Initial commit

comfy/ldm/genmo/joint_model/rope_mixed.py

+#original code from https://github.com/genmoai/models under apache 2.0 license
+
+# import functools
+import math
+
+import torch
+
+
+def centers(start: float, stop, num, dtype=None, device=None):
+    """linspace through bin centers.
+
+    Args:
+        start (float): Start of the range.
+        stop (float): End of the range.
+        num (int): Number of points.
+        dtype (torch.dtype): Data type of the points.
+        device (torch.device): Device of the points.
+
+    Returns:
+        centers (Tensor): Centers of the bins. Shape: (num,).
+    """
+    edges = torch.linspace(start, stop, num + 1, dtype=dtype, device=device)
+    return (edges[:-1] + edges[1:]) / 2
+
+
+# @functools.lru_cache(maxsize=1)
+def create_position_matrix(
+    T: int,
+    pH: int,
+    pW: int,
+    device: torch.device,
+    dtype: torch.dtype,
+    *,
+    target_area: float = 36864,
+):
+    """
+    Args:
+        T: int - Temporal dimension
+        pH: int - Height dimension after patchify
+        pW: int - Width dimension after patchify
+
+    Returns:
+        pos: [T * pH * pW, 3] - position matrix
+    """
+    # Create 1D tensors for each dimension
+    t = torch.arange(T, dtype=dtype)
+
+    # Positionally interpolate to area 36864.
+    # (3072x3072 frame with 16x16 patches = 192x192 latents).
+    # This automatically scales rope positions when the resolution changes.
+    # We use a large target area so the model is more sensitive
+    # to changes in the learned pos_frequencies matrix.
+    scale = math.sqrt(target_area / (pW * pH))
+    w = centers(-pW * scale / 2, pW * scale / 2, pW)
+    h = centers(-pH * scale / 2, pH * scale / 2, pH)
+
+    # Use meshgrid to create 3D grids
+    grid_t, grid_h, grid_w = torch.meshgrid(t, h, w, indexing="ij")
+
+    # Stack and reshape the grids.
+    pos = torch.stack([grid_t, grid_h, grid_w], dim=-1)  # [T, pH, pW, 3]
+    pos = pos.view(-1, 3)  # [T * pH * pW, 3]
+    pos = pos.to(dtype=dtype, device=device)
+
+    return pos
+
+
+def compute_mixed_rotation(
+    freqs: torch.Tensor,
+    pos: torch.Tensor,
+):
+    """
+    Project each 3-dim position into per-head, per-head-dim 1D frequencies.
+
+    Args:
+        freqs: [3, num_heads, num_freqs] - learned rotation frequency (for t, row, col) for each head position
+        pos: [N, 3] - position of each token
+        num_heads: int
+
+    Returns:
+        freqs_cos: [N, num_heads, num_freqs] - cosine components
+        freqs_sin: [N, num_heads, num_freqs] - sine components
+    """
+    assert freqs.ndim == 3
+    freqs_sum = torch.einsum("Nd,dhf->Nhf", pos.to(freqs), freqs)
+    freqs_cos = torch.cos(freqs_sum)
+    freqs_sin = torch.sin(freqs_sum)
+    return freqs_cos, freqs_sin

comfy/ldm/genmo/joint_model/temporal_rope.py

+#original code from https://github.com/genmoai/models under apache 2.0 license
+
+# Based on Llama3 Implementation.
+import torch
+
+
+def apply_rotary_emb_qk_real(
+    xqk: torch.Tensor,
+    freqs_cos: torch.Tensor,
+    freqs_sin: torch.Tensor,
+) -> torch.Tensor:
+    """
+    Apply rotary embeddings to input tensors using the given frequency tensor without complex numbers.
+
+    Args:
+        xqk (torch.Tensor): Query and/or Key tensors to apply rotary embeddings. Shape: (B, S, *, num_heads, D)
+                            Can be either just query or just key, or both stacked along some batch or * dim.
+        freqs_cos (torch.Tensor): Precomputed cosine frequency tensor.
+        freqs_sin (torch.Tensor): Precomputed sine frequency tensor.
+
+    Returns:
+        torch.Tensor: The input tensor with rotary embeddings applied.
+    """
+    # Split the last dimension into even and odd parts
+    xqk_even = xqk[..., 0::2]
+    xqk_odd = xqk[..., 1::2]
+
+    # Apply rotation
+    cos_part = (xqk_even * freqs_cos - xqk_odd * freqs_sin).type_as(xqk)
+    sin_part = (xqk_even * freqs_sin + xqk_odd * freqs_cos).type_as(xqk)
+
+    # Interleave the results back into the original shape
+    out = torch.stack([cos_part, sin_part], dim=-1).flatten(-2)
+    return out

comfy/ldm/genmo/joint_model/utils.py

+#original code from https://github.com/genmoai/models under apache 2.0 license
+#adapted to ComfyUI
+
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def modulate(x, shift, scale):
+    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+
+def pool_tokens(x: torch.Tensor, mask: torch.Tensor, *, keepdim=False) -> torch.Tensor:
+    """
+    Pool tokens in x using mask.
+
+    NOTE: We assume x does not require gradients.
+
+    Args:
+        x: (B, L, D) tensor of tokens.
+        mask: (B, L) boolean tensor indicating which tokens are not padding.
+
+    Returns:
+        pooled: (B, D) tensor of pooled tokens.
+    """
+    assert x.size(1) == mask.size(1)  # Expected mask to have same length as tokens.
+    assert x.size(0) == mask.size(0)  # Expected mask to have same batch size as tokens.
+    mask = mask[:, :, None].to(dtype=x.dtype)
+    mask = mask / mask.sum(dim=1, keepdim=True).clamp(min=1)
+    pooled = (x * mask).sum(dim=1, keepdim=keepdim)
+    return pooled
+
+
+class AttentionPool(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        output_dim: int = None,
+        device: Optional[torch.device] = None,
+        dtype=None,
+        operations=None,
+    ):
+        """
+        Args:
+            spatial_dim (int): Number of tokens in sequence length.
+            embed_dim (int): Dimensionality of input tokens.
+            num_heads (int): Number of attention heads.
+            output_dim (int): Dimensionality of output tokens. Defaults to embed_dim.
+        """
+        super().__init__()
+        self.num_heads = num_heads
+        self.to_kv = operations.Linear(embed_dim, 2 * embed_dim, device=device, dtype=dtype)
+        self.to_q = operations.Linear(embed_dim, embed_dim, device=device, dtype=dtype)
+        self.to_out = operations.Linear(embed_dim, output_dim or embed_dim, device=device, dtype=dtype)
+
+    def forward(self, x, mask):
+        """
+        Args:
+            x (torch.Tensor): (B, L, D) tensor of input tokens.
+            mask (torch.Tensor): (B, L) boolean tensor indicating which tokens are not padding.
+
+        NOTE: We assume x does not require gradients.
+
+        Returns:
+            x (torch.Tensor): (B, D) tensor of pooled tokens.
+        """
+        D = x.size(2)
+
+        # Construct attention mask, shape: (B, 1, num_queries=1, num_keys=1+L).
+        attn_mask = mask[:, None, None, :].bool()  # (B, 1, 1, L).
+        attn_mask = F.pad(attn_mask, (1, 0), value=True)  # (B, 1, 1, 1+L).
+
+        # Average non-padding token features. These will be used as the query.
+        x_pool = pool_tokens(x, mask, keepdim=True)  # (B, 1, D)
+
+        # Concat pooled features to input sequence.
+        x = torch.cat([x_pool, x], dim=1)  # (B, L+1, D)
+
+        # Compute queries, keys, values. Only the mean token is used to create a query.
+        kv = self.to_kv(x)  # (B, L+1, 2 * D)
+        q = self.to_q(x[:, 0])  # (B, D)
+
+        # Extract heads.
+        head_dim = D // self.num_heads
+        kv = kv.unflatten(2, (2, self.num_heads, head_dim))  # (B, 1+L, 2, H, head_dim)
+        kv = kv.transpose(1, 3)  # (B, H, 2, 1+L, head_dim)
+        k, v = kv.unbind(2)  # (B, H, 1+L, head_dim)
+        q = q.unflatten(1, (self.num_heads, head_dim))  # (B, H, head_dim)
+        q = q.unsqueeze(2)  # (B, H, 1, head_dim)
+
+        # Compute attention.
+        x = F.scaled_dot_product_attention(
+            q, k, v, attn_mask=attn_mask, dropout_p=0.0
+        )  # (B, H, 1, head_dim)
+
+        # Concatenate heads and run output.
+        x = x.squeeze(2).flatten(1, 2)  # (B, D = H * head_dim)
+        x = self.to_out(x)
+        return x

