diffusers 0.32.2__py3-none-any.whl → 0.33.1__py3-none-any.whl

diffusers/models/transformers/transformer_2d.py

@@ -18,7 +18,7 @@ import torch.nn.functional as F
 from torch import nn
 
 from ...configuration_utils import LegacyConfigMixin, register_to_config
-from ...utils import deprecate, is_torch_version, logging
+from ...utils import deprecate, logging
 from ..attention import BasicTransformerBlock
 from ..embeddings import ImagePositionalEmbeddings, PatchEmbed, PixArtAlphaTextProjection
 from ..modeling_outputs import Transformer2DModelOutput
@@ -66,6 +66,7 @@ class Transformer2DModel(LegacyModelMixin, LegacyConfigMixin):
 
     _supports_gradient_checkpointing = True
     _no_split_modules = ["BasicTransformerBlock"]
+    _skip_layerwise_casting_patterns = ["latent_image_embedding", "norm"]
 
     @register_to_config
     def __init__(
@@ -210,9 +211,9 @@ class Transformer2DModel(LegacyModelMixin, LegacyConfigMixin):
 
     def _init_vectorized_inputs(self, norm_type):
         assert self.config.sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
-        assert (
-            self.config.num_vector_embeds is not None
-        ), "Transformer2DModel over discrete input must provide num_embed"
+        assert self.config.num_vector_embeds is not None, (
+            "Transformer2DModel over discrete input must provide num_embed"
+        )
 
         self.height = self.config.sample_size
         self.width = self.config.sample_size
@@ -320,10 +321,6 @@ class Transformer2DModel(LegacyModelMixin, LegacyConfigMixin):
                 in_features=self.caption_channels, hidden_size=self.inner_dim
             )
 
-    def _set_gradient_checkpointing(self, module, value=False):
-        if hasattr(module, "gradient_checkpointing"):
-            module.gradient_checkpointing = value
-
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -416,19 +413,8 @@ class Transformer2DModel(LegacyModelMixin, LegacyConfigMixin):
         # 2. Blocks
         for block in self.transformer_blocks:
             if torch.is_grad_enabled() and self.gradient_checkpointing:
-
-                def create_custom_forward(module, return_dict=None):
-                    def custom_forward(*inputs):
-                        if return_dict is not None:
-                            return module(*inputs, return_dict=return_dict)
-                        else:
-                            return module(*inputs)
-
-                    return custom_forward
-
-                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
-                hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(block),
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
                     hidden_states,
                     attention_mask,
                     encoder_hidden_states,
@@ -436,7 +422,6 @@ class Transformer2DModel(LegacyModelMixin, LegacyConfigMixin):
                     timestep,
                     cross_attention_kwargs,
                     class_labels,
-                    **ckpt_kwargs,
                 )
             else:
                 hidden_states = block(

diffusers/models/transformers/transformer_allegro.py

@@ -13,17 +13,18 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any, Dict, Optional, Tuple
+from typing import Optional, Tuple
 
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
-from ...utils import is_torch_version, logging
+from ...utils import logging
 from ...utils.torch_utils import maybe_allow_in_graph
 from ..attention import FeedForward
 from ..attention_processor import AllegroAttnProcessor2_0, Attention
+from ..cache_utils import CacheMixin
 from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
 from ..modeling_outputs import Transformer2DModelOutput
 from ..modeling_utils import ModelMixin
@@ -172,7 +173,7 @@ class AllegroTransformerBlock(nn.Module):
         return hidden_states
 
 
-class AllegroTransformer3DModel(ModelMixin, ConfigMixin):
+class AllegroTransformer3DModel(ModelMixin, ConfigMixin, CacheMixin):
     _supports_gradient_checkpointing = True
 
     """
@@ -221,6 +222,9 @@ class AllegroTransformer3DModel(ModelMixin, ConfigMixin):
             Scaling factor to apply in 3D positional embeddings across time dimension.
     """
 
