Initial commit

comfy/ldm/hydit/controlnet.py

+
+import torch
+import torch.nn as nn
+
+
+from comfy.ldm.modules.diffusionmodules.mmdit import (
+    TimestepEmbedder,
+    PatchEmbed,
+)
+from .poolers import AttentionPool
+
+import comfy.latent_formats
+from .models import HunYuanDiTBlock, calc_rope
+
+
+
+class HunYuanControlNet(nn.Module):
+    """
+    HunYuanDiT: Diffusion model with a Transformer backbone.
+
+    Inherit ModelMixin and ConfigMixin to be compatible with the sampler StableDiffusionPipeline of diffusers.
+
+    Inherit PeftAdapterMixin to be compatible with the PEFT training pipeline.
+
+    Parameters
+    ----------
+    args: argparse.Namespace
+        The arguments parsed by argparse.
+    input_size: tuple
+        The size of the input image.
+    patch_size: int
+        The size of the patch.
+    in_channels: int
+        The number of input channels.
+    hidden_size: int
+        The hidden size of the transformer backbone.
+    depth: int
+        The number of transformer blocks.
+    num_heads: int
+        The number of attention heads.
+    mlp_ratio: float
+        The ratio of the hidden size of the MLP in the transformer block.
+    log_fn: callable
+        The logging function.
+    """
+
+    def __init__(
+        self,
+        input_size: tuple = 128,
+        patch_size: int = 2,
+        in_channels: int = 4,
+        hidden_size: int = 1408,
+        depth: int = 40,
+        num_heads: int = 16,
+        mlp_ratio: float = 4.3637,
+        text_states_dim=1024,
+        text_states_dim_t5=2048,
+        text_len=77,
+        text_len_t5=256,
+        qk_norm=True,  # See http://arxiv.org/abs/2302.05442 for details.
+        size_cond=False,
+        use_style_cond=False,
+        learn_sigma=True,
+        norm="layer",
+        log_fn: callable = print,
+        attn_precision=None,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs,
+    ):
+        super().__init__()
+        self.log_fn = log_fn
+        self.depth = depth
+        self.learn_sigma = learn_sigma
+        self.in_channels = in_channels
+        self.out_channels = in_channels * 2 if learn_sigma else in_channels
+        self.patch_size = patch_size
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.text_states_dim = text_states_dim
+        self.text_states_dim_t5 = text_states_dim_t5
+        self.text_len = text_len
+        self.text_len_t5 = text_len_t5
+        self.size_cond = size_cond
+        self.use_style_cond = use_style_cond
+        self.norm = norm
+        self.dtype = dtype
+        self.latent_format = comfy.latent_formats.SDXL
+
+        self.mlp_t5 = nn.Sequential(
+            nn.Linear(
+                self.text_states_dim_t5,
+                self.text_states_dim_t5 * 4,
+                bias=True,
+                dtype=dtype,
+                device=device,
+            ),
+            nn.SiLU(),
+            nn.Linear(
+                self.text_states_dim_t5 * 4,
+                self.text_states_dim,
+                bias=True,
+                dtype=dtype,
+                device=device,
+            ),
+        )
+        # learnable replace
+        self.text_embedding_padding = nn.Parameter(
+            torch.randn(
+                self.text_len + self.text_len_t5,
+                self.text_states_dim,
+                dtype=dtype,
+                device=device,
+            )
+        )
+
+        # Attention pooling
+        pooler_out_dim = 1024
+        self.pooler = AttentionPool(
+            self.text_len_t5,
+            self.text_states_dim_t5,
+            num_heads=8,
+            output_dim=pooler_out_dim,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        # Dimension of the extra input vectors
+        self.extra_in_dim = pooler_out_dim
+
+        if self.size_cond:
+            # Image size and crop size conditions
+            self.extra_in_dim += 6 * 256
+
+        if self.use_style_cond:
+            # Here we use a default learned embedder layer for future extension.
+            self.style_embedder = nn.Embedding(
+                1, hidden_size, dtype=dtype, device=device
+            )
+            self.extra_in_dim += hidden_size
+
+        # Text embedding for `add`
+        self.x_embedder = PatchEmbed(
+            input_size,
+            patch_size,
+            in_channels,
+            hidden_size,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.t_embedder = TimestepEmbedder(
+            hidden_size, dtype=dtype, device=device, operations=operations
+        )
+        self.extra_embedder = nn.Sequential(
+            operations.Linear(
+                self.extra_in_dim, hidden_size * 4, dtype=dtype, device=device
+            ),
+            nn.SiLU(),
+            operations.Linear(
+                hidden_size * 4, hidden_size, bias=True, dtype=dtype, device=device
+            ),
+        )
+
+        # HUnYuanDiT Blocks
+        self.blocks = nn.ModuleList(
+            [
+                HunYuanDiTBlock(
+                    hidden_size=hidden_size,
+                    c_emb_size=hidden_size,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    text_states_dim=self.text_states_dim,
+                    qk_norm=qk_norm,
+                    norm_type=self.norm,
+                    skip=False,
+                    attn_precision=attn_precision,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations,
+                )
+                for _ in range(19)
+            ]
+        )
+
+        # Input zero linear for the first block
+        self.before_proj = operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
+
+
+        # Output zero linear for the every block
+        self.after_proj_list = nn.ModuleList(
+            [
+
+                    operations.Linear(
+                        self.hidden_size, self.hidden_size, dtype=dtype, device=device
+                    )
+                for _ in range(len(self.blocks))
+            ]
+        )
+
+    def forward(
+        self,
+        x,
+        hint,
+        timesteps,
+        context,#encoder_hidden_states=None,
+        text_embedding_mask=None,
+        encoder_hidden_states_t5=None,
+        text_embedding_mask_t5=None,
+        image_meta_size=None,
+        style=None,
+        return_dict=False,
+        **kwarg,
+    ):
+        """
+        Forward pass of the encoder.
+
+        Parameters
+        ----------
+        x: torch.Tensor
+            (B, D, H, W)
+        t: torch.Tensor
+            (B)
+        encoder_hidden_states: torch.Tensor
+            CLIP text embedding, (B, L_clip, D)
+        text_embedding_mask: torch.Tensor
+            CLIP text embedding mask, (B, L_clip)
+        encoder_hidden_states_t5: torch.Tensor
+            T5 text embedding, (B, L_t5, D)
+        text_embedding_mask_t5: torch.Tensor
+            T5 text embedding mask, (B, L_t5)
+        image_meta_size: torch.Tensor
+            (B, 6)
+        style: torch.Tensor
+            (B)
+        cos_cis_img: torch.Tensor
+        sin_cis_img: torch.Tensor
+        return_dict: bool
+            Whether to return a dictionary.
+        """
+        condition = hint
+        if condition.shape[0] == 1:
+            condition = torch.repeat_interleave(condition, x.shape[0], dim=0)
+
+        text_states = context  # 2,77,1024
+        text_states_t5 = encoder_hidden_states_t5  # 2,256,2048
+        text_states_mask = text_embedding_mask.bool()  # 2,77
+        text_states_t5_mask = text_embedding_mask_t5.bool()  # 2,256
+        b_t5, l_t5, c_t5 = text_states_t5.shape
+        text_states_t5 = self.mlp_t5(text_states_t5.view(-1, c_t5)).view(b_t5, l_t5, -1)
+
+        padding = comfy.ops.cast_to_input(self.text_embedding_padding, text_states)
+
+        text_states[:, -self.text_len :] = torch.where(
+            text_states_mask[:, -self.text_len :].unsqueeze(2),
+            text_states[:, -self.text_len :],
+            padding[: self.text_len],
+        )
+        text_states_t5[:, -self.text_len_t5 :] = torch.where(
+            text_states_t5_mask[:, -self.text_len_t5 :].unsqueeze(2),
+            text_states_t5[:, -self.text_len_t5 :],
+            padding[self.text_len :],
+        )
+
+        text_states = torch.cat([text_states, text_states_t5], dim=1)  # 2,205，1024
+
+        # _, _, oh, ow = x.shape
+        # th, tw = oh // self.patch_size, ow // self.patch_size
+
+        # Get image RoPE embedding according to `reso`lution.
+        freqs_cis_img = calc_rope(
+            x, self.patch_size, self.hidden_size // self.num_heads
+        )  # (cos_cis_img, sin_cis_img)
+
+        # ========================= Build time and image embedding =========================
+        t = self.t_embedder(timesteps, dtype=self.dtype)
+        x = self.x_embedder(x)
+
+        # ========================= Concatenate all extra vectors =========================
+        # Build text tokens with pooling
+        extra_vec = self.pooler(encoder_hidden_states_t5)
+
+        # Build image meta size tokens if applicable
+        # if image_meta_size is not None:
+        #     image_meta_size = timestep_embedding(image_meta_size.view(-1), 256)   # [B * 6, 256]
+        #     if image_meta_size.dtype != self.dtype:
+        #         image_meta_size = image_meta_size.half()
+        #     image_meta_size = image_meta_size.view(-1, 6 * 256)
+        #     extra_vec = torch.cat([extra_vec, image_meta_size], dim=1)  # [B, D + 6 * 256]
+
+        # Build style tokens
+        if style is not None:
+            style_embedding = self.style_embedder(style)
+            extra_vec = torch.cat([extra_vec, style_embedding], dim=1)
+
+        # Concatenate all extra vectors
+        c = t + self.extra_embedder(extra_vec)  # [B, D]
+
+        # ========================= Deal with Condition =========================
+        condition = self.x_embedder(condition)
+
+        # ========================= Forward pass through HunYuanDiT blocks =========================
+        controls = []
+        x = x + self.before_proj(condition)  # add condition
+        for layer, block in enumerate(self.blocks):
+            x = block(x, c, text_states, freqs_cis_img)
+            controls.append(self.after_proj_list[layer](x))  # zero linear for output
+
+        return {"output": controls}

comfy/ldm/hydit/models.py

+
+import torch
+import torch.nn as nn
+
+import comfy.ops
+from comfy.ldm.modules.diffusionmodules.mmdit import Mlp, TimestepEmbedder, PatchEmbed
+from comfy.ldm.modules.diffusionmodules.util import timestep_embedding
+from torch.utils import checkpoint
+
+from .attn_layers import Attention, CrossAttention
+from .poolers import AttentionPool
+from .posemb_layers import get_2d_rotary_pos_embed, get_fill_resize_and_crop
+
+def calc_rope(x, patch_size, head_size):
+    th = (x.shape[2] + (patch_size // 2)) // patch_size
+    tw = (x.shape[3] + (patch_size // 2)) // patch_size
+    base_size = 512 // 8 // patch_size
+    start, stop = get_fill_resize_and_crop((th, tw), base_size)
+    sub_args = [start, stop, (th, tw)]
+    # head_size = HUNYUAN_DIT_CONFIG['DiT-g/2']['hidden_size'] // HUNYUAN_DIT_CONFIG['DiT-g/2']['num_heads']
+    rope = get_2d_rotary_pos_embed(head_size, *sub_args)
+    rope = (rope[0].to(x), rope[1].to(x))
+    return rope
+
+
+def modulate(x, shift, scale):
+    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+
+class HunYuanDiTBlock(nn.Module):
+    """
+    A HunYuanDiT block with `add` conditioning.
+    """
+    def __init__(self,
+                 hidden_size,
+                 c_emb_size,
+                 num_heads,
+                 mlp_ratio=4.0,
+                 text_states_dim=1024,
+                 qk_norm=False,
+                 norm_type="layer",
+                 skip=False,
+                 attn_precision=None,
+                 dtype=None,
+                 device=None,
+                 operations=None,
+                 ):
+        super().__init__()
+        use_ele_affine = True
+
+        if norm_type == "layer":
+            norm_layer = operations.LayerNorm
+        elif norm_type == "rms":
+            norm_layer = operations.RMSNorm
+        else:
+            raise ValueError(f"Unknown norm_type: {norm_type}")
+
+        # ========================= Self-Attention =========================
+        self.norm1 = norm_layer(hidden_size, elementwise_affine=use_ele_affine, eps=1e-6, dtype=dtype, device=device)
+        self.attn1 = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, qk_norm=qk_norm, attn_precision=attn_precision, dtype=dtype, device=device, operations=operations)
+
+        # ========================= FFN =========================
+        self.norm2 = norm_layer(hidden_size, elementwise_affine=use_ele_affine, eps=1e-6, dtype=dtype, device=device)
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        approx_gelu = lambda: nn.GELU(approximate="tanh")
+        self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0, dtype=dtype, device=device, operations=operations)
+
+        # ========================= Add =========================
+        # Simply use add like SDXL.
+        self.default_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(c_emb_size, hidden_size, bias=True, dtype=dtype, device=device)
+        )
+
+        # ========================= Cross-Attention =========================
+        self.attn2 = CrossAttention(hidden_size, text_states_dim, num_heads=num_heads, qkv_bias=True,
+                                        qk_norm=qk_norm, attn_precision=attn_precision, dtype=dtype, device=device, operations=operations)
+        self.norm3 = norm_layer(hidden_size, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device)
+
+        # ========================= Skip Connection =========================
+        if skip:
+            self.skip_norm = norm_layer(2 * hidden_size, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device)
+            self.skip_linear = operations.Linear(2 * hidden_size, hidden_size, dtype=dtype, device=device)
+        else:
+            self.skip_linear = None
+
+        self.gradient_checkpointing = False
+
+    def _forward(self, x, c=None, text_states=None, freq_cis_img=None, skip=None):
+        # Long Skip Connection
+        if self.skip_linear is not None:
+            cat = torch.cat([x, skip], dim=-1)
+            if cat.dtype != x.dtype:
+                cat = cat.to(x.dtype)
+            cat = self.skip_norm(cat)
+            x = self.skip_linear(cat)
+
+        # Self-Attention
+        shift_msa = self.default_modulation(c).unsqueeze(dim=1)
+        attn_inputs = (
+            self.norm1(x) + shift_msa, freq_cis_img,
+        )
+        x = x + self.attn1(*attn_inputs)[0]
+
+        # Cross-Attention
+        cross_inputs = (
+            self.norm3(x), text_states, freq_cis_img
+        )
+        x = x + self.attn2(*cross_inputs)[0]
+
+        # FFN Layer
+        mlp_inputs = self.norm2(x)
+        x = x + self.mlp(mlp_inputs)
+
+        return x
+
+    def forward(self, x, c=None, text_states=None, freq_cis_img=None, skip=None):
+        if self.gradient_checkpointing and self.training:
+            return checkpoint.checkpoint(self._forward, x, c, text_states, freq_cis_img, skip)
+        return self._forward(x, c, text_states, freq_cis_img, skip)
+
+
+class FinalLayer(nn.Module):
+    """
+    The final layer of HunYuanDiT.
+    """
+    def __init__(self, final_hidden_size, c_emb_size, patch_size, out_channels, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(final_hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(final_hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(c_emb_size, 2 * final_hidden_size, bias=True, dtype=dtype, device=device)
+        )
+
+    def forward(self, x, c):
+        shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
+        x = modulate(self.norm_final(x), shift, scale)
+        x = self.linear(x)
+        return x
+
+
+class HunYuanDiT(nn.Module):
+    """
+    HunYuanDiT: Diffusion model with a Transformer backbone.
+
+    Inherit ModelMixin and ConfigMixin to be compatible with the sampler StableDiffusionPipeline of diffusers.
+
+    Inherit PeftAdapterMixin to be compatible with the PEFT training pipeline.
+
+    Parameters
+    ----------
+    args: argparse.Namespace
+        The arguments parsed by argparse.
+    input_size: tuple
+        The size of the input image.
+    patch_size: int
+        The size of the patch.
+    in_channels: int
+        The number of input channels.
+    hidden_size: int
+        The hidden size of the transformer backbone.
+    depth: int
+        The number of transformer blocks.
+    num_heads: int
+        The number of attention heads.
+    mlp_ratio: float
+        The ratio of the hidden size of the MLP in the transformer block.
+    log_fn: callable
+        The logging function.
+    """
+    #@register_to_config
+    def __init__(self,
+                 input_size: tuple = 32,
+                 patch_size: int = 2,
+                 in_channels: int = 4,
+                 hidden_size: int = 1152,
+                 depth: int = 28,
+                 num_heads: int = 16,
+                 mlp_ratio: float = 4.0,
+                 text_states_dim = 1024,
+                 text_states_dim_t5 = 2048,
+                 text_len = 77,
+                 text_len_t5 = 256,
+                 qk_norm = True,# See http://arxiv.org/abs/2302.05442 for details.
+                 size_cond = False,
+                 use_style_cond = False,
+                 learn_sigma = True,
+                 norm = "layer",
+                 log_fn: callable = print,
+                 attn_precision=None,
+                 dtype=None,
+                 device=None,
+                 operations=None,
+                 **kwargs,
+    ):
+        super().__init__()
+        self.log_fn = log_fn
+        self.depth = depth
+        self.learn_sigma = learn_sigma
+        self.in_channels = in_channels
+        self.out_channels = in_channels * 2 if learn_sigma else in_channels
+        self.patch_size = patch_size
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.text_states_dim = text_states_dim
+        self.text_states_dim_t5 = text_states_dim_t5
+        self.text_len = text_len
+        self.text_len_t5 = text_len_t5
+        self.size_cond = size_cond
+        self.use_style_cond = use_style_cond
+        self.norm = norm
+        self.dtype = dtype
+        #import pdb
+        #pdb.set_trace()
+
+        self.mlp_t5 = nn.Sequential(
+            operations.Linear(self.text_states_dim_t5, self.text_states_dim_t5 * 4, bias=True, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(self.text_states_dim_t5 * 4, self.text_states_dim, bias=True, dtype=dtype, device=device),
+        )
+        # learnable replace
+        self.text_embedding_padding = nn.Parameter(
+            torch.empty(self.text_len + self.text_len_t5, self.text_states_dim, dtype=dtype, device=device))
+
+        # Attention pooling
+        pooler_out_dim = 1024
+        self.pooler = AttentionPool(self.text_len_t5, self.text_states_dim_t5, num_heads=8, output_dim=pooler_out_dim, dtype=dtype, device=device, operations=operations)
+
+        # Dimension of the extra input vectors
+        self.extra_in_dim = pooler_out_dim
+
+        if self.size_cond:
+            # Image size and crop size conditions
+            self.extra_in_dim += 6 * 256
+
+        if self.use_style_cond:
+            # Here we use a default learned embedder layer for future extension.
+            self.style_embedder = operations.Embedding(1, hidden_size, dtype=dtype, device=device)
+            self.extra_in_dim += hidden_size
+
+        # Text embedding for `add`
+        self.x_embedder = PatchEmbed(input_size, patch_size, in_channels, hidden_size, dtype=dtype, device=device, operations=operations)
+        self.t_embedder = TimestepEmbedder(hidden_size, dtype=dtype, device=device, operations=operations)
+        self.extra_embedder = nn.Sequential(
+            operations.Linear(self.extra_in_dim, hidden_size * 4, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(hidden_size * 4, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+        # HUnYuanDiT Blocks
+        self.blocks = nn.ModuleList([
+            HunYuanDiTBlock(hidden_size=hidden_size,
+                            c_emb_size=hidden_size,
+                            num_heads=num_heads,
+                            mlp_ratio=mlp_ratio,
+                            text_states_dim=self.text_states_dim,
+                            qk_norm=qk_norm,
+                            norm_type=self.norm,
+                            skip=layer > depth // 2,
+                            attn_precision=attn_precision,
+                            dtype=dtype,
+                            device=device,
+                            operations=operations,
+                            )
+            for layer in range(depth)
+        ])
+
+        self.final_layer = FinalLayer(hidden_size, hidden_size, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations)
+        self.unpatchify_channels = self.out_channels
+
+
+
+    def forward(self,
+                x,
+                t,
+                context,#encoder_hidden_states=None,
+                text_embedding_mask=None,
+                encoder_hidden_states_t5=None,
+                text_embedding_mask_t5=None,
+                image_meta_size=None,
+                style=None,
+                return_dict=False,
+                control=None,
+                transformer_options={},
+                ):
+        """
+        Forward pass of the encoder.
+
+        Parameters
+        ----------
+        x: torch.Tensor
+            (B, D, H, W)
+        t: torch.Tensor
+            (B)
+        encoder_hidden_states: torch.Tensor
+            CLIP text embedding, (B, L_clip, D)
+        text_embedding_mask: torch.Tensor
+            CLIP text embedding mask, (B, L_clip)
+        encoder_hidden_states_t5: torch.Tensor
+            T5 text embedding, (B, L_t5, D)
+        text_embedding_mask_t5: torch.Tensor
+            T5 text embedding mask, (B, L_t5)
+        image_meta_size: torch.Tensor
+            (B, 6)
+        style: torch.Tensor
+            (B)
+        cos_cis_img: torch.Tensor
+        sin_cis_img: torch.Tensor
+        return_dict: bool
+            Whether to return a dictionary.
+        """
+        patches_replace = transformer_options.get("patches_replace", {})
+        encoder_hidden_states = context
+        text_states = encoder_hidden_states                     # 2,77,1024
+        text_states_t5 = encoder_hidden_states_t5               # 2,256,2048
+        text_states_mask = text_embedding_mask.bool()           # 2,77
+        text_states_t5_mask = text_embedding_mask_t5.bool()     # 2,256
+        b_t5, l_t5, c_t5 = text_states_t5.shape
+        text_states_t5 = self.mlp_t5(text_states_t5.view(-1, c_t5)).view(b_t5, l_t5, -1)
+
+        padding = comfy.ops.cast_to_input(self.text_embedding_padding, text_states)
+
+        text_states[:,-self.text_len:] = torch.where(text_states_mask[:,-self.text_len:].unsqueeze(2), text_states[:,-self.text_len:], padding[:self.text_len])
+        text_states_t5[:,-self.text_len_t5:] = torch.where(text_states_t5_mask[:,-self.text_len_t5:].unsqueeze(2), text_states_t5[:,-self.text_len_t5:], padding[self.text_len:])
+
+        text_states = torch.cat([text_states, text_states_t5], dim=1)  # 2,205，1024
+        # clip_t5_mask = torch.cat([text_states_mask, text_states_t5_mask], dim=-1)
+
+        _, _, oh, ow = x.shape
+        th, tw = (oh + (self.patch_size // 2)) // self.patch_size, (ow + (self.patch_size // 2)) // self.patch_size
+
+
+        # Get image RoPE embedding according to `reso`lution.
+        freqs_cis_img = calc_rope(x, self.patch_size, self.hidden_size // self.num_heads) #(cos_cis_img, sin_cis_img)
+
+        # ========================= Build time and image embedding =========================
+        t = self.t_embedder(t, dtype=x.dtype)
+        x = self.x_embedder(x)
+
+        # ========================= Concatenate all extra vectors =========================
+        # Build text tokens with pooling
+        extra_vec = self.pooler(encoder_hidden_states_t5)
+
+        # Build image meta size tokens if applicable
+        if self.size_cond:
+            image_meta_size = timestep_embedding(image_meta_size.view(-1), 256).to(x.dtype)   # [B * 6, 256]
+            image_meta_size = image_meta_size.view(-1, 6 * 256)
+            extra_vec = torch.cat([extra_vec, image_meta_size], dim=1)  # [B, D + 6 * 256]
+
+        # Build style tokens
+        if self.use_style_cond:
+            if style is None:
+                style = torch.zeros((extra_vec.shape[0],), device=x.device, dtype=torch.int)
+            style_embedding = self.style_embedder(style, out_dtype=x.dtype)
+            extra_vec = torch.cat([extra_vec, style_embedding], dim=1)
+
+        # Concatenate all extra vectors
+        c = t + self.extra_embedder(extra_vec)  # [B, D]
+
+        blocks_replace = patches_replace.get("dit", {})
+
+        controls = None
+        if control:
+            controls = control.get("output", None)
+        # ========================= Forward pass through HunYuanDiT blocks =========================
+        skips = []
+        for layer, block in enumerate(self.blocks):
+            if layer > self.depth // 2:
+                if controls is not None:
+                    skip = skips.pop() + controls.pop().to(dtype=x.dtype)
+                else:
+                    skip = skips.pop()
+            else:
+                skip = None
+
+            if ("double_block", layer) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], args["vec"], args["txt"], args["pe"], args["skip"])
+                    return out
+
+                out = blocks_replace[("double_block", layer)]({"img": x, "txt": text_states, "vec": c, "pe": freqs_cis_img, "skip": skip}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, c, text_states, freqs_cis_img, skip)   # (N, L, D)
+
+
+            if layer < (self.depth // 2 - 1):
+                skips.append(x)
+        if controls is not None and len(controls) != 0:
+            raise ValueError("The number of controls is not equal to the number of skip connections.")
+
+        # ========================= Final layer =========================
+        x = self.final_layer(x, c)                              # (N, L, patch_size ** 2 * out_channels)
+        x = self.unpatchify(x, th, tw)                          # (N, out_channels, H, W)
+
+        if return_dict:
+            return {'x': x}
+        if self.learn_sigma:
+            return x[:,:self.out_channels // 2,:oh,:ow]
+        return x[:,:,:oh,:ow]
+
+    def unpatchify(self, x, h, w):
+        """
+        x: (N, T, patch_size**2 * C)
+        imgs: (N, H, W, C)
+        """
+        c = self.unpatchify_channels
+        p = self.x_embedder.patch_size[0]
+        # h = w = int(x.shape[1] ** 0.5)
+        assert h * w == x.shape[1]
+
+        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
+        x = torch.einsum('nhwpqc->nchpwq', x)
+        imgs = x.reshape(shape=(x.shape[0], c, h * p, w * p))
+        return imgs

comfy/ldm/hydit/poolers.py

+import torch
+import torch.nn as nn
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.ops
+
+class AttentionPool(nn.Module):
+    def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(torch.empty(spacial_dim + 1, embed_dim, dtype=dtype, device=device))
+        self.k_proj = operations.Linear(embed_dim, embed_dim, dtype=dtype, device=device)
+        self.q_proj = operations.Linear(embed_dim, embed_dim, dtype=dtype, device=device)
+        self.v_proj = operations.Linear(embed_dim, embed_dim, dtype=dtype, device=device)
+        self.c_proj = operations.Linear(embed_dim, output_dim or embed_dim, dtype=dtype, device=device)
+        self.num_heads = num_heads
+        self.embed_dim = embed_dim
+
+    def forward(self, x):
+        x = x[:,:self.positional_embedding.shape[0] - 1]
+        x = x.permute(1, 0, 2)  # NLC -> LNC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (L+1)NC
+        x = x + comfy.ops.cast_to_input(self.positional_embedding[:, None, :], x) # (L+1)NC
+
+        q = self.q_proj(x[:1])
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+
+        batch_size = q.shape[1]
+        head_dim = self.embed_dim // self.num_heads
+        q = q.view(1, batch_size * self.num_heads, head_dim).transpose(0, 1).view(batch_size, self.num_heads, -1, head_dim)
+        k = k.view(k.shape[0], batch_size * self.num_heads, head_dim).transpose(0, 1).view(batch_size, self.num_heads, -1, head_dim)
+        v = v.view(v.shape[0], batch_size * self.num_heads, head_dim).transpose(0, 1).view(batch_size, self.num_heads, -1, head_dim)
+
+        attn_output = optimized_attention(q, k, v, self.num_heads, skip_reshape=True).transpose(0, 1)
+
+        attn_output = self.c_proj(attn_output)
+        return attn_output.squeeze(0)

