Initial commit

comfy/ldm/cosmos/position_embedding.py

+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional
+
+import torch
+from einops import rearrange, repeat
+from torch import nn
+import math
+
+
+def normalize(x: torch.Tensor, dim: Optional[List[int]] = None, eps: float = 0) -> torch.Tensor:
+    """
+    Normalizes the input tensor along specified dimensions such that the average square norm of elements is adjusted.
+
+    Args:
+        x (torch.Tensor): The input tensor to normalize.
+        dim (list, optional): The dimensions over which to normalize. If None, normalizes over all dimensions except the first.
+        eps (float, optional): A small constant to ensure numerical stability during division.
+
+    Returns:
+        torch.Tensor: The normalized tensor.
+    """
+    if dim is None:
+        dim = list(range(1, x.ndim))
+    norm = torch.linalg.vector_norm(x, dim=dim, keepdim=True, dtype=torch.float32)
+    norm = torch.add(eps, norm, alpha=math.sqrt(norm.numel() / x.numel()))
+    return x / norm.to(x.dtype)
+
+
+class VideoPositionEmb(nn.Module):
+    def forward(self, x_B_T_H_W_C: torch.Tensor, fps=Optional[torch.Tensor], device=None, dtype=None) -> torch.Tensor:
+        """
+        It delegates the embedding generation to generate_embeddings function.
+        """
+        B_T_H_W_C = x_B_T_H_W_C.shape
+        embeddings = self.generate_embeddings(B_T_H_W_C, fps=fps, device=device, dtype=dtype)
+
+        return embeddings
+
+    def generate_embeddings(self, B_T_H_W_C: torch.Size, fps=Optional[torch.Tensor], device=None):
+        raise NotImplementedError
+
+
+class VideoRopePosition3DEmb(VideoPositionEmb):
+    def __init__(
+        self,
+        *,  # enforce keyword arguments
+        head_dim: int,
+        len_h: int,
+        len_w: int,
+        len_t: int,
+        base_fps: int = 24,
+        h_extrapolation_ratio: float = 1.0,
+        w_extrapolation_ratio: float = 1.0,
+        t_extrapolation_ratio: float = 1.0,
+        enable_fps_modulation: bool = True,
+        device=None,
+        **kwargs,  # used for compatibility with other positional embeddings; unused in this class
+    ):
+        del kwargs
+        super().__init__()
+        self.base_fps = base_fps
+        self.max_h = len_h
+        self.max_w = len_w
+        self.enable_fps_modulation = enable_fps_modulation
+
+        dim = head_dim
+        dim_h = dim // 6 * 2
+        dim_w = dim_h
+        dim_t = dim - 2 * dim_h
+        assert dim == dim_h + dim_w + dim_t, f"bad dim: {dim} != {dim_h} + {dim_w} + {dim_t}"
+        self.register_buffer(
+            "dim_spatial_range",
+            torch.arange(0, dim_h, 2, device=device)[: (dim_h // 2)].float() / dim_h,
+            persistent=False,
+        )
+        self.register_buffer(
+            "dim_temporal_range",
+            torch.arange(0, dim_t, 2, device=device)[: (dim_t // 2)].float() / dim_t,
+            persistent=False,
+        )
+
+        self.h_ntk_factor = h_extrapolation_ratio ** (dim_h / (dim_h - 2))
+        self.w_ntk_factor = w_extrapolation_ratio ** (dim_w / (dim_w - 2))
+        self.t_ntk_factor = t_extrapolation_ratio ** (dim_t / (dim_t - 2))
+
+    def generate_embeddings(
+        self,
+        B_T_H_W_C: torch.Size,
+        fps: Optional[torch.Tensor] = None,
+        h_ntk_factor: Optional[float] = None,
+        w_ntk_factor: Optional[float] = None,
+        t_ntk_factor: Optional[float] = None,
+        device=None,
+        dtype=None,
+    ):
+        """
+        Generate embeddings for the given input size.
+
+        Args:
+            B_T_H_W_C (torch.Size): Input tensor size (Batch, Time, Height, Width, Channels).
+            fps (Optional[torch.Tensor], optional): Frames per second. Defaults to None.
+            h_ntk_factor (Optional[float], optional): Height NTK factor. If None, uses self.h_ntk_factor.
+            w_ntk_factor (Optional[float], optional): Width NTK factor. If None, uses self.w_ntk_factor.
+            t_ntk_factor (Optional[float], optional): Time NTK factor. If None, uses self.t_ntk_factor.
+
+        Returns:
+            Not specified in the original code snippet.
+        """
+        h_ntk_factor = h_ntk_factor if h_ntk_factor is not None else self.h_ntk_factor
+        w_ntk_factor = w_ntk_factor if w_ntk_factor is not None else self.w_ntk_factor
+        t_ntk_factor = t_ntk_factor if t_ntk_factor is not None else self.t_ntk_factor
+
+        h_theta = 10000.0 * h_ntk_factor
+        w_theta = 10000.0 * w_ntk_factor
+        t_theta = 10000.0 * t_ntk_factor
+
+        h_spatial_freqs = 1.0 / (h_theta**self.dim_spatial_range.to(device=device))
+        w_spatial_freqs = 1.0 / (w_theta**self.dim_spatial_range.to(device=device))
+        temporal_freqs = 1.0 / (t_theta**self.dim_temporal_range.to(device=device))
+
+        B, T, H, W, _ = B_T_H_W_C
+        seq = torch.arange(max(H, W, T), dtype=torch.float, device=device)
+        uniform_fps = (fps is None) or isinstance(fps, (int, float)) or (fps.min() == fps.max())
+        assert (
+            uniform_fps or B == 1 or T == 1
+        ), "For video batch, batch size should be 1 for non-uniform fps. For image batch, T should be 1"
+        half_emb_h = torch.outer(seq[:H].to(device=device), h_spatial_freqs)
+        half_emb_w = torch.outer(seq[:W].to(device=device), w_spatial_freqs)
+
+        # apply sequence scaling in temporal dimension
+        if fps is None or self.enable_fps_modulation is False:  # image case
+            half_emb_t = torch.outer(seq[:T].to(device=device), temporal_freqs)
+        else:
+            half_emb_t = torch.outer(seq[:T].to(device=device) / fps * self.base_fps, temporal_freqs)
+
+        half_emb_h = torch.stack([torch.cos(half_emb_h), -torch.sin(half_emb_h), torch.sin(half_emb_h), torch.cos(half_emb_h)], dim=-1)
+        half_emb_w = torch.stack([torch.cos(half_emb_w), -torch.sin(half_emb_w), torch.sin(half_emb_w), torch.cos(half_emb_w)], dim=-1)
+        half_emb_t = torch.stack([torch.cos(half_emb_t), -torch.sin(half_emb_t), torch.sin(half_emb_t), torch.cos(half_emb_t)], dim=-1)
+
+        em_T_H_W_D = torch.cat(
+            [
+                repeat(half_emb_t, "t d x -> t h w d x", h=H, w=W),
+                repeat(half_emb_h, "h d x -> t h w d x", t=T, w=W),
+                repeat(half_emb_w, "w d x -> t h w d x", t=T, h=H),
+            ]
+            , dim=-2,
+        )
+
+        return rearrange(em_T_H_W_D, "t h w d (i j) -> (t h w) d i j", i=2, j=2).float()
+
+
+class LearnablePosEmbAxis(VideoPositionEmb):
+    def __init__(
+        self,
+        *,  # enforce keyword arguments
+        interpolation: str,
+        model_channels: int,
+        len_h: int,
+        len_w: int,
+        len_t: int,
+        device=None,
+        dtype=None,
+        **kwargs,
+    ):
+        """
+        Args:
+            interpolation (str): we curretly only support "crop", ideally when we need extrapolation capacity, we should adjust frequency or other more advanced methods. they are not implemented yet.
+        """
+        del kwargs  # unused
+        super().__init__()
+        self.interpolation = interpolation
+        assert self.interpolation in ["crop"], f"Unknown interpolation method {self.interpolation}"
+
+        self.pos_emb_h = nn.Parameter(torch.empty(len_h, model_channels, device=device, dtype=dtype))
+        self.pos_emb_w = nn.Parameter(torch.empty(len_w, model_channels, device=device, dtype=dtype))
+        self.pos_emb_t = nn.Parameter(torch.empty(len_t, model_channels, device=device, dtype=dtype))
+
+    def generate_embeddings(self, B_T_H_W_C: torch.Size, fps=Optional[torch.Tensor], device=None, dtype=None) -> torch.Tensor:
+        B, T, H, W, _ = B_T_H_W_C
+        if self.interpolation == "crop":
+            emb_h_H = self.pos_emb_h[:H].to(device=device, dtype=dtype)
+            emb_w_W = self.pos_emb_w[:W].to(device=device, dtype=dtype)
+            emb_t_T = self.pos_emb_t[:T].to(device=device, dtype=dtype)
+            emb = (
+                repeat(emb_t_T, "t d-> b t h w d", b=B, h=H, w=W)
+                + repeat(emb_h_H, "h d-> b t h w d", b=B, t=T, w=W)
+                + repeat(emb_w_W, "w d-> b t h w d", b=B, t=T, h=H)
+            )
+            assert list(emb.shape)[:4] == [B, T, H, W], f"bad shape: {list(emb.shape)[:4]} != {B, T, H, W}"
+        else:
+            raise ValueError(f"Unknown interpolation method {self.interpolation}")
+
+        return normalize(emb, dim=-1, eps=1e-6)