comfy/ldm/genmo/vae/model.py

+#original code from https://github.com/genmoai/models under apache 2.0 license
+#adapted to ComfyUI
+
+from typing import List, Optional, Tuple, Union
+from functools import partial
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+from comfy.ldm.modules.attention import optimized_attention
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+# import mochi_preview.dit.joint_model.context_parallel as cp
+# from mochi_preview.vae.cp_conv import cp_pass_frames, gather_all_frames
+
+
+def cast_tuple(t, length=1):
+    return t if isinstance(t, tuple) else ((t,) * length)
+
+
+class GroupNormSpatial(ops.GroupNorm):
+    """
+    GroupNorm applied per-frame.
+    """
+
+    def forward(self, x: torch.Tensor, *, chunk_size: int = 8):
+        B, C, T, H, W = x.shape
+        x = rearrange(x, "B C T H W -> (B T) C H W")
+        # Run group norm in chunks.
+        output = torch.empty_like(x)
+        for b in range(0, B * T, chunk_size):
+            output[b : b + chunk_size] = super().forward(x[b : b + chunk_size])
+        return rearrange(output, "(B T) C H W -> B C T H W", B=B, T=T)
+
+class PConv3d(ops.Conv3d):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size: Union[int, Tuple[int, int, int]],
+        stride: Union[int, Tuple[int, int, int]],
+        causal: bool = True,
+        context_parallel: bool = True,
+        **kwargs,
+    ):
+        self.causal = causal
+        self.context_parallel = context_parallel
+        kernel_size = cast_tuple(kernel_size, 3)
+        stride = cast_tuple(stride, 3)
+        height_pad = (kernel_size[1] - 1) // 2
+        width_pad = (kernel_size[2] - 1) // 2
+
+        super().__init__(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            dilation=(1, 1, 1),
+            padding=(0, height_pad, width_pad),
+            **kwargs,
+        )
+
+    def forward(self, x: torch.Tensor):
+        # Compute padding amounts.
+        context_size = self.kernel_size[0] - 1
+        if self.causal:
+            pad_front = context_size
+            pad_back = 0
+        else:
+            pad_front = context_size // 2
+            pad_back = context_size - pad_front
+
+        # Apply padding.
+        assert self.padding_mode == "replicate"  # DEBUG
+        mode = "constant" if self.padding_mode == "zeros" else self.padding_mode
+        x = F.pad(x, (0, 0, 0, 0, pad_front, pad_back), mode=mode)
+        return super().forward(x)
+
+
+class Conv1x1(ops.Linear):
+    """*1x1 Conv implemented with a linear layer."""
+
+    def __init__(self, in_features: int, out_features: int, *args, **kwargs):
+        super().__init__(in_features, out_features, *args, **kwargs)
+
+    def forward(self, x: torch.Tensor):
+        """Forward pass.
+
+        Args:
+            x: Input tensor. Shape: [B, C, *] or [B, *, C].
+
+        Returns:
+            x: Output tensor. Shape: [B, C', *] or [B, *, C'].
+        """
+        x = x.movedim(1, -1)
+        x = super().forward(x)
+        x = x.movedim(-1, 1)
+        return x
+
+
+class DepthToSpaceTime(nn.Module):
+    def __init__(
+        self,
+        temporal_expansion: int,
+        spatial_expansion: int,
+    ):
+        super().__init__()
+        self.temporal_expansion = temporal_expansion
+        self.spatial_expansion = spatial_expansion
+
+    # When printed, this module should show the temporal and spatial expansion factors.
+    def extra_repr(self):
+        return f"texp={self.temporal_expansion}, sexp={self.spatial_expansion}"
+
+    def forward(self, x: torch.Tensor):
+        """Forward pass.
+
+        Args:
+            x: Input tensor. Shape: [B, C, T, H, W].
+
+        Returns:
+            x: Rearranged tensor. Shape: [B, C/(st*s*s), T*st, H*s, W*s].
+        """
+        x = rearrange(
+            x,
+            "B (C st sh sw) T H W -> B C (T st) (H sh) (W sw)",
+            st=self.temporal_expansion,
+            sh=self.spatial_expansion,
+            sw=self.spatial_expansion,
+        )
+
+        # cp_rank, _ = cp.get_cp_rank_size()
+        if self.temporal_expansion > 1: # and cp_rank == 0:
+            # Drop the first self.temporal_expansion - 1 frames.
+            # This is because we always want the 3x3x3 conv filter to only apply
+            # to the first frame, and the first frame doesn't need to be repeated.
+            assert all(x.shape)
+            x = x[:, :, self.temporal_expansion - 1 :]
+            assert all(x.shape)
+
+        return x
+
+
+def norm_fn(
+    in_channels: int,
+    affine: bool = True,
+):
+    return GroupNormSpatial(affine=affine, num_groups=32, num_channels=in_channels)
+
+
+class ResBlock(nn.Module):
+    """Residual block that preserves the spatial dimensions."""
+
+    def __init__(
+        self,
+        channels: int,
+        *,
+        affine: bool = True,
+        attn_block: Optional[nn.Module] = None,
+        causal: bool = True,
+        prune_bottleneck: bool = False,
+        padding_mode: str,
+        bias: bool = True,
+    ):
+        super().__init__()
+        self.channels = channels
+
+        assert causal
+        self.stack = nn.Sequential(
+            norm_fn(channels, affine=affine),
+            nn.SiLU(inplace=True),
+            PConv3d(
+                in_channels=channels,
+                out_channels=channels // 2 if prune_bottleneck else channels,
+                kernel_size=(3, 3, 3),
+                stride=(1, 1, 1),
+                padding_mode=padding_mode,
+                bias=bias,
+                causal=causal,
+            ),
+            norm_fn(channels, affine=affine),
+            nn.SiLU(inplace=True),
+            PConv3d(
+                in_channels=channels // 2 if prune_bottleneck else channels,
+                out_channels=channels,
+                kernel_size=(3, 3, 3),
+                stride=(1, 1, 1),
+                padding_mode=padding_mode,
+                bias=bias,
+                causal=causal,
+            ),
+        )
+
+        self.attn_block = attn_block if attn_block else nn.Identity()
+
+    def forward(self, x: torch.Tensor):
+        """Forward pass.
+
+        Args:
+            x: Input tensor. Shape: [B, C, T, H, W].
+        """
+        residual = x
+        x = self.stack(x)
+        x = x + residual
+        del residual
+
+        return self.attn_block(x)
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        head_dim: int = 32,
+        qkv_bias: bool = False,
+        out_bias: bool = True,
+        qk_norm: bool = True,
+    ) -> None:
+        super().__init__()
+        self.head_dim = head_dim
+        self.num_heads = dim // head_dim
+        self.qk_norm = qk_norm
+
+        self.qkv = nn.Linear(dim, 3 * dim, bias=qkv_bias)
+        self.out = nn.Linear(dim, dim, bias=out_bias)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+    ) -> torch.Tensor:
+        """Compute temporal self-attention.
+
+        Args:
+            x: Input tensor. Shape: [B, C, T, H, W].
+            chunk_size: Chunk size for large tensors.
+
+        Returns:
+            x: Output tensor. Shape: [B, C, T, H, W].
+        """
+        B, _, T, H, W = x.shape
+
+        if T == 1:
+            # No attention for single frame.
+            x = x.movedim(1, -1)  # [B, C, T, H, W] -> [B, T, H, W, C]
+            qkv = self.qkv(x)
+            _, _, x = qkv.chunk(3, dim=-1)  # Throw away queries and keys.
+            x = self.out(x)
+            return x.movedim(-1, 1)  # [B, T, H, W, C] -> [B, C, T, H, W]
+
+        # 1D temporal attention.
+        x = rearrange(x, "B C t h w -> (B h w) t C")
+        qkv = self.qkv(x)
+
+        # Input: qkv with shape [B, t, 3 * num_heads * head_dim]
+        # Output: x with shape [B, num_heads, t, head_dim]
+        q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, self.head_dim).transpose(1, 3).unbind(2)
+
+        if self.qk_norm:
+            q = F.normalize(q, p=2, dim=-1)
+            k = F.normalize(k, p=2, dim=-1)
+
+        x = optimized_attention(q, k, v, self.num_heads, skip_reshape=True)
+
+        assert x.size(0) == q.size(0)
+
+        x = self.out(x)
+        x = rearrange(x, "(B h w) t C -> B C t h w", B=B, h=H, w=W)
+        return x
+
+
+class AttentionBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        **attn_kwargs,
+    ) -> None:
+        super().__init__()
+        self.norm = norm_fn(dim)
+        self.attn = Attention(dim, **attn_kwargs)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x + self.attn(self.norm(x))
+
+
+class CausalUpsampleBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        num_res_blocks: int,
+        *,
+        temporal_expansion: int = 2,
+        spatial_expansion: int = 2,
+        **block_kwargs,
+    ):
+        super().__init__()
+
+        blocks = []
+        for _ in range(num_res_blocks):
+            blocks.append(block_fn(in_channels, **block_kwargs))
+        self.blocks = nn.Sequential(*blocks)
+
+        self.temporal_expansion = temporal_expansion
+        self.spatial_expansion = spatial_expansion
+
+        # Change channels in the final convolution layer.
+        self.proj = Conv1x1(
+            in_channels,
+            out_channels * temporal_expansion * (spatial_expansion**2),
+        )
+
+        self.d2st = DepthToSpaceTime(
+            temporal_expansion=temporal_expansion, spatial_expansion=spatial_expansion
+        )
+
+    def forward(self, x):
+        x = self.blocks(x)
+        x = self.proj(x)
+        x = self.d2st(x)
+        return x
+
+
+def block_fn(channels, *, affine: bool = True, has_attention: bool = False, **block_kwargs):
+    attn_block = AttentionBlock(channels) if has_attention else None
+    return ResBlock(channels, affine=affine, attn_block=attn_block, **block_kwargs)
+
+
+class DownsampleBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        num_res_blocks,
+        *,
+        temporal_reduction=2,
+        spatial_reduction=2,
+        **block_kwargs,
+    ):
+        """
+        Downsample block for the VAE encoder.
+
+        Args:
+            in_channels: Number of input channels.
+            out_channels: Number of output channels.
+            num_res_blocks: Number of residual blocks.
+            temporal_reduction: Temporal reduction factor.
+            spatial_reduction: Spatial reduction factor.
+        """
+        super().__init__()
+        layers = []
+
+        # Change the channel count in the strided convolution.
+        # This lets the ResBlock have uniform channel count,
+        # as in ConvNeXt.
+        assert in_channels != out_channels
+        layers.append(
+            PConv3d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=(temporal_reduction, spatial_reduction, spatial_reduction),
+                stride=(temporal_reduction, spatial_reduction, spatial_reduction),
+                # First layer in each block always uses replicate padding
+                padding_mode="replicate",
+                bias=block_kwargs["bias"],
+            )
+        )
+
+        for _ in range(num_res_blocks):
+            layers.append(block_fn(out_channels, **block_kwargs))
+
+        self.layers = nn.Sequential(*layers)
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+def add_fourier_features(inputs: torch.Tensor, start=6, stop=8, step=1):
+    num_freqs = (stop - start) // step
+    assert inputs.ndim == 5
+    C = inputs.size(1)
+
+    # Create Base 2 Fourier features.
+    freqs = torch.arange(start, stop, step, dtype=inputs.dtype, device=inputs.device)
+    assert num_freqs == len(freqs)
+    w = torch.pow(2.0, freqs) * (2 * torch.pi)  # [num_freqs]
+    C = inputs.shape[1]
+    w = w.repeat(C)[None, :, None, None, None]  # [1, C * num_freqs, 1, 1, 1]
+
+    # Interleaved repeat of input channels to match w.
+    h = inputs.repeat_interleave(num_freqs, dim=1)  # [B, C * num_freqs, T, H, W]
+    # Scale channels by frequency.
+    h = w * h
+
+    return torch.cat(
+        [
+            inputs,
+            torch.sin(h),
+            torch.cos(h),
+        ],
+        dim=1,
+    )
+
+
+class FourierFeatures(nn.Module):
+    def __init__(self, start: int = 6, stop: int = 8, step: int = 1):
+        super().__init__()
+        self.start = start
+        self.stop = stop
+        self.step = step
+
+    def forward(self, inputs):
+        """Add Fourier features to inputs.
+
+        Args:
+            inputs: Input tensor. Shape: [B, C, T, H, W]
+
+        Returns:
+            h: Output tensor. Shape: [B, (1 + 2 * num_freqs) * C, T, H, W]
+        """
+        return add_fourier_features(inputs, self.start, self.stop, self.step)
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        out_channels: int = 3,
+        latent_dim: int,
+        base_channels: int,
+        channel_multipliers: List[int],
+        num_res_blocks: List[int],
+        temporal_expansions: Optional[List[int]] = None,
+        spatial_expansions: Optional[List[int]] = None,
+        has_attention: List[bool],
+        output_norm: bool = True,
+        nonlinearity: str = "silu",
+        output_nonlinearity: str = "silu",
+        causal: bool = True,
+        **block_kwargs,
+    ):
+        super().__init__()
+        self.input_channels = latent_dim
+        self.base_channels = base_channels
+        self.channel_multipliers = channel_multipliers
+        self.num_res_blocks = num_res_blocks
+        self.output_nonlinearity = output_nonlinearity
+        assert nonlinearity == "silu"
+        assert causal
+
+        ch = [mult * base_channels for mult in channel_multipliers]
+        self.num_up_blocks = len(ch) - 1
+        assert len(num_res_blocks) == self.num_up_blocks + 2
+
+        blocks = []
+
+        first_block = [
+            ops.Conv3d(latent_dim, ch[-1], kernel_size=(1, 1, 1))
+        ]  # Input layer.
+        # First set of blocks preserve channel count.
+        for _ in range(num_res_blocks[-1]):
+            first_block.append(
+                block_fn(
+                    ch[-1],
+                    has_attention=has_attention[-1],
+                    causal=causal,
+                    **block_kwargs,
+                )
+            )
+        blocks.append(nn.Sequential(*first_block))
+
+        assert len(temporal_expansions) == len(spatial_expansions) == self.num_up_blocks
+        assert len(num_res_blocks) == len(has_attention) == self.num_up_blocks + 2
+
+        upsample_block_fn = CausalUpsampleBlock
+
+        for i in range(self.num_up_blocks):
+            block = upsample_block_fn(
+                ch[-i - 1],
+                ch[-i - 2],
+                num_res_blocks=num_res_blocks[-i - 2],
+                has_attention=has_attention[-i - 2],
+                temporal_expansion=temporal_expansions[-i - 1],
+                spatial_expansion=spatial_expansions[-i - 1],
+                causal=causal,
+                **block_kwargs,
+            )
+            blocks.append(block)
+
+        assert not output_norm
+
+        # Last block. Preserve channel count.
+        last_block = []
+        for _ in range(num_res_blocks[0]):
+            last_block.append(
+                block_fn(
+                    ch[0], has_attention=has_attention[0], causal=causal, **block_kwargs
+                )
+            )
+        blocks.append(nn.Sequential(*last_block))
+
+        self.blocks = nn.ModuleList(blocks)
+        self.output_proj = Conv1x1(ch[0], out_channels)
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x: Latent tensor. Shape: [B, input_channels, t, h, w]. Scaled [-1, 1].
+
+        Returns:
+            x: Reconstructed video tensor. Shape: [B, C, T, H, W]. Scaled to [-1, 1].
+               T + 1 = (t - 1) * 4.
+               H = h * 16, W = w * 16.
+        """
+        for block in self.blocks:
+            x = block(x)
+
+        if self.output_nonlinearity == "silu":
+            x = F.silu(x, inplace=not self.training)
+        else:
+            assert (
+                not self.output_nonlinearity
+            )  # StyleGAN3 omits the to-RGB nonlinearity.
+
+        return self.output_proj(x).contiguous()
+
+class LatentDistribution:
+    def __init__(self, mean: torch.Tensor, logvar: torch.Tensor):
+        """Initialize latent distribution.
+
+        Args:
+            mean: Mean of the distribution. Shape: [B, C, T, H, W].
+            logvar: Logarithm of variance of the distribution. Shape: [B, C, T, H, W].
+        """
+        assert mean.shape == logvar.shape
+        self.mean = mean
+        self.logvar = logvar
+
+    def sample(self, temperature=1.0, generator: torch.Generator = None, noise=None):
+        if temperature == 0.0:
+            return self.mean
+
+        if noise is None:
+            noise = torch.randn(self.mean.shape, device=self.mean.device, dtype=self.mean.dtype, generator=generator)
+        else:
+            assert noise.device == self.mean.device
+            noise = noise.to(self.mean.dtype)
+
+        if temperature != 1.0:
+            raise NotImplementedError(f"Temperature {temperature} is not supported.")
+
+        # Just Gaussian sample with no scaling of variance.
+        return noise * torch.exp(self.logvar * 0.5) + self.mean
+
+    def mode(self):
+        return self.mean
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels: int,
+        base_channels: int,
+        channel_multipliers: List[int],
+        num_res_blocks: List[int],
+        latent_dim: int,
+        temporal_reductions: List[int],
+        spatial_reductions: List[int],
+        prune_bottlenecks: List[bool],
+        has_attentions: List[bool],
+        affine: bool = True,
+        bias: bool = True,
+        input_is_conv_1x1: bool = False,
+        padding_mode: str,
+    ):
+        super().__init__()
+        self.temporal_reductions = temporal_reductions
+        self.spatial_reductions = spatial_reductions
+        self.base_channels = base_channels
+        self.channel_multipliers = channel_multipliers
+        self.num_res_blocks = num_res_blocks
+        self.latent_dim = latent_dim
+
+        self.fourier_features = FourierFeatures()
+        ch = [mult * base_channels for mult in channel_multipliers]
+        num_down_blocks = len(ch) - 1
+        assert len(num_res_blocks) == num_down_blocks + 2
+
+        layers = (
+            [ops.Conv3d(in_channels, ch[0], kernel_size=(1, 1, 1), bias=True)]
+            if not input_is_conv_1x1
+            else [Conv1x1(in_channels, ch[0])]
+        )
+
+        assert len(prune_bottlenecks) == num_down_blocks + 2
+        assert len(has_attentions) == num_down_blocks + 2
+        block = partial(block_fn, padding_mode=padding_mode, affine=affine, bias=bias)
+
+        for _ in range(num_res_blocks[0]):
+            layers.append(block(ch[0], has_attention=has_attentions[0], prune_bottleneck=prune_bottlenecks[0]))
+        prune_bottlenecks = prune_bottlenecks[1:]
+        has_attentions = has_attentions[1:]
+
+        assert len(temporal_reductions) == len(spatial_reductions) == len(ch) - 1
+        for i in range(num_down_blocks):
+            layer = DownsampleBlock(
+                ch[i],
+                ch[i + 1],
+                num_res_blocks=num_res_blocks[i + 1],
+                temporal_reduction=temporal_reductions[i],
+                spatial_reduction=spatial_reductions[i],
+                prune_bottleneck=prune_bottlenecks[i],
+                has_attention=has_attentions[i],
+                affine=affine,
+                bias=bias,
+                padding_mode=padding_mode,
+            )
+
+            layers.append(layer)
+
+        # Additional blocks.
+        for _ in range(num_res_blocks[-1]):
+            layers.append(block(ch[-1], has_attention=has_attentions[-1], prune_bottleneck=prune_bottlenecks[-1]))
+
+        self.layers = nn.Sequential(*layers)
+
+        # Output layers.
+        self.output_norm = norm_fn(ch[-1])
+        self.output_proj = Conv1x1(ch[-1], 2 * latent_dim, bias=False)
+
+    @property
+    def temporal_downsample(self):
+        return math.prod(self.temporal_reductions)
+
+    @property
+    def spatial_downsample(self):
+        return math.prod(self.spatial_reductions)
+
+    def forward(self, x) -> LatentDistribution:
+        """Forward pass.
+
+        Args:
+            x: Input video tensor. Shape: [B, C, T, H, W]. Scaled to [-1, 1]
+
+        Returns:
+            means: Latent tensor. Shape: [B, latent_dim, t, h, w]. Scaled [-1, 1].
+                   h = H // 8, w = W // 8, t - 1 = (T - 1) // 6
+            logvar: Shape: [B, latent_dim, t, h, w].
+        """
+        assert x.ndim == 5, f"Expected 5D input, got {x.shape}"
+        x = self.fourier_features(x)
+
+        x = self.layers(x)
+
+        x = self.output_norm(x)
+        x = F.silu(x, inplace=True)
+        x = self.output_proj(x)
+
+        means, logvar = torch.chunk(x, 2, dim=1)
+
+        assert means.ndim == 5
+        assert logvar.shape == means.shape
+        assert means.size(1) == self.latent_dim
+
+        return LatentDistribution(means, logvar)
+
+
+class VideoVAE(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.encoder = Encoder(
+            in_channels=15,
+            base_channels=64,
+            channel_multipliers=[1, 2, 4, 6],
+            num_res_blocks=[3, 3, 4, 6, 3],
+            latent_dim=12,
+            temporal_reductions=[1, 2, 3],
+            spatial_reductions=[2, 2, 2],
+            prune_bottlenecks=[False, False, False, False, False],
+            has_attentions=[False, True, True, True, True],
+            affine=True,
+            bias=True,
+            input_is_conv_1x1=True,
+            padding_mode="replicate"
+        )
+        self.decoder = Decoder(
+            out_channels=3,
+            base_channels=128,
+            channel_multipliers=[1, 2, 4, 6],
+            temporal_expansions=[1, 2, 3],
+            spatial_expansions=[2, 2, 2],
+            num_res_blocks=[3, 3, 4, 6, 3],
+            latent_dim=12,
+            has_attention=[False, False, False, False, False],
+            padding_mode="replicate",
+            output_norm=False,
+            nonlinearity="silu",
+            output_nonlinearity="silu",
+            causal=True,
+        )
+
+    def encode(self, x):
+        return self.encoder(x).mode()
+
+    def decode(self, x):
+        return self.decoder(x)