comfy/ldm/hydit/posemb_layers.py

+import torch
+import numpy as np
+from typing import Union
+
+
+def _to_tuple(x):
+    if isinstance(x, int):
+        return x, x
+    else:
+        return x
+
+
+def get_fill_resize_and_crop(src, tgt):
+    th, tw = _to_tuple(tgt)
+    h, w = _to_tuple(src)
+
+    tr = th / tw        # base resolution
+    r = h / w           # target resolution
+
+    # resize
+    if r > tr:
+        resize_height = th
+        resize_width = int(round(th / h * w))
+    else:
+        resize_width = tw
+        resize_height = int(round(tw / w * h))    # resize the target resolution down based on the base resolution
+
+    crop_top = int(round((th - resize_height) / 2.0))
+    crop_left = int(round((tw - resize_width) / 2.0))
+
+    return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)
+
+
+def get_meshgrid(start, *args):
+    if len(args) == 0:
+        # start is grid_size
+        num = _to_tuple(start)
+        start = (0, 0)
+        stop = num
+    elif len(args) == 1:
+        # start is start, args[0] is stop, step is 1
+        start = _to_tuple(start)
+        stop = _to_tuple(args[0])
+        num = (stop[0] - start[0], stop[1] - start[1])
+    elif len(args) == 2:
+        # start is start, args[0] is stop, args[1] is num
+        start = _to_tuple(start)
+        stop = _to_tuple(args[0])
+        num = _to_tuple(args[1])
+    else:
+        raise ValueError(f"len(args) should be 0, 1 or 2, but got {len(args)}")
+
+    grid_h = np.linspace(start[0], stop[0], num[0], endpoint=False, dtype=np.float32)
+    grid_w = np.linspace(start[1], stop[1], num[1], endpoint=False, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)   # [2, W, H]
+    return grid
+
+#################################################################################
+#                   Sine/Cosine Positional Embedding Functions                  #
+#################################################################################
+# https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py
+
+def get_2d_sincos_pos_embed(embed_dim, start, *args, cls_token=False, extra_tokens=0):
+    """
+    grid_size: int of the grid height and width
+    return:
+    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+    """
+    grid = get_meshgrid(start, *args)   # [2, H, w]
+    # grid_h = np.arange(grid_size, dtype=np.float32)
+    # grid_w = np.arange(grid_size, dtype=np.float32)
+    # grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    # grid = np.stack(grid, axis=0)   # [2, W, H]
+
+    grid = grid.reshape([2, 1, *grid.shape[1:]])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if cls_token and extra_tokens > 0:
+        pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
+    return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=1)    # (H*W, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (W,H)
+    out: (M, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float64)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out)   # (M, D/2)
+    emb_cos = np.cos(out)   # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+
+#################################################################################
+#                   Rotary Positional Embedding Functions                       #
+#################################################################################
+# https://github.com/facebookresearch/llama/blob/main/llama/model.py#L443
+
+def get_2d_rotary_pos_embed(embed_dim, start, *args, use_real=True):
+    """
+    This is a 2d version of precompute_freqs_cis, which is a RoPE for image tokens with 2d structure.
+
+    Parameters
+    ----------
+    embed_dim: int
+        embedding dimension size
+    start: int or tuple of int
+        If len(args) == 0, start is num; If len(args) == 1, start is start, args[0] is stop, step is 1;
+        If len(args) == 2, start is start, args[0] is stop, args[1] is num.
+    use_real: bool
+        If True, return real part and imaginary part separately. Otherwise, return complex numbers.
+
+    Returns
+    -------
+    pos_embed: torch.Tensor
+        [HW, D/2]
+    """
+    grid = get_meshgrid(start, *args)   # [2, H, w]
+    grid = grid.reshape([2, 1, *grid.shape[1:]])   # Returns a sampling matrix with the same resolution as the target resolution
+    pos_embed = get_2d_rotary_pos_embed_from_grid(embed_dim, grid, use_real=use_real)
+    return pos_embed
+
+
+def get_2d_rotary_pos_embed_from_grid(embed_dim, grid, use_real=False):
+    assert embed_dim % 4 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_rotary_pos_embed(embed_dim // 2, grid[0].reshape(-1), use_real=use_real)  # (H*W, D/4)
+    emb_w = get_1d_rotary_pos_embed(embed_dim // 2, grid[1].reshape(-1), use_real=use_real)  # (H*W, D/4)
+
+    if use_real:
+        cos = torch.cat([emb_h[0], emb_w[0]], dim=1)    # (H*W, D/2)
+        sin = torch.cat([emb_h[1], emb_w[1]], dim=1)    # (H*W, D/2)
+        return cos, sin
+    else:
+        emb = torch.cat([emb_h, emb_w], dim=1)    # (H*W, D/2)
+        return emb
+
+
+def get_1d_rotary_pos_embed(dim: int, pos: Union[np.ndarray, int], theta: float = 10000.0, use_real=False):
+    """
+    Precompute the frequency tensor for complex exponentials (cis) with given dimensions.
+
+    This function calculates a frequency tensor with complex exponentials using the given dimension 'dim'
+    and the end index 'end'. The 'theta' parameter scales the frequencies.
+    The returned tensor contains complex values in complex64 data type.
+
+    Args:
+        dim (int): Dimension of the frequency tensor.
+        pos (np.ndarray, int): Position indices for the frequency tensor. [S] or scalar
+        theta (float, optional): Scaling factor for frequency computation. Defaults to 10000.0.
+        use_real (bool, optional): If True, return real part and imaginary part separately.
+                                   Otherwise, return complex numbers.
+
+    Returns:
+        torch.Tensor: Precomputed frequency tensor with complex exponentials. [S, D/2]
+
+    """
+    if isinstance(pos, int):
+        pos = np.arange(pos)
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))  # [D/2]
+    t = torch.from_numpy(pos).to(freqs.device)  # type: ignore  # [S]
+    freqs = torch.outer(t, freqs).float()  # type: ignore   # [S, D/2]
+    if use_real:
+        freqs_cos = freqs.cos().repeat_interleave(2, dim=1)  # [S, D]
+        freqs_sin = freqs.sin().repeat_interleave(2, dim=1)  # [S, D]
+        return freqs_cos, freqs_sin
+    else:
+        freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64     # [S, D/2]
+        return freqs_cis
+
+
+
+def calc_sizes(rope_img, patch_size, th, tw):
+    if rope_img == 'extend':
+        # Expansion mode
+        sub_args = [(th, tw)]
+    elif rope_img.startswith('base'):
+        # Based on the specified dimensions, other dimensions are obtained through interpolation.
+        base_size = int(rope_img[4:]) // 8 // patch_size
+        start, stop = get_fill_resize_and_crop((th, tw), base_size)
+        sub_args = [start, stop, (th, tw)]
+    else:
+        raise ValueError(f"Unknown rope_img: {rope_img}")
+    return sub_args
+
+
+def init_image_posemb(rope_img,
+                      resolutions,
+                      patch_size,
+                      hidden_size,
+                      num_heads,
+                      log_fn,
+                      rope_real=True,
+                      ):
+    freqs_cis_img = {}
+    for reso in resolutions:
+        th, tw = reso.height // 8 // patch_size, reso.width // 8 // patch_size
+        sub_args = calc_sizes(rope_img, patch_size, th, tw)
+        freqs_cis_img[str(reso)] = get_2d_rotary_pos_embed(hidden_size // num_heads, *sub_args, use_real=rope_real)
+        log_fn(f"    Using image RoPE ({rope_img}) ({'real' if rope_real else 'complex'}): {sub_args} | ({reso}) "
+               f"{freqs_cis_img[str(reso)][0].shape if rope_real else freqs_cis_img[str(reso)].shape}")
+    return freqs_cis_img

comfy/ldm/lightricks/model.py

+import torch
+from torch import nn
+import comfy.patcher_extension
+import comfy.ldm.modules.attention
+import comfy.ldm.common_dit
+from einops import rearrange
+import math
+from typing import Dict, Optional, Tuple
+
+from .symmetric_patchifier import SymmetricPatchifier, latent_to_pixel_coords
+
+
+def get_timestep_embedding(
+    timesteps: torch.Tensor,
+    embedding_dim: int,
+    flip_sin_to_cos: bool = False,
+    downscale_freq_shift: float = 1,
+    scale: float = 1,
+    max_period: int = 10000,
+):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings.
+
+    Args
+        timesteps (torch.Tensor):
+            a 1-D Tensor of N indices, one per batch element. These may be fractional.
+        embedding_dim (int):
+            the dimension of the output.
+        flip_sin_to_cos (bool):
+            Whether the embedding order should be `cos, sin` (if True) or `sin, cos` (if False)
+        downscale_freq_shift (float):
+            Controls the delta between frequencies between dimensions
+        scale (float):
+            Scaling factor applied to the embeddings.
+        max_period (int):
+            Controls the maximum frequency of the embeddings
+    Returns
+        torch.Tensor: an [N x dim] Tensor of positional embeddings.
+    """
+    assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array"
+
+    half_dim = embedding_dim // 2
+    exponent = -math.log(max_period) * torch.arange(
+        start=0, end=half_dim, dtype=torch.float32, device=timesteps.device
+    )
+    exponent = exponent / (half_dim - downscale_freq_shift)
+
+    emb = torch.exp(exponent)
+    emb = timesteps[:, None].float() * emb[None, :]
+
+    # scale embeddings
+    emb = scale * emb
+
+    # concat sine and cosine embeddings
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
+
+    # flip sine and cosine embeddings
+    if flip_sin_to_cos:
+        emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1)
+
+    # zero pad
+    if embedding_dim % 2 == 1:
+        emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
+    return emb
+
+
+class TimestepEmbedding(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        time_embed_dim: int,
+        act_fn: str = "silu",
+        out_dim: int = None,
+        post_act_fn: Optional[str] = None,
+        cond_proj_dim=None,
+        sample_proj_bias=True,
+        dtype=None, device=None, operations=None,
+    ):
+        super().__init__()
+
+        self.linear_1 = operations.Linear(in_channels, time_embed_dim, sample_proj_bias, dtype=dtype, device=device)
+
+        if cond_proj_dim is not None:
+            self.cond_proj = operations.Linear(cond_proj_dim, in_channels, bias=False, dtype=dtype, device=device)
+        else:
+            self.cond_proj = None
+
+        self.act = nn.SiLU()
+
+        if out_dim is not None:
+            time_embed_dim_out = out_dim
+        else:
+            time_embed_dim_out = time_embed_dim
+        self.linear_2 = operations.Linear(time_embed_dim, time_embed_dim_out, sample_proj_bias, dtype=dtype, device=device)
+
+        if post_act_fn is None:
+            self.post_act = None
+        # else:
+        #     self.post_act = get_activation(post_act_fn)
+
+    def forward(self, sample, condition=None):
+        if condition is not None:
+            sample = sample + self.cond_proj(condition)
+        sample = self.linear_1(sample)
+
+        if self.act is not None:
+            sample = self.act(sample)
+
+        sample = self.linear_2(sample)
+
+        if self.post_act is not None:
+            sample = self.post_act(sample)
+        return sample
+
+
+class Timesteps(nn.Module):
+    def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float, scale: int = 1):
+        super().__init__()
+        self.num_channels = num_channels
+        self.flip_sin_to_cos = flip_sin_to_cos
+        self.downscale_freq_shift = downscale_freq_shift
+        self.scale = scale
+
+    def forward(self, timesteps):
+        t_emb = get_timestep_embedding(
+            timesteps,
+            self.num_channels,
+            flip_sin_to_cos=self.flip_sin_to_cos,
+            downscale_freq_shift=self.downscale_freq_shift,
+            scale=self.scale,
+        )
+        return t_emb
+
+
+class PixArtAlphaCombinedTimestepSizeEmbeddings(nn.Module):
+    """
+    For PixArt-Alpha.
+
+    Reference:
+    https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L164C9-L168C29
+    """
+
+    def __init__(self, embedding_dim, size_emb_dim, use_additional_conditions: bool = False, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        self.outdim = size_emb_dim
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, timestep, resolution, aspect_ratio, batch_size, hidden_dtype):
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, D)
+        return timesteps_emb
+
+
+class AdaLayerNormSingle(nn.Module):
+    r"""
+    Norm layer adaptive layer norm single (adaLN-single).
+
+    As proposed in PixArt-Alpha (see: https://arxiv.org/abs/2310.00426; Section 2.3).
+
+    Parameters:
+        embedding_dim (`int`): The size of each embedding vector.
+        use_additional_conditions (`bool`): To use additional conditions for normalization or not.
+    """
+
+    def __init__(self, embedding_dim: int, use_additional_conditions: bool = False, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings(
+            embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions, dtype=dtype, device=device, operations=operations
+        )
+
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(embedding_dim, 6 * embedding_dim, bias=True, dtype=dtype, device=device)
+
+    def forward(
+        self,
+        timestep: torch.Tensor,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        batch_size: Optional[int] = None,
+        hidden_dtype: Optional[torch.dtype] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        # No modulation happening here.
+        added_cond_kwargs = added_cond_kwargs or {"resolution": None, "aspect_ratio": None}
+        embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype)
+        return self.linear(self.silu(embedded_timestep)), embedded_timestep
+
+class PixArtAlphaTextProjection(nn.Module):
+    """
+    Projects caption embeddings. Also handles dropout for classifier-free guidance.
+
+    Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
+    """
+
+    def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh", dtype=None, device=None, operations=None):
+        super().__init__()
+        if out_features is None:
+            out_features = hidden_size
+        self.linear_1 = operations.Linear(in_features=in_features, out_features=hidden_size, bias=True, dtype=dtype, device=device)
+        if act_fn == "gelu_tanh":
+            self.act_1 = nn.GELU(approximate="tanh")
+        elif act_fn == "silu":
+            self.act_1 = nn.SiLU()
+        else:
+            raise ValueError(f"Unknown activation function: {act_fn}")
+        self.linear_2 = operations.Linear(in_features=hidden_size, out_features=out_features, bias=True, dtype=dtype, device=device)
+
+    def forward(self, caption):
+        hidden_states = self.linear_1(caption)
+        hidden_states = self.act_1(hidden_states)
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+
+class GELU_approx(nn.Module):
+    def __init__(self, dim_in, dim_out, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.proj = operations.Linear(dim_in, dim_out, dtype=dtype, device=device)
+
+    def forward(self, x):
+        return torch.nn.functional.gelu(self.proj(x), approximate="tanh")
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=None):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        project_in = GELU_approx(dim, inner_dim, dtype=dtype, device=device, operations=operations)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            operations.Linear(inner_dim, dim_out, dtype=dtype, device=device)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+def apply_rotary_emb(input_tensor, freqs_cis): #TODO: remove duplicate funcs and pick the best/fastest one
+    cos_freqs = freqs_cis[0]
+    sin_freqs = freqs_cis[1]
+
+    t_dup = rearrange(input_tensor, "... (d r) -> ... d r", r=2)
+    t1, t2 = t_dup.unbind(dim=-1)
+    t_dup = torch.stack((-t2, t1), dim=-1)
+    input_tensor_rot = rearrange(t_dup, "... d r -> ... (d r)")
+
+    out = input_tensor * cos_freqs + input_tensor_rot * sin_freqs
+
+    return out
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., attn_precision=None, dtype=None, device=None, operations=None):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = query_dim if context_dim is None else context_dim
+        self.attn_precision = attn_precision
+
+        self.heads = heads
+        self.dim_head = dim_head
+
+        self.q_norm = operations.RMSNorm(inner_dim, eps=1e-5, dtype=dtype, device=device)
+        self.k_norm = operations.RMSNorm(inner_dim, eps=1e-5, dtype=dtype, device=device)
+
+        self.to_q = operations.Linear(query_dim, inner_dim, bias=True, dtype=dtype, device=device)
+        self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)
+        self.to_v = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)
+
+        self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout))
+
+    def forward(self, x, context=None, mask=None, pe=None):
+        q = self.to_q(x)
+        context = x if context is None else context
+        k = self.to_k(context)
+        v = self.to_v(context)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        if pe is not None:
+            q = apply_rotary_emb(q, pe)
+            k = apply_rotary_emb(k, pe)
+
+        if mask is None:
+            out = comfy.ldm.modules.attention.optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision)
+        else:
+            out = comfy.ldm.modules.attention.optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision)
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(self, dim, n_heads, d_head, context_dim=None, attn_precision=None, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        self.attn_precision = attn_precision
+        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, context_dim=None, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations)
+        self.ff = FeedForward(dim, dim_out=dim, glu=True, dtype=dtype, device=device, operations=operations)
+
+        self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations)
+
+        self.scale_shift_table = nn.Parameter(torch.empty(6, dim, device=device, dtype=dtype))
+
+    def forward(self, x, context=None, attention_mask=None, timestep=None, pe=None):
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + timestep.reshape(x.shape[0], timestep.shape[1], self.scale_shift_table.shape[0], -1)).unbind(dim=2)
+
+        x += self.attn1(comfy.ldm.common_dit.rms_norm(x) * (1 + scale_msa) + shift_msa, pe=pe) * gate_msa
+
+        x += self.attn2(x, context=context, mask=attention_mask)
+
+        y = comfy.ldm.common_dit.rms_norm(x) * (1 + scale_mlp) + shift_mlp
+        x += self.ff(y) * gate_mlp
+
+        return x
+
+def get_fractional_positions(indices_grid, max_pos):
+    fractional_positions = torch.stack(
+        [
+            indices_grid[:, i] / max_pos[i]
+            for i in range(3)
+        ],
+        dim=-1,
+    )
+    return fractional_positions
+
+
+def precompute_freqs_cis(indices_grid, dim, out_dtype, theta=10000.0, max_pos=[20, 2048, 2048]):
+    dtype = torch.float32 #self.dtype
+
+    fractional_positions = get_fractional_positions(indices_grid, max_pos)
+
+    start = 1
+    end = theta
+    device = fractional_positions.device
+
+    indices = theta ** (
+        torch.linspace(
+            math.log(start, theta),
+            math.log(end, theta),
+            dim // 6,
+            device=device,
+            dtype=dtype,
+        )
+    )
+    indices = indices.to(dtype=dtype)
+
+    indices = indices * math.pi / 2
+
+    freqs = (
+        (indices * (fractional_positions.unsqueeze(-1) * 2 - 1))
+        .transpose(-1, -2)
+        .flatten(2)
+    )
+
+    cos_freq = freqs.cos().repeat_interleave(2, dim=-1)
+    sin_freq = freqs.sin().repeat_interleave(2, dim=-1)
+    if dim % 6 != 0:
+        cos_padding = torch.ones_like(cos_freq[:, :, : dim % 6])
+        sin_padding = torch.zeros_like(cos_freq[:, :, : dim % 6])
+        cos_freq = torch.cat([cos_padding, cos_freq], dim=-1)
+        sin_freq = torch.cat([sin_padding, sin_freq], dim=-1)
+    return cos_freq.to(out_dtype), sin_freq.to(out_dtype)
+
+
+class LTXVModel(torch.nn.Module):
+    def __init__(self,
+                 in_channels=128,
+                 cross_attention_dim=2048,
+                 attention_head_dim=64,
+                 num_attention_heads=32,
+
+                 caption_channels=4096,
+                 num_layers=28,
+
+
+                 positional_embedding_theta=10000.0,
+                 positional_embedding_max_pos=[20, 2048, 2048],
+                 causal_temporal_positioning=False,
+                 vae_scale_factors=(8, 32, 32),
+                 dtype=None, device=None, operations=None, **kwargs):
+        super().__init__()
+        self.generator = None
+        self.vae_scale_factors = vae_scale_factors
+        self.dtype = dtype
+        self.out_channels = in_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+        self.causal_temporal_positioning = causal_temporal_positioning
+
+        self.patchify_proj = operations.Linear(in_channels, self.inner_dim, bias=True, dtype=dtype, device=device)
+
+        self.adaln_single = AdaLayerNormSingle(
+            self.inner_dim, use_additional_conditions=False, dtype=dtype, device=device, operations=operations
+        )
+
+        # self.adaln_single.linear = operations.Linear(self.inner_dim, 4 * self.inner_dim, bias=True, dtype=dtype, device=device)
+
+        self.caption_projection = PixArtAlphaTextProjection(
+            in_features=caption_channels, hidden_size=self.inner_dim, dtype=dtype, device=device, operations=operations
+        )
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    context_dim=cross_attention_dim,
+                    # attn_precision=attn_precision,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        self.scale_shift_table = nn.Parameter(torch.empty(2, self.inner_dim, dtype=dtype, device=device))
+        self.norm_out = operations.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.proj_out = operations.Linear(self.inner_dim, self.out_channels, dtype=dtype, device=device)
+
+        self.patchifier = SymmetricPatchifier(1)
+
+    def forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, **kwargs)
+
+    def _forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
+        patches_replace = transformer_options.get("patches_replace", {})
+
+        orig_shape = list(x.shape)
+
+        x, latent_coords = self.patchifier.patchify(x)
+        pixel_coords = latent_to_pixel_coords(
+            latent_coords=latent_coords,
+            scale_factors=self.vae_scale_factors,
+            causal_fix=self.causal_temporal_positioning,
+        )
+
+        if keyframe_idxs is not None:
+            pixel_coords[:, :, -keyframe_idxs.shape[2]:] = keyframe_idxs
+
+        fractional_coords = pixel_coords.to(torch.float32)
+        fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate)
+
+        x = self.patchify_proj(x)
+        timestep = timestep * 1000.0
+
+        if attention_mask is not None and not torch.is_floating_point(attention_mask):
+            attention_mask = (attention_mask - 1).to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])) * torch.finfo(x.dtype).max
+
+        pe = precompute_freqs_cis(fractional_coords, dim=self.inner_dim, out_dtype=x.dtype)
+
+        batch_size = x.shape[0]
+        timestep, embedded_timestep = self.adaln_single(
+            timestep.flatten(),
+            {"resolution": None, "aspect_ratio": None},
+            batch_size=batch_size,
+            hidden_dtype=x.dtype,
+        )
+        # Second dimension is 1 or number of tokens (if timestep_per_token)
+        timestep = timestep.view(batch_size, -1, timestep.shape[-1])
+        embedded_timestep = embedded_timestep.view(
+            batch_size, -1, embedded_timestep.shape[-1]
+        )
+
+        # 2. Blocks
+        if self.caption_projection is not None:
+            batch_size = x.shape[0]
+            context = self.caption_projection(context)
+            context = context.view(
+                batch_size, -1, x.shape[-1]
+            )
+
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.transformer_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], attention_mask=args["attention_mask"], timestep=args["vec"], pe=args["pe"])
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "attention_mask": attention_mask, "vec": timestep, "pe": pe}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(
+                    x,
+                    context=context,
+                    attention_mask=attention_mask,
+                    timestep=timestep,
+                    pe=pe
+                )
+
+        # 3. Output
+        scale_shift_values = (
+            self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + embedded_timestep[:, :, None]
+        )
+        shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1]
+        x = self.norm_out(x)
+        # Modulation
+        x = x * (1 + scale) + shift
+        x = self.proj_out(x)
+
+        x = self.patchifier.unpatchify(
+            latents=x,
+            output_height=orig_shape[3],
+            output_width=orig_shape[4],
+            output_num_frames=orig_shape[2],
+            out_channels=orig_shape[1] // math.prod(self.patchifier.patch_size),
+        )
+
+        return x

comfy/ldm/lightricks/symmetric_patchifier.py

+from abc import ABC, abstractmethod
+from typing import Tuple
+
+import torch
+from einops import rearrange
+from torch import Tensor
+
+
+def latent_to_pixel_coords(
+    latent_coords: Tensor, scale_factors: Tuple[int, int, int], causal_fix: bool = False
+) -> Tensor:
+    """
+    Converts latent coordinates to pixel coordinates by scaling them according to the VAE's
+    configuration.
+    Args:
+        latent_coords (Tensor): A tensor of shape [batch_size, 3, num_latents]
+        containing the latent corner coordinates of each token.
+        scale_factors (Tuple[int, int, int]): The scale factors of the VAE's latent space.
+        causal_fix (bool): Whether to take into account the different temporal scale
+            of the first frame. Default = False for backwards compatibility.
+    Returns:
+        Tensor: A tensor of pixel coordinates corresponding to the input latent coordinates.
+    """
+    pixel_coords = (
+        latent_coords
+        * torch.tensor(scale_factors, device=latent_coords.device)[None, :, None]
+    )
+    if causal_fix:
+        # Fix temporal scale for first frame to 1 due to causality
+        pixel_coords[:, 0] = (pixel_coords[:, 0] + 1 - scale_factors[0]).clamp(min=0)
+    return pixel_coords
+
+
+class Patchifier(ABC):
+    def __init__(self, patch_size: int):
+        super().__init__()
+        self._patch_size = (1, patch_size, patch_size)
+
+    @abstractmethod
+    def patchify(
+        self, latents: Tensor, frame_rates: Tensor, scale_grid: bool
+    ) -> Tuple[Tensor, Tensor]:
+        pass
+
+    @abstractmethod
+    def unpatchify(
+        self,
+        latents: Tensor,
+        output_height: int,
+        output_width: int,
+        output_num_frames: int,
+        out_channels: int,
+    ) -> Tuple[Tensor, Tensor]:
+        pass
+
+    @property
+    def patch_size(self):
+        return self._patch_size
+
+    def get_latent_coords(
+        self, latent_num_frames, latent_height, latent_width, batch_size, device
+    ):
+        """
+        Return a tensor of shape [batch_size, 3, num_patches] containing the
+            top-left corner latent coordinates of each latent patch.
+        The tensor is repeated for each batch element.
+        """
+        latent_sample_coords = torch.meshgrid(
+            torch.arange(0, latent_num_frames, self._patch_size[0], device=device),
+            torch.arange(0, latent_height, self._patch_size[1], device=device),
+            torch.arange(0, latent_width, self._patch_size[2], device=device),
+            indexing="ij",
+        )
+        latent_sample_coords = torch.stack(latent_sample_coords, dim=0)
+        latent_coords = latent_sample_coords.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1)
+        latent_coords = rearrange(
+            latent_coords, "b c f h w -> b c (f h w)", b=batch_size
+        )
+        return latent_coords
+
+
+class SymmetricPatchifier(Patchifier):
+    def patchify(
+        self,
+        latents: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        b, _, f, h, w = latents.shape
+        latent_coords = self.get_latent_coords(f, h, w, b, latents.device)
+        latents = rearrange(
+            latents,
+            "b c (f p1) (h p2) (w p3) -> b (f h w) (c p1 p2 p3)",
+            p1=self._patch_size[0],
+            p2=self._patch_size[1],
+            p3=self._patch_size[2],
+        )
+        return latents, latent_coords
+
+    def unpatchify(
+        self,
+        latents: Tensor,
+        output_height: int,
+        output_width: int,
+        output_num_frames: int,
+        out_channels: int,
+    ) -> Tuple[Tensor, Tensor]:
+        output_height = output_height // self._patch_size[1]
+        output_width = output_width // self._patch_size[2]
+        latents = rearrange(
+            latents,
+            "b (f h w) (c p q) -> b c f (h p) (w q) ",
+            f=output_num_frames,
+            h=output_height,
+            w=output_width,
+            p=self._patch_size[1],
+            q=self._patch_size[2],
+        )
+        return latents

comfy/ldm/lightricks/vae/causal_conv3d.py

+from typing import Tuple, Union
+
+import torch
+import torch.nn as nn
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+
+class CausalConv3d(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size: int = 3,
+        stride: Union[int, Tuple[int]] = 1,
+        dilation: int = 1,
+        groups: int = 1,
+        spatial_padding_mode: str = "zeros",
+        **kwargs,
+    ):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        kernel_size = (kernel_size, kernel_size, kernel_size)
+        self.time_kernel_size = kernel_size[0]
+
+        dilation = (dilation, 1, 1)
+
+        height_pad = kernel_size[1] // 2
+        width_pad = kernel_size[2] // 2
+        padding = (0, height_pad, width_pad)
+
+        self.conv = ops.Conv3d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride=stride,
+            dilation=dilation,
+            padding=padding,
+            padding_mode=spatial_padding_mode,
+            groups=groups,
+        )
+
+    def forward(self, x, causal: bool = True):
+        if causal:
+            first_frame_pad = x[:, :, :1, :, :].repeat(
+                (1, 1, self.time_kernel_size - 1, 1, 1)
+            )
+            x = torch.concatenate((first_frame_pad, x), dim=2)
+        else:
+            first_frame_pad = x[:, :, :1, :, :].repeat(
+                (1, 1, (self.time_kernel_size - 1) // 2, 1, 1)
+            )
+            last_frame_pad = x[:, :, -1:, :, :].repeat(
+                (1, 1, (self.time_kernel_size - 1) // 2, 1, 1)
+            )
+            x = torch.concatenate((first_frame_pad, x, last_frame_pad), dim=2)
+        x = self.conv(x)
+        return x
+
+    @property
+    def weight(self):
+        return self.conv.weight