comfy/ldm/cosmos/predict2.py

+# original code from: https://github.com/nvidia-cosmos/cosmos-predict2
+
+import torch
+from torch import nn
+from einops import rearrange
+from einops.layers.torch import Rearrange
+import logging
+from typing import Callable, Optional, Tuple
+import math
+
+from .position_embedding import VideoRopePosition3DEmb, LearnablePosEmbAxis
+from torchvision import transforms
+
+import comfy.patcher_extension
+from comfy.ldm.modules.attention import optimized_attention
+
+def apply_rotary_pos_emb(
+    t: torch.Tensor,
+    freqs: torch.Tensor,
+) -> torch.Tensor:
+    t_ = t.reshape(*t.shape[:-1], 2, -1).movedim(-2, -1).unsqueeze(-2).float()
+    t_out = freqs[..., 0] * t_[..., 0] + freqs[..., 1] * t_[..., 1]
+    t_out = t_out.movedim(-1, -2).reshape(*t.shape).type_as(t)
+    return t_out
+
+
+# ---------------------- Feed Forward Network -----------------------
+class GPT2FeedForward(nn.Module):
+    def __init__(self, d_model: int, d_ff: int, device=None, dtype=None, operations=None) -> None:
+        super().__init__()
+        self.activation = nn.GELU()
+        self.layer1 = operations.Linear(d_model, d_ff, bias=False, device=device, dtype=dtype)
+        self.layer2 = operations.Linear(d_ff, d_model, bias=False, device=device, dtype=dtype)
+
+        self._layer_id = None
+        self._dim = d_model
+        self._hidden_dim = d_ff
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.layer1(x)
+
+        x = self.activation(x)
+        x = self.layer2(x)
+        return x
+
+
+def torch_attention_op(q_B_S_H_D: torch.Tensor, k_B_S_H_D: torch.Tensor, v_B_S_H_D: torch.Tensor) -> torch.Tensor:
+    """Computes multi-head attention using PyTorch's native implementation.
+
+    This function provides a PyTorch backend alternative to Transformer Engine's attention operation.
+    It rearranges the input tensors to match PyTorch's expected format, computes scaled dot-product
+    attention, and rearranges the output back to the original format.
+
+    The input tensor names use the following dimension conventions:
+
+    - B: batch size
+    - S: sequence length
+    - H: number of attention heads
+    - D: head dimension
+
+    Args:
+        q_B_S_H_D: Query tensor with shape (batch, seq_len, n_heads, head_dim)
+        k_B_S_H_D: Key tensor with shape (batch, seq_len, n_heads, head_dim)
+        v_B_S_H_D: Value tensor with shape (batch, seq_len, n_heads, head_dim)
+
+    Returns:
+        Attention output tensor with shape (batch, seq_len, n_heads * head_dim)
+    """
+    in_q_shape = q_B_S_H_D.shape
+    in_k_shape = k_B_S_H_D.shape
+    q_B_H_S_D = rearrange(q_B_S_H_D, "b ... h k -> b h ... k").view(in_q_shape[0], in_q_shape[-2], -1, in_q_shape[-1])
+    k_B_H_S_D = rearrange(k_B_S_H_D, "b ... h v -> b h ... v").view(in_k_shape[0], in_k_shape[-2], -1, in_k_shape[-1])
+    v_B_H_S_D = rearrange(v_B_S_H_D, "b ... h v -> b h ... v").view(in_k_shape[0], in_k_shape[-2], -1, in_k_shape[-1])
+    return optimized_attention(q_B_H_S_D, k_B_H_S_D, v_B_H_S_D, in_q_shape[-2], skip_reshape=True)
+
+
+class Attention(nn.Module):
+    """
+    A flexible attention module supporting both self-attention and cross-attention mechanisms.
+
+    This module implements a multi-head attention layer that can operate in either self-attention
+    or cross-attention mode. The mode is determined by whether a context dimension is provided.
+    The implementation uses scaled dot-product attention and supports optional bias terms and
+    dropout regularization.
+
+    Args:
+        query_dim (int): The dimensionality of the query vectors.
+        context_dim (int, optional): The dimensionality of the context (key/value) vectors.
+            If None, the module operates in self-attention mode using query_dim. Default: None
+        n_heads (int, optional): Number of attention heads for multi-head attention. Default: 8
+        head_dim (int, optional): The dimension of each attention head. Default: 64
+        dropout (float, optional): Dropout probability applied to the output. Default: 0.0
+        qkv_format (str, optional): Format specification for QKV tensors. Default: "bshd"
+        backend (str, optional): Backend to use for the attention operation. Default: "transformer_engine"
+
+    Examples:
+        >>> # Self-attention with 512 dimensions and 8 heads
+        >>> self_attn = Attention(query_dim=512)
+        >>> x = torch.randn(32, 16, 512)  # (batch_size, seq_len, dim)
+        >>> out = self_attn(x)  # (32, 16, 512)
+
+        >>> # Cross-attention
+        >>> cross_attn = Attention(query_dim=512, context_dim=256)
+        >>> query = torch.randn(32, 16, 512)
+        >>> context = torch.randn(32, 8, 256)
+        >>> out = cross_attn(query, context)  # (32, 16, 512)
+    """
+
+    def __init__(
+        self,
+        query_dim: int,
+        context_dim: Optional[int] = None,
+        n_heads: int = 8,
+        head_dim: int = 64,
+        dropout: float = 0.0,
+        device=None,
+        dtype=None,
+        operations=None,
+    ) -> None:
+        super().__init__()
+        logging.debug(
+            f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, context_dim is {context_dim} and using "
+            f"{n_heads} heads with a dimension of {head_dim}."
+        )
+        self.is_selfattn = context_dim is None  # self attention
+
+        context_dim = query_dim if context_dim is None else context_dim
+        inner_dim = head_dim * n_heads
+
+        self.n_heads = n_heads
+        self.head_dim = head_dim
+        self.query_dim = query_dim
+        self.context_dim = context_dim
+
+        self.q_proj = operations.Linear(query_dim, inner_dim, bias=False, device=device, dtype=dtype)
+        self.q_norm = operations.RMSNorm(self.head_dim, eps=1e-6, device=device, dtype=dtype)
+
+        self.k_proj = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype)
+        self.k_norm = operations.RMSNorm(self.head_dim, eps=1e-6, device=device, dtype=dtype)
+
+        self.v_proj = operations.Linear(context_dim, inner_dim, bias=False, device=device, dtype=dtype)
+        self.v_norm = nn.Identity()
+
+        self.output_proj = operations.Linear(inner_dim, query_dim, bias=False, device=device, dtype=dtype)
+        self.output_dropout = nn.Dropout(dropout) if dropout > 1e-4 else nn.Identity()
+
+        self.attn_op = torch_attention_op
+
+        self._query_dim = query_dim
+        self._context_dim = context_dim
+        self._inner_dim = inner_dim
+
+    def compute_qkv(
+        self,
+        x: torch.Tensor,
+        context: Optional[torch.Tensor] = None,
+        rope_emb: Optional[torch.Tensor] = None,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        q = self.q_proj(x)
+        context = x if context is None else context
+        k = self.k_proj(context)
+        v = self.v_proj(context)
+        q, k, v = map(
+            lambda t: rearrange(t, "b ... (h d) -> b ... h d", h=self.n_heads, d=self.head_dim),
+            (q, k, v),
+        )
+
+        def apply_norm_and_rotary_pos_emb(
+            q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, rope_emb: Optional[torch.Tensor]
+        ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+            q = self.q_norm(q)
+            k = self.k_norm(k)
+            v = self.v_norm(v)
+            if self.is_selfattn and rope_emb is not None:  # only apply to self-attention!
+                q = apply_rotary_pos_emb(q, rope_emb)
+                k = apply_rotary_pos_emb(k, rope_emb)
+            return q, k, v
+
+        q, k, v = apply_norm_and_rotary_pos_emb(q, k, v, rope_emb)
+
+        return q, k, v
+
+    def compute_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> torch.Tensor:
+        result = self.attn_op(q, k, v)  # [B, S, H, D]
+        return self.output_dropout(self.output_proj(result))
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        context: Optional[torch.Tensor] = None,
+        rope_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Args:
+            x (Tensor): The query tensor of shape [B, Mq, K]
+            context (Optional[Tensor]): The key tensor of shape [B, Mk, K] or use x as context [self attention] if None
+        """
+        q, k, v = self.compute_qkv(x, context, rope_emb=rope_emb)
+        return self.compute_attention(q, k, v)
+
+
+class Timesteps(nn.Module):
+    def __init__(self, num_channels: int):
+        super().__init__()
+        self.num_channels = num_channels
+
+    def forward(self, timesteps_B_T: torch.Tensor) -> torch.Tensor:
+        assert timesteps_B_T.ndim == 2, f"Expected 2D input, got {timesteps_B_T.ndim}"
+        timesteps = timesteps_B_T.flatten().float()
+        half_dim = self.num_channels // 2
+        exponent = -math.log(10000) * torch.arange(half_dim, dtype=torch.float32, device=timesteps.device)
+        exponent = exponent / (half_dim - 0.0)
+
+        emb = torch.exp(exponent)
+        emb = timesteps[:, None].float() * emb[None, :]
+
+        sin_emb = torch.sin(emb)
+        cos_emb = torch.cos(emb)
+        emb = torch.cat([cos_emb, sin_emb], dim=-1)
+
+        return rearrange(emb, "(b t) d -> b t d", b=timesteps_B_T.shape[0], t=timesteps_B_T.shape[1])
+
+
+class TimestepEmbedding(nn.Module):
+    def __init__(self, in_features: int, out_features: int, use_adaln_lora: bool = False, device=None, dtype=None, operations=None):
+        super().__init__()
+        logging.debug(
+            f"Using AdaLN LoRA Flag:  {use_adaln_lora}. We enable bias if no AdaLN LoRA for backward compatibility."
+        )
+        self.in_dim = in_features
+        self.out_dim = out_features
+        self.linear_1 = operations.Linear(in_features, out_features, bias=not use_adaln_lora, device=device, dtype=dtype)
+        self.activation = nn.SiLU()
+        self.use_adaln_lora = use_adaln_lora
+        if use_adaln_lora:
+            self.linear_2 = operations.Linear(out_features, 3 * out_features, bias=False, device=device, dtype=dtype)
+        else:
+            self.linear_2 = operations.Linear(out_features, out_features, bias=False, device=device, dtype=dtype)
+
+    def forward(self, sample: torch.Tensor) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        emb = self.linear_1(sample)
+        emb = self.activation(emb)
+        emb = self.linear_2(emb)
+
+        if self.use_adaln_lora:
+            adaln_lora_B_T_3D = emb
+            emb_B_T_D = sample
+        else:
+            adaln_lora_B_T_3D = None
+            emb_B_T_D = emb
+
+        return emb_B_T_D, adaln_lora_B_T_3D
+
+
+class PatchEmbed(nn.Module):
+    """
+    PatchEmbed is a module for embedding patches from an input tensor by applying either 3D or 2D convolutional layers,
+    depending on the . This module can process inputs with temporal (video) and spatial (image) dimensions,
+    making it suitable for video and image processing tasks. It supports dividing the input into patches
+    and embedding each patch into a vector of size `out_channels`.
+
+    Parameters:
+    - spatial_patch_size (int): The size of each spatial patch.
+    - temporal_patch_size (int): The size of each temporal patch.
+    - in_channels (int): Number of input channels. Default: 3.
+    - out_channels (int): The dimension of the embedding vector for each patch. Default: 768.
+    - bias (bool): If True, adds a learnable bias to the output of the convolutional layers. Default: True.
+    """
+
+    def __init__(
+        self,
+        spatial_patch_size: int,
+        temporal_patch_size: int,
+        in_channels: int = 3,
+        out_channels: int = 768,
+        device=None, dtype=None, operations=None
+    ):
+        super().__init__()
+        self.spatial_patch_size = spatial_patch_size
+        self.temporal_patch_size = temporal_patch_size
+
+        self.proj = nn.Sequential(
+            Rearrange(
+                "b c (t r) (h m) (w n) -> b t h w (c r m n)",
+                r=temporal_patch_size,
+                m=spatial_patch_size,
+                n=spatial_patch_size,
+            ),
+            operations.Linear(
+                in_channels * spatial_patch_size * spatial_patch_size * temporal_patch_size, out_channels, bias=False, device=device, dtype=dtype
+            ),
+        )
+        self.dim = in_channels * spatial_patch_size * spatial_patch_size * temporal_patch_size
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Forward pass of the PatchEmbed module.
+
+        Parameters:
+        - x (torch.Tensor): The input tensor of shape (B, C, T, H, W) where
+            B is the batch size,
+            C is the number of channels,
+            T is the temporal dimension,
+            H is the height, and
