Upload folder using huggingface_hub (part 2)

.gitattributes

@@ -6,3 +6,4 @@ unstableRevolution_V3Fp16.safetensors filter=lfs diff=lfs merge=lfs -text
 ae.safetensors filter=lfs diff=lfs merge=lfs -text
 qwen_3_4b.safetensors filter=lfs diff=lfs merge=lfs -text
 pornmaster_proSDXLV8.safetensors filter=lfs diff=lfs merge=lfs -text
+ComfyUI/comfy/text_encoders/t5_pile_tokenizer/tokenizer.model filter=lfs diff=lfs merge=lfs -text

ComfyUI/comfy/ldm/cascade/controlnet.py

@@ -0,0 +1,92 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Stability AI
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
+import torchvision
+from torch import nn
+from .common import LayerNorm2d_op
+
+
+class CNetResBlock(nn.Module):
+    def __init__(self, c, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.blocks = nn.Sequential(
+            LayerNorm2d_op(operations)(c, dtype=dtype, device=device),
+            nn.GELU(),
+            operations.Conv2d(c, c, kernel_size=3, padding=1),
+            LayerNorm2d_op(operations)(c, dtype=dtype, device=device),
+            nn.GELU(),
+            operations.Conv2d(c, c, kernel_size=3, padding=1),
+        )
+
+    def forward(self, x):
+        return x + self.blocks(x)
+
+
+class ControlNet(nn.Module):
+    def __init__(self, c_in=3, c_proj=2048, proj_blocks=None, bottleneck_mode=None, dtype=None, device=None, operations=nn):
+        super().__init__()
+        if bottleneck_mode is None:
+            bottleneck_mode = 'effnet'
+        self.proj_blocks = proj_blocks
+        if bottleneck_mode == 'effnet':
+            embd_channels = 1280
+            self.backbone = torchvision.models.efficientnet_v2_s().features.eval()
+            if c_in != 3:
+                in_weights = self.backbone[0][0].weight.data
+                self.backbone[0][0] = operations.Conv2d(c_in, 24, kernel_size=3, stride=2, bias=False, dtype=dtype, device=device)
+                if c_in > 3:
+                    # nn.init.constant_(self.backbone[0][0].weight, 0)
+                    self.backbone[0][0].weight.data[:, :3] = in_weights[:, :3].clone()
+                else:
+                    self.backbone[0][0].weight.data = in_weights[:, :c_in].clone()
+        elif bottleneck_mode == 'simple':
+            embd_channels = c_in
+            self.backbone = nn.Sequential(
+                operations.Conv2d(embd_channels, embd_channels * 4, kernel_size=3, padding=1, dtype=dtype, device=device),
+                nn.LeakyReLU(0.2, inplace=True),
+                operations.Conv2d(embd_channels * 4, embd_channels, kernel_size=3, padding=1, dtype=dtype, device=device),
+            )
+        elif bottleneck_mode == 'large':
+            self.backbone = nn.Sequential(
+                operations.Conv2d(c_in, 4096 * 4, kernel_size=1, dtype=dtype, device=device),
+                nn.LeakyReLU(0.2, inplace=True),
+                operations.Conv2d(4096 * 4, 1024, kernel_size=1, dtype=dtype, device=device),
+                *[CNetResBlock(1024, dtype=dtype, device=device, operations=operations) for _ in range(8)],
+                operations.Conv2d(1024, 1280, kernel_size=1, dtype=dtype, device=device),
+            )
+            embd_channels = 1280
+        else:
+            raise ValueError(f'Unknown bottleneck mode: {bottleneck_mode}')
+        self.projections = nn.ModuleList()
+        for _ in range(len(proj_blocks)):
+            self.projections.append(nn.Sequential(
+                operations.Conv2d(embd_channels, embd_channels, kernel_size=1, bias=False, dtype=dtype, device=device),
+                nn.LeakyReLU(0.2, inplace=True),
+                operations.Conv2d(embd_channels, c_proj, kernel_size=1, bias=False, dtype=dtype, device=device),
+            ))
+            # nn.init.constant_(self.projections[-1][-1].weight, 0)  # zero output projection
+        self.xl = False
+        self.input_channels = c_in
+        self.unshuffle_amount = 8
+
+    def forward(self, x):
+        x = self.backbone(x)
+        proj_outputs = [None for _ in range(max(self.proj_blocks) + 1)]
+        for i, idx in enumerate(self.proj_blocks):
+            proj_outputs[idx] = self.projections[i](x)
+        return {"input": proj_outputs[::-1]}

ComfyUI/comfy/ldm/cascade/stage_a.py

@@ -0,0 +1,250 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Stability AI
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
+import torch
+from torch import nn
+from torch.autograd import Function
+import comfy.ops
+
+ops = comfy.ops.disable_weight_init
+
+
+class vector_quantize(Function):
+    @staticmethod
+    def forward(ctx, x, codebook):
+        with torch.no_grad():
+            codebook_sqr = torch.sum(codebook ** 2, dim=1)
+            x_sqr = torch.sum(x ** 2, dim=1, keepdim=True)
+
+            dist = torch.addmm(codebook_sqr + x_sqr, x, codebook.t(), alpha=-2.0, beta=1.0)
+            _, indices = dist.min(dim=1)
+
+            ctx.save_for_backward(indices, codebook)
+            ctx.mark_non_differentiable(indices)
+
+            nn = torch.index_select(codebook, 0, indices)
+            return nn, indices
+
+    @staticmethod
+    def backward(ctx, grad_output, grad_indices):
+        grad_inputs, grad_codebook = None, None
+
+        if ctx.needs_input_grad[0]:
+            grad_inputs = grad_output.clone()
+        if ctx.needs_input_grad[1]:
+            # Gradient wrt. the codebook
+            indices, codebook = ctx.saved_tensors
+
+            grad_codebook = torch.zeros_like(codebook)
+            grad_codebook.index_add_(0, indices, grad_output)
+
+        return (grad_inputs, grad_codebook)
+
+
+class VectorQuantize(nn.Module):
+    def __init__(self, embedding_size, k, ema_decay=0.99, ema_loss=False):
+        """
+        Takes an input of variable size (as long as the last dimension matches the embedding size).
+        Returns one tensor containing the nearest neigbour embeddings to each of the inputs,
+        with the same size as the input, vq and commitment components for the loss as a touple
+        in the second output and the indices of the quantized vectors in the third:
+        quantized, (vq_loss, commit_loss), indices
+        """
+        super(VectorQuantize, self).__init__()
+
+        self.codebook = nn.Embedding(k, embedding_size)
+        self.codebook.weight.data.uniform_(-1./k, 1./k)
+        self.vq = vector_quantize.apply
+
+        self.ema_decay = ema_decay
+        self.ema_loss = ema_loss
+        if ema_loss:
+            self.register_buffer('ema_element_count', torch.ones(k))
+            self.register_buffer('ema_weight_sum', torch.zeros_like(self.codebook.weight))
+
+    def _laplace_smoothing(self, x, epsilon):
+        n = torch.sum(x)
+        return ((x + epsilon) / (n + x.size(0) * epsilon) * n)
+
+    def _updateEMA(self, z_e_x, indices):
+        mask = nn.functional.one_hot(indices, self.ema_element_count.size(0)).float()
+        elem_count = mask.sum(dim=0)
+        weight_sum = torch.mm(mask.t(), z_e_x)
+
+        self.ema_element_count = (self.ema_decay * self.ema_element_count) + ((1-self.ema_decay) * elem_count)
+        self.ema_element_count = self._laplace_smoothing(self.ema_element_count, 1e-5)
+        self.ema_weight_sum = (self.ema_decay * self.ema_weight_sum) + ((1-self.ema_decay) * weight_sum)
+
+        self.codebook.weight.data = self.ema_weight_sum / self.ema_element_count.unsqueeze(-1)
+
+    def idx2vq(self, idx, dim=-1):
+        q_idx = self.codebook(idx)
+        if dim != -1:
+            q_idx = q_idx.movedim(-1, dim)
+        return q_idx
+
+    def forward(self, x, get_losses=True, dim=-1):
+        if dim != -1:
+            x = x.movedim(dim, -1)
+        z_e_x = x.contiguous().view(-1, x.size(-1)) if len(x.shape) > 2 else x
+        z_q_x, indices = self.vq(z_e_x, self.codebook.weight.detach())
+        vq_loss, commit_loss = None, None
+        if self.ema_loss and self.training:
+            self._updateEMA(z_e_x.detach(), indices.detach())
+        # pick the graded embeddings after updating the codebook in order to have a more accurate commitment loss
+        z_q_x_grd = torch.index_select(self.codebook.weight, dim=0, index=indices)
+        if get_losses:
+            vq_loss = (z_q_x_grd - z_e_x.detach()).pow(2).mean()
+            commit_loss = (z_e_x - z_q_x_grd.detach()).pow(2).mean()
+
+        z_q_x = z_q_x.view(x.shape)
+        if dim != -1:
+            z_q_x = z_q_x.movedim(-1, dim)
+        return z_q_x, (vq_loss, commit_loss), indices.view(x.shape[:-1])
+
+
+class ResBlock(nn.Module):
+    def __init__(self, c, c_hidden):
+        super().__init__()
+        # depthwise/attention
+        self.norm1 = nn.LayerNorm(c, elementwise_affine=False, eps=1e-6)
+        self.depthwise = nn.Sequential(
+            nn.ReplicationPad2d(1),
+            ops.Conv2d(c, c, kernel_size=3, groups=c)
+        )
+
+        # channelwise
+        self.norm2 = nn.LayerNorm(c, elementwise_affine=False, eps=1e-6)
+        self.channelwise = nn.Sequential(
+            ops.Linear(c, c_hidden),
+            nn.GELU(),
+            ops.Linear(c_hidden, c),
+        )
+
+        self.gammas = nn.Parameter(torch.zeros(6), requires_grad=False)
+
+    def _norm(self, x, norm):
+        return norm(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
+
+    def forward(self, x):
+        mods = self.gammas
+
+        x_temp = self._norm(x, self.norm1) * (1 + mods[0]) + mods[1]
+        try:
+            x = x + self.depthwise(x_temp) * mods[2]
+        except: #operation not implemented for bf16
+            x_temp = self.depthwise[0](x_temp.float()).to(x.dtype)
+            x = x + self.depthwise[1](x_temp) * mods[2]
+
+        x_temp = self._norm(x, self.norm2) * (1 + mods[3]) + mods[4]
+        x = x + self.channelwise(x_temp.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) * mods[5]
+
+        return x
+
+
+class StageA(nn.Module):
+    def __init__(self, levels=2, bottleneck_blocks=12, c_hidden=384, c_latent=4, codebook_size=8192):
+        super().__init__()
+        self.c_latent = c_latent
+        c_levels = [c_hidden // (2 ** i) for i in reversed(range(levels))]
+
+        # Encoder blocks
+        self.in_block = nn.Sequential(
+            nn.PixelUnshuffle(2),
+            ops.Conv2d(3 * 4, c_levels[0], kernel_size=1)
+        )
+        down_blocks = []
+        for i in range(levels):
+            if i > 0:
+                down_blocks.append(ops.Conv2d(c_levels[i - 1], c_levels[i], kernel_size=4, stride=2, padding=1))
+            block = ResBlock(c_levels[i], c_levels[i] * 4)
+            down_blocks.append(block)
+        down_blocks.append(nn.Sequential(
+            ops.Conv2d(c_levels[-1], c_latent, kernel_size=1, bias=False),
+            nn.BatchNorm2d(c_latent),  # then normalize them to have mean 0 and std 1
+        ))
+        self.down_blocks = nn.Sequential(*down_blocks)
+        self.down_blocks[0]
+
+        self.codebook_size = codebook_size
+        self.vquantizer = VectorQuantize(c_latent, k=codebook_size)
+
+        # Decoder blocks
+        up_blocks = [nn.Sequential(
+            ops.Conv2d(c_latent, c_levels[-1], kernel_size=1)
+        )]
+        for i in range(levels):
+            for j in range(bottleneck_blocks if i == 0 else 1):
+                block = ResBlock(c_levels[levels - 1 - i], c_levels[levels - 1 - i] * 4)
+                up_blocks.append(block)
+            if i < levels - 1:
+                up_blocks.append(
+                    ops.ConvTranspose2d(c_levels[levels - 1 - i], c_levels[levels - 2 - i], kernel_size=4, stride=2,
+                                       padding=1))
+        self.up_blocks = nn.Sequential(*up_blocks)
+        self.out_block = nn.Sequential(
+            ops.Conv2d(c_levels[0], 3 * 4, kernel_size=1),
+            nn.PixelShuffle(2),
+        )
+
+    def encode(self, x, quantize=False):
+        x = self.in_block(x)
+        x = self.down_blocks(x)
+        if quantize:
+            qe, (vq_loss, commit_loss), indices = self.vquantizer.forward(x, dim=1)
+            return qe, x, indices, vq_loss + commit_loss * 0.25
+        else:
+            return x
+
+    def decode(self, x):
+        x = self.up_blocks(x)
+        x = self.out_block(x)
+        return x
+
+    def forward(self, x, quantize=False):
+        qe, x, _, vq_loss = self.encode(x, quantize)
+        x = self.decode(qe)
+        return x, vq_loss
+
+
+class Discriminator(nn.Module):
+    def __init__(self, c_in=3, c_cond=0, c_hidden=512, depth=6):
+        super().__init__()
+        d = max(depth - 3, 3)
+        layers = [
+            nn.utils.spectral_norm(ops.Conv2d(c_in, c_hidden // (2 ** d), kernel_size=3, stride=2, padding=1)),
+            nn.LeakyReLU(0.2),
+        ]
+        for i in range(depth - 1):
+            c_in = c_hidden // (2 ** max((d - i), 0))
+            c_out = c_hidden // (2 ** max((d - 1 - i), 0))
+            layers.append(nn.utils.spectral_norm(ops.Conv2d(c_in, c_out, kernel_size=3, stride=2, padding=1)))
+            layers.append(nn.InstanceNorm2d(c_out))
+            layers.append(nn.LeakyReLU(0.2))
+        self.encoder = nn.Sequential(*layers)
+        self.shuffle = ops.Conv2d((c_hidden + c_cond) if c_cond > 0 else c_hidden, 1, kernel_size=1)
+        self.logits = nn.Sigmoid()
+
+    def forward(self, x, cond=None):
+        x = self.encoder(x)
+        if cond is not None:
+            cond = cond.view(cond.size(0), cond.size(1), 1, 1, ).expand(-1, -1, x.size(-2), x.size(-1))
+            x = torch.cat([x, cond], dim=1)
+        x = self.shuffle(x)
+        x = self.logits(x)
+        return x

ComfyUI/comfy/ldm/cascade/stage_b.py

@@ -0,0 +1,256 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Stability AI
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
+import math
+import torch
+from torch import nn
+from .common import AttnBlock, LayerNorm2d_op, ResBlock, FeedForwardBlock, TimestepBlock
+
+class StageB(nn.Module):
+    def __init__(self, c_in=4, c_out=4, c_r=64, patch_size=2, c_cond=1280, c_hidden=[320, 640, 1280, 1280],
+                 nhead=[-1, -1, 20, 20], blocks=[[2, 6, 28, 6], [6, 28, 6, 2]],
+                 block_repeat=[[1, 1, 1, 1], [3, 3, 2, 2]], level_config=['CT', 'CT', 'CTA', 'CTA'], c_clip=1280,
+                 c_clip_seq=4, c_effnet=16, c_pixels=3, kernel_size=3, dropout=[0, 0, 0.0, 0.0], self_attn=True,
+                 t_conds=['sca'], stable_cascade_stage=None, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.dtype = dtype
+        self.c_r = c_r
+        self.t_conds = t_conds
+        self.c_clip_seq = c_clip_seq
+        if not isinstance(dropout, list):
+            dropout = [dropout] * len(c_hidden)
+        if not isinstance(self_attn, list):
+            self_attn = [self_attn] * len(c_hidden)
+
+        # CONDITIONING
+        self.effnet_mapper = nn.Sequential(
+            operations.Conv2d(c_effnet, c_hidden[0] * 4, kernel_size=1, dtype=dtype, device=device),
+            nn.GELU(),
+            operations.Conv2d(c_hidden[0] * 4, c_hidden[0], kernel_size=1, dtype=dtype, device=device),
+            LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        )
+        self.pixels_mapper = nn.Sequential(
+            operations.Conv2d(c_pixels, c_hidden[0] * 4, kernel_size=1, dtype=dtype, device=device),
+            nn.GELU(),
+            operations.Conv2d(c_hidden[0] * 4, c_hidden[0], kernel_size=1, dtype=dtype, device=device),
+            LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        )
+        self.clip_mapper = operations.Linear(c_clip, c_cond * c_clip_seq, dtype=dtype, device=device)
+        self.clip_norm = operations.LayerNorm(c_cond, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+
+        self.embedding = nn.Sequential(
+            nn.PixelUnshuffle(patch_size),
+            operations.Conv2d(c_in * (patch_size ** 2), c_hidden[0], kernel_size=1, dtype=dtype, device=device),
+            LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        )
+
+        def get_block(block_type, c_hidden, nhead, c_skip=0, dropout=0, self_attn=True):
+            if block_type == 'C':
+                return ResBlock(c_hidden, c_skip, kernel_size=kernel_size, dropout=dropout, dtype=dtype, device=device, operations=operations)
+            elif block_type == 'A':
+                return AttnBlock(c_hidden, c_cond, nhead, self_attn=self_attn, dropout=dropout, dtype=dtype, device=device, operations=operations)
+            elif block_type == 'F':
+                return FeedForwardBlock(c_hidden, dropout=dropout, dtype=dtype, device=device, operations=operations)
+            elif block_type == 'T':
+                return TimestepBlock(c_hidden, c_r, conds=t_conds, dtype=dtype, device=device, operations=operations)
+            else:
+                raise Exception(f'Block type {block_type} not supported')
+
+        # BLOCKS
+        # -- down blocks
+        self.down_blocks = nn.ModuleList()
+        self.down_downscalers = nn.ModuleList()
+        self.down_repeat_mappers = nn.ModuleList()
+        for i in range(len(c_hidden)):
+            if i > 0:
+                self.down_downscalers.append(nn.Sequential(
+                    LayerNorm2d_op(operations)(c_hidden[i - 1], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device),
+                    operations.Conv2d(c_hidden[i - 1], c_hidden[i], kernel_size=2, stride=2, dtype=dtype, device=device),
+                ))
+            else:
+                self.down_downscalers.append(nn.Identity())
+            down_block = nn.ModuleList()
+            for _ in range(blocks[0][i]):
+                for block_type in level_config[i]:
+                    block = get_block(block_type, c_hidden[i], nhead[i], dropout=dropout[i], self_attn=self_attn[i])
+                    down_block.append(block)
+            self.down_blocks.append(down_block)
+            if block_repeat is not None:
+                block_repeat_mappers = nn.ModuleList()
+                for _ in range(block_repeat[0][i] - 1):
+                    block_repeat_mappers.append(operations.Conv2d(c_hidden[i], c_hidden[i], kernel_size=1, dtype=dtype, device=device))
+                self.down_repeat_mappers.append(block_repeat_mappers)
+
+        # -- up blocks
+        self.up_blocks = nn.ModuleList()
+        self.up_upscalers = nn.ModuleList()
+        self.up_repeat_mappers = nn.ModuleList()
+        for i in reversed(range(len(c_hidden))):
+            if i > 0:
+                self.up_upscalers.append(nn.Sequential(
+                    LayerNorm2d_op(operations)(c_hidden[i], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device),
+                    operations.ConvTranspose2d(c_hidden[i], c_hidden[i - 1], kernel_size=2, stride=2, dtype=dtype, device=device),
+                ))
+            else:
+                self.up_upscalers.append(nn.Identity())
+            up_block = nn.ModuleList()
+            for j in range(blocks[1][::-1][i]):
+                for k, block_type in enumerate(level_config[i]):
+                    c_skip = c_hidden[i] if i < len(c_hidden) - 1 and j == k == 0 else 0
+                    block = get_block(block_type, c_hidden[i], nhead[i], c_skip=c_skip, dropout=dropout[i],
+                                      self_attn=self_attn[i])
+                    up_block.append(block)
+            self.up_blocks.append(up_block)
+            if block_repeat is not None:
+                block_repeat_mappers = nn.ModuleList()
+                for _ in range(block_repeat[1][::-1][i] - 1):
+                    block_repeat_mappers.append(operations.Conv2d(c_hidden[i], c_hidden[i], kernel_size=1, dtype=dtype, device=device))
+                self.up_repeat_mappers.append(block_repeat_mappers)
+
+        # OUTPUT
+        self.clf = nn.Sequential(
+            LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device),
+            operations.Conv2d(c_hidden[0], c_out * (patch_size ** 2), kernel_size=1, dtype=dtype, device=device),
+            nn.PixelShuffle(patch_size),
+        )
+
+        # --- WEIGHT INIT ---
+    #     self.apply(self._init_weights)  # General init
+    #     nn.init.normal_(self.clip_mapper.weight, std=0.02)  # conditionings
+    #     nn.init.normal_(self.effnet_mapper[0].weight, std=0.02)  # conditionings
+    #     nn.init.normal_(self.effnet_mapper[2].weight, std=0.02)  # conditionings
+    #     nn.init.normal_(self.pixels_mapper[0].weight, std=0.02)  # conditionings
+    #     nn.init.normal_(self.pixels_mapper[2].weight, std=0.02)  # conditionings
+    #     torch.nn.init.xavier_uniform_(self.embedding[1].weight, 0.02)  # inputs
+    #     nn.init.constant_(self.clf[1].weight, 0)  # outputs
+    #
+    #     # blocks
+    #     for level_block in self.down_blocks + self.up_blocks:
+    #         for block in level_block:
+    #             if isinstance(block, ResBlock) or isinstance(block, FeedForwardBlock):
+    #                 block.channelwise[-1].weight.data *= np.sqrt(1 / sum(blocks[0]))
+    #             elif isinstance(block, TimestepBlock):
+    #                 for layer in block.modules():
+    #                     if isinstance(layer, nn.Linear):
+    #                         nn.init.constant_(layer.weight, 0)
+    #
+    # def _init_weights(self, m):
+    #     if isinstance(m, (nn.Conv2d, nn.Linear)):
+    #         torch.nn.init.xavier_uniform_(m.weight)
+    #         if m.bias is not None:
+    #             nn.init.constant_(m.bias, 0)
+
+    def gen_r_embedding(self, r, max_positions=10000):
+        r = r * max_positions
+        half_dim = self.c_r // 2
+        emb = math.log(max_positions) / (half_dim - 1)
+        emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp()
+        emb = r[:, None] * emb[None, :]
+        emb = torch.cat([emb.sin(), emb.cos()], dim=1)
+        if self.c_r % 2 == 1:  # zero pad
+            emb = nn.functional.pad(emb, (0, 1), mode='constant')
+        return emb
+
+    def gen_c_embeddings(self, clip):
+        if len(clip.shape) == 2:
+            clip = clip.unsqueeze(1)
+        clip = self.clip_mapper(clip).view(clip.size(0), clip.size(1) * self.c_clip_seq, -1)
+        clip = self.clip_norm(clip)
+        return clip
+
+    def _down_encode(self, x, r_embed, clip, transformer_options={}):
+        level_outputs = []
+        block_group = zip(self.down_blocks, self.down_downscalers, self.down_repeat_mappers)
+        for down_block, downscaler, repmap in block_group:
+            x = downscaler(x)
+            for i in range(len(repmap) + 1):
+                for block in down_block:
+                    if isinstance(block, ResBlock) or (
+                            hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
+                                                                                  ResBlock)):
+                        x = block(x)
+                    elif isinstance(block, AttnBlock) or (
+                            hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
+                                                                                  AttnBlock)):
+                        x = block(x, clip, transformer_options=transformer_options)
+                    elif isinstance(block, TimestepBlock) or (
+                            hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
+                                                                                  TimestepBlock)):
+                        x = block(x, r_embed)
+                    else:
+                        x = block(x)
+                if i < len(repmap):
+                    x = repmap[i](x)
+            level_outputs.insert(0, x)
+        return level_outputs
+
+    def _up_decode(self, level_outputs, r_embed, clip, transformer_options={}):
+        x = level_outputs[0]
+        block_group = zip(self.up_blocks, self.up_upscalers, self.up_repeat_mappers)
+        for i, (up_block, upscaler, repmap) in enumerate(block_group):
+            for j in range(len(repmap) + 1):
+                for k, block in enumerate(up_block):
+                    if isinstance(block, ResBlock) or (
+                            hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
+                                                                                  ResBlock)):
+                        skip = level_outputs[i] if k == 0 and i > 0 else None
+                        if skip is not None and (x.size(-1) != skip.size(-1) or x.size(-2) != skip.size(-2)):
+                            x = torch.nn.functional.interpolate(x, skip.shape[-2:], mode='bilinear',
+                                                                align_corners=True)
+                        x = block(x, skip)
+                    elif isinstance(block, AttnBlock) or (
+                            hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
+                                                                                  AttnBlock)):
+                        x = block(x, clip, transformer_options=transformer_options)
+                    elif isinstance(block, TimestepBlock) or (
+                            hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
+                                                                                  TimestepBlock)):
+                        x = block(x, r_embed)
+                    else:
+                        x = block(x)
+                if j < len(repmap):
+                    x = repmap[j](x)
+            x = upscaler(x)
+        return x
+
+    def forward(self, x, r, effnet, clip, pixels=None, transformer_options={}, **kwargs):
+        if pixels is None:
+            pixels = x.new_zeros(x.size(0), 3, 8, 8)
+
+        # Process the conditioning embeddings
+        r_embed = self.gen_r_embedding(r).to(dtype=x.dtype)
+        for c in self.t_conds:
+            t_cond = kwargs.get(c, torch.zeros_like(r))
+            r_embed = torch.cat([r_embed, self.gen_r_embedding(t_cond).to(dtype=x.dtype)], dim=1)
+        clip = self.gen_c_embeddings(clip)
+
+        # Model Blocks
+        x = self.embedding(x)
+        x = x + self.effnet_mapper(
+            nn.functional.interpolate(effnet, size=x.shape[-2:], mode='bilinear', align_corners=True))
+        x = x + nn.functional.interpolate(self.pixels_mapper(pixels), size=x.shape[-2:], mode='bilinear',
+                                          align_corners=True)
+        level_outputs = self._down_encode(x, r_embed, clip, transformer_options=transformer_options)
+        x = self._up_decode(level_outputs, r_embed, clip, transformer_options=transformer_options)
+        return self.clf(x)
+
+    def update_weights_ema(self, src_model, beta=0.999):
+        for self_params, src_params in zip(self.parameters(), src_model.parameters()):
+            self_params.data = self_params.data * beta + src_params.data.clone().to(self_params.device) * (1 - beta)
+        for self_buffers, src_buffers in zip(self.buffers(), src_model.buffers()):
+            self_buffers.data = self_buffers.data * beta + src_buffers.data.clone().to(self_buffers.device) * (1 - beta)

