsample=bool(
+                    add_spatial_downsample or add_time_downsample),
+                downsample_stride=downsample_stride,
+                resnet_eps=1e-6,
+                downsample_padding=0,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                attention_head_dim=output_channel,
+                temb_channels=None,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = UNetMidBlockCausal3D(
+            in_channels=block_out_channels[-1],
+            resnet_eps=1e-6,
+            resnet_act_fn=act_fn,
+            output_scale_factor=1,
+            resnet_time_scale_shift="default",
+            attention_head_dim=block_out_channels[-1],
+            resnet_groups=norm_num_groups,
+            temb_channels=None,
+            add_attention=mid_block_add_attention,
+        )
+
+        # out
+        self.conv_norm_out = nn.GroupNorm(
+            num_channels=block_out_channels[-1], num_groups=norm_num_groups, eps=1e-6)
+        self.conv_act = nn.SiLU()
+
+        conv_out_channels = 2 * out_channels if double_z else out_channels
+        self.conv_out = CausalConv3d(
+            block_out_channels[-1], conv_out_channels, kernel_size=3)
+
+    def forward(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+        r"""The forward method of the `EncoderCausal3D` class."""
+        assert len(sample.shape) == 5, "The input tensor should have 5 dimensions"
+
+        sample = self.conv_in(sample)
+
+        # down
+        for down_block in self.down_blocks:
+            sample = down_block(sample)
+
+        # middle
+        sample = self.mid_block(sample)
+
+        # post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        return sample
+
+
+class DecoderCausal3D(nn.Module):
+    r"""
+    The `DecoderCausal3D` layer of a variational autoencoder that decodes its latent representation into an output sample.
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        up_block_types: Tuple[str, ...] = ("UpDecoderBlockCausal3D",),
+        block_out_channels: Tuple[int, ...] = (64,),
+        layers_per_block: int = 2,
+        norm_num_groups: int = 32,
+        act_fn: str = "silu",
+        norm_type: str = "group",  # group, spatial
+        mid_block_add_attention=True,
+        time_compression_ratio: int = 4,
+        spatial_compression_ratio: int = 8,
+    ):
+        super().__init__()
+        self.layers_per_block = layers_per_block
+
+        self.conv_in = CausalConv3d(
+            in_channels, block_out_channels[-1], kernel_size=3, stride=1)
+        self.mid_block = None
+        self.up_blocks = nn.ModuleList([])
+
+        temb_channels = in_channels if norm_type == "spatial" else None
+
+        # mid
+        self.mid_block = UNetMidBlockCausal3D(
+            in_channels=block_out_channels[-1],
+            resnet_eps=1e-6,
+            resnet_act_fn=act_fn,
+            output_scale_factor=1,
+            resnet_time_scale_shift="default" if norm_type == "group" else norm_type,
+            attention_head_dim=block_out_channels[-1],
+            resnet_groups=norm_num_groups,
+            temb_channels=temb_channels,
+            add_attention=mid_block_add_attention,
+        )
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+            num_spatial_upsample_layers = int(
+                np.log2(spatial_compression_ratio))
+            num_time_upsample_layers = int(np.log2(time_compression_ratio))
+
+            if time_compression_ratio == 4:
+                add_spatial_upsample = bool(i < num_spatial_upsample_layers)
+                add_time_upsample = bool(i >= len(
+                    block_out_channels) - 1 - num_time_upsample_layers and not is_final_block)
+            else:
+                raise ValueError(
+                    f"Unsupported time_compression_ratio: {time_compression_ratio}")
+
+            upsample_scale_factor_HW = (
+                2, 2) if add_spatial_upsample else (1, 1)
+            upsample_scale_factor_T = (2, ) if add_time_upsample else (1, )
+            upsample_scale_factor = tuple(
+                upsample_scale_factor_T + upsample_scale_factor_HW)
+            up_block = get_up_block3d(
+                up_block_type,
+                num_layers=self.layers_per_block + 1,
+                in_channels=prev_output_channel,
+                out_channels=output_channel,
+                prev_output_channel=None,
+                add_upsample=bool(add_spatial_upsample or add_time_upsample),
+                upsample_scale_factor=upsample_scale_factor,
+                resnet_eps=1e-6,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                attention_head_dim=output_channel,
+                temb_channels=temb_channels,
+                resnet_time_scale_shift=norm_type,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        if norm_type == "spatial":
+            self.conv_norm_out = SpatialNorm(
+                block_out_channels[0], temb_channels)
+        else:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-6)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CausalConv3d(
+            block_out_channels[0], out_channels, kernel_size=3)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        latent_embeds: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        r"""The forward method of the `DecoderCausal3D` class."""
+        assert len(sample.shape) == 5, "The input tensor should have 5 dimensions"
+
+        sample = self.conv_in(sample)
+
+        upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
+        if self.training and self.gradient_checkpointing:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    return module(*inputs)
+
+                return custom_forward
+
+            if is_torch_version(">=", "1.11.0"):
+                # middle
+                sample = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(self.mid_block),
+                    sample,
+                    latent_embeds,
+                    use_reentrant=False,
+                )
+                sample = sample.to(upscale_dtype)
+
+                # up
+                for up_block in self.up_blocks:
+                    sample = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(up_block),
+                        sample,
+                        latent_embeds,
+                        use_reentrant=False,
+                    )
+            else:
+                # middle
+                sample = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(
+                        self.mid_block), sample, latent_embeds
+                )
+                sample = sample.to(upscale_dtype)
+
+                # up
+                for up_block in self.up_blocks:
+                    sample = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(up_block), sample, latent_embeds)
+        else:
+            # middle
+            sample = self.mid_block(sample, latent_embeds)
+            sample = sample.to(upscale_dtype)
+
+            # up
+            for up_block in self.up_blocks:
+                sample = up_block(sample, latent_embeds)
+
+        # post-process
+        if latent_embeds is None:
+            sample = self.conv_norm_out(sample)
+        else:
+            sample = self.conv_norm_out(sample, latent_embeds)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        return sample
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters: torch.Tensor, deterministic: bool = False):
+        if parameters.ndim == 3:
+            dim = 2  # (B, L, C)
+        elif parameters.ndim == 5 or parameters.ndim == 4:
+            dim = 1  # (B, C, T, H ,W) / (B, C, H, W)
+        else:
+            raise NotImplementedError
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=dim)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(
+                self.mean, device=self.parameters.device, dtype=self.parameters.dtype
+            )
+
+    def sample(self, generator: Optional[torch.Generator] = None) -> torch.FloatTensor:
+        # make sure sample is on the same device as the parameters and has same dtype
+        sample = randn_tensor(
+            self.mean.shape,
+            generator=generator,
+            device=self.parameters.device,
+            dtype=self.parameters.dtype,
+        )
+        x = self.mean + self.std * sample
+        return x
+
+    def kl(self, other: "DiagonalGaussianDistribution" = None) -> torch.Tensor:
+        if self.deterministic:
+            return torch.Tensor([0.0])
+        else:
+            reduce_dim = list(range(1, self.mean.ndim))
+            if other is None:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean, 2) + self.var - 1.0 - self.logvar,
+                    dim=reduce_dim,
+                )
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var
+                    - 1.0
+                    - self.logvar
+                    + other.logvar,
+                    dim=reduce_dim,
+                )
+
+    def nll(self, sample: torch.Tensor, dims: Tuple[int, ...] = [1, 2, 3]) -> torch.Tensor:
+        if self.deterministic:
+            return torch.Tensor([0.0])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar +
+            torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims,
+        )
+
+    def mode(self) -> torch.Tensor:
+        return self.mean