+            W is the width of the input.
+
+        Returns:
+        - torch.Tensor: The embedded patches as a tensor, with shape b t h w c.
+        """
+        assert x.dim() == 5
+        _, _, T, H, W = x.shape
+        assert (
+            H % self.spatial_patch_size == 0 and W % self.spatial_patch_size == 0
+        ), f"H,W {(H, W)} should be divisible by spatial_patch_size {self.spatial_patch_size}"
+        assert T % self.temporal_patch_size == 0
+        x = self.proj(x)
+        return x
+
+
+class FinalLayer(nn.Module):
+    """
+    The final layer of video DiT.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        spatial_patch_size: int,
+        temporal_patch_size: int,
+        out_channels: int,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 256,
+        device=None, dtype=None, operations=None
+    ):
+        super().__init__()
+        self.layer_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = operations.Linear(
+            hidden_size, spatial_patch_size * spatial_patch_size * temporal_patch_size * out_channels, bias=False, device=device, dtype=dtype
+        )
+        self.hidden_size = hidden_size
+        self.n_adaln_chunks = 2
+        self.use_adaln_lora = use_adaln_lora
+        self.adaln_lora_dim = adaln_lora_dim
+        if use_adaln_lora:
+            self.adaln_modulation = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(hidden_size, adaln_lora_dim, bias=False, device=device, dtype=dtype),
+                operations.Linear(adaln_lora_dim, self.n_adaln_chunks * hidden_size, bias=False, device=device, dtype=dtype),
+            )
+        else:
+            self.adaln_modulation = nn.Sequential(
+                nn.SiLU(), operations.Linear(hidden_size, self.n_adaln_chunks * hidden_size, bias=False, device=device, dtype=dtype)
+            )
+
+    def forward(
+        self,
+        x_B_T_H_W_D: torch.Tensor,
+        emb_B_T_D: torch.Tensor,
+        adaln_lora_B_T_3D: Optional[torch.Tensor] = None,
+    ):
+        if self.use_adaln_lora:
+            assert adaln_lora_B_T_3D is not None
+            shift_B_T_D, scale_B_T_D = (
+                self.adaln_modulation(emb_B_T_D) + adaln_lora_B_T_3D[:, :, : 2 * self.hidden_size]
+            ).chunk(2, dim=-1)
+        else:
+            shift_B_T_D, scale_B_T_D = self.adaln_modulation(emb_B_T_D).chunk(2, dim=-1)
+
+        shift_B_T_1_1_D, scale_B_T_1_1_D = rearrange(shift_B_T_D, "b t d -> b t 1 1 d"), rearrange(
+            scale_B_T_D, "b t d -> b t 1 1 d"
+        )
+
+        def _fn(
+            _x_B_T_H_W_D: torch.Tensor,
+            _norm_layer: nn.Module,
+            _scale_B_T_1_1_D: torch.Tensor,
+            _shift_B_T_1_1_D: torch.Tensor,
+        ) -> torch.Tensor:
+            return _norm_layer(_x_B_T_H_W_D) * (1 + _scale_B_T_1_1_D) + _shift_B_T_1_1_D
+
+        x_B_T_H_W_D = _fn(x_B_T_H_W_D, self.layer_norm, scale_B_T_1_1_D, shift_B_T_1_1_D)
+        x_B_T_H_W_O = self.linear(x_B_T_H_W_D)
+        return x_B_T_H_W_O
+
+
+class Block(nn.Module):
+    """
+    A transformer block that combines self-attention, cross-attention and MLP layers with AdaLN modulation.
+    Each component (self-attention, cross-attention, MLP) has its own layer normalization and AdaLN modulation.
+
+    Parameters:
+        x_dim (int): Dimension of input features
+        context_dim (int): Dimension of context features for cross-attention
+        num_heads (int): Number of attention heads
+        mlp_ratio (float): Multiplier for MLP hidden dimension. Default: 4.0
+        use_adaln_lora (bool): Whether to use AdaLN-LoRA modulation. Default: False
+        adaln_lora_dim (int): Hidden dimension for AdaLN-LoRA layers. Default: 256
+
+    The block applies the following sequence:
+    1. Self-attention with AdaLN modulation
+    2. Cross-attention with AdaLN modulation
+    3. MLP with AdaLN modulation
+
+    Each component uses skip connections and layer normalization.
+    """
+
+    def __init__(
+        self,
+        x_dim: int,
+        context_dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 256,
+        device=None,
+        dtype=None,
+        operations=None,
+    ):
+        super().__init__()
+        self.x_dim = x_dim
+        self.layer_norm_self_attn = operations.LayerNorm(x_dim, elementwise_affine=False, eps=1e-6, device=device, dtype=dtype)
+        self.self_attn = Attention(x_dim, None, num_heads, x_dim // num_heads, device=device, dtype=dtype, operations=operations)
+
+        self.layer_norm_cross_attn = operations.LayerNorm(x_dim, elementwise_affine=False, eps=1e-6, device=device, dtype=dtype)
+        self.cross_attn = Attention(
+            x_dim, context_dim, num_heads, x_dim // num_heads, device=device, dtype=dtype, operations=operations
+        )
+
+        self.layer_norm_mlp = operations.LayerNorm(x_dim, elementwise_affine=False, eps=1e-6, device=device, dtype=dtype)
+        self.mlp = GPT2FeedForward(x_dim, int(x_dim * mlp_ratio), device=device, dtype=dtype, operations=operations)
+
+        self.use_adaln_lora = use_adaln_lora
+        if self.use_adaln_lora:
+            self.adaln_modulation_self_attn = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(x_dim, adaln_lora_dim, bias=False, device=device, dtype=dtype),
+                operations.Linear(adaln_lora_dim, 3 * x_dim, bias=False, device=device, dtype=dtype),
+            )
+            self.adaln_modulation_cross_attn = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(x_dim, adaln_lora_dim, bias=False, device=device, dtype=dtype),
+                operations.Linear(adaln_lora_dim, 3 * x_dim, bias=False, device=device, dtype=dtype),
+            )
+            self.adaln_modulation_mlp = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(x_dim, adaln_lora_dim, bias=False, device=device, dtype=dtype),
+                operations.Linear(adaln_lora_dim, 3 * x_dim, bias=False, device=device, dtype=dtype),
+            )
+        else:
+            self.adaln_modulation_self_attn = nn.Sequential(nn.SiLU(), operations.Linear(x_dim, 3 * x_dim, bias=False, device=device, dtype=dtype))
+            self.adaln_modulation_cross_attn = nn.Sequential(nn.SiLU(), operations.Linear(x_dim, 3 * x_dim, bias=False, device=device, dtype=dtype))
+            self.adaln_modulation_mlp = nn.Sequential(nn.SiLU(), operations.Linear(x_dim, 3 * x_dim, bias=False, device=device, dtype=dtype))
+
+    def forward(
+        self,
+        x_B_T_H_W_D: torch.Tensor,
+        emb_B_T_D: torch.Tensor,
+        crossattn_emb: torch.Tensor,
+        rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
+        adaln_lora_B_T_3D: Optional[torch.Tensor] = None,
+        extra_per_block_pos_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if extra_per_block_pos_emb is not None:
+            x_B_T_H_W_D = x_B_T_H_W_D + extra_per_block_pos_emb
+
+        if self.use_adaln_lora:
+            shift_self_attn_B_T_D, scale_self_attn_B_T_D, gate_self_attn_B_T_D = (
+                self.adaln_modulation_self_attn(emb_B_T_D) + adaln_lora_B_T_3D
+            ).chunk(3, dim=-1)
+            shift_cross_attn_B_T_D, scale_cross_attn_B_T_D, gate_cross_attn_B_T_D = (
+                self.adaln_modulation_cross_attn(emb_B_T_D) + adaln_lora_B_T_3D
+            ).chunk(3, dim=-1)
+            shift_mlp_B_T_D, scale_mlp_B_T_D, gate_mlp_B_T_D = (
+                self.adaln_modulation_mlp(emb_B_T_D) + adaln_lora_B_T_3D
+            ).chunk(3, dim=-1)
+        else:
+            shift_self_attn_B_T_D, scale_self_attn_B_T_D, gate_self_attn_B_T_D = self.adaln_modulation_self_attn(
+                emb_B_T_D
+            ).chunk(3, dim=-1)
+            shift_cross_attn_B_T_D, scale_cross_attn_B_T_D, gate_cross_attn_B_T_D = self.adaln_modulation_cross_attn(
+                emb_B_T_D
+            ).chunk(3, dim=-1)
+            shift_mlp_B_T_D, scale_mlp_B_T_D, gate_mlp_B_T_D = self.adaln_modulation_mlp(emb_B_T_D).chunk(3, dim=-1)
+
+        # Reshape tensors from (B, T, D) to (B, T, 1, 1, D) for broadcasting
+        shift_self_attn_B_T_1_1_D = rearrange(shift_self_attn_B_T_D, "b t d -> b t 1 1 d")
+        scale_self_attn_B_T_1_1_D = rearrange(scale_self_attn_B_T_D, "b t d -> b t 1 1 d")
+        gate_self_attn_B_T_1_1_D = rearrange(gate_self_attn_B_T_D, "b t d -> b t 1 1 d")
+
+        shift_cross_attn_B_T_1_1_D = rearrange(shift_cross_attn_B_T_D, "b t d -> b t 1 1 d")
+        scale_cross_attn_B_T_1_1_D = rearrange(scale_cross_attn_B_T_D, "b t d -> b t 1 1 d")
+        gate_cross_attn_B_T_1_1_D = rearrange(gate_cross_attn_B_T_D, "b t d -> b t 1 1 d")
+
+        shift_mlp_B_T_1_1_D = rearrange(shift_mlp_B_T_D, "b t d -> b t 1 1 d")
+        scale_mlp_B_T_1_1_D = rearrange(scale_mlp_B_T_D, "b t d -> b t 1 1 d")
+        gate_mlp_B_T_1_1_D = rearrange(gate_mlp_B_T_D, "b t d -> b t 1 1 d")
+
+        B, T, H, W, D = x_B_T_H_W_D.shape
+
+        def _fn(_x_B_T_H_W_D, _norm_layer, _scale_B_T_1_1_D, _shift_B_T_1_1_D):
+            return _norm_layer(_x_B_T_H_W_D) * (1 + _scale_B_T_1_1_D) + _shift_B_T_1_1_D
+
+        normalized_x_B_T_H_W_D = _fn(
+            x_B_T_H_W_D,
+            self.layer_norm_self_attn,
+            scale_self_attn_B_T_1_1_D,
+            shift_self_attn_B_T_1_1_D,
+        )
+        result_B_T_H_W_D = rearrange(
+            self.self_attn(
+                # normalized_x_B_T_HW_D,
+                rearrange(normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
+                None,
+                rope_emb=rope_emb_L_1_1_D,
+            ),
+            "b (t h w) d -> b t h w d",
+            t=T,
+            h=H,
+            w=W,
+        )
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D * result_B_T_H_W_D
+
+        def _x_fn(
+            _x_B_T_H_W_D: torch.Tensor,
+            layer_norm_cross_attn: Callable,
+            _scale_cross_attn_B_T_1_1_D: torch.Tensor,
+            _shift_cross_attn_B_T_1_1_D: torch.Tensor,
+        ) -> torch.Tensor:
+            _normalized_x_B_T_H_W_D = _fn(
+                _x_B_T_H_W_D, layer_norm_cross_attn, _scale_cross_attn_B_T_1_1_D, _shift_cross_attn_B_T_1_1_D
+            )
+            _result_B_T_H_W_D = rearrange(
+                self.cross_attn(
+                    rearrange(_normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
+                    crossattn_emb,
+                    rope_emb=rope_emb_L_1_1_D,
+                ),
+                "b (t h w) d -> b t h w d",
+                t=T,
+                h=H,
+                w=W,
+            )
+            return _result_B_T_H_W_D
+
+        result_B_T_H_W_D = _x_fn(
+            x_B_T_H_W_D,
+            self.layer_norm_cross_attn,
+            scale_cross_attn_B_T_1_1_D,
+            shift_cross_attn_B_T_1_1_D,
+        )
+        x_B_T_H_W_D = result_B_T_H_W_D * gate_cross_attn_B_T_1_1_D + x_B_T_H_W_D
+
+        normalized_x_B_T_H_W_D = _fn(
+            x_B_T_H_W_D,
+            self.layer_norm_mlp,
+            scale_mlp_B_T_1_1_D,
+            shift_mlp_B_T_1_1_D,
+        )
+        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D)
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D * result_B_T_H_W_D
+        return x_B_T_H_W_D
+
+
+class MiniTrainDIT(nn.Module):
+    """
+    A clean impl of DIT that can load and  reproduce the training results of the original DIT model in~(cosmos 1)
+    A general implementation of adaln-modulated VIT-like~(DiT) transformer for video processing.
+
+    Args:
+        max_img_h (int): Maximum height of the input images.
+        max_img_w (int): Maximum width of the input images.
+        max_frames (int): Maximum number of frames in the video sequence.
+        in_channels (int): Number of input channels (e.g., RGB channels for color images).
+        out_channels (int): Number of output channels.
+        patch_spatial (tuple): Spatial resolution of patches for input processing.
+        patch_temporal (int): Temporal resolution of patches for input processing.
+        concat_padding_mask (bool): If True, includes a mask channel in the input to handle padding.
+        model_channels (int): Base number of channels used throughout the model.
+        num_blocks (int): Number of transformer blocks.
+        num_heads (int): Number of heads in the multi-head attention layers.
+        mlp_ratio (float): Expansion ratio for MLP blocks.
+        crossattn_emb_channels (int): Number of embedding channels for cross-attention.
+        pos_emb_cls (str): Type of positional embeddings.
+        pos_emb_learnable (bool): Whether positional embeddings are learnable.
+        pos_emb_interpolation (str): Method for interpolating positional embeddings.
+        min_fps (int): Minimum frames per second.