ComfyUI/comfy/ldm/cascade/stage_c.py

@@ -0,0 +1,273 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Stability AI
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
+import torch
+from torch import nn
+import math
+from .common import AttnBlock, LayerNorm2d_op, ResBlock, FeedForwardBlock, TimestepBlock
+# from .controlnet import ControlNetDeliverer
+
+class UpDownBlock2d(nn.Module):
+    def __init__(self, c_in, c_out, mode, enabled=True, dtype=None, device=None, operations=None):
+        super().__init__()
+        assert mode in ['up', 'down']
+        interpolation = nn.Upsample(scale_factor=2 if mode == 'up' else 0.5, mode='bilinear',
+                                    align_corners=True) if enabled else nn.Identity()
+        mapping = operations.Conv2d(c_in, c_out, kernel_size=1, dtype=dtype, device=device)
+        self.blocks = nn.ModuleList([interpolation, mapping] if mode == 'up' else [mapping, interpolation])
+
+    def forward(self, x):
+        for block in self.blocks:
+            x = block(x)
+        return x
+
+
+class StageC(nn.Module):
+    def __init__(self, c_in=16, c_out=16, c_r=64, patch_size=1, c_cond=2048, c_hidden=[2048, 2048], nhead=[32, 32],
+                 blocks=[[8, 24], [24, 8]], block_repeat=[[1, 1], [1, 1]], level_config=['CTA', 'CTA'],
+                 c_clip_text=1280, c_clip_text_pooled=1280, c_clip_img=768, c_clip_seq=4, kernel_size=3,
+                 dropout=[0.0, 0.0], self_attn=True, t_conds=['sca', 'crp'], switch_level=[False], stable_cascade_stage=None,
+                 dtype=None, device=None, operations=None):
+        super().__init__()
+        self.dtype = dtype
+        self.c_r = c_r
+        self.t_conds = t_conds
+        self.c_clip_seq = c_clip_seq
+        if not isinstance(dropout, list):
+            dropout = [dropout] * len(c_hidden)
+        if not isinstance(self_attn, list):
+            self_attn = [self_attn] * len(c_hidden)
+
+        # CONDITIONING
+        self.clip_txt_mapper = operations.Linear(c_clip_text, c_cond, dtype=dtype, device=device)
+        self.clip_txt_pooled_mapper = operations.Linear(c_clip_text_pooled, c_cond * c_clip_seq, dtype=dtype, device=device)
+        self.clip_img_mapper = operations.Linear(c_clip_img, c_cond * c_clip_seq, dtype=dtype, device=device)
+        self.clip_norm = operations.LayerNorm(c_cond, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+
+        self.embedding = nn.Sequential(
+            nn.PixelUnshuffle(patch_size),
+            operations.Conv2d(c_in * (patch_size ** 2), c_hidden[0], kernel_size=1, dtype=dtype, device=device),
+            LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6)
+        )
+
+        def get_block(block_type, c_hidden, nhead, c_skip=0, dropout=0, self_attn=True):
+            if block_type == 'C':
+                return ResBlock(c_hidden, c_skip, kernel_size=kernel_size, dropout=dropout, dtype=dtype, device=device, operations=operations)
+            elif block_type == 'A':
+                return AttnBlock(c_hidden, c_cond, nhead, self_attn=self_attn, dropout=dropout, dtype=dtype, device=device, operations=operations)
+            elif block_type == 'F':
+                return FeedForwardBlock(c_hidden, dropout=dropout, dtype=dtype, device=device, operations=operations)
+            elif block_type == 'T':
+                return TimestepBlock(c_hidden, c_r, conds=t_conds, dtype=dtype, device=device, operations=operations)
+            else:
+                raise Exception(f'Block type {block_type} not supported')
+
+        # BLOCKS
+        # -- down blocks
+        self.down_blocks = nn.ModuleList()
+        self.down_downscalers = nn.ModuleList()
+        self.down_repeat_mappers = nn.ModuleList()
+        for i in range(len(c_hidden)):
+            if i > 0:
+                self.down_downscalers.append(nn.Sequential(
+                    LayerNorm2d_op(operations)(c_hidden[i - 1], elementwise_affine=False, eps=1e-6),
+                    UpDownBlock2d(c_hidden[i - 1], c_hidden[i], mode='down', enabled=switch_level[i - 1], dtype=dtype, device=device, operations=operations)
+                ))
+            else:
+                self.down_downscalers.append(nn.Identity())
+            down_block = nn.ModuleList()
+            for _ in range(blocks[0][i]):
+                for block_type in level_config[i]:
+                    block = get_block(block_type, c_hidden[i], nhead[i], dropout=dropout[i], self_attn=self_attn[i])
+                    down_block.append(block)
+            self.down_blocks.append(down_block)
+            if block_repeat is not None:
+                block_repeat_mappers = nn.ModuleList()
+                for _ in range(block_repeat[0][i] - 1):
+                    block_repeat_mappers.append(operations.Conv2d(c_hidden[i], c_hidden[i], kernel_size=1, dtype=dtype, device=device))
+                self.down_repeat_mappers.append(block_repeat_mappers)
+
+        # -- up blocks
+        self.up_blocks = nn.ModuleList()
+        self.up_upscalers = nn.ModuleList()
+        self.up_repeat_mappers = nn.ModuleList()
+        for i in reversed(range(len(c_hidden))):
+            if i > 0:
+                self.up_upscalers.append(nn.Sequential(
+                    LayerNorm2d_op(operations)(c_hidden[i], elementwise_affine=False, eps=1e-6),
+                    UpDownBlock2d(c_hidden[i], c_hidden[i - 1], mode='up', enabled=switch_level[i - 1], dtype=dtype, device=device, operations=operations)
+                ))
+            else:
+                self.up_upscalers.append(nn.Identity())
+            up_block = nn.ModuleList()
+            for j in range(blocks[1][::-1][i]):
+                for k, block_type in enumerate(level_config[i]):
+                    c_skip = c_hidden[i] if i < len(c_hidden) - 1 and j == k == 0 else 0
+                    block = get_block(block_type, c_hidden[i], nhead[i], c_skip=c_skip, dropout=dropout[i],
+                                      self_attn=self_attn[i])
+                    up_block.append(block)
+            self.up_blocks.append(up_block)
+            if block_repeat is not None:
+                block_repeat_mappers = nn.ModuleList()
+                for _ in range(block_repeat[1][::-1][i] - 1):
+                    block_repeat_mappers.append(operations.Conv2d(c_hidden[i], c_hidden[i], kernel_size=1, dtype=dtype, device=device))
+                self.up_repeat_mappers.append(block_repeat_mappers)
+
+        # OUTPUT
+        self.clf = nn.Sequential(
+            LayerNorm2d_op(operations)(c_hidden[0], elementwise_affine=False, eps=1e-6, dtype=dtype, device=device),
+            operations.Conv2d(c_hidden[0], c_out * (patch_size ** 2), kernel_size=1, dtype=dtype, device=device),
+            nn.PixelShuffle(patch_size),
+        )
+
+        # --- WEIGHT INIT ---
+    #     self.apply(self._init_weights)  # General init
+    #     nn.init.normal_(self.clip_txt_mapper.weight, std=0.02)  # conditionings
+    #     nn.init.normal_(self.clip_txt_pooled_mapper.weight, std=0.02)  # conditionings
+    #     nn.init.normal_(self.clip_img_mapper.weight, std=0.02)  # conditionings
+    #     torch.nn.init.xavier_uniform_(self.embedding[1].weight, 0.02)  # inputs
+    #     nn.init.constant_(self.clf[1].weight, 0)  # outputs
+    #
+    #     # blocks
+    #     for level_block in self.down_blocks + self.up_blocks:
+    #         for block in level_block:
+    #             if isinstance(block, ResBlock) or isinstance(block, FeedForwardBlock):
+    #                 block.channelwise[-1].weight.data *= np.sqrt(1 / sum(blocks[0]))
+    #             elif isinstance(block, TimestepBlock):
+    #                 for layer in block.modules():
+    #                     if isinstance(layer, nn.Linear):
+    #                         nn.init.constant_(layer.weight, 0)
+    #
+    # def _init_weights(self, m):
+    #     if isinstance(m, (nn.Conv2d, nn.Linear)):
+    #         torch.nn.init.xavier_uniform_(m.weight)
+    #         if m.bias is not None:
+    #             nn.init.constant_(m.bias, 0)
+
+    def gen_r_embedding(self, r, max_positions=10000):
+        r = r * max_positions
+        half_dim = self.c_r // 2
+        emb = math.log(max_positions) / (half_dim - 1)
+        emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp()
+        emb = r[:, None] * emb[None, :]
+        emb = torch.cat([emb.sin(), emb.cos()], dim=1)
+        if self.c_r % 2 == 1:  # zero pad
+            emb = nn.functional.pad(emb, (0, 1), mode='constant')
+        return emb
+
+    def gen_c_embeddings(self, clip_txt, clip_txt_pooled, clip_img):
+        clip_txt = self.clip_txt_mapper(clip_txt)
+        if len(clip_txt_pooled.shape) == 2:
+            clip_txt_pooled = clip_txt_pooled.unsqueeze(1)
+        if len(clip_img.shape) == 2:
+            clip_img = clip_img.unsqueeze(1)
+        clip_txt_pool = self.clip_txt_pooled_mapper(clip_txt_pooled).view(clip_txt_pooled.size(0), clip_txt_pooled.size(1) * self.c_clip_seq, -1)
+        clip_img = self.clip_img_mapper(clip_img).view(clip_img.size(0), clip_img.size(1) * self.c_clip_seq, -1)
+        clip = torch.cat([clip_txt, clip_txt_pool, clip_img], dim=1)
+        clip = self.clip_norm(clip)
+        return clip
+
+    def _down_encode(self, x, r_embed, clip, cnet=None, transformer_options={}):
+        level_outputs = []
+        block_group = zip(self.down_blocks, self.down_downscalers, self.down_repeat_mappers)
+        for down_block, downscaler, repmap in block_group:
+            x = downscaler(x)
+            for i in range(len(repmap) + 1):
+                for block in down_block:
+                    if isinstance(block, ResBlock) or (
+                            hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
+                                                                                  ResBlock)):
+                        if cnet is not None:
+                            next_cnet = cnet.pop()
+                            if next_cnet is not None:
+                                x = x + nn.functional.interpolate(next_cnet, size=x.shape[-2:], mode='bilinear',
+                                                                  align_corners=True).to(x.dtype)
+                        x = block(x)
+                    elif isinstance(block, AttnBlock) or (
+                            hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
+                                                                                  AttnBlock)):
+                        x = block(x, clip, transformer_options=transformer_options)
+                    elif isinstance(block, TimestepBlock) or (
+                            hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
+                                                                                  TimestepBlock)):
+                        x = block(x, r_embed)
+                    else:
+                        x = block(x)
+                if i < len(repmap):
+                    x = repmap[i](x)
+            level_outputs.insert(0, x)
+        return level_outputs
+
+    def _up_decode(self, level_outputs, r_embed, clip, cnet=None, transformer_options={}):
+        x = level_outputs[0]
+        block_group = zip(self.up_blocks, self.up_upscalers, self.up_repeat_mappers)
+        for i, (up_block, upscaler, repmap) in enumerate(block_group):
+            for j in range(len(repmap) + 1):
+                for k, block in enumerate(up_block):
+                    if isinstance(block, ResBlock) or (
+                            hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
+                                                                                  ResBlock)):
+                        skip = level_outputs[i] if k == 0 and i > 0 else None
+                        if skip is not None and (x.size(-1) != skip.size(-1) or x.size(-2) != skip.size(-2)):
+                            x = torch.nn.functional.interpolate(x, skip.shape[-2:], mode='bilinear',
+                                                                align_corners=True)
+                        if cnet is not None:
+                            next_cnet = cnet.pop()
+                            if next_cnet is not None:
+                                x = x + nn.functional.interpolate(next_cnet, size=x.shape[-2:], mode='bilinear',
+                                                                  align_corners=True).to(x.dtype)
+                        x = block(x, skip)
+                    elif isinstance(block, AttnBlock) or (
+                            hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
+                                                                                  AttnBlock)):
+                        x = block(x, clip, transformer_options=transformer_options)
+                    elif isinstance(block, TimestepBlock) or (
+                            hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
+                                                                                  TimestepBlock)):
+                        x = block(x, r_embed)
+                    else:
+                        x = block(x)
+                if j < len(repmap):
+                    x = repmap[j](x)
+            x = upscaler(x)
+        return x
+
+    def forward(self, x, r, clip_text, clip_text_pooled, clip_img, control=None, transformer_options={}, **kwargs):
+        # Process the conditioning embeddings
+        r_embed = self.gen_r_embedding(r).to(dtype=x.dtype)
+        for c in self.t_conds:
+            t_cond = kwargs.get(c, torch.zeros_like(r))
+            r_embed = torch.cat([r_embed, self.gen_r_embedding(t_cond).to(dtype=x.dtype)], dim=1)
+        clip = self.gen_c_embeddings(clip_text, clip_text_pooled, clip_img)
+
+        if control is not None:
+            cnet = control.get("input")
+        else:
+            cnet = None
+
+        # Model Blocks
+        x = self.embedding(x)
+        level_outputs = self._down_encode(x, r_embed, clip, cnet, transformer_options=transformer_options)
+        x = self._up_decode(level_outputs, r_embed, clip, cnet, transformer_options=transformer_options)
+        return self.clf(x)
+
+    def update_weights_ema(self, src_model, beta=0.999):
+        for self_params, src_params in zip(self.parameters(), src_model.parameters()):
+            self_params.data = self_params.data * beta + src_params.data.clone().to(self_params.device) * (1 - beta)
+        for self_buffers, src_buffers in zip(self.buffers(), src_model.buffers()):
+            self_buffers.data = self_buffers.data * beta + src_buffers.data.clone().to(self_buffers.device) * (1 - beta)

ComfyUI/comfy/ldm/cascade/stage_c_coder.py

@@ -0,0 +1,98 @@
+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Stability AI
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+import torch
+import torchvision
+from torch import nn
+
+import comfy.ops
+
+ops = comfy.ops.disable_weight_init
+
+# EfficientNet
+class EfficientNetEncoder(nn.Module):
+    def __init__(self, c_latent=16):
+        super().__init__()
+        self.backbone = torchvision.models.efficientnet_v2_s().features.eval()
+        self.mapper = nn.Sequential(
+            ops.Conv2d(1280, c_latent, kernel_size=1, bias=False),
+            nn.BatchNorm2d(c_latent, affine=False),  # then normalize them to have mean 0 and std 1
+        )
+        self.mean = nn.Parameter(torch.tensor([0.485, 0.456, 0.406]))
+        self.std = nn.Parameter(torch.tensor([0.229, 0.224, 0.225]))
+
+    def forward(self, x):
+        x = x * 0.5 + 0.5
+        x = (x - self.mean.view([3,1,1]).to(device=x.device, dtype=x.dtype)) / self.std.view([3,1,1]).to(device=x.device, dtype=x.dtype)
+        o = self.mapper(self.backbone(x))
+        return o
+
+
+# Fast Decoder for Stage C latents. E.g. 16 x 24 x 24 -> 3 x 192 x 192
+class Previewer(nn.Module):
+    def __init__(self, c_in=16, c_hidden=512, c_out=3):
+        super().__init__()
+        self.blocks = nn.Sequential(
+            ops.Conv2d(c_in, c_hidden, kernel_size=1),  # 16 channels to 512 channels
+            nn.GELU(),
+            nn.BatchNorm2d(c_hidden),
+
+            ops.Conv2d(c_hidden, c_hidden, kernel_size=3, padding=1),
+            nn.GELU(),
+            nn.BatchNorm2d(c_hidden),
+
+            ops.ConvTranspose2d(c_hidden, c_hidden // 2, kernel_size=2, stride=2),  # 16 -> 32
+            nn.GELU(),
+            nn.BatchNorm2d(c_hidden // 2),
+
+            ops.Conv2d(c_hidden // 2, c_hidden // 2, kernel_size=3, padding=1),
+            nn.GELU(),
+            nn.BatchNorm2d(c_hidden // 2),
+
+            ops.ConvTranspose2d(c_hidden // 2, c_hidden // 4, kernel_size=2, stride=2),  # 32 -> 64
+            nn.GELU(),
+            nn.BatchNorm2d(c_hidden // 4),
+
+            ops.Conv2d(c_hidden // 4, c_hidden // 4, kernel_size=3, padding=1),
+            nn.GELU(),
+            nn.BatchNorm2d(c_hidden // 4),
+
+            ops.ConvTranspose2d(c_hidden // 4, c_hidden // 4, kernel_size=2, stride=2),  # 64 -> 128
+            nn.GELU(),
+            nn.BatchNorm2d(c_hidden // 4),
+
+            ops.Conv2d(c_hidden // 4, c_hidden // 4, kernel_size=3, padding=1),
+            nn.GELU(),
+            nn.BatchNorm2d(c_hidden // 4),
+
+            ops.Conv2d(c_hidden // 4, c_out, kernel_size=1),
+        )
+
+    def forward(self, x):
+        return (self.blocks(x) - 0.5) * 2.0
+
+class StageC_coder(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.previewer = Previewer()
+        self.encoder = EfficientNetEncoder()
+
+    def encode(self, x):
+        return self.encoder(x)
+
+    def decode(self, x):
+        return self.previewer(x)

ComfyUI/comfy/ldm/chroma/layers.py

@@ -0,0 +1,62 @@
+import torch
+from torch import Tensor, nn
+
+from comfy.ldm.flux.layers import (
+    MLPEmbedder,
+    ModulationOut,
+)
+
+# TODO: remove this in a few months
+SingleStreamBlock = None
+DoubleStreamBlock = None
+
+
+class ChromaModulationOut(ModulationOut):
+    @classmethod
+    def from_offset(cls, tensor: torch.Tensor, offset: int = 0) -> ModulationOut:
+        return cls(
+            shift=tensor[:, offset : offset + 1, :],
+            scale=tensor[:, offset + 1 : offset + 2, :],
+            gate=tensor[:, offset + 2 : offset + 3, :],
+        )
+
+
+
+
+class Approximator(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int, hidden_dim: int, n_layers = 5, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.in_proj = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.layers = nn.ModuleList([MLPEmbedder(hidden_dim, hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
+        self.norms = nn.ModuleList([operations.RMSNorm(hidden_dim, dtype=dtype, device=device) for x in range( n_layers)])
+        self.out_proj = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device)
+
+    @property
+    def device(self):
+        # Get the device of the module (assumes all parameters are on the same device)
+        return next(self.parameters()).device
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.in_proj(x)
+
+        for layer, norms in zip(self.layers, self.norms):
+            x = x + layer(norms(x))
+
+        x = self.out_proj(x)
+
+        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, out_channels, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x: Tensor, vec: Tensor) -> Tensor:
+        shift, scale = vec
+        shift = shift.squeeze(1)
+        scale = scale.squeeze(1)
+        x = torch.addcmul(shift[:, None, :], 1 + scale[:, None, :], self.norm_final(x))
+        x = self.linear(x)
+        return x

ComfyUI/comfy/ldm/chroma/model.py

@@ -0,0 +1,295 @@
+#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.patcher_extension
+import comfy.ldm.common_dit
+
+from comfy.ldm.flux.layers import (
+    EmbedND,
+    timestep_embedding,
+    DoubleStreamBlock,
+    SingleStreamBlock,
+)
+
+from .layers import (
+    LastLayer,
+    Approximator,
+    ChromaModulationOut,
+)
+
+
+@dataclass
+class ChromaParams:
+    in_channels: int
+    out_channels: 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
+    in_dim: int
+    out_dim: int
+    hidden_dim: int
+    n_layers: int
+    txt_ids_dims: list
+    vec_in_dim: int
+
+
+
+class Chroma(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 = ChromaParams(**kwargs)
+        self.params = params
+        self.patch_size = params.patch_size
+        self.in_channels = params.in_channels
+        self.out_channels = params.out_channels
+        if params.hidden_size % params.num_heads != 0:
+            raise ValueError(
+                f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
+            )
+        pe_dim = params.hidden_size // params.num_heads
+        if sum(params.axes_dim) != pe_dim:
+            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
+        self.hidden_size = params.hidden_size
+        self.num_heads = params.num_heads
+        self.in_dim = params.in_dim
+        self.out_dim = params.out_dim
+        self.hidden_dim = params.hidden_dim
+        self.n_layers = params.n_layers
+        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.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
+        # set as nn identity for now, will overwrite it later.
+        self.distilled_guidance_layer = Approximator(
+                    in_dim=self.in_dim,
+                    hidden_dim=self.hidden_dim,
+                    out_dim=self.out_dim,
+                    n_layers=self.n_layers,
+                    dtype=dtype, device=device, operations=operations
+                )
+
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                    modulation=False,
+                    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, modulation=False, 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)
+
+        self.skip_mmdit = []
+        self.skip_dit = []
+        self.lite = False
+
+    def get_modulations(self, tensor: torch.Tensor, block_type: str, *, idx: int = 0):
+        # This function slices up the modulations tensor which has the following layout:
+        #   single     : num_single_blocks * 3 elements
+        #   double_img : num_double_blocks * 6 elements
+        #   double_txt : num_double_blocks * 6 elements
+        #   final      : 2 elements
+        if block_type == "final":
+            return (tensor[:, -2:-1, :], tensor[:, -1:, :])
+        single_block_count = self.params.depth_single_blocks
+        double_block_count = self.params.depth
+        offset = 3 * idx
+        if block_type == "single":
+            return ChromaModulationOut.from_offset(tensor, offset)
+        # Double block modulations are 6 elements so we double 3 * idx.
+        offset *= 2
+        if block_type in {"double_img", "double_txt"}:
+            # Advance past the single block modulations.
+            offset += 3 * single_block_count
+            if block_type == "double_txt":
+                # Advance past the double block img modulations.
+                offset += 6 * double_block_count
+            return (
+                ChromaModulationOut.from_offset(tensor, offset),
+                ChromaModulationOut.from_offset(tensor, offset + 3),
+            )
+        raise ValueError("Bad block_type")
+
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        guidance: Tensor = None,
+        control = None,
+        transformer_options={},
+        attn_mask: Tensor = None,
+    ) -> Tensor:
+        transformer_options = transformer_options.copy()
+        patches_replace = transformer_options.get("patches_replace", {})
+
+        # running on sequences img
+        img = self.img_in(img)
+
+        # distilled vector guidance
+        mod_index_length = 344
+        distill_timestep = timestep_embedding(timesteps.detach().clone(), 16).to(img.device, img.dtype)
+        # guidance = guidance *
+        distil_guidance = timestep_embedding(guidance.detach().clone(), 16).to(img.device, img.dtype)
+
+        # get all modulation index
+        modulation_index = timestep_embedding(torch.arange(mod_index_length, device=img.device), 32).to(img.device, img.dtype)
+        # we need to broadcast the modulation index here so each batch has all of the index
+        modulation_index = modulation_index.unsqueeze(0).repeat(img.shape[0], 1, 1).to(img.device, img.dtype)
+        # and we need to broadcast timestep and guidance along too
+        timestep_guidance = torch.cat([distill_timestep, distil_guidance], dim=1).unsqueeze(1).repeat(1, mod_index_length, 1).to(img.dtype).to(img.device, img.dtype)
+        # then and only then we could concatenate it together
+        input_vec = torch.cat([timestep_guidance, modulation_index], dim=-1).to(img.device, img.dtype)
+
+        mod_vectors = self.distilled_guidance_layer(input_vec)
+
+        txt = self.txt_in(txt)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.double_blocks)
+        transformer_options["block_type"] = "double"
+        for i, block in enumerate(self.double_blocks):
+            transformer_options["block_index"] = i
+            if i not in self.skip_mmdit:
+                double_mod = (
+                    self.get_modulations(mod_vectors, "double_img", idx=i),
+                    self.get_modulations(mod_vectors, "double_txt", idx=i),
+                )
+                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"),
+                                                       transformer_options=args.get("transformer_options"))
+                        return out
+
+                    out = blocks_replace[("double_block", i)]({"img": img,
+                                                               "txt": txt,
+                                                               "vec": double_mod,
+                                                               "pe": pe,
+                                                               "attn_mask": attn_mask,
+                                                               "transformer_options": transformer_options},
+                                                              {"original_block": block_wrap})
+                    txt = out["txt"]
+                    img = out["img"]
+                else:
+                    img, txt = block(img=img,
+                                     txt=txt,
+                                     vec=double_mod,
+                                     pe=pe,
+                                     attn_mask=attn_mask,
+                                     transformer_options=transformer_options)
+
+                if control is not None: # Controlnet
+                    control_i = control.get("input")
+                    if i < len(control_i):
+                        add = control_i[i]
+                        if add is not None:
+                            img += add
+
+        img = torch.cat((txt, img), 1)
+
+        transformer_options["total_blocks"] = len(self.single_blocks)
+        transformer_options["block_type"] = "single"
+        transformer_options["img_slice"] = [txt.shape[1], img.shape[1]]
+        for i, block in enumerate(self.single_blocks):
+            transformer_options["block_index"] = i
+            if i not in self.skip_dit:
+                single_mod = self.get_modulations(mod_vectors, "single", idx=i)
+                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"),
+                                           transformer_options=args.get("transformer_options"))
+                        return out
+
+                    out = blocks_replace[("single_block", i)]({"img": img,
+                                                               "vec": single_mod,
+                                                               "pe": pe,
+                                                               "attn_mask": attn_mask,
+                                                               "transformer_options": transformer_options},
+                                                              {"original_block": block_wrap})
+                    img = out["img"]
+                else:
+                    img = block(img, vec=single_mod, pe=pe, attn_mask=attn_mask, transformer_options=transformer_options)
+
+                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] :, ...] += add
+
+        img = img[:, txt.shape[1] :, ...]
+        if hasattr(self, "final_layer"):
+            final_mod = self.get_modulations(mod_vectors, "final")
+            img = self.final_layer(img, vec=final_mod)  # (N, T, patch_size ** 2 * out_channels)
+        return img
+
+    def forward(self, x, timestep, context, guidance, 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, guidance, control, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs):
+        bs, c, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+
+        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=self.patch_size, pw=self.patch_size)
+
+        if img.ndim != 3 or context.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        h_len = ((h + (self.patch_size // 2)) // self.patch_size)
+        w_len = ((w + (self.patch_size // 2)) // self.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).unsqueeze(1)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        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)
+        out = self.forward_orig(img, img_ids, context, txt_ids, timestep, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
+        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=self.patch_size, pw=self.patch_size)[:,:,:h,:w]

ComfyUI/comfy/ldm/chroma_radiance/layers.py

@@ -0,0 +1,204 @@
+# Adapted from https://github.com/lodestone-rock/flow
+from functools import lru_cache
+
+import torch
+from torch import nn
+
+
+class NerfEmbedder(nn.Module):
+    """
+    An embedder module that combines input features with a 2D positional
+    encoding that mimics the Discrete Cosine Transform (DCT).
+
+    This module takes an input tensor of shape (B, P^2, C), where P is the
+    patch size, and enriches it with positional information before projecting
+    it to a new hidden size.
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        hidden_size_input: int,
+        max_freqs: int,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        """
+        Initializes the NerfEmbedder.
+
+        Args:
+            in_channels (int): The number of channels in the input tensor.
+            hidden_size_input (int): The desired dimension of the output embedding.
+            max_freqs (int): The number of frequency components to use for both
+                             the x and y dimensions of the positional encoding.
+                             The total number of positional features will be max_freqs^2.
+        """
+        super().__init__()
+        self.dtype = dtype
+        self.max_freqs = max_freqs
+        self.hidden_size_input = hidden_size_input
+
+        # A linear layer to project the concatenated input features and
+        # positional encodings to the final output dimension.
+        self.embedder = nn.Sequential(
+            operations.Linear(in_channels + max_freqs**2, hidden_size_input, dtype=dtype, device=device)
+        )
+
+    @lru_cache(maxsize=4)
+    def fetch_pos(self, patch_size: int, device: torch.device, dtype: torch.dtype) -> torch.Tensor:
+        """
+        Generates and caches 2D DCT-like positional embeddings for a given patch size.
+
+        The LRU cache is a performance optimization that avoids recomputing the
+        same positional grid on every forward pass.
+
+        Args:
+            patch_size (int): The side length of the square input patch.
+            device: The torch device to create the tensors on.
+            dtype: The torch dtype for the tensors.
+
+        Returns:
+            A tensor of shape (1, patch_size^2, max_freqs^2) containing the
+            positional embeddings.
+        """
+        # Create normalized 1D coordinate grids from 0 to 1.
+        pos_x = torch.linspace(0, 1, patch_size, device=device, dtype=dtype)
+        pos_y = torch.linspace(0, 1, patch_size, device=device, dtype=dtype)
+
+        # Create a 2D meshgrid of coordinates.
+        pos_y, pos_x = torch.meshgrid(pos_y, pos_x, indexing="ij")
+
+        # Reshape positions to be broadcastable with frequencies.
+        # Shape becomes (patch_size^2, 1, 1).
+        pos_x = pos_x.reshape(-1, 1, 1)
+        pos_y = pos_y.reshape(-1, 1, 1)
+
+        # Create a 1D tensor of frequency values from 0 to max_freqs-1.
+        freqs = torch.linspace(0, self.max_freqs - 1, self.max_freqs, dtype=dtype, device=device)
+
+        # Reshape frequencies to be broadcastable for creating 2D basis functions.
+        # freqs_x shape: (1, max_freqs, 1)
+        # freqs_y shape: (1, 1, max_freqs)
+        freqs_x = freqs[None, :, None]
+        freqs_y = freqs[None, None, :]
+
+        # A custom weighting coefficient, not part of standard DCT.
+        # This seems to down-weight the contribution of higher-frequency interactions.
+        coeffs = (1 + freqs_x * freqs_y) ** -1
+
+        # Calculate the 1D cosine basis functions for x and y coordinates.
+        # This is the core of the DCT formulation.
+        dct_x = torch.cos(pos_x * freqs_x * torch.pi)
+        dct_y = torch.cos(pos_y * freqs_y * torch.pi)
+
+        # Combine the 1D basis functions to create 2D basis functions by element-wise
+        # multiplication, and apply the custom coefficients. Broadcasting handles the
+        # combination of all (pos_x, freqs_x) with all (pos_y, freqs_y).
+        # The result is flattened into a feature vector for each position.
+        dct = (dct_x * dct_y * coeffs).view(1, -1, self.max_freqs ** 2)
+
+        return dct
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        """
+        Forward pass for the embedder.
+
+        Args:
+            inputs (Tensor): The input tensor of shape (B, P^2, C).
+
+        Returns:
+            Tensor: The output tensor of shape (B, P^2, hidden_size_input).
+        """
+        # Get the batch size, number of pixels, and number of channels.
+        B, P2, C = inputs.shape
+
+        # Infer the patch side length from the number of pixels (P^2).
+        patch_size = int(P2 ** 0.5)
+
+        input_dtype = inputs.dtype
+        inputs = inputs.to(dtype=self.dtype)
+
+        # Fetch the pre-computed or cached positional embeddings.
+        dct = self.fetch_pos(patch_size, inputs.device, self.dtype)
+
+        # Repeat the positional embeddings for each item in the batch.
+        dct = dct.repeat(B, 1, 1)
+
+        # Concatenate the original input features with the positional embeddings
+        # along the feature dimension.
+        inputs = torch.cat((inputs, dct), dim=-1)
+
+        # Project the combined tensor to the target hidden size.
+        return self.embedder(inputs).to(dtype=input_dtype)
+
+
+class NerfGLUBlock(nn.Module):
+    """
+    A NerfBlock using a Gated Linear Unit (GLU) like MLP.
+    """
+    def __init__(self, hidden_size_s: int, hidden_size_x: int, mlp_ratio, dtype=None, device=None, operations=None):
+        super().__init__()
+        # The total number of parameters for the MLP is increased to accommodate
+        # the gate, value, and output projection matrices.
+        # We now need to generate parameters for 3 matrices.
+        total_params = 3 * hidden_size_x**2 * mlp_ratio
+        self.param_generator = operations.Linear(hidden_size_s, total_params, dtype=dtype, device=device)
+        self.norm = operations.RMSNorm(hidden_size_x, dtype=dtype, device=device)
+        self.mlp_ratio = mlp_ratio
+
+
+    def forward(self, x: torch.Tensor, s: torch.Tensor) -> torch.Tensor:
+        batch_size, num_x, hidden_size_x = x.shape
+        mlp_params = self.param_generator(s)
+
+        # Split the generated parameters into three parts for the gate, value, and output projection.
+        fc1_gate_params, fc1_value_params, fc2_params = mlp_params.chunk(3, dim=-1)
+
+        # Reshape the parameters into matrices for batch matrix multiplication.
+        fc1_gate = fc1_gate_params.view(batch_size, hidden_size_x, hidden_size_x * self.mlp_ratio)
+        fc1_value = fc1_value_params.view(batch_size, hidden_size_x, hidden_size_x * self.mlp_ratio)
+        fc2 = fc2_params.view(batch_size, hidden_size_x * self.mlp_ratio, hidden_size_x)
+
+        # Normalize the generated weight matrices as in the original implementation.
+        fc1_gate = torch.nn.functional.normalize(fc1_gate, dim=-2)
+        fc1_value = torch.nn.functional.normalize(fc1_value, dim=-2)
+        fc2 = torch.nn.functional.normalize(fc2, dim=-2)
+
+        res_x = x
+        x = self.norm(x)
+
+        # Apply the final output projection.
+        x = torch.bmm(torch.nn.functional.silu(torch.bmm(x, fc1_gate)) * torch.bmm(x, fc1_value), fc2)
+
+        return x + res_x
+
+
+class NerfFinalLayer(nn.Module):
+    def __init__(self, hidden_size, out_channels, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm = operations.RMSNorm(hidden_size, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, out_channels, dtype=dtype, device=device)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # RMSNorm normalizes over the last dimension, but our channel dim (C) is at dim=1.
+        # So we temporarily move the channel dimension to the end for the norm operation.
+        return self.linear(self.norm(x.movedim(1, -1))).movedim(-1, 1)
+
+
+class NerfFinalLayerConv(nn.Module):
+    def __init__(self, hidden_size: int, out_channels: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm = operations.RMSNorm(hidden_size, dtype=dtype, device=device)
+        self.conv = operations.Conv2d(
+            in_channels=hidden_size,
+            out_channels=out_channels,
+            kernel_size=3,
+            padding=1,
+            dtype=dtype,
+            device=device,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # RMSNorm normalizes over the last dimension, but our channel dim (C) is at dim=1.
+        # So we temporarily move the channel dimension to the end for the norm operation.
+        return self.conv(self.norm(x.movedim(1, -1)).movedim(-1, 1))

ComfyUI/comfy/ldm/chroma_radiance/model.py

@@ -0,0 +1,343 @@
+# Credits:
+# Original Flux code can be found on: https://github.com/black-forest-labs/flux
+# Chroma Radiance adaption referenced from https://github.com/lodestone-rock/flow
+
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+from torch import Tensor, nn
+from einops import repeat
+import comfy.ldm.common_dit
+
+from comfy.ldm.flux.layers import EmbedND, DoubleStreamBlock, SingleStreamBlock
+
+from comfy.ldm.chroma.model import Chroma, ChromaParams
+from comfy.ldm.chroma.layers import (
+    Approximator,
+)
+from .layers import (
+    NerfEmbedder,
+    NerfGLUBlock,
+    NerfFinalLayer,
+    NerfFinalLayerConv,
+)
+
+
+@dataclass
+class ChromaRadianceParams(ChromaParams):
+    patch_size: int
+    nerf_hidden_size: int
+    nerf_mlp_ratio: int
+    nerf_depth: int
+    nerf_max_freqs: int
+    # Setting nerf_tile_size to 0 disables tiling.
+    nerf_tile_size: int
+    # Currently one of linear (legacy) or conv.
+    nerf_final_head_type: str
+    # None means use the same dtype as the model.
+    nerf_embedder_dtype: Optional[torch.dtype]
+    use_x0: bool
+    # Use sequential txt_ids instead of zeros
+    use_sequential_txt_ids: bool
+
+class ChromaRadiance(Chroma):
+    """
+    Transformer model for flow matching on sequences.
+    """
+
+    def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
+        if operations is None:
+            raise RuntimeError("Attempt to create ChromaRadiance object without setting operations")
+        nn.Module.__init__(self)
+        self.dtype = dtype
+        params = ChromaRadianceParams(**kwargs)
+        self.params = params
+        self.patch_size = params.patch_size
+        self.in_channels = params.in_channels
+        self.out_channels = params.out_channels
+        if params.hidden_size % params.num_heads != 0:
+            raise ValueError(
+                f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
+            )
+        pe_dim = params.hidden_size // params.num_heads
+        if sum(params.axes_dim) != pe_dim:
+            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
+        self.hidden_size = params.hidden_size
+        self.num_heads = params.num_heads
+        self.in_dim = params.in_dim
+        self.out_dim = params.out_dim
+        self.hidden_dim = params.hidden_dim
+        self.n_layers = params.n_layers
+        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
+        self.img_in_patch = operations.Conv2d(
+            params.in_channels,
+            params.hidden_size,
+            kernel_size=params.patch_size,
+            stride=params.patch_size,
+            bias=True,
+            dtype=dtype,
+            device=device,
+        )
+        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
+        # set as nn identity for now, will overwrite it later.
+        self.distilled_guidance_layer = Approximator(
+                    in_dim=self.in_dim,
+                    hidden_dim=self.hidden_dim,
+                    out_dim=self.out_dim,
+                    n_layers=self.n_layers,
+                    dtype=dtype, device=device, operations=operations
+                )
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                    modulation=False,
+                    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,
+                    modulation=False,
+                    dtype=dtype, device=device, operations=operations,
+                )
+                for _ in range(params.depth_single_blocks)
+            ]
+        )
+
+        # pixel channel concat with DCT
+        self.nerf_image_embedder = NerfEmbedder(
+            in_channels=params.in_channels,
+            hidden_size_input=params.nerf_hidden_size,
+            max_freqs=params.nerf_max_freqs,
+            dtype=params.nerf_embedder_dtype or dtype,
+            device=device,
+            operations=operations,
+        )
+
+        self.nerf_blocks = nn.ModuleList([
+            NerfGLUBlock(
+                hidden_size_s=params.hidden_size,
+                hidden_size_x=params.nerf_hidden_size,
+                mlp_ratio=params.nerf_mlp_ratio,
+                dtype=dtype,
+                device=device,
+                operations=operations,
+            ) for _ in range(params.nerf_depth)
+        ])
+
+        if params.nerf_final_head_type == "linear":
+            self.nerf_final_layer = NerfFinalLayer(
+                params.nerf_hidden_size,
+                out_channels=params.in_channels,
+                dtype=dtype,
+                device=device,
+                operations=operations,
+            )
+        elif params.nerf_final_head_type == "conv":
+            self.nerf_final_layer_conv = NerfFinalLayerConv(
+                params.nerf_hidden_size,
+                out_channels=params.in_channels,
+                dtype=dtype,
+                device=device,
+                operations=operations,
+            )
+        else:
+            errstr = f"Unsupported nerf_final_head_type {params.nerf_final_head_type}"
+            raise ValueError(errstr)
+
+        self.skip_mmdit = []
+        self.skip_dit = []
+        self.lite = False
+
+        if params.use_x0:
+            self.register_buffer("__x0__", torch.tensor([]))
+
+        if params.use_sequential_txt_ids:
+            self.register_buffer("__sequential__", torch.tensor([]))
+
+    @property
+    def _nerf_final_layer(self) -> nn.Module:
+        if self.params.nerf_final_head_type == "linear":
+            return self.nerf_final_layer
+        if self.params.nerf_final_head_type == "conv":
+            return self.nerf_final_layer_conv
+        # Impossible to get here as we raise an error on unexpected types on initialization.
+        raise NotImplementedError
+
+    def img_in(self, img: Tensor) -> Tensor:
+        img = self.img_in_patch(img) # -> [B, Hidden, H/P, W/P]
+        # flatten into a sequence for the transformer.
+        return img.flatten(2).transpose(1, 2) # -> [B, NumPatches, Hidden]
+
+    def forward_nerf(
+        self,
+        img_orig: Tensor,
+        img_out: Tensor,
+        params: ChromaRadianceParams,
+    ) -> Tensor:
+        B, C, H, W = img_orig.shape
+        num_patches = img_out.shape[1]
+        patch_size = params.patch_size
+
+        # Store the raw pixel values of each patch for the NeRF head later.
+        # unfold creates patches: [B, C * P * P, NumPatches]
+        nerf_pixels = nn.functional.unfold(img_orig, kernel_size=patch_size, stride=patch_size)
+        nerf_pixels = nerf_pixels.transpose(1, 2) # -> [B, NumPatches, C * P * P]
+
+        # Reshape for per-patch processing
+        nerf_hidden = img_out.reshape(B * num_patches, params.hidden_size)
+        nerf_pixels = nerf_pixels.reshape(B * num_patches, C, patch_size**2).transpose(1, 2)
+
+        if params.nerf_tile_size > 0 and num_patches > params.nerf_tile_size:
+            # Enable tiling if nerf_tile_size isn't 0 and we actually have more patches than
+            # the tile size.
+            img_dct = self.forward_tiled_nerf(nerf_hidden, nerf_pixels, B, C, num_patches, patch_size, params)
+        else:
+            # Get DCT-encoded pixel embeddings [pixel-dct]
+            img_dct = self.nerf_image_embedder(nerf_pixels)
+
+            # Pass through the dynamic MLP blocks (the NeRF)
+            for block in self.nerf_blocks:
+                img_dct = block(img_dct, nerf_hidden)
+
+        # Reassemble the patches into the final image.
+        img_dct = img_dct.transpose(1, 2) # -> [B*NumPatches, C, P*P]
+        # Reshape to combine with batch dimension for fold
+        img_dct = img_dct.reshape(B, num_patches, -1) # -> [B, NumPatches, C*P*P]
+        img_dct = img_dct.transpose(1, 2) # -> [B, C*P*P, NumPatches]
+        img_dct = nn.functional.fold(
+            img_dct,
+            output_size=(H, W),
+            kernel_size=patch_size,
+            stride=patch_size,
+        )
+        return self._nerf_final_layer(img_dct)
+
+    def forward_tiled_nerf(
+        self,
+        nerf_hidden: Tensor,
+        nerf_pixels: Tensor,
+        batch: int,
+        channels: int,
+        num_patches: int,
+        patch_size: int,
+        params: ChromaRadianceParams,
+    ) -> Tensor:
+        """
+        Processes the NeRF head in tiles to save memory.
+        nerf_hidden has shape [B, L, D]
+        nerf_pixels has shape [B, L, C * P * P]
+        """
+        tile_size = params.nerf_tile_size
+        output_tiles = []
+        # Iterate over the patches in tiles. The dimension L (num_patches) is at index 1.
+        for i in range(0, num_patches, tile_size):
+            end = min(i + tile_size, num_patches)
+
+            # Slice the current tile from the input tensors
+            nerf_hidden_tile = nerf_hidden[i * batch:end * batch]
+            nerf_pixels_tile = nerf_pixels[i * batch:end * batch]
+
+            # get DCT-encoded pixel embeddings [pixel-dct]
+            img_dct_tile = self.nerf_image_embedder(nerf_pixels_tile)
+
+            # pass through the dynamic MLP blocks (the NeRF)
+            for block in self.nerf_blocks:
+                img_dct_tile = block(img_dct_tile, nerf_hidden_tile)
+
+            output_tiles.append(img_dct_tile)
+
+        # Concatenate the processed tiles along the patch dimension
+        return torch.cat(output_tiles, dim=0)
+
+    def radiance_get_override_params(self, overrides: dict) -> ChromaRadianceParams:
+        params = self.params
+        if not overrides:
+            return params
+        params_dict = {k: getattr(params, k) for k in params.__dataclass_fields__}
+        nullable_keys = frozenset(("nerf_embedder_dtype",))
+        bad_keys = tuple(k for k in overrides if k not in params_dict)
+        if bad_keys:
+            e = f"Unknown key(s) in transformer_options chroma_radiance_options: {', '.join(bad_keys)}"
+            raise ValueError(e)
+        bad_keys = tuple(
+            k
+            for k, v in overrides.items()
+            if not isinstance(v, type(getattr(params, k))) and (v is not None or k not in nullable_keys)
+        )
+        if bad_keys:
+            e = f"Invalid value(s) in transformer_options chroma_radiance_options: {', '.join(bad_keys)}"
+            raise ValueError(e)
+        # At this point it's all valid keys and values so we can merge with the existing params.
+        params_dict |= overrides
+        return params.__class__(**params_dict)
+
+    def _apply_x0_residual(self, predicted, noisy, timesteps):
+
+        # non zero during training to prevent 0 div
+        eps = 0.0
+        return (noisy - predicted) / (timesteps.view(-1,1,1,1) + eps)
+
+    def _forward(
+        self,
+        x: Tensor,
+        timestep: Tensor,
+        context: Tensor,
+        guidance: Optional[Tensor],
+        control: Optional[dict]=None,
+        transformer_options: dict={},
+        **kwargs: dict,
+    ) -> Tensor:
+        bs, c, h, w = x.shape
+        img = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+
+        if img.ndim != 4:
+            raise ValueError("Input img tensor must be in [B, C, H, W] format.")
+        if context.ndim != 3:
+            raise ValueError("Input txt tensors must have 3 dimensions.")
+
+        params = self.radiance_get_override_params(transformer_options.get("chroma_radiance_options", {}))
+
+        h_len = (img.shape[-2] // self.patch_size)
+        w_len = (img.shape[-1] // self.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).unsqueeze(1)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        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)
+        # Radiance after 2026-05-22 uses sequential txt_ids instead of zeros
+        if params.use_sequential_txt_ids:
+            txt_ids[:, :, 0] = torch.arange(context.shape[1], device=x.device, dtype=x.dtype).unsqueeze(0).expand(bs, -1)
+
+        img_out = self.forward_orig(
+            img,
+            img_ids,
+            context,
+            txt_ids,
+            timestep,
+            guidance,
+            control,
+            transformer_options,
+            attn_mask=kwargs.get("attention_mask", None),
+        )
+
+        out = self.forward_nerf(img, img_out, params)[:, :, :h, :w]
+
+        # If x0 variant → v-pred, just return this instead
+        if hasattr(self, "__x0__"):
+            out = self._apply_x0_residual(out, img, timestep)
+        return out
+