comfy/ldm/hidream/model.py

+from typing import Optional, Tuple, List
+
+import torch
+import torch.nn as nn
+import einops
+from einops import repeat
+
+from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps
+import torch.nn.functional as F
+
+from comfy.ldm.flux.math import apply_rope, rope
+from comfy.ldm.flux.layers import LastLayer
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+import comfy.patcher_extension
+import comfy.ldm.common_dit
+
+
+# Copied from https://github.com/black-forest-labs/flux/blob/main/src/flux/modules/layers.py
+class EmbedND(nn.Module):
+    def __init__(self, theta: int, axes_dim: List[int]):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        n_axes = ids.shape[-1]
+        emb = torch.cat(
+            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
+            dim=-3,
+        )
+        return emb.unsqueeze(2)
+
+
+class PatchEmbed(nn.Module):
+    def __init__(
+        self,
+        patch_size=2,
+        in_channels=4,
+        out_channels=1024,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.out_channels = out_channels
+        self.proj = operations.Linear(in_channels * patch_size * patch_size, out_channels, bias=True, dtype=dtype, device=device)
+
+    def forward(self, latent):
+        latent = self.proj(latent)
+        return latent
+
+
+class PooledEmbed(nn.Module):
+    def __init__(self, text_emb_dim, hidden_size, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.pooled_embedder = TimestepEmbedding(in_channels=text_emb_dim, time_embed_dim=hidden_size, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, pooled_embed):
+        return self.pooled_embedder(pooled_embed)
+
+
+class TimestepEmbed(nn.Module):
+    def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=frequency_embedding_size, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=frequency_embedding_size, time_embed_dim=hidden_size, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, timesteps, wdtype):
+        t_emb = self.time_proj(timesteps).to(dtype=wdtype)
+        t_emb = self.timestep_embedder(t_emb)
+        return t_emb
+
+
+def attention(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor):
+    return optimized_attention(query.view(query.shape[0], -1, query.shape[-1] * query.shape[-2]), key.view(key.shape[0], -1, key.shape[-1] * key.shape[-2]), value.view(value.shape[0], -1, value.shape[-1] * value.shape[-2]), query.shape[2])
+
+
+class HiDreamAttnProcessor_flashattn:
+    """Attention processor used typically in processing the SD3-like self-attention projections."""
+
+    def __call__(
+        self,
+        attn,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        rope: torch.FloatTensor = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        dtype = image_tokens.dtype
+        batch_size = image_tokens.shape[0]
+
+        query_i = attn.q_rms_norm(attn.to_q(image_tokens)).to(dtype=dtype)
+        key_i = attn.k_rms_norm(attn.to_k(image_tokens)).to(dtype=dtype)
+        value_i = attn.to_v(image_tokens)
+
+        inner_dim = key_i.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query_i = query_i.view(batch_size, -1, attn.heads, head_dim)
+        key_i = key_i.view(batch_size, -1, attn.heads, head_dim)
+        value_i = value_i.view(batch_size, -1, attn.heads, head_dim)
+        if image_tokens_masks is not None:
+            key_i = key_i * image_tokens_masks.view(batch_size, -1, 1, 1)
+
+        if not attn.single:
+            query_t = attn.q_rms_norm_t(attn.to_q_t(text_tokens)).to(dtype=dtype)
+            key_t = attn.k_rms_norm_t(attn.to_k_t(text_tokens)).to(dtype=dtype)
+            value_t = attn.to_v_t(text_tokens)
+
+            query_t = query_t.view(batch_size, -1, attn.heads, head_dim)
+            key_t = key_t.view(batch_size, -1, attn.heads, head_dim)
+            value_t = value_t.view(batch_size, -1, attn.heads, head_dim)
+
+            num_image_tokens = query_i.shape[1]
+            num_text_tokens = query_t.shape[1]
+            query = torch.cat([query_i, query_t], dim=1)
+            key = torch.cat([key_i, key_t], dim=1)
+            value = torch.cat([value_i, value_t], dim=1)
+        else:
+            query = query_i
+            key = key_i
+            value = value_i
+
+        if query.shape[-1] == rope.shape[-3] * 2:
+            query, key = apply_rope(query, key, rope)
+        else:
+            query_1, query_2 = query.chunk(2, dim=-1)
+            key_1, key_2 = key.chunk(2, dim=-1)
+            query_1, key_1 = apply_rope(query_1, key_1, rope)
+            query = torch.cat([query_1, query_2], dim=-1)
+            key = torch.cat([key_1, key_2], dim=-1)
+
+        hidden_states = attention(query, key, value)
+
+        if not attn.single:
+            hidden_states_i, hidden_states_t = torch.split(hidden_states, [num_image_tokens, num_text_tokens], dim=1)
+            hidden_states_i = attn.to_out(hidden_states_i)
+            hidden_states_t = attn.to_out_t(hidden_states_t)
+            return hidden_states_i, hidden_states_t
+        else:
+            hidden_states = attn.to_out(hidden_states)
+            return hidden_states
+
+class HiDreamAttention(nn.Module):
+    def __init__(
+        self,
+        query_dim: int,
+        heads: int = 8,
+        dim_head: int = 64,
+        upcast_attention: bool = False,
+        upcast_softmax: bool = False,
+        scale_qk: bool = True,
+        eps: float = 1e-5,
+        processor = None,
+        out_dim: int = None,
+        single: bool = False,
+        dtype=None, device=None, operations=None
+    ):
+        # super(Attention, self).__init__()
+        super().__init__()
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.query_dim = query_dim
+        self.upcast_attention = upcast_attention
+        self.upcast_softmax = upcast_softmax
+        self.out_dim = out_dim if out_dim is not None else query_dim
+
+        self.scale_qk = scale_qk
+        self.scale = dim_head**-0.5 if self.scale_qk else 1.0
+
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+        self.sliceable_head_dim = heads
+        self.single = single
+
+        linear_cls = operations.Linear
+        self.linear_cls = linear_cls
+        self.to_q = linear_cls(query_dim, self.inner_dim, dtype=dtype, device=device)
+        self.to_k = linear_cls(self.inner_dim, self.inner_dim, dtype=dtype, device=device)
+        self.to_v = linear_cls(self.inner_dim, self.inner_dim, dtype=dtype, device=device)
+        self.to_out = linear_cls(self.inner_dim, self.out_dim, dtype=dtype, device=device)
+        self.q_rms_norm = operations.RMSNorm(self.inner_dim, eps, dtype=dtype, device=device)
+        self.k_rms_norm = operations.RMSNorm(self.inner_dim, eps, dtype=dtype, device=device)
+
+        if not single:
+            self.to_q_t = linear_cls(query_dim, self.inner_dim, dtype=dtype, device=device)
+            self.to_k_t = linear_cls(self.inner_dim, self.inner_dim, dtype=dtype, device=device)
+            self.to_v_t = linear_cls(self.inner_dim, self.inner_dim, dtype=dtype, device=device)
+            self.to_out_t = linear_cls(self.inner_dim, self.out_dim, dtype=dtype, device=device)
+            self.q_rms_norm_t = operations.RMSNorm(self.inner_dim, eps, dtype=dtype, device=device)
+            self.k_rms_norm_t = operations.RMSNorm(self.inner_dim, eps, dtype=dtype, device=device)
+
+        self.processor = processor
+
+    def forward(
+        self,
+        norm_image_tokens: torch.FloatTensor,
+        image_tokens_masks: torch.FloatTensor = None,
+        norm_text_tokens: torch.FloatTensor = None,
+        rope: torch.FloatTensor = None,
+    ) -> torch.Tensor:
+        return self.processor(
+            self,
+            image_tokens = norm_image_tokens,
+            image_tokens_masks = image_tokens_masks,
+            text_tokens = norm_text_tokens,
+            rope = rope,
+        )
+
+
+class FeedForwardSwiGLU(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+        dtype=None, device=None, operations=None
+    ):
+        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 = operations.Linear(dim, hidden_dim, bias=False, dtype=dtype, device=device)
+        self.w2 = operations.Linear(hidden_dim, dim, bias=False, dtype=dtype, device=device)
+        self.w3 = operations.Linear(dim, hidden_dim, bias=False, dtype=dtype, device=device)
+
+    def forward(self, x):
+        return self.w2(torch.nn.functional.silu(self.w1(x)) * self.w3(x))
+
+
+# Modified from https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/model.py
+class MoEGate(nn.Module):
+    def __init__(self, embed_dim, num_routed_experts=4, num_activated_experts=2, aux_loss_alpha=0.01, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.top_k = num_activated_experts
+        self.n_routed_experts = num_routed_experts
+
+        self.scoring_func = 'softmax'
+        self.alpha = aux_loss_alpha
+        self.seq_aux = False
+
+        # topk selection algorithm
+        self.norm_topk_prob = False
+        self.gating_dim = embed_dim
+        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim), dtype=dtype, device=device))
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        pass
+        # import torch.nn.init  as init
+        # init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+
+    def forward(self, hidden_states):
+        bsz, seq_len, h = hidden_states.shape
+
+        ### compute gating score
+        hidden_states = hidden_states.view(-1, h)
+        logits = F.linear(hidden_states, comfy.model_management.cast_to(self.weight, dtype=hidden_states.dtype, device=hidden_states.device), None)
+        if self.scoring_func == 'softmax':
+            scores = logits.softmax(dim=-1)
+        else:
+            raise NotImplementedError(f'insupportable scoring function for MoE gating: {self.scoring_func}')
+
+        ### select top-k experts
+        topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False)
+
+        ### norm gate to sum 1
+        if self.top_k > 1 and self.norm_topk_prob:
+            denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weight = topk_weight / denominator
+
+        aux_loss = None
+        return topk_idx, topk_weight, aux_loss
+
+
+# Modified from https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/model.py
+class MOEFeedForwardSwiGLU(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        num_routed_experts: int,
+        num_activated_experts: int,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.shared_experts = FeedForwardSwiGLU(dim, hidden_dim // 2, dtype=dtype, device=device, operations=operations)
+        self.experts = nn.ModuleList([FeedForwardSwiGLU(dim, hidden_dim, dtype=dtype, device=device, operations=operations) for i in range(num_routed_experts)])
+        self.gate = MoEGate(
+            embed_dim = dim,
+            num_routed_experts = num_routed_experts,
+            num_activated_experts = num_activated_experts,
+            dtype=dtype, device=device, operations=operations
+        )
+        self.num_activated_experts = num_activated_experts
+
+    def forward(self, x):
+        wtype = x.dtype
+        identity = x
+        orig_shape = x.shape
+        topk_idx, topk_weight, aux_loss = self.gate(x)
+        x = x.view(-1, x.shape[-1])
+        flat_topk_idx = topk_idx.view(-1)
+        if True:  # self.training: # TODO: check which branch performs faster
+            x = x.repeat_interleave(self.num_activated_experts, dim=0)
+            y = torch.empty_like(x, dtype=wtype)
+            for i, expert in enumerate(self.experts):
+                y[flat_topk_idx == i] = expert(x[flat_topk_idx == i]).to(dtype=wtype)
+            y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
+            y =  y.view(*orig_shape).to(dtype=wtype)
+            #y = AddAuxiliaryLoss.apply(y, aux_loss)
+        else:
+            y = self.moe_infer(x, flat_topk_idx, topk_weight.view(-1, 1)).view(*orig_shape)
+        y = y + self.shared_experts(identity)
+        return y
+
+    @torch.no_grad()
+    def moe_infer(self, x, flat_expert_indices, flat_expert_weights):
+        expert_cache = torch.zeros_like(x)
+        idxs = flat_expert_indices.argsort()
+        tokens_per_expert = flat_expert_indices.bincount().cpu().numpy().cumsum(0)
+        token_idxs = idxs // self.num_activated_experts
+        for i, end_idx in enumerate(tokens_per_expert):
+            start_idx = 0 if i == 0 else tokens_per_expert[i-1]
+            if start_idx == end_idx:
+                continue
+            expert = self.experts[i]
+            exp_token_idx = token_idxs[start_idx:end_idx]
+            expert_tokens = x[exp_token_idx]
+            expert_out = expert(expert_tokens)
+            expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]])
+
+            # for fp16 and other dtype
+            expert_cache = expert_cache.to(expert_out.dtype)
+            expert_cache.scatter_reduce_(0, exp_token_idx.view(-1, 1).repeat(1, x.shape[-1]), expert_out, reduce='sum')
+        return expert_cache
+
+
+class TextProjection(nn.Module):
+    def __init__(self, in_features, hidden_size, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.linear = operations.Linear(in_features=in_features, out_features=hidden_size, bias=False, dtype=dtype, device=device)
+
+    def forward(self, caption):
+        hidden_states = self.linear(caption)
+        return hidden_states
+
+
+class BlockType:
+    TransformerBlock = 1
+    SingleTransformerBlock = 2
+
+
+class HiDreamImageSingleTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(dim, 6 * dim, bias=True, dtype=dtype, device=device)
+        )
+
+        # 1. Attention
+        self.norm1_i = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        self.attn1 = HiDreamAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            processor = HiDreamAttnProcessor_flashattn(),
+            single = True,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        # 3. Feed-forward
+        self.norm3_i = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        if num_routed_experts > 0:
+            self.ff_i = MOEFeedForwardSwiGLU(
+                dim = dim,
+                hidden_dim = 4 * dim,
+                num_routed_experts = num_routed_experts,
+                num_activated_experts = num_activated_experts,
+                dtype=dtype, device=device, operations=operations
+            )
+        else:
+            self.ff_i = FeedForwardSwiGLU(dim = dim, hidden_dim = 4 * dim, dtype=dtype, device=device, operations=operations)
+
+    def forward(
+        self,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        adaln_input: Optional[torch.FloatTensor] = None,
+        rope: torch.FloatTensor = None,
+
+    ) -> torch.FloatTensor:
+        wtype = image_tokens.dtype
+        shift_msa_i, scale_msa_i, gate_msa_i, shift_mlp_i, scale_mlp_i, gate_mlp_i = \
+            self.adaLN_modulation(adaln_input)[:,None].chunk(6, dim=-1)
+
+        # 1. MM-Attention
+        norm_image_tokens = self.norm1_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_msa_i) + shift_msa_i
+        attn_output_i = self.attn1(
+            norm_image_tokens,
+            image_tokens_masks,
+            rope = rope,
+        )
+        image_tokens = gate_msa_i * attn_output_i + image_tokens
+
+        # 2. Feed-forward
+        norm_image_tokens = self.norm3_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_mlp_i) + shift_mlp_i