+        max_fps (int): Maximum frames per second.
+        use_adaln_lora (bool): Whether to use AdaLN-LoRA.
+        adaln_lora_dim (int): Dimension for AdaLN-LoRA.
+        rope_h_extrapolation_ratio (float): Height extrapolation ratio for RoPE.
+        rope_w_extrapolation_ratio (float): Width extrapolation ratio for RoPE.
+        rope_t_extrapolation_ratio (float): Temporal extrapolation ratio for RoPE.
+        extra_per_block_abs_pos_emb (bool): Whether to use extra per-block absolute positional embeddings.
+        extra_h_extrapolation_ratio (float): Height extrapolation ratio for extra embeddings.
+        extra_w_extrapolation_ratio (float): Width extrapolation ratio for extra embeddings.
+        extra_t_extrapolation_ratio (float): Temporal extrapolation ratio for extra embeddings.
+    """
+
+    def __init__(
+        self,
+        max_img_h: int,
+        max_img_w: int,
+        max_frames: int,
+        in_channels: int,
+        out_channels: int,
+        patch_spatial: int,  # tuple,
+        patch_temporal: int,
+        concat_padding_mask: bool = True,
+        # attention settings
+        model_channels: int = 768,
+        num_blocks: int = 10,
+        num_heads: int = 16,
+        mlp_ratio: float = 4.0,
+        # cross attention settings
+        crossattn_emb_channels: int = 1024,
+        # positional embedding settings
+        pos_emb_cls: str = "sincos",
+        pos_emb_learnable: bool = False,
+        pos_emb_interpolation: str = "crop",
+        min_fps: int = 1,
+        max_fps: int = 30,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 256,
+        rope_h_extrapolation_ratio: float = 1.0,
+        rope_w_extrapolation_ratio: float = 1.0,
+        rope_t_extrapolation_ratio: float = 1.0,
+        extra_per_block_abs_pos_emb: bool = False,
+        extra_h_extrapolation_ratio: float = 1.0,
+        extra_w_extrapolation_ratio: float = 1.0,
+        extra_t_extrapolation_ratio: float = 1.0,
+        rope_enable_fps_modulation: bool = True,
+        image_model=None,
+        device=None,
+        dtype=None,
+        operations=None,
+    ) -> None:
+        super().__init__()
+        self.dtype = dtype
+        self.max_img_h = max_img_h
+        self.max_img_w = max_img_w
+        self.max_frames = max_frames
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.patch_spatial = patch_spatial
+        self.patch_temporal = patch_temporal
+        self.num_heads = num_heads
+        self.num_blocks = num_blocks
+        self.model_channels = model_channels
+        self.concat_padding_mask = concat_padding_mask
+        # positional embedding settings
+        self.pos_emb_cls = pos_emb_cls
+        self.pos_emb_learnable = pos_emb_learnable
+        self.pos_emb_interpolation = pos_emb_interpolation
+        self.min_fps = min_fps
+        self.max_fps = max_fps
+        self.rope_h_extrapolation_ratio = rope_h_extrapolation_ratio
+        self.rope_w_extrapolation_ratio = rope_w_extrapolation_ratio
+        self.rope_t_extrapolation_ratio = rope_t_extrapolation_ratio
+        self.extra_per_block_abs_pos_emb = extra_per_block_abs_pos_emb
+        self.extra_h_extrapolation_ratio = extra_h_extrapolation_ratio
+        self.extra_w_extrapolation_ratio = extra_w_extrapolation_ratio
+        self.extra_t_extrapolation_ratio = extra_t_extrapolation_ratio
+        self.rope_enable_fps_modulation = rope_enable_fps_modulation
+
+        self.build_pos_embed(device=device, dtype=dtype)
+        self.use_adaln_lora = use_adaln_lora
+        self.adaln_lora_dim = adaln_lora_dim
+        self.t_embedder = nn.Sequential(
+            Timesteps(model_channels),
+            TimestepEmbedding(model_channels, model_channels, use_adaln_lora=use_adaln_lora, device=device, dtype=dtype, operations=operations,),
+        )
+
+        in_channels = in_channels + 1 if concat_padding_mask else in_channels
+        self.x_embedder = PatchEmbed(
+            spatial_patch_size=patch_spatial,
+            temporal_patch_size=patch_temporal,
+            in_channels=in_channels,
+            out_channels=model_channels,
+            device=device, dtype=dtype, operations=operations,
+        )
+
+        self.blocks = nn.ModuleList(
+            [
+                Block(
+                    x_dim=model_channels,
+                    context_dim=crossattn_emb_channels,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    use_adaln_lora=use_adaln_lora,
+                    adaln_lora_dim=adaln_lora_dim,
+                    device=device, dtype=dtype, operations=operations,
+                )
+                for _ in range(num_blocks)
+            ]
+        )
+
+        self.final_layer = FinalLayer(
+            hidden_size=self.model_channels,
+            spatial_patch_size=self.patch_spatial,
+            temporal_patch_size=self.patch_temporal,
+            out_channels=self.out_channels,
+            use_adaln_lora=self.use_adaln_lora,
+            adaln_lora_dim=self.adaln_lora_dim,
+            device=device, dtype=dtype, operations=operations,
+        )
+
+        self.t_embedding_norm = operations.RMSNorm(model_channels, eps=1e-6, device=device, dtype=dtype)
+
+    def build_pos_embed(self, device=None, dtype=None) -> None:
+        if self.pos_emb_cls == "rope3d":
+            cls_type = VideoRopePosition3DEmb
+        else:
+            raise ValueError(f"Unknown pos_emb_cls {self.pos_emb_cls}")
+
+        logging.debug(f"Building positional embedding with {self.pos_emb_cls} class, impl {cls_type}")
+        kwargs = dict(
+            model_channels=self.model_channels,
+            len_h=self.max_img_h // self.patch_spatial,
+            len_w=self.max_img_w // self.patch_spatial,
+            len_t=self.max_frames // self.patch_temporal,
+            max_fps=self.max_fps,
+            min_fps=self.min_fps,
+            is_learnable=self.pos_emb_learnable,
+            interpolation=self.pos_emb_interpolation,
+            head_dim=self.model_channels // self.num_heads,
+            h_extrapolation_ratio=self.rope_h_extrapolation_ratio,
+            w_extrapolation_ratio=self.rope_w_extrapolation_ratio,
+            t_extrapolation_ratio=self.rope_t_extrapolation_ratio,
+            enable_fps_modulation=self.rope_enable_fps_modulation,
+            device=device,
+        )
+        self.pos_embedder = cls_type(
+            **kwargs,  # type: ignore
+        )
+
+        if self.extra_per_block_abs_pos_emb:
+            kwargs["h_extrapolation_ratio"] = self.extra_h_extrapolation_ratio
+            kwargs["w_extrapolation_ratio"] = self.extra_w_extrapolation_ratio
+            kwargs["t_extrapolation_ratio"] = self.extra_t_extrapolation_ratio
+            kwargs["device"] = device
+            kwargs["dtype"] = dtype
+            self.extra_pos_embedder = LearnablePosEmbAxis(
+                **kwargs,  # type: ignore
+            )
+
+    def prepare_embedded_sequence(
+        self,
+        x_B_C_T_H_W: torch.Tensor,
+        fps: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]]:
+        """
+        Prepares an embedded sequence tensor by applying positional embeddings and handling padding masks.
+
+        Args:
+            x_B_C_T_H_W (torch.Tensor): video
+            fps (Optional[torch.Tensor]): Frames per second tensor to be used for positional embedding when required.
+                                    If None, a default value (`self.base_fps`) will be used.
+            padding_mask (Optional[torch.Tensor]): current it is not used
+
+        Returns:
+            Tuple[torch.Tensor, Optional[torch.Tensor]]:
+                - A tensor of shape (B, T, H, W, D) with the embedded sequence.
+                - An optional positional embedding tensor, returned only if the positional embedding class
+                (`self.pos_emb_cls`) includes 'rope'. Otherwise, None.
+
+        Notes:
+            - If `self.concat_padding_mask` is True, a padding mask channel is concatenated to the input tensor.
+            - The method of applying positional embeddings depends on the value of `self.pos_emb_cls`.
+            - If 'rope' is in `self.pos_emb_cls` (case insensitive), the positional embeddings are generated using
+                the `self.pos_embedder` with the shape [T, H, W].
+            - If "fps_aware" is in `self.pos_emb_cls`, the positional embeddings are generated using the
+            `self.pos_embedder` with the fps tensor.
+            - Otherwise, the positional embeddings are generated without considering fps.
+        """
+        if self.concat_padding_mask:
+            if padding_mask is None:
+                padding_mask = torch.zeros(x_B_C_T_H_W.shape[0], 1, x_B_C_T_H_W.shape[3], x_B_C_T_H_W.shape[4], dtype=x_B_C_T_H_W.dtype, device=x_B_C_T_H_W.device)
+            else:
+                padding_mask = transforms.functional.resize(
+                    padding_mask, list(x_B_C_T_H_W.shape[-2:]), interpolation=transforms.InterpolationMode.NEAREST
+                )
+            x_B_C_T_H_W = torch.cat(
+                [x_B_C_T_H_W, padding_mask.unsqueeze(1).repeat(1, 1, x_B_C_T_H_W.shape[2], 1, 1)], dim=1
+            )
+        x_B_T_H_W_D = self.x_embedder(x_B_C_T_H_W)
+
+        if self.extra_per_block_abs_pos_emb:
+            extra_pos_emb = self.extra_pos_embedder(x_B_T_H_W_D, fps=fps, device=x_B_C_T_H_W.device, dtype=x_B_C_T_H_W.dtype)
+        else:
+            extra_pos_emb = None
+
+        if "rope" in self.pos_emb_cls.lower():
+            return x_B_T_H_W_D, self.pos_embedder(x_B_T_H_W_D, fps=fps, device=x_B_C_T_H_W.device), extra_pos_emb
+        x_B_T_H_W_D = x_B_T_H_W_D + self.pos_embedder(x_B_T_H_W_D, device=x_B_C_T_H_W.device)  # [B, T, H, W, D]
+
+        return x_B_T_H_W_D, None, extra_pos_emb
+
+    def unpatchify(self, x_B_T_H_W_M: torch.Tensor) -> torch.Tensor:
+        x_B_C_Tt_Hp_Wp = rearrange(
+            x_B_T_H_W_M,
+            "B T H W (p1 p2 t C) -> B C (T t) (H p1) (W p2)",
+            p1=self.patch_spatial,
+            p2=self.patch_spatial,
+            t=self.patch_temporal,
+        )
+        return x_B_C_Tt_Hp_Wp
+
+    def forward(self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        context: torch.Tensor,
+        fps: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, kwargs.get("transformer_options", {}))
+        ).execute(x, timesteps, context, fps, padding_mask, **kwargs)
+
+    def _forward(
+        self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        context: torch.Tensor,
+        fps: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ):
+        x_B_C_T_H_W = x
+        timesteps_B_T = timesteps
+        crossattn_emb = context
+        """
+        Args:
+            x: (B, C, T, H, W) tensor of spatial-temp inputs
+            timesteps: (B, ) tensor of timesteps
+            crossattn_emb: (B, N, D) tensor of cross-attention embeddings
+        """
+        x_B_T_H_W_D, rope_emb_L_1_1_D, extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D = self.prepare_embedded_sequence(
+            x_B_C_T_H_W,
+            fps=fps,
+            padding_mask=padding_mask,
+        )
+
+        if timesteps_B_T.ndim == 1:
+            timesteps_B_T = timesteps_B_T.unsqueeze(1)
+        t_embedding_B_T_D, adaln_lora_B_T_3D = self.t_embedder[1](self.t_embedder[0](timesteps_B_T).to(x_B_T_H_W_D.dtype))
+        t_embedding_B_T_D = self.t_embedding_norm(t_embedding_B_T_D)
+
+        # for logging purpose
+        affline_scale_log_info = {}
+        affline_scale_log_info["t_embedding_B_T_D"] = t_embedding_B_T_D.detach()
+        self.affline_scale_log_info = affline_scale_log_info
+        self.affline_emb = t_embedding_B_T_D
+        self.crossattn_emb = crossattn_emb
+
+        if extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D is not None:
+            assert (
+                x_B_T_H_W_D.shape == extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D.shape
+            ), f"{x_B_T_H_W_D.shape} != {extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D.shape}"
+
+        block_kwargs = {
+            "rope_emb_L_1_1_D": rope_emb_L_1_1_D.unsqueeze(1).unsqueeze(0),
+            "adaln_lora_B_T_3D": adaln_lora_B_T_3D,
+            "extra_per_block_pos_emb": extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D,
+        }
+        for block in self.blocks:
+            x_B_T_H_W_D = block(
+                x_B_T_H_W_D,
+                t_embedding_B_T_D,
+                crossattn_emb,
+                **block_kwargs,
+            )
+
+        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D, t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
+        x_B_C_Tt_Hp_Wp = self.unpatchify(x_B_T_H_W_O)
+        return x_B_C_Tt_Hp_Wp