ComfyUI/comfy/ldm/cogvideo/model.py

@@ -0,0 +1,573 @@
+# CogVideoX 3D Transformer - ported to ComfyUI native ops
+# Architecture reference: diffusers CogVideoXTransformer3DModel
+# Style reference: comfy/ldm/wan/model.py
+
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.patcher_extension
+import comfy.ldm.common_dit
+
+
+def _get_1d_rotary_pos_embed(dim, pos, theta=10000.0):
+    """Returns (cos, sin) each with shape [seq_len, dim].
+
+    Frequencies are computed at dim//2 resolution then repeat_interleaved
+    to full dim, matching CogVideoX's interleaved (real, imag) pair format.
+    """
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float32, device=pos.device) / dim))
+    angles = torch.outer(pos.float(), freqs.float())
+    cos = angles.cos().repeat_interleave(2, dim=-1).float()
+    sin = angles.sin().repeat_interleave(2, dim=-1).float()
+    return (cos, sin)
+
+
+def apply_rotary_emb(x, freqs_cos_sin):
+    """Apply CogVideoX rotary embedding to query or key tensor.
+
+    x: [B, heads, seq_len, head_dim]
+    freqs_cos_sin: (cos, sin) each [seq_len, head_dim//2]
+
+    Uses interleaved pair rotation (same as diffusers CogVideoX/Flux).
+    head_dim is reshaped to (-1, 2) pairs, rotated, then flattened back.
+    """
+    cos, sin = freqs_cos_sin
+    cos = cos[None, None, :, :].to(x.device)
+    sin = sin[None, None, :, :].to(x.device)
+
+    # Interleaved pairs: [B, H, S, D] -> [B, H, S, D//2, 2] -> (real, imag)
+    x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)
+    x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+
+    return (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+
+def get_timestep_embedding(timesteps, dim, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1, max_period=10000):
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=timesteps.device) / half)
+    args = timesteps[:, None].float() * freqs[None] * scale
+    embedding = torch.cat([torch.sin(args), torch.cos(args)], dim=-1)
+    if flip_sin_to_cos:
+        embedding = torch.cat([embedding[:, half:], embedding[:, :half]], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+def get_3d_sincos_pos_embed(embed_dim, spatial_size, temporal_size, spatial_interpolation_scale=1.0, temporal_interpolation_scale=1.0, device=None):
+    if isinstance(spatial_size, int):
+        spatial_size = (spatial_size, spatial_size)
+
+    grid_w = torch.arange(spatial_size[0], dtype=torch.float32, device=device) / spatial_interpolation_scale
+    grid_h = torch.arange(spatial_size[1], dtype=torch.float32, device=device) / spatial_interpolation_scale
+    grid_t = torch.arange(temporal_size, dtype=torch.float32, device=device) / temporal_interpolation_scale
+
+    grid_t, grid_h, grid_w = torch.meshgrid(grid_t, grid_h, grid_w, indexing="ij")
+
+    embed_dim_spatial = 2 * (embed_dim // 3)
+    embed_dim_temporal = embed_dim // 3
+
+    pos_embed_spatial = _get_2d_sincos_pos_embed(embed_dim_spatial, grid_h, grid_w, device=device)
+    pos_embed_temporal = _get_1d_sincos_pos_embed(embed_dim_temporal, grid_t[:, 0, 0], device=device)
+
+    T, H, W = grid_t.shape
+    pos_embed_temporal = pos_embed_temporal.unsqueeze(1).unsqueeze(1).expand(-1, H, W, -1)
+    pos_embed = torch.cat([pos_embed_temporal, pos_embed_spatial], dim=-1)
+
+    return pos_embed
+
+
+def _get_2d_sincos_pos_embed(embed_dim, grid_h, grid_w, device=None):
+    T, H, W = grid_h.shape
+    half_dim = embed_dim // 2
+    pos_h = _get_1d_sincos_pos_embed(half_dim, grid_h.reshape(-1), device=device).reshape(T, H, W, half_dim)
+    pos_w = _get_1d_sincos_pos_embed(half_dim, grid_w.reshape(-1), device=device).reshape(T, H, W, half_dim)
+    return torch.cat([pos_h, pos_w], dim=-1)
+
+
+def _get_1d_sincos_pos_embed(embed_dim, pos, device=None):
+    half = embed_dim // 2
+    freqs = torch.exp(-math.log(10000.0) * torch.arange(start=0, end=half, dtype=torch.float32, device=device) / half)
+    args = pos.float().reshape(-1)[:, None] * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if embed_dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+
+class CogVideoXPatchEmbed(nn.Module):
+    def __init__(self, patch_size=2, patch_size_t=None, in_channels=16, dim=1920,
+                 text_dim=4096, bias=True, sample_width=90, sample_height=60,
+                 sample_frames=49, temporal_compression_ratio=4,
+                 max_text_seq_length=226, spatial_interpolation_scale=1.875,
+                 temporal_interpolation_scale=1.0, use_positional_embeddings=True,
+                 use_learned_positional_embeddings=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
+        self.dim = dim
+        self.sample_height = sample_height
+        self.sample_width = sample_width
+        self.sample_frames = sample_frames
+        self.temporal_compression_ratio = temporal_compression_ratio
+        self.max_text_seq_length = max_text_seq_length
+        self.spatial_interpolation_scale = spatial_interpolation_scale
+        self.temporal_interpolation_scale = temporal_interpolation_scale
+        self.use_positional_embeddings = use_positional_embeddings
+        self.use_learned_positional_embeddings = use_learned_positional_embeddings
+
+        if patch_size_t is None:
+            self.proj = operations.Conv2d(in_channels, dim, kernel_size=patch_size, stride=patch_size, bias=bias, device=device, dtype=dtype)
+        else:
+            self.proj = operations.Linear(in_channels * patch_size * patch_size * patch_size_t, dim, device=device, dtype=dtype)
+
+        self.text_proj = operations.Linear(text_dim, dim, device=device, dtype=dtype)
+
+        if use_positional_embeddings or use_learned_positional_embeddings:
+            persistent = use_learned_positional_embeddings
+            pos_embedding = self._get_positional_embeddings(sample_height, sample_width, sample_frames)
+            self.register_buffer("pos_embedding", pos_embedding, persistent=persistent)
+
+    def _get_positional_embeddings(self, sample_height, sample_width, sample_frames, device=None):
+        post_patch_height = sample_height // self.patch_size
+        post_patch_width = sample_width // self.patch_size
+        post_time_compression_frames = (sample_frames - 1) // self.temporal_compression_ratio + 1
+        if self.patch_size_t is not None:
+            post_time_compression_frames = post_time_compression_frames // self.patch_size_t
+        num_patches = post_patch_height * post_patch_width * post_time_compression_frames
+
+        pos_embedding = get_3d_sincos_pos_embed(
+            self.dim,
+            (post_patch_width, post_patch_height),
+            post_time_compression_frames,
+            self.spatial_interpolation_scale,
+            self.temporal_interpolation_scale,
+            device=device,
+        )
+        pos_embedding = pos_embedding.reshape(-1, self.dim)
+        joint_pos_embedding = pos_embedding.new_zeros(
+            1, self.max_text_seq_length + num_patches, self.dim, requires_grad=False
+        )
+        joint_pos_embedding.data[:, self.max_text_seq_length:].copy_(pos_embedding)
+        return joint_pos_embedding
+
+    def forward(self, text_embeds, image_embeds):
+        input_dtype = text_embeds.dtype
+        text_embeds = self.text_proj(text_embeds.to(self.text_proj.weight.dtype)).to(input_dtype)
+        batch_size, num_frames, channels, height, width = image_embeds.shape
+
+        proj_dtype = self.proj.weight.dtype
+        if self.patch_size_t is None:
+            image_embeds = image_embeds.reshape(-1, channels, height, width)
+            image_embeds = self.proj(image_embeds.to(proj_dtype)).to(input_dtype)
+            image_embeds = image_embeds.view(batch_size, num_frames, *image_embeds.shape[1:])
+            image_embeds = image_embeds.flatten(3).transpose(2, 3)
+            image_embeds = image_embeds.flatten(1, 2)
+        else:
+            p = self.patch_size
+            p_t = self.patch_size_t
+            image_embeds = image_embeds.permute(0, 1, 3, 4, 2)
+            image_embeds = image_embeds.reshape(
+                batch_size, num_frames // p_t, p_t, height // p, p, width // p, p, channels
+            )
+            image_embeds = image_embeds.permute(0, 1, 3, 5, 7, 2, 4, 6).flatten(4, 7).flatten(1, 3)
+            image_embeds = self.proj(image_embeds.to(proj_dtype)).to(input_dtype)
+
+        embeds = torch.cat([text_embeds, image_embeds], dim=1).contiguous()
+
+        if self.use_positional_embeddings or self.use_learned_positional_embeddings:
+            text_seq_length = text_embeds.shape[1]
+            num_image_patches = image_embeds.shape[1]
+
+            if self.use_learned_positional_embeddings:
+                image_pos = self.pos_embedding[
+                    :, self.max_text_seq_length:self.max_text_seq_length + num_image_patches
+                ].to(device=embeds.device, dtype=embeds.dtype)
+            else:
+                image_pos = get_3d_sincos_pos_embed(
+                    self.dim,
+                    (width // self.patch_size, height // self.patch_size),
+                    num_image_patches // ((height // self.patch_size) * (width // self.patch_size)),
+                    self.spatial_interpolation_scale,
+                    self.temporal_interpolation_scale,
+                    device=embeds.device,
+                ).reshape(1, num_image_patches, self.dim).to(dtype=embeds.dtype)
+
+            # Build joint: zeros for text + sincos for image
+            joint_pos = torch.zeros(1, text_seq_length + num_image_patches, self.dim, device=embeds.device, dtype=embeds.dtype)
+            joint_pos[:, text_seq_length:] = image_pos
+            embeds = embeds + joint_pos
+
+        return embeds
+
+
+class CogVideoXLayerNormZero(nn.Module):
+    def __init__(self, time_dim, dim, elementwise_affine=True, eps=1e-5, bias=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(time_dim, 6 * dim, bias=bias, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
+
+    def forward(self, hidden_states, encoder_hidden_states, temb):
+        shift, scale, gate, enc_shift, enc_scale, enc_gate = self.linear(self.silu(temb)).chunk(6, dim=1)
+        hidden_states = self.norm(hidden_states) * (1 + scale)[:, None, :] + shift[:, None, :]
+        encoder_hidden_states = self.norm(encoder_hidden_states) * (1 + enc_scale)[:, None, :] + enc_shift[:, None, :]
+        return hidden_states, encoder_hidden_states, gate[:, None, :], enc_gate[:, None, :]
+
+
+class CogVideoXAdaLayerNorm(nn.Module):
+    def __init__(self, time_dim, dim, elementwise_affine=True, eps=1e-5,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(time_dim, 2 * dim, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
+
+    def forward(self, x, temb):
+        temb = self.linear(self.silu(temb))
+        shift, scale = temb.chunk(2, dim=1)
+        x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
+        return x
+
+
+class CogVideoXBlock(nn.Module):
+    def __init__(self, dim, num_heads, head_dim, time_dim,
+                 eps=1e-5, ff_inner_dim=None, ff_bias=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = head_dim
+
+        self.norm1 = CogVideoXLayerNormZero(time_dim, dim, eps=eps, device=device, dtype=dtype, operations=operations)
+
+        # Self-attention (joint text + latent)
+        self.q = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.k = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.v = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.norm_q = operations.LayerNorm(head_dim, eps=1e-6, elementwise_affine=True, device=device, dtype=dtype)
+        self.norm_k = operations.LayerNorm(head_dim, eps=1e-6, elementwise_affine=True, device=device, dtype=dtype)
+        self.attn_out = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+
+        self.norm2 = CogVideoXLayerNormZero(time_dim, dim, eps=eps, device=device, dtype=dtype, operations=operations)
+
+        # Feed-forward (GELU approximate)
+        inner_dim = ff_inner_dim or dim * 4
+        self.ff_proj = operations.Linear(dim, inner_dim, bias=ff_bias, device=device, dtype=dtype)
+        self.ff_out = operations.Linear(inner_dim, dim, bias=ff_bias, device=device, dtype=dtype)
+
+    def forward(self, hidden_states, encoder_hidden_states, temb, image_rotary_emb=None, transformer_options=None):
+        if transformer_options is None:
+            transformer_options = {}
+        text_seq_length = encoder_hidden_states.size(1)
+
+        # Norm & modulate
+        norm_hidden, norm_encoder, gate_msa, enc_gate_msa = self.norm1(hidden_states, encoder_hidden_states, temb)
+
+        # Joint self-attention
+        qkv_input = torch.cat([norm_encoder, norm_hidden], dim=1)
+        b, s, _ = qkv_input.shape
+        n, d = self.num_heads, self.head_dim
+
+        q = self.q(qkv_input).view(b, s, n, d)
+        k = self.k(qkv_input).view(b, s, n, d)
+        v = self.v(qkv_input)
+
+        q = self.norm_q(q).view(b, s, n, d)
+        k = self.norm_k(k).view(b, s, n, d)
+
+        # Apply rotary embeddings to image tokens only (diffusers format: [B, heads, seq, head_dim])
+        if image_rotary_emb is not None:
+            q_img = q[:, text_seq_length:].transpose(1, 2)  # [B, heads, img_seq, head_dim]
+            k_img = k[:, text_seq_length:].transpose(1, 2)
+            q_img = apply_rotary_emb(q_img, image_rotary_emb)
+            k_img = apply_rotary_emb(k_img, image_rotary_emb)
+            q = torch.cat([q[:, :text_seq_length], q_img.transpose(1, 2)], dim=1)
+            k = torch.cat([k[:, :text_seq_length], k_img.transpose(1, 2)], dim=1)
+
+        attn_out = optimized_attention(
+            q.reshape(b, s, n * d),
+            k.reshape(b, s, n * d),
+            v,
+            heads=self.num_heads,
+            transformer_options=transformer_options,
+        )
+
+        attn_out = self.attn_out(attn_out)
+
+        attn_encoder, attn_hidden = attn_out.split([text_seq_length, s - text_seq_length], dim=1)
+
+        hidden_states = hidden_states + gate_msa * attn_hidden
+        encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder
+
+        # Norm & modulate for FF
+        norm_hidden, norm_encoder, gate_ff, enc_gate_ff = self.norm2(hidden_states, encoder_hidden_states, temb)
+
+        # Feed-forward (GELU on concatenated text + latent)
+        ff_input = torch.cat([norm_encoder, norm_hidden], dim=1)
+        ff_output = self.ff_out(F.gelu(self.ff_proj(ff_input), approximate="tanh"))
+
+        hidden_states = hidden_states + gate_ff * ff_output[:, text_seq_length:]
+        encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]
+
+        return hidden_states, encoder_hidden_states
+
+
+class CogVideoXTransformer3DModel(nn.Module):
+    def __init__(self,
+                 num_attention_heads=30,
+                 attention_head_dim=64,
+                 in_channels=16,
+                 out_channels=16,
+                 flip_sin_to_cos=True,
+                 freq_shift=0,
+                 time_embed_dim=512,
+                 ofs_embed_dim=None,
+                 text_embed_dim=4096,
+                 num_layers=30,
+                 dropout=0.0,
+                 attention_bias=True,
+                 sample_width=90,
+                 sample_height=60,
+                 sample_frames=49,
+                 patch_size=2,
+                 patch_size_t=None,
+                 temporal_compression_ratio=4,
+                 max_text_seq_length=226,
+                 spatial_interpolation_scale=1.875,
+                 temporal_interpolation_scale=1.0,
+                 use_rotary_positional_embeddings=False,
+                 use_learned_positional_embeddings=False,
+                 patch_bias=True,
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+        super().__init__()
+        self.dtype = dtype
+        dim = num_attention_heads * attention_head_dim
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
+        self.max_text_seq_length = max_text_seq_length
+        self.use_rotary_positional_embeddings = use_rotary_positional_embeddings
+
+        # 1. Patch embedding
+        self.patch_embed = CogVideoXPatchEmbed(
+            patch_size=patch_size,
+            patch_size_t=patch_size_t,
+            in_channels=in_channels,
+            dim=dim,
+            text_dim=text_embed_dim,
+            bias=patch_bias,
+            sample_width=sample_width,
+            sample_height=sample_height,
+            sample_frames=sample_frames,
+            temporal_compression_ratio=temporal_compression_ratio,
+            max_text_seq_length=max_text_seq_length,
+            spatial_interpolation_scale=spatial_interpolation_scale,
+            temporal_interpolation_scale=temporal_interpolation_scale,
+            use_positional_embeddings=not use_rotary_positional_embeddings,
+            use_learned_positional_embeddings=use_learned_positional_embeddings,
+            device=device, dtype=torch.float32, operations=operations,
+        )
+
+        # 2. Time embedding
+        self.time_proj_dim = dim
+        self.time_proj_flip = flip_sin_to_cos
+        self.time_proj_shift = freq_shift
+        self.time_embedding_linear_1 = operations.Linear(dim, time_embed_dim, device=device, dtype=dtype)
+        self.time_embedding_act = nn.SiLU()
+        self.time_embedding_linear_2 = operations.Linear(time_embed_dim, time_embed_dim, device=device, dtype=dtype)
+
+        # Optional OFS embedding (CogVideoX 1.5 I2V)
+        self.ofs_proj_dim = ofs_embed_dim
+        if ofs_embed_dim:
+            self.ofs_embedding_linear_1 = operations.Linear(ofs_embed_dim, ofs_embed_dim, device=device, dtype=dtype)
+            self.ofs_embedding_act = nn.SiLU()
+            self.ofs_embedding_linear_2 = operations.Linear(ofs_embed_dim, ofs_embed_dim, device=device, dtype=dtype)
+        else:
+            self.ofs_embedding_linear_1 = None
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList([
+            CogVideoXBlock(
+                dim=dim,
+                num_heads=num_attention_heads,
+                head_dim=attention_head_dim,
+                time_dim=time_embed_dim,
+                eps=1e-5,
+                device=device, dtype=dtype, operations=operations,
+            )
+            for _ in range(num_layers)
+        ])
+
+        self.norm_final = operations.LayerNorm(dim, eps=1e-5, elementwise_affine=True, device=device, dtype=dtype)
+
+        # 4. Output
+        self.norm_out = CogVideoXAdaLayerNorm(
+            time_dim=time_embed_dim, dim=dim, eps=1e-5,
+            device=device, dtype=dtype, operations=operations,
+        )
+
+        if patch_size_t is None:
+            output_dim = patch_size * patch_size * out_channels
+        else:
+            output_dim = patch_size * patch_size * patch_size_t * out_channels
+
+        self.proj_out = operations.Linear(dim, output_dim, device=device, dtype=dtype)
+
+        self.spatial_interpolation_scale = spatial_interpolation_scale
+        self.temporal_interpolation_scale = temporal_interpolation_scale
+        self.temporal_compression_ratio = temporal_compression_ratio
+
+    def forward(self, x, timestep, context, ofs=None, transformer_options=None, **kwargs):
+        if transformer_options is None:
+            transformer_options = {}
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, ofs, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, ofs=None, transformer_options=None, **kwargs):
+        if transformer_options is None:
+            transformer_options = {}
+        # ComfyUI passes [B, C, T, H, W]
+        batch_size, channels, t, h, w = x.shape
+
+        # Pad to patch size (temporal + spatial), same pattern as WAN
+        p_t = self.patch_size_t if self.patch_size_t is not None else 1
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (p_t, self.patch_size, self.patch_size))
+
+        # CogVideoX expects [B, T, C, H, W]
+        x = x.permute(0, 2, 1, 3, 4)
+        batch_size, num_frames, channels, height, width = x.shape
+
+        # Time embedding
+        t_emb = get_timestep_embedding(timestep, self.time_proj_dim, self.time_proj_flip, self.time_proj_shift)
+        t_emb = t_emb.to(dtype=x.dtype)
+        emb = self.time_embedding_linear_2(self.time_embedding_act(self.time_embedding_linear_1(t_emb)))
+
+        if self.ofs_embedding_linear_1 is not None and ofs is not None:
+            ofs_emb = get_timestep_embedding(ofs, self.ofs_proj_dim, self.time_proj_flip, self.time_proj_shift)
+            ofs_emb = ofs_emb.to(dtype=x.dtype)
+            ofs_emb = self.ofs_embedding_linear_2(self.ofs_embedding_act(self.ofs_embedding_linear_1(ofs_emb)))
+            emb = emb + ofs_emb
+
+        # Patch embedding
+        hidden_states = self.patch_embed(context, x)
+
+        text_seq_length = context.shape[1]
+        encoder_hidden_states = hidden_states[:, :text_seq_length]
+        hidden_states = hidden_states[:, text_seq_length:]
+
+        # Rotary embeddings (if used)
+        image_rotary_emb = None
+        if self.use_rotary_positional_embeddings:
+            post_patch_height = height // self.patch_size
+            post_patch_width = width // self.patch_size
+            if self.patch_size_t is None:
+                post_time = num_frames
+            else:
+                post_time = num_frames // self.patch_size_t
+            image_rotary_emb = self._get_rotary_emb(post_patch_height, post_patch_width, post_time, device=x.device)
+
+        # Transformer blocks
+        for i, block in enumerate(self.blocks):
+            hidden_states, encoder_hidden_states = block(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                temb=emb,
+                image_rotary_emb=image_rotary_emb,
+                transformer_options=transformer_options,
+            )
+
+        hidden_states = self.norm_final(hidden_states)
+
+        # Output projection
+        hidden_states = self.norm_out(hidden_states, temb=emb)
+        hidden_states = self.proj_out(hidden_states)
+
+        # Unpatchify
+        p = self.patch_size
+        p_t = self.patch_size_t
+
+        if p_t is None:
+            output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
+            output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
+        else:
+            output = hidden_states.reshape(
+                batch_size, (num_frames + p_t - 1) // p_t, height // p, width // p, -1, p_t, p, p
+            )
+            output = output.permute(0, 1, 5, 4, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(1, 2)
+
+        # Back to ComfyUI format [B, C, T, H, W] and crop padding
+        output = output.permute(0, 2, 1, 3, 4)[:, :, :t, :h, :w]
+        return output
+
+    def _get_rotary_emb(self, h, w, t, device):
+        """Compute CogVideoX 3D rotary positional embeddings.
+
+        For CogVideoX 1.5 (patch_size_t != None): uses "slice" mode — grid positions
+        are integer arange computed at max_size, then sliced to actual size.
+        For CogVideoX 1.0 (patch_size_t == None): uses "linspace" mode with crop coords
+        scaled by spatial_interpolation_scale.
+        """
+        d = self.attention_head_dim
+        dim_t = d // 4
+        dim_h = d // 8 * 3
+        dim_w = d // 8 * 3
+
+        if self.patch_size_t is not None:
+            # CogVideoX 1.5: "slice" mode — positions are simple integer indices
+            # Compute at max(sample_size, actual_size) then slice to actual
+            base_h = self.patch_embed.sample_height // self.patch_size
+            base_w = self.patch_embed.sample_width // self.patch_size
+            max_h = max(base_h, h)
+            max_w = max(base_w, w)
+
+            grid_h = torch.arange(max_h, device=device, dtype=torch.float32)
+            grid_w = torch.arange(max_w, device=device, dtype=torch.float32)
+            grid_t = torch.arange(t, device=device, dtype=torch.float32)
+        else:
+            # CogVideoX 1.0: "linspace" mode with interpolation scale
+            grid_h = torch.linspace(0, h - 1, h, device=device, dtype=torch.float32) * self.spatial_interpolation_scale
+            grid_w = torch.linspace(0, w - 1, w, device=device, dtype=torch.float32) * self.spatial_interpolation_scale
+            grid_t = torch.arange(t, device=device, dtype=torch.float32)
+
+        freqs_t = _get_1d_rotary_pos_embed(dim_t, grid_t)
+        freqs_h = _get_1d_rotary_pos_embed(dim_h, grid_h)
+        freqs_w = _get_1d_rotary_pos_embed(dim_w, grid_w)
+
+        t_cos, t_sin = freqs_t
+        h_cos, h_sin = freqs_h
+        w_cos, w_sin = freqs_w
+
+        # Slice to actual size (for "slice" mode where grids may be larger)
+        t_cos, t_sin = t_cos[:t], t_sin[:t]
+        h_cos, h_sin = h_cos[:h], h_sin[:h]
+        w_cos, w_sin = w_cos[:w], w_sin[:w]
+
+        # Broadcast and concatenate into [T*H*W, head_dim]
+        t_cos = t_cos[:, None, None, :].expand(-1, h, w, -1)
+        t_sin = t_sin[:, None, None, :].expand(-1, h, w, -1)
+        h_cos = h_cos[None, :, None, :].expand(t, -1, w, -1)
+        h_sin = h_sin[None, :, None, :].expand(t, -1, w, -1)
+        w_cos = w_cos[None, None, :, :].expand(t, h, -1, -1)
+        w_sin = w_sin[None, None, :, :].expand(t, h, -1, -1)
+
+        cos = torch.cat([t_cos, h_cos, w_cos], dim=-1).reshape(t * h * w, -1)
+        sin = torch.cat([t_sin, h_sin, w_sin], dim=-1).reshape(t * h * w, -1)
+        return (cos, sin)