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

+from __future__ import annotations
+import torch
+from torch import nn
+from functools import partial
+import math
+from einops import rearrange
+from typing import List, Optional, Tuple, Union
+from .conv_nd_factory import make_conv_nd, make_linear_nd
+from .pixel_norm import PixelNorm
+from ..model import PixArtAlphaCombinedTimestepSizeEmbeddings
+import comfy.ops
+
+ops = comfy.ops.disable_weight_init
+
+class Encoder(nn.Module):
+    r"""
+    The `Encoder` layer of a variational autoencoder that encodes its input into a latent representation.
+
+    Args:
+        dims (`int` or `Tuple[int, int]`, *optional*, defaults to 3):
+            The number of dimensions to use in convolutions.
+        in_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        out_channels (`int`, *optional*, defaults to 3):
+            The number of output channels.
+        blocks (`List[Tuple[str, int]]`, *optional*, defaults to `[("res_x", 1)]`):
+            The blocks to use. Each block is a tuple of the block name and the number of layers.
+        base_channels (`int`, *optional*, defaults to 128):
+            The number of output channels for the first convolutional layer.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups for normalization.
+        patch_size (`int`, *optional*, defaults to 1):
+            The patch size to use. Should be a power of 2.
+        norm_layer (`str`, *optional*, defaults to `group_norm`):
+            The normalization layer to use. Can be either `group_norm` or `pixel_norm`.
+        latent_log_var (`str`, *optional*, defaults to `per_channel`):
+            The number of channels for the log variance. Can be either `per_channel`, `uniform`, `constant` or `none`.
+    """
+
+    def __init__(
+        self,
+        dims: Union[int, Tuple[int, int]] = 3,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        blocks: List[Tuple[str, int | dict]] = [("res_x", 1)],
+        base_channels: int = 128,
+        norm_num_groups: int = 32,
+        patch_size: Union[int, Tuple[int]] = 1,
+        norm_layer: str = "group_norm",  # group_norm, pixel_norm
+        latent_log_var: str = "per_channel",
+        spatial_padding_mode: str = "zeros",
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.norm_layer = norm_layer
+        self.latent_channels = out_channels
+        self.latent_log_var = latent_log_var
+        self.blocks_desc = blocks
+
+        in_channels = in_channels * patch_size**2
+        output_channel = base_channels
+
+        self.conv_in = make_conv_nd(
+            dims=dims,
+            in_channels=in_channels,
+            out_channels=output_channel,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            causal=True,
+            spatial_padding_mode=spatial_padding_mode,
+        )
+
+        self.down_blocks = nn.ModuleList([])
+
+        for block_name, block_params in blocks:
+            input_channel = output_channel
+            if isinstance(block_params, int):
+                block_params = {"num_layers": block_params}
+
+            if block_name == "res_x":
+                block = UNetMidBlock3D(
+                    dims=dims,
+                    in_channels=input_channel,
+                    num_layers=block_params["num_layers"],
+                    resnet_eps=1e-6,
+                    resnet_groups=norm_num_groups,
+                    norm_layer=norm_layer,
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "res_x_y":
+                output_channel = block_params.get("multiplier", 2) * output_channel
+                block = ResnetBlock3D(
+                    dims=dims,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    eps=1e-6,
+                    groups=norm_num_groups,
+                    norm_layer=norm_layer,
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_time":
+                block = make_conv_nd(
+                    dims=dims,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    kernel_size=3,
+                    stride=(2, 1, 1),
+                    causal=True,
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_space":
+                block = make_conv_nd(
+                    dims=dims,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    kernel_size=3,
+                    stride=(1, 2, 2),
+                    causal=True,
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_all":
+                block = make_conv_nd(
+                    dims=dims,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    kernel_size=3,
+                    stride=(2, 2, 2),
+                    causal=True,
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_all_x_y":
+                output_channel = block_params.get("multiplier", 2) * output_channel
+                block = make_conv_nd(
+                    dims=dims,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    kernel_size=3,
+                    stride=(2, 2, 2),
+                    causal=True,
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_all_res":
+                output_channel = block_params.get("multiplier", 2) * output_channel
+                block = SpaceToDepthDownsample(
+                    dims=dims,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    stride=(2, 2, 2),
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_space_res":
+                output_channel = block_params.get("multiplier", 2) * output_channel
+                block = SpaceToDepthDownsample(
+                    dims=dims,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    stride=(1, 2, 2),
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_time_res":
+                output_channel = block_params.get("multiplier", 2) * output_channel
+                block = SpaceToDepthDownsample(
+                    dims=dims,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    stride=(2, 1, 1),
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            else:
+                raise ValueError(f"unknown block: {block_name}")
+
+            self.down_blocks.append(block)
+
+        # out
+        if norm_layer == "group_norm":
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=output_channel, num_groups=norm_num_groups, eps=1e-6
+            )
+        elif norm_layer == "pixel_norm":
+            self.conv_norm_out = PixelNorm()
+        elif norm_layer == "layer_norm":
+            self.conv_norm_out = LayerNorm(output_channel, eps=1e-6)
+
+        self.conv_act = nn.SiLU()
+
+        conv_out_channels = out_channels
+        if latent_log_var == "per_channel":
+            conv_out_channels *= 2
+        elif latent_log_var == "uniform":
+            conv_out_channels += 1
+        elif latent_log_var == "constant":
+            conv_out_channels += 1
+        elif latent_log_var != "none":
+            raise ValueError(f"Invalid latent_log_var: {latent_log_var}")
+        self.conv_out = make_conv_nd(
+            dims,
+            output_channel,
+            conv_out_channels,
+            3,
+            padding=1,
+            causal=True,
+            spatial_padding_mode=spatial_padding_mode,
+        )
+
+        self.gradient_checkpointing = False
+
+    def forward(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+        r"""The forward method of the `Encoder` class."""
+
+        sample = patchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
+        sample = self.conv_in(sample)
+
+        checkpoint_fn = (
+            partial(torch.utils.checkpoint.checkpoint, use_reentrant=False)
+            if self.gradient_checkpointing and self.training
+            else lambda x: x
+        )
+
+        for down_block in self.down_blocks:
+            sample = checkpoint_fn(down_block)(sample)
+
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if self.latent_log_var == "uniform":
+            last_channel = sample[:, -1:, ...]
+            num_dims = sample.dim()
+
+            if num_dims == 4:
+                # For shape (B, C, H, W)
+                repeated_last_channel = last_channel.repeat(
+                    1, sample.shape[1] - 2, 1, 1
+                )
+                sample = torch.cat([sample, repeated_last_channel], dim=1)
+            elif num_dims == 5:
+                # For shape (B, C, F, H, W)
+                repeated_last_channel = last_channel.repeat(
+                    1, sample.shape[1] - 2, 1, 1, 1
+                )
+                sample = torch.cat([sample, repeated_last_channel], dim=1)
+            else:
+                raise ValueError(f"Invalid input shape: {sample.shape}")
+        elif self.latent_log_var == "constant":
+            sample = sample[:, :-1, ...]
+            approx_ln_0 = (
+                -30
+            )  # this is the minimal clamp value in DiagonalGaussianDistribution objects
+            sample = torch.cat(
+                [sample, torch.ones_like(sample, device=sample.device) * approx_ln_0],
+                dim=1,
+            )
+
+        return sample
+
+
+class Decoder(nn.Module):
+    r"""
+    The `Decoder` layer of a variational autoencoder that decodes its latent representation into an output sample.
+
+    Args:
+        dims (`int` or `Tuple[int, int]`, *optional*, defaults to 3):
+            The number of dimensions to use in convolutions.
+        in_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        out_channels (`int`, *optional*, defaults to 3):
+            The number of output channels.
+        blocks (`List[Tuple[str, int]]`, *optional*, defaults to `[("res_x", 1)]`):
+            The blocks to use. Each block is a tuple of the block name and the number of layers.
+        base_channels (`int`, *optional*, defaults to 128):
+            The number of output channels for the first convolutional layer.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups for normalization.
+        patch_size (`int`, *optional*, defaults to 1):
+            The patch size to use. Should be a power of 2.
+        norm_layer (`str`, *optional*, defaults to `group_norm`):
+            The normalization layer to use. Can be either `group_norm` or `pixel_norm`.
+        causal (`bool`, *optional*, defaults to `True`):
+            Whether to use causal convolutions or not.
+    """
+
+    def __init__(
+        self,
+        dims,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        blocks: List[Tuple[str, int | dict]] = [("res_x", 1)],
+        base_channels: int = 128,
+        layers_per_block: int = 2,
+        norm_num_groups: int = 32,
+        patch_size: int = 1,
+        norm_layer: str = "group_norm",
+        causal: bool = True,
+        timestep_conditioning: bool = False,
+        spatial_padding_mode: str = "zeros",
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.layers_per_block = layers_per_block
+        out_channels = out_channels * patch_size**2
+        self.causal = causal
+        self.blocks_desc = blocks
+
+        # Compute output channel to be product of all channel-multiplier blocks
+        output_channel = base_channels
+        for block_name, block_params in list(reversed(blocks)):
+            block_params = block_params if isinstance(block_params, dict) else {}
+            if block_name == "res_x_y":
+                output_channel = output_channel * block_params.get("multiplier", 2)
+            if block_name == "compress_all":
+                output_channel = output_channel * block_params.get("multiplier", 1)
+
+        self.conv_in = make_conv_nd(
+            dims,
+            in_channels,
+            output_channel,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            causal=True,
+            spatial_padding_mode=spatial_padding_mode,
+        )
+
+        self.up_blocks = nn.ModuleList([])
+
+        for block_name, block_params in list(reversed(blocks)):
+            input_channel = output_channel
+            if isinstance(block_params, int):
+                block_params = {"num_layers": block_params}
+
+            if block_name == "res_x":
+                block = UNetMidBlock3D(
+                    dims=dims,
+                    in_channels=input_channel,
+                    num_layers=block_params["num_layers"],
+                    resnet_eps=1e-6,
+                    resnet_groups=norm_num_groups,
+                    norm_layer=norm_layer,
+                    inject_noise=block_params.get("inject_noise", False),
+                    timestep_conditioning=timestep_conditioning,
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "attn_res_x":
+                block = UNetMidBlock3D(
+                    dims=dims,
+                    in_channels=input_channel,
+                    num_layers=block_params["num_layers"],
+                    resnet_groups=norm_num_groups,
+                    norm_layer=norm_layer,
+                    inject_noise=block_params.get("inject_noise", False),
+                    timestep_conditioning=timestep_conditioning,
+                    attention_head_dim=block_params["attention_head_dim"],
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "res_x_y":
+                output_channel = output_channel // block_params.get("multiplier", 2)
+                block = ResnetBlock3D(
+                    dims=dims,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    eps=1e-6,
+                    groups=norm_num_groups,
+                    norm_layer=norm_layer,
+                    inject_noise=block_params.get("inject_noise", False),
+                    timestep_conditioning=False,
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_time":
+                block = DepthToSpaceUpsample(
+                    dims=dims,
+                    in_channels=input_channel,
+                    stride=(2, 1, 1),
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_space":
+                block = DepthToSpaceUpsample(
+                    dims=dims,
+                    in_channels=input_channel,
+                    stride=(1, 2, 2),
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_all":
+                output_channel = output_channel // block_params.get("multiplier", 1)
+                block = DepthToSpaceUpsample(
+                    dims=dims,
+                    in_channels=input_channel,
+                    stride=(2, 2, 2),
+                    residual=block_params.get("residual", False),
+                    out_channels_reduction_factor=block_params.get("multiplier", 1),
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            else:
+                raise ValueError(f"unknown layer: {block_name}")
+
+            self.up_blocks.append(block)
+
+        if norm_layer == "group_norm":
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=output_channel, num_groups=norm_num_groups, eps=1e-6
+            )
+        elif norm_layer == "pixel_norm":
+            self.conv_norm_out = PixelNorm()
+        elif norm_layer == "layer_norm":
+            self.conv_norm_out = LayerNorm(output_channel, eps=1e-6)
+
+        self.conv_act = nn.SiLU()
+        self.conv_out = make_conv_nd(
+            dims,
+            output_channel,
+            out_channels,
+            3,
+            padding=1,
+            causal=True,
+            spatial_padding_mode=spatial_padding_mode,
+        )
+
+        self.gradient_checkpointing = False
+
+        self.timestep_conditioning = timestep_conditioning
+
+        if timestep_conditioning:
+            self.timestep_scale_multiplier = nn.Parameter(
+                torch.tensor(1000.0, dtype=torch.float32)
+            )
+            self.last_time_embedder = PixArtAlphaCombinedTimestepSizeEmbeddings(
+                output_channel * 2, 0, operations=ops,
+            )
+            self.last_scale_shift_table = nn.Parameter(torch.empty(2, output_channel))
+
+    # def forward(self, sample: torch.FloatTensor, target_shape) -> torch.FloatTensor:
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Optional[torch.Tensor] = None,
+    ) -> torch.FloatTensor:
+        r"""The forward method of the `Decoder` class."""
+        batch_size = sample.shape[0]
+
+        sample = self.conv_in(sample, causal=self.causal)
+
+        checkpoint_fn = (
+            partial(torch.utils.checkpoint.checkpoint, use_reentrant=False)
+            if self.gradient_checkpointing and self.training
+            else lambda x: x
+        )
+
+        scaled_timestep = None
+        if self.timestep_conditioning:
+            assert (
+                timestep is not None
+            ), "should pass timestep with timestep_conditioning=True"
+            scaled_timestep = timestep * self.timestep_scale_multiplier.to(dtype=sample.dtype, device=sample.device)
+
+        for up_block in self.up_blocks:
+            if self.timestep_conditioning and isinstance(up_block, UNetMidBlock3D):
+                sample = checkpoint_fn(up_block)(
+                    sample, causal=self.causal, timestep=scaled_timestep
+                )
+            else:
+                sample = checkpoint_fn(up_block)(sample, causal=self.causal)
+
+        sample = self.conv_norm_out(sample)
+
+        if self.timestep_conditioning:
+            embedded_timestep = self.last_time_embedder(
+                timestep=scaled_timestep.flatten(),
+                resolution=None,
+                aspect_ratio=None,
+                batch_size=sample.shape[0],
+                hidden_dtype=sample.dtype,
+            )
+            embedded_timestep = embedded_timestep.view(
+                batch_size, embedded_timestep.shape[-1], 1, 1, 1
+            )
+            ada_values = self.last_scale_shift_table[
+                None, ..., None, None, None
+            ].to(device=sample.device, dtype=sample.dtype) + embedded_timestep.reshape(
+                batch_size,
+                2,
+                -1,
+                embedded_timestep.shape[-3],
+                embedded_timestep.shape[-2],
+                embedded_timestep.shape[-1],
+            )
+            shift, scale = ada_values.unbind(dim=1)
+            sample = sample * (1 + scale) + shift
+
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample, causal=self.causal)
+
+        sample = unpatchify(sample, patch_size_hw=self.patch_size, patch_size_t=1)
+
+        return sample
+
+
+class UNetMidBlock3D(nn.Module):
+    """
+    A 3D UNet mid-block [`UNetMidBlock3D`] with multiple residual blocks.
+
+    Args:
+        in_channels (`int`): The number of input channels.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
+        num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
+        resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
+        resnet_groups (`int`, *optional*, defaults to 32):
+            The number of groups to use in the group normalization layers of the resnet blocks.
+        norm_layer (`str`, *optional*, defaults to `group_norm`):
+            The normalization layer to use. Can be either `group_norm` or `pixel_norm`.
+        inject_noise (`bool`, *optional*, defaults to `False`):
+            Whether to inject noise into the hidden states.
+        timestep_conditioning (`bool`, *optional*, defaults to `False`):
+            Whether to condition the hidden states on the timestep.
+
+    Returns:
+        `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
+        in_channels, height, width)`.
+
+    """
+
+    def __init__(
+        self,
+        dims: Union[int, Tuple[int, int]],
+        in_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_groups: int = 32,
+        norm_layer: str = "group_norm",
+        inject_noise: bool = False,
+        timestep_conditioning: bool = False,
+        spatial_padding_mode: str = "zeros",
+    ):
+        super().__init__()
+        resnet_groups = (
+            resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        )
+
+        self.timestep_conditioning = timestep_conditioning
+
+        if timestep_conditioning:
+            self.time_embedder = PixArtAlphaCombinedTimestepSizeEmbeddings(
+                in_channels * 4, 0, operations=ops,
+            )
+
+        self.res_blocks = nn.ModuleList(
+            [
+                ResnetBlock3D(
+                    dims=dims,
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    norm_layer=norm_layer,
+                    inject_noise=inject_noise,
+                    timestep_conditioning=timestep_conditioning,
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        causal: bool = True,
+        timestep: Optional[torch.Tensor] = None,
+    ) -> torch.FloatTensor:
+        timestep_embed = None
+        if self.timestep_conditioning:
+            assert (
+                timestep is not None
+            ), "should pass timestep with timestep_conditioning=True"
+            batch_size = hidden_states.shape[0]
+            timestep_embed = self.time_embedder(
+                timestep=timestep.flatten(),
+                resolution=None,
+                aspect_ratio=None,
+                batch_size=batch_size,
+                hidden_dtype=hidden_states.dtype,
+            )
+            timestep_embed = timestep_embed.view(
+                batch_size, timestep_embed.shape[-1], 1, 1, 1
+            )
+
+        for resnet in self.res_blocks:
+            hidden_states = resnet(hidden_states, causal=causal, timestep=timestep_embed)
+
+        return hidden_states
+
+
+class SpaceToDepthDownsample(nn.Module):
+    def __init__(self, dims, in_channels, out_channels, stride, spatial_padding_mode):
+        super().__init__()
+        self.stride = stride
+        self.group_size = in_channels * math.prod(stride) // out_channels
+        self.conv = make_conv_nd(
+            dims=dims,
+            in_channels=in_channels,
+            out_channels=out_channels // math.prod(stride),
+            kernel_size=3,
+            stride=1,
+            causal=True,
+            spatial_padding_mode=spatial_padding_mode,
+        )
+
+    def forward(self, x, causal: bool = True):
+        if self.stride[0] == 2:
+            x = torch.cat(
+                [x[:, :, :1, :, :], x], dim=2
+            )  # duplicate first frames for padding
+
+        # skip connection
+        x_in = rearrange(
+            x,
+            "b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w",
+            p1=self.stride[0],
+            p2=self.stride[1],
+            p3=self.stride[2],
+        )
+        x_in = rearrange(x_in, "b (c g) d h w -> b c g d h w", g=self.group_size)
+        x_in = x_in.mean(dim=2)
+
+        # conv
+        x = self.conv(x, causal=causal)
+        x = rearrange(
+            x,
+            "b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w",
+            p1=self.stride[0],
+            p2=self.stride[1],
+            p3=self.stride[2],
+        )
+
+        x = x + x_in
+
+        return x
+
+
+class DepthToSpaceUpsample(nn.Module):
+    def __init__(
+        self,
+        dims,
+        in_channels,
+        stride,
+        residual=False,
+        out_channels_reduction_factor=1,
+        spatial_padding_mode="zeros",
+    ):
+        super().__init__()
+        self.stride = stride
+        self.out_channels = (
+            math.prod(stride) * in_channels // out_channels_reduction_factor
+        )
+        self.conv = make_conv_nd(
+            dims=dims,
+            in_channels=in_channels,
+            out_channels=self.out_channels,
+            kernel_size=3,
+            stride=1,
+            causal=True,
+            spatial_padding_mode=spatial_padding_mode,
+        )
+        self.residual = residual
+        self.out_channels_reduction_factor = out_channels_reduction_factor
+
+    def forward(self, x, causal: bool = True, timestep: Optional[torch.Tensor] = None):
+        if self.residual:
+            # Reshape and duplicate the input to match the output shape
+            x_in = rearrange(
+                x,
+                "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)",
+                p1=self.stride[0],
+                p2=self.stride[1],
+                p3=self.stride[2],
+            )
+            num_repeat = math.prod(self.stride) // self.out_channels_reduction_factor
+            x_in = x_in.repeat(1, num_repeat, 1, 1, 1)
+            if self.stride[0] == 2:
+                x_in = x_in[:, :, 1:, :, :]
+        x = self.conv(x, causal=causal)
+        x = rearrange(
+            x,
+            "b (c p1 p2 p3) d h w -> b c (d p1) (h p2) (w p3)",
+            p1=self.stride[0],
+            p2=self.stride[1],
+            p3=self.stride[2],
+        )
+        if self.stride[0] == 2:
+            x = x[:, :, 1:, :, :]
+        if self.residual:
+            x = x + x_in
+        return x
+
+class LayerNorm(nn.Module):
+    def __init__(self, dim, eps, elementwise_affine=True) -> None:
+        super().__init__()
+        self.norm = ops.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine)
+
+    def forward(self, x):
+        x = rearrange(x, "b c d h w -> b d h w c")
+        x = self.norm(x)
+        x = rearrange(x, "b d h w c -> b c d h w")
+        return x
+
+
+class ResnetBlock3D(nn.Module):
+    r"""
+    A Resnet block.
+
+    Parameters:
+        in_channels (`int`): The number of channels in the input.
+        out_channels (`int`, *optional*, default to be `None`):
+            The number of output channels for the first conv layer. If None, same as `in_channels`.
+        dropout (`float`, *optional*, defaults to `0.0`): The dropout probability to use.
+        groups (`int`, *optional*, default to `32`): The number of groups to use for the first normalization layer.
+        eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the normalization.
+    """
+
+    def __init__(
+        self,
+        dims: Union[int, Tuple[int, int]],
+        in_channels: int,
+        out_channels: Optional[int] = None,
+        dropout: float = 0.0,
+        groups: int = 32,
+        eps: float = 1e-6,
+        norm_layer: str = "group_norm",
+        inject_noise: bool = False,
+        timestep_conditioning: bool = False,
+        spatial_padding_mode: str = "zeros",
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.inject_noise = inject_noise
+
+        if norm_layer == "group_norm":
+            self.norm1 = nn.GroupNorm(
+                num_groups=groups, num_channels=in_channels, eps=eps, affine=True
+            )
+        elif norm_layer == "pixel_norm":
+            self.norm1 = PixelNorm()
+        elif norm_layer == "layer_norm":
+            self.norm1 = LayerNorm(in_channels, eps=eps, elementwise_affine=True)
+
+        self.non_linearity = nn.SiLU()
+
+        self.conv1 = make_conv_nd(
+            dims,
+            in_channels,
+            out_channels,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            causal=True,
+            spatial_padding_mode=spatial_padding_mode,
+        )
+
+        if inject_noise:
+            self.per_channel_scale1 = nn.Parameter(torch.zeros((in_channels, 1, 1)))
+
+        if norm_layer == "group_norm":
+            self.norm2 = nn.GroupNorm(
+                num_groups=groups, num_channels=out_channels, eps=eps, affine=True
+            )
+        elif norm_layer == "pixel_norm":
+            self.norm2 = PixelNorm()
+        elif norm_layer == "layer_norm":
+            self.norm2 = LayerNorm(out_channels, eps=eps, elementwise_affine=True)
+
+        self.dropout = torch.nn.Dropout(dropout)
+
+        self.conv2 = make_conv_nd(
+            dims,
+            out_channels,
+            out_channels,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+            causal=True,
+            spatial_padding_mode=spatial_padding_mode,
+        )
+
+        if inject_noise:
+            self.per_channel_scale2 = nn.Parameter(torch.zeros((in_channels, 1, 1)))
+
+        self.conv_shortcut = (
+            make_linear_nd(
+                dims=dims, in_channels=in_channels, out_channels=out_channels
+            )
+            if in_channels != out_channels
+            else nn.Identity()
+        )
+
+        self.norm3 = (
+            LayerNorm(in_channels, eps=eps, elementwise_affine=True)
+            if in_channels != out_channels
+            else nn.Identity()
+        )
+
+        self.timestep_conditioning = timestep_conditioning
+
+        if timestep_conditioning:
+            self.scale_shift_table = nn.Parameter(
+                torch.randn(4, in_channels) / in_channels**0.5
+            )
+
+    def _feed_spatial_noise(
+        self, hidden_states: torch.FloatTensor, per_channel_scale: torch.FloatTensor
+    ) -> torch.FloatTensor:
+        spatial_shape = hidden_states.shape[-2:]
+        device = hidden_states.device
+        dtype = hidden_states.dtype
+
+        # similar to the "explicit noise inputs" method in style-gan
+        spatial_noise = torch.randn(spatial_shape, device=device, dtype=dtype)[None]
+        scaled_noise = (spatial_noise * per_channel_scale)[None, :, None, ...]
+        hidden_states = hidden_states + scaled_noise
+
+        return hidden_states
+
+    def forward(
+        self,
+        input_tensor: torch.FloatTensor,
+        causal: bool = True,
+        timestep: Optional[torch.Tensor] = None,
+    ) -> torch.FloatTensor:
+        hidden_states = input_tensor
+        batch_size = hidden_states.shape[0]
+
+        hidden_states = self.norm1(hidden_states)
+        if self.timestep_conditioning:
+            assert (
+                timestep is not None
+            ), "should pass timestep with timestep_conditioning=True"
+            ada_values = self.scale_shift_table[
+                None, ..., None, None, None
+            ].to(device=hidden_states.device, dtype=hidden_states.dtype) + timestep.reshape(
+                batch_size,
+                4,
+                -1,
+                timestep.shape[-3],
+                timestep.shape[-2],
+                timestep.shape[-1],
+            )
+            shift1, scale1, shift2, scale2 = ada_values.unbind(dim=1)
+
+            hidden_states = hidden_states * (1 + scale1) + shift1
+
+        hidden_states = self.non_linearity(hidden_states)
+
+        hidden_states = self.conv1(hidden_states, causal=causal)
+
+        if self.inject_noise:
+            hidden_states = self._feed_spatial_noise(
+                hidden_states, self.per_channel_scale1.to(device=hidden_states.device, dtype=hidden_states.dtype)
+            )
+
+        hidden_states = self.norm2(hidden_states)
+
+        if self.timestep_conditioning:
+            hidden_states = hidden_states * (1 + scale2) + shift2
+
+        hidden_states = self.non_linearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+
+        hidden_states = self.conv2(hidden_states, causal=causal)
+
+        if self.inject_noise:
+            hidden_states = self._feed_spatial_noise(
+                hidden_states, self.per_channel_scale2.to(device=hidden_states.device, dtype=hidden_states.dtype)
+            )
+
+        input_tensor = self.norm3(input_tensor)
+
+        batch_size = input_tensor.shape[0]
+
+        input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = input_tensor + hidden_states
+
+        return output_tensor
+
+
+def patchify(x, patch_size_hw, patch_size_t=1):
+    if patch_size_hw == 1 and patch_size_t == 1:
+        return x
+    if x.dim() == 4:
+        x = rearrange(
+            x, "b c (h q) (w r) -> b (c r q) h w", q=patch_size_hw, r=patch_size_hw
+        )
+    elif x.dim() == 5:
+        x = rearrange(
+            x,
+            "b c (f p) (h q) (w r) -> b (c p r q) f h w",
+            p=patch_size_t,
+            q=patch_size_hw,
+            r=patch_size_hw,
+        )
+    else:
+        raise ValueError(f"Invalid input shape: {x.shape}")
+
+    return x
+
+
+def unpatchify(x, patch_size_hw, patch_size_t=1):
+    if patch_size_hw == 1 and patch_size_t == 1:
+        return x
+
+    if x.dim() == 4:
+        x = rearrange(
+            x, "b (c r q) h w -> b c (h q) (w r)", q=patch_size_hw, r=patch_size_hw
+        )
+    elif x.dim() == 5:
+        x = rearrange(
+            x,
+            "b (c p r q) f h w -> b c (f p) (h q) (w r)",
+            p=patch_size_t,
+            q=patch_size_hw,
+            r=patch_size_hw,
+        )
+
+    return x
+
+class processor(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.register_buffer("std-of-means", torch.empty(128))
+        self.register_buffer("mean-of-means", torch.empty(128))
+        self.register_buffer("mean-of-stds", torch.empty(128))
+        self.register_buffer("mean-of-stds_over_std-of-means", torch.empty(128))
+        self.register_buffer("channel", torch.empty(128))
+
+    def un_normalize(self, x):
+        return (x * self.get_buffer("std-of-means").view(1, -1, 1, 1, 1).to(x)) + self.get_buffer("mean-of-means").view(1, -1, 1, 1, 1).to(x)
+
+    def normalize(self, x):
+        return (x - self.get_buffer("mean-of-means").view(1, -1, 1, 1, 1).to(x)) / self.get_buffer("std-of-means").view(1, -1, 1, 1, 1).to(x)
+
+class VideoVAE(nn.Module):
+    def __init__(self, version=0, config=None):
+        super().__init__()
+
+        if config is None:
+            config = self.guess_config(version)
+
+        self.timestep_conditioning = config.get("timestep_conditioning", False)
+        double_z = config.get("double_z", True)
+        latent_log_var = config.get(
+            "latent_log_var", "per_channel" if double_z else "none"
+        )
+
+        self.encoder = Encoder(
+            dims=config["dims"],
+            in_channels=config.get("in_channels", 3),
+            out_channels=config["latent_channels"],
+            blocks=config.get("encoder_blocks", config.get("encoder_blocks", config.get("blocks"))),
+            patch_size=config.get("patch_size", 1),
+            latent_log_var=latent_log_var,
+            norm_layer=config.get("norm_layer", "group_norm"),
+            spatial_padding_mode=config.get("spatial_padding_mode", "zeros"),
+        )
+
+        self.decoder = Decoder(
+            dims=config["dims"],
+            in_channels=config["latent_channels"],
+            out_channels=config.get("out_channels", 3),
+            blocks=config.get("decoder_blocks", config.get("decoder_blocks", config.get("blocks"))),
+            patch_size=config.get("patch_size", 1),
+            norm_layer=config.get("norm_layer", "group_norm"),
+            causal=config.get("causal_decoder", False),
+            timestep_conditioning=self.timestep_conditioning,
+            spatial_padding_mode=config.get("spatial_padding_mode", "reflect"),
+        )
+
+        self.per_channel_statistics = processor()
+
+    def guess_config(self, version):
+        if version == 0:
+            config = {
+                "_class_name": "CausalVideoAutoencoder",
+                "dims": 3,
+                "in_channels": 3,
+                "out_channels": 3,
+                "latent_channels": 128,
+                "blocks": [
+                    ["res_x", 4],
+                    ["compress_all", 1],
+                    ["res_x_y", 1],
+                    ["res_x", 3],
+                    ["compress_all", 1],
+                    ["res_x_y", 1],
+                    ["res_x", 3],
+                    ["compress_all", 1],
+                    ["res_x", 3],
+                    ["res_x", 4],
+                ],
+                "scaling_factor": 1.0,
+                "norm_layer": "pixel_norm",
+                "patch_size": 4,
+                "latent_log_var": "uniform",
+                "use_quant_conv": False,
+                "causal_decoder": False,
+            }
+        elif version == 1:
+            config = {
+                "_class_name": "CausalVideoAutoencoder",
+                "dims": 3,
+                "in_channels": 3,
+                "out_channels": 3,
+                "latent_channels": 128,
+                "decoder_blocks": [
+                    ["res_x", {"num_layers": 5, "inject_noise": True}],
+                    ["compress_all", {"residual": True, "multiplier": 2}],
+                    ["res_x", {"num_layers": 6, "inject_noise": True}],
+                    ["compress_all", {"residual": True, "multiplier": 2}],
+                    ["res_x", {"num_layers": 7, "inject_noise": True}],
+                    ["compress_all", {"residual": True, "multiplier": 2}],
+                    ["res_x", {"num_layers": 8, "inject_noise": False}]
+                ],
+                "encoder_blocks": [
+                    ["res_x", {"num_layers": 4}],
+                    ["compress_all", {}],
+                    ["res_x_y", 1],
+                    ["res_x", {"num_layers": 3}],
+                    ["compress_all", {}],
+                    ["res_x_y", 1],
+                    ["res_x", {"num_layers": 3}],
+                    ["compress_all", {}],
+                    ["res_x", {"num_layers": 3}],
+                    ["res_x", {"num_layers": 4}]
+                ],
+                "scaling_factor": 1.0,
+                "norm_layer": "pixel_norm",
+                "patch_size": 4,
+                "latent_log_var": "uniform",
+                "use_quant_conv": False,
+                "causal_decoder": False,
+                "timestep_conditioning": True,
+            }
+        else:
+            config = {
+                "_class_name": "CausalVideoAutoencoder",
+                "dims": 3,
+                "in_channels": 3,
+                "out_channels": 3,
+                "latent_channels": 128,
+                "encoder_blocks": [
+                    ["res_x", {"num_layers": 4}],
+                    ["compress_space_res", {"multiplier": 2}],
+                    ["res_x", {"num_layers": 6}],
+                    ["compress_time_res", {"multiplier": 2}],
+                    ["res_x", {"num_layers": 6}],
+                    ["compress_all_res", {"multiplier": 2}],
+                    ["res_x", {"num_layers": 2}],
+                    ["compress_all_res", {"multiplier": 2}],
+                    ["res_x", {"num_layers": 2}]
+                ],
+                "decoder_blocks": [
+                    ["res_x", {"num_layers": 5, "inject_noise": False}],
+                    ["compress_all", {"residual": True, "multiplier": 2}],
+                    ["res_x", {"num_layers": 5, "inject_noise": False}],
+                    ["compress_all", {"residual": True, "multiplier": 2}],
+                    ["res_x", {"num_layers": 5, "inject_noise": False}],
+                    ["compress_all", {"residual": True, "multiplier": 2}],
+                    ["res_x", {"num_layers": 5, "inject_noise": False}]
+                ],
+                "scaling_factor": 1.0,
+                "norm_layer": "pixel_norm",
+                "patch_size": 4,
+                "latent_log_var": "uniform",
+                "use_quant_conv": False,
+                "causal_decoder": False,
+                "timestep_conditioning": True
+            }
+        return config
+
+    def encode(self, x):
+        frames_count = x.shape[2]
+        if ((frames_count - 1) % 8) != 0:
+            raise ValueError("Invalid number of frames: Encode input must have 1 + 8 * x frames (e.g., 1, 9, 17, ...). Please check your input.")
+        means, logvar = torch.chunk(self.encoder(x), 2, dim=1)
+        return self.per_channel_statistics.normalize(means)
+
+    def decode(self, x, timestep=0.05, noise_scale=0.025):
+        if self.timestep_conditioning: #TODO: seed
+            x = torch.randn_like(x) * noise_scale + (1.0 - noise_scale) * x
+        return self.decoder(self.per_channel_statistics.un_normalize(x), timestep=timestep)
+