comfy/ldm/cosmos/vae.py

+# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""The causal continuous video tokenizer with VAE or AE formulation for 3D data.."""
+
+import logging
+import torch
+from torch import nn
+from enum import Enum
+import math
+
+from .cosmos_tokenizer.layers3d import (
+    EncoderFactorized,
+    DecoderFactorized,
+    CausalConv3d,
+)
+
+
+class IdentityDistribution(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, parameters):
+        return parameters, (torch.tensor([0.0]), torch.tensor([0.0]))
+
+
+class GaussianDistribution(torch.nn.Module):
+    def __init__(self, min_logvar: float = -30.0, max_logvar: float = 20.0):
+        super().__init__()
+        self.min_logvar = min_logvar
+        self.max_logvar = max_logvar
+
+    def sample(self, mean, logvar):
+        std = torch.exp(0.5 * logvar)
+        return mean + std * torch.randn_like(mean)
+
+    def forward(self, parameters):
+        mean, logvar = torch.chunk(parameters, 2, dim=1)
+        logvar = torch.clamp(logvar, self.min_logvar, self.max_logvar)
+        return self.sample(mean, logvar), (mean, logvar)
+
+
+class ContinuousFormulation(Enum):
+    VAE = GaussianDistribution
+    AE = IdentityDistribution
+
+
+class CausalContinuousVideoTokenizer(nn.Module):
+    def __init__(
+        self, z_channels: int, z_factor: int, latent_channels: int, **kwargs
+    ) -> None:
+        super().__init__()
+        self.name = kwargs.get("name", "CausalContinuousVideoTokenizer")
+        self.latent_channels = latent_channels
+        self.sigma_data = 0.5
+
+        # encoder_name = kwargs.get("encoder", Encoder3DType.BASE.name)
+        self.encoder = EncoderFactorized(
+            z_channels=z_factor * z_channels, **kwargs
+        )
+        if kwargs.get("temporal_compression", 4) == 4:
+            kwargs["channels_mult"] = [2, 4]
+        # decoder_name = kwargs.get("decoder", Decoder3DType.BASE.name)
+        self.decoder = DecoderFactorized(
+            z_channels=z_channels, **kwargs
+        )
+
+        self.quant_conv = CausalConv3d(
+            z_factor * z_channels,
+            z_factor * latent_channels,
+            kernel_size=1,
+            padding=0,
+        )
+        self.post_quant_conv = CausalConv3d(
+            latent_channels, z_channels, kernel_size=1, padding=0
+        )
+
+        # formulation_name = kwargs.get("formulation", ContinuousFormulation.AE.name)
+        self.distribution = IdentityDistribution()  # ContinuousFormulation[formulation_name].value()
+
+        num_parameters = sum(param.numel() for param in self.parameters())
+        logging.debug(f"model={self.name}, num_parameters={num_parameters:,}")
+        logging.debug(
+            f"z_channels={z_channels}, latent_channels={self.latent_channels}."
+        )
+
+        latent_temporal_chunk = 16
+        self.latent_mean = nn.Parameter(torch.zeros([self.latent_channels * latent_temporal_chunk], dtype=torch.float32))
+        self.latent_std = nn.Parameter(torch.ones([self.latent_channels * latent_temporal_chunk], dtype=torch.float32))
+
+
+    def encode(self, x):
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        z, posteriors = self.distribution(moments)
+        latent_ch = z.shape[1]
+        latent_t = z.shape[2]
+        in_dtype = z.dtype
+        mean = self.latent_mean.view(latent_ch, -1)
+        std = self.latent_std.view(latent_ch, -1)
+
+        mean = mean.repeat(1, math.ceil(latent_t / mean.shape[-1]))[:, : latent_t].reshape([1, latent_ch, -1, 1, 1]).to(dtype=in_dtype, device=z.device)
+        std = std.repeat(1, math.ceil(latent_t / std.shape[-1]))[:, : latent_t].reshape([1, latent_ch, -1, 1, 1]).to(dtype=in_dtype, device=z.device)
+        return ((z - mean) / std) * self.sigma_data
+
+    def decode(self, z):
+        in_dtype = z.dtype
+        latent_ch = z.shape[1]
+        latent_t = z.shape[2]
+        mean = self.latent_mean.view(latent_ch, -1)
+        std = self.latent_std.view(latent_ch, -1)
+
+        mean = mean.repeat(1, math.ceil(latent_t / mean.shape[-1]))[:, : latent_t].reshape([1, latent_ch, -1, 1, 1]).to(dtype=in_dtype, device=z.device)
+        std = std.repeat(1, math.ceil(latent_t / std.shape[-1]))[:, : latent_t].reshape([1, latent_ch, -1, 1, 1]).to(dtype=in_dtype, device=z.device)
+
+        z = z / self.sigma_data
+        z = z * std + mean
+        z = self.post_quant_conv(z)
+        return self.decoder(z)
+

comfy/ldm/flux/controlnet.py

+#Original code can be found on: https://github.com/XLabs-AI/x-flux/blob/main/src/flux/controlnet.py
+#modified to support different types of flux controlnets
+
+import torch
+import math
+from torch import Tensor, nn
+from einops import rearrange, repeat
+
+from .layers import (timestep_embedding)
+
+from .model import Flux
+import comfy.ldm.common_dit
+
+class MistolineCondDownsamplBlock(nn.Module):
+    def __init__(self, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.encoder = nn.Sequential(
+            operations.Conv2d(3, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 1, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device)
+        )
+
+    def forward(self, x):
+        return self.encoder(x)
+
+class MistolineControlnetBlock(nn.Module):
+    def __init__(self, hidden_size, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.linear = operations.Linear(hidden_size, hidden_size, dtype=dtype, device=device)
+        self.act = nn.SiLU()
+
+    def forward(self, x):
+        return self.act(self.linear(x))
+
+
+class ControlNetFlux(Flux):
+    def __init__(self, latent_input=False, num_union_modes=0, mistoline=False, control_latent_channels=None, image_model=None, dtype=None, device=None, operations=None, **kwargs):
+        super().__init__(final_layer=False, dtype=dtype, device=device, operations=operations, **kwargs)
+
+        self.main_model_double = 19
+        self.main_model_single = 38
+
+        self.mistoline = mistoline
+        # add ControlNet blocks
+        if self.mistoline:
+            control_block = lambda : MistolineControlnetBlock(self.hidden_size, dtype=dtype, device=device, operations=operations)
+        else:
+            control_block = lambda : operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
+
+        self.controlnet_blocks = nn.ModuleList([])
+        for _ in range(self.params.depth):
+            self.controlnet_blocks.append(control_block())
+
+        self.controlnet_single_blocks = nn.ModuleList([])
+        for _ in range(self.params.depth_single_blocks):
+            self.controlnet_single_blocks.append(control_block())
+
+        self.num_union_modes = num_union_modes
+        self.controlnet_mode_embedder = None
+        if self.num_union_modes > 0:
+            self.controlnet_mode_embedder = operations.Embedding(self.num_union_modes, self.hidden_size, dtype=dtype, device=device)
+
+        self.gradient_checkpointing = False
+        self.latent_input = latent_input
+        if control_latent_channels is None:
+            control_latent_channels = self.in_channels
+        else:
+            control_latent_channels *= 2 * 2 #patch size
+
+        self.pos_embed_input = operations.Linear(control_latent_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
+        if not self.latent_input:
+            if self.mistoline:
+                self.input_cond_block = MistolineCondDownsamplBlock(dtype=dtype, device=device, operations=operations)
+            else:
+                self.input_hint_block = nn.Sequential(
+                    operations.Conv2d(3, 16, 3, padding=1, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, stride=2, dtype=dtype, device=device),
+                    nn.SiLU(),
+                    operations.Conv2d(16, 16, 3, padding=1, dtype=dtype, device=device)
+                )
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        controlnet_cond: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        y: Tensor,
+        guidance: Tensor = None,
+        control_type: Tensor = None,
+    ) -> Tensor:
+        if img.ndim != 3 or txt.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        if y is None:
+            y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype)
+        else:
+            y = y[:, :self.params.vec_in_dim]
+
+        # running on sequences img
+        img = self.img_in(img)
+
+        controlnet_cond = self.pos_embed_input(controlnet_cond)
+        img = img + controlnet_cond
+        vec = self.time_in(timestep_embedding(timesteps, 256))
+        if self.params.guidance_embed:
+            vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
+        vec = vec + self.vector_in(y)
+        txt = self.txt_in(txt)
+
+        if self.controlnet_mode_embedder is not None and len(control_type) > 0:
+            control_cond = self.controlnet_mode_embedder(torch.tensor(control_type, device=img.device), out_dtype=img.dtype).unsqueeze(0).repeat((txt.shape[0], 1, 1))
+            txt = torch.cat([control_cond, txt], dim=1)
+            txt_ids = torch.cat([txt_ids[:,:1], txt_ids], dim=1)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        controlnet_double = ()
+
+        for i in range(len(self.double_blocks)):
+            img, txt = self.double_blocks[i](img=img, txt=txt, vec=vec, pe=pe)
+            controlnet_double = controlnet_double + (self.controlnet_blocks[i](img),)
+
+        img = torch.cat((txt, img), 1)
+
+        controlnet_single = ()
+
+        for i in range(len(self.single_blocks)):
+            img = self.single_blocks[i](img, vec=vec, pe=pe)
+            controlnet_single = controlnet_single + (self.controlnet_single_blocks[i](img[:, txt.shape[1] :, ...]),)
+
+        repeat = math.ceil(self.main_model_double / len(controlnet_double))
+        if self.latent_input:
+            out_input = ()
+            for x in controlnet_double:
+                    out_input += (x,) * repeat
+        else:
+            out_input = (controlnet_double * repeat)
+
+        out = {"input": out_input[:self.main_model_double]}
+        if len(controlnet_single) > 0:
+            repeat = math.ceil(self.main_model_single / len(controlnet_single))
+            out_output = ()
+            if self.latent_input:
+                for x in controlnet_single:
+                        out_output += (x,) * repeat
+            else:
+                out_output = (controlnet_single * repeat)
+            out["output"] = out_output[:self.main_model_single]
+        return out
+
+    def forward(self, x, timesteps, context, y=None, guidance=None, hint=None, **kwargs):
+        patch_size = 2
+        if self.latent_input:
+            hint = comfy.ldm.common_dit.pad_to_patch_size(hint, (patch_size, patch_size))
+        elif self.mistoline:
+            hint = hint * 2.0 - 1.0
+            hint = self.input_cond_block(hint)
+        else:
+            hint = hint * 2.0 - 1.0
+            hint = self.input_hint_block(hint)
+
+        hint = rearrange(hint, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
+
+        bs, c, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
+
+        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
+
+        h_len = ((h + (patch_size // 2)) // patch_size)
+        w_len = ((w + (patch_size // 2)) // patch_size)
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[..., 1] = img_ids[..., 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype)[:, None]
+        img_ids[..., 2] = img_ids[..., 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype)[None, :]
+        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        return self.forward_orig(img, img_ids, hint, context, txt_ids, timesteps, y, guidance, control_type=kwargs.get("control_type", []))