+    _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["pos_embed", "norm", "adaln_single"]
+
     @register_to_config
     def __init__(
         self,
@@ -300,9 +304,6 @@ class AllegroTransformer3DModel(ModelMixin, ConfigMixin):
 
         self.gradient_checkpointing = False
 
-    def _set_gradient_checkpointing(self, module, value=False):
-        self.gradient_checkpointing = value
-
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -372,23 +373,14 @@ class AllegroTransformer3DModel(ModelMixin, ConfigMixin):
         for i, block in enumerate(self.transformer_blocks):
             # TODO(aryan): Implement gradient checkpointing
             if torch.is_grad_enabled() and self.gradient_checkpointing:
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        return module(*inputs)
-
-                    return custom_forward
-
-                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
-                hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(block),
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
                     hidden_states,
                     encoder_hidden_states,
                     timestep,
                     attention_mask,
                     encoder_attention_mask,
                     image_rotary_emb,
-                    **ckpt_kwargs,
                 )
             else:
                 hidden_states = block(

diffusers/models/transformers/transformer_cogview3plus.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 
-from typing import Any, Dict, Union
+from typing import Dict, Union
 
 import torch
 import torch.nn as nn
@@ -27,7 +27,7 @@ from ...models.attention_processor import (
 )
 from ...models.modeling_utils import ModelMixin
 from ...models.normalization import AdaLayerNormContinuous
-from ...utils import is_torch_version, logging
+from ...utils import logging
 from ..embeddings import CogView3CombinedTimestepSizeEmbeddings, CogView3PlusPatchEmbed
 from ..modeling_outputs import Transformer2DModelOutput
 from ..normalization import CogView3PlusAdaLayerNormZeroTextImage
@@ -166,6 +166,8 @@ class CogView3PlusTransformer2DModel(ModelMixin, ConfigMixin):
     """
 
     _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["patch_embed", "norm"]
+    _no_split_modules = ["CogView3PlusTransformerBlock", "CogView3PlusPatchEmbed"]
 
     @register_to_config
     def __init__(
@@ -287,10 +289,6 @@ class CogView3PlusTransformer2DModel(ModelMixin, ConfigMixin):
         for name, module in self.named_children():
             fn_recursive_attn_processor(name, module, processor)
 
-    def _set_gradient_checkpointing(self, module, value=False):
-        if hasattr(module, "gradient_checkpointing"):
-            module.gradient_checkpointing = value
-
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -342,20 +340,11 @@ class CogView3PlusTransformer2DModel(ModelMixin, ConfigMixin):
 
         for index_block, block in enumerate(self.transformer_blocks):
             if torch.is_grad_enabled() and self.gradient_checkpointing:
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        return module(*inputs)
-
-                    return custom_forward
-
-                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
-                hidden_states, encoder_hidden_states = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(block),
+                hidden_states, encoder_hidden_states = self._gradient_checkpointing_func(
+                    block,
                     hidden_states,
                     encoder_hidden_states,
                     emb,
-                    **ckpt_kwargs,
                 )
             else:
                 hidden_states, encoder_hidden_states = block(