comfy/ldm/lightricks/vae/conv_nd_factory.py

+from typing import Tuple, Union
+
+
+from .dual_conv3d import DualConv3d
+from .causal_conv3d import CausalConv3d
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+def make_conv_nd(
+    dims: Union[int, Tuple[int, int]],
+    in_channels: int,
+    out_channels: int,
+    kernel_size: int,
+    stride=1,
+    padding=0,
+    dilation=1,
+    groups=1,
+    bias=True,
+    causal=False,
+    spatial_padding_mode="zeros",
+    temporal_padding_mode="zeros",
+):
+    if not (spatial_padding_mode == temporal_padding_mode or causal):
+        raise NotImplementedError("spatial and temporal padding modes must be equal")
+    if dims == 2:
+        return ops.Conv2d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            groups=groups,
+            bias=bias,
+            padding_mode=spatial_padding_mode,
+        )
+    elif dims == 3:
+        if causal:
+            return CausalConv3d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=kernel_size,
+                stride=stride,
+                padding=padding,
+                dilation=dilation,
+                groups=groups,
+                bias=bias,
+                spatial_padding_mode=spatial_padding_mode,
+            )
+        return ops.Conv3d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            groups=groups,
+            bias=bias,
+            padding_mode=spatial_padding_mode,
+        )
+    elif dims == (2, 1):
+        return DualConv3d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            bias=bias,
+            padding_mode=spatial_padding_mode,
+        )
+    else:
+        raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def make_linear_nd(
+    dims: int,
+    in_channels: int,
+    out_channels: int,
+    bias=True,
+):
+    if dims == 2:
+        return ops.Conv2d(
+            in_channels=in_channels, out_channels=out_channels, kernel_size=1, bias=bias
+        )
+    elif dims == 3 or dims == (2, 1):
+        return ops.Conv3d(
+            in_channels=in_channels, out_channels=out_channels, kernel_size=1, bias=bias
+        )
+    else:
+        raise ValueError(f"unsupported dimensions: {dims}")

comfy/ldm/lightricks/vae/dual_conv3d.py

+import math
+from typing import Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+
+class DualConv3d(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride: Union[int, Tuple[int, int, int]] = 1,
+        padding: Union[int, Tuple[int, int, int]] = 0,
+        dilation: Union[int, Tuple[int, int, int]] = 1,
+        groups=1,
+        bias=True,
+        padding_mode="zeros",
+    ):
+        super(DualConv3d, self).__init__()
+
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.padding_mode = padding_mode
+        # Ensure kernel_size, stride, padding, and dilation are tuples of length 3
+        if isinstance(kernel_size, int):
+            kernel_size = (kernel_size, kernel_size, kernel_size)
+        if kernel_size == (1, 1, 1):
+            raise ValueError(
+                "kernel_size must be greater than 1. Use make_linear_nd instead."
+            )
+        if isinstance(stride, int):
+            stride = (stride, stride, stride)
+        if isinstance(padding, int):
+            padding = (padding, padding, padding)
+        if isinstance(dilation, int):
+            dilation = (dilation, dilation, dilation)
+
+        # Set parameters for convolutions
+        self.groups = groups
+        self.bias = bias
+
+        # Define the size of the channels after the first convolution
+        intermediate_channels = (
+            out_channels if in_channels < out_channels else in_channels
+        )
+
+        # Define parameters for the first convolution
+        self.weight1 = nn.Parameter(
+            torch.Tensor(
+                intermediate_channels,
+                in_channels // groups,
+                1,
+                kernel_size[1],
+                kernel_size[2],
+            )
+        )
+        self.stride1 = (1, stride[1], stride[2])
+        self.padding1 = (0, padding[1], padding[2])
+        self.dilation1 = (1, dilation[1], dilation[2])
+        if bias:
+            self.bias1 = nn.Parameter(torch.Tensor(intermediate_channels))
+        else:
+            self.register_parameter("bias1", None)
+
+        # Define parameters for the second convolution
+        self.weight2 = nn.Parameter(
+            torch.Tensor(
+                out_channels, intermediate_channels // groups, kernel_size[0], 1, 1
+            )
+        )
+        self.stride2 = (stride[0], 1, 1)
+        self.padding2 = (padding[0], 0, 0)
+        self.dilation2 = (dilation[0], 1, 1)
+        if bias:
+            self.bias2 = nn.Parameter(torch.Tensor(out_channels))
+        else:
+            self.register_parameter("bias2", None)
+
+        # Initialize weights and biases
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        nn.init.kaiming_uniform_(self.weight1, a=math.sqrt(5))
+        nn.init.kaiming_uniform_(self.weight2, a=math.sqrt(5))
+        if self.bias:
+            fan_in1, _ = nn.init._calculate_fan_in_and_fan_out(self.weight1)
+            bound1 = 1 / math.sqrt(fan_in1)
+            nn.init.uniform_(self.bias1, -bound1, bound1)
+            fan_in2, _ = nn.init._calculate_fan_in_and_fan_out(self.weight2)
+            bound2 = 1 / math.sqrt(fan_in2)
+            nn.init.uniform_(self.bias2, -bound2, bound2)
+
+    def forward(self, x, use_conv3d=False, skip_time_conv=False):
+        if use_conv3d:
+            return self.forward_with_3d(x=x, skip_time_conv=skip_time_conv)
+        else:
+            return self.forward_with_2d(x=x, skip_time_conv=skip_time_conv)
+
+    def forward_with_3d(self, x, skip_time_conv):
+        # First convolution
+        x = F.conv3d(
+            x,
+            self.weight1,
+            self.bias1,
+            self.stride1,
+            self.padding1,
+            self.dilation1,
+            self.groups,
+            padding_mode=self.padding_mode,
+        )
+
+        if skip_time_conv:
+            return x
+
+        # Second convolution
+        x = F.conv3d(
+            x,
+            self.weight2,
+            self.bias2,
+            self.stride2,
+            self.padding2,
+            self.dilation2,
+            self.groups,
+            padding_mode=self.padding_mode,
+        )
+
+        return x
+
+    def forward_with_2d(self, x, skip_time_conv):
+        b, c, d, h, w = x.shape
+
+        # First 2D convolution
+        x = rearrange(x, "b c d h w -> (b d) c h w")
+        # Squeeze the depth dimension out of weight1 since it's 1
+        weight1 = self.weight1.squeeze(2)
+        # Select stride, padding, and dilation for the 2D convolution
+        stride1 = (self.stride1[1], self.stride1[2])
+        padding1 = (self.padding1[1], self.padding1[2])
+        dilation1 = (self.dilation1[1], self.dilation1[2])
+        x = F.conv2d(
+            x,
+            weight1,
+            self.bias1,
+            stride1,
+            padding1,
+            dilation1,
+            self.groups,
+            padding_mode=self.padding_mode,
+        )
+
+        _, _, h, w = x.shape
+
+        if skip_time_conv:
+            x = rearrange(x, "(b d) c h w -> b c d h w", b=b)
+            return x
+
+        # Second convolution which is essentially treated as a 1D convolution across the 'd' dimension
+        x = rearrange(x, "(b d) c h w -> (b h w) c d", b=b)
+
+        # Reshape weight2 to match the expected dimensions for conv1d
+        weight2 = self.weight2.squeeze(-1).squeeze(-1)
+        # Use only the relevant dimension for stride, padding, and dilation for the 1D convolution
+        stride2 = self.stride2[0]
+        padding2 = self.padding2[0]
+        dilation2 = self.dilation2[0]
+        x = F.conv1d(
+            x,
+            weight2,
+            self.bias2,
+            stride2,
+            padding2,
+            dilation2,
+            self.groups,
+            padding_mode=self.padding_mode,
+        )
+        x = rearrange(x, "(b h w) c d -> b c d h w", b=b, h=h, w=w)
+
+        return x
+
+    @property
+    def weight(self):
+        return self.weight2
+
+
+def test_dual_conv3d_consistency():
+    # Initialize parameters
+    in_channels = 3
+    out_channels = 5
+    kernel_size = (3, 3, 3)
+    stride = (2, 2, 2)
+    padding = (1, 1, 1)
+
+    # Create an instance of the DualConv3d class
+    dual_conv3d = DualConv3d(
+        in_channels=in_channels,
+        out_channels=out_channels,
+        kernel_size=kernel_size,
+        stride=stride,
+        padding=padding,
+        bias=True,
+    )
+
+    # Example input tensor
+    test_input = torch.randn(1, 3, 10, 10, 10)
+
+    # Perform forward passes with both 3D and 2D settings
+    output_conv3d = dual_conv3d(test_input, use_conv3d=True)
+    output_2d = dual_conv3d(test_input, use_conv3d=False)
+
+    # Assert that the outputs from both methods are sufficiently close
+    assert torch.allclose(
+        output_conv3d, output_2d, atol=1e-6
+    ), "Outputs are not consistent between 3D and 2D convolutions."

comfy/ldm/lightricks/vae/pixel_norm.py

+import torch
+from torch import nn
+
+
+class PixelNorm(nn.Module):
+    def __init__(self, dim=1, eps=1e-8):
+        super(PixelNorm, self).__init__()
+        self.dim = dim
+        self.eps = eps
+
+    def forward(self, x):
+        return x / torch.sqrt(torch.mean(x**2, dim=self.dim, keepdim=True) + self.eps)

comfy/ldm/lumina/model.py

+# Code from: https://github.com/Alpha-VLLM/Lumina-Image-2.0/blob/main/models/model.py
+from __future__ import annotations
+
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import comfy.ldm.common_dit
+
+from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder
+from comfy.ldm.modules.attention import optimized_attention_masked
+from comfy.ldm.flux.layers import EmbedND
+import comfy.patcher_extension
+
+
+def modulate(x, scale):
+    return x * (1 + scale.unsqueeze(1))
+
+#############################################################################
+#                               Core NextDiT Model                              #
+#############################################################################
+
+
+class JointAttention(nn.Module):
+    """Multi-head attention module."""
+
+    def __init__(
+        self,
+        dim: int,
+        n_heads: int,
+        n_kv_heads: Optional[int],
+        qk_norm: bool,
+        operation_settings={},
+    ):
+        """
+        Initialize the Attention module.
+
+        Args:
+            dim (int): Number of input dimensions.
+            n_heads (int): Number of heads.
+            n_kv_heads (Optional[int]): Number of kv heads, if using GQA.
+
+        """
+        super().__init__()
+        self.n_kv_heads = n_heads if n_kv_heads is None else n_kv_heads
+        self.n_local_heads = n_heads
+        self.n_local_kv_heads = self.n_kv_heads
+        self.n_rep = self.n_local_heads // self.n_local_kv_heads
+        self.head_dim = dim // n_heads
+
+        self.qkv = operation_settings.get("operations").Linear(
+            dim,
+            (n_heads + self.n_kv_heads + self.n_kv_heads) * self.head_dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.out = operation_settings.get("operations").Linear(
+            n_heads * self.head_dim,
+            dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+        if qk_norm:
+            self.q_norm = operation_settings.get("operations").RMSNorm(self.head_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+            self.k_norm = operation_settings.get("operations").RMSNorm(self.head_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        else:
+            self.q_norm = self.k_norm = nn.Identity()
+
+    @staticmethod
+    def apply_rotary_emb(
+        x_in: torch.Tensor,
+        freqs_cis: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Apply rotary embeddings to input tensors using the given frequency
+        tensor.
+
+        This function applies rotary embeddings to the given query 'xq' and
+        key 'xk' tensors using the provided frequency tensor 'freqs_cis'. The
+        input tensors are reshaped as complex numbers, and the frequency tensor
+        is reshaped for broadcasting compatibility. The resulting tensors
+        contain rotary embeddings and are returned as real tensors.
+
+        Args:
+            x_in (torch.Tensor): Query or Key tensor to apply rotary embeddings.
+            freqs_cis (torch.Tensor): Precomputed frequency tensor for complex
+                exponentials.
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor
+                and key tensor with rotary embeddings.
+        """
+
+        t_ = x_in.reshape(*x_in.shape[:-1], -1, 1, 2)
+        t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1]
+        return t_out.reshape(*x_in.shape)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        freqs_cis: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+
+        Args:
+            x:
+            x_mask:
+            freqs_cis:
+
+        Returns:
+
+        """
+        bsz, seqlen, _ = x.shape
+
+        xq, xk, xv = torch.split(
+            self.qkv(x),
+            [
+                self.n_local_heads * self.head_dim,
+                self.n_local_kv_heads * self.head_dim,
+                self.n_local_kv_heads * self.head_dim,
+            ],
+            dim=-1,
+        )
+        xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim)
+        xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
+        xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
+
+        xq = self.q_norm(xq)
+        xk = self.k_norm(xk)
+
+        xq = JointAttention.apply_rotary_emb(xq, freqs_cis=freqs_cis)
+        xk = JointAttention.apply_rotary_emb(xk, freqs_cis=freqs_cis)
+
+        n_rep = self.n_local_heads // self.n_local_kv_heads
+        if n_rep >= 1:
+            xk = xk.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
+            xv = xv.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
+        output = optimized_attention_masked(xq.movedim(1, 2), xk.movedim(1, 2), xv.movedim(1, 2), self.n_local_heads, x_mask, skip_reshape=True)
+
+        return self.out(output)
+
+
+class FeedForward(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        multiple_of: int,
+        ffn_dim_multiplier: Optional[float],
+        operation_settings={},
+    ):
+        """
+        Initialize the FeedForward module.
+
+        Args:
+            dim (int): Input dimension.
+            hidden_dim (int): Hidden dimension of the feedforward layer.
+            multiple_of (int): Value to ensure hidden dimension is a multiple
+                of this value.
+            ffn_dim_multiplier (float, optional): Custom multiplier for hidden
+                dimension. Defaults to None.
+
+        """
+        super().__init__()
+        # custom dim factor multiplier
+        if ffn_dim_multiplier is not None:
+            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
+        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+        self.w1 = operation_settings.get("operations").Linear(
+            dim,
+            hidden_dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.w2 = operation_settings.get("operations").Linear(
+            hidden_dim,
+            dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.w3 = operation_settings.get("operations").Linear(
+            dim,
+            hidden_dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+    # @torch.compile
+    def _forward_silu_gating(self, x1, x3):
+        return F.silu(x1) * x3
+
+    def forward(self, x):
+        return self.w2(self._forward_silu_gating(self.w1(x), self.w3(x)))
+
+
+class JointTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        layer_id: int,
+        dim: int,
+        n_heads: int,
+        n_kv_heads: int,
+        multiple_of: int,
+        ffn_dim_multiplier: float,
+        norm_eps: float,
+        qk_norm: bool,
+        modulation=True,
+        operation_settings={},
+    ) -> None:
+        """
+        Initialize a TransformerBlock.
+
+        Args:
+            layer_id (int): Identifier for the layer.
+            dim (int): Embedding dimension of the input features.
+            n_heads (int): Number of attention heads.
+            n_kv_heads (Optional[int]): Number of attention heads in key and
+                value features (if using GQA), or set to None for the same as
+                query.
+            multiple_of (int):
+            ffn_dim_multiplier (float):
+            norm_eps (float):
+
+        """
+        super().__init__()
+        self.dim = dim
+        self.head_dim = dim // n_heads
+        self.attention = JointAttention(dim, n_heads, n_kv_heads, qk_norm, operation_settings=operation_settings)
+        self.feed_forward = FeedForward(
+            dim=dim,
+            hidden_dim=4 * dim,
+            multiple_of=multiple_of,
+            ffn_dim_multiplier=ffn_dim_multiplier,
+            operation_settings=operation_settings,
+        )
+        self.layer_id = layer_id
+        self.attention_norm1 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.ffn_norm1 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        self.attention_norm2 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.ffn_norm2 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        self.modulation = modulation
+        if modulation:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                operation_settings.get("operations").Linear(
+                    min(dim, 1024),
+                    4 * dim,
+                    bias=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+            )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        freqs_cis: torch.Tensor,
+        adaln_input: Optional[torch.Tensor]=None,
+    ):
+        """
+        Perform a forward pass through the TransformerBlock.
+
+        Args:
+            x (torch.Tensor): Input tensor.
+            freqs_cis (torch.Tensor): Precomputed cosine and sine frequencies.
+
+        Returns:
+            torch.Tensor: Output tensor after applying attention and
+                feedforward layers.
+
+        """
+        if self.modulation:
+            assert adaln_input is not None
+            scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).chunk(4, dim=1)
+
+            x = x + gate_msa.unsqueeze(1).tanh() * self.attention_norm2(
+                self.attention(
+                    modulate(self.attention_norm1(x), scale_msa),
+                    x_mask,
+                    freqs_cis,
+                )
+            )
+            x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(
+                self.feed_forward(
+                    modulate(self.ffn_norm1(x), scale_mlp),
+                )
+            )
+        else:
+            assert adaln_input is None
+            x = x + self.attention_norm2(
+                self.attention(
+                    self.attention_norm1(x),
+                    x_mask,
+                    freqs_cis,
+                )
+            )
+            x = x + self.ffn_norm2(
+                self.feed_forward(
+                    self.ffn_norm1(x),
+                )
+            )
+        return x
+
+
+class FinalLayer(nn.Module):
+    """
+    The final layer of NextDiT.
+    """
+
+    def __init__(self, hidden_size, patch_size, out_channels, operation_settings={}):
+        super().__init__()
+        self.norm_final = operation_settings.get("operations").LayerNorm(
+            hidden_size,
+            elementwise_affine=False,
+            eps=1e-6,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.linear = operation_settings.get("operations").Linear(
+            hidden_size,
+            patch_size * patch_size * out_channels,
+            bias=True,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operation_settings.get("operations").Linear(
+                min(hidden_size, 1024),
+                hidden_size,
+                bias=True,
+                device=operation_settings.get("device"),
+                dtype=operation_settings.get("dtype"),
+            ),
+        )
+
+    def forward(self, x, c):
+        scale = self.adaLN_modulation(c)
+        x = modulate(self.norm_final(x), scale)
+        x = self.linear(x)
+        return x
+
+
+class NextDiT(nn.Module):
+    """
+    Diffusion model with a Transformer backbone.
+    """
+
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 4,
+        dim: int = 4096,
+        n_layers: int = 32,
+        n_refiner_layers: int = 2,
+        n_heads: int = 32,
+        n_kv_heads: Optional[int] = None,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+        norm_eps: float = 1e-5,
+        qk_norm: bool = False,
+        cap_feat_dim: int = 5120,
+        axes_dims: List[int] = (16, 56, 56),
+        axes_lens: List[int] = (1, 512, 512),
+        image_model=None,
+        device=None,
+        dtype=None,
+        operations=None,
+    ) -> None:
+        super().__init__()
+        self.dtype = dtype
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+        self.patch_size = patch_size
+
+        self.x_embedder = operation_settings.get("operations").Linear(
+            in_features=patch_size * patch_size * in_channels,
+            out_features=dim,
+            bias=True,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+        self.noise_refiner = nn.ModuleList(
+            [
+                JointTransformerBlock(
+                    layer_id,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    qk_norm,
+                    modulation=True,
+                    operation_settings=operation_settings,
+                )
+                for layer_id in range(n_refiner_layers)
+            ]
+        )
+        self.context_refiner = nn.ModuleList(
+            [
+                JointTransformerBlock(
+                    layer_id,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    qk_norm,
+                    modulation=False,
+                    operation_settings=operation_settings,
+                )
+                for layer_id in range(n_refiner_layers)
+            ]
+        )
+
+        self.t_embedder = TimestepEmbedder(min(dim, 1024), **operation_settings)
+        self.cap_embedder = nn.Sequential(
+            operation_settings.get("operations").RMSNorm(cap_feat_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+            operation_settings.get("operations").Linear(
+                cap_feat_dim,
+                dim,
+                bias=True,
+                device=operation_settings.get("device"),
+                dtype=operation_settings.get("dtype"),
+            ),
+        )
+
+        self.layers = nn.ModuleList(
+            [
+                JointTransformerBlock(
+                    layer_id,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    qk_norm,
+                    operation_settings=operation_settings,
+                )
+                for layer_id in range(n_layers)
+            ]
+        )
+        self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.final_layer = FinalLayer(dim, patch_size, self.out_channels, operation_settings=operation_settings)
+
+        assert (dim // n_heads) == sum(axes_dims)
+        self.axes_dims = axes_dims
+        self.axes_lens = axes_lens
+        self.rope_embedder = EmbedND(dim=dim // n_heads, theta=10000.0, axes_dim=axes_dims)
+        self.dim = dim
+        self.n_heads = n_heads
+
+    def unpatchify(
+        self, x: torch.Tensor, img_size: List[Tuple[int, int]], cap_size: List[int], return_tensor=False
+    ) -> List[torch.Tensor]:
+        """
+        x: (N, T, patch_size**2 * C)
+        imgs: (N, H, W, C)
+        """
+        pH = pW = self.patch_size
+        imgs = []
+        for i in range(x.size(0)):
+            H, W = img_size[i]
+            begin = cap_size[i]
+            end = begin + (H // pH) * (W // pW)
+            imgs.append(
+                x[i][begin:end]
+                .view(H // pH, W // pW, pH, pW, self.out_channels)
+                .permute(4, 0, 2, 1, 3)
+                .flatten(3, 4)
+                .flatten(1, 2)
+            )
+
+        if return_tensor:
+            imgs = torch.stack(imgs, dim=0)
+        return imgs
+
+    def patchify_and_embed(
+        self, x: List[torch.Tensor] | torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], List[int], torch.Tensor]:
+        bsz = len(x)
+        pH = pW = self.patch_size
+        device = x[0].device
+        dtype = x[0].dtype
+
+        if cap_mask is not None:
+            l_effective_cap_len = cap_mask.sum(dim=1).tolist()
+        else:
+            l_effective_cap_len = [num_tokens] * bsz
+
+        if cap_mask is not None and not torch.is_floating_point(cap_mask):
+            cap_mask = (cap_mask - 1).to(dtype) * torch.finfo(dtype).max
+
+        img_sizes = [(img.size(1), img.size(2)) for img in x]
+        l_effective_img_len = [(H // pH) * (W // pW) for (H, W) in img_sizes]
+
+        max_seq_len = max(
+            (cap_len+img_len for cap_len, img_len in zip(l_effective_cap_len, l_effective_img_len))
+        )
+        max_cap_len = max(l_effective_cap_len)
+        max_img_len = max(l_effective_img_len)
+
+        position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.int32, device=device)
+
+        for i in range(bsz):
+            cap_len = l_effective_cap_len[i]
+            img_len = l_effective_img_len[i]
+            H, W = img_sizes[i]
+            H_tokens, W_tokens = H // pH, W // pW
+            assert H_tokens * W_tokens == img_len
+
+            position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.int32, device=device)
+            position_ids[i, cap_len:cap_len+img_len, 0] = cap_len
+            row_ids = torch.arange(H_tokens, dtype=torch.int32, device=device).view(-1, 1).repeat(1, W_tokens).flatten()
+            col_ids = torch.arange(W_tokens, dtype=torch.int32, device=device).view(1, -1).repeat(H_tokens, 1).flatten()
+            position_ids[i, cap_len:cap_len+img_len, 1] = row_ids
+            position_ids[i, cap_len:cap_len+img_len, 2] = col_ids
+
+        freqs_cis = self.rope_embedder(position_ids).movedim(1, 2).to(dtype)
+
+        # build freqs_cis for cap and image individually
+        cap_freqs_cis_shape = list(freqs_cis.shape)
+        # cap_freqs_cis_shape[1] = max_cap_len
+        cap_freqs_cis_shape[1] = cap_feats.shape[1]
+        cap_freqs_cis = torch.zeros(*cap_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        img_freqs_cis_shape = list(freqs_cis.shape)
+        img_freqs_cis_shape[1] = max_img_len
+        img_freqs_cis = torch.zeros(*img_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        for i in range(bsz):
+            cap_len = l_effective_cap_len[i]
+            img_len = l_effective_img_len[i]
+            cap_freqs_cis[i, :cap_len] = freqs_cis[i, :cap_len]
+            img_freqs_cis[i, :img_len] = freqs_cis[i, cap_len:cap_len+img_len]
+
+        # refine context
+        for layer in self.context_refiner:
+            cap_feats = layer(cap_feats, cap_mask, cap_freqs_cis)
+
+        # refine image
+        flat_x = []
+        for i in range(bsz):
+            img = x[i]
+            C, H, W = img.size()
+            img = img.view(C, H // pH, pH, W // pW, pW).permute(1, 3, 2, 4, 0).flatten(2).flatten(0, 1)
+            flat_x.append(img)
+        x = flat_x
+        padded_img_embed = torch.zeros(bsz, max_img_len, x[0].shape[-1], device=device, dtype=x[0].dtype)
+        padded_img_mask = torch.zeros(bsz, max_img_len, dtype=dtype, device=device)
+        for i in range(bsz):
+            padded_img_embed[i, :l_effective_img_len[i]] = x[i]
+            padded_img_mask[i, l_effective_img_len[i]:] = -torch.finfo(dtype).max
+
+        padded_img_embed = self.x_embedder(padded_img_embed)
+        padded_img_mask = padded_img_mask.unsqueeze(1)
+        for layer in self.noise_refiner:
+            padded_img_embed = layer(padded_img_embed, padded_img_mask, img_freqs_cis, t)
+
+        if cap_mask is not None:
+            mask = torch.zeros(bsz, max_seq_len, dtype=dtype, device=device)
+            mask[:, :max_cap_len] = cap_mask[:, :max_cap_len]
+        else:
+            mask = None
+
+        padded_full_embed = torch.zeros(bsz, max_seq_len, self.dim, device=device, dtype=x[0].dtype)
+        for i in range(bsz):
+            cap_len = l_effective_cap_len[i]
+            img_len = l_effective_img_len[i]
+
+            padded_full_embed[i, :cap_len] = cap_feats[i, :cap_len]
+            padded_full_embed[i, cap_len:cap_len+img_len] = padded_img_embed[i, :img_len]
+
+        return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis
+
+    def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, kwargs.get("transformer_options", {}))
+        ).execute(x, timesteps, context, num_tokens, attention_mask, **kwargs)
+
+    # def forward(self, x, t, cap_feats, cap_mask):
+    def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
+        t = 1.0 - timesteps
+        cap_feats = context
+        cap_mask = attention_mask
+        bs, c, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+        """
+        Forward pass of NextDiT.
+        t: (N,) tensor of diffusion timesteps
+        y: (N,) tensor of text tokens/features
+        """
+
+        t = self.t_embedder(t, dtype=x.dtype)  # (N, D)
+        adaln_input = t
+
+        cap_feats = self.cap_embedder(cap_feats)  # (N, L, D)  # todo check if able to batchify w.o. redundant compute
+
+        x_is_tensor = isinstance(x, torch.Tensor)
+        x, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens)
+        freqs_cis = freqs_cis.to(x.device)
+
+        for layer in self.layers:
+            x = layer(x, mask, freqs_cis, adaln_input)
+
+        x = self.final_layer(x, adaln_input)
+        x = self.unpatchify(x, img_size, cap_size, return_tensor=x_is_tensor)[:,:,:h,:w]
+
+        return -x
+