comfy/ldm/flux/layers.py

+import math
+from dataclasses import dataclass
+
+import torch
+from torch import Tensor, nn
+
+from .math import attention, rope
+import comfy.ops
+import comfy.ldm.common_dit
+
+
+class EmbedND(nn.Module):
+    def __init__(self, dim: int, theta: int, axes_dim: list):
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+        self.axes_dim = axes_dim
+
+    def forward(self, ids: Tensor) -> 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(1)
+
+
+def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0):
+    """
+    Create sinusoidal timestep embeddings.
+    :param t: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an (N, D) Tensor of positional embeddings.
+    """
+    t = time_factor * t
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=t.device) / half)
+
+    args = t[:, None].float() * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    if torch.is_floating_point(t):
+        embedding = embedding.to(t)
+    return embedding
+
+class MLPEmbedder(nn.Module):
+    def __init__(self, in_dim: int, hidden_dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.in_layer = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.silu = nn.SiLU()
+        self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.out_layer(self.silu(self.in_layer(x)))
+
+
+class RMSNorm(torch.nn.Module):
+    def __init__(self, dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.scale = nn.Parameter(torch.empty((dim), dtype=dtype, device=device))
+
+    def forward(self, x: Tensor):
+        return comfy.ldm.common_dit.rms_norm(x, self.scale, 1e-6)
+
+
+class QKNorm(torch.nn.Module):
+    def __init__(self, dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.query_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
+        self.key_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple:
+        q = self.query_norm(q)
+        k = self.key_norm(k)
+        return q.to(v), k.to(v)
+
+
+class SelfAttention(nn.Module):
+    def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+
+        self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+        self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
+        self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
+
+
+@dataclass
+class ModulationOut:
+    shift: Tensor
+    scale: Tensor
+    gate: Tensor
+
+
+class Modulation(nn.Module):
+    def __init__(self, dim: int, double: bool, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.is_double = double
+        self.multiplier = 6 if double else 3
+        self.lin = operations.Linear(dim, self.multiplier * dim, bias=True, dtype=dtype, device=device)
+
+    def forward(self, vec: Tensor) -> tuple:
+        if vec.ndim == 2:
+            vec = vec[:, None, :]
+        out = self.lin(nn.functional.silu(vec)).chunk(self.multiplier, dim=-1)
+
+        return (
+            ModulationOut(*out[:3]),
+            ModulationOut(*out[3:]) if self.is_double else None,
+        )
+
+
+def apply_mod(tensor, m_mult, m_add=None, modulation_dims=None):
+    if modulation_dims is None:
+        if m_add is not None:
+            return torch.addcmul(m_add, tensor, m_mult)
+        else:
+            return tensor * m_mult
+    else:
+        for d in modulation_dims:
+            tensor[:, d[0]:d[1]] *= m_mult[:, d[2]]
+            if m_add is not None:
+                tensor[:, d[0]:d[1]] += m_add[:, d[2]]
+        return tensor
+
+
+class DoubleStreamBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.img_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
+        self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+
+        self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.img_mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+        self.txt_mod = Modulation(hidden_size, double=True, dtype=dtype, device=device, operations=operations)
+        self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+
+        self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.txt_mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+        self.flipped_img_txt = flipped_img_txt
+
+    def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None):
+        img_mod1, img_mod2 = self.img_mod(vec)
+        txt_mod1, txt_mod2 = self.txt_mod(vec)
+
+        # prepare image for attention
+        img_modulated = self.img_norm1(img)
+        img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img)
+        img_qkv = self.img_attn.qkv(img_modulated)
+        img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
+
+        # prepare txt for attention
+        txt_modulated = self.txt_norm1(txt)
+        txt_modulated = apply_mod(txt_modulated, (1 + txt_mod1.scale), txt_mod1.shift, modulation_dims_txt)
+        txt_qkv = self.txt_attn.qkv(txt_modulated)
+        txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
+
+        if self.flipped_img_txt:
+            # run actual attention
+            attn = attention(torch.cat((img_q, txt_q), dim=2),
+                             torch.cat((img_k, txt_k), dim=2),
+                             torch.cat((img_v, txt_v), dim=2),
+                             pe=pe, mask=attn_mask)
+
+            img_attn, txt_attn = attn[:, : img.shape[1]], attn[:, img.shape[1]:]
+        else:
+            # run actual attention
+            attn = attention(torch.cat((txt_q, img_q), dim=2),
+                             torch.cat((txt_k, img_k), dim=2),
+                             torch.cat((txt_v, img_v), dim=2),
+                             pe=pe, mask=attn_mask)
+
+            txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
+
+        # calculate the img bloks
+        img = img + apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
+        img = img + apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img)
+
+        # calculate the txt bloks
+        txt += apply_mod(self.txt_attn.proj(txt_attn), txt_mod1.gate, None, modulation_dims_txt)
+        txt += apply_mod(self.txt_mlp(apply_mod(self.txt_norm2(txt), (1 + txt_mod2.scale), txt_mod2.shift, modulation_dims_txt)), txt_mod2.gate, None, modulation_dims_txt)
+
+        if txt.dtype == torch.float16:
+            txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504)
+
+        return img, txt
+
+
+class SingleStreamBlock(nn.Module):
+    """
+    A DiT block with parallel linear layers as described in
+    https://arxiv.org/abs/2302.05442 and adapted modulation interface.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qk_scale: float = None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.hidden_dim = hidden_size
+        self.num_heads = num_heads
+        head_dim = hidden_size // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        # qkv and mlp_in
+        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device)
+        # proj and mlp_out
+        self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device)
+
+        self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
+
+        self.hidden_size = hidden_size
+        self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+
+        self.mlp_act = nn.GELU(approximate="tanh")
+        self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims=None) -> Tensor:
+        mod, _ = self.modulation(vec)
+        qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+
+        q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k = self.norm(q, k, v)
+
+        # compute attention
+        attn = attention(q, k, v, pe=pe, mask=attn_mask)
+        # compute activation in mlp stream, cat again and run second linear layer
+        output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
+        x += apply_mod(output, mod.gate, None, modulation_dims)
+        if x.dtype == torch.float16:
+            x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
+        return x
+
+
+class LastLayer(nn.Module):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device))
+
+    def forward(self, x: Tensor, vec: Tensor, modulation_dims=None) -> Tensor:
+        if vec.ndim == 2:
+            vec = vec[:, None, :]
+
+        shift, scale = self.adaLN_modulation(vec).chunk(2, dim=-1)
+        x = apply_mod(self.norm_final(x), (1 + scale), shift, modulation_dims)
+        x = self.linear(x)
+        return x

comfy/ldm/flux/math.py

+import torch
+from einops import rearrange
+from torch import Tensor
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+
+
+def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None) -> Tensor:
+    q_shape = q.shape
+    k_shape = k.shape
+
+    if pe is not None:
+        q = q.to(dtype=pe.dtype).reshape(*q.shape[:-1], -1, 1, 2)
+        k = k.to(dtype=pe.dtype).reshape(*k.shape[:-1], -1, 1, 2)
+        q = (pe[..., 0] * q[..., 0] + pe[..., 1] * q[..., 1]).reshape(*q_shape).type_as(v)
+        k = (pe[..., 0] * k[..., 0] + pe[..., 1] * k[..., 1]).reshape(*k_shape).type_as(v)
+
+    heads = q.shape[1]
+    x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask)
+    return x
+
+
+def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
+    assert dim % 2 == 0
+    if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled():
+        device = torch.device("cpu")
+    else:
+        device = pos.device
+
+    scale = torch.linspace(0, (dim - 2) / dim, steps=dim//2, dtype=torch.float64, device=device)
+    omega = 1.0 / (theta**scale)
+    out = torch.einsum("...n,d->...nd", pos.to(dtype=torch.float32, device=device), omega)
+    out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1)
+    out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
+    return out.to(dtype=torch.float32, device=pos.device)
+
+
+def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
+    # import pdb
+    # pdb.set_trace()
+    # from lightop import op
+
+    # op.rotary_embedding_fuse()
+
+    xq_ = xq.to(dtype=freqs_cis.dtype).reshape(*xq.shape[:-1], -1, 1, 2)
+    xk_ = xk.to(dtype=freqs_cis.dtype).reshape(*xk.shape[:-1], -1, 1, 2)
+    xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
+    xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
+    return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)
+

comfy/ldm/flux/model.py

+#Original code can be found on: https://github.com/black-forest-labs/flux
+
+from dataclasses import dataclass
+
+import torch
+from torch import Tensor, nn
+from einops import rearrange, repeat
+import comfy.ldm.common_dit
+import comfy.patcher_extension
+
+from .layers import (
+    DoubleStreamBlock,
+    EmbedND,
+    LastLayer,
+    MLPEmbedder,
+    SingleStreamBlock,
+    timestep_embedding,
+)
+
+@dataclass
+class FluxParams:
+    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: int
+    qkv_bias: bool
+    guidance_embed: bool
+
+
+class Flux(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 = FluxParams(**kwargs)
+        self.params = params
+        self.patch_size = params.patch_size
+        self.in_channels = params.in_channels * params.patch_size * params.patch_size
+        self.out_channels = params.out_channels * params.patch_size * params.patch_size
+        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 = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
+        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 = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                    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, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        y: Tensor,
+        guidance: Tensor = None,
+        control = None,
+        transformer_options={},
+        attn_mask: Tensor = None,
+    ) -> Tensor:
+
+        if y is None:
+            y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype)
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        if img.ndim != 3 or txt.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        # running on sequences img
+        img = self.img_in(img)
+        vec = self.time_in(timestep_embedding(timesteps, 256).to(img.dtype))
+        if self.params.guidance_embed:
+            if guidance is not None:
+                vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
+
+        vec = vec + self.vector_in(y[:,:self.params.vec_in_dim])
+        txt = self.txt_in(txt)
+
+        if img_ids is not None:
+            ids = torch.cat((txt_ids, img_ids), dim=1)
+            pe = self.pe_embedder(ids)
+        else:
+            pe = 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.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)
+
+            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.shape[1]] += add
+
+        if img.dtype == torch.float16:
+            img = torch.nan_to_num(img, nan=0.0, posinf=65504, neginf=-65504)
+
+        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)
+
+            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[:, txt.shape[1] : txt.shape[1] + add.shape[1], ...] += add
+
+        img = img[:, txt.shape[1] :, ...]
+
+        img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
+        return img
+
+    def process_img(self, x, index=0, h_offset=0, w_offset=0):
+        bs, c, h, w = x.shape
+        patch_size = self.patch_size
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
+
+        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
+        h_len = ((h + (patch_size // 2)) // patch_size)
+        w_len = ((w + (patch_size // 2)) // patch_size)
+
+        h_offset = ((h_offset + (patch_size // 2)) // patch_size)
+        w_offset = ((w_offset + (patch_size // 2)) // patch_size)
+
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[:, :, 0] = img_ids[:, :, 1] + index
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        return img, repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+    def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=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, ref_latents, control, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs):
+        bs, c, h_orig, w_orig = x.shape
+        patch_size = self.patch_size
+
+        h_len = ((h_orig + (patch_size // 2)) // patch_size)
+        w_len = ((w_orig + (patch_size // 2)) // patch_size)
+        img, img_ids = self.process_img(x)
+        img_tokens = img.shape[1]
+        if ref_latents is not None:
+            h = 0
+            w = 0
+            index = 0
+            index_ref_method = kwargs.get("ref_latents_method", "offset") == "index"
+            for ref in ref_latents:
+                if index_ref_method:
+                    index += 1
+                    h_offset = 0
+                    w_offset = 0
+                else:
+                    index = 1
+                    h_offset = 0
+                    w_offset = 0
+                    if ref.shape[-2] + h > ref.shape[-1] + w:
+                        w_offset = w
+                    else:
+                        h_offset = h
+                    h = max(h, ref.shape[-2] + h_offset)
+                    w = max(w, ref.shape[-1] + w_offset)
+
+                kontext, kontext_ids = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
+                img = torch.cat([img, kontext], dim=1)
+                img_ids = torch.cat([img_ids, kontext_ids], dim=1)
+
+        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
+        out = out[:, :img_tokens]
+        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h_orig,:w_orig]

comfy/ldm/flux/redux.py

+import torch
+import comfy.ops
+
+ops = comfy.ops.manual_cast
+
+class ReduxImageEncoder(torch.nn.Module):
+    def __init__(
+        self,
+        redux_dim: int = 1152,
+        txt_in_features: int = 4096,
+        device=None,
+        dtype=None,
+    ) -> None:
+        super().__init__()
+
+        self.redux_dim = redux_dim
+        self.device = device
+        self.dtype = dtype
+
+        self.redux_up = ops.Linear(redux_dim, txt_in_features * 3, dtype=dtype)
+        self.redux_down = ops.Linear(txt_in_features * 3, txt_in_features, dtype=dtype)
+
+    def forward(self, sigclip_embeds) -> torch.Tensor:
+        projected_x = self.redux_down(torch.nn.functional.silu(self.redux_up(sigclip_embeds)))
+        return projected_x