diffusers/models/transformers/transformer_cogview4.py

@@ -0,0 +1,462 @@
+# Copyright 2024 The CogView team, Tsinghua University & ZhipuAI and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# 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.
+
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from ..attention import FeedForward
+from ..attention_processor import Attention
+from ..cache_utils import CacheMixin
+from ..embeddings import CogView3CombinedTimestepSizeEmbeddings
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormContinuous
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class CogView4PatchEmbed(nn.Module):
+    def __init__(
+        self,
+        in_channels: int = 16,
+        hidden_size: int = 2560,
+        patch_size: int = 2,
+        text_hidden_size: int = 4096,
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+
+        self.proj = nn.Linear(in_channels * patch_size**2, hidden_size)
+        self.text_proj = nn.Linear(text_hidden_size, hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, channel, height, width = hidden_states.shape
+        post_patch_height = height // self.patch_size
+        post_patch_width = width // self.patch_size
+
+        hidden_states = hidden_states.reshape(
+            batch_size, channel, post_patch_height, self.patch_size, post_patch_width, self.patch_size
+        )
+        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5).flatten(3, 5).flatten(1, 2)
+        hidden_states = self.proj(hidden_states)
+        encoder_hidden_states = self.text_proj(encoder_hidden_states)
+
+        return hidden_states, encoder_hidden_states
+
+
+class CogView4AdaLayerNormZero(nn.Module):
+    def __init__(self, embedding_dim: int, dim: int) -> None:
+        super().__init__()
+
+        self.norm = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5)
+        self.norm_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5)
+        self.linear = nn.Linear(embedding_dim, 12 * dim, bias=True)
+
+    def forward(
+        self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor, temb: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        norm_hidden_states = self.norm(hidden_states)
+        norm_encoder_hidden_states = self.norm_context(encoder_hidden_states)
+
+        emb = self.linear(temb)
+        (
+            shift_msa,
+            c_shift_msa,
+            scale_msa,
+            c_scale_msa,
+            gate_msa,
+            c_gate_msa,
+            shift_mlp,
+            c_shift_mlp,
+            scale_mlp,
+            c_scale_mlp,
+            gate_mlp,
+            c_gate_mlp,
+        ) = emb.chunk(12, dim=1)
+
+        hidden_states = norm_hidden_states * (1 + scale_msa.unsqueeze(1)) + shift_msa.unsqueeze(1)
+        encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_msa.unsqueeze(1)) + c_shift_msa.unsqueeze(1)
+
+        return (
+            hidden_states,
+            gate_msa,
+            shift_mlp,
+            scale_mlp,
+            gate_mlp,
+            encoder_hidden_states,
+            c_gate_msa,
+            c_shift_mlp,
+            c_scale_mlp,
+            c_gate_mlp,
+        )
+
+
+class CogView4AttnProcessor:
+    """
+    Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
+    query and key vectors, but does not include spatial normalization.
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("CogView4AttnProcessor requires PyTorch 2.0. To use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, text_seq_length, embed_dim = encoder_hidden_states.shape
+        batch_size, image_seq_length, embed_dim = hidden_states.shape
+        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+
+        # 1. QKV projections
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        query = query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        key = key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        value = value.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+        # 2. QK normalization
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # 3. Rotational positional embeddings applied to latent stream
+        if image_rotary_emb is not None:
+            from ..embeddings import apply_rotary_emb
+
+            query[:, :, text_seq_length:, :] = apply_rotary_emb(
+                query[:, :, text_seq_length:, :], image_rotary_emb, use_real_unbind_dim=-2
+            )
+            key[:, :, text_seq_length:, :] = apply_rotary_emb(
+                key[:, :, text_seq_length:, :], image_rotary_emb, use_real_unbind_dim=-2
+            )
+
+        # 4. Attention
+        if attention_mask is not None:
+            text_attention_mask = attention_mask.float().to(query.device)
+            actual_text_seq_length = text_attention_mask.size(1)
+            new_attention_mask = torch.zeros((batch_size, text_seq_length + image_seq_length), device=query.device)
+            new_attention_mask[:, :actual_text_seq_length] = text_attention_mask
+            new_attention_mask = new_attention_mask.unsqueeze(2)
+            attention_mask_matrix = new_attention_mask @ new_attention_mask.transpose(1, 2)
+            attention_mask = (attention_mask_matrix > 0).unsqueeze(1).to(query.dtype)
+
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+        hidden_states = hidden_states.transpose(1, 2).flatten(2, 3)