ComfyUI/comfy/ldm/cogvideo/vae.py

@@ -0,0 +1,566 @@
+# CogVideoX VAE - ported to ComfyUI native ops
+# Architecture reference: diffusers AutoencoderKLCogVideoX
+# Style reference: comfy/ldm/wan/vae.py
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+
+class CausalConv3d(nn.Module):
+    """Causal 3D convolution with temporal padding.
+
+    Uses comfy.ops.Conv3d with autopad='causal_zero' fast path: when input has
+    a single temporal frame and no cache, the 3D conv weight is sliced to act
+    as a 2D conv, avoiding computation on zero-padded temporal dimensions.
+    """
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, pad_mode="constant"):
+        super().__init__()
+        if isinstance(kernel_size, int):
+            kernel_size = (kernel_size,) * 3
+
+        time_kernel, height_kernel, width_kernel = kernel_size
+        self.time_kernel_size = time_kernel
+        self.pad_mode = pad_mode
+
+        height_pad = (height_kernel - 1) // 2
+        width_pad = (width_kernel - 1) // 2
+        self.time_causal_padding = (width_pad, width_pad, height_pad, height_pad, time_kernel - 1, 0)
+
+        stride = stride if isinstance(stride, tuple) else (stride, 1, 1)
+        dilation = (dilation, 1, 1)
+        self.conv = ops.Conv3d(
+            in_channels, out_channels, kernel_size,
+            stride=stride, dilation=dilation,
+            padding=(0, height_pad, width_pad),
+        )
+
+    def forward(self, x, conv_cache=None):
+        if self.pad_mode == "replicate":
+            x = F.pad(x, self.time_causal_padding, mode="replicate")
+            conv_cache = None
+        else:
+            kernel_t = self.time_kernel_size
+            if kernel_t > 1:
+                if conv_cache is None and x.shape[2] == 1:
+                    # Fast path: single frame, no cache. All temporal padding
+                    # frames are copies of the input (replicate-style), so the
+                    # 3D conv reduces to a 2D conv with summed temporal kernel.
+                    w = comfy.ops.cast_to_input(self.conv.weight, x)
+                    b = comfy.ops.cast_to_input(self.conv.bias, x) if self.conv.bias is not None else None
+                    w2d = w.sum(dim=2, keepdim=True)
+                    out = F.conv3d(x, w2d, b,
+                                   self.conv.stride, self.conv.padding,
+                                   self.conv.dilation, self.conv.groups)
+                    return out, None
+                cached = [conv_cache] if conv_cache is not None else [x[:, :, :1]] * (kernel_t - 1)
+                x = torch.cat(cached + [x], dim=2)
+            conv_cache = x[:, :, -self.time_kernel_size + 1:].clone() if self.time_kernel_size > 1 else None
+
+        out = self.conv(x)
+        return out, conv_cache
+
+
+def _interpolate_zq(zq, target_size):
+    """Interpolate latent z to target (T, H, W), matching CogVideoX's first-frame-special handling."""
+    t = target_size[0]
+    if t > 1 and t % 2 == 1:
+        z_first = F.interpolate(zq[:, :, :1], size=(1, target_size[1], target_size[2]))
+        z_rest = F.interpolate(zq[:, :, 1:], size=(t - 1, target_size[1], target_size[2]))
+        return torch.cat([z_first, z_rest], dim=2)
+    return F.interpolate(zq, size=target_size)
+
+
+class SpatialNorm3D(nn.Module):
+    """Spatially conditioned normalization."""
+    def __init__(self, f_channels, zq_channels, groups=32):
+        super().__init__()
+        self.norm_layer = ops.GroupNorm(num_channels=f_channels, num_groups=groups, eps=1e-6, affine=True)
+        self.conv_y = CausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+        self.conv_b = CausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+
+    def forward(self, f, zq, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        if zq.shape[-3:] != f.shape[-3:]:
+            zq = _interpolate_zq(zq, f.shape[-3:])
+
+        conv_y, new_cache["conv_y"] = self.conv_y(zq, conv_cache=conv_cache.get("conv_y"))
+        conv_b, new_cache["conv_b"] = self.conv_b(zq, conv_cache=conv_cache.get("conv_b"))
+
+        return self.norm_layer(f) * conv_y + conv_b, new_cache
+
+
+class ResnetBlock3D(nn.Module):
+    """3D ResNet block with optional spatial norm."""
+    def __init__(self, in_channels, out_channels=None, temb_channels=512, groups=32,
+                 eps=1e-6, act_fn="silu", spatial_norm_dim=None, pad_mode="first"):
+        super().__init__()
+        out_channels = out_channels or in_channels
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.spatial_norm_dim = spatial_norm_dim
+
+        if act_fn == "silu":
+            self.nonlinearity = nn.SiLU()
+        elif act_fn == "swish":
+            self.nonlinearity = nn.SiLU()
+        else:
+            self.nonlinearity = nn.SiLU()
+
+        if spatial_norm_dim is None:
+            self.norm1 = ops.GroupNorm(num_channels=in_channels, num_groups=groups, eps=eps)
+            self.norm2 = ops.GroupNorm(num_channels=out_channels, num_groups=groups, eps=eps)
+        else:
+            self.norm1 = SpatialNorm3D(in_channels, spatial_norm_dim, groups=groups)
+            self.norm2 = SpatialNorm3D(out_channels, spatial_norm_dim, groups=groups)
+
+        self.conv1 = CausalConv3d(in_channels, out_channels, kernel_size=3, pad_mode=pad_mode)
+
+        if temb_channels > 0:
+            self.temb_proj = ops.Linear(temb_channels, out_channels)
+
+        self.conv2 = CausalConv3d(out_channels, out_channels, kernel_size=3, pad_mode=pad_mode)
+
+        if in_channels != out_channels:
+            self.conv_shortcut = ops.Conv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+        else:
+            self.conv_shortcut = None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        residual = x
+
+        if zq is not None:
+            x, new_cache["norm1"] = self.norm1(x, zq, conv_cache=conv_cache.get("norm1"))
+        else:
+            x = self.norm1(x)
+
+        x = self.nonlinearity(x)
+        x, new_cache["conv1"] = self.conv1(x, conv_cache=conv_cache.get("conv1"))
+
+        if temb is not None and hasattr(self, "temb_proj"):
+            x = x + self.temb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+
+        if zq is not None:
+            x, new_cache["norm2"] = self.norm2(x, zq, conv_cache=conv_cache.get("norm2"))
+        else:
+            x = self.norm2(x)
+
+        x = self.nonlinearity(x)
+        x, new_cache["conv2"] = self.conv2(x, conv_cache=conv_cache.get("conv2"))
+
+        if self.conv_shortcut is not None:
+            residual = self.conv_shortcut(residual)
+
+        return x + residual, new_cache
+
+
+class Downsample3D(nn.Module):
+    """3D downsampling with optional temporal compression."""
+    def __init__(self, in_channels, out_channels, kernel_size=3, stride=2, padding=0, compress_time=False):
+        super().__init__()
+        self.conv = ops.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x):
+        if self.compress_time:
+            b, c, t, h, w = x.shape
+            x = x.permute(0, 3, 4, 1, 2).reshape(b * h * w, c, t)
+            if t % 2 == 1:
+                x_first, x_rest = x[..., 0], x[..., 1:]
+                if x_rest.shape[-1] > 0:
+                    x_rest = F.avg_pool1d(x_rest, kernel_size=2, stride=2)
+                x = torch.cat([x_first[..., None], x_rest], dim=-1)
+                x = x.reshape(b, h, w, c, x.shape[-1]).permute(0, 3, 4, 1, 2)
+            else:
+                x = F.avg_pool1d(x, kernel_size=2, stride=2)
+                x = x.reshape(b, h, w, c, x.shape[-1]).permute(0, 3, 4, 1, 2)
+
+        pad = (0, 1, 0, 1)
+        x = F.pad(x, pad, mode="constant", value=0)
+        b, c, t, h, w = x.shape
+        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        x = self.conv(x)
+        x = x.reshape(b, t, x.shape[1], x.shape[2], x.shape[3]).permute(0, 2, 1, 3, 4)
+        return x
+
+
+class Upsample3D(nn.Module):
+    """3D upsampling with optional temporal decompression."""
+    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, padding=1, compress_time=False):
+        super().__init__()
+        self.conv = ops.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x):
+        if self.compress_time:
+            if x.shape[2] > 1 and x.shape[2] % 2 == 1:
+                x_first, x_rest = x[:, :, 0], x[:, :, 1:]
+                x_first = F.interpolate(x_first, scale_factor=2.0)
+                x_rest = F.interpolate(x_rest, scale_factor=2.0)
+                x = torch.cat([x_first[:, :, None, :, :], x_rest], dim=2)
+            elif x.shape[2] > 1:
+                x = F.interpolate(x, scale_factor=2.0)
+            else:
+                x = x.squeeze(2)
+                x = F.interpolate(x, scale_factor=2.0)
+                x = x[:, :, None, :, :]
+        else:
+            b, c, t, h, w = x.shape
+            x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+            x = F.interpolate(x, scale_factor=2.0)
+            x = x.reshape(b, t, c, *x.shape[2:]).permute(0, 2, 1, 3, 4)
+
+        b, c, t, h, w = x.shape
+        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        x = self.conv(x)
+        x = x.reshape(b, t, *x.shape[1:]).permute(0, 2, 1, 3, 4)
+        return x
+
+
+class DownBlock3D(nn.Module):
+    def __init__(self, in_channels, out_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, add_downsample=True,
+                 compress_time=False, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels if i == 0 else out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels,
+                groups=groups, eps=eps, act_fn=act_fn, pad_mode=pad_mode,
+            )
+            for i in range(num_layers)
+        ])
+        self.downsamplers = nn.ModuleList([Downsample3D(out_channels, out_channels, compress_time=compress_time)]) if add_downsample else None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        if self.downsamplers is not None:
+            for ds in self.downsamplers:
+                x = ds(x)
+        return x, new_cache
+
+
+class MidBlock3D(nn.Module):
+    def __init__(self, in_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, spatial_norm_dim=None, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels, out_channels=in_channels,
+                temb_channels=temb_channels, groups=groups, eps=eps,
+                act_fn=act_fn, spatial_norm_dim=spatial_norm_dim, pad_mode=pad_mode,
+            )
+            for _ in range(num_layers)
+        ])
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        return x, new_cache
+
+
+class UpBlock3D(nn.Module):
+    def __init__(self, in_channels, out_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, spatial_norm_dim=16,
+                 add_upsample=True, compress_time=False, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels if i == 0 else out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels, groups=groups, eps=eps,
+                act_fn=act_fn, spatial_norm_dim=spatial_norm_dim, pad_mode=pad_mode,
+            )
+            for i in range(num_layers)
+        ])
+        self.upsamplers = nn.ModuleList([Upsample3D(out_channels, out_channels, compress_time=compress_time)]) if add_upsample else None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        if self.upsamplers is not None:
+            for us in self.upsamplers:
+                x = us(x)
+        return x, new_cache
+
+
+class Encoder3D(nn.Module):
+    def __init__(self, in_channels=3, out_channels=16,
+                 block_out_channels=(128, 256, 256, 512),
+                 layers_per_block=3, act_fn="silu",
+                 eps=1e-6, groups=32, pad_mode="first",
+                 temporal_compression_ratio=4):
+        super().__init__()
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        self.conv_in = CausalConv3d(in_channels, block_out_channels[0], kernel_size=3, pad_mode=pad_mode)
+
+        self.down_blocks = nn.ModuleList()
+        output_channel = block_out_channels[0]
+        for i in range(len(block_out_channels)):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            self.down_blocks.append(DownBlock3D(
+                in_channels=input_channel, out_channels=output_channel,
+                temb_channels=0, num_layers=layers_per_block,
+                eps=eps, act_fn=act_fn, groups=groups,
+                add_downsample=not is_final, compress_time=compress_time,
+            ))
+
+        self.mid_block = MidBlock3D(
+            in_channels=block_out_channels[-1], temb_channels=0,
+            num_layers=2, eps=eps, act_fn=act_fn, groups=groups, pad_mode=pad_mode,
+        )
+
+        self.norm_out = ops.GroupNorm(groups, block_out_channels[-1], eps=1e-6)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CausalConv3d(block_out_channels[-1], 2 * out_channels, kernel_size=3, pad_mode=pad_mode)
+
+    def forward(self, x, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        x, new_cache["conv_in"] = self.conv_in(x, conv_cache=conv_cache.get("conv_in"))
+
+        for i, block in enumerate(self.down_blocks):
+            key = f"down_block_{i}"
+            x, new_cache[key] = block(x, None, None, conv_cache.get(key))
+
+        x, new_cache["mid_block"] = self.mid_block(x, None, None, conv_cache=conv_cache.get("mid_block"))
+
+        x = self.norm_out(x)
+        x = self.conv_act(x)
+        x, new_cache["conv_out"] = self.conv_out(x, conv_cache=conv_cache.get("conv_out"))
+
+        return x, new_cache
+
+
+class Decoder3D(nn.Module):
+    def __init__(self, in_channels=16, out_channels=3,
+                 block_out_channels=(128, 256, 256, 512),
+                 layers_per_block=3, act_fn="silu",
+                 eps=1e-6, groups=32, pad_mode="first",
+                 temporal_compression_ratio=4):
+        super().__init__()
+        reversed_channels = list(reversed(block_out_channels))
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        self.conv_in = CausalConv3d(in_channels, reversed_channels[0], kernel_size=3, pad_mode=pad_mode)
+
+        self.mid_block = MidBlock3D(
+            in_channels=reversed_channels[0], temb_channels=0,
+            num_layers=2, eps=eps, act_fn=act_fn, groups=groups,
+            spatial_norm_dim=in_channels, pad_mode=pad_mode,
+        )
+
+        self.up_blocks = nn.ModuleList()
+        output_channel = reversed_channels[0]
+        for i in range(len(block_out_channels)):
+            prev_channel = output_channel
+            output_channel = reversed_channels[i]
+            is_final = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            self.up_blocks.append(UpBlock3D(
+                in_channels=prev_channel, out_channels=output_channel,
+                temb_channels=0, num_layers=layers_per_block + 1,
+                eps=eps, act_fn=act_fn, groups=groups,
+                spatial_norm_dim=in_channels,
+                add_upsample=not is_final, compress_time=compress_time,
+            ))
+
+        self.norm_out = SpatialNorm3D(reversed_channels[-1], in_channels, groups=groups)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CausalConv3d(reversed_channels[-1], out_channels, kernel_size=3, pad_mode=pad_mode)
+
+    def forward(self, sample, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        x, new_cache["conv_in"] = self.conv_in(sample, conv_cache=conv_cache.get("conv_in"))
+
+        x, new_cache["mid_block"] = self.mid_block(x, None, sample, conv_cache=conv_cache.get("mid_block"))
+
+        for i, block in enumerate(self.up_blocks):
+            key = f"up_block_{i}"
+            x, new_cache[key] = block(x, None, sample, conv_cache=conv_cache.get(key))
+
+        x, new_cache["norm_out"] = self.norm_out(x, sample, conv_cache=conv_cache.get("norm_out"))
+        x = self.conv_act(x)
+        x, new_cache["conv_out"] = self.conv_out(x, conv_cache=conv_cache.get("conv_out"))
+
+        return x, new_cache
+
+
+
+class AutoencoderKLCogVideoX(nn.Module):
+    """CogVideoX VAE. Spatial tiling/slicing handled by ComfyUI's VAE wrapper.
+
+    Uses rolling temporal decode: conv_in + mid_block + temporal up_blocks run
+    on the full (low-res) tensor, then the expensive spatial-only up_blocks +
+    norm_out + conv_out are processed in small temporal chunks with conv_cache
+    carrying causal state between chunks. This keeps peak VRAM proportional to
+    chunk_size rather than total frame count.
+    """
+
+    def __init__(self,
+                 in_channels=3, out_channels=3,
+                 block_out_channels=(128, 256, 256, 512),
+                 latent_channels=16, layers_per_block=3,
+                 act_fn="silu", eps=1e-6, groups=32,
+                 temporal_compression_ratio=4,
+                 ):
+        super().__init__()
+        self.latent_channels = latent_channels
+        self.temporal_compression_ratio = temporal_compression_ratio
+
+        self.encoder = Encoder3D(
+            in_channels=in_channels, out_channels=latent_channels,
+            block_out_channels=block_out_channels, layers_per_block=layers_per_block,
+            act_fn=act_fn, eps=eps, groups=groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+        self.decoder = Decoder3D(
+            in_channels=latent_channels, out_channels=out_channels,
+            block_out_channels=block_out_channels, layers_per_block=layers_per_block,
+            act_fn=act_fn, eps=eps, groups=groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+
+        self.num_latent_frames_batch_size = 2
+        self.num_sample_frames_batch_size = 8
+
+    def encode(self, x):
+        t = x.shape[2]
+        frame_batch = self.num_sample_frames_batch_size
+        remainder = t % frame_batch
+        conv_cache = None
+        enc = []
+
+        # Process remainder frames first so only the first chunk can have an
+        # odd temporal dimension — where Downsample3D's first-frame-special
+        # handling in temporal compression is actually correct.
+        if remainder > 0:
+            chunk, conv_cache = self.encoder(x[:, :, :remainder], conv_cache=conv_cache)
+            enc.append(chunk.to(x.device))
+
+        for start in range(remainder, t, frame_batch):
+            chunk, conv_cache = self.encoder(x[:, :, start:start + frame_batch], conv_cache=conv_cache)
+            enc.append(chunk.to(x.device))
+
+        enc = torch.cat(enc, dim=2)
+        mean, _ = enc.chunk(2, dim=1)
+        return mean
+
+    def decode(self, z):
+        return self._decode_rolling(z)
+
+    def _decode_batched(self, z):
+        """Original batched decode - processes 2 latent frames through full decoder."""
+        t = z.shape[2]
+        frame_batch = self.num_latent_frames_batch_size
+        num_batches = max(t // frame_batch, 1)
+        conv_cache = None
+        dec = []
+        for i in range(num_batches):
+            remaining = t % frame_batch
+            start = frame_batch * i + (0 if i == 0 else remaining)
+            end = frame_batch * (i + 1) + remaining
+            chunk, conv_cache = self.decoder(z[:, :, start:end], conv_cache=conv_cache)
+            dec.append(chunk.cpu())
+        return torch.cat(dec, dim=2).to(z.device)
+
+    def _decode_rolling(self, z):
+        """Rolling decode - processes low-res layers on full tensor, then rolls
+        through expensive high-res layers in temporal chunks."""
+        decoder = self.decoder
+        device = z.device
+
+        # Determine which up_blocks have temporal upsample vs spatial-only.
+        # Temporal up_blocks are cheap (low res), spatial-only are expensive.
+        temporal_compress_level = int(np.log2(self.temporal_compression_ratio))
+        split_at = temporal_compress_level  # first N up_blocks do temporal upsample
+
+        # Phase 1: conv_in + mid_block + temporal up_blocks on full tensor (low/medium res)
+        x, _ = decoder.conv_in(z)
+        x, _ = decoder.mid_block(x, None, z)
+
+        for i in range(split_at):
+            x, _ = decoder.up_blocks[i](x, None, z)
+
+        # Phase 2: remaining spatial-only up_blocks + norm_out + conv_out in temporal chunks
+        remaining_blocks = list(range(split_at, len(decoder.up_blocks)))
+        chunk_size = 4  # pixel frames per chunk through high-res layers
+        t_expanded = x.shape[2]
+
+        if t_expanded <= chunk_size or len(remaining_blocks) == 0:
+            # Small enough to process in one go
+            for i in remaining_blocks:
+                x, _ = decoder.up_blocks[i](x, None, z)
+            x, _ = decoder.norm_out(x, z)
+            x = decoder.conv_act(x)
+            x, _ = decoder.conv_out(x)
+            return x
+
+        # Expand z temporally once to match Phase 2's time dimension.
+        # z stays at latent spatial resolution so this is small (~16 MB vs ~1.3 GB
+        # for the old approach of pre-interpolating to every pixel resolution).
+        z_time_expanded = _interpolate_zq(z, (t_expanded, z.shape[3], z.shape[4]))
+
+        # Process in temporal chunks, interpolating spatially per-chunk to avoid
+        # allocating full [B, C, t_expanded, H, W] tensors at each resolution.
+        dec_out = []
+        conv_caches = {}
+
+        for chunk_start in range(0, t_expanded, chunk_size):
+            chunk_end = min(chunk_start + chunk_size, t_expanded)
+            x_chunk = x[:, :, chunk_start:chunk_end]
+            z_t_chunk = z_time_expanded[:, :, chunk_start:chunk_end]
+            z_spatial_cache = {}
+
+            for i in remaining_blocks:
+                block = decoder.up_blocks[i]
+                cache_key = f"up_block_{i}"
+                hw_key = (x_chunk.shape[3], x_chunk.shape[4])
+                if hw_key not in z_spatial_cache:
+                    if z_t_chunk.shape[3] == hw_key[0] and z_t_chunk.shape[4] == hw_key[1]:
+                        z_spatial_cache[hw_key] = z_t_chunk
+                    else:
+                        z_spatial_cache[hw_key] = F.interpolate(z_t_chunk, size=(z_t_chunk.shape[2], hw_key[0], hw_key[1]))
+                x_chunk, new_cache = block(x_chunk, None, z_spatial_cache[hw_key], conv_cache=conv_caches.get(cache_key))
+                conv_caches[cache_key] = new_cache
+
+            hw_key = (x_chunk.shape[3], x_chunk.shape[4])
+            if hw_key not in z_spatial_cache:
+                z_spatial_cache[hw_key] = F.interpolate(z_t_chunk, size=(z_t_chunk.shape[2], hw_key[0], hw_key[1]))
+            x_chunk, new_cache = decoder.norm_out(x_chunk, z_spatial_cache[hw_key], conv_cache=conv_caches.get("norm_out"))
+            conv_caches["norm_out"] = new_cache
+            x_chunk = decoder.conv_act(x_chunk)
+            x_chunk, new_cache = decoder.conv_out(x_chunk, conv_cache=conv_caches.get("conv_out"))
+            conv_caches["conv_out"] = new_cache
+
+            dec_out.append(x_chunk.cpu())
+            del z_spatial_cache
+
+        del x, z_time_expanded
+        return torch.cat(dec_out, dim=2).to(device)

ComfyUI/comfy/ldm/colormap.py

@@ -0,0 +1,25 @@
+"""Colormap utilities for depth and geometry visualisation."""
+
+from __future__ import annotations
+
+import torch
+
+
+def turbo(x: torch.Tensor) -> torch.Tensor:
+    """Anton Mikhailov polynomial approximation of the Turbo colormap.
+
+    Args:
+        x: Float tensor with values in [0, 1].
+
+    Returns:
+        RGB tensor of the same shape as ``x`` with a trailing size-3 dimension.
+    """
+    x = x.clamp(0.0, 1.0)
+    x2 = x * x
+    x3 = x2 * x
+    x4 = x2 * x2
+    x5 = x4 * x
+    r = 0.13572138 + 4.61539260*x - 42.66032258*x2 + 132.13108234*x3 - 152.94239396*x4 + 59.28637943*x5
+    g = 0.09140261 + 2.19418839*x + 4.84296658*x2 - 14.18503333*x3 +   4.27729857*x4 +  2.82956604*x5
+    b = 0.10667330 + 12.64194608*x - 60.58204836*x2 + 110.36276771*x3 - 89.90310912*x4 + 27.34824973*x5
+    return torch.stack([r, g, b], dim=-1).clamp(0.0, 1.0)

ComfyUI/comfy/ldm/common_dit.py

@@ -0,0 +1,16 @@
+import torch
+import comfy.rmsnorm
+
+
+def pad_to_patch_size(img, patch_size=(2, 2), padding_mode="circular"):
+    if padding_mode == "circular" and (torch.jit.is_tracing() or torch.jit.is_scripting()):
+        padding_mode = "reflect"
+
+    pad = ()
+    for i in range(img.ndim - 2):
+        pad = (0, (patch_size[i] - img.shape[i + 2] % patch_size[i]) % patch_size[i]) + pad
+
+    return torch.nn.functional.pad(img, pad, mode=padding_mode)
+
+
+rms_norm = comfy.rmsnorm.rms_norm