comfy/ldm/models/autoencoder.py

+import logging
+import math
+import torch
+from contextlib import contextmanager
+from typing import Any, Dict, Tuple, Union
+
+from comfy.ldm.modules.distributions.distributions import DiagonalGaussianDistribution
+
+from comfy.ldm.util import get_obj_from_str, instantiate_from_config
+from comfy.ldm.modules.ema import LitEma
+import comfy.ops
+
+class DiagonalGaussianRegularizer(torch.nn.Module):
+    def __init__(self, sample: bool = False):
+        super().__init__()
+        self.sample = sample
+
+    def get_trainable_parameters(self) -> Any:
+        yield from ()
+
+    def forward(self, z: torch.Tensor) -> Tuple[torch.Tensor, dict]:
+        posterior = DiagonalGaussianDistribution(z)
+        if self.sample:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        return z, None
+
+
+class AbstractAutoencoder(torch.nn.Module):
+    """
+    This is the base class for all autoencoders, including image autoencoders, image autoencoders with discriminators,
+    unCLIP models, etc. Hence, it is fairly general, and specific features
+    (e.g. discriminator training, encoding, decoding) must be implemented in subclasses.
+    """
+
+    def __init__(
+        self,
+        ema_decay: Union[None, float] = None,
+        monitor: Union[None, str] = None,
+        input_key: str = "jpg",
+        **kwargs,
+    ):
+        super().__init__()
+
+        self.input_key = input_key
+        self.use_ema = ema_decay is not None
+        if monitor is not None:
+            self.monitor = monitor
+
+        if self.use_ema:
+            self.model_ema = LitEma(self, decay=ema_decay)
+            logging.info(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+    def get_input(self, batch) -> Any:
+        raise NotImplementedError()
+
+    def on_train_batch_end(self, *args, **kwargs):
+        # for EMA computation
+        if self.use_ema:
+            self.model_ema(self)
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.parameters())
+            self.model_ema.copy_to(self)
+            if context is not None:
+                logging.info(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.parameters())
+                if context is not None:
+                    logging.info(f"{context}: Restored training weights")
+
+    def encode(self, *args, **kwargs) -> torch.Tensor:
+        raise NotImplementedError("encode()-method of abstract base class called")
+
+    def decode(self, *args, **kwargs) -> torch.Tensor:
+        raise NotImplementedError("decode()-method of abstract base class called")
+
+    def instantiate_optimizer_from_config(self, params, lr, cfg):
+        logging.info(f"loading >>> {cfg['target']} <<< optimizer from config")
+        return get_obj_from_str(cfg["target"])(
+            params, lr=lr, **cfg.get("params", dict())
+        )
+
+    def configure_optimizers(self) -> Any:
+        raise NotImplementedError()
+
+
+class AutoencodingEngine(AbstractAutoencoder):
+    """
+    Base class for all image autoencoders that we train, like VQGAN or AutoencoderKL
+    (we also restore them explicitly as special cases for legacy reasons).
+    Regularizations such as KL or VQ are moved to the regularizer class.
+    """
+
+    def __init__(
+        self,
+        *args,
+        encoder_config: Dict,
+        decoder_config: Dict,
+        regularizer_config: Dict,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+
+        self.encoder: torch.nn.Module = instantiate_from_config(encoder_config)
+        self.decoder: torch.nn.Module = instantiate_from_config(decoder_config)
+        self.regularization = instantiate_from_config(
+            regularizer_config
+        )
+
+    def get_last_layer(self):
+        return self.decoder.get_last_layer()
+
+    def encode(
+        self,
+        x: torch.Tensor,
+        return_reg_log: bool = False,
+        unregularized: bool = False,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, dict]]:
+        z = self.encoder(x)
+        if unregularized:
+            return z, dict()
+        z, reg_log = self.regularization(z)
+        if return_reg_log:
+            return z, reg_log
+        return z
+
+    def decode(self, z: torch.Tensor, **kwargs) -> torch.Tensor:
+        x = self.decoder(z, **kwargs)
+        return x
+
+    def forward(
+        self, x: torch.Tensor, **additional_decode_kwargs
+    ) -> Tuple[torch.Tensor, torch.Tensor, dict]:
+        z, reg_log = self.encode(x, return_reg_log=True)
+        dec = self.decode(z, **additional_decode_kwargs)
+        return z, dec, reg_log
+
+
+class AutoencodingEngineLegacy(AutoencodingEngine):
+    def __init__(self, embed_dim: int, **kwargs):
+        self.max_batch_size = kwargs.pop("max_batch_size", None)
+        ddconfig = kwargs.pop("ddconfig")
+        super().__init__(
+            encoder_config={
+                "target": "comfy.ldm.modules.diffusionmodules.model.Encoder",
+                "params": ddconfig,
+            },
+            decoder_config={
+                "target": "comfy.ldm.modules.diffusionmodules.model.Decoder",
+                "params": ddconfig,
+            },
+            **kwargs,
+        )
+
+        if ddconfig.get("conv3d", False):
+            conv_op = comfy.ops.disable_weight_init.Conv3d
+        else:
+            conv_op = comfy.ops.disable_weight_init.Conv2d
+
+        self.quant_conv = conv_op(
+            (1 + ddconfig["double_z"]) * ddconfig["z_channels"],
+            (1 + ddconfig["double_z"]) * embed_dim,
+            1,
+        )
+
+        self.post_quant_conv = conv_op(embed_dim, ddconfig["z_channels"], 1)
+        self.embed_dim = embed_dim
+
+    def get_autoencoder_params(self) -> list:
+        params = super().get_autoencoder_params()
+        return params
+
+    def encode(
+        self, x: torch.Tensor, return_reg_log: bool = False
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, dict]]:
+        if self.max_batch_size is None:
+            z = self.encoder(x)
+            z = self.quant_conv(z)
+        else:
+            N = x.shape[0]
+            bs = self.max_batch_size
+            n_batches = int(math.ceil(N / bs))
+            z = list()
+            for i_batch in range(n_batches):
+                z_batch = self.encoder(x[i_batch * bs : (i_batch + 1) * bs])
+                z_batch = self.quant_conv(z_batch)
+                z.append(z_batch)
+            z = torch.cat(z, 0)
+
+        z, reg_log = self.regularization(z)
+        if return_reg_log:
+            return z, reg_log
+        return z
+
+    def decode(self, z: torch.Tensor, **decoder_kwargs) -> torch.Tensor:
+        if self.max_batch_size is None:
+            dec = self.post_quant_conv(z)
+            dec = self.decoder(dec, **decoder_kwargs)
+        else:
+            N = z.shape[0]
+            bs = self.max_batch_size
+            n_batches = int(math.ceil(N / bs))
+            dec = list()
+            for i_batch in range(n_batches):
+                dec_batch = self.post_quant_conv(z[i_batch * bs : (i_batch + 1) * bs])
+                dec_batch = self.decoder(dec_batch, **decoder_kwargs)
+                dec.append(dec_batch)
+            dec = torch.cat(dec, 0)
+
+        return dec
+
+
+class AutoencoderKL(AutoencodingEngineLegacy):
+    def __init__(self, **kwargs):
+        if "lossconfig" in kwargs:
+            kwargs["loss_config"] = kwargs.pop("lossconfig")
+        super().__init__(
+            regularizer_config={
+                "target": (
+                    "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"
+                )
+            },
+            **kwargs,
+        )

comfy/ldm/modules/attention.py

+import math
+import sys
+
+import torch
+import torch.nn.functional as F
+from torch import nn, einsum
+from einops import rearrange, repeat
+from typing import Optional
+import logging
+
+from .diffusionmodules.util import AlphaBlender, timestep_embedding
+from .sub_quadratic_attention import efficient_dot_product_attention
+
+from comfy import model_management
+
+if model_management.xformers_enabled():
+    import xformers
+    import xformers.ops
+
+if model_management.sage_attention_enabled():
+    try:
+        from sageattention import sageattn
+    except ModuleNotFoundError as e:
+        if e.name == "sageattention":
+            logging.error(f"\n\nTo use the `--use-sage-attention` feature, the `sageattention` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install sageattention")
+        else:
+            raise e
+        exit(-1)
+
+if model_management.flash_attention_enabled():
+    try:
+        from flash_attn import flash_attn_func
+    except ModuleNotFoundError:
+        logging.error(f"\n\nTo use the `--use-flash-attention` feature, the `flash-attn` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install flash-attn")
+        exit(-1)
+
+from comfy.cli_args import args
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+FORCE_UPCAST_ATTENTION_DTYPE = model_management.force_upcast_attention_dtype()
+
+def get_attn_precision(attn_precision, current_dtype):
+    if args.dont_upcast_attention:
+        return None
+
+    if FORCE_UPCAST_ATTENTION_DTYPE is not None and current_dtype in FORCE_UPCAST_ATTENTION_DTYPE:
+        return FORCE_UPCAST_ATTENTION_DTYPE[current_dtype]
+    return attn_precision
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out, dtype=None, device=None, operations=ops):
+        super().__init__()
+        self.proj = operations.Linear(dim_in, dim_out * 2, dtype=dtype, device=device)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=ops):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            operations.Linear(dim, inner_dim, dtype=dtype, device=device),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim, dtype=dtype, device=device, operations=operations)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            operations.Linear(inner_dim, dim_out, dtype=dtype, device=device)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+def Normalize(in_channels, dtype=None, device=None):
+    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
+
+def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+    attn_precision = get_attn_precision(attn_precision, q.dtype)
+
+    if skip_reshape:
+        b, _, _, dim_head = q.shape
+    else:
+        b, _, dim_head = q.shape
+        dim_head //= heads
+
+    scale = dim_head ** -0.5
+
+    h = heads
+    if skip_reshape:
+         q, k, v = map(
+            lambda t: t.reshape(b * heads, -1, dim_head),
+            (q, k, v),
+        )
+    else:
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(b, -1, heads, dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b * heads, -1, dim_head)
+            .contiguous(),
+            (q, k, v),
+        )
+
+    # force cast to fp32 to avoid overflowing
+    if attn_precision == torch.float32:
+        sim = einsum('b i d, b j d -> b i j', q.float(), k.float()) * scale
+    else:
+        sim = einsum('b i d, b j d -> b i j', q, k) * scale
+
+    del q, k
+
+    if exists(mask):
+        if mask.dtype == torch.bool:
+            mask = rearrange(mask, 'b ... -> b (...)') #TODO: check if this bool part matches pytorch attention
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b j -> (b h) () j', h=h)
+            sim.masked_fill_(~mask, max_neg_value)
+        else:
+            if len(mask.shape) == 2:
+                bs = 1
+            else:
+                bs = mask.shape[0]
+            mask = mask.reshape(bs, -1, mask.shape[-2], mask.shape[-1]).expand(b, heads, -1, -1).reshape(-1, mask.shape[-2], mask.shape[-1])
+            sim.add_(mask)
+
+    # attention, what we cannot get enough of
+    sim = sim.softmax(dim=-1)
+
+    out = einsum('b i j, b j d -> b i d', sim.to(v.dtype), v)
+
+    if skip_output_reshape:
+        out = (
+            out.unsqueeze(0)
+            .reshape(b, heads, -1, dim_head)
+        )
+    else:
+        out = (
+            out.unsqueeze(0)
+            .reshape(b, heads, -1, dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b, -1, heads * dim_head)
+        )
+    return out
+
+
+def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+    attn_precision = get_attn_precision(attn_precision, query.dtype)
+
+    if skip_reshape:
+        b, _, _, dim_head = query.shape
+    else:
+        b, _, dim_head = query.shape
+        dim_head //= heads
+
+    if skip_reshape:
+        query = query.reshape(b * heads, -1, dim_head)
+        value = value.reshape(b * heads, -1, dim_head)
+        key = key.reshape(b * heads, -1, dim_head).movedim(1, 2)
+    else:
+        query = query.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 1, 3).reshape(b * heads, -1, dim_head)
+        value = value.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 1, 3).reshape(b * heads, -1, dim_head)
+        key = key.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 3, 1).reshape(b * heads, dim_head, -1)
+
+
+    dtype = query.dtype
+    upcast_attention = attn_precision == torch.float32 and query.dtype != torch.float32
+    if upcast_attention:
+        bytes_per_token = torch.finfo(torch.float32).bits//8
+    else:
+        bytes_per_token = torch.finfo(query.dtype).bits//8
+    batch_x_heads, q_tokens, _ = query.shape
+    _, _, k_tokens = key.shape
+
+    mem_free_total, _ = model_management.get_free_memory(query.device, True)
+
+    kv_chunk_size_min = None
+    kv_chunk_size = None
+    query_chunk_size = None
+
+    for x in [4096, 2048, 1024, 512, 256]:
+        count = mem_free_total / (batch_x_heads * bytes_per_token * x * 4.0)
+        if count >= k_tokens:
+            kv_chunk_size = k_tokens
+            query_chunk_size = x
+            break
+
+    if query_chunk_size is None:
+        query_chunk_size = 512
+
+    if mask is not None:
+        if len(mask.shape) == 2:
+            bs = 1
+        else:
+            bs = mask.shape[0]
+        mask = mask.reshape(bs, -1, mask.shape[-2], mask.shape[-1]).expand(b, heads, -1, -1).reshape(-1, mask.shape[-2], mask.shape[-1])
+
+    hidden_states = efficient_dot_product_attention(
+        query,
+        key,
+        value,
+        query_chunk_size=query_chunk_size,
+        kv_chunk_size=kv_chunk_size,
+        kv_chunk_size_min=kv_chunk_size_min,
+        use_checkpoint=False,
+        upcast_attention=upcast_attention,
+        mask=mask,
+    )
+
+    hidden_states = hidden_states.to(dtype)
+    if skip_output_reshape:
+        hidden_states = hidden_states.unflatten(0, (-1, heads))
+    else:
+        hidden_states = hidden_states.unflatten(0, (-1, heads)).transpose(1,2).flatten(start_dim=2)
+    return hidden_states
+
+def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+    attn_precision = get_attn_precision(attn_precision, q.dtype)
+
+    if skip_reshape:
+        b, _, _, dim_head = q.shape
+    else:
+        b, _, dim_head = q.shape
+        dim_head //= heads
+
+    scale = dim_head ** -0.5
+
+    if skip_reshape:
+         q, k, v = map(
+            lambda t: t.reshape(b * heads, -1, dim_head),
+            (q, k, v),
+        )
+    else:
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(b, -1, heads, dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b * heads, -1, dim_head)
+            .contiguous(),
+            (q, k, v),
+        )
+
+    r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
+
+    mem_free_total = model_management.get_free_memory(q.device)
+
+    if attn_precision == torch.float32:
+        element_size = 4
+        upcast = True
+    else:
+        element_size = q.element_size()
+        upcast = False
+
+    gb = 1024 ** 3
+    tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * element_size
+    modifier = 3
+    mem_required = tensor_size * modifier
+    steps = 1
+
+
+    if mem_required > mem_free_total:
+        steps = 2**(math.ceil(math.log(mem_required / mem_free_total, 2)))
+        # print(f"Expected tensor size:{tensor_size/gb:0.1f}GB, cuda free:{mem_free_cuda/gb:0.1f}GB "
+        #      f"torch free:{mem_free_torch/gb:0.1f} total:{mem_free_total/gb:0.1f} steps:{steps}")
+
+    if steps > 64:
+        max_res = math.floor(math.sqrt(math.sqrt(mem_free_total / 2.5)) / 8) * 64
+        raise RuntimeError(f'Not enough memory, use lower resolution (max approx. {max_res}x{max_res}). '
+                            f'Need: {mem_required/64/gb:0.1f}GB free, Have:{mem_free_total/gb:0.1f}GB free')
+
+    if mask is not None:
+        if len(mask.shape) == 2:
+            bs = 1
+        else:
+            bs = mask.shape[0]
+        mask = mask.reshape(bs, -1, mask.shape[-2], mask.shape[-1]).expand(b, heads, -1, -1).reshape(-1, mask.shape[-2], mask.shape[-1])
+
+    # print("steps", steps, mem_required, mem_free_total, modifier, q.element_size(), tensor_size)
+    first_op_done = False
+    cleared_cache = False
+    while True:
+        try:
+            slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1]
+            for i in range(0, q.shape[1], slice_size):
+                end = i + slice_size
+                if upcast:
+                    with torch.autocast(enabled=False, device_type = 'cuda'):
+                        s1 = einsum('b i d, b j d -> b i j', q[:, i:end].float(), k.float()) * scale
+                else:
+                    s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k) * scale
+
+                if mask is not None:
+                    if len(mask.shape) == 2:
+                        s1 += mask[i:end]
+                    else:
+                        if mask.shape[1] == 1:
+                            s1 += mask
+                        else:
+                            s1 += mask[:, i:end]
+
+                s2 = s1.softmax(dim=-1).to(v.dtype)
+                del s1
+                first_op_done = True
+
+                r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
+                del s2
+            break
+        except model_management.OOM_EXCEPTION as e:
+            if first_op_done == False:
+                model_management.soft_empty_cache(True)
+                if cleared_cache == False:
+                    cleared_cache = True
+                    logging.warning("out of memory error, emptying cache and trying again")
+                    continue
+                steps *= 2
+                if steps > 64:
+                    raise e
+                logging.warning("out of memory error, increasing steps and trying again {}".format(steps))
+            else:
+                raise e
+
+    del q, k, v
+
+    if skip_output_reshape:
+        r1 = (
+            r1.unsqueeze(0)
+            .reshape(b, heads, -1, dim_head)
+        )
+    else:
+        r1 = (
+            r1.unsqueeze(0)
+            .reshape(b, heads, -1, dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b, -1, heads * dim_head)
+        )
+    return r1
+
+BROKEN_XFORMERS = False
+try:
+    x_vers = xformers.__version__
+    # XFormers bug confirmed on all versions from 0.0.21 to 0.0.26 (q with bs bigger than 65535 gives CUDA error)
+    BROKEN_XFORMERS = x_vers.startswith("0.0.2") and not x_vers.startswith("0.0.20")
+except:
+    pass
+
+def attention_xformers(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+    b = q.shape[0]
+    dim_head = q.shape[-1]
+    # check to make sure xformers isn't broken
+    disabled_xformers = False
+
+    if BROKEN_XFORMERS:
+        if b * heads > 65535:
+            disabled_xformers = True
+
+    if not disabled_xformers:
+        if torch.jit.is_tracing() or torch.jit.is_scripting():
+            disabled_xformers = True
+
+    if disabled_xformers:
+        return attention_pytorch(q, k, v, heads, mask, skip_reshape=skip_reshape)
+
+    if skip_reshape:
+        # b h k d -> b k h d
+        q, k, v = map(
+            lambda t: t.permute(0, 2, 1, 3),
+            (q, k, v),
+        )
+    # actually do the reshaping
+    else:
+        dim_head //= heads
+        q, k, v = map(
+            lambda t: t.reshape(b, -1, heads, dim_head),
+            (q, k, v),
+        )
+
+    if mask is not None:
+        # add a singleton batch dimension
+        if mask.ndim == 2:
+            mask = mask.unsqueeze(0)
+        # add a singleton heads dimension
+        if mask.ndim == 3:
+            mask = mask.unsqueeze(1)
+        # pad to a multiple of 8
+        pad = 8 - mask.shape[-1] % 8
+        # the xformers docs says that it's allowed to have a mask of shape (1, Nq, Nk)
+        # but when using separated heads, the shape has to be (B, H, Nq, Nk)
+        # in flux, this matrix ends up being over 1GB
+        # here, we create a mask with the same batch/head size as the input mask (potentially singleton or full)
+        mask_out = torch.empty([mask.shape[0], mask.shape[1], q.shape[1], mask.shape[-1] + pad], dtype=q.dtype, device=q.device)
+
+        mask_out[..., :mask.shape[-1]] = mask
+        # doesn't this remove the padding again??
+        mask = mask_out[..., :mask.shape[-1]]
+        mask = mask.expand(b, heads, -1, -1)
+
+    out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=mask)
+
+    if skip_output_reshape:
+        out = out.permute(0, 2, 1, 3)
+    else:
+        out = (
+            out.reshape(b, -1, heads * dim_head)
+        )
+
+    return out
+
+if model_management.is_nvidia(): #pytorch 2.3 and up seem to have this issue.
+    SDP_BATCH_LIMIT = 2**15
+else:
+    #TODO: other GPUs ?
+    SDP_BATCH_LIMIT = 2**31
+
+
+def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+    
+    if skip_reshape:
+        b, _, _, dim_head = q.shape
+        # print("11111111111111", b, dim_head, q.shape)
+    else:
+        b, _, dim_head = q.shape
+        dim_head //= heads
+        q, k, v = map(
+            lambda t: t.view(b, -1, heads, dim_head).transpose(1, 2),
+            (q, k, v),
+        )
+        # print("22222222222222", b, dim_head, heads, q.shape, k.shape, v.shape)
+
+    if mask is not None:
+        # print("33333333333333")
+        # add a batch dimension if there isn't already one
+        if mask.ndim == 2:
+            mask = mask.unsqueeze(0)
+        # add a heads dimension if there isn't already one
+        if mask.ndim == 3:
+            mask = mask.unsqueeze(1)
+    
+    if SDP_BATCH_LIMIT >= b:
+        # print("44444444444444", SDP_BATCH_LIMIT)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+        if not skip_output_reshape:
+            out = (
+                out.transpose(1, 2).reshape(b, -1, heads * dim_head)
+            )
+    else:
+        # print("55555555555555", SDP_BATCH_LIMIT)
+        out = torch.empty((b, q.shape[2], heads * dim_head), dtype=q.dtype, layout=q.layout, device=q.device)
+        for i in range(0, b, SDP_BATCH_LIMIT):
+            m = mask
+            if mask is not None:
+                if mask.shape[0] > 1:
+                    m = mask[i : i + SDP_BATCH_LIMIT]
+
+            out[i : i + SDP_BATCH_LIMIT] = comfy.ops.scaled_dot_product_attention(
+                q[i : i + SDP_BATCH_LIMIT],
+                k[i : i + SDP_BATCH_LIMIT],
+                v[i : i + SDP_BATCH_LIMIT],
+                attn_mask=m,
+                dropout_p=0.0, is_causal=False
+            ).transpose(1, 2).reshape(-1, q.shape[2], heads * dim_head)
+    return out
+
+
+def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+    if skip_reshape:
+        b, _, _, dim_head = q.shape
+        tensor_layout = "HND"
+    else:
+        b, _, dim_head = q.shape
+        dim_head //= heads
+        q, k, v = map(
+            lambda t: t.view(b, -1, heads, dim_head),
+            (q, k, v),
+        )
+        tensor_layout = "NHD"
+
+    if mask is not None:
+        # add a batch dimension if there isn't already one
+        if mask.ndim == 2:
+            mask = mask.unsqueeze(0)
+        # add a heads dimension if there isn't already one
+        if mask.ndim == 3:
+            mask = mask.unsqueeze(1)
+
+    try:
+        out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout)
+    except Exception as e:
+        logging.error("Error running sage attention: {}, using pytorch attention instead.".format(e))
+        if tensor_layout == "NHD":
+            q, k, v = map(
+                lambda t: t.transpose(1, 2),
+                (q, k, v),
+            )
+        return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=True, skip_output_reshape=skip_output_reshape)
+
+    if tensor_layout == "HND":
+        if not skip_output_reshape:
+            out = (
+                out.transpose(1, 2).reshape(b, -1, heads * dim_head)
+            )
+    else:
+        if skip_output_reshape:
+            out = out.transpose(1, 2)
+        else:
+            out = out.reshape(b, -1, heads * dim_head)
+    return out
+
+
+try:
+    @torch.library.custom_op("flash_attention::flash_attn", mutates_args=())
+    def flash_attn_wrapper(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
+                    dropout_p: float = 0.0, causal: bool = False) -> torch.Tensor:
+        return flash_attn_func(q, k, v, dropout_p=dropout_p, causal=causal)
+
+
+    @flash_attn_wrapper.register_fake
+    def flash_attn_fake(q, k, v, dropout_p=0.0, causal=False):
+        # Output shape is the same as q
+        return q.new_empty(q.shape)
+except AttributeError as error:
+    FLASH_ATTN_ERROR = error
+
+    def flash_attn_wrapper(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
+                    dropout_p: float = 0.0, causal: bool = False) -> torch.Tensor:
+        assert False, f"Could not define flash_attn_wrapper: {FLASH_ATTN_ERROR}"
+
+
+def attention_flash(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+    # import pdb
+    # pdb.set_trace()
+    if skip_reshape:
+        b, _, _, dim_head = q.shape
+    else:
+        b, _, dim_head = q.shape
+        dim_head //= heads
+        q, k, v = map(
+            lambda t: t.view(b, -1, heads, dim_head).transpose(1, 2),
+            (q, k, v),
+        )
+
+    if mask is not None:
+        # add a batch dimension if there isn't already one
+        if mask.ndim == 2:
+            mask = mask.unsqueeze(0)
+        # add a heads dimension if there isn't already one
+        if mask.ndim == 3:
+            mask = mask.unsqueeze(1)
+
+    try:
+        assert mask is None
+        out = flash_attn_wrapper(
+            q.transpose(1, 2),
+            k.transpose(1, 2),
+            v.transpose(1, 2),
+            dropout_p=0.0,
+            causal=False,
+        ).transpose(1, 2)
+    except Exception as e:
+        logging.warning(f"Flash Attention failed, using default SDPA: {e}")
+        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+    if not skip_output_reshape:
+        out = (
+            out.transpose(1, 2).reshape(b, -1, heads * dim_head)
+        )
+    return out
+
+
+optimized_attention = attention_basic
+
+if model_management.sage_attention_enabled():
+    logging.info("Using sage attention")
+    optimized_attention = attention_sage
+elif model_management.xformers_enabled():
+    logging.info("Using xformers attention")
+    optimized_attention = attention_xformers
+elif model_management.flash_attention_enabled():
+    logging.info("Using Flash Attention")
+    optimized_attention = attention_flash
+elif model_management.pytorch_attention_enabled():
+    logging.info("Using pytorch attention")
+    optimized_attention = attention_pytorch
+else:
+    if args.use_split_cross_attention:
+        logging.info("Using split optimization for attention")
+        optimized_attention = attention_split
+    else:
+        logging.info("Using sub quadratic optimization for attention, if you have memory or speed issues try using: --use-split-cross-attention")
+        optimized_attention = attention_sub_quad
+
+optimized_attention_masked = optimized_attention
+
+def optimized_attention_for_device(device, mask=False, small_input=False):
+    if small_input:
+        if model_management.pytorch_attention_enabled():
+            return attention_pytorch #TODO: need to confirm but this is probably slightly faster for small inputs in all cases
+        else:
+            return attention_basic
+
+    if device == torch.device("cpu"):
+        return attention_sub_quad
+
+    if mask:
+        return optimized_attention_masked
+
+    return optimized_attention
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., attn_precision=None, dtype=None, device=None, operations=ops):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+        self.attn_precision = attn_precision
+
+        self.heads = heads
+        self.dim_head = dim_head
+
+        self.to_q = operations.Linear(query_dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.to_k = operations.Linear(context_dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.to_v = operations.Linear(context_dim, inner_dim, bias=False, dtype=dtype, device=device)
+
+        self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout))
+
+    def forward(self, x, context=None, value=None, mask=None):
+        q = self.to_q(x)
+        context = default(context, x)
+        k = self.to_k(context)
+        if value is not None:
+            v = self.to_v(value)
+            del value
+        else:
+            v = self.to_v(context)
+
+        if mask is None:
+            out = optimized_attention(q, k, v, self.heads, attn_precision=self.attn_precision)
+        else:
+            out = optimized_attention_masked(q, k, v, self.heads, mask, attn_precision=self.attn_precision)
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, ff_in=False, inner_dim=None,
+                 disable_self_attn=False, disable_temporal_crossattention=False, switch_temporal_ca_to_sa=False, attn_precision=None, dtype=None, device=None, operations=ops):
+        super().__init__()
+
+        self.ff_in = ff_in or inner_dim is not None
+        if inner_dim is None:
+            inner_dim = dim
+
+        self.is_res = inner_dim == dim
+        self.attn_precision = attn_precision
+
+        if self.ff_in:
+            self.norm_in = operations.LayerNorm(dim, dtype=dtype, device=device)
+            self.ff_in = FeedForward(dim, dim_out=inner_dim, dropout=dropout, glu=gated_ff, dtype=dtype, device=device, operations=operations)
+
+        self.disable_self_attn = disable_self_attn
+        self.attn1 = CrossAttention(query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout,
+                              context_dim=context_dim if self.disable_self_attn else None, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations)  # is a self-attention if not self.disable_self_attn
+        self.ff = FeedForward(inner_dim, dim_out=dim, dropout=dropout, glu=gated_ff, dtype=dtype, device=device, operations=operations)
+
+        if disable_temporal_crossattention:
+            if switch_temporal_ca_to_sa:
+                raise ValueError
+            else:
+                self.attn2 = None
+        else:
+            context_dim_attn2 = None
+            if not switch_temporal_ca_to_sa:
+                context_dim_attn2 = context_dim
+
+            self.attn2 = CrossAttention(query_dim=inner_dim, context_dim=context_dim_attn2,
+                                heads=n_heads, dim_head=d_head, dropout=dropout, attn_precision=self.attn_precision, dtype=dtype, device=device, operations=operations)  # is self-attn if context is none
+            self.norm2 = operations.LayerNorm(inner_dim, dtype=dtype, device=device)
+
+        self.norm1 = operations.LayerNorm(inner_dim, dtype=dtype, device=device)
+        self.norm3 = operations.LayerNorm(inner_dim, dtype=dtype, device=device)
+        self.n_heads = n_heads
+        self.d_head = d_head
+        self.switch_temporal_ca_to_sa = switch_temporal_ca_to_sa
+
+    def forward(self, x, context=None, transformer_options={}):
+        # logging.info("=====================111111111111111111111=============================")
+        extra_options = {}
+        block = transformer_options.get("block", None)
+        block_index = transformer_options.get("block_index", 0)
+        transformer_patches = {}
+        transformer_patches_replace = {}
+
+        for k in transformer_options:
+            if k == "patches":
+                transformer_patches = transformer_options[k]
+            elif k == "patches_replace":
+                transformer_patches_replace = transformer_options[k]
+            else:
+                extra_options[k] = transformer_options[k]
+
+        extra_options["n_heads"] = self.n_heads
+        extra_options["dim_head"] = self.d_head
+        extra_options["attn_precision"] = self.attn_precision
+
+        if self.ff_in:
+            x_skip = x
+            x = self.ff_in(self.norm_in(x))
+            if self.is_res:
+                x += x_skip
+
+        n = self.norm1(x)
+        if self.disable_self_attn:
+            context_attn1 = context
+        else:
+            context_attn1 = None
+        value_attn1 = None
+
+        if "attn1_patch" in transformer_patches:
+            patch = transformer_patches["attn1_patch"]
+            if context_attn1 is None:
+                context_attn1 = n
+            value_attn1 = context_attn1
+            for p in patch:
+                n, context_attn1, value_attn1 = p(n, context_attn1, value_attn1, extra_options)
+
+        if block is not None:
+            transformer_block = (block[0], block[1], block_index)
+        else:
+            transformer_block = None
+        attn1_replace_patch = transformer_patches_replace.get("attn1", {})
+        block_attn1 = transformer_block
+        if block_attn1 not in attn1_replace_patch:
+            block_attn1 = block
+
+        if block_attn1 in attn1_replace_patch:
+            if context_attn1 is None:
+                context_attn1 = n
+                value_attn1 = n
+            n = self.attn1.to_q(n)
+            context_attn1 = self.attn1.to_k(context_attn1)
+            value_attn1 = self.attn1.to_v(value_attn1)
+            n = attn1_replace_patch[block_attn1](n, context_attn1, value_attn1, extra_options)
+            n = self.attn1.to_out(n)
+        else:
+            n = self.attn1(n, context=context_attn1, value=value_attn1)
+
+        if "attn1_output_patch" in transformer_patches:
+            patch = transformer_patches["attn1_output_patch"]
+            for p in patch:
+                n = p(n, extra_options)
+
+        x = n + x
+        if "middle_patch" in transformer_patches:
+            patch = transformer_patches["middle_patch"]
+            for p in patch:
+                x = p(x, extra_options)
+
+        if self.attn2 is not None:
+            n = self.norm2(x)
+            if self.switch_temporal_ca_to_sa:
+                context_attn2 = n
+            else:
+                context_attn2 = context
+            value_attn2 = None
+            if "attn2_patch" in transformer_patches:
+                patch = transformer_patches["attn2_patch"]
+                value_attn2 = context_attn2
+                for p in patch:
+                    n, context_attn2, value_attn2 = p(n, context_attn2, value_attn2, extra_options)
+
+            attn2_replace_patch = transformer_patches_replace.get("attn2", {})
+            block_attn2 = transformer_block
+            if block_attn2 not in attn2_replace_patch:
+                block_attn2 = block
+
+            if block_attn2 in attn2_replace_patch:
+                if value_attn2 is None:
+                    value_attn2 = context_attn2
+                n = self.attn2.to_q(n)
+                context_attn2 = self.attn2.to_k(context_attn2)
+                value_attn2 = self.attn2.to_v(value_attn2)
+                n = attn2_replace_patch[block_attn2](n, context_attn2, value_attn2, extra_options)
+                n = self.attn2.to_out(n)
+            else:
+                n = self.attn2(n, context=context_attn2, value=value_attn2)
+
+        if "attn2_output_patch" in transformer_patches:
+            patch = transformer_patches["attn2_output_patch"]
+            for p in patch:
+                n = p(n, extra_options)
+
+        x = n + x
+        if self.is_res:
+            x_skip = x
+        x = self.ff(self.norm3(x))
+        if self.is_res:
+            x = x_skip + x
+
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    NEW: use_linear for more efficiency instead of the 1x1 convs
+    """
+    def __init__(self, in_channels, n_heads, d_head,
+                 depth=1, dropout=0., context_dim=None,
+                 disable_self_attn=False, use_linear=False,
+                 use_checkpoint=True, attn_precision=None, dtype=None, device=None, operations=ops):
+        super().__init__()
+        if exists(context_dim) and not isinstance(context_dim, list):
+            context_dim = [context_dim] * depth
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = operations.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
+        if not use_linear:
+            self.proj_in = operations.Conv2d(in_channels,
+                                     inner_dim,
+                                     kernel_size=1,
+                                     stride=1,
+                                     padding=0, dtype=dtype, device=device)
+        else:
+            self.proj_in = operations.Linear(in_channels, inner_dim, dtype=dtype, device=device)
+
+        self.transformer_blocks = nn.ModuleList(
+            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d],
+                                   disable_self_attn=disable_self_attn, checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=dtype, device=device, operations=operations)
+                for d in range(depth)]
+        )
+        if not use_linear:
+            self.proj_out = operations.Conv2d(inner_dim,in_channels,
+                                                  kernel_size=1,
+                                                  stride=1,
+                                                  padding=0, dtype=dtype, device=device)
+        else:
+            self.proj_out = operations.Linear(in_channels, inner_dim, dtype=dtype, device=device)
+        self.use_linear = use_linear
+
+    def forward(self, x, context=None, transformer_options={}):
+        # note: if no context is given, cross-attention defaults to self-attention
+        if not isinstance(context, list):
+            context = [context] * len(self.transformer_blocks)
+        b, c, h, w = x.shape
+        transformer_options["activations_shape"] = list(x.shape)
+        x_in = x
+        x = self.norm(x)
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = x.movedim(1, 3).flatten(1, 2).contiguous()
+        if self.use_linear:
+            x = self.proj_in(x)
+        for i, block in enumerate(self.transformer_blocks):
+            transformer_options["block_index"] = i
+            x = block(x, context=context[i], transformer_options=transformer_options)
+        if self.use_linear:
+            x = self.proj_out(x)
+        x = x.reshape(x.shape[0], h, w, x.shape[-1]).movedim(3, 1).contiguous()
+        if not self.use_linear:
+            x = self.proj_out(x)
+        return x + x_in
+
+
+class SpatialVideoTransformer(SpatialTransformer):
+    def __init__(
+        self,
+        in_channels,
+        n_heads,
+        d_head,
+        depth=1,
+        dropout=0.0,
+        use_linear=False,
+        context_dim=None,
+        use_spatial_context=False,
+        timesteps=None,
+        merge_strategy: str = "fixed",
+        merge_factor: float = 0.5,
+        time_context_dim=None,
+        ff_in=False,
+        checkpoint=False,
+        time_depth=1,
+        disable_self_attn=False,
+        disable_temporal_crossattention=False,
+        max_time_embed_period: int = 10000,
+        attn_precision=None,
+        dtype=None, device=None, operations=ops
+    ):
+        super().__init__(
+            in_channels,
+            n_heads,
+            d_head,
+            depth=depth,
+            dropout=dropout,
+            use_checkpoint=checkpoint,
+            context_dim=context_dim,
+            use_linear=use_linear,
+            disable_self_attn=disable_self_attn,
+            attn_precision=attn_precision,
+            dtype=dtype, device=device, operations=operations
+        )
+        self.time_depth = time_depth
+        self.depth = depth
+        self.max_time_embed_period = max_time_embed_period
+
+        time_mix_d_head = d_head
+        n_time_mix_heads = n_heads
+
+        time_mix_inner_dim = int(time_mix_d_head * n_time_mix_heads)
+
+        inner_dim = n_heads * d_head
+        if use_spatial_context:
+            time_context_dim = context_dim
+
+        self.time_stack = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    n_time_mix_heads,
+                    time_mix_d_head,
+                    dropout=dropout,
+                    context_dim=time_context_dim,
+                    # timesteps=timesteps,
+                    checkpoint=checkpoint,
+                    ff_in=ff_in,
+                    inner_dim=time_mix_inner_dim,
+                    disable_self_attn=disable_self_attn,
+                    disable_temporal_crossattention=disable_temporal_crossattention,
+                    attn_precision=attn_precision,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(self.depth)
+            ]
+        )
+
+        assert len(self.time_stack) == len(self.transformer_blocks)
+
+        self.use_spatial_context = use_spatial_context
+        self.in_channels = in_channels
+
+        time_embed_dim = self.in_channels * 4
+        self.time_pos_embed = nn.Sequential(
+            operations.Linear(self.in_channels, time_embed_dim, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(time_embed_dim, self.in_channels, dtype=dtype, device=device),
+        )
+
+        self.time_mixer = AlphaBlender(
+            alpha=merge_factor, merge_strategy=merge_strategy
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        context: Optional[torch.Tensor] = None,
+        time_context: Optional[torch.Tensor] = None,
+        timesteps: Optional[int] = None,
+        image_only_indicator: Optional[torch.Tensor] = None,
+        transformer_options={}
+    ) -> torch.Tensor:
+        _, _, h, w = x.shape
+        transformer_options["activations_shape"] = list(x.shape)
+        x_in = x
+        spatial_context = None
+        if exists(context):
+            spatial_context = context
+
+        if self.use_spatial_context:
+            assert (
+                context.ndim == 3
+            ), f"n dims of spatial context should be 3 but are {context.ndim}"
+
+            if time_context is None:
+                time_context = context
+            time_context_first_timestep = time_context[::timesteps]
+            time_context = repeat(
+                time_context_first_timestep, "b ... -> (b n) ...", n=h * w
+            )
+        elif time_context is not None and not self.use_spatial_context:
+            time_context = repeat(time_context, "b ... -> (b n) ...", n=h * w)
+            if time_context.ndim == 2:
+                time_context = rearrange(time_context, "b c -> b 1 c")
+
+        x = self.norm(x)
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = rearrange(x, "b c h w -> b (h w) c")
+        if self.use_linear:
+            x = self.proj_in(x)
+
+        num_frames = torch.arange(timesteps, device=x.device)
+        num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps)
+        num_frames = rearrange(num_frames, "b t -> (b t)")
+        t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False, max_period=self.max_time_embed_period).to(x.dtype)
+        emb = self.time_pos_embed(t_emb)
+        emb = emb[:, None, :]
+
+        for it_, (block, mix_block) in enumerate(
+            zip(self.transformer_blocks, self.time_stack)
+        ):
+            transformer_options["block_index"] = it_
+            x = block(
+                x,
+                context=spatial_context,
+                transformer_options=transformer_options,
+            )
+
+            x_mix = x
+            x_mix = x_mix + emb
+
+            B, S, C = x_mix.shape
+            x_mix = rearrange(x_mix, "(b t) s c -> (b s) t c", t=timesteps)
+            x_mix = mix_block(x_mix, context=time_context) #TODO: transformer_options
+            x_mix = rearrange(
+                x_mix, "(b s) t c -> (b t) s c", s=S, b=B // timesteps, c=C, t=timesteps
+            )
+
+            x = self.time_mixer(x_spatial=x, x_temporal=x_mix, image_only_indicator=image_only_indicator)
+
+        if self.use_linear:
+            x = self.proj_out(x)
+        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
+        if not self.use_linear:
+            x = self.proj_out(x)
+        out = x + x_in
+        return out
+
+