+        hidden_states = hidden_states.type_as(query)
+
+        # 5. Output projection
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+
+        encoder_hidden_states, hidden_states = hidden_states.split(
+            [text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
+        )
+        return hidden_states, encoder_hidden_states
+
+
+class CogView4TransformerBlock(nn.Module):
+    def __init__(
+        self, dim: int = 2560, num_attention_heads: int = 64, attention_head_dim: int = 40, time_embed_dim: int = 512
+    ) -> None:
+        super().__init__()
+
+        # 1. Attention
+        self.norm1 = CogView4AdaLayerNormZero(time_embed_dim, dim)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            out_dim=dim,
+            bias=True,
+            qk_norm="layer_norm",
+            elementwise_affine=False,
+            eps=1e-5,
+            processor=CogView4AttnProcessor(),
+        )
+
+        # 2. Feedforward
+        self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5)
+        self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-5)
+        self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        # 1. Timestep conditioning
+        (
+            norm_hidden_states,
+            gate_msa,
+            shift_mlp,
+            scale_mlp,
+            gate_mlp,
+            norm_encoder_hidden_states,
+            c_gate_msa,
+            c_shift_mlp,
+            c_scale_mlp,
+            c_gate_mlp,
+        ) = self.norm1(hidden_states, encoder_hidden_states, temb)
+
+        # 2. Attention
+        attn_hidden_states, attn_encoder_hidden_states = self.attn1(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            image_rotary_emb=image_rotary_emb,
+            attention_mask=attention_mask,
+            **kwargs,
+        )
+        hidden_states = hidden_states + attn_hidden_states * gate_msa.unsqueeze(1)
+        encoder_hidden_states = encoder_hidden_states + attn_encoder_hidden_states * c_gate_msa.unsqueeze(1)
+
+        # 3. Feedforward
+        norm_hidden_states = self.norm2(hidden_states) * (1 + scale_mlp.unsqueeze(1)) + shift_mlp.unsqueeze(1)
+        norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states) * (
+            1 + c_scale_mlp.unsqueeze(1)
+        ) + c_shift_mlp.unsqueeze(1)
+
+        ff_output = self.ff(norm_hidden_states)
+        ff_output_context = self.ff(norm_encoder_hidden_states)
+        hidden_states = hidden_states + ff_output * gate_mlp.unsqueeze(1)
+        encoder_hidden_states = encoder_hidden_states + ff_output_context * c_gate_mlp.unsqueeze(1)
+
+        return hidden_states, encoder_hidden_states
+
+
+class CogView4RotaryPosEmbed(nn.Module):
+    def __init__(self, dim: int, patch_size: int, rope_axes_dim: Tuple[int, int], theta: float = 10000.0) -> None:
+        super().__init__()
+
+        self.dim = dim
+        self.patch_size = patch_size
+        self.rope_axes_dim = rope_axes_dim
+        self.theta = theta
+
+    def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        batch_size, num_channels, height, width = hidden_states.shape
+        height, width = height // self.patch_size, width // self.patch_size
+
+        dim_h, dim_w = self.dim // 2, self.dim // 2
+        h_inv_freq = 1.0 / (
+            self.theta ** (torch.arange(0, dim_h, 2, dtype=torch.float32)[: (dim_h // 2)].float() / dim_h)
+        )
+        w_inv_freq = 1.0 / (
+            self.theta ** (torch.arange(0, dim_w, 2, dtype=torch.float32)[: (dim_w // 2)].float() / dim_w)
+        )
+        h_seq = torch.arange(self.rope_axes_dim[0])
+        w_seq = torch.arange(self.rope_axes_dim[1])
+        freqs_h = torch.outer(h_seq, h_inv_freq)
+        freqs_w = torch.outer(w_seq, w_inv_freq)
+
+        h_idx = torch.arange(height, device=freqs_h.device)
+        w_idx = torch.arange(width, device=freqs_w.device)
+        inner_h_idx = h_idx * self.rope_axes_dim[0] // height
+        inner_w_idx = w_idx * self.rope_axes_dim[1] // width
+
+        freqs_h = freqs_h[inner_h_idx]
+        freqs_w = freqs_w[inner_w_idx]
+
+        # Create position matrices for height and width
+        # [height, 1, dim//4] and [1, width, dim//4]
+        freqs_h = freqs_h.unsqueeze(1)
+        freqs_w = freqs_w.unsqueeze(0)
+        # Broadcast freqs_h and freqs_w to [height, width, dim//4]
+        freqs_h = freqs_h.expand(height, width, -1)
+        freqs_w = freqs_w.expand(height, width, -1)
+
+        # Concatenate along last dimension to get [height, width, dim//2]
+        freqs = torch.cat([freqs_h, freqs_w], dim=-1)
+        freqs = torch.cat([freqs, freqs], dim=-1)  # [height, width, dim]
+        freqs = freqs.reshape(height * width, -1)
+        return (freqs.cos(), freqs.sin())
+
+
+class CogView4Transformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, CacheMixin):
+    r"""
+    Args:
+        patch_size (`int`, defaults to `2`):
+            The size of the patches to use in the patch embedding layer.
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        num_layers (`int`, defaults to `30`):
+            The number of layers of Transformer blocks to use.
+        attention_head_dim (`int`, defaults to `40`):
+            The number of channels in each head.