ComfyUI/comfy/ldm/cosmos/blocks.py

@@ -0,0 +1,805 @@
+# 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.
+
+import math
+from typing import Optional
+import logging
+
+import numpy as np
+import torch
+from einops import rearrange, repeat
+from einops.layers.torch import Rearrange
+from torch import nn
+
+from comfy.ldm.modules.attention import optimized_attention
+
+
+def get_normalization(name: str, channels: int, weight_args={}, operations=None):
+    if name == "I":
+        return nn.Identity()
+    elif name == "R":
+        return operations.RMSNorm(channels, elementwise_affine=True, eps=1e-6, **weight_args)
+    else:
+        raise ValueError(f"Normalization {name} not found")
+
+
+class BaseAttentionOp(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+
+class Attention(nn.Module):
+    """
+    Generalized attention impl.
+
+    Allowing for both self-attention and cross-attention configurations depending on whether a `context_dim` is provided.
+    If `context_dim` is None, self-attention is assumed.
+
+    Parameters:
+        query_dim (int): Dimension of each query vector.
+        context_dim (int, optional): Dimension of each context vector. If None, self-attention is assumed.
+        heads (int, optional): Number of attention heads. Defaults to 8.
+        dim_head (int, optional): Dimension of each head. Defaults to 64.
+        dropout (float, optional): Dropout rate applied to the output of the attention block. Defaults to 0.0.
+        attn_op (BaseAttentionOp, optional): Custom attention operation to be used instead of the default.
+        qkv_bias (bool, optional): If True, adds a learnable bias to query, key, and value projections. Defaults to False.
+        out_bias (bool, optional): If True, adds a learnable bias to the output projection. Defaults to False.
+        qkv_norm (str, optional): A string representing normalization strategies for query, key, and value projections.
+                                  Defaults to "SSI".
+        qkv_norm_mode (str, optional): A string representing normalization mode for query, key, and value projections.
+                                        Defaults to 'per_head'. Only support 'per_head'.
+
+    Examples:
+        >>> attn = Attention(query_dim=128, context_dim=256, heads=4, dim_head=32, dropout=0.1)
+        >>> query = torch.randn(10, 128)  # Batch size of 10
+        >>> context = torch.randn(10, 256)  # Batch size of 10
+        >>> output = attn(query, context)  # Perform the attention operation
+
+    Note:
+        https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
+    """
+
+    def __init__(
+        self,
+        query_dim: int,
+        context_dim=None,
+        heads=8,
+        dim_head=64,
+        dropout=0.0,
+        attn_op: Optional[BaseAttentionOp] = None,
+        qkv_bias: bool = False,
+        out_bias: bool = False,
+        qkv_norm: str = "SSI",
+        qkv_norm_mode: str = "per_head",
+        backend: str = "transformer_engine",
+        qkv_format: str = "bshd",
+        weight_args={},
+        operations=None,
+    ) -> None:
+        super().__init__()
+
+        self.is_selfattn = context_dim is None  # self attention
+
+        inner_dim = dim_head * heads
+        context_dim = query_dim if context_dim is None else context_dim
+
+        self.heads = heads
+        self.dim_head = dim_head
+        self.qkv_norm_mode = qkv_norm_mode
+        self.qkv_format = qkv_format
+
+        if self.qkv_norm_mode == "per_head":
+            norm_dim = dim_head
+        else:
+            raise ValueError(f"Normalization mode {self.qkv_norm_mode} not found, only support 'per_head'")
+
+        self.backend = backend
+
+        self.to_q = nn.Sequential(
+            operations.Linear(query_dim, inner_dim, bias=qkv_bias, **weight_args),
+            get_normalization(qkv_norm[0], norm_dim, weight_args=weight_args, operations=operations),
+        )
+        self.to_k = nn.Sequential(
+            operations.Linear(context_dim, inner_dim, bias=qkv_bias, **weight_args),
+            get_normalization(qkv_norm[1], norm_dim, weight_args=weight_args, operations=operations),
+        )
+        self.to_v = nn.Sequential(
+            operations.Linear(context_dim, inner_dim, bias=qkv_bias, **weight_args),
+            get_normalization(qkv_norm[2], norm_dim, weight_args=weight_args, operations=operations),
+        )
+
+        self.to_out = nn.Sequential(
+            operations.Linear(inner_dim, query_dim, bias=out_bias, **weight_args),
+            nn.Dropout(dropout),
+        )
+
+    def cal_qkv(
+        self, x, context=None, mask=None, rope_emb=None, **kwargs
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        del kwargs
+
+
+        """
+        self.to_q, self.to_k, self.to_v are nn.Sequential with projection + normalization layers.
+        Before 07/24/2024, these modules normalize across all heads.
+        After 07/24/2024, to support tensor parallelism and follow the common practice in the community,
+        we support to normalize per head.
+        To keep the checkpoint copatibility with the previous code,
+        we keep the nn.Sequential but call the projection and the normalization layers separately.
+        We use a flag `self.qkv_norm_mode` to control the normalization behavior.
+        The default value of `self.qkv_norm_mode` is "per_head", which means we normalize per head.
+        """
+        if self.qkv_norm_mode == "per_head":
+            q = self.to_q[0](x)
+            context = x if context is None else context
+            k = self.to_k[0](context)
+            v = self.to_v[0](context)
+            q, k, v = map(
+                lambda t: rearrange(t, "s b (n c) -> b n s c", n=self.heads, c=self.dim_head),
+                (q, k, v),
+            )
+        else:
+            raise ValueError(f"Normalization mode {self.qkv_norm_mode} not found, only support 'per_head'")
+
+        q = self.to_q[1](q)
+        k = self.to_k[1](k)
+        v = self.to_v[1](v)
+        if self.is_selfattn and rope_emb is not None:  # only apply to self-attention!
+            # apply_rotary_pos_emb inlined
+            q_shape = q.shape
+            q = q.reshape(*q.shape[:-1], 2, -1).movedim(-2, -1).unsqueeze(-2)
+            q = rope_emb[..., 0] * q[..., 0] + rope_emb[..., 1] * q[..., 1]
+            q = q.movedim(-1, -2).reshape(*q_shape).to(x.dtype)
+
+            # apply_rotary_pos_emb inlined
+            k_shape = k.shape
+            k = k.reshape(*k.shape[:-1], 2, -1).movedim(-2, -1).unsqueeze(-2)
+            k = rope_emb[..., 0] * k[..., 0] + rope_emb[..., 1] * k[..., 1]
+            k = k.movedim(-1, -2).reshape(*k_shape).to(x.dtype)
+        return q, k, v
+
+    def forward(
+        self,
+        x,
+        context=None,
+        mask=None,
+        rope_emb=None,
+        transformer_options={},
+        **kwargs,
+    ):
+        """
+        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.cal_qkv(x, context, mask, rope_emb=rope_emb, **kwargs)
+        out = optimized_attention(q, k, v, self.heads, skip_reshape=True, mask=mask, skip_output_reshape=True, transformer_options=transformer_options)
+        del q, k, v
+        out = rearrange(out, " b n s c -> s b (n c)")
+        return self.to_out(out)
+
+
+class FeedForward(nn.Module):
+    """
+    Transformer FFN with optional gating
+
+    Parameters:
+        d_model (int): Dimensionality of input features.
+        d_ff (int): Dimensionality of the hidden layer.
+        dropout (float, optional): Dropout rate applied after the activation function. Defaults to 0.1.
+        activation (callable, optional): The activation function applied after the first linear layer.
+                                         Defaults to nn.ReLU().
+        is_gated (bool, optional): If set to True, incorporates gating mechanism to the feed-forward layer.
+                                   Defaults to False.
+        bias (bool, optional): If set to True, adds a bias to the linear layers. Defaults to True.
+
+    Example:
+        >>> ff = FeedForward(d_model=512, d_ff=2048)
+        >>> x = torch.randn(64, 10, 512)  # Example input tensor
+        >>> output = ff(x)
+        >>> print(output.shape)  # Expected shape: (64, 10, 512)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        d_ff: int,
+        dropout: float = 0.1,
+        activation=nn.ReLU(),
+        is_gated: bool = False,
+        bias: bool = False,
+        weight_args={},
+        operations=None,
+    ) -> None:
+        super().__init__()
+
+        self.layer1 = operations.Linear(d_model, d_ff, bias=bias, **weight_args)
+        self.layer2 = operations.Linear(d_ff, d_model, bias=bias, **weight_args)
+
+        self.dropout = nn.Dropout(dropout)
+        self.activation = activation
+        self.is_gated = is_gated
+        if is_gated:
+            self.linear_gate = operations.Linear(d_model, d_ff, bias=False, **weight_args)
+
+    def forward(self, x: torch.Tensor):
+        g = self.activation(self.layer1(x))
+        if self.is_gated:
+            x = g * self.linear_gate(x)
+        else:
+            x = g
+        assert self.dropout.p == 0.0, "we skip dropout"
+        return self.layer2(x)
+
+
+class GPT2FeedForward(FeedForward):
+    def __init__(self, d_model: int, d_ff: int, dropout: float = 0.1, bias: bool = False, weight_args={}, operations=None):
+        super().__init__(
+            d_model=d_model,
+            d_ff=d_ff,
+            dropout=dropout,
+            activation=nn.GELU(),
+            is_gated=False,
+            bias=bias,
+            weight_args=weight_args,
+            operations=operations,
+        )
+
+    def forward(self, x: torch.Tensor):
+        assert self.dropout.p == 0.0, "we skip dropout"
+
+        x = self.layer1(x)
+        x = self.activation(x)
+        x = self.layer2(x)
+
+        return x
+
+
+def modulate(x, shift, scale):
+    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+
+class Timesteps(nn.Module):
+    def __init__(self, num_channels):
+        super().__init__()
+        self.num_channels = num_channels
+
+    def forward(self, timesteps):
+        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 emb
+
+
+class TimestepEmbedding(nn.Module):
+    def __init__(self, in_features: int, out_features: int, use_adaln_lora: bool = False, weight_args={}, 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.linear_1 = operations.Linear(in_features, out_features, bias=not use_adaln_lora, **weight_args)
+        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, **weight_args)
+        else:
+            self.linear_2 = operations.Linear(out_features, out_features, bias=True, **weight_args)
+
+    def forward(self, sample: torch.Tensor) -> torch.Tensor:
+        emb = self.linear_1(sample)
+        emb = self.activation(emb)
+        emb = self.linear_2(emb)
+
+        if self.use_adaln_lora:
+            adaln_lora_B_3D = emb
+            emb_B_D = sample
+        else:
+            emb_B_D = emb
+            adaln_lora_B_3D = None
+
+        return emb_B_D, adaln_lora_B_3D
+
+
+class FourierFeatures(nn.Module):
+    """
+    Implements a layer that generates Fourier features from input tensors, based on randomly sampled
+    frequencies and phases. This can help in learning high-frequency functions in low-dimensional problems.
+
+    [B] -> [B, D]
+
+    Parameters:
+        num_channels (int): The number of Fourier features to generate.
+        bandwidth (float, optional): The scaling factor for the frequency of the Fourier features. Defaults to 1.
+        normalize (bool, optional): If set to True, the outputs are scaled by sqrt(2), usually to normalize
+                                    the variance of the features. Defaults to False.
+
+    Example:
+        >>> layer = FourierFeatures(num_channels=256, bandwidth=0.5, normalize=True)
+        >>> x = torch.randn(10, 256)  # Example input tensor
+        >>> output = layer(x)
+        >>> print(output.shape)  # Expected shape: (10, 256)
+    """
+
+    def __init__(self, num_channels, bandwidth=1, normalize=False):
+        super().__init__()
+        self.register_buffer("freqs", 2 * np.pi * bandwidth * torch.randn(num_channels), persistent=True)
+        self.register_buffer("phases", 2 * np.pi * torch.rand(num_channels), persistent=True)
+        self.gain = np.sqrt(2) if normalize else 1
+
+    def forward(self, x, gain: float = 1.0):
+        """
+        Apply the Fourier feature transformation to the input tensor.
+
+        Args:
+            x (torch.Tensor): The input tensor.
+            gain (float, optional): An additional gain factor applied during the forward pass. Defaults to 1.
+
+        Returns:
+            torch.Tensor: The transformed tensor, with Fourier features applied.
+        """
+        in_dtype = x.dtype
+        x = x.to(torch.float32).ger(self.freqs.to(torch.float32)).add(self.phases.to(torch.float32))
+        x = x.cos().mul(self.gain * gain).to(in_dtype)
+        return x
+
+
+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,
+        temporal_patch_size,
+        in_channels=3,
+        out_channels=768,
+        bias=True,
+        weight_args={},
+        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=bias, **weight_args
+            ),
+        )
+        self.out = nn.Identity()
+
+    def forward(self, x):
+        """
+        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
+        assert T % self.temporal_patch_size == 0
+        x = self.proj(x)
+        return self.out(x)
+
+
+class FinalLayer(nn.Module):
+    """
+    The final layer of video DiT.
+    """
+
+    def __init__(
+        self,
+        hidden_size,
+        spatial_patch_size,
+        temporal_patch_size,
+        out_channels,
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 256,
+        weight_args={},
+        operations=None,
+    ):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, **weight_args)
+        self.linear = operations.Linear(
+            hidden_size, spatial_patch_size * spatial_patch_size * temporal_patch_size * out_channels, bias=False, **weight_args
+        )
+        self.hidden_size = hidden_size
+        self.n_adaln_chunks = 2
+        self.use_adaln_lora = use_adaln_lora
+        if use_adaln_lora:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(hidden_size, adaln_lora_dim, bias=False, **weight_args),
+                operations.Linear(adaln_lora_dim, self.n_adaln_chunks * hidden_size, bias=False, **weight_args),
+            )
+        else:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(), operations.Linear(hidden_size, self.n_adaln_chunks * hidden_size, bias=False, **weight_args)
+            )
+
+    def forward(
+        self,
+        x_BT_HW_D,
+        emb_B_D,
+        adaln_lora_B_3D: Optional[torch.Tensor] = None,
+    ):
+        if self.use_adaln_lora:
+            assert adaln_lora_B_3D is not None
+            shift_B_D, scale_B_D = (self.adaLN_modulation(emb_B_D) + adaln_lora_B_3D[:, : 2 * self.hidden_size]).chunk(
+                2, dim=1
+            )
+        else:
+            shift_B_D, scale_B_D = self.adaLN_modulation(emb_B_D).chunk(2, dim=1)
+
+        B = emb_B_D.shape[0]
+        T = x_BT_HW_D.shape[0] // B
+        shift_BT_D, scale_BT_D = repeat(shift_B_D, "b d -> (b t) d", t=T), repeat(scale_B_D, "b d -> (b t) d", t=T)
+        x_BT_HW_D = modulate(self.norm_final(x_BT_HW_D), shift_BT_D, scale_BT_D)
+
+        x_BT_HW_D = self.linear(x_BT_HW_D)
+        return x_BT_HW_D
+
+
+class VideoAttn(nn.Module):
+    """
+    Implements video attention with optional cross-attention capabilities.
+
+    This module processes video features while maintaining their spatio-temporal structure. It can perform
+    self-attention within the video features or cross-attention with external context features.
+
+    Parameters:
+        x_dim (int): Dimension of input feature vectors
+        context_dim (Optional[int]): Dimension of context features for cross-attention. None for self-attention
+        num_heads (int): Number of attention heads
+        bias (bool): Whether to include bias in attention projections. Default: False
+        qkv_norm_mode (str): Normalization mode for query/key/value projections. Must be "per_head". Default: "per_head"
+        x_format (str): Format of input tensor. Must be "BTHWD". Default: "BTHWD"
+
+    Input shape:
+        - x: (T, H, W, B, D) video features
+        - context (optional): (M, B, D) context features for cross-attention
+        where:
+            T: temporal dimension
+            H: height
+            W: width
+            B: batch size
+            D: feature dimension
+            M: context sequence length
+    """
+
+    def __init__(
+        self,
+        x_dim: int,
+        context_dim: Optional[int],
+        num_heads: int,
+        bias: bool = False,
+        qkv_norm_mode: str = "per_head",
+        x_format: str = "BTHWD",
+        weight_args={},
+        operations=None,
+    ) -> None:
+        super().__init__()
+        self.x_format = x_format
+
+        self.attn = Attention(
+            x_dim,
+            context_dim,
+            num_heads,
+            x_dim // num_heads,
+            qkv_bias=bias,
+            qkv_norm="RRI",
+            out_bias=bias,
+            qkv_norm_mode=qkv_norm_mode,
+            qkv_format="sbhd",
+            weight_args=weight_args,
+            operations=operations,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        context: Optional[torch.Tensor] = None,
+        crossattn_mask: Optional[torch.Tensor] = None,
+        rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
+        transformer_options: Optional[dict] = {},
+    ) -> torch.Tensor:
+        """
+        Forward pass for video attention.
+
+        Args:
+            x (Tensor): Input tensor of shape (B, T, H, W, D) or (T, H, W, B, D) representing batches of video data.
+            context (Tensor): Context tensor of shape (B, M, D) or (M, B, D),
+            where M is the sequence length of the context.
+            crossattn_mask (Optional[Tensor]): An optional mask for cross-attention mechanisms.
+            rope_emb_L_1_1_D (Optional[Tensor]):
+            Rotary positional embedding tensor of shape (L, 1, 1, D). L == THW for current video training.
+
+        Returns:
+            Tensor: The output tensor with applied attention, maintaining the input shape.
+        """
+
+        x_T_H_W_B_D = x
+        context_M_B_D = context
+        T, H, W, B, D = x_T_H_W_B_D.shape
+        x_THW_B_D = rearrange(x_T_H_W_B_D, "t h w b d -> (t h w) b d")
+        x_THW_B_D = self.attn(
+            x_THW_B_D,
+            context_M_B_D,
+            crossattn_mask,
+            rope_emb=rope_emb_L_1_1_D,
+            transformer_options=transformer_options,
+        )
+        x_T_H_W_B_D = rearrange(x_THW_B_D, "(t h w) b d -> t h w b d", h=H, w=W)
+        return x_T_H_W_B_D
+
+
+def adaln_norm_state(norm_state, x, scale, shift):
+    normalized = norm_state(x)
+    return normalized * (1 + scale) + shift
+
+
+class DITBuildingBlock(nn.Module):
+    """
+    A building block for the DiT (Diffusion Transformer) architecture that supports different types of
+    attention and MLP operations with adaptive layer normalization.
+
+    Parameters:
+        block_type (str): Type of block - one of:
+            - "cross_attn"/"ca": Cross-attention
+            - "full_attn"/"fa": Full self-attention
+            - "mlp"/"ff": MLP/feedforward block
+        x_dim (int): Dimension of input features
+        context_dim (Optional[int]): Dimension of context features for cross-attention
+        num_heads (int): Number of attention heads
+        mlp_ratio (float): MLP hidden dimension multiplier. Default: 4.0
+        bias (bool): Whether to use bias in layers. Default: False
+        mlp_dropout (float): Dropout rate for MLP. Default: 0.0
+        qkv_norm_mode (str): QKV normalization mode. Default: "per_head"
+        x_format (str): Input tensor format. Default: "BTHWD"
+        use_adaln_lora (bool): Whether to use AdaLN-LoRA. Default: False
+        adaln_lora_dim (int): Dimension for AdaLN-LoRA. Default: 256
+    """
+
+    def __init__(
+        self,
+        block_type: str,
+        x_dim: int,
+        context_dim: Optional[int],
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        bias: bool = False,
+        mlp_dropout: float = 0.0,
+        qkv_norm_mode: str = "per_head",
+        x_format: str = "BTHWD",
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 256,
+        weight_args={},
+        operations=None
+    ) -> None:
+        block_type = block_type.lower()
+
+        super().__init__()
+        self.x_format = x_format
+        if block_type in ["cross_attn", "ca"]:
+            self.block = VideoAttn(
+                x_dim,
+                context_dim,
+                num_heads,
+                bias=bias,
+                qkv_norm_mode=qkv_norm_mode,
+                x_format=self.x_format,
+                weight_args=weight_args,
+                operations=operations,
+            )
+        elif block_type in ["full_attn", "fa"]:
+            self.block = VideoAttn(
+                x_dim, None, num_heads, bias=bias, qkv_norm_mode=qkv_norm_mode, x_format=self.x_format, weight_args=weight_args, operations=operations
+            )
+        elif block_type in ["mlp", "ff"]:
+            self.block = GPT2FeedForward(x_dim, int(x_dim * mlp_ratio), dropout=mlp_dropout, bias=bias, weight_args=weight_args, operations=operations)
+        else:
+            raise ValueError(f"Unknown block type: {block_type}")
+
+        self.block_type = block_type
+        self.use_adaln_lora = use_adaln_lora
+
+        self.norm_state = nn.LayerNorm(x_dim, elementwise_affine=False, eps=1e-6)
+        self.n_adaln_chunks = 3
+        if use_adaln_lora:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(x_dim, adaln_lora_dim, bias=False, **weight_args),
+                operations.Linear(adaln_lora_dim, self.n_adaln_chunks * x_dim, bias=False, **weight_args),
+            )
+        else:
+            self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(x_dim, self.n_adaln_chunks * x_dim, bias=False, **weight_args))
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        emb_B_D: torch.Tensor,
+        crossattn_emb: torch.Tensor,
+        crossattn_mask: Optional[torch.Tensor] = None,
+        rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
+        adaln_lora_B_3D: Optional[torch.Tensor] = None,
+        transformer_options: Optional[dict] = {},
+    ) -> torch.Tensor:
+        """
+        Forward pass for dynamically configured blocks with adaptive normalization.
+
+        Args:
+            x (Tensor): Input tensor of shape (B, T, H, W, D) or (T, H, W, B, D).
+            emb_B_D (Tensor): Embedding tensor for adaptive layer normalization modulation.
+            crossattn_emb (Tensor): Tensor for cross-attention blocks.
+            crossattn_mask (Optional[Tensor]): Optional mask for cross-attention.
+            rope_emb_L_1_1_D (Optional[Tensor]):
+            Rotary positional embedding tensor of shape (L, 1, 1, D). L == THW for current video training.
+
+        Returns:
+            Tensor: The output tensor after processing through the configured block and adaptive normalization.
+        """
+        if self.use_adaln_lora:
+            shift_B_D, scale_B_D, gate_B_D = (self.adaLN_modulation(emb_B_D) + adaln_lora_B_3D).chunk(
+                self.n_adaln_chunks, dim=1
+            )
+        else:
+            shift_B_D, scale_B_D, gate_B_D = self.adaLN_modulation(emb_B_D).chunk(self.n_adaln_chunks, dim=1)
+
+        shift_1_1_1_B_D, scale_1_1_1_B_D, gate_1_1_1_B_D = (
+            shift_B_D.unsqueeze(0).unsqueeze(0).unsqueeze(0),
+            scale_B_D.unsqueeze(0).unsqueeze(0).unsqueeze(0),
+            gate_B_D.unsqueeze(0).unsqueeze(0).unsqueeze(0),
+        )
+
+        if self.block_type in ["mlp", "ff"]:
+            x = x + gate_1_1_1_B_D * self.block(
+                adaln_norm_state(self.norm_state, x, scale_1_1_1_B_D, shift_1_1_1_B_D),
+            )
+        elif self.block_type in ["full_attn", "fa"]:
+            x = x + gate_1_1_1_B_D * self.block(
+                adaln_norm_state(self.norm_state, x, scale_1_1_1_B_D, shift_1_1_1_B_D),
+                context=None,
+                rope_emb_L_1_1_D=rope_emb_L_1_1_D,
+                transformer_options=transformer_options,
+            )
+        elif self.block_type in ["cross_attn", "ca"]:
+            x = x + gate_1_1_1_B_D * self.block(
+                adaln_norm_state(self.norm_state, x, scale_1_1_1_B_D, shift_1_1_1_B_D),
+                context=crossattn_emb,
+                crossattn_mask=crossattn_mask,
+                rope_emb_L_1_1_D=rope_emb_L_1_1_D,
+                transformer_options=transformer_options,
+            )
+        else:
+            raise ValueError(f"Unknown block type: {self.block_type}")
+
+        return x
+
+
+class GeneralDITTransformerBlock(nn.Module):
+    """
+    A wrapper module that manages a sequence of DITBuildingBlocks to form a complete transformer layer.
+    Each block in the sequence is specified by a block configuration string.
+
+    Parameters:
+        x_dim (int): Dimension of input features
+        context_dim (int): Dimension of context features for cross-attention blocks
+        num_heads (int): Number of attention heads
+        block_config (str): String specifying block sequence (e.g. "ca-fa-mlp" for cross-attention,
+                          full-attention, then MLP)
+        mlp_ratio (float): MLP hidden dimension multiplier. Default: 4.0
+        x_format (str): Input tensor format. Default: "BTHWD"
+        use_adaln_lora (bool): Whether to use AdaLN-LoRA. Default: False
+        adaln_lora_dim (int): Dimension for AdaLN-LoRA. Default: 256
+
+    The block_config string uses "-" to separate block types:
+        - "ca"/"cross_attn": Cross-attention block
+        - "fa"/"full_attn": Full self-attention block
+        - "mlp"/"ff": MLP/feedforward block
+
+    Example:
+        block_config = "ca-fa-mlp" creates a sequence of:
+        1. Cross-attention block
+        2. Full self-attention block
+        3. MLP block
+    """
+
+    def __init__(
+        self,
+        x_dim: int,
+        context_dim: int,
+        num_heads: int,
+        block_config: str,
+        mlp_ratio: float = 4.0,
+        x_format: str = "BTHWD",
+        use_adaln_lora: bool = False,
+        adaln_lora_dim: int = 256,
+        weight_args={},
+        operations=None
+    ):
+        super().__init__()
+        self.blocks = nn.ModuleList()
+        self.x_format = x_format
+        for block_type in block_config.split("-"):
+            self.blocks.append(
+                DITBuildingBlock(
+                    block_type,
+                    x_dim,
+                    context_dim,
+                    num_heads,
+                    mlp_ratio,
+                    x_format=self.x_format,
+                    use_adaln_lora=use_adaln_lora,
+                    adaln_lora_dim=adaln_lora_dim,
+                    weight_args=weight_args,
+                    operations=operations,
+                )
+            )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        emb_B_D: torch.Tensor,
+        crossattn_emb: torch.Tensor,
+        crossattn_mask: Optional[torch.Tensor] = None,
+        rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
+        adaln_lora_B_3D: Optional[torch.Tensor] = None,
+        transformer_options: Optional[dict] = {},
+    ) -> torch.Tensor:
+        for block in self.blocks:
+            x = block(
+                x,
+                emb_B_D,
+                crossattn_emb,
+                crossattn_mask,
+                rope_emb_L_1_1_D=rope_emb_L_1_1_D,
+                adaln_lora_B_3D=adaln_lora_B_3D,
+                transformer_options=transformer_options,
+            )
+        return x

ComfyUI/comfy/ldm/cosmos/cosmos_tokenizer/layers3d.py

@@ -0,0 +1,1041 @@
+# 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 model definition for 3D layers
+
+Adapted from: https://github.com/lucidrains/magvit2-pytorch/blob/
+9f49074179c912736e617d61b32be367eb5f993a/magvit2_pytorch/magvit2_pytorch.py#L889
+
+[MIT License Copyright (c) 2023 Phil Wang]
+https://github.com/lucidrains/magvit2-pytorch/blob/
+9f49074179c912736e617d61b32be367eb5f993a/LICENSE
+"""
+import math
+from typing import Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import logging
+
+from comfy.ldm.modules.diffusionmodules.model import vae_attention
+
+from .patching import (
+    Patcher,
+    Patcher3D,
+    UnPatcher,
+    UnPatcher3D,
+)
+from .utils import (
+    CausalNormalize,
+    batch2space,
+    batch2time,
+    cast_tuple,
+    is_odd,
+    nonlinearity,
+    replication_pad,
+    space2batch,
+    time2batch,
+)
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+_LEGACY_NUM_GROUPS = 32
+
+
+class CausalConv3d(nn.Module):
+    def __init__(
+        self,
+        chan_in: int = 1,
+        chan_out: int = 1,
+        kernel_size: Union[int, Tuple[int, int, int]] = 3,
+        pad_mode: str = "constant",
+        **kwargs,
+    ):
+        super().__init__()
+        kernel_size = cast_tuple(kernel_size, 3)
+
+        time_kernel_size, height_kernel_size, width_kernel_size = kernel_size
+
+        assert is_odd(height_kernel_size) and is_odd(width_kernel_size)
+
+        dilation = kwargs.pop("dilation", 1)
+        stride = kwargs.pop("stride", 1)
+        time_stride = kwargs.pop("time_stride", 1)
+        time_dilation = kwargs.pop("time_dilation", 1)
+        padding = kwargs.pop("padding", 1)
+
+        self.pad_mode = pad_mode
+        time_pad = time_dilation * (time_kernel_size - 1) + (1 - time_stride)
+        self.time_pad = time_pad
+
+        self.spatial_pad = (padding, padding, padding, padding)
+
+        stride = (time_stride, stride, stride)
+        dilation = (time_dilation, dilation, dilation)
+        self.conv3d = ops.Conv3d(
+            chan_in,
+            chan_out,
+            kernel_size,
+            stride=stride,
+            dilation=dilation,
+            **kwargs,
+        )
+
+    def _replication_pad(self, x: torch.Tensor) -> torch.Tensor:
+        x_prev = x[:, :, :1, ...].repeat(1, 1, self.time_pad, 1, 1)
+        x = torch.cat([x_prev, x], dim=2)
+        padding = self.spatial_pad + (0, 0)
+        return F.pad(x, padding, mode=self.pad_mode, value=0.0)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self._replication_pad(x)
+        return self.conv3d(x)
+
+
+class CausalUpsample3d(nn.Module):
+    def __init__(self, in_channels: int) -> None:
+        super().__init__()
+        self.conv = CausalConv3d(
+            in_channels, in_channels, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = x.repeat_interleave(2, dim=3).repeat_interleave(2, dim=4)
+        time_factor = 1.0 + 1.0 * (x.shape[2] > 1)
+        if isinstance(time_factor, torch.Tensor):
+            time_factor = time_factor.item()
+        x = x.repeat_interleave(int(time_factor), dim=2)
+        # TODO(freda): Check if this causes temporal inconsistency.
+        # Shoule reverse the order of the following two ops,
+        # better perf and better temporal smoothness.
+        x = self.conv(x)
+        return x[..., int(time_factor - 1) :, :, :]
+
+
+class CausalDownsample3d(nn.Module):
+    def __init__(self, in_channels: int) -> None:
+        super().__init__()
+        self.conv = CausalConv3d(
+            in_channels,
+            in_channels,
+            kernel_size=3,
+            stride=2,
+            time_stride=2,
+            padding=0,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        pad = (0, 1, 0, 1, 0, 0)
+        x = F.pad(x, pad, mode="constant", value=0)
+        x = replication_pad(x)
+        x = self.conv(x)
+        return x
+
+
+class CausalHybridUpsample3d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        spatial_up: bool = True,
+        temporal_up: bool = True,
+        **kwargs,
+    ) -> None:
+        super().__init__()
+        self.spatial_up = spatial_up
+        self.temporal_up = temporal_up
+        if not self.spatial_up and not self.temporal_up:
+            return
+
+        self.conv1 = CausalConv3d(
+            in_channels,
+            in_channels,
+            kernel_size=(3, 1, 1),
+            stride=1,
+            time_stride=1,
+            padding=0,
+        )
+        self.conv2 = CausalConv3d(
+            in_channels,
+            in_channels,
+            kernel_size=(1, 3, 3),
+            stride=1,
+            time_stride=1,
+            padding=1,
+        )
+        self.conv3 = CausalConv3d(
+            in_channels,
+            in_channels,
+            kernel_size=1,
+            stride=1,
+            time_stride=1,
+            padding=0,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if not self.spatial_up and not self.temporal_up:
+            return x
+
+        # hybrid upsample temporally.
+        if self.temporal_up:
+            time_factor = 1.0 + 1.0 * (x.shape[2] > 1)
+            if isinstance(time_factor, torch.Tensor):
+                time_factor = time_factor.item()
+            x = x.repeat_interleave(int(time_factor), dim=2)
+            x = x[..., int(time_factor - 1) :, :, :]
+            x = self.conv1(x) + x
+
+        # hybrid upsample spatially.
+        if self.spatial_up:
+            x = x.repeat_interleave(2, dim=3).repeat_interleave(2, dim=4)
+            x = self.conv2(x) + x
+
+        # final 1x1x1 conv.
+        x = self.conv3(x)
+        return x
+
+
+class CausalHybridDownsample3d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        spatial_down: bool = True,
+        temporal_down: bool = True,
+        **kwargs,
+    ) -> None:
+        super().__init__()
+        self.spatial_down = spatial_down
+        self.temporal_down = temporal_down
+        if not self.spatial_down and not self.temporal_down:
+            return
+
+        self.conv1 = CausalConv3d(
+            in_channels,
+            in_channels,
+            kernel_size=(1, 3, 3),
+            stride=2,
+            time_stride=1,
+            padding=0,
+        )
+        self.conv2 = CausalConv3d(
+            in_channels,
+            in_channels,
+            kernel_size=(3, 1, 1),
+            stride=1,
+            time_stride=2,
+            padding=0,
+        )
+        self.conv3 = CausalConv3d(
+            in_channels,
+            in_channels,
+            kernel_size=1,
+            stride=1,
+            time_stride=1,
+            padding=0,
+        )
+
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if not self.spatial_down and not self.temporal_down:
+            return x
+
+        # hybrid downsample spatially.
+        if self.spatial_down:
+            pad = (0, 1, 0, 1, 0, 0)
+            x = F.pad(x, pad, mode="constant", value=0)
+            x1 = self.conv1(x)
+            x2 = F.avg_pool3d(x, kernel_size=(1, 2, 2), stride=(1, 2, 2))
+            x = x1 + x2
+
+        # hybrid downsample temporally.
+        if self.temporal_down:
+            x = replication_pad(x)
+            x1 = self.conv2(x)
+            x2 = F.avg_pool3d(x, kernel_size=(2, 1, 1), stride=(2, 1, 1))
+            x = x1 + x2
+
+        # final 1x1x1 conv.
+        x = self.conv3(x)
+        return x
+
+
+class CausalResnetBlock3d(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels: int,
+        out_channels: int = None,
+        dropout: float,
+        num_groups: int,
+    ) -> None:
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+
+        self.norm1 = CausalNormalize(in_channels, num_groups=num_groups)
+        self.conv1 = CausalConv3d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        self.norm2 = CausalNormalize(out_channels, num_groups=num_groups)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = CausalConv3d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        self.nin_shortcut = (
+            CausalConv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+            if in_channels != out_channels
+            else nn.Identity()
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+        x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class CausalResnetBlockFactorized3d(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels: int,
+        out_channels: int = None,
+        dropout: float,
+        num_groups: int,
+    ) -> None:
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+
+        self.norm1 = CausalNormalize(in_channels, num_groups=1)
+        self.conv1 = nn.Sequential(
+            CausalConv3d(
+                in_channels,
+                out_channels,
+                kernel_size=(1, 3, 3),
+                stride=1,
+                padding=1,
+            ),
+            CausalConv3d(
+                out_channels,
+                out_channels,
+                kernel_size=(3, 1, 1),
+                stride=1,
+                padding=0,
+            ),
+        )
+        self.norm2 = CausalNormalize(out_channels, num_groups=num_groups)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = nn.Sequential(
+            CausalConv3d(
+                out_channels,
+                out_channels,
+                kernel_size=(1, 3, 3),
+                stride=1,
+                padding=1,
+            ),
+            CausalConv3d(
+                out_channels,
+                out_channels,
+                kernel_size=(3, 1, 1),
+                stride=1,
+                padding=0,
+            ),
+        )
+        self.nin_shortcut = (
+            CausalConv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+            if in_channels != out_channels
+            else nn.Identity()
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+        x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class CausalAttnBlock(nn.Module):
+    def __init__(self, in_channels: int, num_groups: int) -> None:
+        super().__init__()
+
+        self.norm = CausalNormalize(in_channels, num_groups=num_groups)
+        self.q = CausalConv3d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.k = CausalConv3d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.v = CausalConv3d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.proj_out = CausalConv3d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+
+        self.optimized_attention = vae_attention()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        q, batch_size = time2batch(q)
+        k, batch_size = time2batch(k)
+        v, batch_size = time2batch(v)
+
+        b, c, h, w = q.shape
+        h_ = self.optimized_attention(q, k, v)
+
+        h_ = batch2time(h_, batch_size)
+        h_ = self.proj_out(h_)
+        return x + h_
+
+
+class CausalTemporalAttnBlock(nn.Module):
+    def __init__(self, in_channels: int, num_groups: int) -> None:
+        super().__init__()
+
+        self.norm = CausalNormalize(in_channels, num_groups=num_groups)
+        self.q = CausalConv3d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.k = CausalConv3d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.v = CausalConv3d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.proj_out = CausalConv3d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        q, batch_size, height = space2batch(q)
+        k, _, _ = space2batch(k)
+        v, _, _ = space2batch(v)
+
+        bhw, c, t = q.shape
+        q = q.permute(0, 2, 1)  # (bhw, t, c)
+        k = k.permute(0, 2, 1)  # (bhw, t, c)
+        v = v.permute(0, 2, 1)  # (bhw, t, c)
+
+        w_ = torch.bmm(q, k.permute(0, 2, 1))  # (bhw, t, t)
+        w_ = w_ * (int(c) ** (-0.5))
+
+        # Apply causal mask
+        mask = torch.tril(torch.ones_like(w_))
+        w_ = w_.masked_fill(mask == 0, float("-inf"))
+        w_ = F.softmax(w_, dim=2)
+
+        # attend to values
+        h_ = torch.bmm(w_, v)  # (bhw, t, c)
+        h_ = h_.permute(0, 2, 1).reshape(bhw, c, t)  # (bhw, c, t)
+
+        h_ = batch2space(h_, batch_size, height)
+        h_ = self.proj_out(h_)
+        return x + h_
+
+
+class EncoderBase(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        channels: int,
+        channels_mult: list[int],
+        num_res_blocks: int,
+        attn_resolutions: list[int],
+        dropout: float,
+        resolution: int,
+        z_channels: int,
+        **ignore_kwargs,
+    ) -> None:
+        super().__init__()
+        self.num_resolutions = len(channels_mult)
+        self.num_res_blocks = num_res_blocks
+
+        # Patcher.
+        patch_size = ignore_kwargs.get("patch_size", 1)
+        self.patcher = Patcher(
+            patch_size, ignore_kwargs.get("patch_method", "rearrange")
+        )
+        in_channels = in_channels * patch_size * patch_size
+
+        # downsampling
+        self.conv_in = CausalConv3d(
+            in_channels, channels, kernel_size=3, stride=1, padding=1
+        )
+
+        # num of groups for GroupNorm, num_groups=1 for LayerNorm.
+        num_groups = ignore_kwargs.get("num_groups", _LEGACY_NUM_GROUPS)
+        curr_res = resolution // patch_size
+        in_ch_mult = (1,) + tuple(channels_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = channels * in_ch_mult[i_level]
+            block_out = channels * channels_mult[i_level]
+            for _ in range(self.num_res_blocks):
+                block.append(
+                    CausalResnetBlock3d(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        dropout=dropout,
+                        num_groups=num_groups,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(CausalAttnBlock(block_in, num_groups=num_groups))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = CausalDownsample3d(block_in)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = CausalResnetBlock3d(
+            in_channels=block_in,
+            out_channels=block_in,
+            dropout=dropout,
+            num_groups=num_groups,
+        )
+        self.mid.attn_1 = CausalAttnBlock(block_in, num_groups=num_groups)
+        self.mid.block_2 = CausalResnetBlock3d(
+            in_channels=block_in,
+            out_channels=block_in,
+            dropout=dropout,
+            num_groups=num_groups,
+        )
+
+        # end
+        self.norm_out = CausalNormalize(block_in, num_groups=num_groups)
+        self.conv_out = CausalConv3d(
+            block_in, z_channels, kernel_size=3, stride=1, padding=1
+        )
+
+    def patcher3d(self, x: torch.Tensor) -> torch.Tensor:
+        x, batch_size = time2batch(x)
+        x = self.patcher(x)
+        x = batch2time(x, batch_size)
+        return x
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.patcher3d(x)
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1])
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions - 1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+            else:
+                # temporal downsample (last level)
+                time_factor = 1 + 1 * (hs[-1].shape[2] > 1)
+                if isinstance(time_factor, torch.Tensor):
+                    time_factor = time_factor.item()
+                hs[-1] = replication_pad(hs[-1])
+                hs.append(
+                    F.avg_pool3d(
+                        hs[-1],
+                        kernel_size=[time_factor, 1, 1],
+                        stride=[2, 1, 1],
+                    )
+                )
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class DecoderBase(nn.Module):
+    def __init__(
+        self,
+        out_channels: int,
+        channels: int,
+        channels_mult: list[int],
+        num_res_blocks: int,
+        attn_resolutions: list[int],
+        dropout: float,
+        resolution: int,
+        z_channels: int,
+        **ignore_kwargs,
+    ):
+        super().__init__()
+        self.num_resolutions = len(channels_mult)
+        self.num_res_blocks = num_res_blocks
+
+        # UnPatcher.
+        patch_size = ignore_kwargs.get("patch_size", 1)
+        self.unpatcher = UnPatcher(
+            patch_size, ignore_kwargs.get("patch_method", "rearrange")
+        )
+        out_ch = out_channels * patch_size * patch_size
+
+        block_in = channels * channels_mult[self.num_resolutions - 1]
+        curr_res = (resolution // patch_size) // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+        logging.debug(
+            "Working with z of shape {} = {} dimensions.".format(
+                self.z_shape, np.prod(self.z_shape)
+            )
+        )
+
+        # z to block_in
+        self.conv_in = CausalConv3d(
+            z_channels, block_in, kernel_size=3, stride=1, padding=1
+        )
+
+        # num of groups for GroupNorm, num_groups=1 for LayerNorm.
+        num_groups = ignore_kwargs.get("num_groups", _LEGACY_NUM_GROUPS)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = CausalResnetBlock3d(
+            in_channels=block_in,
+            out_channels=block_in,
+            dropout=dropout,
+            num_groups=num_groups,
+        )
+        self.mid.attn_1 = CausalAttnBlock(block_in, num_groups=num_groups)
+        self.mid.block_2 = CausalResnetBlock3d(
+            in_channels=block_in,
+            out_channels=block_in,
+            dropout=dropout,
+            num_groups=num_groups,
+        )
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = channels * channels_mult[i_level]
+            for _ in range(self.num_res_blocks + 1):
+                block.append(
+                    CausalResnetBlock3d(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        dropout=dropout,
+                        num_groups=num_groups,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(CausalAttnBlock(block_in, num_groups=num_groups))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = CausalUpsample3d(block_in)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = CausalNormalize(block_in, num_groups=num_groups)
+        self.conv_out = CausalConv3d(
+            block_in, out_ch, kernel_size=3, stride=1, padding=1
+        )
+
+    def unpatcher3d(self, x: torch.Tensor) -> torch.Tensor:
+        x, batch_size = time2batch(x)
+        x = self.unpatcher(x)
+        x = batch2time(x, batch_size)
+
+        return x
+
+    def forward(self, z):
+        h = self.conv_in(z)
+
+        # middle block.
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+
+        # decoder blocks.
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+            else:
+                # temporal upsample (last level)
+                time_factor = 1.0 + 1.0 * (h.shape[2] > 1)
+                if isinstance(time_factor, torch.Tensor):
+                    time_factor = time_factor.item()
+                h = h.repeat_interleave(int(time_factor), dim=2)
+                h = h[..., int(time_factor - 1) :, :, :]
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        h = self.unpatcher3d(h)
+        return h
+
+
+class EncoderFactorized(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        channels: int,
+        channels_mult: list[int],
+        num_res_blocks: int,
+        attn_resolutions: list[int],
+        dropout: float,
+        resolution: int,
+        z_channels: int,
+        spatial_compression: int = 8,
+        temporal_compression: int = 8,
+        **ignore_kwargs,
+    ) -> None:
+        super().__init__()
+        self.num_resolutions = len(channels_mult)
+        self.num_res_blocks = num_res_blocks
+
+        # Patcher.
+        patch_size = ignore_kwargs.get("patch_size", 1)
+        self.patcher3d = Patcher3D(
+            patch_size, ignore_kwargs.get("patch_method", "haar")
+        )
+        in_channels = in_channels * patch_size * patch_size * patch_size
+
+        # calculate the number of downsample operations
+        self.num_spatial_downs = int(math.log2(spatial_compression)) - int(
+            math.log2(patch_size)
+        )
+        assert (
+            self.num_spatial_downs <= self.num_resolutions
+        ), f"Spatially downsample {self.num_resolutions} times at most"
+
+        self.num_temporal_downs = int(math.log2(temporal_compression)) - int(
+            math.log2(patch_size)
+        )
+        assert (
+            self.num_temporal_downs <= self.num_resolutions
+        ), f"Temporally downsample {self.num_resolutions} times at most"
+
+        # downsampling
+        self.conv_in = nn.Sequential(
+            CausalConv3d(
+                in_channels,
+                channels,
+                kernel_size=(1, 3, 3),
+                stride=1,
+                padding=1,
+            ),
+            CausalConv3d(
+                channels, channels, kernel_size=(3, 1, 1), stride=1, padding=0
+            ),
+        )
+
+        curr_res = resolution // patch_size
+        in_ch_mult = (1,) + tuple(channels_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = channels * in_ch_mult[i_level]
+            block_out = channels * channels_mult[i_level]
+            for _ in range(self.num_res_blocks):
+                block.append(
+                    CausalResnetBlockFactorized3d(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        dropout=dropout,
+                        num_groups=1,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(
+                        nn.Sequential(
+                            CausalAttnBlock(block_in, num_groups=1),
+                            CausalTemporalAttnBlock(block_in, num_groups=1),
+                        )
+                    )
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                spatial_down = i_level < self.num_spatial_downs
+                temporal_down = i_level < self.num_temporal_downs
+                down.downsample = CausalHybridDownsample3d(
+                    block_in,
+                    spatial_down=spatial_down,
+                    temporal_down=temporal_down,
+                )
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = CausalResnetBlockFactorized3d(
+            in_channels=block_in,
+            out_channels=block_in,
+            dropout=dropout,
+            num_groups=1,
+        )
+        self.mid.attn_1 = nn.Sequential(
+            CausalAttnBlock(block_in, num_groups=1),
+            CausalTemporalAttnBlock(block_in, num_groups=1),
+        )
+        self.mid.block_2 = CausalResnetBlockFactorized3d(
+            in_channels=block_in,
+            out_channels=block_in,
+            dropout=dropout,
+            num_groups=1,
+        )
+
+        # end
+        self.norm_out = CausalNormalize(block_in, num_groups=1)
+        self.conv_out = nn.Sequential(
+            CausalConv3d(
+                block_in, z_channels, kernel_size=(1, 3, 3), stride=1, padding=1
+            ),
+            CausalConv3d(
+                z_channels,
+                z_channels,
+                kernel_size=(3, 1, 1),
+                stride=1,
+                padding=0,
+            ),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.patcher3d(x)
+
+        # downsampling
+        h = self.conv_in(x)
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](h)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+            if i_level != self.num_resolutions - 1:
+                h = self.down[i_level].downsample(h)
+
+        # middle
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class DecoderFactorized(nn.Module):
+    def __init__(
+        self,
+        out_channels: int,
+        channels: int,
+        channels_mult: list[int],
+        num_res_blocks: int,
+        attn_resolutions: list[int],
+        dropout: float,
+        resolution: int,
+        z_channels: int,
+        spatial_compression: int = 8,
+        temporal_compression: int = 8,
+        **ignore_kwargs,
+    ):
+        super().__init__()
+        self.num_resolutions = len(channels_mult)
+        self.num_res_blocks = num_res_blocks
+
+        # UnPatcher.
+        patch_size = ignore_kwargs.get("patch_size", 1)
+        self.unpatcher3d = UnPatcher3D(
+            patch_size, ignore_kwargs.get("patch_method", "haar")
+        )
+        out_ch = out_channels * patch_size * patch_size * patch_size
+
+        # calculate the number of upsample operations
+        self.num_spatial_ups = int(math.log2(spatial_compression)) - int(
+            math.log2(patch_size)
+        )
+        assert (
+            self.num_spatial_ups <= self.num_resolutions
+        ), f"Spatially upsample {self.num_resolutions} times at most"
+        self.num_temporal_ups = int(math.log2(temporal_compression)) - int(
+            math.log2(patch_size)
+        )
+        assert (
+            self.num_temporal_ups <= self.num_resolutions
+        ), f"Temporally upsample {self.num_resolutions} times at most"
+
+        block_in = channels * channels_mult[self.num_resolutions - 1]
+        curr_res = (resolution // patch_size) // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+        logging.debug(
+            "Working with z of shape {} = {} dimensions.".format(
+                self.z_shape, np.prod(self.z_shape)
+            )
+        )
+
+        # z to block_in
+        self.conv_in = nn.Sequential(
+            CausalConv3d(
+                z_channels, block_in, kernel_size=(1, 3, 3), stride=1, padding=1
+            ),
+            CausalConv3d(
+                block_in, block_in, kernel_size=(3, 1, 1), stride=1, padding=0
+            ),
+        )
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = CausalResnetBlockFactorized3d(
+            in_channels=block_in,
+            out_channels=block_in,
+            dropout=dropout,
+            num_groups=1,
+        )
+        self.mid.attn_1 = nn.Sequential(
+            CausalAttnBlock(block_in, num_groups=1),
+            CausalTemporalAttnBlock(block_in, num_groups=1),
+        )
+        self.mid.block_2 = CausalResnetBlockFactorized3d(
+            in_channels=block_in,
+            out_channels=block_in,
+            dropout=dropout,
+            num_groups=1,
+        )
+
+        legacy_mode = ignore_kwargs.get("legacy_mode", False)
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = channels * channels_mult[i_level]
+            for _ in range(self.num_res_blocks + 1):
+                block.append(
+                    CausalResnetBlockFactorized3d(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        dropout=dropout,
+                        num_groups=1,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(
+                        nn.Sequential(
+                            CausalAttnBlock(block_in, num_groups=1),
+                            CausalTemporalAttnBlock(block_in, num_groups=1),
+                        )
+                    )
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                # The layer index for temporal/spatial downsampling performed
+                # in the encoder should correspond to the layer index in
+                # reverse order where upsampling is performed in the decoder.
+                # If you've a pre-trained model, you can simply finetune.
+                i_level_reverse = self.num_resolutions - i_level - 1
+                if legacy_mode:
+                    temporal_up = i_level_reverse < self.num_temporal_ups
+                else:
+                    temporal_up = 0 < i_level_reverse < self.num_temporal_ups + 1
+                spatial_up = temporal_up or (
+                    i_level_reverse < self.num_spatial_ups
+                    and self.num_spatial_ups > self.num_temporal_ups
+                )
+                up.upsample = CausalHybridUpsample3d(
+                    block_in, spatial_up=spatial_up, temporal_up=temporal_up
+                )
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = CausalNormalize(block_in, num_groups=1)
+        self.conv_out = nn.Sequential(
+            CausalConv3d(block_in, out_ch, kernel_size=(1, 3, 3), stride=1, padding=1),
+            CausalConv3d(out_ch, out_ch, kernel_size=(3, 1, 1), stride=1, padding=0),
+        )
+
+    def forward(self, z):
+        h = self.conv_in(z)
+
+        # middle block.
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+
+        # decoder blocks.
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        h = self.unpatcher3d(h)
+        return h