comfy/ldm/modules/diffusionmodules/mmdit.py

+from functools import partial
+from typing import Dict, Optional, List
+
+import numpy as np
+import torch
+import torch.nn as nn
+from ..attention import optimized_attention
+from einops import rearrange, repeat
+from .util import timestep_embedding
+import comfy.ops
+import comfy.ldm.common_dit
+
+def default(x, y):
+    if x is not None:
+        return x
+    return y
+
+class Mlp(nn.Module):
+    """ MLP as used in Vision Transformer, MLP-Mixer and related networks
+    """
+    def __init__(
+            self,
+            in_features,
+            hidden_features=None,
+            out_features=None,
+            act_layer=nn.GELU,
+            norm_layer=None,
+            bias=True,
+            drop=0.,
+            use_conv=False,
+            dtype=None,
+            device=None,
+            operations=None,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        drop_probs = drop
+        linear_layer = partial(operations.Conv2d, kernel_size=1) if use_conv else operations.Linear
+
+        self.fc1 = linear_layer(in_features, hidden_features, bias=bias, dtype=dtype, device=device)
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs)
+        self.norm = norm_layer(hidden_features) if norm_layer is not None else nn.Identity()
+        self.fc2 = linear_layer(hidden_features, out_features, bias=bias, dtype=dtype, device=device)
+        self.drop2 = nn.Dropout(drop_probs)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop1(x)
+        x = self.norm(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x
+
+class PatchEmbed(nn.Module):
+    """ 2D Image to Patch Embedding
+    """
+    dynamic_img_pad: torch.jit.Final[bool]
+
+    def __init__(
+            self,
+            img_size: Optional[int] = 224,
+            patch_size: int = 16,
+            in_chans: int = 3,
+            embed_dim: int = 768,
+            norm_layer = None,
+            flatten: bool = True,
+            bias: bool = True,
+            strict_img_size: bool = True,
+            dynamic_img_pad: bool = True,
+            padding_mode='circular',
+            conv3d=False,
+            dtype=None,
+            device=None,
+            operations=None,
+    ):
+        super().__init__()
+        try:
+            len(patch_size)
+            self.patch_size = patch_size
+        except:
+            if conv3d:
+                self.patch_size = (patch_size, patch_size, patch_size)
+            else:
+                self.patch_size = (patch_size, patch_size)
+        self.padding_mode = padding_mode
+
+        # flatten spatial dim and transpose to channels last, kept for bwd compat
+        self.flatten = flatten
+        self.strict_img_size = strict_img_size
+        self.dynamic_img_pad = dynamic_img_pad
+        if conv3d:
+            self.proj = operations.Conv3d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias, dtype=dtype, device=device)
+        else:
+            self.proj = operations.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias, dtype=dtype, device=device)
+        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
+
+    def forward(self, x):
+        if self.dynamic_img_pad:
+            x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size, padding_mode=self.padding_mode)
+        x = self.proj(x)
+        if self.flatten:
+            x = x.flatten(2).transpose(1, 2)  # NCHW -> NLC
+        x = self.norm(x)
+        return x
+
+def modulate(x, shift, scale):
+    if shift is None:
+        shift = torch.zeros_like(scale)
+    return torch.addcmul(shift.unsqueeze(1), x, 1+ scale.unsqueeze(1))
+
+
+#################################################################################
+#                   Sine/Cosine Positional Embedding Functions                  #
+#################################################################################
+
+
+def get_2d_sincos_pos_embed(
+    embed_dim,
+    grid_size,
+    cls_token=False,
+    extra_tokens=0,
+    scaling_factor=None,
+    offset=None,
+):
+    """
+    grid_size: int of the grid height and width
+    return:
+    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+    """
+    grid_h = np.arange(grid_size, dtype=np.float32)
+    grid_w = np.arange(grid_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+    if scaling_factor is not None:
+        grid = grid / scaling_factor
+    if offset is not None:
+        grid = grid - offset
+
+    grid = grid.reshape([2, 1, grid_size, grid_size])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if cls_token and extra_tokens > 0:
+        pos_embed = np.concatenate(
+            [np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0
+        )
+    return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (M,)
+    out: (M, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float64)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out)  # (M, D/2)
+    emb_cos = np.cos(out)  # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+def get_1d_sincos_pos_embed_from_grid_torch(embed_dim, pos, device=None, dtype=torch.float32):
+    omega = torch.arange(embed_dim // 2, device=device, dtype=dtype)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D/2,)
+    pos = pos.reshape(-1)  # (M,)
+    out = torch.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
+    emb = torch.cat([emb_sin, emb_cos], dim=1)  # (M, D)
+    return emb
+
+def get_2d_sincos_pos_embed_torch(embed_dim, w, h, val_center=7.5, val_magnitude=7.5, device=None, dtype=torch.float32):
+    small = min(h, w)
+    val_h = (h / small) * val_magnitude
+    val_w = (w / small) * val_magnitude
+    grid_h, grid_w = torch.meshgrid(torch.linspace(-val_h + val_center, val_h + val_center, h, device=device, dtype=dtype), torch.linspace(-val_w + val_center, val_w + val_center, w, device=device, dtype=dtype), indexing='ij')
+    emb_h = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_h, device=device, dtype=dtype)
+    emb_w = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_w, device=device, dtype=dtype)
+    emb = torch.cat([emb_w, emb_h], dim=1)  # (H*W, D)
+    return emb
+
+
+#################################################################################
+#               Embedding Layers for Timesteps and Class Labels                 #
+#################################################################################
+
+
+class TimestepEmbedder(nn.Module):
+    """
+    Embeds scalar timesteps into vector representations.
+    """
+
+    def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            operations.Linear(frequency_embedding_size, hidden_size, bias=True, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(hidden_size, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+
+    def forward(self, t, dtype, **kwargs):
+        t_freq = timestep_embedding(t, self.frequency_embedding_size).to(dtype)
+        t_emb = self.mlp(t_freq)
+        return t_emb
+
+
+class VectorEmbedder(nn.Module):
+    """
+    Embeds a flat vector of dimension input_dim
+    """
+
+    def __init__(self, input_dim: int, hidden_size: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            operations.Linear(input_dim, hidden_size, bias=True, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(hidden_size, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        emb = self.mlp(x)
+        return emb
+
+
+#################################################################################
+#                                 Core DiT Model                                #
+#################################################################################
+
+
+def split_qkv(qkv, head_dim):
+    qkv = qkv.reshape(qkv.shape[0], qkv.shape[1], 3, -1, head_dim).movedim(2, 0)
+    return qkv[0], qkv[1], qkv[2]
+
+
+class SelfAttention(nn.Module):
+    ATTENTION_MODES = ("xformers", "torch", "torch-hb", "math", "debug")
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = False,
+        qk_scale: Optional[float] = None,
+        proj_drop: float = 0.0,
+        attn_mode: str = "xformers",
+        pre_only: bool = False,
+        qk_norm: Optional[str] = None,
+        rmsnorm: bool = False,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+        if not pre_only:
+            self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
+            self.proj_drop = nn.Dropout(proj_drop)
+        assert attn_mode in self.ATTENTION_MODES
+        self.attn_mode = attn_mode
+        self.pre_only = pre_only
+
+        if qk_norm == "rms":
+            self.ln_q = RMSNorm(self.head_dim, elementwise_affine=True, eps=1.0e-6, dtype=dtype, device=device)
+            self.ln_k = RMSNorm(self.head_dim, elementwise_affine=True, eps=1.0e-6, dtype=dtype, device=device)
+        elif qk_norm == "ln":
+            self.ln_q = operations.LayerNorm(self.head_dim, elementwise_affine=True, eps=1.0e-6, dtype=dtype, device=device)
+            self.ln_k = operations.LayerNorm(self.head_dim, elementwise_affine=True, eps=1.0e-6, dtype=dtype, device=device)
+        elif qk_norm is None:
+            self.ln_q = nn.Identity()
+            self.ln_k = nn.Identity()
+        else:
+            raise ValueError(qk_norm)
+
+    def pre_attention(self, x: torch.Tensor) -> torch.Tensor:
+        B, L, C = x.shape
+        qkv = self.qkv(x)
+        q, k, v = split_qkv(qkv, self.head_dim)
+        q = self.ln_q(q).reshape(q.shape[0], q.shape[1], -1)
+        k = self.ln_k(k).reshape(q.shape[0], q.shape[1], -1)
+        return (q, k, v)
+
+    def post_attention(self, x: torch.Tensor) -> torch.Tensor:
+        assert not self.pre_only
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        q, k, v = self.pre_attention(x)
+        x = optimized_attention(
+            q, k, v, heads=self.num_heads
+        )
+        x = self.post_attention(x)
+        return x
+
+
+class RMSNorm(torch.nn.Module):
+    def __init__(
+        self, dim: int, elementwise_affine: bool = False, eps: float = 1e-6, device=None, dtype=None, **kwargs
+    ):
+        """
+        Initialize the RMSNorm normalization layer.
+        Args:
+            dim (int): The dimension of the input tensor.
+            eps (float, optional): A small value added to the denominator for numerical stability. Default is 1e-6.
+        Attributes:
+            eps (float): A small value added to the denominator for numerical stability.
+            weight (nn.Parameter): Learnable scaling parameter.
+        """
+        super().__init__()
+        self.eps = eps
+        self.learnable_scale = elementwise_affine
+        if self.learnable_scale:
+            self.weight = nn.Parameter(torch.empty(dim, device=device, dtype=dtype))
+        else:
+            self.register_parameter("weight", None)
+
+    def forward(self, x):
+        return comfy.ldm.common_dit.rms_norm(x, self.weight, self.eps)
+
+
+
+class SwiGLUFeedForward(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        multiple_of: int,
+        ffn_dim_multiplier: Optional[float] = None,
+    ):
+        """
+        Initialize the FeedForward module.
+
+        Args:
+            dim (int): Input dimension.
+            hidden_dim (int): Hidden dimension of the feedforward layer.
+            multiple_of (int): Value to ensure hidden dimension is a multiple of this value.
+            ffn_dim_multiplier (float, optional): Custom multiplier for hidden dimension. Defaults to None.
+
+        Attributes:
+            w1 (ColumnParallelLinear): Linear transformation for the first layer.
+            w2 (RowParallelLinear): Linear transformation for the second layer.
+            w3 (ColumnParallelLinear): Linear transformation for the third layer.
+
+        """
+        super().__init__()
+        hidden_dim = int(2 * hidden_dim / 3)
+        # custom dim factor multiplier
+        if ffn_dim_multiplier is not None:
+            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
+        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+        self.w1 = nn.Linear(dim, hidden_dim, bias=False)
+        self.w2 = nn.Linear(hidden_dim, dim, bias=False)
+        self.w3 = nn.Linear(dim, hidden_dim, bias=False)
+
+    def forward(self, x):
+        return self.w2(nn.functional.silu(self.w1(x)) * self.w3(x))
+
+
+class DismantledBlock(nn.Module):
+    """
+    A DiT block with gated adaptive layer norm (adaLN) conditioning.
+    """
+
+    ATTENTION_MODES = ("xformers", "torch", "torch-hb", "math", "debug")
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        attn_mode: str = "xformers",
+        qkv_bias: bool = False,
+        pre_only: bool = False,
+        rmsnorm: bool = False,
+        scale_mod_only: bool = False,
+        swiglu: bool = False,
+        qk_norm: Optional[str] = None,
+        x_block_self_attn: bool = False,
+        dtype=None,
+        device=None,
+        operations=None,
+        **block_kwargs,
+    ):
+        super().__init__()
+        assert attn_mode in self.ATTENTION_MODES
+        if not rmsnorm:
+            self.norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        else:
+            self.norm1 = RMSNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.attn = SelfAttention(
+            dim=hidden_size,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            attn_mode=attn_mode,
+            pre_only=pre_only,
+            qk_norm=qk_norm,
+            rmsnorm=rmsnorm,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+        if x_block_self_attn:
+            assert not pre_only
+            assert not scale_mod_only
+            self.x_block_self_attn = True
+            self.attn2 = SelfAttention(
+                dim=hidden_size,
+                num_heads=num_heads,
+                qkv_bias=qkv_bias,
+                attn_mode=attn_mode,
+                pre_only=False,
+                qk_norm=qk_norm,
+                rmsnorm=rmsnorm,
+                dtype=dtype,
+                device=device,
+                operations=operations
+            )
+        else:
+            self.x_block_self_attn = False
+        if not pre_only:
+            if not rmsnorm:
+                self.norm2 = operations.LayerNorm(
+                    hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device
+                )
+            else:
+                self.norm2 = RMSNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        if not pre_only:
+            if not swiglu:
+                self.mlp = Mlp(
+                    in_features=hidden_size,
+                    hidden_features=mlp_hidden_dim,
+                    act_layer=lambda: nn.GELU(approximate="tanh"),
+                    drop=0,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations
+                )
+            else:
+                self.mlp = SwiGLUFeedForward(
+                    dim=hidden_size,
+                    hidden_dim=mlp_hidden_dim,
+                    multiple_of=256,
+                )
+        self.scale_mod_only = scale_mod_only
+        if x_block_self_attn:
+            assert not pre_only
+            assert not scale_mod_only
+            n_mods = 9
+        elif not scale_mod_only:
+            n_mods = 6 if not pre_only else 2
+        else:
+            n_mods = 4 if not pre_only else 1
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(), operations.Linear(hidden_size, n_mods * hidden_size, bias=True, dtype=dtype, device=device)
+        )
+        self.pre_only = pre_only
+
+    def pre_attention(self, x: torch.Tensor, c: torch.Tensor) -> torch.Tensor:
+        if not self.pre_only:
+            if not self.scale_mod_only:
+                (
+                    shift_msa,
+                    scale_msa,
+                    gate_msa,
+                    shift_mlp,
+                    scale_mlp,
+                    gate_mlp,
+                ) = self.adaLN_modulation(c).chunk(6, dim=1)
+            else:
+                shift_msa = None
+                shift_mlp = None
+                (
+                    scale_msa,
+                    gate_msa,
+                    scale_mlp,
+                    gate_mlp,
+                ) = self.adaLN_modulation(
+                    c
+                ).chunk(4, dim=1)
+            qkv = self.attn.pre_attention(modulate(self.norm1(x), shift_msa, scale_msa))
+            return qkv, (
+                x,
+                gate_msa,
+                shift_mlp,
+                scale_mlp,
+                gate_mlp,
+            )
+        else:
+            if not self.scale_mod_only:
+                (
+                    shift_msa,
+                    scale_msa,
+                ) = self.adaLN_modulation(
+                    c
+                ).chunk(2, dim=1)
+            else:
+                shift_msa = None
+                scale_msa = self.adaLN_modulation(c)
+            qkv = self.attn.pre_attention(modulate(self.norm1(x), shift_msa, scale_msa))
+            return qkv, None
+
+    def post_attention(self, attn, x, gate_msa, shift_mlp, scale_mlp, gate_mlp):
+        assert not self.pre_only
+        x = x + gate_msa.unsqueeze(1) * self.attn.post_attention(attn)
+        x = x + gate_mlp.unsqueeze(1) * self.mlp(
+            modulate(self.norm2(x), shift_mlp, scale_mlp)
+        )
+        return x
+
+    def pre_attention_x(self, x: torch.Tensor, c: torch.Tensor) -> torch.Tensor:
+        assert self.x_block_self_attn
+        (
+            shift_msa,
+            scale_msa,
+            gate_msa,
+            shift_mlp,
+            scale_mlp,
+            gate_mlp,
+            shift_msa2,
+            scale_msa2,
+            gate_msa2,
+        ) = self.adaLN_modulation(c).chunk(9, dim=1)
+        x_norm = self.norm1(x)
+        qkv = self.attn.pre_attention(modulate(x_norm, shift_msa, scale_msa))
+        qkv2 = self.attn2.pre_attention(modulate(x_norm, shift_msa2, scale_msa2))
+        return qkv, qkv2, (
+            x,
+            gate_msa,
+            shift_mlp,
+            scale_mlp,
+            gate_mlp,
+            gate_msa2,
+        )
+
+    def post_attention_x(self, attn, attn2, x, gate_msa, shift_mlp, scale_mlp, gate_mlp, gate_msa2):
+        assert not self.pre_only
+        attn1 = self.attn.post_attention(attn)
+        attn2 = self.attn2.post_attention(attn2)
+        x = gate_cat(x, gate_msa, gate_msa2, attn1, attn2)
+        x = x + gate_mlp.unsqueeze(1) * self.mlp(
+            modulate(self.norm2(x), shift_mlp, scale_mlp)
+        )
+        return x
+
+    def forward(self, x: torch.Tensor, c: torch.Tensor) -> torch.Tensor:
+        assert not self.pre_only
+        if self.x_block_self_attn:
+            qkv, qkv2, intermediates = self.pre_attention_x(x, c)
+            attn, _ = optimized_attention(
+                qkv[0], qkv[1], qkv[2],
+                num_heads=self.attn.num_heads,
+            )
+            attn2, _ = optimized_attention(
+                qkv2[0], qkv2[1], qkv2[2],
+                num_heads=self.attn2.num_heads,
+            )
+            return self.post_attention_x(attn, attn2, *intermediates)
+        else:
+            qkv, intermediates = self.pre_attention(x, c)
+            attn = optimized_attention(
+                qkv[0], qkv[1], qkv[2],
+                heads=self.attn.num_heads,
+            )
+            return self.post_attention(attn, *intermediates)
+
+def gate_cat(x, gate_msa, gate_msa2, attn1, attn2):
+    out1 = gate_msa.unsqueeze(1) * attn1
+    out2 = gate_msa2.unsqueeze(1) * attn2
+    x = torch.stack([x, out1, out2], dim=0).sum(dim=0)
+    return x
+
+def block_mixing(*args, use_checkpoint=True, **kwargs):
+    if use_checkpoint:
+        return torch.utils.checkpoint.checkpoint(
+            _block_mixing, *args, use_reentrant=False, **kwargs
+        )
+    else:
+        return _block_mixing(*args, **kwargs)
+
+
+def _block_mixing(context, x, context_block, x_block, c):
+    context_qkv, context_intermediates = context_block.pre_attention(context, c)
+
+    if x_block.x_block_self_attn:
+        x_qkv, x_qkv2, x_intermediates = x_block.pre_attention_x(x, c)
+    else:
+        x_qkv, x_intermediates = x_block.pre_attention(x, c)
+
+    o = []
+    for t in range(3):
+        o.append(torch.cat((context_qkv[t], x_qkv[t]), dim=1))
+    qkv = tuple(o)
+
+    attn = optimized_attention(
+        qkv[0], qkv[1], qkv[2],
+        heads=x_block.attn.num_heads,
+    )
+    context_attn, x_attn = (
+        attn[:, : context_qkv[0].shape[1]],
+        attn[:, context_qkv[0].shape[1] :],
+    )
+
+    if not context_block.pre_only:
+        context = context_block.post_attention(context_attn, *context_intermediates)
+
+    else:
+        context = None
+    if x_block.x_block_self_attn:
+        attn2 = optimized_attention(
+                x_qkv2[0], x_qkv2[1], x_qkv2[2],
+                heads=x_block.attn2.num_heads,
+            )
+        x = x_block.post_attention_x(x_attn, attn2, *x_intermediates)
+    else:
+        x = x_block.post_attention(x_attn, *x_intermediates)
+    return context, x
+
+
+class JointBlock(nn.Module):
+    """just a small wrapper to serve as a fsdp unit"""
+
+    def __init__(
+        self,
+        *args,
+        **kwargs,
+    ):
+        super().__init__()
+        pre_only = kwargs.pop("pre_only")
+        qk_norm = kwargs.pop("qk_norm", None)
+        x_block_self_attn = kwargs.pop("x_block_self_attn", False)
+        self.context_block = DismantledBlock(*args, pre_only=pre_only, qk_norm=qk_norm, **kwargs)
+        self.x_block = DismantledBlock(*args,
+                                       pre_only=False,
+                                       qk_norm=qk_norm,
+                                       x_block_self_attn=x_block_self_attn,
+                                       **kwargs)
+
+    def forward(self, *args, **kwargs):
+        return block_mixing(
+            *args, context_block=self.context_block, x_block=self.x_block, **kwargs
+        )
+
+
+class FinalLayer(nn.Module):
+    """
+    The final layer of DiT.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        patch_size: int,
+        out_channels: int,
+        total_out_channels: Optional[int] = None,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = (
+            operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
+            if (total_out_channels is None)
+            else operations.Linear(hidden_size, total_out_channels, bias=True, dtype=dtype, device=device)
+        )
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device)
+        )
+
+    def forward(self, x: torch.Tensor, c: torch.Tensor) -> torch.Tensor:
+        shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
+        x = modulate(self.norm_final(x), shift, scale)
+        x = self.linear(x)
+        return x
+
+class SelfAttentionContext(nn.Module):
+    def __init__(self, dim, heads=8, dim_head=64, dtype=None, device=None, operations=None):
+        super().__init__()
+        dim_head = dim // heads
+        inner_dim = dim
+
+        self.heads = heads
+        self.dim_head = dim_head
+
+        self.qkv = operations.Linear(dim, dim * 3, bias=True, dtype=dtype, device=device)
+
+        self.proj = operations.Linear(inner_dim, dim, dtype=dtype, device=device)
+
+    def forward(self, x):
+        qkv = self.qkv(x)
+        q, k, v = split_qkv(qkv, self.dim_head)
+        x = optimized_attention(q.reshape(q.shape[0], q.shape[1], -1), k, v, heads=self.heads)
+        return self.proj(x)
+
+class ContextProcessorBlock(nn.Module):
+    def __init__(self, context_size, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm1 = operations.LayerNorm(context_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.attn = SelfAttentionContext(context_size, dtype=dtype, device=device, operations=operations)
+        self.norm2 = operations.LayerNorm(context_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.mlp = Mlp(in_features=context_size, hidden_features=(context_size * 4), act_layer=lambda: nn.GELU(approximate="tanh"), drop=0, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, x):
+        x += self.attn(self.norm1(x))
+        x += self.mlp(self.norm2(x))
+        return x
+
+class ContextProcessor(nn.Module):
+    def __init__(self, context_size, num_layers, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.layers = torch.nn.ModuleList([ContextProcessorBlock(context_size, dtype=dtype, device=device, operations=operations) for i in range(num_layers)])
+        self.norm = operations.LayerNorm(context_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+
+    def forward(self, x):
+        for i, l in enumerate(self.layers):
+            x = l(x)
+        return self.norm(x)
+
+class MMDiT(nn.Module):
+    """
+    Diffusion model with a Transformer backbone.
+    """
+
+    def __init__(
+        self,
+        input_size: int = 32,
+        patch_size: int = 2,
+        in_channels: int = 4,
+        depth: int = 28,
+        # hidden_size: Optional[int] = None,
+        # num_heads: Optional[int] = None,
+        mlp_ratio: float = 4.0,
+        learn_sigma: bool = False,
+        adm_in_channels: Optional[int] = None,
+        context_embedder_config: Optional[Dict] = None,
+        compile_core: bool = False,
+        use_checkpoint: bool = False,
+        register_length: int = 0,
+        attn_mode: str = "torch",
+        rmsnorm: bool = False,
+        scale_mod_only: bool = False,
+        swiglu: bool = False,
+        out_channels: Optional[int] = None,
+        pos_embed_scaling_factor: Optional[float] = None,
+        pos_embed_offset: Optional[float] = None,
+        pos_embed_max_size: Optional[int] = None,
+        num_patches = None,
+        qk_norm: Optional[str] = None,
+        qkv_bias: bool = True,
+        context_processor_layers = None,
+        x_block_self_attn: bool = False,
+        x_block_self_attn_layers: Optional[List[int]] = [],
+        context_size = 4096,
+        num_blocks = None,
+        final_layer = True,
+        skip_blocks = False,
+        dtype = None, #TODO
+        device = None,
+        operations = None,
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.learn_sigma = learn_sigma
+        self.in_channels = in_channels
+        default_out_channels = in_channels * 2 if learn_sigma else in_channels
+        self.out_channels = default(out_channels, default_out_channels)
+        self.patch_size = patch_size
+        self.pos_embed_scaling_factor = pos_embed_scaling_factor
+        self.pos_embed_offset = pos_embed_offset
+        self.pos_embed_max_size = pos_embed_max_size
+        self.x_block_self_attn_layers = x_block_self_attn_layers
+
+        # hidden_size = default(hidden_size, 64 * depth)
+        # num_heads = default(num_heads, hidden_size // 64)
+
+        # apply magic --> this defines a head_size of 64
+        self.hidden_size = 64 * depth
+        num_heads = depth
+        if num_blocks is None:
+            num_blocks = depth
+
+        self.depth = depth
+        self.num_heads = num_heads
+
+        self.x_embedder = PatchEmbed(
+            input_size,
+            patch_size,
+            in_channels,
+            self.hidden_size,
+            bias=True,
+            strict_img_size=self.pos_embed_max_size is None,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+        self.t_embedder = TimestepEmbedder(self.hidden_size, dtype=dtype, device=device, operations=operations)
+
+        self.y_embedder = None
+        if adm_in_channels is not None:
+            assert isinstance(adm_in_channels, int)
+            self.y_embedder = VectorEmbedder(adm_in_channels, self.hidden_size, dtype=dtype, device=device, operations=operations)
+
+        if context_processor_layers is not None:
+            self.context_processor = ContextProcessor(context_size, context_processor_layers, dtype=dtype, device=device, operations=operations)
+        else:
+            self.context_processor = None
+
+        self.context_embedder = nn.Identity()
+        if context_embedder_config is not None:
+            if context_embedder_config["target"] == "torch.nn.Linear":
+                self.context_embedder = operations.Linear(**context_embedder_config["params"], dtype=dtype, device=device)
+
+        self.register_length = register_length
+        if self.register_length > 0:
+            self.register = nn.Parameter(torch.randn(1, register_length, self.hidden_size, dtype=dtype, device=device))
+
+        # num_patches = self.x_embedder.num_patches
+        # Will use fixed sin-cos embedding:
+        # just use a buffer already
+        if num_patches is not None:
+            self.register_buffer(
+                "pos_embed",
+                torch.empty(1, num_patches, self.hidden_size, dtype=dtype, device=device),
+            )
+        else:
+            self.pos_embed = None
+
+        self.use_checkpoint = use_checkpoint
+        if not skip_blocks:
+            self.joint_blocks = nn.ModuleList(
+                [
+                    JointBlock(
+                        self.hidden_size,
+                        num_heads,
+                        mlp_ratio=mlp_ratio,
+                        qkv_bias=qkv_bias,
+                        attn_mode=attn_mode,
+                        pre_only=(i == num_blocks - 1) and final_layer,
+                        rmsnorm=rmsnorm,
+                        scale_mod_only=scale_mod_only,
+                        swiglu=swiglu,
+                        qk_norm=qk_norm,
+                        x_block_self_attn=(i in self.x_block_self_attn_layers) or x_block_self_attn,
+                        dtype=dtype,
+                        device=device,
+                        operations=operations,
+                    )
+                    for i in range(num_blocks)
+                ]
+            )
+
+        if final_layer:
+            self.final_layer = FinalLayer(self.hidden_size, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations)
+
+        if compile_core:
+            assert False
+            self.forward_core_with_concat = torch.compile(self.forward_core_with_concat)
+
+    def cropped_pos_embed(self, hw, device=None):
+        p = self.x_embedder.patch_size[0]
+        h, w = hw
+        # patched size
+        h = (h + 1) // p
+        w = (w + 1) // p
+        if self.pos_embed is None:
+            return get_2d_sincos_pos_embed_torch(self.hidden_size, w, h, device=device)
+        assert self.pos_embed_max_size is not None
+        assert h <= self.pos_embed_max_size, (h, self.pos_embed_max_size)
+        assert w <= self.pos_embed_max_size, (w, self.pos_embed_max_size)
+        top = (self.pos_embed_max_size - h) // 2
+        left = (self.pos_embed_max_size - w) // 2
+        spatial_pos_embed = rearrange(
+            self.pos_embed,
+            "1 (h w) c -> 1 h w c",
+            h=self.pos_embed_max_size,
+            w=self.pos_embed_max_size,
+        )
+        spatial_pos_embed = spatial_pos_embed[:, top : top + h, left : left + w, :]
+        spatial_pos_embed = rearrange(spatial_pos_embed, "1 h w c -> 1 (h w) c")
+        # print(spatial_pos_embed, top, left, h, w)
+        # # t = get_2d_sincos_pos_embed_torch(self.hidden_size, w, h, 7.875, 7.875, device=device) #matches exactly for 1024 res
+        # t = get_2d_sincos_pos_embed_torch(self.hidden_size, w, h, 7.5, 7.5, device=device) #scales better
+        # # print(t)
+        # return t
+        return spatial_pos_embed
+
+    def unpatchify(self, x, hw=None):
+        """
+        x: (N, T, patch_size**2 * C)
+        imgs: (N, H, W, C)
+        """
+        c = self.out_channels
+        p = self.x_embedder.patch_size[0]
+        if hw is None:
+            h = w = int(x.shape[1] ** 0.5)
+        else:
+            h, w = hw
+            h = (h + 1) // p
+            w = (w + 1) // p
+        assert h * w == x.shape[1]
+
+        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
+        x = torch.einsum("nhwpqc->nchpwq", x)
+        imgs = x.reshape(shape=(x.shape[0], c, h * p, w * p))
+        return imgs
+
+    def forward_core_with_concat(
+        self,
+        x: torch.Tensor,
+        c_mod: torch.Tensor,
+        context: Optional[torch.Tensor] = None,
+        control = None,
+        transformer_options = {},
+    ) -> torch.Tensor:
+        patches_replace = transformer_options.get("patches_replace", {})
+        if self.register_length > 0:
+            context = torch.cat(
+                (
+                    repeat(self.register, "1 ... -> b ...", b=x.shape[0]),
+                    default(context, torch.Tensor([]).type_as(x)),
+                ),
+                1,
+            )
+
+        # context is B, L', D
+        # x is B, L, D
+        blocks_replace = patches_replace.get("dit", {})
+        blocks = len(self.joint_blocks)
+        for i in range(blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["txt"], out["img"] = self.joint_blocks[i](args["txt"], args["img"], c=args["vec"])
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": c_mod}, {"original_block": block_wrap})
+                context = out["txt"]
+                x = out["img"]
+            else:
+                context, x = self.joint_blocks[i](
+                    context,
+                    x,
+                    c=c_mod,
+                    use_checkpoint=self.use_checkpoint,
+                )
+            if control is not None:
+                control_o = control.get("output")
+                if i < len(control_o):
+                    add = control_o[i]
+                    if add is not None:
+                        x += add
+
+        x = self.final_layer(x, c_mod)  # (N, T, patch_size ** 2 * out_channels)
+        return x
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        t: torch.Tensor,
+        y: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        control = None,
+        transformer_options = {},
+    ) -> torch.Tensor:
+        """
+        Forward pass of DiT.
+        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
+        t: (N,) tensor of diffusion timesteps
+        y: (N,) tensor of class labels
+        """
+
+        if self.context_processor is not None:
+            context = self.context_processor(context)
+
+        hw = x.shape[-2:]
+        x = self.x_embedder(x) + comfy.ops.cast_to_input(self.cropped_pos_embed(hw, device=x.device), x)
+        c = self.t_embedder(t, dtype=x.dtype)  # (N, D)
+        if y is not None and self.y_embedder is not None:
+            y = self.y_embedder(y)  # (N, D)
+            c = c + y  # (N, D)
+
+        if context is not None:
+            context = self.context_embedder(context)
+
+        x = self.forward_core_with_concat(x, c, context, control, transformer_options)
+
+        x = self.unpatchify(x, hw=hw)  # (N, out_channels, H, W)
+        return x[:,:,:hw[-2],:hw[-1]]
+
+
+class OpenAISignatureMMDITWrapper(MMDiT):
+    def forward(
+        self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        context: Optional[torch.Tensor] = None,
+        y: Optional[torch.Tensor] = None,
+        control = None,
+        transformer_options = {},
+        **kwargs,
+    ) -> torch.Tensor:
+        return super().forward(x, timesteps, context=context, y=y, control=control, transformer_options=transformer_options)
+