+        ff_output_i = gate_mlp_i * self.ff_i(norm_image_tokens.to(dtype=wtype))
+        image_tokens = ff_output_i + image_tokens
+        return image_tokens
+
+
+class HiDreamImageTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(dim, 12 * dim, bias=True, dtype=dtype, device=device)
+        )
+        # nn.init.zeros_(self.adaLN_modulation[1].weight)
+        # nn.init.zeros_(self.adaLN_modulation[1].bias)
+
+        # 1. Attention
+        self.norm1_i = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        self.norm1_t = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        self.attn1 = HiDreamAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            processor = HiDreamAttnProcessor_flashattn(),
+            single = False,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        # 3. Feed-forward
+        self.norm3_i = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        if num_routed_experts > 0:
+            self.ff_i = MOEFeedForwardSwiGLU(
+                dim = dim,
+                hidden_dim = 4 * dim,
+                num_routed_experts = num_routed_experts,
+                num_activated_experts = num_activated_experts,
+                dtype=dtype, device=device, operations=operations
+            )
+        else:
+            self.ff_i = FeedForwardSwiGLU(dim = dim, hidden_dim = 4 * dim, dtype=dtype, device=device, operations=operations)
+        self.norm3_t = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False)
+        self.ff_t = FeedForwardSwiGLU(dim = dim, hidden_dim = 4 * dim, dtype=dtype, device=device, operations=operations)
+
+    def forward(
+        self,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        adaln_input: Optional[torch.FloatTensor] = None,
+        rope: torch.FloatTensor = None,
+    ) -> torch.FloatTensor:
+        wtype = image_tokens.dtype
+        shift_msa_i, scale_msa_i, gate_msa_i, shift_mlp_i, scale_mlp_i, gate_mlp_i, \
+        shift_msa_t, scale_msa_t, gate_msa_t, shift_mlp_t, scale_mlp_t, gate_mlp_t = \
+            self.adaLN_modulation(adaln_input)[:,None].chunk(12, dim=-1)
+
+        # 1. MM-Attention
+        norm_image_tokens = self.norm1_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_msa_i) + shift_msa_i
+        norm_text_tokens = self.norm1_t(text_tokens).to(dtype=wtype)
+        norm_text_tokens = norm_text_tokens * (1 + scale_msa_t) + shift_msa_t
+
+        attn_output_i, attn_output_t = self.attn1(
+            norm_image_tokens,
+            image_tokens_masks,
+            norm_text_tokens,
+            rope = rope,
+        )
+
+        image_tokens = gate_msa_i * attn_output_i + image_tokens
+        text_tokens = gate_msa_t * attn_output_t + text_tokens
+
+        # 2. Feed-forward
+        norm_image_tokens = self.norm3_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_mlp_i) + shift_mlp_i
+        norm_text_tokens = self.norm3_t(text_tokens).to(dtype=wtype)
+        norm_text_tokens = norm_text_tokens * (1 + scale_mlp_t) + shift_mlp_t
+
+        ff_output_i = gate_mlp_i * self.ff_i(norm_image_tokens)
+        ff_output_t = gate_mlp_t * self.ff_t(norm_text_tokens)
+        image_tokens = ff_output_i + image_tokens
+        text_tokens = ff_output_t + text_tokens
+        return image_tokens, text_tokens
+
+
+class HiDreamImageBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        block_type: BlockType = BlockType.TransformerBlock,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        block_classes = {
+            BlockType.TransformerBlock: HiDreamImageTransformerBlock,
+            BlockType.SingleTransformerBlock: HiDreamImageSingleTransformerBlock,
+        }
+        self.block = block_classes[block_type](
+            dim,
+            num_attention_heads,
+            attention_head_dim,
+            num_routed_experts,
+            num_activated_experts,
+            dtype=dtype, device=device, operations=operations
+        )
+
+    def forward(
+        self,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        adaln_input: torch.FloatTensor = None,
+        rope: torch.FloatTensor = None,
+    ) -> torch.FloatTensor:
+        return self.block(
+            image_tokens,
+            image_tokens_masks,
+            text_tokens,
+            adaln_input,
+            rope,
+        )
+
+
+class HiDreamImageTransformer2DModel(nn.Module):
+    def __init__(
+        self,
+        patch_size: Optional[int] = None,
+        in_channels: int = 64,
+        out_channels: Optional[int] = None,
+        num_layers: int = 16,
+        num_single_layers: int = 32,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 20,
+        caption_channels: List[int] = None,
+        text_emb_dim: int = 2048,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        axes_dims_rope: Tuple[int, int] = (32, 32),
+        max_resolution: Tuple[int, int] = (128, 128),
+        llama_layers: List[int] = None,
+        image_model=None,
+        dtype=None, device=None, operations=None
+    ):
+        self.patch_size = patch_size
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.num_layers = num_layers
+        self.num_single_layers = num_single_layers
+
+        self.gradient_checkpointing = False
+
+        super().__init__()
+        self.dtype = dtype
+        self.out_channels = out_channels or in_channels
+        self.inner_dim = self.num_attention_heads * self.attention_head_dim
+        self.llama_layers = llama_layers
+
+        self.t_embedder = TimestepEmbed(self.inner_dim, dtype=dtype, device=device, operations=operations)
+        self.p_embedder = PooledEmbed(text_emb_dim, self.inner_dim, dtype=dtype, device=device, operations=operations)
+        self.x_embedder = PatchEmbed(
+            patch_size = patch_size,
+            in_channels = in_channels,
+            out_channels = self.inner_dim,
+            dtype=dtype, device=device, operations=operations
+        )
+        self.pe_embedder = EmbedND(theta=10000, axes_dim=axes_dims_rope)
+
+        self.double_stream_blocks = nn.ModuleList(
+            [
+                HiDreamImageBlock(
+                    dim = self.inner_dim,
+                    num_attention_heads = self.num_attention_heads,
+                    attention_head_dim = self.attention_head_dim,
+                    num_routed_experts = num_routed_experts,
+                    num_activated_experts = num_activated_experts,
+                    block_type = BlockType.TransformerBlock,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for i in range(self.num_layers)
+            ]
+        )
+
+        self.single_stream_blocks = nn.ModuleList(
+            [
+                HiDreamImageBlock(
+                    dim = self.inner_dim,
+                    num_attention_heads = self.num_attention_heads,
+                    attention_head_dim = self.attention_head_dim,
+                    num_routed_experts = num_routed_experts,
+                    num_activated_experts = num_activated_experts,
+                    block_type = BlockType.SingleTransformerBlock,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for i in range(self.num_single_layers)
+            ]
+        )
+
+        self.final_layer = LastLayer(self.inner_dim, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations)
+
+        caption_channels = [caption_channels[1], ] * (num_layers + num_single_layers) + [caption_channels[0], ]
+        caption_projection = []
+        for caption_channel in caption_channels:
+            caption_projection.append(TextProjection(in_features=caption_channel, hidden_size=self.inner_dim, dtype=dtype, device=device, operations=operations))
+        self.caption_projection = nn.ModuleList(caption_projection)
+        self.max_seq = max_resolution[0] * max_resolution[1] // (patch_size * patch_size)
+
+    def expand_timesteps(self, timesteps, batch_size, device):
+        if not torch.is_tensor(timesteps):
+            is_mps = device.type == "mps"
+            if isinstance(timesteps, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(device)
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(batch_size)
+        return timesteps
+
+    def unpatchify(self, x: torch.Tensor, img_sizes: List[Tuple[int, int]]) -> List[torch.Tensor]:
+        x_arr = []
+        for i, img_size in enumerate(img_sizes):
+            pH, pW = img_size
+            x_arr.append(
+                einops.rearrange(x[i, :pH*pW].reshape(1, pH, pW, -1), 'B H W (p1 p2 C) -> B C (H p1) (W p2)',
+                    p1=self.patch_size, p2=self.patch_size)
+            )
+        x = torch.cat(x_arr, dim=0)
+        return x
+
+    def patchify(self, x, max_seq, img_sizes=None):
+        pz2 = self.patch_size * self.patch_size
+        if isinstance(x, torch.Tensor):
+            B = x.shape[0]
+            device = x.device
+            dtype = x.dtype
+        else:
+            B = len(x)
+            device = x[0].device
+            dtype = x[0].dtype
+        x_masks = torch.zeros((B, max_seq), dtype=dtype, device=device)
+
+        if img_sizes is not None:
+            for i, img_size in enumerate(img_sizes):
+                x_masks[i, 0:img_size[0] * img_size[1]] = 1
+            x = einops.rearrange(x, 'B C S p -> B S (p C)', p=pz2)
+        elif isinstance(x, torch.Tensor):
+            pH, pW = x.shape[-2] // self.patch_size, x.shape[-1] // self.patch_size
+            x = einops.rearrange(x, 'B C (H p1) (W p2) -> B (H W) (p1 p2 C)', p1=self.patch_size, p2=self.patch_size)
+            img_sizes = [[pH, pW]] * B
+            x_masks = None
+        else:
+            raise NotImplementedError
+        return x, x_masks, img_sizes
+
+    def forward(self,
+        x: torch.Tensor,
+        t: torch.Tensor,
+        y: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        encoder_hidden_states_llama3=None,
+        image_cond=None,
+        control = None,
+        transformer_options = {},
+    ):
+        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, t, y, context, encoder_hidden_states_llama3, image_cond, control, transformer_options)
+
+    def _forward(
+        self,
+        x: torch.Tensor,
+        t: torch.Tensor,
+        y: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        encoder_hidden_states_llama3=None,
+        image_cond=None,
+        control = None,
+        transformer_options = {},
+    ) -> torch.Tensor:
+        bs, c, h, w = x.shape
+        if image_cond is not None:
+            x = torch.cat([x, image_cond], dim=-1)
+        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+        timesteps = t
+        pooled_embeds = y
+        T5_encoder_hidden_states = context
+
+        img_sizes = None
+
+        # spatial forward
+        batch_size = hidden_states.shape[0]
+        hidden_states_type = hidden_states.dtype
+
+        # 0. time
+        timesteps = self.expand_timesteps(timesteps, batch_size, hidden_states.device)
+        timesteps = self.t_embedder(timesteps, hidden_states_type)
+        p_embedder = self.p_embedder(pooled_embeds)
+        adaln_input = timesteps + p_embedder
+
+        hidden_states, image_tokens_masks, img_sizes = self.patchify(hidden_states, self.max_seq, img_sizes)
+        if image_tokens_masks is None:
+            pH, pW = img_sizes[0]
+            img_ids = torch.zeros(pH, pW, 3, device=hidden_states.device)
+            img_ids[..., 1] = img_ids[..., 1] + torch.arange(pH, device=hidden_states.device)[:, None]
+            img_ids[..., 2] = img_ids[..., 2] + torch.arange(pW, device=hidden_states.device)[None, :]
+            img_ids = repeat(img_ids, "h w c -> b (h w) c", b=batch_size)
+        hidden_states = self.x_embedder(hidden_states)
+
+        # T5_encoder_hidden_states = encoder_hidden_states[0]
+        encoder_hidden_states = encoder_hidden_states_llama3.movedim(1, 0)
+        encoder_hidden_states = [encoder_hidden_states[k] for k in self.llama_layers]
+
+        if self.caption_projection is not None:
+            new_encoder_hidden_states = []
+            for i, enc_hidden_state in enumerate(encoder_hidden_states):
+                enc_hidden_state = self.caption_projection[i](enc_hidden_state)
+                enc_hidden_state = enc_hidden_state.view(batch_size, -1, hidden_states.shape[-1])
+                new_encoder_hidden_states.append(enc_hidden_state)
+            encoder_hidden_states = new_encoder_hidden_states
+            T5_encoder_hidden_states = self.caption_projection[-1](T5_encoder_hidden_states)
+            T5_encoder_hidden_states = T5_encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+            encoder_hidden_states.append(T5_encoder_hidden_states)
+
+        txt_ids = torch.zeros(
+            batch_size,
+            encoder_hidden_states[-1].shape[1] + encoder_hidden_states[-2].shape[1] + encoder_hidden_states[0].shape[1],
+            3,
+            device=img_ids.device, dtype=img_ids.dtype
+        )
+        ids = torch.cat((img_ids, txt_ids), dim=1)
+        rope = self.pe_embedder(ids)
+
+        # 2. Blocks
+        block_id = 0
+        initial_encoder_hidden_states = torch.cat([encoder_hidden_states[-1], encoder_hidden_states[-2]], dim=1)
+        initial_encoder_hidden_states_seq_len = initial_encoder_hidden_states.shape[1]
+        for bid, block in enumerate(self.double_stream_blocks):
+            cur_llama31_encoder_hidden_states = encoder_hidden_states[block_id]
+            cur_encoder_hidden_states = torch.cat([initial_encoder_hidden_states, cur_llama31_encoder_hidden_states], dim=1)
+            hidden_states, initial_encoder_hidden_states = block(
+                image_tokens = hidden_states,
+                image_tokens_masks = image_tokens_masks,
+                text_tokens = cur_encoder_hidden_states,
+                adaln_input = adaln_input,
+                rope = rope,
+            )
+            initial_encoder_hidden_states = initial_encoder_hidden_states[:, :initial_encoder_hidden_states_seq_len]
+            block_id += 1
+
+        image_tokens_seq_len = hidden_states.shape[1]
+        hidden_states = torch.cat([hidden_states, initial_encoder_hidden_states], dim=1)
+        hidden_states_seq_len = hidden_states.shape[1]
+        if image_tokens_masks is not None:
+            encoder_attention_mask_ones = torch.ones(
+                (batch_size, initial_encoder_hidden_states.shape[1] + cur_llama31_encoder_hidden_states.shape[1]),
+                device=image_tokens_masks.device, dtype=image_tokens_masks.dtype
+            )
+            image_tokens_masks = torch.cat([image_tokens_masks, encoder_attention_mask_ones], dim=1)
+
+        for bid, block in enumerate(self.single_stream_blocks):
+            cur_llama31_encoder_hidden_states = encoder_hidden_states[block_id]
+            hidden_states = torch.cat([hidden_states, cur_llama31_encoder_hidden_states], dim=1)
+            hidden_states = block(
+                image_tokens=hidden_states,
+                image_tokens_masks=image_tokens_masks,
+                text_tokens=None,
+                adaln_input=adaln_input,
+                rope=rope,
+            )
+            hidden_states = hidden_states[:, :hidden_states_seq_len]
+            block_id += 1
+
+        hidden_states = hidden_states[:, :image_tokens_seq_len, ...]
+        output = self.final_layer(hidden_states, adaln_input)
+        output = self.unpatchify(output, img_sizes)
+        return -output[:, :, :h, :w]