ComfyUI/comfy/ldm/cosmos/cosmos_tokenizer/patching.py

@@ -0,0 +1,377 @@
+# 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 patcher and unpatcher implementation for 2D and 3D data.
+
+The idea of Haar wavelet is to compute LL, LH, HL, HH component as two 1D convolutions.
+One on the rows and one on the columns.
+For example, in 1D signal, we have [a, b], then the low-freq compoenent is [a + b] / 2 and high-freq is [a - b] / 2.
+We can use a 1D convolution with kernel [1, 1] and stride 2 to represent the L component.
+For H component, we can use a 1D convolution with kernel [1, -1] and stride 2.
+Although in principle, we typically only do additional Haar wavelet over the LL component. But here we do it for all
+   as we need to support downsampling for more than 2x.
+For example, 4x downsampling can be done by 2x Haar and additional 2x Haar, and the shape would be.
+   [3, 256, 256] -> [12, 128, 128] -> [48, 64, 64]
+"""
+
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+
+_WAVELETS = {
+    "haar": torch.tensor([0.7071067811865476, 0.7071067811865476]),
+    "rearrange": torch.tensor([1.0, 1.0]),
+}
+_PERSISTENT = False
+
+
+class Patcher(torch.nn.Module):
+    """A module to convert image tensors into patches using torch operations.
+
+    The main difference from `class Patching` is that this module implements
+    all operations using torch, rather than python or numpy, for efficiency purpose.
+
+    It's bit-wise identical to the Patching module outputs, with the added
+    benefit of being torch.jit scriptable.
+    """
+
+    def __init__(self, patch_size=1, patch_method="haar"):
+        super().__init__()
+        self.patch_size = patch_size
+        self.patch_method = patch_method
+        self.register_buffer(
+            "wavelets", _WAVELETS[patch_method], persistent=_PERSISTENT
+        )
+        self.range = range(int(torch.log2(torch.tensor(self.patch_size)).item()))
+        self.register_buffer(
+            "_arange",
+            torch.arange(_WAVELETS[patch_method].shape[0]),
+            persistent=_PERSISTENT,
+        )
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, x):
+        if self.patch_method == "haar":
+            return self._haar(x)
+        elif self.patch_method == "rearrange":
+            return self._arrange(x)
+        else:
+            raise ValueError("Unknown patch method: " + self.patch_method)
+
+    def _dwt(self, x, mode="reflect", rescale=False):
+        dtype = x.dtype
+        h = self.wavelets.to(device=x.device)
+
+        n = h.shape[0]
+        g = x.shape[1]
+        hl = h.flip(0).reshape(1, 1, -1).repeat(g, 1, 1)
+        hh = (h * ((-1) ** self._arange.to(device=x.device))).reshape(1, 1, -1).repeat(g, 1, 1)
+        hh = hh.to(dtype=dtype)
+        hl = hl.to(dtype=dtype)
+
+        x = F.pad(x, pad=(n - 2, n - 1, n - 2, n - 1), mode=mode).to(dtype)
+        xl = F.conv2d(x, hl.unsqueeze(2), groups=g, stride=(1, 2))
+        xh = F.conv2d(x, hh.unsqueeze(2), groups=g, stride=(1, 2))
+        xll = F.conv2d(xl, hl.unsqueeze(3), groups=g, stride=(2, 1))
+        xlh = F.conv2d(xl, hh.unsqueeze(3), groups=g, stride=(2, 1))
+        xhl = F.conv2d(xh, hl.unsqueeze(3), groups=g, stride=(2, 1))
+        xhh = F.conv2d(xh, hh.unsqueeze(3), groups=g, stride=(2, 1))
+
+        out = torch.cat([xll, xlh, xhl, xhh], dim=1)
+        if rescale:
+            out = out / 2
+        return out
+
+    def _haar(self, x):
+        for _ in self.range:
+            x = self._dwt(x, rescale=True)
+        return x
+
+    def _arrange(self, x):
+        x = rearrange(
+            x,
+            "b c (h p1) (w p2) -> b (c p1 p2) h w",
+            p1=self.patch_size,
+            p2=self.patch_size,
+        ).contiguous()
+        return x
+
+
+class Patcher3D(Patcher):
+    """A 3D discrete wavelet transform for video data, expects 5D tensor, i.e. a batch of videos."""
+
+    def __init__(self, patch_size=1, patch_method="haar"):
+        super().__init__(patch_method=patch_method, patch_size=patch_size)
+        self.register_buffer(
+            "patch_size_buffer",
+            patch_size * torch.ones([1], dtype=torch.int32),
+            persistent=_PERSISTENT,
+        )
+
+    def _dwt(self, x, wavelet, mode="reflect", rescale=False):
+        dtype = x.dtype
+        h = self.wavelets.to(device=x.device)
+
+        n = h.shape[0]
+        g = x.shape[1]
+        hl = h.flip(0).reshape(1, 1, -1).repeat(g, 1, 1)
+        hh = (h * ((-1) ** self._arange.to(device=x.device))).reshape(1, 1, -1).repeat(g, 1, 1)
+        hh = hh.to(dtype=dtype)
+        hl = hl.to(dtype=dtype)
+
+        # Handles temporal axis.
+        x = F.pad(
+            x, pad=(max(0, n - 2), n - 1, n - 2, n - 1, n - 2, n - 1), mode=mode
+        ).to(dtype)
+        xl = F.conv3d(x, hl.unsqueeze(3).unsqueeze(4), groups=g, stride=(2, 1, 1))
+        xh = F.conv3d(x, hh.unsqueeze(3).unsqueeze(4), groups=g, stride=(2, 1, 1))
+
+        # Handles spatial axes.
+        xll = F.conv3d(xl, hl.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1))
+        xlh = F.conv3d(xl, hh.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1))
+        xhl = F.conv3d(xh, hl.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1))
+        xhh = F.conv3d(xh, hh.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1))
+
+        xlll = F.conv3d(xll, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
+        xllh = F.conv3d(xll, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
+        xlhl = F.conv3d(xlh, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
+        xlhh = F.conv3d(xlh, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
+        xhll = F.conv3d(xhl, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
+        xhlh = F.conv3d(xhl, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
+        xhhl = F.conv3d(xhh, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
+        xhhh = F.conv3d(xhh, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
+
+        out = torch.cat([xlll, xllh, xlhl, xlhh, xhll, xhlh, xhhl, xhhh], dim=1)
+        if rescale:
+            out = out / (2 * torch.sqrt(torch.tensor(2.0)))
+        return out
+
+    def _haar(self, x):
+        xi, xv = torch.split(x, [1, x.shape[2] - 1], dim=2)
+        x = torch.cat([xi.repeat_interleave(self.patch_size, dim=2), xv], dim=2)
+        for _ in self.range:
+            x = self._dwt(x, "haar", rescale=True)
+        return x
+
+    def _arrange(self, x):
+        xi, xv = torch.split(x, [1, x.shape[2] - 1], dim=2)
+        x = torch.cat([xi.repeat_interleave(self.patch_size, dim=2), xv], dim=2)
+        x = rearrange(
+            x,
+            "b c (t p1) (h p2) (w p3) -> b (c p1 p2 p3) t h w",
+            p1=self.patch_size,
+            p2=self.patch_size,
+            p3=self.patch_size,
+        ).contiguous()
+        return x
+
+
+class UnPatcher(torch.nn.Module):
+    """A module to convert patches into image tensorsusing torch operations.
+
+    The main difference from `class Unpatching` is that this module implements
+    all operations using torch, rather than python or numpy, for efficiency purpose.
+
+    It's bit-wise identical to the Unpatching module outputs, with the added
+    benefit of being torch.jit scriptable.
+    """
+
+    def __init__(self, patch_size=1, patch_method="haar"):
+        super().__init__()
+        self.patch_size = patch_size
+        self.patch_method = patch_method
+        self.register_buffer(
+            "wavelets", _WAVELETS[patch_method], persistent=_PERSISTENT
+        )
+        self.range = range(int(torch.log2(torch.tensor(self.patch_size)).item()))
+        self.register_buffer(
+            "_arange",
+            torch.arange(_WAVELETS[patch_method].shape[0]),
+            persistent=_PERSISTENT,
+        )
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, x):
+        if self.patch_method == "haar":
+            return self._ihaar(x)
+        elif self.patch_method == "rearrange":
+            return self._iarrange(x)
+        else:
+            raise ValueError("Unknown patch method: " + self.patch_method)
+
+    def _idwt(self, x, wavelet="haar", mode="reflect", rescale=False):
+        dtype = x.dtype
+        h = self.wavelets.to(device=x.device)
+        n = h.shape[0]
+
+        g = x.shape[1] // 4
+        hl = h.flip([0]).reshape(1, 1, -1).repeat([g, 1, 1])
+        hh = (h * ((-1) ** self._arange.to(device=x.device))).reshape(1, 1, -1).repeat(g, 1, 1)
+        hh = hh.to(dtype=dtype)
+        hl = hl.to(dtype=dtype)
+
+        xll, xlh, xhl, xhh = torch.chunk(x.to(dtype), 4, dim=1)
+
+        # Inverse transform.
+        yl = torch.nn.functional.conv_transpose2d(
+            xll, hl.unsqueeze(3), groups=g, stride=(2, 1), padding=(n - 2, 0)
+        )
+        yl += torch.nn.functional.conv_transpose2d(
+            xlh, hh.unsqueeze(3), groups=g, stride=(2, 1), padding=(n - 2, 0)
+        )
+        yh = torch.nn.functional.conv_transpose2d(
+            xhl, hl.unsqueeze(3), groups=g, stride=(2, 1), padding=(n - 2, 0)
+        )
+        yh += torch.nn.functional.conv_transpose2d(
+            xhh, hh.unsqueeze(3), groups=g, stride=(2, 1), padding=(n - 2, 0)
+        )
+        y = torch.nn.functional.conv_transpose2d(
+            yl, hl.unsqueeze(2), groups=g, stride=(1, 2), padding=(0, n - 2)
+        )
+        y += torch.nn.functional.conv_transpose2d(
+            yh, hh.unsqueeze(2), groups=g, stride=(1, 2), padding=(0, n - 2)
+        )
+
+        if rescale:
+            y = y * 2
+        return y
+
+    def _ihaar(self, x):
+        for _ in self.range:
+            x = self._idwt(x, "haar", rescale=True)
+        return x
+
+    def _iarrange(self, x):
+        x = rearrange(
+            x,
+            "b (c p1 p2) h w -> b c (h p1) (w p2)",
+            p1=self.patch_size,
+            p2=self.patch_size,
+        )
+        return x
+
+
+class UnPatcher3D(UnPatcher):
+    """A 3D inverse discrete wavelet transform for video wavelet decompositions."""
+
+    def __init__(self, patch_size=1, patch_method="haar"):
+        super().__init__(patch_method=patch_method, patch_size=patch_size)
+
+    def _idwt(self, x, wavelet="haar", mode="reflect", rescale=False):
+        dtype = x.dtype
+        h = self.wavelets.to(device=x.device)
+
+        g = x.shape[1] // 8  # split into 8 spatio-temporal filtered tesnors.
+        hl = h.flip([0]).reshape(1, 1, -1).repeat([g, 1, 1])
+        hh = (h * ((-1) ** self._arange.to(device=x.device))).reshape(1, 1, -1).repeat(g, 1, 1)
+        hl = hl.to(dtype=dtype)
+        hh = hh.to(dtype=dtype)
+
+        xlll, xllh, xlhl, xlhh, xhll, xhlh, xhhl, xhhh = torch.chunk(x, 8, dim=1)
+        del x
+
+        # Height height transposed convolutions.
+        xll = F.conv_transpose3d(
+            xlll, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2)
+        )
+        del xlll
+
+        xll += F.conv_transpose3d(
+            xllh, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2)
+        )
+        del xllh
+
+        xlh = F.conv_transpose3d(
+            xlhl, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2)
+        )
+        del xlhl
+
+        xlh += F.conv_transpose3d(
+            xlhh, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2)
+        )
+        del xlhh
+
+        xhl = F.conv_transpose3d(
+            xhll, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2)
+        )
+        del xhll
+
+        xhl += F.conv_transpose3d(
+            xhlh, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2)
+        )
+        del xhlh
+
+        xhh = F.conv_transpose3d(
+            xhhl, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2)
+        )
+        del xhhl
+
+        xhh += F.conv_transpose3d(
+            xhhh, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2)
+        )
+        del xhhh
+
+        # Handles width transposed convolutions.
+        xl = F.conv_transpose3d(
+            xll, hl.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1)
+        )
+        del xll
+
+        xl += F.conv_transpose3d(
+            xlh, hh.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1)
+        )
+        del xlh
+
+        xh = F.conv_transpose3d(
+            xhl, hl.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1)
+        )
+        del xhl
+
+        xh += F.conv_transpose3d(
+            xhh, hh.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1)
+        )
+        del xhh
+
+        # Handles time axis transposed convolutions.
+        x = F.conv_transpose3d(
+            xl, hl.unsqueeze(3).unsqueeze(4), groups=g, stride=(2, 1, 1)
+        )
+        del xl
+
+        x += F.conv_transpose3d(
+            xh, hh.unsqueeze(3).unsqueeze(4), groups=g, stride=(2, 1, 1)
+        )
+
+        if rescale:
+            x = x * (2 * torch.sqrt(torch.tensor(2.0)))
+        return x
+
+    def _ihaar(self, x):
+        for _ in self.range:
+            x = self._idwt(x, "haar", rescale=True)
+        x = x[:, :, self.patch_size - 1 :, ...]
+        return x
+
+    def _iarrange(self, x):
+        x = rearrange(
+            x,
+            "b (c p1 p2 p3) t h w -> b c (t p1) (h p2) (w p3)",
+            p1=self.patch_size,
+            p2=self.patch_size,
+            p3=self.patch_size,
+        )
+        x = x[:, :, self.patch_size - 1 :, ...]
+        return x

ComfyUI/comfy/ldm/cosmos/cosmos_tokenizer/utils.py

@@ -0,0 +1,113 @@
+# 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.
+"""Shared utilities for the networks module."""
+
+from typing import Any
+
+import torch
+from einops import rearrange
+
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+def time2batch(x: torch.Tensor) -> tuple[torch.Tensor, int]:
+    batch_size = x.shape[0]
+    return rearrange(x, "b c t h w -> (b t) c h w"), batch_size
+
+
+def batch2time(x: torch.Tensor, batch_size: int) -> torch.Tensor:
+    return rearrange(x, "(b t) c h w -> b c t h w", b=batch_size)
+
+
+def space2batch(x: torch.Tensor) -> tuple[torch.Tensor, int]:
+    batch_size, height = x.shape[0], x.shape[-2]
+    return rearrange(x, "b c t h w -> (b h w) c t"), batch_size, height
+
+
+def batch2space(x: torch.Tensor, batch_size: int, height: int) -> torch.Tensor:
+    return rearrange(x, "(b h w) c t -> b c t h w", b=batch_size, h=height)
+
+
+def cast_tuple(t: Any, length: int = 1) -> Any:
+    return t if isinstance(t, tuple) else ((t,) * length)
+
+
+def replication_pad(x):
+    return torch.cat([x[:, :, :1, ...], x], dim=2)
+
+
+def divisible_by(num: int, den: int) -> bool:
+    return (num % den) == 0
+
+
+def is_odd(n: int) -> bool:
+    return not divisible_by(n, 2)
+
+
+def nonlinearity(x):
+    # x * sigmoid(x)
+    return torch.nn.functional.silu(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return ops.GroupNorm(
+        num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True
+    )
+
+
+class CausalNormalize(torch.nn.Module):
+    def __init__(self, in_channels, num_groups=1):
+        super().__init__()
+        self.norm = ops.GroupNorm(
+            num_groups=num_groups,
+            num_channels=in_channels,
+            eps=1e-6,
+            affine=True,
+        )
+        self.num_groups = num_groups
+
+    def forward(self, x):
+        # if num_groups !=1, we apply a spatio-temporal groupnorm for backward compatibility purpose.
+        # All new models should use num_groups=1, otherwise causality is not guaranteed.
+        if self.num_groups == 1:
+            x, batch_size = time2batch(x)
+            return batch2time(self.norm(x), batch_size)
+        return self.norm(x)
+
+
+def exists(v):
+    return v is not None
+
+
+def default(*args):
+    for arg in args:
+        if exists(arg):
+            return arg
+    return None
+
+
+def round_ste(z: torch.Tensor) -> torch.Tensor:
+    """Round with straight through gradients."""
+    zhat = z.round()
+    return z + (zhat - z).detach()
+
+
+def log(t, eps=1e-5):
+    return t.clamp(min=eps).log()
+
+
+def entropy(prob):
+    return (-prob * log(prob)).sum(dim=-1)

ComfyUI/comfy/ldm/cosmos/model.py

@@ -0,0 +1,552 @@
+# 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.
+
+"""
+A general implementation of adaln-modulated VIT-like~(DiT) transformer for video processing.
+"""
+
+from typing import Optional, Tuple
+
+import torch
+from einops import rearrange
+from torch import nn
+from torchvision import transforms
+
+from enum import Enum
+import logging
+
+import comfy.patcher_extension
+
+from .blocks import (
+    FinalLayer,
+    GeneralDITTransformerBlock,
+    PatchEmbed,
+    TimestepEmbedding,
+    Timesteps,
+)
+
+from .position_embedding import LearnablePosEmbAxis, VideoRopePosition3DEmb
+
+
+class DataType(Enum):
+    IMAGE = "image"
+    VIDEO = "video"
+
+
+class GeneralDIT(nn.Module):
+    """
+    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.
+        block_config (str): Configuration of the transformer block. See Notes for supported block types.
+        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.
+        block_x_format (str): Format of input tensor for transformer blocks ('BTHWD' or 'THWBD').
+        crossattn_emb_channels (int): Number of embedding channels for cross-attention.
+        use_cross_attn_mask (bool): Whether to use mask in 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.
+        affline_emb_norm (bool): Whether to normalize affine embeddings.
+        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_per_block_abs_pos_emb_type (str): Type of extra per-block 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.
+
+    Notes:
+        Supported block types in block_config:
+        * cross_attn, ca: Cross attention
+        * full_attn: Full attention on all flattened tokens
+        * mlp, ff: Feed forward block
+    """
+
+    def __init__(
+        self,
+        max_img_h: int,
+        max_img_w: int,
+        max_frames: int,
+        in_channels: int,
+        out_channels: int,
+        patch_spatial: tuple,
+        patch_temporal: int,
+        concat_padding_mask: bool = True,
+        # attention settings
+        block_config: str = "FA-CA-MLP",
+        model_channels: int = 768,
+        num_blocks: int = 10,
+        num_heads: int = 16,
+        mlp_ratio: float = 4.0,
+        block_x_format: str = "BTHWD",
+        # cross attention settings
+        crossattn_emb_channels: int = 1024,
+        use_cross_attn_mask: bool = False,
+        # positional embedding settings
+        pos_emb_cls: str = "sincos",
+        pos_emb_learnable: bool = False,
+        pos_emb_interpolation: str = "crop",
+        affline_emb_norm: bool = False,  # whether or not to normalize the affine embedding
+        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_per_block_abs_pos_emb_type: str = "sincos",
+        extra_h_extrapolation_ratio: float = 1.0,
+        extra_w_extrapolation_ratio: float = 1.0,
+        extra_t_extrapolation_ratio: float = 1.0,
+        image_model=None,
+        device=None,
+        dtype=None,
+        operations=None,
+    ) -> None:
+        super().__init__()
+        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.use_cross_attn_mask = use_cross_attn_mask
+        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.affline_emb_norm = affline_emb_norm
+        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_per_block_abs_pos_emb_type = extra_per_block_abs_pos_emb_type.lower()
+        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.dtype = dtype
+        weight_args = {"device": device, "dtype": dtype}
+
+        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,
+            bias=False,
+            weight_args=weight_args,
+            operations=operations,
+        )
+
+        self.build_pos_embed(device=device, dtype=dtype)
+        self.block_x_format = block_x_format
+        self.use_adaln_lora = use_adaln_lora
+        self.adaln_lora_dim = adaln_lora_dim
+        self.t_embedder = nn.ModuleList(
+            [Timesteps(model_channels),
+             TimestepEmbedding(model_channels, model_channels, use_adaln_lora=use_adaln_lora, weight_args=weight_args, operations=operations),]
+        )
+
+        self.blocks = nn.ModuleDict()
+
+        for idx in range(num_blocks):
+            self.blocks[f"block{idx}"] = GeneralDITTransformerBlock(
+                x_dim=model_channels,
+                context_dim=crossattn_emb_channels,
+                num_heads=num_heads,
+                block_config=block_config,
+                mlp_ratio=mlp_ratio,
+                x_format=self.block_x_format,
+                use_adaln_lora=use_adaln_lora,
+                adaln_lora_dim=adaln_lora_dim,
+                weight_args=weight_args,
+                operations=operations,
+            )
+
+        if self.affline_emb_norm:
+            logging.debug("Building affine embedding normalization layer")
+            self.affline_norm = operations.RMSNorm(model_channels, elementwise_affine=True, eps=1e-6, device=device, dtype=dtype)
+        else:
+            self.affline_norm = nn.Identity()
+
+        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,
+            weight_args=weight_args,
+            operations=operations,
+        )
+
+    def build_pos_embed(self, device=None, dtype=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,
+            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,
+            device=device,
+        )
+        self.pos_embedder = cls_type(
+            **kwargs,
+        )
+
+        if self.extra_per_block_abs_pos_emb:
+            assert self.extra_per_block_abs_pos_emb_type in [
+                "learnable",
+            ], f"Unknown extra_per_block_abs_pos_emb_type {self.extra_per_block_abs_pos_emb_type}"
+            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,
+            )
+
+    def prepare_embedded_sequence(
+        self,
+        x_B_C_T_H_W: torch.Tensor,
+        fps: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        latent_condition: Optional[torch.Tensor] = None,
+        latent_condition_sigma: Optional[torch.Tensor] = None,
+    ) -> Tuple[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 not None:
+                padding_mask = transforms.functional.resize(
+                    padding_mask, list(x_B_C_T_H_W.shape[-2:]), interpolation=transforms.InterpolationMode.NEAREST
+                )
+            else:
+                padding_mask = torch.zeros((x_B_C_T_H_W.shape[0], 1, x_B_C_T_H_W.shape[-2], x_B_C_T_H_W.shape[-1]), dtype=x_B_C_T_H_W.dtype, device=x_B_C_T_H_W.device)
+
+            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
+
+        if "fps_aware" in self.pos_emb_cls:
+            x_B_T_H_W_D = 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)  # [B, T, H, W, D]
+        else:
+            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 decoder_head(
+        self,
+        x_B_T_H_W_D: torch.Tensor,
+        emb_B_D: torch.Tensor,
+        crossattn_emb: torch.Tensor,
+        origin_shape: Tuple[int, int, int, int, int],  # [B, C, T, H, W]
+        crossattn_mask: Optional[torch.Tensor] = None,
+        adaln_lora_B_3D: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        del crossattn_emb, crossattn_mask
+        B, C, T_before_patchify, H_before_patchify, W_before_patchify = origin_shape
+        x_BT_HW_D = rearrange(x_B_T_H_W_D, "B T H W D -> (B T) (H W) D")
+        x_BT_HW_D = self.final_layer(x_BT_HW_D, emb_B_D, adaln_lora_B_3D=adaln_lora_B_3D)
+        # This is to ensure x_BT_HW_D has the correct shape because
+        # when we merge T, H, W into one dimension, x_BT_HW_D has shape (B * T * H * W, 1*1, D).
+        x_BT_HW_D = x_BT_HW_D.view(
+            B * T_before_patchify // self.patch_temporal,
+            H_before_patchify // self.patch_spatial * W_before_patchify // self.patch_spatial,
+            -1,
+        )
+        x_B_D_T_H_W = rearrange(
+            x_BT_HW_D,
+            "(B T) (H W) (p1 p2 t C) -> B C (T t) (H p1) (W p2)",
+            p1=self.patch_spatial,
+            p2=self.patch_spatial,
+            H=H_before_patchify // self.patch_spatial,
+            W=W_before_patchify // self.patch_spatial,
+            t=self.patch_temporal,
+            B=B,
+        )
+        return x_B_D_T_H_W
+
+    def forward_before_blocks(
+        self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        crossattn_emb: torch.Tensor,
+        crossattn_mask: Optional[torch.Tensor] = None,
+        fps: Optional[torch.Tensor] = None,
+        image_size: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        scalar_feature: Optional[torch.Tensor] = None,
+        data_type: Optional[DataType] = DataType.VIDEO,
+        latent_condition: Optional[torch.Tensor] = None,
+        latent_condition_sigma: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        """
+        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
+            crossattn_mask: (B, N) tensor of cross-attention masks
+        """
+        del kwargs
+        assert isinstance(
+            data_type, DataType
+        ), f"Expected DataType, got {type(data_type)}. We need discuss this flag later."
+        original_shape = x.shape
+        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,
+            fps=fps,
+            padding_mask=padding_mask,
+            latent_condition=latent_condition,
+            latent_condition_sigma=latent_condition_sigma,
+        )
+        # logging affline scale information
+        affline_scale_log_info = {}
+
+        timesteps_B_D, adaln_lora_B_3D = self.t_embedder[1](self.t_embedder[0](timesteps.flatten()).to(x.dtype))
+        affline_emb_B_D = timesteps_B_D
+        affline_scale_log_info["timesteps_B_D"] = timesteps_B_D.detach()
+
+        if scalar_feature is not None:
+            raise NotImplementedError("Scalar feature is not implemented yet.")
+
+        affline_scale_log_info["affline_emb_B_D"] = affline_emb_B_D.detach()
+        affline_emb_B_D = self.affline_norm(affline_emb_B_D)
+
+        if self.use_cross_attn_mask:
+            if crossattn_mask is not None and not torch.is_floating_point(crossattn_mask):
+                crossattn_mask = (crossattn_mask - 1).to(x.dtype) * torch.finfo(x.dtype).max
+            crossattn_mask = crossattn_mask[:, None, None, :]  # .to(dtype=torch.bool)  # [B, 1, 1, length]
+        else:
+            crossattn_mask = None
+
+        if self.blocks["block0"].x_format == "THWBD":
+            x = rearrange(x_B_T_H_W_D, "B T H W D -> T H W B D")
+            if extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D is not None:
+                extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D = rearrange(
+                    extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D, "B T H W D -> T H W B D"
+                )
+            crossattn_emb = rearrange(crossattn_emb, "B M D -> M B D")
+
+            if crossattn_mask:
+                crossattn_mask = rearrange(crossattn_mask, "B M -> M B")
+
+        elif self.blocks["block0"].x_format == "BTHWD":
+            x = x_B_T_H_W_D
+        else:
+            raise ValueError(f"Unknown x_format {self.blocks[0].x_format}")
+        output = {
+            "x": x,
+            "affline_emb_B_D": affline_emb_B_D,
+            "crossattn_emb": crossattn_emb,
+            "crossattn_mask": crossattn_mask,
+            "rope_emb_L_1_1_D": rope_emb_L_1_1_D,
+            "adaln_lora_B_3D": adaln_lora_B_3D,
+            "original_shape": original_shape,
+            "extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D": extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D,
+        }
+        return output
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        context: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        # crossattn_emb: torch.Tensor,
+        # crossattn_mask: Optional[torch.Tensor] = None,
+        fps: Optional[torch.Tensor] = None,
+        image_size: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        scalar_feature: Optional[torch.Tensor] = None,
+        data_type: Optional[DataType] = DataType.VIDEO,
+        latent_condition: Optional[torch.Tensor] = None,
+        latent_condition_sigma: Optional[torch.Tensor] = None,
+        condition_video_augment_sigma: 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,
+                attention_mask,
+                fps,
+                image_size,
+                padding_mask,
+                scalar_feature,
+                data_type,
+                latent_condition,
+                latent_condition_sigma,
+                condition_video_augment_sigma,
+                **kwargs)
+
+    def _forward(
+        self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        context: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        # crossattn_emb: torch.Tensor,
+        # crossattn_mask: Optional[torch.Tensor] = None,
+        fps: Optional[torch.Tensor] = None,
+        image_size: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        scalar_feature: Optional[torch.Tensor] = None,
+        data_type: Optional[DataType] = DataType.VIDEO,
+        latent_condition: Optional[torch.Tensor] = None,
+        latent_condition_sigma: Optional[torch.Tensor] = None,
+        condition_video_augment_sigma: Optional[torch.Tensor] = None,
+        **kwargs,
+    ):
+        """
+        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
+            crossattn_mask: (B, N) tensor of cross-attention masks
+            condition_video_augment_sigma: (B,) used in lvg(long video generation), we add noise with this sigma to
+                augment condition input, the lvg model will condition on the condition_video_augment_sigma value;
+                we need forward_before_blocks pass to the forward_before_blocks function.
+        """
+
+        crossattn_emb = context
+        crossattn_mask = attention_mask
+
+        inputs = self.forward_before_blocks(
+            x=x,
+            timesteps=timesteps,
+            crossattn_emb=crossattn_emb,
+            crossattn_mask=crossattn_mask,
+            fps=fps,
+            image_size=image_size,
+            padding_mask=padding_mask,
+            scalar_feature=scalar_feature,
+            data_type=data_type,
+            latent_condition=latent_condition,
+            latent_condition_sigma=latent_condition_sigma,
+            condition_video_augment_sigma=condition_video_augment_sigma,
+            **kwargs,
+        )
+        x, affline_emb_B_D, crossattn_emb, crossattn_mask, rope_emb_L_1_1_D, adaln_lora_B_3D, original_shape = (
+            inputs["x"],
+            inputs["affline_emb_B_D"],
+            inputs["crossattn_emb"],
+            inputs["crossattn_mask"],
+            inputs["rope_emb_L_1_1_D"],
+            inputs["adaln_lora_B_3D"],
+            inputs["original_shape"],
+        )
+        extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D = inputs["extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D"].to(x.dtype)
+        del inputs
+
+        if extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D is not None:
+            assert (
+                x.shape == extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D.shape
+            ), f"{x.shape} != {extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D.shape} {original_shape}"
+
+        transformer_options = kwargs.get("transformer_options", {})
+        for _, block in self.blocks.items():
+            assert (
+                self.blocks["block0"].x_format == block.x_format
+            ), f"First block has x_format {self.blocks[0].x_format}, got {block.x_format}"
+
+            if extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D is not None:
+                x += extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D
+            x = block(
+                x,
+                affline_emb_B_D,
+                crossattn_emb,
+                crossattn_mask,
+                rope_emb_L_1_1_D=rope_emb_L_1_1_D,
+                adaln_lora_B_3D=adaln_lora_B_3D,
+                transformer_options=transformer_options,
+            )
+
+        x_B_T_H_W_D = rearrange(x, "T H W B D -> B T H W D")
+
+        x_B_D_T_H_W = self.decoder_head(
+            x_B_T_H_W_D=x_B_T_H_W_D,
+            emb_B_D=affline_emb_B_D,
+            crossattn_emb=None,
+            origin_shape=original_shape,
+            crossattn_mask=None,
+            adaln_lora_B_3D=adaln_lora_B_3D,
+        )
+
+        return x_B_D_T_H_W