comfy/ldm/modules/diffusionmodules/model.py

+# pytorch_diffusion + derived encoder decoder
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+import logging
+
+from comfy import model_management
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+if model_management.xformers_enabled_vae():
+    import xformers
+    import xformers.ops
+
+def get_timestep_embedding(timesteps, embedding_dim):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models:
+    From Fairseq.
+    Build sinusoidal embeddings.
+    This matches the implementation in tensor2tensor, but differs slightly
+    from the description in Section 3.5 of "Attention Is All You Need".
+    """
+    assert len(timesteps.shape) == 1
+
+    half_dim = embedding_dim // 2
+    emb = math.log(10000) / (half_dim - 1)
+    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
+    emb = emb.to(device=timesteps.device)
+    emb = timesteps.float()[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0,1,0,0))
+    return emb
+
+
+def nonlinearity(x):
+    # swish
+    return torch.nn.functional.silu(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return ops.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class VideoConv3d(nn.Module):
+    def __init__(self, n_channels, out_channels, kernel_size, stride=1, dilation=1, padding_mode='replicate', padding=1, **kwargs):
+        super().__init__()
+
+        self.padding_mode = padding_mode
+        if padding != 0:
+            padding = (padding, padding, padding, padding, kernel_size - 1, 0)
+        else:
+            kwargs["padding"] = padding
+
+        self.padding = padding
+        self.conv = ops.Conv3d(n_channels, out_channels, kernel_size, stride=stride, dilation=dilation, **kwargs)
+
+    def forward(self, x):
+        if self.padding != 0:
+            x = torch.nn.functional.pad(x, self.padding, mode=self.padding_mode)
+        return self.conv(x)
+
+def interpolate_up(x, scale_factor):
+    try:
+        return torch.nn.functional.interpolate(x, scale_factor=scale_factor, mode="nearest")
+    except: #operation not implemented for bf16
+        orig_shape = list(x.shape)
+        out_shape = orig_shape[:2]
+        for i in range(len(orig_shape) - 2):
+            out_shape.append(round(orig_shape[i + 2] * scale_factor[i]))
+        out = torch.empty(out_shape, dtype=x.dtype, layout=x.layout, device=x.device)
+        split = 8
+        l = out.shape[1] // split
+        for i in range(0, out.shape[1], l):
+            out[:,i:i+l] = torch.nn.functional.interpolate(x[:,i:i+l].to(torch.float32), scale_factor=scale_factor, mode="nearest").to(x.dtype)
+        return out
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv, conv_op=ops.Conv2d, scale_factor=2.0):
+        super().__init__()
+        self.with_conv = with_conv
+        self.scale_factor = scale_factor
+
+        if self.with_conv:
+            self.conv = conv_op(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        scale_factor = self.scale_factor
+        if isinstance(scale_factor, (int, float)):
+            scale_factor = (scale_factor,) * (x.ndim - 2)
+
+        if x.ndim == 5 and scale_factor[0] > 1.0:
+            t = x.shape[2]
+            if t > 1:
+                a, b = x.split((1, t - 1), dim=2)
+                del x
+                b = interpolate_up(b, scale_factor)
+            else:
+                a = x
+
+            a = interpolate_up(a.squeeze(2), scale_factor=scale_factor[1:]).unsqueeze(2)
+            if t > 1:
+                x = torch.cat((a, b), dim=2)
+            else:
+                x = a
+        else:
+            x = interpolate_up(x, scale_factor)
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv, stride=2, conv_op=ops.Conv2d):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = conv_op(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=stride,
+                                        padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            if x.ndim == 4:
+                pad = (0, 1, 0, 1)
+                mode = "constant"
+                x = torch.nn.functional.pad(x, pad, mode=mode, value=0)
+            elif x.ndim == 5:
+                pad = (1, 1, 1, 1, 2, 0)
+                mode = "replicate"
+                x = torch.nn.functional.pad(x, pad, mode=mode)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
+                 dropout, temb_channels=512, conv_op=ops.Conv2d):
+        super().__init__()
+        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.swish = torch.nn.SiLU(inplace=True)
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = conv_op(in_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if temb_channels > 0:
+            self.temb_proj = ops.Linear(temb_channels,
+                                             out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout, inplace=True)
+        self.conv2 = conv_op(out_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = conv_op(in_channels,
+                                                     out_channels,
+                                                     kernel_size=3,
+                                                     stride=1,
+                                                     padding=1)
+            else:
+                self.nin_shortcut = conv_op(in_channels,
+                                                    out_channels,
+                                                    kernel_size=1,
+                                                    stride=1,
+                                                    padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = self.swish(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(self.swish(temb))[:,:,None,None]
+
+        h = self.norm2(h)
+        h = self.swish(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x+h
+
+def slice_attention(q, k, v):
+    r1 = torch.zeros_like(k, device=q.device)
+    scale = (int(q.shape[-1])**(-0.5))
+
+    mem_free_total = model_management.get_free_memory(q.device)
+
+    tensor_size = q.shape[0] * q.shape[1] * k.shape[2] * q.element_size()
+    modifier = 3 if q.element_size() == 2 else 2.5
+    mem_required = tensor_size * modifier
+    steps = 1
+
+    if mem_required > mem_free_total:
+        steps = 2**(math.ceil(math.log(mem_required / mem_free_total, 2)))
+
+    while True:
+        try:
+            slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1]
+            for i in range(0, q.shape[1], slice_size):
+                end = i + slice_size
+                s1 = torch.bmm(q[:, i:end], k) * scale
+
+                s2 = torch.nn.functional.softmax(s1, dim=2).permute(0,2,1)
+                del s1
+
+                r1[:, :, i:end] = torch.bmm(v, s2)
+                del s2
+            break
+        except model_management.OOM_EXCEPTION as e:
+            model_management.soft_empty_cache(True)
+            steps *= 2
+            if steps > 128:
+                raise e
+            logging.warning("out of memory error, increasing steps and trying again {}".format(steps))
+
+    return r1
+
+def normal_attention(q, k, v):
+    # compute attention
+    orig_shape = q.shape
+    b = orig_shape[0]
+    c = orig_shape[1]
+
+    q = q.reshape(b, c, -1)
+    q = q.permute(0, 2, 1)   # b,hw,c
+    k = k.reshape(b, c, -1) # b,c,hw
+    v = v.reshape(b, c, -1)
+
+    r1 = slice_attention(q, k, v)
+    h_ = r1.reshape(orig_shape)
+    del r1
+    return h_
+
+def xformers_attention(q, k, v):
+    # compute attention
+    orig_shape = q.shape
+    B = orig_shape[0]
+    C = orig_shape[1]
+    q, k, v = map(
+        lambda t: t.view(B, C, -1).transpose(1, 2).contiguous(),
+        (q, k, v),
+    )
+
+    try:
+        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None)
+        out = out.transpose(1, 2).reshape(orig_shape)
+    except NotImplementedError:
+        out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(orig_shape)
+    return out
+
+def pytorch_attention(q, k, v):
+    # compute attention
+    orig_shape = q.shape
+    B = orig_shape[0]
+    C = orig_shape[1]
+    q, k, v = map(
+        lambda t: t.view(B, 1, C, -1).transpose(2, 3).contiguous(),
+        (q, k, v),
+    )
+
+    try:
+        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
+        out = out.transpose(2, 3).reshape(orig_shape)
+    except model_management.OOM_EXCEPTION:
+        logging.warning("scaled_dot_product_attention OOMed: switched to slice attention")
+        out = slice_attention(q.view(B, -1, C), k.view(B, -1, C).transpose(1, 2), v.view(B, -1, C).transpose(1, 2)).reshape(orig_shape)
+    return out
+
+
+def vae_attention():
+    if model_management.xformers_enabled_vae():
+        logging.info("Using xformers attention in VAE")
+        return xformers_attention
+    elif model_management.pytorch_attention_enabled_vae():
+        logging.info("Using pytorch attention in VAE")
+        return pytorch_attention
+    else:
+        logging.info("Using split attention in VAE")
+        return normal_attention
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels, conv_op=ops.Conv2d):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = conv_op(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = conv_op(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = conv_op(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = conv_op(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+        self.optimized_attention = vae_attention()
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        h_ = self.optimized_attention(q, k, v)
+
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None, conv_op=ops.Conv2d):
+    return AttnBlock(in_channels, conv_op=conv_op)
+
+
+class Model(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = self.ch*4
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        self.use_timestep = use_timestep
+        if self.use_timestep:
+            # timestep embedding
+            self.temb = nn.Module()
+            self.temb.dense = nn.ModuleList([
+                ops.Linear(self.ch,
+                                self.temb_ch),
+                ops.Linear(self.temb_ch,
+                                self.temb_ch),
+            ])
+
+        # downsampling
+        self.conv_in = ops.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            skip_in = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                if i_block == self.num_res_blocks:
+                    skip_in = ch*in_ch_mult[i_level]
+                block.append(ResnetBlock(in_channels=block_in+skip_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = ops.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x, t=None, context=None):
+        #assert x.shape[2] == x.shape[3] == self.resolution
+        if context is not None:
+            # assume aligned context, cat along channel axis
+            x = torch.cat((x, context), dim=1)
+        if self.use_timestep:
+            # timestep embedding
+            assert t is not None
+            temb = get_timestep_embedding(t, self.ch)
+            temb = self.temb.dense[0](temb)
+            temb = nonlinearity(temb)
+            temb = self.temb.dense[1](temb)
+        else:
+            temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](
+                    torch.cat([h, hs.pop()], dim=1), temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+    def get_last_layer(self):
+        return self.conv_out.weight
+
+
+class Encoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
+                 conv3d=False, time_compress=None,
+                 **ignore_kwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        if conv3d:
+            conv_op = VideoConv3d
+            mid_attn_conv_op = ops.Conv3d
+        else:
+            conv_op = ops.Conv2d
+            mid_attn_conv_op = ops.Conv2d
+
+        # downsampling
+        self.conv_in = conv_op(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout,
+                                         conv_op=conv_op))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type, conv_op=conv_op))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                stride = 2
+                if time_compress is not None:
+                    if (self.num_resolutions - 1 - i_level) > math.log2(time_compress):
+                        stride = (1, 2, 2)
+                down.downsample = Downsample(block_in, resamp_with_conv, stride=stride, conv_op=conv_op)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout,
+                                       conv_op=conv_op)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type, conv_op=mid_attn_conv_op)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout,
+                                       conv_op=conv_op)
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = conv_op(block_in,
+                                        2*z_channels if double_z else z_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # timestep embedding
+        temb = None
+        # downsampling
+        h = self.conv_in(x)
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](h, temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+            if i_level != self.num_resolutions-1:
+                h = self.down[i_level].downsample(h)
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
+                 conv_out_op=ops.Conv2d,
+                 resnet_op=ResnetBlock,
+                 attn_op=AttnBlock,
+                 conv3d=False,
+                 time_compress=None,
+                **ignorekwargs):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+
+        if conv3d:
+            conv_op = VideoConv3d
+            conv_out_op = VideoConv3d
+            mid_attn_conv_op = ops.Conv3d
+        else:
+            conv_op = ops.Conv2d
+            mid_attn_conv_op = ops.Conv2d
+
+        # compute block_in and curr_res at lowest res
+        block_in = ch*ch_mult[self.num_resolutions-1]
+        curr_res = resolution // 2**(self.num_resolutions-1)
+        self.z_shape = (1,z_channels,curr_res,curr_res)
+        logging.debug("Working with z of shape {} = {} dimensions.".format(
+            self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = conv_op(z_channels,
+                                       block_in,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = resnet_op(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout,
+                                       conv_op=conv_op)
+        self.mid.attn_1 = attn_op(block_in, conv_op=mid_attn_conv_op)
+        self.mid.block_2 = resnet_op(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout,
+                                       conv_op=conv_op)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                block.append(resnet_op(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout,
+                                         conv_op=conv_op))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(attn_op(block_in, conv_op=conv_op))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                scale_factor = 2.0
+                if time_compress is not None:
+                    if i_level > math.log2(time_compress):
+                        scale_factor = (1.0, 2.0, 2.0)
+
+                up.upsample = Upsample(block_in, resamp_with_conv, conv_op=conv_op, scale_factor=scale_factor)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = conv_out_op(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, z, **kwargs):
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb, **kwargs)
+        h = self.mid.attn_1(h, **kwargs)
+        h = self.mid.block_2(h, temb, **kwargs)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](h, temb, **kwargs)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h, **kwargs)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h, **kwargs)
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h