comfy/ldm/hunyuan3d/model.py

+import torch
+from torch import nn
+from comfy.ldm.flux.layers import (
+    DoubleStreamBlock,
+    LastLayer,
+    MLPEmbedder,
+    SingleStreamBlock,
+    timestep_embedding,
+)
+import comfy.patcher_extension
+
+
+class Hunyuan3Dv2(nn.Module):
+    def __init__(
+        self,
+        in_channels=64,
+        context_in_dim=1536,
+        hidden_size=1024,
+        mlp_ratio=4.0,
+        num_heads=16,
+        depth=16,
+        depth_single_blocks=32,
+        qkv_bias=True,
+        guidance_embed=False,
+        image_model=None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.dtype = dtype
+
+        if hidden_size % num_heads != 0:
+            raise ValueError(
+                f"Hidden size {hidden_size} must be divisible by num_heads {num_heads}"
+            )
+
+        self.max_period = 1000  # While reimplementing the model I noticed that they messed up. This 1000 value was meant to be the time_factor but they set the max_period instead
+        self.latent_in = operations.Linear(in_channels, hidden_size, bias=True, dtype=dtype, device=device)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=hidden_size, dtype=dtype, device=device, operations=operations)
+        self.guidance_in = (
+            MLPEmbedder(in_dim=256, hidden_dim=hidden_size, dtype=dtype, device=device, operations=operations) if guidance_embed else None
+        )
+        self.cond_in = operations.Linear(context_in_dim, hidden_size, dtype=dtype, device=device)
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(depth)
+            ]
+        )
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(depth_single_blocks)
+            ]
+        )
+        self.final_layer = LastLayer(hidden_size, 1, in_channels, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, x, timestep, context, guidance=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, timestep, context, guidance, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, guidance=None, transformer_options={}, **kwargs):
+        x = x.movedim(-1, -2)
+        timestep = 1.0 - timestep
+        txt = context
+        img = self.latent_in(x)
+
+        vec = self.time_in(timestep_embedding(timestep, 256, self.max_period).to(dtype=img.dtype))
+        if self.guidance_in is not None:
+            if guidance is not None:
+                vec = vec + self.guidance_in(timestep_embedding(guidance, 256, self.max_period).to(img.dtype))
+
+        txt = self.cond_in(txt)
+        pe = None
+        attn_mask = None
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.double_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"], out["txt"] = block(img=args["img"],
+                                                   txt=args["txt"],
+                                                   vec=args["vec"],
+                                                   pe=args["pe"],
+                                                   attn_mask=args.get("attn_mask"))
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": img,
+                                                           "txt": txt,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask},
+                                                          {"original_block": block_wrap})
+                txt = out["txt"]
+                img = out["img"]
+            else:
+                img, txt = block(img=img,
+                                 txt=txt,
+                                 vec=vec,
+                                 pe=pe,
+                                 attn_mask=attn_mask)
+
+        img = torch.cat((txt, img), 1)
+
+        for i, block in enumerate(self.single_blocks):
+            if ("single_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"],
+                                       vec=args["vec"],
+                                       pe=args["pe"],
+                                       attn_mask=args.get("attn_mask"))
+                    return out
+
+                out = blocks_replace[("single_block", i)]({"img": img,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask},
+                                                          {"original_block": block_wrap})
+                img = out["img"]
+            else:
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
+
+        img = img[:, txt.shape[1]:, ...]
+        img = self.final_layer(img, vec)
+        return img.movedim(-2, -1) * (-1.0)