ComfyUI/comfy/ldm/cosmos/position_embedding.py

@@ -0,0 +1,207 @@
+# 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)

ComfyUI/comfy/ldm/cosmos/predict2.py

@@ -0,0 +1,932 @@
+# 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
+import comfy.ldm.common_dit
+import comfy.ops
+import comfy.quant_ops
+
+
+# ---------------------- 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, transformer_options: Optional[dict] = {}) -> 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, transformer_options=transformer_options)
+
+
+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,
+        transformer_options: Optional[dict] = {},
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        context = x if context is None else context
+        q_input = x
+        k_input = context
+        v_input = context
+
+        transformer_patches = transformer_options.get("patches", {})
+        patch_name = "attn1_patch" if self.is_selfattn else "attn2_patch"
+        if patch_name in transformer_patches:
+            extra_options = transformer_options.copy()
+            extra_options["n_heads"] = self.n_heads
+            extra_options["dim_head"] = self.head_dim
+            for patch in transformer_patches[patch_name]:
+                out = patch(q_input, k_input, v_input, pe=rope_emb, attn_mask=None, extra_options=extra_options)
+                q_input = out.get("q", q_input)
+                k_input = out.get("k", k_input)
+                v_input = out.get("v", v_input)
+                rope_emb = out.get("pe", rope_emb)
+
+        q = self.q_proj(q_input)
+        k = self.k_proj(k_input)
+        v = self.v_proj(v_input)
+        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]:
+            v = self.v_norm(v)
+            if self.is_selfattn and rope_emb is not None:  # only apply to self-attention!
+                q_scale, _, q_offload_stream = comfy.ops.cast_bias_weight(self.q_norm, q, offloadable=True)
+                k_scale, _, k_offload_stream = comfy.ops.cast_bias_weight(self.k_norm, k, offloadable=True)
+                q, k = comfy.quant_ops.ck.rms_rope_split_half(q, k, rope_emb, q_scale, k_scale, self.q_norm.eps)
+                comfy.ops.uncast_bias_weight(self.q_norm, q_scale, None, q_offload_stream)
+                comfy.ops.uncast_bias_weight(self.k_norm, k_scale, None, k_offload_stream)
+            else:
+                q = self.q_norm(q)
+                k = self.k_norm(k)
+            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, transformer_options: Optional[dict] = {}) -> torch.Tensor:
+        result = self.attn_op(q, k, v, transformer_options=transformer_options)  # [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,
+        transformer_options: Optional[dict] = {},
+    ) -> 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, transformer_options=transformer_options)
+        return self.compute_attention(q, k, v, transformer_options=transformer_options)
+
+
+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 = operations.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,
+        transformer_options: Optional[dict] = {},
+    ) -> torch.Tensor:
+        residual_dtype = x_B_T_H_W_D.dtype
+        compute_dtype = emb_B_T_D.dtype
+        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.to(compute_dtype), "b t h w d -> b (t h w) d"),
+                None,
+                rope_emb=rope_emb_L_1_1_D,
+                transformer_options=transformer_options,
+            ),
+            "b (t h w) d -> b t h w d",
+            t=T,
+            h=H,
+            w=W,
+        )
+        x_B_T_H_W_D = torch.addcmul(x_B_T_H_W_D, gate_self_attn_B_T_1_1_D.to(residual_dtype), result_B_T_H_W_D.to(residual_dtype))
+
+        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,
+            transformer_options: Optional[dict] = {},
+        ) -> 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.to(compute_dtype), "b t h w d -> b (t h w) d"),
+                    crossattn_emb,
+                    rope_emb=rope_emb_L_1_1_D,
+                    transformer_options=transformer_options,
+                ),
+                "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,
+            transformer_options=transformer_options,
+        )
+        x_B_T_H_W_D = torch.addcmul(x_B_T_H_W_D, gate_cross_attn_B_T_1_1_D.to(residual_dtype), result_B_T_H_W_D.to(residual_dtype))
+
+        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,
+        ).to(compute_dtype)
+        patches = transformer_options.get("patches", {})
+        if "mlp_patch" in patches:
+            args = {"x": normalized_x_B_T_H_W_D, "transformer_options": transformer_options}
+            for patch in patches["mlp_patch"]:
+                args = patch(args)
+            normalized_x_B_T_H_W_D = args["x"]
+        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D)
+        x_B_T_H_W_D = torch.addcmul(x_B_T_H_W_D, gate_mlp_B_T_1_1_D.to(residual_dtype), result_B_T_H_W_D.to(residual_dtype))
+        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,
+    ):
+        orig_shape = list(x.shape)
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_temporal, self.patch_spatial, self.patch_spatial))
+        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}"
+
+        transformer_options = kwargs.get("transformer_options", {})
+        patches = transformer_options.get("patches", {})
+        if "post_input" in patches:
+            transformer_options = transformer_options.copy()
+            transformer_options["model_patch_data"] = {}
+
+        if "post_input" in patches:
+            for patch in patches["post_input"]:
+                out = patch({"img": x_B_T_H_W_D, "x": x_B_C_T_H_W, "transformer_options": transformer_options})
+                x_B_T_H_W_D = out["img"]
+
+        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,
+            "transformer_options": transformer_options,
+        }
+
+        # The residual stream for this model has large values. To make fp16 compute_dtype work, we keep the residual stream
+        # in fp32, but run attention and MLP modules in fp16.
+        # An alternate method that clamps fp16 values "works" in the sense that it makes coherent images, but there is noticeable
+        # quality degradation and visual artifacts.
+        if x_B_T_H_W_D.dtype == torch.float16:
+            x_B_T_H_W_D = x_B_T_H_W_D.float()
+
+        for block_index, block in enumerate(self.blocks):
+            transformer_options["block_index"] = block_index
+            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.to(crossattn_emb.dtype), 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)[:, :, :orig_shape[-3], :orig_shape[-2], :orig_shape[-1]]
+        return x_B_C_Tt_Hp_Wp

ComfyUI/comfy/ldm/cosmos/vae.py

@@ -0,0 +1,131 @@
+# 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)
+

ComfyUI/comfy/ldm/depth_anything_3/camera.py

@@ -0,0 +1,177 @@
+"""Camera-token encoder and decoder for Depth Anything 3."""
+
+from __future__ import annotations
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from comfy.ldm.modules.attention import optimized_attention_for_device
+from .transform import affine_inverse, extri_intri_to_pose_encoding
+
+
+# -----------------------------------------------------------------------
+# Building blocks (mirror depth_anything_3.model.utils.{attention,block})
+# -----------------------------------------------------------------------
+
+
+class _Mlp(nn.Module):
+    """Standard 2-layer MLP with GELU. Matches upstream ``utils.attention.Mlp``."""
+
+    def __init__(self, in_features, hidden_features=None, out_features=None, *, device=None, dtype=None, operations=None):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = operations.Linear(in_features, hidden_features, bias=True, device=device, dtype=dtype)
+        self.fc2 = operations.Linear(hidden_features, out_features, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        return self.fc2(F.gelu(self.fc1(x)))
+
+
+class _LayerScale(nn.Module):
+    """Per-channel learnable scaling. Matches upstream LayerScale."""
+
+    def __init__(self, dim, *, device=None, dtype=None):
+        super().__init__()
+        self.gamma = nn.Parameter(torch.empty(dim, device=device, dtype=dtype))
+
+    def forward(self, x):
+        return x * self.gamma.to(dtype=x.dtype, device=x.device)
+
+
+class _Attention(nn.Module):
+    """ Self-attention with fused QKV projection. Mirrors upstream utils.attention.Attention;
+    Layout matches the HF safetensors (attn.qkv.{weight,bias} and attn.proj.{weight,bias})."""
+
+    def __init__(self, dim, num_heads, *, device=None, dtype=None, operations=None):
+        super().__init__()
+        assert dim % num_heads == 0
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.qkv = operations.Linear(dim, dim * 3, bias=True, device=device, dtype=dtype)
+        self.proj = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, C)
+        q, k, v = qkv.unbind(2)                      # each (B, N, C)
+        attn_fn = optimized_attention_for_device(x.device, small_input=True)
+        out = attn_fn(q, k, v, heads=self.num_heads)
+        return self.proj(out)
+
+
+class _Block(nn.Module):
+    """Pre-norm transformer block with LayerScale. Used by :class:CameraEnc. Layout follows upstream utils.block.Block."""
+
+    def __init__(self, dim, num_heads, mlp_ratio=4, init_values=0.01, *, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.norm1 = operations.LayerNorm(dim, device=device, dtype=dtype)
+        self.attn = _Attention(dim, num_heads, device=device, dtype=dtype, operations=operations)
+        self.ls1 = _LayerScale(dim, device=device, dtype=dtype) if init_values else nn.Identity()
+        self.norm2 = operations.LayerNorm(dim, device=device, dtype=dtype)
+        self.mlp = _Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), device=device, dtype=dtype, operations=operations)
+        self.ls2 = _LayerScale(dim, device=device, dtype=dtype) if init_values else nn.Identity()
+
+    def forward(self, x):
+        x = x + self.ls1(self.attn(self.norm1(x)))
+        x = x + self.ls2(self.mlp(self.norm2(x)))
+        return x
+
+
+class CameraEnc(nn.Module):
+    """Encode per-view (extrinsics, intrinsics) into a camera token.
+
+    Maps a 9-D pose-encoding vector through a small MLP up to the backbone's
+    ``embed_dim``, then runs ``trunk_depth`` transformer blocks. The output
+    has shape ``(B, S, embed_dim)`` and is injected at block ``alt_start``
+    of the DINOv2 backbone in place of the cls token.
+
+    Parameters mirror the upstream ``cam_enc.py`` so HF weights load directly.
+    """
+
+    def __init__(
+        self,
+        dim_out: int = 1024,
+        dim_in: int = 9,
+        trunk_depth: int = 4,
+        target_dim: int = 9,
+        num_heads: int = 16,
+        mlp_ratio: int = 4,
+        init_values: float = 0.01,
+        *,
+        device=None, dtype=None, operations=None,
+        **_kwargs,
+    ):
+        super().__init__()
+        self.target_dim = target_dim
+        self.trunk_depth = trunk_depth
+        self.trunk = nn.Sequential(*[
+            _Block(dim_out, num_heads=num_heads, mlp_ratio=mlp_ratio,
+                   init_values=init_values,
+                   device=device, dtype=dtype, operations=operations)
+            for _ in range(trunk_depth)
+        ])
+        self.token_norm = operations.LayerNorm(dim_out, device=device, dtype=dtype)
+        self.trunk_norm = operations.LayerNorm(dim_out, device=device, dtype=dtype)
+        self.pose_branch = _Mlp(
+            in_features=dim_in,
+            hidden_features=dim_out // 2,
+            out_features=dim_out,
+            device=device, dtype=dtype, operations=operations,
+        )
+
+    def forward(self, extrinsics: torch.Tensor, intrinsics: torch.Tensor,
+                image_size_hw) -> torch.Tensor:
+        """Encode camera parameters into ``(B, S, dim_out)`` tokens."""
+        c2ws = affine_inverse(extrinsics)
+        pose_encoding = extri_intri_to_pose_encoding(c2ws, intrinsics, image_size_hw)
+        tokens = self.pose_branch(pose_encoding.to(self.pose_branch.fc1.weight.dtype))
+        tokens = self.token_norm(tokens)
+        tokens = self.trunk(tokens)
+        tokens = self.trunk_norm(tokens)
+        return tokens
+
+
+class CameraDec(nn.Module):
+    """Decode the final cam token into a 9-D pose encoding.
+
+    Output layout: ``[T(3), quat_xyzw(4), fov_h, fov_w]``. The translation is
+    always predicted by the network; the quaternion and FoV can either be
+    predicted or supplied via ``camera_encoding`` (used at training time
+    when GT cameras are available -- not exercised at inference here).
+
+    Parameters mirror the upstream ``cam_dec.py`` so HF weights load directly.
+    """
+
+    def __init__(self, dim_in: int = 1536,
+                 *, device=None, dtype=None, operations=None, **_kwargs):
+        super().__init__()
+        d = dim_in
+        self.backbone = nn.Sequential(
+            operations.Linear(d, d, device=device, dtype=dtype),
+            nn.ReLU(),
+            operations.Linear(d, d, device=device, dtype=dtype),
+            nn.ReLU(),
+        )
+        self.fc_t = operations.Linear(d, 3, device=device, dtype=dtype)
+        self.fc_qvec = operations.Linear(d, 4, device=device, dtype=dtype)
+        self.fc_fov = nn.Sequential(
+            operations.Linear(d, 2, device=device, dtype=dtype),
+            nn.ReLU(),
+        )
+
+    def forward(self, feat: torch.Tensor,
+                camera_encoding: "torch.Tensor | None" = None) -> torch.Tensor:
+        """Decode ``(B, N, dim_in)`` cam tokens into ``(B, N, 9)`` pose enc."""
+        B, N = feat.shape[:2]
+        feat = feat.reshape(B * N, -1)
+        feat = self.backbone(feat)
+        out_t = self.fc_t(feat.float()).reshape(B, N, 3)
+        if camera_encoding is None:
+            out_qvec = self.fc_qvec(feat.float()).reshape(B, N, 4)
+            out_fov = self.fc_fov(feat.float()).reshape(B, N, 2)
+        else:
+            out_qvec = camera_encoding[..., 3:7]
+            out_fov = camera_encoding[..., -2:]
+        return torch.cat([out_t, out_qvec, out_fov], dim=-1)

ComfyUI/comfy/ldm/depth_anything_3/dpt.py

@@ -0,0 +1,489 @@
+"""DPT / DualDPT heads for Depth Anything 3."""
+
+from __future__ import annotations
+
+from typing import List, Optional, Sequence, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Permute(nn.Module):
+    def __init__(self, dims: Tuple[int, ...]):
+        super().__init__()
+        self.dims = dims
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x.permute(*self.dims)
+
+
+def _custom_interpolate(
+    x: torch.Tensor,
+    size: Optional[Tuple[int, int]] = None,
+    scale_factor: Optional[float] = None,
+    mode: str = "bilinear",
+    align_corners: bool = True,
+) -> torch.Tensor:
+    if size is None:
+        assert scale_factor is not None
+        size = (int(x.shape[-2] * scale_factor), int(x.shape[-1] * scale_factor))
+    INT_MAX = 1610612736
+    total = size[0] * size[1] * x.shape[0] * x.shape[1]
+    if total > INT_MAX:
+        chunks = torch.chunk(x, chunks=(total // INT_MAX) + 1, dim=0)
+        outs = [F.interpolate(c, size=size, mode=mode, align_corners=align_corners) for c in chunks]
+        return torch.cat(outs, dim=0).contiguous()
+    return F.interpolate(x, size=size, mode=mode, align_corners=align_corners)
+
+
+def _create_uv_grid(width: int, height: int, aspect_ratio: float, dtype, device) -> torch.Tensor:
+    """Normalised UV grid spanning (-x_span, -y_span)..(x_span, y_span)."""
+    diag_factor = (aspect_ratio ** 2 + 1.0) ** 0.5
+    span_x = aspect_ratio / diag_factor
+    span_y = 1.0 / diag_factor
+    left_x = -span_x * (width - 1) / width
+    right_x = span_x * (width - 1) / width
+    top_y = -span_y * (height - 1) / height
+    bottom_y = span_y * (height - 1) / height
+    x_coords = torch.linspace(left_x, right_x, steps=width, dtype=dtype, device=device)
+    y_coords = torch.linspace(top_y, bottom_y, steps=height, dtype=dtype, device=device)
+    uu, vv = torch.meshgrid(x_coords, y_coords, indexing="xy")
+    return torch.stack((uu, vv), dim=-1)  # (H, W, 2)
+
+
+def _make_sincos_pos_embed(embed_dim: int, pos: torch.Tensor, omega_0: float = 100.0) -> torch.Tensor:
+    omega = torch.arange(embed_dim // 2, dtype=torch.float32, device=pos.device)
+    omega = 1.0 / omega_0 ** (omega / (embed_dim / 2.0))
+    pos = pos.reshape(-1)
+    out = torch.einsum("m,d->md", pos, omega)
+    return torch.cat([out.sin(), out.cos()], dim=1).float()
+
+
+def _position_grid_to_embed(pos_grid: torch.Tensor, embed_dim: int, omega_0: float = 100.0) -> torch.Tensor:
+    H, W, _ = pos_grid.shape
+    pos_flat = pos_grid.reshape(-1, 2)
+    emb_x = _make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 0], omega_0=omega_0)
+    emb_y = _make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 1], omega_0=omega_0)
+    emb = torch.cat([emb_x, emb_y], dim=-1)
+    return emb.view(H, W, embed_dim)
+
+
+def _add_pos_embed(x: torch.Tensor, W: int, H: int, ratio: float = 0.1) -> torch.Tensor:
+    """Stateless UV positional embedding added to a feature map (B, C, h, w)."""
+    pw, ph = x.shape[-1], x.shape[-2]
+    pe = _create_uv_grid(pw, ph, aspect_ratio=W / H, dtype=x.dtype, device=x.device)
+    pe = _position_grid_to_embed(pe, x.shape[1]) * ratio
+    pe = pe.permute(2, 0, 1)[None].expand(x.shape[0], -1, -1, -1).to(dtype=x.dtype)
+    return x + pe
+
+
+def _apply_activation(x: torch.Tensor, activation: str) -> torch.Tensor:
+    act = (activation or "linear").lower()
+    if act == "exp":
+        return torch.exp(x)
+    if act == "expp1":
+        return torch.exp(x) + 1
+    if act == "expm1":
+        return torch.expm1(x)
+    if act == "relu":
+        return torch.relu(x)
+    if act == "sigmoid":
+        return torch.sigmoid(x)
+    if act == "softplus":
+        return F.softplus(x)
+    if act == "tanh":
+        return torch.tanh(x)
+    return x
+
+
+# -----------------------------------------------------------------------------
+# Fusion building blocks
+# -----------------------------------------------------------------------------
+
+
+class ResidualConvUnit(nn.Module):
+    def __init__(self, features: int, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.conv1 = operations.Conv2d(features, features, 3, 1, 1, bias=True, device=device, dtype=dtype)
+        self.conv2 = operations.Conv2d(features, features, 3, 1, 1, bias=True, device=device, dtype=dtype)
+        self.activation = nn.ReLU(inplace=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = self.activation(x)
+        out = self.conv1(out)
+        out = self.activation(out)
+        out = self.conv2(out)
+        return out + x
+
+
+class FeatureFusionBlock(nn.Module):
+    def __init__(self, features: int, has_residual: bool = True, align_corners: bool = True, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.align_corners = align_corners
+        self.has_residual = has_residual
+        if has_residual:
+            self.resConfUnit1 = ResidualConvUnit(features, device=device, dtype=dtype, operations=operations)
+        else:
+            self.resConfUnit1 = None
+        self.resConfUnit2 = ResidualConvUnit(features, device=device, dtype=dtype, operations=operations)
+        self.out_conv = operations.Conv2d(features, features, 1, 1, 0, bias=True, device=device, dtype=dtype)
+
+    def forward(self, *xs: torch.Tensor, size: Optional[Tuple[int, int]] = None) -> torch.Tensor:
+        y = xs[0]
+        if self.has_residual and len(xs) > 1 and self.resConfUnit1 is not None:
+            y = y + self.resConfUnit1(xs[1])
+        y = self.resConfUnit2(y)
+        if size is None:
+            up_kwargs = {"scale_factor": 2.0}
+        else:
+            up_kwargs = {"size": size}
+        y = _custom_interpolate(y, **up_kwargs, mode="bilinear", align_corners=self.align_corners)
+        y = self.out_conv(y)
+        return y
+
+
+class _Scratch(nn.Module):
+    """Container that mirrors upstream ``scratch`` attribute layout."""
+
+
+def _make_scratch(in_shape: List[int], out_shape: int, device=None, dtype=None, operations=None) -> _Scratch:
+    scratch = _Scratch()
+    scratch.layer1_rn = operations.Conv2d(in_shape[0], out_shape, 3, 1, 1, bias=False, device=device, dtype=dtype)
+    scratch.layer2_rn = operations.Conv2d(in_shape[1], out_shape, 3, 1, 1, bias=False, device=device, dtype=dtype)
+    scratch.layer3_rn = operations.Conv2d(in_shape[2], out_shape, 3, 1, 1, bias=False, device=device, dtype=dtype)
+    scratch.layer4_rn = operations.Conv2d(in_shape[3], out_shape, 3, 1, 1, bias=False, device=device, dtype=dtype)
+    return scratch
+
+
+def _make_fusion_block(features: int, has_residual: bool = True, device=None, dtype=None, operations=None) -> FeatureFusionBlock:
+    return FeatureFusionBlock(features, has_residual=has_residual, align_corners=True, device=device, dtype=dtype, operations=operations)
+
+
+# -----------------------------------------------------------------------------
+# DPT (single head + optional sky head) -- used by DA3Mono/Metric
+# -----------------------------------------------------------------------------
+
+
+class DPT(nn.Module):
+    """Single-head DPT used by DA3Mono-Large and DA3Metric-Large."""
+
+    def __init__(
+        self,
+        dim_in: int,
+        patch_size: int = 14,
+        output_dim: int = 1,
+        activation: str = "exp",
+        conf_activation: str = "expp1",
+        features: int = 256,
+        out_channels: Sequence[int] = (256, 512, 1024, 1024),
+        pos_embed: bool = False,
+        down_ratio: int = 1,
+        head_name: str = "depth",
+        use_sky_head: bool = True,
+        sky_name: str = "sky",
+        sky_activation: str = "relu",
+        norm_type: str = "idt",
+        device=None, dtype=None, operations=None,
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.activation = activation
+        self.conf_activation = conf_activation
+        self.pos_embed = pos_embed
+        self.down_ratio = down_ratio
+        self.head_main = head_name
+        self.sky_name = sky_name
+        self.out_dim = output_dim
+        self.has_conf = output_dim > 1
+        self.use_sky_head = use_sky_head
+        self.sky_activation = sky_activation
+        self.intermediate_layer_idx: Tuple[int, int, int, int] = (0, 1, 2, 3)
+
+        if norm_type == "layer":
+            self.norm = operations.LayerNorm(dim_in, device=device, dtype=dtype)
+        else:
+            self.norm = nn.Identity()
+
+        out_channels = list(out_channels)
+        self.projects = nn.ModuleList([
+            operations.Conv2d(dim_in, oc, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype)
+            for oc in out_channels
+        ])
+        self.resize_layers = nn.ModuleList([
+            operations.ConvTranspose2d(out_channels[0], out_channels[0], kernel_size=4, stride=4, padding=0, device=device, dtype=dtype),
+            operations.ConvTranspose2d(out_channels[1], out_channels[1], kernel_size=2, stride=2, padding=0, device=device, dtype=dtype),
+            nn.Identity(),
+            operations.Conv2d(out_channels[3], out_channels[3], kernel_size=3, stride=2, padding=1, device=device, dtype=dtype),
+        ])
+
+        self.scratch = _make_scratch(out_channels, features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet1 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet2 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet3 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet4 = _make_fusion_block(features, has_residual=False, device=device, dtype=dtype, operations=operations)
+
+        head_features_1 = features
+        head_features_2 = 32
+        self.scratch.output_conv1 = operations.Conv2d(
+            head_features_1, head_features_1 // 2, kernel_size=3, stride=1, padding=1,
+            device=device, dtype=dtype,
+        )
+        self.scratch.output_conv2 = nn.Sequential(
+            operations.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1, device=device, dtype=dtype),
+            nn.ReLU(inplace=False),
+            operations.Conv2d(head_features_2, output_dim, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype),
+        )
+
+        if self.use_sky_head:
+            self.scratch.sky_output_conv2 = nn.Sequential(
+                operations.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1, device=device, dtype=dtype),
+                nn.ReLU(inplace=False),
+                operations.Conv2d(head_features_2, 1, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype),
+            )
+
+    def forward(self, feats: List[torch.Tensor], H: int, W: int, patch_start_idx: int = 0, **_kwargs) -> dict:
+        # feats[i][0] is the patch-token tensor with shape (B, S, N_patch, C)
+        B, S, N, C = feats[0][0].shape
+        feats_flat = [feat[0].reshape(B * S, N, C) for feat in feats]
+
+        ph, pw = H // self.patch_size, W // self.patch_size
+        resized = []
+        for stage_idx, take_idx in enumerate(self.intermediate_layer_idx):
+            x = feats_flat[take_idx][:, patch_start_idx:]
+            x = self.norm(x)
+            x = x.permute(0, 2, 1).contiguous().reshape(B * S, C, ph, pw)
+            x = self.projects[stage_idx](x)
+            if self.pos_embed:
+                x = _add_pos_embed(x, W, H)
+            x = self.resize_layers[stage_idx](x)
+            resized.append(x)
+
+        l1_rn = self.scratch.layer1_rn(resized[0])
+        l2_rn = self.scratch.layer2_rn(resized[1])
+        l3_rn = self.scratch.layer3_rn(resized[2])
+        l4_rn = self.scratch.layer4_rn(resized[3])
+
+        out = self.scratch.refinenet4(l4_rn, size=l3_rn.shape[2:])
+        out = self.scratch.refinenet3(out, l3_rn, size=l2_rn.shape[2:])
+        out = self.scratch.refinenet2(out, l2_rn, size=l1_rn.shape[2:])
+        out = self.scratch.refinenet1(out, l1_rn)
+
+        h_out = int(ph * self.patch_size / self.down_ratio)
+        w_out = int(pw * self.patch_size / self.down_ratio)
+
+        fused = self.scratch.output_conv1(out)
+        fused = _custom_interpolate(fused, (h_out, w_out), mode="bilinear", align_corners=True)
+        if self.pos_embed:
+            fused = _add_pos_embed(fused, W, H)
+        feat = fused
+
+        main_logits = self.scratch.output_conv2(feat)
+        outs = {}
+        if self.has_conf:
+            fmap = main_logits.permute(0, 2, 3, 1)
+            pred = _apply_activation(fmap[..., :-1], self.activation)
+            conf = _apply_activation(fmap[..., -1], self.conf_activation)
+            outs[self.head_main] = pred.squeeze(-1).view(B, S, *pred.shape[1:-1])
+            outs[f"{self.head_main}_conf"] = conf.view(B, S, *conf.shape[1:])
+        else:
+            pred = _apply_activation(main_logits, self.activation)
+            outs[self.head_main] = pred.squeeze(1).view(B, S, *pred.shape[2:])
+
+        if self.use_sky_head:
+            sky_logits = self.scratch.sky_output_conv2(feat)
+            if self.sky_activation.lower() == "sigmoid":
+                sky = torch.sigmoid(sky_logits)
+            elif self.sky_activation.lower() == "relu":
+                sky = F.relu(sky_logits)
+            else:
+                sky = sky_logits
+            outs[self.sky_name] = sky.squeeze(1).view(B, S, *sky.shape[2:])
+
+        return outs
+
+
+# -----------------------------------------------------------------------------
+# DualDPT (depth + auxiliary "ray" head) -- used by DA3-Small / DA3-Base
+# -----------------------------------------------------------------------------
+
+
+class DualDPT(nn.Module):
+    """Two-head DPT used by DA3-Small / DA3-Base."""
+
+    def __init__(
+        self,
+        dim_in: int,
+        patch_size: int = 14,
+        output_dim: int = 2,
+        activation: str = "exp",
+        conf_activation: str = "expp1",
+        features: int = 256,
+        out_channels: Sequence[int] = (256, 512, 1024, 1024),
+        pos_embed: bool = True,
+        down_ratio: int = 1,
+        aux_pyramid_levels: int = 4,
+        aux_out1_conv_num: int = 5,
+        head_names: Tuple[str, str] = ("depth", "ray"),
+        device=None, dtype=None, operations=None,
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.activation = activation
+        self.conf_activation = conf_activation
+        self.pos_embed = pos_embed
+        self.down_ratio = down_ratio
+        self.aux_levels = aux_pyramid_levels
+        self.aux_out1_conv_num = aux_out1_conv_num
+        self.head_main, self.head_aux = head_names
+        self.intermediate_layer_idx: Tuple[int, int, int, int] = (0, 1, 2, 3)
+        # Toggle the auxiliary ray branch at runtime. Default off (mono path).
+        # DepthAnything3Net flips this on when running multi-view + ray-pose.
+        self.enable_aux: bool = False
+
+        self.norm = operations.LayerNorm(dim_in, device=device, dtype=dtype)
+        out_channels = list(out_channels)
+        self.projects = nn.ModuleList([
+            operations.Conv2d(dim_in, oc, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype)
+            for oc in out_channels
+        ])
+        self.resize_layers = nn.ModuleList([
+            operations.ConvTranspose2d(out_channels[0], out_channels[0], kernel_size=4, stride=4, padding=0, device=device, dtype=dtype),
+            operations.ConvTranspose2d(out_channels[1], out_channels[1], kernel_size=2, stride=2, padding=0, device=device, dtype=dtype),
+            nn.Identity(),
+            operations.Conv2d(out_channels[3], out_channels[3], kernel_size=3, stride=2, padding=1, device=device, dtype=dtype),
+        ])
+
+        self.scratch = _make_scratch(out_channels, features, device=device, dtype=dtype, operations=operations)
+        # Main fusion chain
+        self.scratch.refinenet1 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet2 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet3 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet4 = _make_fusion_block(features, has_residual=False, device=device, dtype=dtype, operations=operations)
+        # Auxiliary fusion chain (separate copies)
+        self.scratch.refinenet1_aux = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet2_aux = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet3_aux = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet4_aux = _make_fusion_block(features, has_residual=False, device=device, dtype=dtype, operations=operations)
+
+        head_features_1 = features
+        head_features_2 = 32
+
+        # Main head neck + final projection
+        self.scratch.output_conv1 = operations.Conv2d(
+            head_features_1, head_features_1 // 2, kernel_size=3, stride=1, padding=1,
+            device=device, dtype=dtype,
+        )
+        self.scratch.output_conv2 = nn.Sequential(
+            operations.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1, device=device, dtype=dtype),
+            nn.ReLU(inplace=False),
+            operations.Conv2d(head_features_2, output_dim, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype),
+        )
+
+        # Aux pre-head per level (multi-level pyramid)
+        self.scratch.output_conv1_aux = nn.ModuleList([
+            self._make_aux_out1_block(head_features_1, device=device, dtype=dtype, operations=operations)
+            for _ in range(self.aux_levels)
+        ])
+
+        # Aux final projection per level (includes LayerNorm permute path).
+        ln_seq = [Permute((0, 2, 3, 1)),
+                  operations.LayerNorm(head_features_2, device=device, dtype=dtype),
+                  Permute((0, 3, 1, 2))]
+        self.scratch.output_conv2_aux = nn.ModuleList([
+            nn.Sequential(
+                operations.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1, device=device, dtype=dtype),
+                *ln_seq,
+                nn.ReLU(inplace=False),
+                operations.Conv2d(head_features_2, 7, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype),
+            )
+            for _ in range(self.aux_levels)
+        ])
+
+    @staticmethod
+    def _make_aux_out1_block(in_ch: int, *, device=None, dtype=None, operations=None) -> nn.Sequential:
+        # aux_out1_conv_num=5 in all Apache-2.0 variants.
+        return nn.Sequential(
+            operations.Conv2d(in_ch, in_ch // 2, 3, 1, 1, device=device, dtype=dtype),
+            operations.Conv2d(in_ch // 2, in_ch, 3, 1, 1, device=device, dtype=dtype),
+            operations.Conv2d(in_ch, in_ch // 2, 3, 1, 1, device=device, dtype=dtype),
+            operations.Conv2d(in_ch // 2, in_ch, 3, 1, 1, device=device, dtype=dtype),
+            operations.Conv2d(in_ch, in_ch // 2, 3, 1, 1, device=device, dtype=dtype),
+        )
+
+    def forward(self, feats: List[torch.Tensor], H: int, W: int, patch_start_idx: int = 0, **_kwargs) -> dict:
+        B, S, N, C = feats[0][0].shape
+        feats_flat = [feat[0].reshape(B * S, N, C) for feat in feats]
+
+        ph, pw = H // self.patch_size, W // self.patch_size
+        resized = []
+        for stage_idx, take_idx in enumerate(self.intermediate_layer_idx):
+            x = feats_flat[take_idx][:, patch_start_idx:]
+            x = self.norm(x)
+            x = x.permute(0, 2, 1).contiguous().reshape(B * S, C, ph, pw)
+            x = self.projects[stage_idx](x)
+            if self.pos_embed:
+                x = _add_pos_embed(x, W, H)
+            x = self.resize_layers[stage_idx](x)
+            resized.append(x)
+
+        l1_rn = self.scratch.layer1_rn(resized[0])
+        l2_rn = self.scratch.layer2_rn(resized[1])
+        l3_rn = self.scratch.layer3_rn(resized[2])
+        l4_rn = self.scratch.layer4_rn(resized[3])
+
+        # Main pyramid (output_conv1 is applied inside the upstream `_fuse`,
+        # before interpolation -- replicate that order here).
+        m = self.scratch.refinenet4(l4_rn, size=l3_rn.shape[2:])
+        if self.enable_aux:
+            a4 = self.scratch.refinenet4_aux(l4_rn, size=l3_rn.shape[2:])
+            aux_pyr = [a4]
+        m = self.scratch.refinenet3(m, l3_rn, size=l2_rn.shape[2:])
+        if self.enable_aux:
+            aux_pyr.append(self.scratch.refinenet3_aux(aux_pyr[-1], l3_rn, size=l2_rn.shape[2:]))
+        m = self.scratch.refinenet2(m, l2_rn, size=l1_rn.shape[2:])
+        if self.enable_aux:
+            aux_pyr.append(self.scratch.refinenet2_aux(aux_pyr[-1], l2_rn, size=l1_rn.shape[2:]))
+        m = self.scratch.refinenet1(m, l1_rn)
+        if self.enable_aux:
+            aux_pyr.append(self.scratch.refinenet1_aux(aux_pyr[-1], l1_rn))
+        m = self.scratch.output_conv1(m)
+
+        h_out = int(ph * self.patch_size / self.down_ratio)
+        w_out = int(pw * self.patch_size / self.down_ratio)
+
+        m = _custom_interpolate(m, (h_out, w_out), mode="bilinear", align_corners=True)
+        if self.pos_embed:
+            m = _add_pos_embed(m, W, H)
+        main_logits = self.scratch.output_conv2(m)
+        fmap = main_logits.permute(0, 2, 3, 1)
+        depth_pred = _apply_activation(fmap[..., :-1], self.activation)
+        depth_conf = _apply_activation(fmap[..., -1], self.conf_activation)
+
+        outs = {
+            self.head_main: depth_pred.squeeze(-1).view(B, S, *depth_pred.shape[1:-1]),
+            f"{self.head_main}_conf": depth_conf.view(B, S, *depth_conf.shape[1:]),
+        }
+
+        if self.enable_aux:
+            # Auxiliary "ray" head (multi-level inside) -- only the last level
+            # is returned. Mirrors upstream ``DualDPT._fuse`` + ``_forward_impl``:
+            # each aux pyramid level goes through ``output_conv1_aux[i]``
+            # (5-layer conv stack that ends at ``features // 2`` channels),
+            # then the last level optionally gets a pos-embed and finally
+            # ``output_conv2_aux[-1]``.
+            aux_processed = [
+                self.scratch.output_conv1_aux[i](a) for i, a in enumerate(aux_pyr)
+            ]
+            last_aux = aux_processed[-1]
+            if self.pos_embed:
+                last_aux = _add_pos_embed(last_aux, W, H)
+            last_aux_logits = self.scratch.output_conv2_aux[-1](last_aux)
+            fmap_last = last_aux_logits.permute(0, 2, 3, 1)
+            # Channels: [ray(6), ray_conf(1)]; ray uses 'linear' activation.
+            aux_pred = fmap_last[..., :-1]
+            aux_conf = _apply_activation(fmap_last[..., -1], self.conf_activation)
+            outs[self.head_aux] = aux_pred.view(B, S, *aux_pred.shape[1:])
+            outs[f"{self.head_aux}_conf"] = aux_conf.view(B, S, *aux_conf.shape[1:])
+
+        return outs