comfy/ldm/modules/diffusionmodules/openaimodel.py

+from abc import abstractmethod
+
+import torch as th
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+import logging
+
+from .util import (
+    checkpoint,
+    avg_pool_nd,
+    timestep_embedding,
+    AlphaBlender,
+)
+from ..attention import SpatialTransformer, SpatialVideoTransformer, default
+from comfy.ldm.util import exists
+import comfy.patcher_extension
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+#This is needed because accelerate makes a copy of transformer_options which breaks "transformer_index"
+def forward_timestep_embed(ts, x, emb, context=None, transformer_options={}, output_shape=None, time_context=None, num_video_frames=None, image_only_indicator=None):
+    for layer in ts:
+        if isinstance(layer, VideoResBlock):
+            x = layer(x, emb, num_video_frames, image_only_indicator)
+        elif isinstance(layer, TimestepBlock):
+            x = layer(x, emb)
+        elif isinstance(layer, SpatialVideoTransformer):
+            x = layer(x, context, time_context, num_video_frames, image_only_indicator, transformer_options)
+            if "transformer_index" in transformer_options:
+                transformer_options["transformer_index"] += 1
+        elif isinstance(layer, SpatialTransformer):
+            x = layer(x, context, transformer_options)
+            if "transformer_index" in transformer_options:
+                transformer_options["transformer_index"] += 1
+        elif isinstance(layer, Upsample):
+            x = layer(x, output_shape=output_shape)
+        else:
+            if "patches" in transformer_options and "forward_timestep_embed_patch" in transformer_options["patches"]:
+                found_patched = False
+                for class_type, handler in transformer_options["patches"]["forward_timestep_embed_patch"]:
+                    if isinstance(layer, class_type):
+                        x = handler(layer, x, emb, context, transformer_options, output_shape, time_context, num_video_frames, image_only_indicator)
+                        found_patched = True
+                        break
+                if found_patched:
+                    continue
+            x = layer(x)
+    return x
+
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, *args, **kwargs):
+        return forward_timestep_embed(self, *args, **kwargs)
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None, operations=ops):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = operations.conv_nd(dims, self.channels, self.out_channels, 3, padding=padding, dtype=dtype, device=device)
+
+    def forward(self, x, output_shape=None):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            shape = [x.shape[2], x.shape[3] * 2, x.shape[4] * 2]
+            if output_shape is not None:
+                shape[1] = output_shape[3]
+                shape[2] = output_shape[4]
+        else:
+            shape = [x.shape[2] * 2, x.shape[3] * 2]
+            if output_shape is not None:
+                shape[0] = output_shape[2]
+                shape[1] = output_shape[3]
+
+        x = F.interpolate(x, size=shape, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1, dtype=None, device=None, operations=ops):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = operations.conv_nd(
+                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding, dtype=dtype, device=device
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param use_checkpoint: if True, use gradient checkpointing on this module.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_conv=False,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        up=False,
+        down=False,
+        kernel_size=3,
+        exchange_temb_dims=False,
+        skip_t_emb=False,
+        dtype=None,
+        device=None,
+        operations=ops
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+        self.exchange_temb_dims = exchange_temb_dims
+
+        if isinstance(kernel_size, list):
+            padding = [k // 2 for k in kernel_size]
+        else:
+            padding = kernel_size // 2
+
+        self.in_layers = nn.Sequential(
+            operations.GroupNorm(32, channels, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.conv_nd(dims, channels, self.out_channels, kernel_size, padding=padding, dtype=dtype, device=device),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims, dtype=dtype, device=device)
+            self.x_upd = Upsample(channels, False, dims, dtype=dtype, device=device)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims, dtype=dtype, device=device)
+            self.x_upd = Downsample(channels, False, dims, dtype=dtype, device=device)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.skip_t_emb = skip_t_emb
+        if self.skip_t_emb:
+            self.emb_layers = None
+            self.exchange_temb_dims = False
+        else:
+            self.emb_layers = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(
+                    emb_channels,
+                    2 * self.out_channels if use_scale_shift_norm else self.out_channels, dtype=dtype, device=device
+                ),
+            )
+        self.out_layers = nn.Sequential(
+            operations.GroupNorm(32, self.out_channels, dtype=dtype, device=device),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            operations.conv_nd(dims, self.out_channels, self.out_channels, kernel_size, padding=padding, dtype=dtype, device=device)
+            ,
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = operations.conv_nd(
+                dims, channels, self.out_channels, kernel_size, padding=padding, dtype=dtype, device=device
+            )
+        else:
+            self.skip_connection = operations.conv_nd(dims, channels, self.out_channels, 1, dtype=dtype, device=device)
+
+    def forward(self, x, emb):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        return checkpoint(
+            self._forward, (x, emb), self.parameters(), self.use_checkpoint
+        )
+
+
+    def _forward(self, x, emb):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+
+        emb_out = None
+        if not self.skip_t_emb:
+            emb_out = self.emb_layers(emb).type(h.dtype)
+            while len(emb_out.shape) < len(h.shape):
+                emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            h = out_norm(h)
+            if emb_out is not None:
+                scale, shift = th.chunk(emb_out, 2, dim=1)
+                h *= (1 + scale)
+                h += shift
+            h = out_rest(h)
+        else:
+            if emb_out is not None:
+                if self.exchange_temb_dims:
+                    emb_out = emb_out.movedim(1, 2)
+                h = h + emb_out
+            h = self.out_layers(h)
+        return self.skip_connection(x) + h
+
+
+class VideoResBlock(ResBlock):
+    def __init__(
+        self,
+        channels: int,
+        emb_channels: int,
+        dropout: float,
+        video_kernel_size=3,
+        merge_strategy: str = "fixed",
+        merge_factor: float = 0.5,
+        out_channels=None,
+        use_conv: bool = False,
+        use_scale_shift_norm: bool = False,
+        dims: int = 2,
+        use_checkpoint: bool = False,
+        up: bool = False,
+        down: bool = False,
+        dtype=None,
+        device=None,
+        operations=ops
+    ):
+        super().__init__(
+            channels,
+            emb_channels,
+            dropout,
+            out_channels=out_channels,
+            use_conv=use_conv,
+            use_scale_shift_norm=use_scale_shift_norm,
+            dims=dims,
+            use_checkpoint=use_checkpoint,
+            up=up,
+            down=down,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+
+        self.time_stack = ResBlock(
+            default(out_channels, channels),
+            emb_channels,
+            dropout=dropout,
+            dims=3,
+            out_channels=default(out_channels, channels),
+            use_scale_shift_norm=False,
+            use_conv=False,
+            up=False,
+            down=False,
+            kernel_size=video_kernel_size,
+            use_checkpoint=use_checkpoint,
+            exchange_temb_dims=True,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+        self.time_mixer = AlphaBlender(
+            alpha=merge_factor,
+            merge_strategy=merge_strategy,
+            rearrange_pattern="b t -> b 1 t 1 1",
+        )
+
+    def forward(
+        self,
+        x: th.Tensor,
+        emb: th.Tensor,
+        num_video_frames: int,
+        image_only_indicator = None,
+    ) -> th.Tensor:
+        x = super().forward(x, emb)
+
+        x_mix = rearrange(x, "(b t) c h w -> b c t h w", t=num_video_frames)
+        x = rearrange(x, "(b t) c h w -> b c t h w", t=num_video_frames)
+
+        x = self.time_stack(
+            x, rearrange(emb, "(b t) ... -> b t ...", t=num_video_frames)
+        )
+        x = self.time_mixer(
+            x_spatial=x_mix, x_temporal=x, image_only_indicator=image_only_indicator
+        )
+        x = rearrange(x, "b c t h w -> (b t) c h w")
+        return x
+
+
+class Timestep(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, t):
+        return timestep_embedding(t, self.dim)
+
+def apply_control(h, control, name):
+    if control is not None and name in control and len(control[name]) > 0:
+        ctrl = control[name].pop()
+        if ctrl is not None:
+            try:
+                h += ctrl
+            except:
+                logging.warning("warning control could not be applied {} {}".format(h.shape, ctrl.shape))
+    return h
+
+class UNetModel(nn.Module):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        use_checkpoint=False,
+        dtype=th.float32,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,    # custom transformer support
+        transformer_depth=1,              # custom transformer support
+        context_dim=None,                 # custom transformer support
+        n_embed=None,                     # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+        disable_self_attentions=None,
+        num_attention_blocks=None,
+        disable_middle_self_attn=False,
+        use_linear_in_transformer=False,
+        adm_in_channels=None,
+        transformer_depth_middle=None,
+        transformer_depth_output=None,
+        use_temporal_resblock=False,
+        use_temporal_attention=False,
+        time_context_dim=None,
+        extra_ff_mix_layer=False,
+        use_spatial_context=False,
+        merge_strategy=None,
+        merge_factor=0.0,
+        video_kernel_size=None,
+        disable_temporal_crossattention=False,
+        max_ddpm_temb_period=10000,
+        attn_precision=None,
+        device=None,
+        operations=ops,
+    ):
+        super().__init__()
+
+        if context_dim is not None:
+            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
+            # from omegaconf.listconfig import ListConfig
+            # if type(context_dim) == ListConfig:
+            #     context_dim = list(context_dim)
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+
+        if isinstance(num_res_blocks, int):
+            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
+        else:
+            if len(num_res_blocks) != len(channel_mult):
+                raise ValueError("provide num_res_blocks either as an int (globally constant) or "
+                                 "as a list/tuple (per-level) with the same length as channel_mult")
+            self.num_res_blocks = num_res_blocks
+
+        if disable_self_attentions is not None:
+            # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
+            assert len(disable_self_attentions) == len(channel_mult)
+        if num_attention_blocks is not None:
+            assert len(num_attention_blocks) == len(self.num_res_blocks)
+
+        transformer_depth = transformer_depth[:]
+        transformer_depth_output = transformer_depth_output[:]
+
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+        self.dtype = dtype
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.use_temporal_resblocks = use_temporal_resblock
+        self.predict_codebook_ids = n_embed is not None
+
+        self.default_num_video_frames = None
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            operations.Linear(model_channels, time_embed_dim, dtype=self.dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
+        )
+
+        if self.num_classes is not None:
+            if isinstance(self.num_classes, int):
+                self.label_emb = nn.Embedding(num_classes, time_embed_dim, dtype=self.dtype, device=device)
+            elif self.num_classes == "continuous":
+                logging.debug("setting up linear c_adm embedding layer")
+                self.label_emb = nn.Linear(1, time_embed_dim)
+            elif self.num_classes == "sequential":
+                assert adm_in_channels is not None
+                self.label_emb = nn.Sequential(
+                    nn.Sequential(
+                        operations.Linear(adm_in_channels, time_embed_dim, dtype=self.dtype, device=device),
+                        nn.SiLU(),
+                        operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
+                    )
+                )
+            else:
+                raise ValueError()
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    operations.conv_nd(dims, in_channels, model_channels, 3, padding=1, dtype=self.dtype, device=device)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+
+        def get_attention_layer(
+            ch,
+            num_heads,
+            dim_head,
+            depth=1,
+            context_dim=None,
+            use_checkpoint=False,
+            disable_self_attn=False,
+        ):
+            if use_temporal_attention:
+                return SpatialVideoTransformer(
+                    ch,
+                    num_heads,
+                    dim_head,
+                    depth=depth,
+                    context_dim=context_dim,
+                    time_context_dim=time_context_dim,
+                    dropout=dropout,
+                    ff_in=extra_ff_mix_layer,
+                    use_spatial_context=use_spatial_context,
+                    merge_strategy=merge_strategy,
+                    merge_factor=merge_factor,
+                    checkpoint=use_checkpoint,
+                    use_linear=use_linear_in_transformer,
+                    disable_self_attn=disable_self_attn,
+                    disable_temporal_crossattention=disable_temporal_crossattention,
+                    max_time_embed_period=max_ddpm_temb_period,
+                    attn_precision=attn_precision,
+                    dtype=self.dtype, device=device, operations=operations
+                )
+            else:
+                return SpatialTransformer(
+                                ch, num_heads, dim_head, depth=depth, context_dim=context_dim,
+                                disable_self_attn=disable_self_attn, use_linear=use_linear_in_transformer,
+                                use_checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=self.dtype, device=device, operations=operations
+                            )
+
+        def get_resblock(
+            merge_factor,
+            merge_strategy,
+            video_kernel_size,
+            ch,
+            time_embed_dim,
+            dropout,
+            out_channels,
+            dims,
+            use_checkpoint,
+            use_scale_shift_norm,
+            down=False,
+            up=False,
+            dtype=None,
+            device=None,
+            operations=ops
+        ):
+            if self.use_temporal_resblocks:
+                return VideoResBlock(
+                    merge_factor=merge_factor,
+                    merge_strategy=merge_strategy,
+                    video_kernel_size=video_kernel_size,
+                    channels=ch,
+                    emb_channels=time_embed_dim,
+                    dropout=dropout,
+                    out_channels=out_channels,
+                    dims=dims,
+                    use_checkpoint=use_checkpoint,
+                    use_scale_shift_norm=use_scale_shift_norm,
+                    down=down,
+                    up=up,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations
+                )
+            else:
+                return ResBlock(
+                    channels=ch,
+                    emb_channels=time_embed_dim,
+                    dropout=dropout,
+                    out_channels=out_channels,
+                    use_checkpoint=use_checkpoint,
+                    dims=dims,
+                    use_scale_shift_norm=use_scale_shift_norm,
+                    down=down,
+                    up=up,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations
+                )
+
+        for level, mult in enumerate(channel_mult):
+            for nr in range(self.num_res_blocks[level]):
+                layers = [
+                    get_resblock(
+                        merge_factor=merge_factor,
+                        merge_strategy=merge_strategy,
+                        video_kernel_size=video_kernel_size,
+                        ch=ch,
+                        time_embed_dim=time_embed_dim,
+                        dropout=dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                        dtype=self.dtype,
+                        device=device,
+                        operations=operations,
+                    )
+                ]
+                ch = mult * model_channels
+                num_transformers = transformer_depth.pop(0)
+                if num_transformers > 0:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
+                        layers.append(get_attention_layer(
+                                ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
+                                disable_self_attn=disabled_sa, use_checkpoint=use_checkpoint)
+                        )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        get_resblock(
+                            merge_factor=merge_factor,
+                            merge_strategy=merge_strategy,
+                            video_kernel_size=video_kernel_size,
+                            ch=ch,
+                            time_embed_dim=time_embed_dim,
+                            dropout=dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                            dtype=self.dtype,
+                            device=device,
+                            operations=operations
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            #num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        mid_block = [
+            get_resblock(
+                merge_factor=merge_factor,
+                merge_strategy=merge_strategy,
+                video_kernel_size=video_kernel_size,
+                ch=ch,
+                time_embed_dim=time_embed_dim,
+                dropout=dropout,
+                out_channels=None,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+                dtype=self.dtype,
+                device=device,
+                operations=operations
+            )]
+
+        self.middle_block = None
+        if transformer_depth_middle >= -1:
+            if transformer_depth_middle >= 0:
+                mid_block += [get_attention_layer(  # always uses a self-attn
+                                ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim,
+                                disable_self_attn=disable_middle_self_attn, use_checkpoint=use_checkpoint
+                            ),
+                get_resblock(
+                    merge_factor=merge_factor,
+                    merge_strategy=merge_strategy,
+                    video_kernel_size=video_kernel_size,
+                    ch=ch,
+                    time_embed_dim=time_embed_dim,
+                    dropout=dropout,
+                    out_channels=None,
+                    dims=dims,
+                    use_checkpoint=use_checkpoint,
+                    use_scale_shift_norm=use_scale_shift_norm,
+                    dtype=self.dtype,
+                    device=device,
+                    operations=operations
+                )]
+            self.middle_block = TimestepEmbedSequential(*mid_block)
+        self._feature_size += ch
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(self.num_res_blocks[level] + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    get_resblock(
+                        merge_factor=merge_factor,
+                        merge_strategy=merge_strategy,
+                        video_kernel_size=video_kernel_size,
+                        ch=ch + ich,
+                        time_embed_dim=time_embed_dim,
+                        dropout=dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                        dtype=self.dtype,
+                        device=device,
+                        operations=operations
+                    )
+                ]
+                ch = model_channels * mult
+                num_transformers = transformer_depth_output.pop()
+                if num_transformers > 0:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or i < num_attention_blocks[level]:
+                        layers.append(
+                            get_attention_layer(
+                                ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
+                                disable_self_attn=disabled_sa, use_checkpoint=use_checkpoint
+                            )
+                        )
+                if level and i == self.num_res_blocks[level]:
+                    out_ch = ch
+                    layers.append(
+                        get_resblock(
+                            merge_factor=merge_factor,
+                            merge_strategy=merge_strategy,
+                            video_kernel_size=video_kernel_size,
+                            ch=ch,
+                            time_embed_dim=time_embed_dim,
+                            dropout=dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                            dtype=self.dtype,
+                            device=device,
+                            operations=operations
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+
+        self.out = nn.Sequential(
+            operations.GroupNorm(32, ch, dtype=self.dtype, device=device),
+            nn.SiLU(),
+            operations.conv_nd(dims, model_channels, out_channels, 3, padding=1, dtype=self.dtype, device=device),
+        )
+        if self.predict_codebook_ids:
+            self.id_predictor = nn.Sequential(
+            operations.GroupNorm(32, ch, dtype=self.dtype, device=device),
+            operations.conv_nd(dims, model_channels, n_embed, 1, dtype=self.dtype, device=device),
+            #nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
+        )
+
+    def forward(self, x, timesteps=None, context=None, y=None, control=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timesteps, context, y, control, transformer_options, **kwargs)
+
+    def _forward(self, x, timesteps=None, context=None, y=None, control=None, transformer_options={}, **kwargs):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :param context: conditioning plugged in via crossattn
+        :param y: an [N] Tensor of labels, if class-conditional.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        transformer_options["original_shape"] = list(x.shape)
+        transformer_options["transformer_index"] = 0
+        transformer_patches = transformer_options.get("patches", {})
+
+        num_video_frames = kwargs.get("num_video_frames", self.default_num_video_frames)
+        image_only_indicator = kwargs.get("image_only_indicator", None)
+        time_context = kwargs.get("time_context", None)
+
+        assert (y is not None) == (
+            self.num_classes is not None
+        ), "must specify y if and only if the model is class-conditional"
+        hs = []
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
+        emb = self.time_embed(t_emb)
+
+        if "emb_patch" in transformer_patches:
+            patch = transformer_patches["emb_patch"]
+            for p in patch:
+                emb = p(emb, self.model_channels, transformer_options)
+
+        if self.num_classes is not None:
+            assert y.shape[0] == x.shape[0]
+            emb = emb + self.label_emb(y)
+
+        h = x
+        for id, module in enumerate(self.input_blocks):
+            transformer_options["block"] = ("input", id)
+            h = forward_timestep_embed(module, h, emb, context, transformer_options, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
+            h = apply_control(h, control, 'input')
+            if "input_block_patch" in transformer_patches:
+                patch = transformer_patches["input_block_patch"]
+                for p in patch:
+                    h = p(h, transformer_options)
+
+            hs.append(h)
+            if "input_block_patch_after_skip" in transformer_patches:
+                patch = transformer_patches["input_block_patch_after_skip"]
+                for p in patch:
+                    h = p(h, transformer_options)
+
+        transformer_options["block"] = ("middle", 0)
+        if self.middle_block is not None:
+            h = forward_timestep_embed(self.middle_block, h, emb, context, transformer_options, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
+        h = apply_control(h, control, 'middle')
+
+
+        for id, module in enumerate(self.output_blocks):
+            transformer_options["block"] = ("output", id)
+            hsp = hs.pop()
+            hsp = apply_control(hsp, control, 'output')
+
+            if "output_block_patch" in transformer_patches:
+                patch = transformer_patches["output_block_patch"]
+                for p in patch:
+                    h, hsp = p(h, hsp, transformer_options)
+
+            h = th.cat([h, hsp], dim=1)
+            del hsp
+            if len(hs) > 0:
+                output_shape = hs[-1].shape
+            else:
+                output_shape = None
+            h = forward_timestep_embed(module, h, emb, context, transformer_options, output_shape, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
+        h = h.type(x.dtype)
+        if self.predict_codebook_ids:
+            return self.id_predictor(h)
+        else:
+            return self.out(h)

comfy/ldm/modules/diffusionmodules/upscaling.py

+import torch
+import torch.nn as nn
+import numpy as np
+from functools import partial
+
+from .util import extract_into_tensor, make_beta_schedule
+
+
+class AbstractLowScaleModel(nn.Module):
+    # for concatenating a downsampled image to the latent representation
+    def __init__(self, noise_schedule_config=None):
+        super(AbstractLowScaleModel, self).__init__()
+        if noise_schedule_config is not None:
+            self.register_schedule(**noise_schedule_config)
+
+    def register_schedule(self, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
+                                   cosine_s=cosine_s)
+        alphas = 1. - betas
+        alphas_cumprod = np.cumprod(alphas, axis=0)
+        alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
+
+        timesteps, = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.linear_start = linear_start
+        self.linear_end = linear_end
+        assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
+
+        to_torch = partial(torch.tensor, dtype=torch.float32)
+
+        self.register_buffer('betas', to_torch(betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
+
+    def q_sample(self, x_start, t, noise=None, seed=None):
+        if noise is None:
+            if seed is None:
+                noise = torch.randn_like(x_start)
+            else:
+                noise = torch.randn(x_start.size(), dtype=x_start.dtype, layout=x_start.layout, generator=torch.manual_seed(seed)).to(x_start.device)
+        return (extract_into_tensor(self.sqrt_alphas_cumprod.to(x_start.device), t, x_start.shape) * x_start +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod.to(x_start.device), t, x_start.shape) * noise)
+
+    def forward(self, x):
+        return x, None
+
+    def decode(self, x):
+        return x
+
+
+class SimpleImageConcat(AbstractLowScaleModel):
+    # no noise level conditioning
+    def __init__(self):
+        super(SimpleImageConcat, self).__init__(noise_schedule_config=None)
+        self.max_noise_level = 0
+
+    def forward(self, x):
+        # fix to constant noise level
+        return x, torch.zeros(x.shape[0], device=x.device).long()
+
+
+class ImageConcatWithNoiseAugmentation(AbstractLowScaleModel):
+    def __init__(self, noise_schedule_config, max_noise_level=1000, to_cuda=False):
+        super().__init__(noise_schedule_config=noise_schedule_config)
+        self.max_noise_level = max_noise_level
+
+    def forward(self, x, noise_level=None, seed=None):
+        if noise_level is None:
+            noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long()
+        else:
+            assert isinstance(noise_level, torch.Tensor)
+        z = self.q_sample(x, noise_level, seed=seed)
+        return z, noise_level
+
+
+

comfy/ldm/modules/diffusionmodules/util.py

+# adopted from
+# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+# and
+# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+# and
+# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
+#
+# thanks!
+
+
+import math
+import logging
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import repeat, rearrange
+
+from comfy.ldm.util import instantiate_from_config
+
+class AlphaBlender(nn.Module):
+    strategies = ["learned", "fixed", "learned_with_images"]
+
+    def __init__(
+        self,
+        alpha: float,
+        merge_strategy: str = "learned_with_images",
+        rearrange_pattern: str = "b t -> (b t) 1 1",
+    ):
+        super().__init__()
+        self.merge_strategy = merge_strategy
+        self.rearrange_pattern = rearrange_pattern
+
+        assert (
+            merge_strategy in self.strategies
+        ), f"merge_strategy needs to be in {self.strategies}"
+
+        if self.merge_strategy == "fixed":
+            self.register_buffer("mix_factor", torch.Tensor([alpha]))
+        elif (
+            self.merge_strategy == "learned"
+            or self.merge_strategy == "learned_with_images"
+        ):
+            self.register_parameter(
+                "mix_factor", torch.nn.Parameter(torch.Tensor([alpha]))
+            )
+        else:
+            raise ValueError(f"unknown merge strategy {self.merge_strategy}")
+
+    def get_alpha(self, image_only_indicator: torch.Tensor, device) -> torch.Tensor:
+        # skip_time_mix = rearrange(repeat(skip_time_mix, 'b -> (b t) () () ()', t=t), '(b t) 1 ... -> b 1 t ...', t=t)
+        if self.merge_strategy == "fixed":
+            # make shape compatible
+            # alpha = repeat(self.mix_factor, '1 -> b () t  () ()', t=t, b=bs)
+            alpha = self.mix_factor.to(device)
+        elif self.merge_strategy == "learned":
+            alpha = torch.sigmoid(self.mix_factor.to(device))
+            # make shape compatible
+            # alpha = repeat(alpha, '1 -> s () ()', s = t * bs)
+        elif self.merge_strategy == "learned_with_images":
+            if image_only_indicator is None:
+                alpha = rearrange(torch.sigmoid(self.mix_factor.to(device)), "... -> ... 1")
+            else:
+                alpha = torch.where(
+                    image_only_indicator.bool(),
+                    torch.ones(1, 1, device=image_only_indicator.device),
+                    rearrange(torch.sigmoid(self.mix_factor.to(image_only_indicator.device)), "... -> ... 1"),
+                )
+            alpha = rearrange(alpha, self.rearrange_pattern)
+            # make shape compatible
+            # alpha = repeat(alpha, '1 -> s () ()', s = t * bs)
+        else:
+            raise NotImplementedError()
+        return alpha
+
+    def forward(
+        self,
+        x_spatial,
+        x_temporal,
+        image_only_indicator=None,
+    ) -> torch.Tensor:
+        alpha = self.get_alpha(image_only_indicator, x_spatial.device)
+        x = (
+            alpha.to(x_spatial.dtype) * x_spatial
+            + (1.0 - alpha).to(x_spatial.dtype) * x_temporal
+        )
+        return x
+
+
+def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+    if schedule == "linear":
+        betas = (
+                torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
+        )
+
+    elif schedule == "cosine":
+        timesteps = (
+                torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
+        )
+        alphas = timesteps / (1 + cosine_s) * np.pi / 2
+        alphas = torch.cos(alphas).pow(2)
+        alphas = alphas / alphas[0]
+        betas = 1 - alphas[1:] / alphas[:-1]
+        betas = torch.clamp(betas, min=0, max=0.999)
+
+    elif schedule == "squaredcos_cap_v2":  # used for karlo prior
+        # return early
+        return betas_for_alpha_bar(
+            n_timestep,
+            lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2,
+        )
+
+    elif schedule == "sqrt_linear":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
+    elif schedule == "sqrt":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
+    else:
+        raise ValueError(f"schedule '{schedule}' unknown.")
+    return betas
+
+
+def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True):
+    if ddim_discr_method == 'uniform':
+        c = num_ddpm_timesteps // num_ddim_timesteps
+        ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
+    elif ddim_discr_method == 'quad':
+        ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int)
+    else:
+        raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"')
+
+    # assert ddim_timesteps.shape[0] == num_ddim_timesteps
+    # add one to get the final alpha values right (the ones from first scale to data during sampling)
+    steps_out = ddim_timesteps + 1
+    if verbose:
+        logging.info(f'Selected timesteps for ddim sampler: {steps_out}')
+    return steps_out
+
+
+def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
+    # select alphas for computing the variance schedule
+    alphas = alphacums[ddim_timesteps]
+    alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
+
+    # according the the formula provided in https://arxiv.org/abs/2010.02502
+    sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
+    if verbose:
+        logging.info(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')
+        logging.info(f'For the chosen value of eta, which is {eta}, '
+              f'this results in the following sigma_t schedule for ddim sampler {sigmas}')
+    return sigmas, alphas, alphas_prev
+
+
+def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function,
+    which defines the cumulative product of (1-beta) over time from t = [0,1].
+    :param num_diffusion_timesteps: the number of betas to produce.
+    :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
+                      produces the cumulative product of (1-beta) up to that
+                      part of the diffusion process.
+    :param max_beta: the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+    """
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
+    return np.array(betas)
+
+
+def extract_into_tensor(a, t, x_shape):
+    b, *_ = t.shape
+    out = a.gather(-1, t)
+    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+
+def checkpoint(func, inputs, params, flag):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    """
+    if flag:
+        args = tuple(inputs) + tuple(params)
+        return CheckpointFunction.apply(func, len(inputs), *args)
+    else:
+        return func(*inputs)
+
+
+class CheckpointFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, run_function, length, *args):
+        ctx.run_function = run_function
+        ctx.input_tensors = list(args[:length])
+        ctx.input_params = list(args[length:])
+        ctx.gpu_autocast_kwargs = {"enabled": torch.is_autocast_enabled(),
+                                   "dtype": torch.get_autocast_gpu_dtype(),
+                                   "cache_enabled": torch.is_autocast_cache_enabled()}
+        with torch.no_grad():
+            output_tensors = ctx.run_function(*ctx.input_tensors)
+        return output_tensors
+
+    @staticmethod
+    def backward(ctx, *output_grads):
+        ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
+        with torch.enable_grad(), \
+                torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs):
+            # Fixes a bug where the first op in run_function modifies the
+            # Tensor storage in place, which is not allowed for detach()'d
+            # Tensors.
+            shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
+            output_tensors = ctx.run_function(*shallow_copies)
+        input_grads = torch.autograd.grad(
+            output_tensors,
+            ctx.input_tensors + ctx.input_params,
+            output_grads,
+            allow_unused=True,
+        )
+        del ctx.input_tensors
+        del ctx.input_params
+        del output_tensors
+        return (None, None) + input_grads
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=timesteps.device) / half
+        )
+        args = timesteps[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    else:
+        embedding = repeat(timesteps, 'b -> b d', d=dim)
+    return embedding
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def mean_flat(tensor):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+class HybridConditioner(nn.Module):
+
+    def __init__(self, c_concat_config, c_crossattn_config):
+        super().__init__()
+        self.concat_conditioner = instantiate_from_config(c_concat_config)
+        self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
+
+    def forward(self, c_concat, c_crossattn):
+        c_concat = self.concat_conditioner(c_concat)
+        c_crossattn = self.crossattn_conditioner(c_crossattn)
+        return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]}
+
+
+def noise_like(shape, device, repeat=False):
+    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
+    noise = lambda: torch.randn(shape, device=device)
+    return repeat_noise() if repeat else noise()