comfy/ldm/genmo/joint_model/asymm_models_joint.py

+#original code from https://github.com/genmoai/models under apache 2.0 license
+#adapted to ComfyUI
+
+from typing import Dict, List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+# from flash_attn import flash_attn_varlen_qkvpacked_func
+from comfy.ldm.modules.attention import optimized_attention
+
+from .layers import (
+    FeedForward,
+    PatchEmbed,
+    TimestepEmbedder,
+)
+
+from .rope_mixed import (
+    compute_mixed_rotation,
+    create_position_matrix,
+)
+from .temporal_rope import apply_rotary_emb_qk_real
+from .utils import (
+    AttentionPool,
+    modulate,
+)
+
+import comfy.ldm.common_dit
+import comfy.ops
+
+
+def modulated_rmsnorm(x, scale, eps=1e-6):
+    # Normalize and modulate
+    x_normed = comfy.ldm.common_dit.rms_norm(x, eps=eps)
+    x_modulated = x_normed * (1 + scale.unsqueeze(1))
+
+    return x_modulated
+
+
+def residual_tanh_gated_rmsnorm(x, x_res, gate, eps=1e-6):
+    # Apply tanh to gate
+    tanh_gate = torch.tanh(gate).unsqueeze(1)
+
+    # Normalize and apply gated scaling
+    x_normed = comfy.ldm.common_dit.rms_norm(x_res, eps=eps) * tanh_gate
+
+    # Apply residual connection
+    output = x + x_normed
+
+    return output
+
+class AsymmetricAttention(nn.Module):
+    def __init__(
+        self,
+        dim_x: int,
+        dim_y: int,
+        num_heads: int = 8,
+        qkv_bias: bool = True,
+        qk_norm: bool = False,
+        attn_drop: float = 0.0,
+        update_y: bool = True,
+        out_bias: bool = True,
+        attend_to_padding: bool = False,
+        softmax_scale: Optional[float] = None,
+        device: Optional[torch.device] = None,
+        dtype=None,
+        operations=None,
+    ):
+        super().__init__()
+        self.dim_x = dim_x
+        self.dim_y = dim_y
+        self.num_heads = num_heads
+        self.head_dim = dim_x // num_heads
+        self.attn_drop = attn_drop
+        self.update_y = update_y
+        self.attend_to_padding = attend_to_padding
+        self.softmax_scale = softmax_scale
+        if dim_x % num_heads != 0:
+            raise ValueError(
+                f"dim_x={dim_x} should be divisible by num_heads={num_heads}"
+            )
+
+        # Input layers.
+        self.qkv_bias = qkv_bias
+        self.qkv_x = operations.Linear(dim_x, 3 * dim_x, bias=qkv_bias, device=device, dtype=dtype)
+        # Project text features to match visual features (dim_y -> dim_x)
+        self.qkv_y = operations.Linear(dim_y, 3 * dim_x, bias=qkv_bias, device=device, dtype=dtype)
+
+        # Query and key normalization for stability.
+        assert qk_norm
+        self.q_norm_x = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype)
+        self.k_norm_x = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype)
+        self.q_norm_y = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype)
+        self.k_norm_y = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype)
+
+        # Output layers. y features go back down from dim_x -> dim_y.
+        self.proj_x = operations.Linear(dim_x, dim_x, bias=out_bias, device=device, dtype=dtype)
+        self.proj_y = (
+            operations.Linear(dim_x, dim_y, bias=out_bias, device=device, dtype=dtype)
+            if update_y
+            else nn.Identity()
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,  # (B, N, dim_x)
+        y: torch.Tensor,  # (B, L, dim_y)
+        scale_x: torch.Tensor,  # (B, dim_x), modulation for pre-RMSNorm.
+        scale_y: torch.Tensor,  # (B, dim_y), modulation for pre-RMSNorm.
+        crop_y,
+        **rope_rotation,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        rope_cos = rope_rotation.get("rope_cos")
+        rope_sin = rope_rotation.get("rope_sin")
+        # Pre-norm for visual features
+        x = modulated_rmsnorm(x, scale_x)  # (B, M, dim_x) where M = N / cp_group_size
+
+        # Process visual features
+        # qkv_x = self.qkv_x(x)  # (B, M, 3 * dim_x)
+        # assert qkv_x.dtype == torch.bfloat16
+        # qkv_x = all_to_all_collect_tokens(
+        #     qkv_x, self.num_heads
+        # )  # (3, B, N, local_h, head_dim)
+
+        # Process text features
+        y = modulated_rmsnorm(y, scale_y)  # (B, L, dim_y)
+        q_y, k_y, v_y = self.qkv_y(y).view(y.shape[0], y.shape[1], 3, self.num_heads, -1).unbind(2)  # (B, N, local_h, head_dim)
+
+        q_y = self.q_norm_y(q_y)
+        k_y = self.k_norm_y(k_y)
+
+        # Split qkv_x into q, k, v
+        q_x, k_x, v_x = self.qkv_x(x).view(x.shape[0], x.shape[1], 3, self.num_heads, -1).unbind(2)  # (B, N, local_h, head_dim)
+        q_x = self.q_norm_x(q_x)
+        q_x = apply_rotary_emb_qk_real(q_x, rope_cos, rope_sin)
+        k_x = self.k_norm_x(k_x)
+        k_x = apply_rotary_emb_qk_real(k_x, rope_cos, rope_sin)
+
+        q = torch.cat([q_x, q_y[:, :crop_y]], dim=1).transpose(1, 2)
+        k = torch.cat([k_x, k_y[:, :crop_y]], dim=1).transpose(1, 2)
+        v = torch.cat([v_x, v_y[:, :crop_y]], dim=1).transpose(1, 2)
+
+        xy = optimized_attention(q,
+                                 k,
+                                 v, self.num_heads, skip_reshape=True)
+
+        x, y = torch.tensor_split(xy, (q_x.shape[1],), dim=1)
+        x = self.proj_x(x)
+        o = torch.zeros(y.shape[0], q_y.shape[1], y.shape[-1], device=y.device, dtype=y.dtype)
+        o[:, :y.shape[1]] = y
+
+        y = self.proj_y(o)
+        # print("ox", x)
+        # print("oy", y)
+        return x, y
+
+
+class AsymmetricJointBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size_x: int,
+        hidden_size_y: int,
+        num_heads: int,
+        *,
+        mlp_ratio_x: float = 8.0,  # Ratio of hidden size to d_model for MLP for visual tokens.
+        mlp_ratio_y: float = 4.0,  # Ratio of hidden size to d_model for MLP for text tokens.
+        update_y: bool = True,  # Whether to update text tokens in this block.
+        device: Optional[torch.device] = None,
+        dtype=None,
+        operations=None,
+        **block_kwargs,
+    ):
+        super().__init__()
+        self.update_y = update_y
+        self.hidden_size_x = hidden_size_x
+        self.hidden_size_y = hidden_size_y
+        self.mod_x = operations.Linear(hidden_size_x, 4 * hidden_size_x, device=device, dtype=dtype)
+        if self.update_y:
+            self.mod_y = operations.Linear(hidden_size_x, 4 * hidden_size_y, device=device, dtype=dtype)
+        else:
+            self.mod_y = operations.Linear(hidden_size_x, hidden_size_y, device=device, dtype=dtype)
+
+        # Self-attention:
+        self.attn = AsymmetricAttention(
+            hidden_size_x,
+            hidden_size_y,
+            num_heads=num_heads,
+            update_y=update_y,
+            device=device,
+            dtype=dtype,
+            operations=operations,
+            **block_kwargs,
+        )
+
+        # MLP.
+        mlp_hidden_dim_x = int(hidden_size_x * mlp_ratio_x)
+        assert mlp_hidden_dim_x == int(1536 * 8)
+        self.mlp_x = FeedForward(
+            in_features=hidden_size_x,
+            hidden_size=mlp_hidden_dim_x,
+            multiple_of=256,
+            ffn_dim_multiplier=None,
+            device=device,
+            dtype=dtype,
+            operations=operations,
+        )
+
+        # MLP for text not needed in last block.
+        if self.update_y:
+            mlp_hidden_dim_y = int(hidden_size_y * mlp_ratio_y)
+            self.mlp_y = FeedForward(
+                in_features=hidden_size_y,
+                hidden_size=mlp_hidden_dim_y,
+                multiple_of=256,
+                ffn_dim_multiplier=None,
+                device=device,
+                dtype=dtype,
+                operations=operations,
+            )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        c: torch.Tensor,
+        y: torch.Tensor,
+        **attn_kwargs,
+    ):
+        """Forward pass of a block.
+
+        Args:
+            x: (B, N, dim) tensor of visual tokens
+            c: (B, dim) tensor of conditioned features
+            y: (B, L, dim) tensor of text tokens
+            num_frames: Number of frames in the video. N = num_frames * num_spatial_tokens
+
+        Returns:
+            x: (B, N, dim) tensor of visual tokens after block
+            y: (B, L, dim) tensor of text tokens after block
+        """
+        N = x.size(1)
+
+        c = F.silu(c)
+        mod_x = self.mod_x(c)
+        scale_msa_x, gate_msa_x, scale_mlp_x, gate_mlp_x = mod_x.chunk(4, dim=1)
+
+        mod_y = self.mod_y(c)
+        if self.update_y:
+            scale_msa_y, gate_msa_y, scale_mlp_y, gate_mlp_y = mod_y.chunk(4, dim=1)
+        else:
+            scale_msa_y = mod_y
+
+        # Self-attention block.
+        x_attn, y_attn = self.attn(
+            x,
+            y,
+            scale_x=scale_msa_x,
+            scale_y=scale_msa_y,
+            **attn_kwargs,
+        )
+
+        assert x_attn.size(1) == N
+        x = residual_tanh_gated_rmsnorm(x, x_attn, gate_msa_x)
+        if self.update_y:
+            y = residual_tanh_gated_rmsnorm(y, y_attn, gate_msa_y)
+
+        # MLP block.
+        x = self.ff_block_x(x, scale_mlp_x, gate_mlp_x)
+        if self.update_y:
+            y = self.ff_block_y(y, scale_mlp_y, gate_mlp_y)
+
+        return x, y
+
+    def ff_block_x(self, x, scale_x, gate_x):
+        x_mod = modulated_rmsnorm(x, scale_x)
+        x_res = self.mlp_x(x_mod)
+        x = residual_tanh_gated_rmsnorm(x, x_res, gate_x)  # Sandwich norm
+        return x
+
+    def ff_block_y(self, y, scale_y, gate_y):
+        y_mod = modulated_rmsnorm(y, scale_y)
+        y_res = self.mlp_y(y_mod)
+        y = residual_tanh_gated_rmsnorm(y, y_res, gate_y)  # Sandwich norm
+        return y
+
+
+class FinalLayer(nn.Module):
+    """
+    The final layer of DiT.
+    """
+
+    def __init__(
+        self,
+        hidden_size,
+        patch_size,
+        out_channels,
+        device: Optional[torch.device] = None,
+        dtype=None,
+        operations=None,
+    ):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(
+            hidden_size, elementwise_affine=False, eps=1e-6, device=device, dtype=dtype
+        )
+        self.mod = operations.Linear(hidden_size, 2 * hidden_size, device=device, dtype=dtype)
+        self.linear = operations.Linear(
+            hidden_size, patch_size * patch_size * out_channels, device=device, dtype=dtype
+        )
+
+    def forward(self, x, c):
+        c = F.silu(c)
+        shift, scale = self.mod(c).chunk(2, dim=1)
+        x = modulate(self.norm_final(x), shift, scale)
+        x = self.linear(x)
+        return x
+
+
+class AsymmDiTJoint(nn.Module):
+    """
+    Diffusion model with a Transformer backbone.
+
+    Ingests text embeddings instead of a label.
+    """
+
+    def __init__(
+        self,
+        *,
+        patch_size=2,
+        in_channels=4,
+        hidden_size_x=1152,
+        hidden_size_y=1152,
+        depth=48,
+        num_heads=16,
+        mlp_ratio_x=8.0,
+        mlp_ratio_y=4.0,
+        use_t5: bool = False,
+        t5_feat_dim: int = 4096,
+        t5_token_length: int = 256,
+        learn_sigma=True,
+        patch_embed_bias: bool = True,
+        timestep_mlp_bias: bool = True,
+        attend_to_padding: bool = False,
+        timestep_scale: Optional[float] = None,
+        use_extended_posenc: bool = False,
+        posenc_preserve_area: bool = False,
+        rope_theta: float = 10000.0,
+        image_model=None,
+        device: Optional[torch.device] = None,
+        dtype=None,
+        operations=None,
+        **block_kwargs,
+    ):
+        super().__init__()
+
+        self.dtype = dtype
+        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_x = hidden_size_x
+        self.hidden_size_y = hidden_size_y
+        self.head_dim = (
+            hidden_size_x // num_heads
+        )  # Head dimension and count is determined by visual.
+        self.attend_to_padding = attend_to_padding
+        self.use_extended_posenc = use_extended_posenc
+        self.posenc_preserve_area = posenc_preserve_area
+        self.use_t5 = use_t5
+        self.t5_token_length = t5_token_length
+        self.t5_feat_dim = t5_feat_dim
+        self.rope_theta = (
+            rope_theta  # Scaling factor for frequency computation for temporal RoPE.
+        )
+
+        self.x_embedder = PatchEmbed(
+            patch_size=patch_size,
+            in_chans=in_channels,
+            embed_dim=hidden_size_x,
+            bias=patch_embed_bias,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+        # Conditionings
+        # Timestep
+        self.t_embedder = TimestepEmbedder(
+            hidden_size_x, bias=timestep_mlp_bias, timestep_scale=timestep_scale, dtype=dtype, device=device, operations=operations
+        )
+
+        if self.use_t5:
+            # Caption Pooling (T5)
+            self.t5_y_embedder = AttentionPool(
+                t5_feat_dim, num_heads=8, output_dim=hidden_size_x, dtype=dtype, device=device, operations=operations
+            )
+
+            # Dense Embedding Projection (T5)
+            self.t5_yproj = operations.Linear(
+                t5_feat_dim, hidden_size_y, bias=True, dtype=dtype, device=device
+            )
+
+        # Initialize pos_frequencies as an empty parameter.
+        self.pos_frequencies = nn.Parameter(
+            torch.empty(3, self.num_heads, self.head_dim // 2, dtype=dtype, device=device)
+        )
+
+        assert not self.attend_to_padding
+
+        # for depth 48:
+        #  b =  0: AsymmetricJointBlock, update_y=True
+        #  b =  1: AsymmetricJointBlock, update_y=True
+        #  ...
+        #  b = 46: AsymmetricJointBlock, update_y=True
+        #  b = 47: AsymmetricJointBlock, update_y=False. No need to update text features.
+        blocks = []
+        for b in range(depth):
+            # Joint multi-modal block
+            update_y = b < depth - 1
+            block = AsymmetricJointBlock(
+                hidden_size_x,
+                hidden_size_y,
+                num_heads,
+                mlp_ratio_x=mlp_ratio_x,
+                mlp_ratio_y=mlp_ratio_y,
+                update_y=update_y,
+                attend_to_padding=attend_to_padding,
+                device=device,
+                dtype=dtype,
+                operations=operations,
+                **block_kwargs,
+            )
+
+            blocks.append(block)
+        self.blocks = nn.ModuleList(blocks)
+
+        self.final_layer = FinalLayer(
+            hidden_size_x, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations
+        )
+
+    def embed_x(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            x: (B, C=12, T, H, W) tensor of visual tokens
+
+        Returns:
+            x: (B, C=3072, N) tensor of visual tokens with positional embedding.
+        """
+        return self.x_embedder(x)  # Convert BcTHW to BCN
+
+    def prepare(
+        self,
+        x: torch.Tensor,
+        sigma: torch.Tensor,
+        t5_feat: torch.Tensor,
+        t5_mask: torch.Tensor,
+    ):
+        """Prepare input and conditioning embeddings."""
+        # Visual patch embeddings with positional encoding.
+        T, H, W = x.shape[-3:]
+        pH, pW = H // self.patch_size, W // self.patch_size
+        x = self.embed_x(x)  # (B, N, D), where N = T * H * W / patch_size ** 2
+        assert x.ndim == 3
+
+        pH, pW = H // self.patch_size, W // self.patch_size
+        N = T * pH * pW
+        assert x.size(1) == N
+        pos = create_position_matrix(
+            T, pH=pH, pW=pW, device=x.device, dtype=torch.float32
+        )  # (N, 3)
+        rope_cos, rope_sin = compute_mixed_rotation(
+            freqs=comfy.ops.cast_to(self.pos_frequencies, dtype=x.dtype, device=x.device), pos=pos
+        )  # Each are (N, num_heads, dim // 2)
+
+        c_t = self.t_embedder(1 - sigma, out_dtype=x.dtype)  # (B, D)
+
+        t5_y_pool = self.t5_y_embedder(t5_feat, t5_mask)  # (B, D)
+
+        c = c_t + t5_y_pool
+
+        y_feat = self.t5_yproj(t5_feat)  # (B, L, t5_feat_dim) --> (B, L, D)
+
+        return x, c, y_feat, rope_cos, rope_sin
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        timestep: torch.Tensor,
+        context: List[torch.Tensor],
+        attention_mask: List[torch.Tensor],
+        num_tokens=256,
+        packed_indices: Dict[str, torch.Tensor] = None,
+        rope_cos: torch.Tensor = None,
+        rope_sin: torch.Tensor = None,
+        control=None, transformer_options={}, **kwargs
+    ):
+        patches_replace = transformer_options.get("patches_replace", {})
+        y_feat = context
+        y_mask = attention_mask
+        sigma = timestep
+        """Forward pass of DiT.
+
+        Args:
+            x: (B, C, T, H, W) tensor of spatial inputs (images or latent representations of images)
+            sigma: (B,) tensor of noise standard deviations
+            y_feat: List((B, L, y_feat_dim) tensor of caption token features. For SDXL text encoders: L=77, y_feat_dim=2048)
+            y_mask: List((B, L) boolean tensor indicating which tokens are not padding)
+            packed_indices: Dict with keys for Flash Attention. Result of compute_packed_indices.
+        """
+        B, _, T, H, W = x.shape
+
+        x, c, y_feat, rope_cos, rope_sin = self.prepare(
+            x, sigma, y_feat, y_mask
+        )
+        del y_mask
+
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"], out["txt"] = block(
+                                                    args["img"],
+                                                    args["vec"],
+                                                    args["txt"],
+                                                    rope_cos=args["rope_cos"],
+                                                    rope_sin=args["rope_sin"],
+                                                    crop_y=args["num_tokens"]
+                                                    )
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": y_feat, "vec": c, "rope_cos": rope_cos, "rope_sin": rope_sin, "num_tokens": num_tokens}, {"original_block": block_wrap})
+                y_feat = out["txt"]
+                x = out["img"]
+            else:
+                x, y_feat = block(
+                    x,
+                    c,
+                    y_feat,
+                    rope_cos=rope_cos,
+                    rope_sin=rope_sin,
+                    crop_y=num_tokens,
+                )  # (B, M, D), (B, L, D)
+        del y_feat  # Final layers don't use dense text features.
+
+        x = self.final_layer(x, c)  # (B, M, patch_size ** 2 * out_channels)
+        x = rearrange(
+            x,
+            "B (T hp wp) (p1 p2 c) -> B c T (hp p1) (wp p2)",
+            T=T,
+            hp=H // self.patch_size,
+            wp=W // self.patch_size,
+            p1=self.patch_size,
+            p2=self.patch_size,
+            c=self.out_channels,
+        )
+
+        return -x