comfy/ldm/hunyuan3d/vae.py

+# Original: https://github.com/Tencent/Hunyuan3D-2/blob/main/hy3dgen/shapegen/models/autoencoders/model.py
+# Since the header on their VAE source file was a bit confusing we asked for permission to use this code from tencent under the GPL license used in ComfyUI.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+from typing import Union, Tuple, List, Callable, Optional
+
+import numpy as np
+from einops import repeat, rearrange
+from tqdm import tqdm
+import logging
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+def generate_dense_grid_points(
+    bbox_min: np.ndarray,
+    bbox_max: np.ndarray,
+    octree_resolution: int,
+    indexing: str = "ij",
+):
+    length = bbox_max - bbox_min
+    num_cells = octree_resolution
+
+    x = np.linspace(bbox_min[0], bbox_max[0], int(num_cells) + 1, dtype=np.float32)
+    y = np.linspace(bbox_min[1], bbox_max[1], int(num_cells) + 1, dtype=np.float32)
+    z = np.linspace(bbox_min[2], bbox_max[2], int(num_cells) + 1, dtype=np.float32)
+    [xs, ys, zs] = np.meshgrid(x, y, z, indexing=indexing)
+    xyz = np.stack((xs, ys, zs), axis=-1)
+    grid_size = [int(num_cells) + 1, int(num_cells) + 1, int(num_cells) + 1]
+
+    return xyz, grid_size, length
+
+
+class VanillaVolumeDecoder:
+    @torch.no_grad()
+    def __call__(
+        self,
+        latents: torch.FloatTensor,
+        geo_decoder: Callable,
+        bounds: Union[Tuple[float], List[float], float] = 1.01,
+        num_chunks: int = 10000,
+        octree_resolution: int = None,
+        enable_pbar: bool = True,
+        **kwargs,
+    ):
+        device = latents.device
+        dtype = latents.dtype
+        batch_size = latents.shape[0]
+
+        # 1. generate query points
+        if isinstance(bounds, float):
+            bounds = [-bounds, -bounds, -bounds, bounds, bounds, bounds]
+
+        bbox_min, bbox_max = np.array(bounds[0:3]), np.array(bounds[3:6])
+        xyz_samples, grid_size, length = generate_dense_grid_points(
+            bbox_min=bbox_min,
+            bbox_max=bbox_max,
+            octree_resolution=octree_resolution,
+            indexing="ij"
+        )
+        xyz_samples = torch.from_numpy(xyz_samples).to(device, dtype=dtype).contiguous().reshape(-1, 3)
+
+        # 2. latents to 3d volume
+        batch_logits = []
+        for start in tqdm(range(0, xyz_samples.shape[0], num_chunks), desc="Volume Decoding",
+                          disable=not enable_pbar):
+            chunk_queries = xyz_samples[start: start + num_chunks, :]
+            chunk_queries = repeat(chunk_queries, "p c -> b p c", b=batch_size)
+            logits = geo_decoder(queries=chunk_queries, latents=latents)
+            batch_logits.append(logits)
+
+        grid_logits = torch.cat(batch_logits, dim=1)
+        grid_logits = grid_logits.view((batch_size, *grid_size)).float()
+
+        return grid_logits
+
+
+class FourierEmbedder(nn.Module):
+    """The sin/cosine positional embedding. Given an input tensor `x` of shape [n_batch, ..., c_dim], it converts
+    each feature dimension of `x[..., i]` into:
+        [
+            sin(x[..., i]),
+            sin(f_1*x[..., i]),
+            sin(f_2*x[..., i]),
+            ...
+            sin(f_N * x[..., i]),
+            cos(x[..., i]),
+            cos(f_1*x[..., i]),
+            cos(f_2*x[..., i]),
+            ...
+            cos(f_N * x[..., i]),
+            x[..., i]     # only present if include_input is True.
+        ], here f_i is the frequency.
+
+    Denote the space is [0 / num_freqs, 1 / num_freqs, 2 / num_freqs, 3 / num_freqs, ..., (num_freqs - 1) / num_freqs].
+    If logspace is True, then the frequency f_i is [2^(0 / num_freqs), ..., 2^(i / num_freqs), ...];
+    Otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)].
+
+    Args:
+        num_freqs (int): the number of frequencies, default is 6;
+        logspace (bool): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+            otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)];
+        input_dim (int): the input dimension, default is 3;
+        include_input (bool): include the input tensor or not, default is True.
+
+    Attributes:
+        frequencies (torch.Tensor): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+                otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1);
+
+        out_dim (int): the embedding size, if include_input is True, it is input_dim * (num_freqs * 2 + 1),
+            otherwise, it is input_dim * num_freqs * 2.
+
+    """
+
+    def __init__(self,
+                 num_freqs: int = 6,
+                 logspace: bool = True,
+                 input_dim: int = 3,
+                 include_input: bool = True,
+                 include_pi: bool = True) -> None:
+
+        """The initialization"""
+
+        super().__init__()
+
+        if logspace:
+            frequencies = 2.0 ** torch.arange(
+                num_freqs,
+                dtype=torch.float32
+            )
+        else:
+            frequencies = torch.linspace(
+                1.0,
+                2.0 ** (num_freqs - 1),
+                num_freqs,
+                dtype=torch.float32
+            )
+
+        if include_pi:
+            frequencies *= torch.pi
+
+        self.register_buffer("frequencies", frequencies, persistent=False)
+        self.include_input = include_input
+        self.num_freqs = num_freqs
+
+        self.out_dim = self.get_dims(input_dim)
+
+    def get_dims(self, input_dim):
+        temp = 1 if self.include_input or self.num_freqs == 0 else 0
+        out_dim = input_dim * (self.num_freqs * 2 + temp)
+
+        return out_dim
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """ Forward process.
+
+        Args:
+            x: tensor of shape [..., dim]
+
+        Returns:
+            embedding: an embedding of `x` of shape [..., dim * (num_freqs * 2 + temp)]
+                where temp is 1 if include_input is True and 0 otherwise.
+        """
+
+        if self.num_freqs > 0:
+            embed = (x[..., None].contiguous() * self.frequencies.to(device=x.device, dtype=x.dtype)).view(*x.shape[:-1], -1)
+            if self.include_input:
+                return torch.cat((x, embed.sin(), embed.cos()), dim=-1)
+            else:
+                return torch.cat((embed.sin(), embed.cos()), dim=-1)
+        else:
+            return x
+
+
+class CrossAttentionProcessor:
+    def __call__(self, attn, q, k, v):
+        out = comfy.ops.scaled_dot_product_attention(q, k, v)
+        return out
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+
+    def __init__(self, drop_prob: float = 0., scale_by_keep: bool = True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+        This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+        the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+        See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+        changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+        'survival rate' as the argument.
+
+        """
+        if self.drop_prob == 0. or not self.training:
+            return x
+        keep_prob = 1 - self.drop_prob
+        shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+        random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+        if keep_prob > 0.0 and self.scale_by_keep:
+            random_tensor.div_(keep_prob)
+        return x * random_tensor
+
+    def extra_repr(self):
+        return f'drop_prob={round(self.drop_prob, 3):0.3f}'
+
+
+class MLP(nn.Module):
+    def __init__(
+        self, *,
+        width: int,
+        expand_ratio: int = 4,
+        output_width: int = None,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.c_fc = ops.Linear(width, width * expand_ratio)
+        self.c_proj = ops.Linear(width * expand_ratio, output_width if output_width is not None else width)
+        self.gelu = nn.GELU()
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
+
+    def forward(self, x):
+        return self.drop_path(self.c_proj(self.gelu(self.c_fc(x))))
+
+
+class QKVMultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        heads: int,
+        width=None,
+        qk_norm=False,
+        norm_layer=ops.LayerNorm
+    ):
+        super().__init__()
+        self.heads = heads
+        self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+
+        self.attn_processor = CrossAttentionProcessor()
+
+    def forward(self, q, kv):
+        _, n_ctx, _ = q.shape
+        bs, n_data, width = kv.shape
+        attn_ch = width // self.heads // 2
+        q = q.view(bs, n_ctx, self.heads, -1)
+        kv = kv.view(bs, n_data, self.heads, -1)
+        k, v = torch.split(kv, attn_ch, dim=-1)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+        q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v))
+        out = self.attn_processor(self, q, k, v)
+        out = out.transpose(1, 2).reshape(bs, n_ctx, -1)
+        return out
+
+
+class MultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        data_width: Optional[int] = None,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        kv_cache: bool = False,
+    ):
+        super().__init__()
+        self.width = width
+        self.heads = heads
+        self.data_width = width if data_width is None else data_width
+        self.c_q = ops.Linear(width, width, bias=qkv_bias)
+        self.c_kv = ops.Linear(self.data_width, width * 2, bias=qkv_bias)
+        self.c_proj = ops.Linear(width, width)
+        self.attention = QKVMultiheadCrossAttention(
+            heads=heads,
+            width=width,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.kv_cache = kv_cache
+        self.data = None
+
+    def forward(self, x, data):
+        x = self.c_q(x)
+        if self.kv_cache:
+            if self.data is None:
+                self.data = self.c_kv(data)
+                logging.info('Save kv cache,this should be called only once for one mesh')
+            data = self.data
+        else:
+            data = self.c_kv(data)
+        x = self.attention(x, data)
+        x = self.c_proj(x)
+        return x
+
+
+class ResidualCrossAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        mlp_expand_ratio: int = 4,
+        data_width: Optional[int] = None,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False
+    ):
+        super().__init__()
+
+        if data_width is None:
+            data_width = width
+
+        self.attn = MultiheadCrossAttention(
+            width=width,
+            heads=heads,
+            data_width=data_width,
+            qkv_bias=qkv_bias,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.ln_2 = norm_layer(data_width, elementwise_affine=True, eps=1e-6)
+        self.ln_3 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.mlp = MLP(width=width, expand_ratio=mlp_expand_ratio)
+
+    def forward(self, x: torch.Tensor, data: torch.Tensor):
+        x = x + self.attn(self.ln_1(x), self.ln_2(data))
+        x = x + self.mlp(self.ln_3(x))
+        return x
+
+
+class QKVMultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        heads: int,
+        width=None,
+        qk_norm=False,
+        norm_layer=ops.LayerNorm
+    ):
+        super().__init__()
+        self.heads = heads
+        self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+
+    def forward(self, qkv):
+        bs, n_ctx, width = qkv.shape
+        attn_ch = width // self.heads // 3
+        qkv = qkv.view(bs, n_ctx, self.heads, -1)
+        q, k, v = torch.split(qkv, attn_ch, dim=-1)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v))
+        out = F.scaled_dot_product_attention(q, k, v).transpose(1, 2).reshape(bs, n_ctx, -1)
+        return out
+
+
+class MultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.heads = heads
+        self.c_qkv = ops.Linear(width, width * 3, bias=qkv_bias)
+        self.c_proj = ops.Linear(width, width)
+        self.attention = QKVMultiheadAttention(
+            heads=heads,
+            width=width,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
+
+    def forward(self, x):
+        x = self.c_qkv(x)
+        x = self.attention(x)
+        x = self.drop_path(self.c_proj(x))
+        return x
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0,
+    ):
+        super().__init__()
+        self.attn = MultiheadAttention(
+            width=width,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm,
+            drop_path_rate=drop_path_rate
+        )
+        self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.mlp = MLP(width=width, drop_path_rate=drop_path_rate)
+        self.ln_2 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+
+    def forward(self, x: torch.Tensor):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        layers: int,
+        heads: int,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(
+                    width=width,
+                    heads=heads,
+                    qkv_bias=qkv_bias,
+                    norm_layer=norm_layer,
+                    qk_norm=qk_norm,
+                    drop_path_rate=drop_path_rate
+                )
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor):
+        for block in self.resblocks:
+            x = block(x)
+        return x
+
+
+class CrossAttentionDecoder(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        out_channels: int,
+        fourier_embedder: FourierEmbedder,
+        width: int,
+        heads: int,
+        mlp_expand_ratio: int = 4,
+        downsample_ratio: int = 1,
+        enable_ln_post: bool = True,
+        qkv_bias: bool = True,
+        qk_norm: bool = False,
+        label_type: str = "binary"
+    ):
+        super().__init__()
+
+        self.enable_ln_post = enable_ln_post
+        self.fourier_embedder = fourier_embedder
+        self.downsample_ratio = downsample_ratio
+        self.query_proj = ops.Linear(self.fourier_embedder.out_dim, width)
+        if self.downsample_ratio != 1:
+            self.latents_proj = ops.Linear(width * downsample_ratio, width)
+        if self.enable_ln_post == False:
+            qk_norm = False
+        self.cross_attn_decoder = ResidualCrossAttentionBlock(
+            width=width,
+            mlp_expand_ratio=mlp_expand_ratio,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm
+        )
+
+        if self.enable_ln_post:
+            self.ln_post = ops.LayerNorm(width)
+        self.output_proj = ops.Linear(width, out_channels)
+        self.label_type = label_type
+        self.count = 0
+
+    def forward(self, queries=None, query_embeddings=None, latents=None):
+        if query_embeddings is None:
+            query_embeddings = self.query_proj(self.fourier_embedder(queries).to(latents.dtype))
+        self.count += query_embeddings.shape[1]
+        if self.downsample_ratio != 1:
+            latents = self.latents_proj(latents)
+        x = self.cross_attn_decoder(query_embeddings, latents)
+        if self.enable_ln_post:
+            x = self.ln_post(x)
+        occ = self.output_proj(x)
+        return occ
+
+
+class ShapeVAE(nn.Module):
+    def __init__(
+        self,
+        *,
+        embed_dim: int,
+        width: int,
+        heads: int,
+        num_decoder_layers: int,
+        geo_decoder_downsample_ratio: int = 1,
+        geo_decoder_mlp_expand_ratio: int = 4,
+        geo_decoder_ln_post: bool = True,
+        num_freqs: int = 8,
+        include_pi: bool = True,
+        qkv_bias: bool = True,
+        qk_norm: bool = False,
+        label_type: str = "binary",
+        drop_path_rate: float = 0.0,
+        scale_factor: float = 1.0,
+    ):
+        super().__init__()
+        self.geo_decoder_ln_post = geo_decoder_ln_post
+
+        self.fourier_embedder = FourierEmbedder(num_freqs=num_freqs, include_pi=include_pi)
+
+        self.post_kl = ops.Linear(embed_dim, width)
+
+        self.transformer = Transformer(
+            width=width,
+            layers=num_decoder_layers,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm,
+            drop_path_rate=drop_path_rate
+        )
+
+        self.geo_decoder = CrossAttentionDecoder(
+            fourier_embedder=self.fourier_embedder,
+            out_channels=1,
+            mlp_expand_ratio=geo_decoder_mlp_expand_ratio,
+            downsample_ratio=geo_decoder_downsample_ratio,
+            enable_ln_post=self.geo_decoder_ln_post,
+            width=width // geo_decoder_downsample_ratio,
+            heads=heads // geo_decoder_downsample_ratio,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm,
+            label_type=label_type,
+        )
+
+        self.volume_decoder = VanillaVolumeDecoder()
+        self.scale_factor = scale_factor
+
+    def decode(self, latents, **kwargs):
+        latents = self.post_kl(latents.movedim(-2, -1))
+        latents = self.transformer(latents)
+
+        bounds = kwargs.get("bounds", 1.01)
+        num_chunks = kwargs.get("num_chunks", 8000)
+        octree_resolution = kwargs.get("octree_resolution", 256)
+        enable_pbar = kwargs.get("enable_pbar", True)
+
+        grid_logits = self.volume_decoder(latents, self.geo_decoder, bounds=bounds, num_chunks=num_chunks, octree_resolution=octree_resolution, enable_pbar=enable_pbar)
+        return grid_logits.movedim(-2, -1)
+
+    def encode(self, x):
+        return None