+        num_attention_heads (`int`, defaults to `64`):
+            The number of heads to use for multi-head attention.
+        out_channels (`int`, defaults to `16`):
+            The number of channels in the output.
+        text_embed_dim (`int`, defaults to `4096`):
+            Input dimension of text embeddings from the text encoder.
+        time_embed_dim (`int`, defaults to `512`):
+            Output dimension of timestep embeddings.
+        condition_dim (`int`, defaults to `256`):
+            The embedding dimension of the input SDXL-style resolution conditions (original_size, target_size,
+            crop_coords).
+        pos_embed_max_size (`int`, defaults to `128`):
+            The maximum resolution of the positional embeddings, from which slices of shape `H x W` are taken and added
+            to input patched latents, where `H` and `W` are the latent height and width respectively. A value of 128
+            means that the maximum supported height and width for image generation is `128 * vae_scale_factor *
+            patch_size => 128 * 8 * 2 => 2048`.
+        sample_size (`int`, defaults to `128`):
+            The base resolution of input latents. If height/width is not provided during generation, this value is used
+            to determine the resolution as `sample_size * vae_scale_factor => 128 * 8 => 1024`
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["CogView4TransformerBlock", "CogView4PatchEmbed", "CogView4PatchEmbed"]
+    _skip_layerwise_casting_patterns = ["patch_embed", "norm", "proj_out"]
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        out_channels: int = 16,
+        num_layers: int = 30,
+        attention_head_dim: int = 40,
+        num_attention_heads: int = 64,
+        text_embed_dim: int = 4096,
+        time_embed_dim: int = 512,
+        condition_dim: int = 256,
+        pos_embed_max_size: int = 128,
+        sample_size: int = 128,
+        rope_axes_dim: Tuple[int, int] = (256, 256),
+    ):
+        super().__init__()
+
+        # CogView4 uses 3 additional SDXL-like conditions - original_size, target_size, crop_coords
+        # Each of these are sincos embeddings of shape 2 * condition_dim
+        pooled_projection_dim = 3 * 2 * condition_dim
+        inner_dim = num_attention_heads * attention_head_dim
+        out_channels = out_channels
+
+        # 1. RoPE
+        self.rope = CogView4RotaryPosEmbed(attention_head_dim, patch_size, rope_axes_dim, theta=10000.0)
+
+        # 2. Patch & Text-timestep embedding
+        self.patch_embed = CogView4PatchEmbed(in_channels, inner_dim, patch_size, text_embed_dim)
+
+        self.time_condition_embed = CogView3CombinedTimestepSizeEmbeddings(
+            embedding_dim=time_embed_dim,
+            condition_dim=condition_dim,
+            pooled_projection_dim=pooled_projection_dim,
+            timesteps_dim=inner_dim,
+        )
+
+        # 3. Transformer blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                CogView4TransformerBlock(inner_dim, num_attention_heads, attention_head_dim, time_embed_dim)
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Output projection
+        self.norm_out = AdaLayerNormContinuous(inner_dim, time_embed_dim, elementwise_affine=False)
+        self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * out_channels, bias=True)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        original_size: torch.Tensor,
+        target_size: torch.Tensor,
+        crop_coords: torch.Tensor,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Union[torch.Tensor, Transformer2DModelOutput]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        batch_size, num_channels, height, width = hidden_states.shape
+
+        # 1. RoPE
+        image_rotary_emb = self.rope(hidden_states)
+
+        # 2. Patch & Timestep embeddings
+        p = self.config.patch_size
+        post_patch_height = height // p
+        post_patch_width = width // p
+
+        hidden_states, encoder_hidden_states = self.patch_embed(hidden_states, encoder_hidden_states)
+
+        temb = self.time_condition_embed(timestep, original_size, target_size, crop_coords, hidden_states.dtype)
+        temb = F.silu(temb)
+
+        # 3. Transformer blocks
+        for block in self.transformer_blocks:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                hidden_states, encoder_hidden_states = self._gradient_checkpointing_func(
+                    block, hidden_states, encoder_hidden_states, temb, image_rotary_emb, attention_mask, **kwargs
+                )
+            else:
+                hidden_states, encoder_hidden_states = block(
+                    hidden_states, encoder_hidden_states, temb, image_rotary_emb, attention_mask, **kwargs
+                )
+
+        # 4. Output norm & projection
+        hidden_states = self.norm_out(hidden_states, temb)
+        hidden_states = self.proj_out(hidden_states)
+
+        # 5. Unpatchify
+        hidden_states = hidden_states.reshape(batch_size, post_patch_height, post_patch_width, -1, p, p)
+        output = hidden_states.permute(0, 3, 1, 4, 2, 5).flatten(4, 5).flatten(2, 3)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+        return Transformer2DModelOutput(sample=output)