comfy/ldm/genmo/joint_model/layers.py

+#original code from https://github.com/genmoai/models under apache 2.0 license
+#adapted to ComfyUI
+
+import collections.abc
+import math
+from itertools import repeat
+from typing import Callable, Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+import comfy.ldm.common_dit
+
+
+# From PyTorch internals
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
+            return tuple(x)
+        return tuple(repeat(x, n))
+
+    return parse
+
+
+to_2tuple = _ntuple(2)
+
+
+class TimestepEmbedder(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        frequency_embedding_size: int = 256,
+        *,
+        bias: bool = True,
+        timestep_scale: Optional[float] = None,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            operations.Linear(frequency_embedding_size, hidden_size, bias=bias, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(hidden_size, hidden_size, bias=bias, dtype=dtype, device=device),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+        self.timestep_scale = timestep_scale
+
+    @staticmethod
+    def timestep_embedding(t, dim, max_period=10000):
+        half = dim // 2
+        freqs = torch.arange(start=0, end=half, dtype=torch.float32, device=t.device)
+        freqs.mul_(-math.log(max_period) / half).exp_()
+        args = t[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat(
+                [embedding, torch.zeros_like(embedding[:, :1])], dim=-1
+            )
+        return embedding
+
+    def forward(self, t, out_dtype):
+        if self.timestep_scale is not None:
+            t = t * self.timestep_scale
+        t_freq = self.timestep_embedding(t, self.frequency_embedding_size).to(dtype=out_dtype)
+        t_emb = self.mlp(t_freq)
+        return t_emb
+
+
+class FeedForward(nn.Module):
+    def __init__(
+        self,
+        in_features: int,
+        hidden_size: int,
+        multiple_of: int,
+        ffn_dim_multiplier: Optional[float],
+        device: Optional[torch.device] = None,
+        dtype=None,
+        operations=None,
+    ):
+        super().__init__()
+        # keep parameter count and computation constant compared to standard FFN
+        hidden_size = int(2 * hidden_size / 3)
+        # custom dim factor multiplier
+        if ffn_dim_multiplier is not None:
+            hidden_size = int(ffn_dim_multiplier * hidden_size)
+        hidden_size = multiple_of * ((hidden_size + multiple_of - 1) // multiple_of)
+
+        self.hidden_dim = hidden_size
+        self.w1 = operations.Linear(in_features, 2 * hidden_size, bias=False, device=device, dtype=dtype)
+        self.w2 = operations.Linear(hidden_size, in_features, bias=False, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x, gate = self.w1(x).chunk(2, dim=-1)
+        x = self.w2(F.silu(x) * gate)
+        return x
+
+
+class PatchEmbed(nn.Module):
+    def __init__(
+        self,
+        patch_size: int = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        norm_layer: Optional[Callable] = None,
+        flatten: bool = True,
+        bias: bool = True,
+        dynamic_img_pad: bool = False,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        super().__init__()
+        self.patch_size = to_2tuple(patch_size)
+        self.flatten = flatten
+        self.dynamic_img_pad = dynamic_img_pad
+
+        self.proj = operations.Conv2d(
+            in_chans,
+            embed_dim,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=bias,
+            device=device,
+            dtype=dtype,
+        )
+        assert norm_layer is None
+        self.norm = (
+            norm_layer(embed_dim, device=device) if norm_layer else nn.Identity()
+        )
+
+    def forward(self, x):
+        B, _C, T, H, W = x.shape
+        if not self.dynamic_img_pad:
+            assert H % self.patch_size[0] == 0, f"Input height ({H}) should be divisible by patch size ({self.patch_size[0]})."
+            assert W % self.patch_size[1] == 0, f"Input width ({W}) should be divisible by patch size ({self.patch_size[1]})."
+        else:
+            pad_h = (self.patch_size[0] - H % self.patch_size[0]) % self.patch_size[0]
+            pad_w = (self.patch_size[1] - W % self.patch_size[1]) % self.patch_size[1]
+            x = F.pad(x, (0, pad_w, 0, pad_h))
+
+        x = rearrange(x, "B C T H W -> (B T) C H W", B=B, T=T)
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size, padding_mode='circular')
+        x = self.proj(x)
+
+        # Flatten temporal and spatial dimensions.
+        if not self.flatten:
+            raise NotImplementedError("Must flatten output.")
+        x = rearrange(x, "(B T) C H W -> B (T H W) C", B=B, T=T)
+
+        x = self.norm(x)
+        return x