comfy/ldm/hunyuan_video/model.py

+#Based on Flux code because of weird hunyuan video code license.
+
+import torch
+import comfy.patcher_extension
+import comfy.ldm.flux.layers
+import comfy.ldm.modules.diffusionmodules.mmdit
+from comfy.ldm.modules.attention import optimized_attention
+
+
+from dataclasses import dataclass
+from einops import repeat
+
+from torch import Tensor, nn
+
+from comfy.ldm.flux.layers import (
+    DoubleStreamBlock,
+    EmbedND,
+    LastLayer,
+    MLPEmbedder,
+    SingleStreamBlock,
+    timestep_embedding
+)
+
+import comfy.ldm.common_dit
+
+
+@dataclass
+class HunyuanVideoParams:
+    in_channels: int
+    out_channels: int
+    vec_in_dim: int
+    context_in_dim: int
+    hidden_size: int
+    mlp_ratio: float
+    num_heads: int
+    depth: int
+    depth_single_blocks: int
+    axes_dim: list
+    theta: int
+    patch_size: list
+    qkv_bias: bool
+    guidance_embed: bool
+
+
+class SelfAttentionRef(nn.Module):
+    def __init__(self, dim: int, qkv_bias: bool = False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+        self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
+
+
+class TokenRefinerBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        heads,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.heads = heads
+        mlp_hidden_dim = hidden_size * 4
+
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+        self.norm1 = operations.LayerNorm(hidden_size, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device)
+        self.self_attn = SelfAttentionRef(hidden_size, True, dtype=dtype, device=device, operations=operations)
+
+        self.norm2 = operations.LayerNorm(hidden_size, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device)
+
+        self.mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+    def forward(self, x, c, mask):
+        mod1, mod2 = self.adaLN_modulation(c).chunk(2, dim=1)
+
+        norm_x = self.norm1(x)
+        qkv = self.self_attn.qkv(norm_x)
+        q, k, v = qkv.reshape(qkv.shape[0], qkv.shape[1], 3, self.heads, -1).permute(2, 0, 3, 1, 4)
+        attn = optimized_attention(q, k, v, self.heads, mask=mask, skip_reshape=True)
+
+        x = x + self.self_attn.proj(attn) * mod1.unsqueeze(1)
+        x = x + self.mlp(self.norm2(x)) * mod2.unsqueeze(1)
+        return x
+
+
+class IndividualTokenRefiner(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        heads,
+        num_blocks,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.blocks = nn.ModuleList(
+            [
+                TokenRefinerBlock(
+                    hidden_size=hidden_size,
+                    heads=heads,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations
+                )
+                for _ in range(num_blocks)
+            ]
+        )
+
+    def forward(self, x, c, mask):
+        m = None
+        if mask is not None:
+            m = mask.view(mask.shape[0], 1, 1, mask.shape[1]).repeat(1, 1, mask.shape[1], 1)
+            m = m + m.transpose(2, 3)
+
+        for block in self.blocks:
+            x = block(x, c, m)
+        return x
+
+
+
+class TokenRefiner(nn.Module):
+    def __init__(
+        self,
+        text_dim,
+        hidden_size,
+        heads,
+        num_blocks,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+
+        self.input_embedder = operations.Linear(text_dim, hidden_size, bias=True, dtype=dtype, device=device)
+        self.t_embedder = MLPEmbedder(256, hidden_size, dtype=dtype, device=device, operations=operations)
+        self.c_embedder = MLPEmbedder(text_dim, hidden_size, dtype=dtype, device=device, operations=operations)
+        self.individual_token_refiner = IndividualTokenRefiner(hidden_size, heads, num_blocks, dtype=dtype, device=device, operations=operations)
+
+    def forward(
+        self,
+        x,
+        timesteps,
+        mask,
+    ):
+        t = self.t_embedder(timestep_embedding(timesteps, 256, time_factor=1.0).to(x.dtype))
+        # m = mask.float().unsqueeze(-1)
+        # c = (x.float() * m).sum(dim=1) / m.sum(dim=1) #TODO: the following works when the x.shape is the same length as the tokens but might break otherwise
+        c = x.sum(dim=1) / x.shape[1]
+
+        c = t + self.c_embedder(c.to(x.dtype))
+        x = self.input_embedder(x)
+        x = self.individual_token_refiner(x, c, mask)
+        return x
+
+class HunyuanVideo(nn.Module):
+    """
+    Transformer model for flow matching on sequences.
+    """
+
+    def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
+        super().__init__()
+        self.dtype = dtype
+        params = HunyuanVideoParams(**kwargs)
+        self.params = params
+        self.patch_size = params.patch_size
+        self.in_channels = params.in_channels
+        self.out_channels = params.out_channels
+        if params.hidden_size % params.num_heads != 0:
+            raise ValueError(
+                f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
+            )
+        pe_dim = params.hidden_size // params.num_heads
+        if sum(params.axes_dim) != pe_dim:
+            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
+        self.hidden_size = params.hidden_size
+        self.num_heads = params.num_heads
+        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
+
+        self.img_in = comfy.ldm.modules.diffusionmodules.mmdit.PatchEmbed(None, self.patch_size, self.in_channels, self.hidden_size, conv3d=True, dtype=dtype, device=device, operations=operations)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations)
+        self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        self.guidance_in = (
+            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, dtype=dtype, device=device, operations=operations) if params.guidance_embed else nn.Identity()
+        )
+
+        self.txt_in = TokenRefiner(params.context_in_dim, self.hidden_size, self.num_heads, 2, dtype=dtype, device=device, operations=operations)
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                    flipped_img_txt=True,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(params.depth)
+            ]
+        )
+
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
+                for _ in range(params.depth_single_blocks)
+            ]
+        )
+
+        if final_layer:
+            self.final_layer = LastLayer(self.hidden_size, self.patch_size[-1], self.out_channels, dtype=dtype, device=device, operations=operations)
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        txt_mask: Tensor,
+        timesteps: Tensor,
+        y: Tensor,
+        guidance: Tensor = None,
+        guiding_frame_index=None,
+        ref_latent=None,
+        control=None,
+        transformer_options={},
+    ) -> Tensor:
+        patches_replace = transformer_options.get("patches_replace", {})
+
+        initial_shape = list(img.shape)
+        # running on sequences img
+        img = self.img_in(img)
+        vec = self.time_in(timestep_embedding(timesteps, 256, time_factor=1.0).to(img.dtype))
+
+        if ref_latent is not None:
+            ref_latent_ids = self.img_ids(ref_latent)
+            ref_latent = self.img_in(ref_latent)
+            img = torch.cat([ref_latent, img], dim=-2)
+            ref_latent_ids[..., 0] = -1
+            ref_latent_ids[..., 2] += (initial_shape[-1] // self.patch_size[-1])
+            img_ids = torch.cat([ref_latent_ids, img_ids], dim=-2)
+
+        if guiding_frame_index is not None:
+            token_replace_vec = self.time_in(timestep_embedding(guiding_frame_index, 256, time_factor=1.0))
+            vec_ = self.vector_in(y[:, :self.params.vec_in_dim])
+            vec = torch.cat([(vec_ + token_replace_vec).unsqueeze(1), (vec_ + vec).unsqueeze(1)], dim=1)
+            frame_tokens = (initial_shape[-1] // self.patch_size[-1]) * (initial_shape[-2] // self.patch_size[-2])
+            modulation_dims = [(0, frame_tokens, 0), (frame_tokens, None, 1)]
+            modulation_dims_txt = [(0, None, 1)]
+        else:
+            vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
+            modulation_dims = None
+            modulation_dims_txt = None
+
+        if self.params.guidance_embed:
+            if guidance is not None:
+                vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
+
+        if txt_mask is not None and not torch.is_floating_point(txt_mask):
+            txt_mask = (txt_mask - 1).to(img.dtype) * torch.finfo(img.dtype).max
+
+        txt = self.txt_in(txt, timesteps, txt_mask)
+
+        ids = torch.cat((img_ids, txt_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        img_len = img.shape[1]
+        if txt_mask is not None:
+            attn_mask_len = img_len + txt.shape[1]
+            attn_mask = torch.zeros((1, 1, attn_mask_len), dtype=img.dtype, device=img.device)
+            attn_mask[:, 0, img_len:] = txt_mask
+        else:
+            attn_mask = None
+
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.double_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"], out["txt"] = block(img=args["img"], txt=args["txt"], vec=args["vec"], pe=args["pe"], attn_mask=args["attention_mask"], modulation_dims_img=args["modulation_dims_img"], modulation_dims_txt=args["modulation_dims_txt"])
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": img, "txt": txt, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims_img': modulation_dims, 'modulation_dims_txt': modulation_dims_txt}, {"original_block": block_wrap})
+                txt = out["txt"]
+                img = out["img"]
+            else:
+                img, txt = block(img=img, txt=txt, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims_img=modulation_dims, modulation_dims_txt=modulation_dims_txt)
+
+            if control is not None: # Controlnet
+                control_i = control.get("input")
+                if i < len(control_i):
+                    add = control_i[i]
+                    if add is not None:
+                        img += add
+
+        img = torch.cat((img, txt), 1)
+
+        for i, block in enumerate(self.single_blocks):
+            if ("single_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], vec=args["vec"], pe=args["pe"], attn_mask=args["attention_mask"], modulation_dims=args["modulation_dims"])
+                    return out
+
+                out = blocks_replace[("single_block", i)]({"img": img, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims': modulation_dims}, {"original_block": block_wrap})
+                img = out["img"]
+            else:
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims=modulation_dims)
+
+            if control is not None: # Controlnet
+                control_o = control.get("output")
+                if i < len(control_o):
+                    add = control_o[i]
+                    if add is not None:
+                        img[:, : img_len] += add
+
+        img = img[:, : img_len]
+        if ref_latent is not None:
+            img = img[:, ref_latent.shape[1]:]
+
+        img = self.final_layer(img, vec, modulation_dims=modulation_dims)  # (N, T, patch_size ** 2 * out_channels)
+
+        shape = initial_shape[-3:]
+        for i in range(len(shape)):
+            shape[i] = shape[i] // self.patch_size[i]
+        img = img.reshape([img.shape[0]] + shape + [self.out_channels] + self.patch_size)
+        img = img.permute(0, 4, 1, 5, 2, 6, 3, 7)
+        img = img.reshape(initial_shape[0], self.out_channels, initial_shape[2], initial_shape[3], initial_shape[4])
+        return img
+
+    def img_ids(self, x):
+        bs, c, t, h, w = x.shape
+        patch_size = self.patch_size
+        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
+        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
+        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
+        img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1)
+        return repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+
+    def forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=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, timestep, context, y, guidance, attention_mask, guiding_frame_index, ref_latent, control, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, control=None, transformer_options={}, **kwargs):
+        bs, c, t, h, w = x.shape
+        img_ids = self.img_ids(x)
+        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, guidance, guiding_frame_index, ref_latent, control=control, transformer_options=transformer_options)
+        return out

comfy/ldm/hydit/attn_layers.py

+import torch
+import torch.nn as nn
+from typing import Tuple, Union, Optional
+from comfy.ldm.modules.attention import optimized_attention
+
+
+def reshape_for_broadcast(freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]], x: torch.Tensor, head_first=False):
+    """
+    Reshape frequency tensor for broadcasting it with another tensor.
+
+    This function reshapes the frequency tensor to have the same shape as the target tensor 'x'
+    for the purpose of broadcasting the frequency tensor during element-wise operations.
+
+    Args:
+        freqs_cis (Union[torch.Tensor, Tuple[torch.Tensor]]): Frequency tensor to be reshaped.
+        x (torch.Tensor): Target tensor for broadcasting compatibility.
+        head_first (bool): head dimension first (except batch dim) or not.
+
+    Returns:
+        torch.Tensor: Reshaped frequency tensor.
+
+    Raises:
+        AssertionError: If the frequency tensor doesn't match the expected shape.
+        AssertionError: If the target tensor 'x' doesn't have the expected number of dimensions.
+    """
+    ndim = x.ndim
+    assert 0 <= 1 < ndim
+
+    if isinstance(freqs_cis, tuple):
+        # freqs_cis: (cos, sin) in real space
+        if head_first:
+            assert freqs_cis[0].shape == (x.shape[-2], x.shape[-1]), f'freqs_cis shape {freqs_cis[0].shape} does not match x shape {x.shape}'
+            shape = [d if i == ndim - 2 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
+        else:
+            assert freqs_cis[0].shape == (x.shape[1], x.shape[-1]), f'freqs_cis shape {freqs_cis[0].shape} does not match x shape {x.shape}'
+            shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
+        return freqs_cis[0].view(*shape), freqs_cis[1].view(*shape)
+    else:
+        # freqs_cis: values in complex space
+        if head_first:
+            assert freqs_cis.shape == (x.shape[-2], x.shape[-1]), f'freqs_cis shape {freqs_cis.shape} does not match x shape {x.shape}'
+            shape = [d if i == ndim - 2 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
+        else:
+            assert freqs_cis.shape == (x.shape[1], x.shape[-1]), f'freqs_cis shape {freqs_cis.shape} does not match x shape {x.shape}'
+            shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
+        return freqs_cis.view(*shape)
+
+
+def rotate_half(x):
+    x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+    return torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+
+
+def apply_rotary_emb(
+        xq: torch.Tensor,
+        xk: Optional[torch.Tensor],
+        freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
+        head_first: bool = False,
+) -> Tuple[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:
+        xq (torch.Tensor): Query tensor to apply rotary embeddings. [B, S, H, D]
+        xk (torch.Tensor): Key tensor to apply rotary embeddings.   [B, S, H, D]
+        freqs_cis (Union[torch.Tensor, Tuple[torch.Tensor]]): Precomputed frequency tensor for complex exponentials.
+        head_first (bool): head dimension first (except batch dim) or not.
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
+
+    """
+    xk_out = None
+    if isinstance(freqs_cis, tuple):
+        cos, sin = reshape_for_broadcast(freqs_cis, xq, head_first)    # [S, D]
+        xq_out = (xq * cos + rotate_half(xq) * sin)
+        if xk is not None:
+            xk_out = (xk * cos + rotate_half(xk) * sin)
+    else:
+        xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))  # [B, S, H, D//2]
+        freqs_cis = reshape_for_broadcast(freqs_cis, xq_, head_first).to(xq.device)   # [S, D//2] --> [1, S, 1, D//2]
+        xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3).type_as(xq)
+        if xk is not None:
+            xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))  # [B, S, H, D//2]
+            xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3).type_as(xk)
+
+    return xq_out, xk_out
+
+
+
+class CrossAttention(nn.Module):
+    """
+    Use QK Normalization.
+    """
+    def __init__(self,
+                 qdim,
+                 kdim,
+                 num_heads,
+                 qkv_bias=True,
+                 qk_norm=False,
+                 attn_drop=0.0,
+                 proj_drop=0.0,
+                 attn_precision=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+        factory_kwargs = {'device': device, 'dtype': dtype}
+        super().__init__()
+        self.attn_precision = attn_precision
+        self.qdim = qdim
+        self.kdim = kdim
+        self.num_heads = num_heads
+        assert self.qdim % num_heads == 0, "self.qdim must be divisible by num_heads"
+        self.head_dim = self.qdim // num_heads
+        assert self.head_dim % 8 == 0 and self.head_dim <= 128, "Only support head_dim <= 128 and divisible by 8"
+        self.scale = self.head_dim ** -0.5
+
+        self.q_proj = operations.Linear(qdim, qdim, bias=qkv_bias, **factory_kwargs)
+        self.kv_proj = operations.Linear(kdim, 2 * qdim, bias=qkv_bias, **factory_kwargs)
+
+        # TODO: eps should be 1 / 65530 if using fp16
+        self.q_norm = operations.LayerNorm(self.head_dim, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device) if qk_norm else nn.Identity()
+        self.k_norm = operations.LayerNorm(self.head_dim, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device) if qk_norm else nn.Identity()
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.out_proj = operations.Linear(qdim, qdim, bias=qkv_bias, **factory_kwargs)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x, y, freqs_cis_img=None):
+        """
+        Parameters
+        ----------
+        x: torch.Tensor
+            (batch, seqlen1, hidden_dim) (where hidden_dim = num heads * head dim)
+        y: torch.Tensor
+            (batch, seqlen2, hidden_dim2)
+        freqs_cis_img: torch.Tensor
+            (batch, hidden_dim // 2), RoPE for image
+        """
+        b, s1, c = x.shape     # [b, s1, D]
+        _, s2, c = y.shape     # [b, s2, 1024]
+
+        q = self.q_proj(x).view(b, s1, self.num_heads, self.head_dim)   # [b, s1, h, d]
+        kv = self.kv_proj(y).view(b, s2, 2, self.num_heads, self.head_dim)    # [b, s2, 2, h, d]
+        k, v = kv.unbind(dim=2) # [b, s, h, d]
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        # Apply RoPE if needed
+        if freqs_cis_img is not None:
+            qq, _ = apply_rotary_emb(q, None, freqs_cis_img)
+            assert qq.shape == q.shape, f'qq: {qq.shape}, q: {q.shape}'
+            q = qq
+
+        q = q.transpose(-2, -3).contiguous()        # q ->  B, L1, H, C - B, H, L1, C
+        k = k.transpose(-2, -3).contiguous()      # k ->  B, L2, H, C - B, H, C, L2
+        v = v.transpose(-2, -3).contiguous()
+
+        context = optimized_attention(q, k, v, self.num_heads, skip_reshape=True, attn_precision=self.attn_precision)
+
+        out = self.out_proj(context)  # context.reshape - B, L1, -1
+        out = self.proj_drop(out)
+
+        out_tuple = (out,)
+
+        return out_tuple
+
+
+class Attention(nn.Module):
+    """
+    We rename some layer names to align with flash attention
+    """
+    def __init__(self, dim, num_heads, qkv_bias=True, qk_norm=False, attn_drop=0., proj_drop=0., attn_precision=None, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.attn_precision = attn_precision
+        self.dim = dim
+        self.num_heads = num_heads
+        assert self.dim % num_heads == 0, 'dim should be divisible by num_heads'
+        self.head_dim = self.dim // num_heads
+        # This assertion is aligned with flash attention
+        assert self.head_dim % 8 == 0 and self.head_dim <= 128, "Only support head_dim <= 128 and divisible by 8"
+        self.scale = self.head_dim ** -0.5
+
+        # qkv --> Wqkv
+        self.Wqkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+        # TODO: eps should be 1 / 65530 if using fp16
+        self.q_norm = operations.LayerNorm(self.head_dim, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device) if qk_norm else nn.Identity()
+        self.k_norm = operations.LayerNorm(self.head_dim, elementwise_affine=True, eps=1e-6, dtype=dtype, device=device) if qk_norm else nn.Identity()
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.out_proj = operations.Linear(dim, dim, dtype=dtype, device=device)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x, freqs_cis_img=None):
+        B, N, C = x.shape
+        qkv = self.Wqkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)   # [3, b, h, s, d]
+        q, k, v = qkv.unbind(0)     # [b, h, s, d]
+        q = self.q_norm(q)          # [b, h, s, d]
+        k = self.k_norm(k)          # [b, h, s, d]
+
+        # Apply RoPE if needed
+        if freqs_cis_img is not None:
+            qq, kk = apply_rotary_emb(q, k, freqs_cis_img, head_first=True)
+            assert qq.shape == q.shape and kk.shape == k.shape, \
+                f'qq: {qq.shape}, q: {q.shape}, kk: {kk.shape}, k: {k.shape}'
+            q, k = qq, kk
+
+        x = optimized_attention(q, k, v, self.num_heads, skip_reshape=True, attn_precision=self.attn_precision)
+        x = self.out_proj(x)
+        x = self.proj_drop(x)
+
+        out_tuple = (x,)
+
+        return out_tuple

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}