ComfyUI/comfy/ldm/depth_anything_3/model.py

@@ -0,0 +1,236 @@
+from __future__ import annotations
+
+from typing import Dict, Optional, Sequence
+
+import torch
+import torch.nn as nn
+
+from comfy.image_encoders.dino2 import Dinov2Model
+
+from .camera import CameraDec, CameraEnc
+from .dpt import DPT, DualDPT
+from .ray_pose import get_extrinsic_from_camray
+from .transform import affine_inverse, pose_encoding_to_extri_intri
+
+
+_HEAD_REGISTRY = {
+    "dpt": DPT,
+    "dualdpt": DualDPT,
+}
+
+
+# Backbone presets (mirror the upstream DINOv2 ViT variants).
+_BACKBONE_PRESETS = {
+    "vits": dict(hidden_size=384,  num_hidden_layers=12, num_attention_heads=6,  use_swiglu_ffn=False),
+    "vitb": dict(hidden_size=768,  num_hidden_layers=12, num_attention_heads=12, use_swiglu_ffn=False),
+    "vitl": dict(hidden_size=1024, num_hidden_layers=24, num_attention_heads=16, use_swiglu_ffn=False),
+    "vitg": dict(hidden_size=1536, num_hidden_layers=40, num_attention_heads=24, use_swiglu_ffn=True),
+}
+
+
+def _build_backbone_config(
+    backbone_name: str,
+    *,
+    alt_start: int,
+    qknorm_start: int,
+    rope_start: int,
+    cat_token: bool,
+) -> dict:
+    if backbone_name not in _BACKBONE_PRESETS:
+        raise ValueError(f"Unknown DINOv2 backbone variant: {backbone_name!r}")
+    cfg = dict(_BACKBONE_PRESETS[backbone_name])
+    cfg.update(dict(
+        layer_norm_eps=1e-6,
+        patch_size=14,
+        image_size=518,
+        # No mask_token in DA3 weights; omit param to avoid load warnings.
+        use_mask_token=False,
+        alt_start=alt_start,
+        qknorm_start=qknorm_start,
+        rope_start=rope_start,
+        cat_token=cat_token,
+        rope_freq=100.0,
+    ))
+    return cfg
+
+
+class DepthAnything3Net(nn.Module):
+
+    PATCH_SIZE = 14
+
+    def __init__(
+        self,
+        # --- Backbone ---
+        backbone_name: str = "vitl",
+        out_layers: Sequence[int] = (4, 11, 17, 23),
+        alt_start: int = -1,
+        qknorm_start: int = -1,
+        rope_start: int = -1,
+        cat_token: bool = False,
+        # --- Head ---
+        head_type: str = "dpt",  # dpt or dualdpt
+        head_dim_in: int = 1024,
+        head_output_dim: int = 1,  # 1 = depth only, 2 = depth+conf
+        head_features: int = 256,
+        head_out_channels: Sequence[int] = (256, 512, 1024, 1024),
+        head_use_sky_head: bool = True, # ignored by DualDPT
+        head_pos_embed: Optional[bool] = None,  # default: True for DualDPT, False for DPT
+        # --- Camera (multi-view) ---
+        has_cam_enc: bool = False,
+        has_cam_dec: bool = False,
+        cam_dim_out: Optional[int] = None,  # CameraEnc dim_out (defaults to embed_dim)
+        cam_dec_dim_in: Optional[int] = None,  # CameraDec dim_in  (defaults to 2*embed_dim with cat_token)
+        # ComfyUI plumbing
+        device=None, dtype=None, operations=None,
+        **_ignored,
+    ):
+        super().__init__()
+        head_cls = _HEAD_REGISTRY[head_type.lower()]
+        self.head_type = head_type.lower()
+        self.has_sky = (self.head_type == "dpt") and head_use_sky_head
+        self.has_conf = head_output_dim > 1
+        self.out_layers = list(out_layers)
+
+        backbone_cfg = _build_backbone_config(
+            backbone_name,
+            alt_start=alt_start,
+            qknorm_start=qknorm_start,
+            rope_start=rope_start,
+            cat_token=cat_token,
+        )
+        self.backbone = Dinov2Model(backbone_cfg, dtype, device, operations)
+
+        head_kwargs = dict(
+            dim_in=head_dim_in,
+            patch_size=self.PATCH_SIZE,
+            output_dim=head_output_dim,
+            features=head_features,
+            out_channels=tuple(head_out_channels),
+            device=device, dtype=dtype, operations=operations,
+        )
+        if self.head_type == "dpt":
+            head_kwargs.update(
+                use_sky_head=head_use_sky_head,
+                pos_embed=(False if head_pos_embed is None else head_pos_embed),
+            )
+        else:  # dualdpt
+            head_kwargs.update(
+                pos_embed=(True if head_pos_embed is None else head_pos_embed),
+            )
+        self.head = head_cls(**head_kwargs)
+
+        # Built only if checkpoint has weights; cam_enc output dim == embed_dim.
+        embed_dim = backbone_cfg["hidden_size"]
+        if has_cam_enc:
+            self.cam_enc = CameraEnc(
+                dim_out=cam_dim_out if cam_dim_out is not None else embed_dim,
+                num_heads=max(1, embed_dim // 64),
+                device=device, dtype=dtype, operations=operations,
+            )
+        else:
+            self.cam_enc = None
+        if has_cam_dec:
+            default_dim = embed_dim * (2 if cat_token else 1)
+            self.cam_dec = CameraDec(
+                dim_in=cam_dec_dim_in if cam_dec_dim_in is not None else default_dim,
+                device=device, dtype=dtype, operations=operations,
+            )
+        else:
+            self.cam_dec = None
+
+        self.dtype = dtype
+
+    def forward(
+        self,
+        image: torch.Tensor,
+        extrinsics: Optional[torch.Tensor] = None,
+        intrinsics: Optional[torch.Tensor] = None,
+        *,
+        use_ray_pose: bool = False,
+        ref_view_strategy: str = "saddle_balanced",
+        export_feat_layers: Optional[Sequence[int]] = None,
+        **_unused,
+    ) -> Dict[str, torch.Tensor]:
+        """Run depth and optionally pose prediction."""
+        if image.ndim == 4:
+            image = image.unsqueeze(1)  # (B, 1, 3, H, W)
+        assert image.ndim == 5 and image.shape[2] == 3, \
+            f"image must be (B,3,H,W) or (B,S,3,H,W); got {tuple(image.shape)}"
+
+        B, S, _, H, W = image.shape
+        assert H % self.PATCH_SIZE == 0 and W % self.PATCH_SIZE == 0, \
+            f"image H,W must be multiples of {self.PATCH_SIZE}; got {(H, W)}"
+
+        # Camera-token preparation (multi-view path).
+        cam_token = None
+        if extrinsics is not None and intrinsics is not None and self.cam_enc is not None:
+            cam_token = self.cam_enc(extrinsics, intrinsics, (H, W))
+
+        # Toggle aux ray output on/off depending on what the caller asked for.
+        if isinstance(self.head, DualDPT):
+            self.head.enable_aux = bool(use_ray_pose)
+
+        feats, aux_feats = self.backbone.get_intermediate_layers_da3(
+            image, self.out_layers, cam_token=cam_token,
+            ref_view_strategy=ref_view_strategy,
+            export_feat_layers=export_feat_layers,
+        )
+        head_out = self.head(feats, H=H, W=W, patch_start_idx=0)
+
+        # Pose prediction.
+        out: Dict[str, torch.Tensor] = {}
+        if use_ray_pose and "ray" in head_out and "ray_conf" in head_out:
+            ray = head_out["ray"]
+            ray_conf = head_out["ray_conf"]
+            extr_c2w, focal, pp = get_extrinsic_from_camray(
+                ray, ray_conf, ray.shape[-3], ray.shape[-2],
+            )
+            # Match the upstream output: w2c, drop the homogeneous row.
+            extr_w2c = affine_inverse(extr_c2w)[:, :, :3, :]
+            # Build pixel-space intrinsics from the normalised focal/pp output.
+            intr = torch.eye(3, device=ray.device, dtype=ray.dtype)
+            intr = intr[None, None].expand(extr_c2w.shape[0], extr_c2w.shape[1], 3, 3).clone()
+            intr[:, :, 0, 0] = focal[:, :, 0] / 2 * W
+            intr[:, :, 1, 1] = focal[:, :, 1] / 2 * H
+            intr[:, :, 0, 2] = pp[:, :, 0] * W * 0.5
+            intr[:, :, 1, 2] = pp[:, :, 1] * H * 0.5
+            out["extrinsics"] = extr_w2c
+            out["intrinsics"] = intr
+        elif self.cam_dec is not None and S > 1:
+            # Decode the cam-token of the final out_layer into a pose encoding.
+            cam_feat = feats[-1][1]  # (B, S, dim_in_to_cam_dec)
+            pose_enc = self.cam_dec(cam_feat)
+            c2w_3x4, intr = pose_encoding_to_extri_intri(pose_enc, (H, W))
+            # Match the upstream output convention: w2c (world->camera), 3x4.
+            c2w_4x4 = torch.cat([
+                c2w_3x4,
+                torch.tensor([0, 0, 0, 1], device=c2w_3x4.device, dtype=c2w_3x4.dtype)
+                    .view(1, 1, 1, 4).expand(B, S, 1, 4),
+            ], dim=-2)
+            out["extrinsics"] = affine_inverse(c2w_4x4)[:, :, :3, :]
+            out["intrinsics"] = intr
+
+        # Flatten the views axis for per-pixel outputs (depth/conf/sky) so the
+        # per-image consumer keeps its (B*S, H, W) interface.
+        for k, v in head_out.items():
+            if k in ("ray", "ray_conf"):
+                # Keep multi-view shape for downstream pose work.
+                out[k] = v
+            elif v.ndim >= 3 and v.shape[0] == B and v.shape[1] == S:
+                out[k] = v.reshape(B * S, *v.shape[2:])
+            else:
+                out[k] = v
+
+        if export_feat_layers:
+            out["aux_features"] = self._reshape_aux_features(aux_feats, H, W)
+        return out
+
+    def _reshape_aux_features(self, aux_feats, H: int, W: int):
+        """Reshape (B, S, N, C) aux features into (B, S, h_p, w_p, C)."""
+        ph, pw = H // self.PATCH_SIZE, W // self.PATCH_SIZE
+        out = []
+        for f in aux_feats:
+            B, S, N, C = f.shape
+            assert N == ph * pw, f"aux feature seq mismatch: {N} != {ph}*{pw}"
+            out.append(f.reshape(B, S, ph, pw, C))
+        return out

ComfyUI/comfy/ldm/depth_anything_3/preprocess.py

@@ -0,0 +1,128 @@
+"""Input/output preprocessing helpers for Depth Anything 3."""
+
+from __future__ import annotations
+
+from typing import Tuple
+
+import torch
+
+import comfy.utils
+
+PATCH_SIZE = 14
+
+# ImageNet normalization constants used during DA3 training.
+_IMAGENET_MEAN = torch.tensor([0.485, 0.456, 0.406])
+_IMAGENET_STD = torch.tensor([0.229, 0.224, 0.225])
+
+
+def _round_to_patch(x: int, patch: int = PATCH_SIZE) -> int:
+    down = (x // patch) * patch
+    up = down + patch
+    return up if abs(up - x) <= abs(x - down) else down
+
+
+def compute_target_size(orig_h: int, orig_w: int, process_res: int, method: str = "upper_bound_resize") -> Tuple[int, int]:
+    """Compute (target_h, target_w) for a single image.
+    upper_bound_resize: scale longest side to process_res, then round each dim to nearest multiple of 14 (default upstream method).
+    lower_bound_resize: scale shortest side to process_res, then round."""
+
+    if method == "upper_bound_resize":
+        longest = max(orig_h, orig_w)
+        scale = process_res / float(longest)
+    elif method == "lower_bound_resize":
+        shortest = min(orig_h, orig_w)
+        scale = process_res / float(shortest)
+    else:
+        raise ValueError(f"Unsupported process_res_method: {method}")
+
+    new_w = max(1, _round_to_patch(int(round(orig_w * scale))))
+    new_h = max(1, _round_to_patch(int(round(orig_h * scale))))
+    return new_h, new_w
+
+
+def preprocess_image(image: torch.Tensor, process_res: int = 504, method: str = "upper_bound_resize") -> torch.Tensor:
+    assert image.ndim == 4 and image.shape[-1] == 3, f"expected (B,H,W,3) IMAGE; got {tuple(image.shape)}"
+    B, H, W, _ = image.shape
+    target_h, target_w = compute_target_size(H, W, process_res, method)
+
+    # (B, H, W, 3) -> (B, 3, H, W)
+    x = image.movedim(-1, 1).contiguous()
+    if (target_h, target_w) != (H, W):
+        # Upstream uses cv2 INTER_CUBIC (upscale) / INTER_AREA (downscale).
+        # Lanczos in ``common_upscale`` is anti-aliased and produces the
+        # closest pixel-wise match in a sweep across {bilinear, bicubic,
+        # area, lanczos, bislerp}. Used in both directions for simplicity.
+        x = comfy.utils.common_upscale(x.float(), target_w, target_h, "lanczos", "disabled",)
+    x = x.clamp(0.0, 1.0)
+
+    mean = _IMAGENET_MEAN.to(device=x.device, dtype=x.dtype).view(1, 3, 1, 1)
+    std = _IMAGENET_STD.to(device=x.device, dtype=x.dtype).view(1, 3, 1, 1)
+    x = (x - mean) / std
+    return x
+
+
+# -----------------------------------------------------------------------------
+# Output post-processing (sky-aware clipping for Mono/Metric variants)
+# -----------------------------------------------------------------------------
+
+
+def compute_non_sky_mask(sky_prediction: torch.Tensor, threshold: float = 0.3) -> torch.Tensor:
+    """Boolean mask: True for non-sky pixels (sky probability < threshold)."""
+    return sky_prediction < threshold
+
+
+def apply_sky_aware_clip(depth: torch.Tensor, sky: torch.Tensor, threshold: float = 0.3, quantile: float = 0.99) -> torch.Tensor:
+    """Clips sky regions to the 99th percentile of non-sky depth. Returns a new depth tensor."""
+    non_sky = compute_non_sky_mask(sky, threshold=threshold)
+    if non_sky.sum() <= 10 or (~non_sky).sum() <= 10:
+        return depth.clone()
+
+    non_sky_depth = depth[non_sky]
+    if non_sky_depth.numel() > 100_000:
+        idx = torch.randint(0, non_sky_depth.numel(), (100_000,), device=non_sky_depth.device)
+        sampled = non_sky_depth[idx]
+    else:
+        sampled = non_sky_depth
+
+    max_depth = torch.quantile(sampled, quantile)
+    out = depth.clone()
+    out[~non_sky] = max_depth
+    return out
+
+
+def normalize_depth_v2_style(depth: torch.Tensor, sky: torch.Tensor | None = None, low_quantile: float = 0.01, high_quantile: float = 0.99) -> torch.Tensor:
+    """V2-style normalization computes percentile bounds over non-sky pixels (when available), then maps depth into [0, 1] with near = white (1.0)."""
+    if sky is not None:
+        mask = compute_non_sky_mask(sky)
+        if mask.any():
+            valid = depth[mask]
+        else:
+            valid = depth.flatten()
+    else:
+        valid = depth.flatten()
+
+    if valid.numel() > 100_000:
+        idx = torch.randint(0, valid.numel(), (100_000,), device=valid.device)
+        sample = valid[idx]
+    else:
+        sample = valid
+
+    lo = torch.quantile(sample, low_quantile)
+    hi = torch.quantile(sample, high_quantile)
+    rng = (hi - lo).clamp(min=1e-6)
+    norm = ((depth - lo) / rng).clamp(0.0, 1.0)
+    # Nearer pixels are brighter (1.0)
+    norm = 1.0 - norm
+    if sky is not None:
+        # Sky pixels become black (far / unknown)
+        sky_mask = ~compute_non_sky_mask(sky)
+        norm = torch.where(sky_mask, torch.zeros_like(norm), norm)
+    return norm
+
+
+def normalize_depth_min_max(depth: torch.Tensor) -> torch.Tensor:
+    """Simple per-frame min/max normalization with near=1.0 convention."""
+    lo = depth.amin(dim=(-2, -1), keepdim=True)
+    hi = depth.amax(dim=(-2, -1), keepdim=True)
+    rng = (hi - lo).clamp(min=1e-6)
+    return 1.0 - ((depth - lo) / rng).clamp(0.0, 1.0)

ComfyUI/comfy/ldm/depth_anything_3/ray_pose.py

@@ -0,0 +1,272 @@
+"""Ray-to-pose conversion for the multi-view path of Depth Anything 3."""
+
+from __future__ import annotations
+
+from typing import Optional, Tuple
+
+import torch
+
+
+# qr/svd use fp32: CUDA often has no fp16/bf16 kernels for these ops.
+
+
+def _ql_decomposition(A: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Decompose A = Q @ L with Q orthogonal and L lower-triangular.
+    Implemented in terms of QR by reversing the columns/rows; the standard
+    trick from the upstream reference. Inputs A are (3, 3)."""
+    P = torch.tensor([[0, 0, 1], [0, 1, 0], [1, 0, 0]], device=A.device, dtype=A.dtype)
+    A_tilde = A @ P
+    # CUDA QR is not implemented for fp16/bf16; upcast just for this call.
+    Q_tilde, R_tilde = torch.linalg.qr(A_tilde.float())
+    Q_tilde = Q_tilde.to(A.dtype)
+    R_tilde = R_tilde.to(A.dtype)
+    Q = Q_tilde @ P
+    L = P @ R_tilde @ P
+    d = torch.diag(L)
+    sign = torch.sign(d)
+    Q = Q * sign[None, :]  # scale columns of Q
+    L = L * sign[:, None]  # scale rows of L
+    return Q, L
+
+
+def _homogenize_points(points: torch.Tensor) -> torch.Tensor:
+    return torch.cat([points, torch.ones_like(points[..., :1])], dim=-1)
+
+
+# -----------------------------------------------------------------------------
+# Weighted-LSQ + RANSAC homography (batched)
+# -----------------------------------------------------------------------------
+
+
+def _find_homography_weighted_lsq(src_pts: torch.Tensor, dst_pts: torch.Tensor, confident_weight: torch.Tensor,) -> torch.Tensor:
+    """Solve a single H with weighted least-squares (DLT)."""
+    N = src_pts.shape[0]
+    if N < 4:
+        raise ValueError("At least 4 points are required to compute a homography.")
+    w = confident_weight.sqrt().unsqueeze(1)  # (N, 1)
+    x = src_pts[:, 0:1]
+    y = src_pts[:, 1:2]
+    u = dst_pts[:, 0:1]
+    v = dst_pts[:, 1:2]
+    zeros = torch.zeros_like(x)
+    A1 = torch.cat([-x * w, -y * w, -w, zeros, zeros, zeros, x * u * w, y * u * w, u * w], dim=1)
+    A2 = torch.cat([zeros, zeros, zeros, -x * w, -y * w, -w, x * v * w, y * v * w, v * w], dim=1)
+    A = torch.cat([A1, A2], dim=0)        # (2N, 9)
+    # CUDA SVD is not implemented for fp16/bf16; upcast just for this call.
+    _, _, Vh = torch.linalg.svd(A.float())
+    Vh = Vh.to(A.dtype)
+    H = Vh[-1].reshape(3, 3)
+    return H / H[-1, -1]
+
+
+def _find_homography_weighted_lsq_batched(src_pts_batch: torch.Tensor, dst_pts_batch: torch.Tensor, confident_weight_batch: torch.Tensor) -> torch.Tensor:
+    """Batched DLT solver. Inputs (B, K, 2) / (B, K); output (B, 3, 3)."""
+    B, K, _ = src_pts_batch.shape
+    w = confident_weight_batch.sqrt().unsqueeze(2)
+    x = src_pts_batch[:, :, 0:1]
+    y = src_pts_batch[:, :, 1:2]
+    u = dst_pts_batch[:, :, 0:1]
+    v = dst_pts_batch[:, :, 1:2]
+    zeros = torch.zeros_like(x)
+    A1 = torch.cat([-x * w, -y * w, -w, zeros, zeros, zeros, x * u * w, y * u * w, u * w], dim=2)
+    A2 = torch.cat([zeros, zeros, zeros, -x * w, -y * w, -w, x * v * w, y * v * w, v * w], dim=2)
+    A = torch.cat([A1, A2], dim=1)        # (B, 2K, 9)
+    # CUDA SVD is not implemented for fp16/bf16; upcast just for this call.
+    _, _, Vh = torch.linalg.svd(A.float())
+    Vh = Vh.to(A.dtype)
+    H = Vh[:, -1].reshape(B, 3, 3)
+    return H / H[:, 2:3, 2:3]
+
+
+def _ransac_find_homography_weighted_batched(
+    src_pts: torch.Tensor,                # (B, N, 2)
+    dst_pts: torch.Tensor,                # (B, N, 2)
+    confident_weight: torch.Tensor,       # (B, N)
+    n_sample: int,
+    n_iter: int = 100,
+    reproj_threshold: float = 3.0,
+    num_sample_for_ransac: int = 8,
+    max_inlier_num: int = 10000,
+    rand_sample_iters_idx: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Batched weighted-RANSAC homography estimator. Returns (B, 3, 3) homography matrices."""
+    B, N, _ = src_pts.shape
+    assert N >= 4
+    device = src_pts.device
+
+    sorted_idx = torch.argsort(confident_weight, descending=True, dim=1)
+    candidate_idx = sorted_idx[:, :n_sample]                  # (B, n_sample)
+
+    if rand_sample_iters_idx is None:
+        rand_sample_iters_idx = torch.stack(
+            [torch.randperm(n_sample, device=device)[:num_sample_for_ransac]
+             for _ in range(n_iter)],
+            dim=0,
+        )
+
+    rand_idx = candidate_idx[:, rand_sample_iters_idx]        # (B, n_iter, k)
+    b_idx = (
+        torch.arange(B, device=device)
+        .view(B, 1, 1)
+        .expand(B, n_iter, num_sample_for_ransac)
+    )
+    src_b = src_pts[b_idx, rand_idx]
+    dst_b = dst_pts[b_idx, rand_idx]
+    w_b = confident_weight[b_idx, rand_idx]
+
+    cB, cN = src_b.shape[:2]
+    H_batch = _find_homography_weighted_lsq_batched(
+        src_b.flatten(0, 1), dst_b.flatten(0, 1), w_b.flatten(0, 1),
+    ).unflatten(0, (cB, cN))                                  # (B, n_iter, 3, 3)
+
+    src_homo = torch.cat([src_pts, torch.ones(B, N, 1, device=device, dtype=src_pts.dtype)], dim=2)
+    proj = torch.bmm(
+        src_homo.unsqueeze(1).expand(B, n_iter, N, 3).reshape(-1, N, 3),
+        H_batch.reshape(-1, 3, 3).transpose(1, 2),
+    )                                                          # (B*n_iter, N, 3)
+    proj_xy = (proj[:, :, :2] / proj[:, :, 2:3]).reshape(B, n_iter, N, 2)
+    err = ((proj_xy - dst_pts.unsqueeze(1)) ** 2).sum(-1).sqrt()  # (B, n_iter, N)
+    inlier_mask = err < reproj_threshold
+    score = (inlier_mask * confident_weight.unsqueeze(1)).sum(dim=2)
+    best_idx = torch.argmax(score, dim=1)
+    best_inlier_mask = inlier_mask[torch.arange(B, device=device), best_idx]
+
+    # Refit with the inlier set (per-batch, since the inlier counts vary).
+    H_inlier_list = []
+    for b in range(B):
+        mask = best_inlier_mask[b]
+        in_src = src_pts[b][mask]
+        in_dst = dst_pts[b][mask]
+        in_w = confident_weight[b][mask]
+        if in_src.shape[0] < 4:
+            # Fall back to identity when RANSAC fails to find enough inliers.
+            H_inlier_list.append(torch.eye(3, device=device, dtype=src_pts.dtype))
+            continue
+        sorted_w = torch.argsort(in_w, descending=True)
+        if len(sorted_w) > max_inlier_num:
+            keep = max(int(len(sorted_w) * 0.95), max_inlier_num)
+            sorted_w = sorted_w[:keep][torch.randperm(keep, device=device)[:max_inlier_num]]
+        H_inlier_list.append(
+            _find_homography_weighted_lsq(in_src[sorted_w], in_dst[sorted_w], in_w[sorted_w])
+        )
+    return torch.stack(H_inlier_list, dim=0)
+
+
+# -----------------------------------------------------------------------------
+# Camera-ray utilities
+# -----------------------------------------------------------------------------
+
+
+def _unproject_identity(num_y: int, num_x: int, B: int, S: int, device, dtype) -> torch.Tensor:
+    """Camera-space unit rays for an identity intrinsic on a 2x2 image plane."""
+    dx = 1.0 / num_x
+    dy = 1.0 / num_y
+    # Centered camera-space coords directly (skip the K^-1 step since it's
+    # just a translation by -1 on x and y when K is identity-with-center=1).
+    y = torch.linspace(-(1 - dy), (1 - dy), num_y, device=device, dtype=dtype)
+    x = torch.linspace(-(1 - dx), (1 - dx), num_x, device=device, dtype=dtype)
+    yy, xx = torch.meshgrid(y, x, indexing="ij")
+    grid = torch.stack((xx, yy), dim=-1)            # (h, w, 2)
+    grid = grid.unsqueeze(0).unsqueeze(0).expand(B, S, num_y, num_x, 2)
+    return torch.cat([grid, torch.ones_like(grid[..., :1])], dim=-1)
+
+
+def _camray_to_caminfo(
+    camray: torch.Tensor,  # (B, S, h, w, 6)
+    confidence: Optional[torch.Tensor] = None,  # (B, S, h, w)
+    reproj_threshold: float = 0.2,
+) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+    """Convert per-pixel camera rays to per-view (R, T, focal, principal)."""
+    if confidence is None:
+        confidence = torch.ones_like(camray[..., 0])
+    B, S, h, w, _ = camray.shape
+    device = camray.device
+    dtype = camray.dtype
+
+    rays_target = camray[..., :3]                           # (B, S, h, w, 3)
+    rays_origin = _unproject_identity(h, w, B, S, device, dtype)
+
+    # Flatten (B*S, h*w, *) for the RANSAC routine.
+    rays_target = rays_target.flatten(0, 1).flatten(1, 2)
+    rays_origin = rays_origin.flatten(0, 1).flatten(1, 2)
+    weights = confidence.flatten(0, 1).flatten(1, 2).clone()
+
+    # Project to 2D in homogeneous form (the upstream calls this "perspective division").
+    z_thresh = 1e-4
+    mask = (rays_target[:, :, 2].abs() > z_thresh) & (rays_origin[:, :, 2].abs() > z_thresh)
+    weights = torch.where(mask, weights, torch.zeros_like(weights))
+    src = rays_origin.clone()
+    dst = rays_target.clone()
+    src[..., 0] = torch.where(mask, src[..., 0] / src[..., 2], src[..., 0])
+    src[..., 1] = torch.where(mask, src[..., 1] / src[..., 2], src[..., 1])
+    dst[..., 0] = torch.where(mask, dst[..., 0] / dst[..., 2], dst[..., 0])
+    dst[..., 1] = torch.where(mask, dst[..., 1] / dst[..., 2], dst[..., 1])
+    src = src[..., :2]
+    dst = dst[..., :2]
+
+    N = src.shape[1]
+    n_iter = 100
+    sample_ratio = 0.3
+    num_sample_for_ransac = 8
+    n_sample = max(num_sample_for_ransac, int(N * sample_ratio))
+    rand_idx = torch.stack(
+        [torch.randperm(n_sample, device=device)[:num_sample_for_ransac] for _ in range(n_iter)],
+        dim=0,
+    )
+
+    # Chunk along the view axis to keep peak memory predictable.
+    chunk = 2
+    A_list = []
+    for i in range(0, src.shape[0], chunk):
+        A = _ransac_find_homography_weighted_batched(
+            src[i:i + chunk], dst[i:i + chunk], weights[i:i + chunk],
+            n_sample=n_sample, n_iter=n_iter,
+            num_sample_for_ransac=num_sample_for_ransac,
+            reproj_threshold=reproj_threshold,
+            rand_sample_iters_idx=rand_idx,
+            max_inlier_num=8000,
+        )
+        # Flip sign on dets that come out < 0 (so that the QL produces a
+        # right-handed rotation). ``det`` lacks fp16/bf16 CUDA kernels, so
+        # do the comparison in fp32.
+        flip = torch.linalg.det(A.float()) < 0
+        A = torch.where(flip[:, None, None], -A, A)
+        A_list.append(A)
+    A = torch.cat(A_list, dim=0)                            # (B*S, 3, 3)
+
+    R_list, f_list, pp_list = [], [], []
+    for i in range(A.shape[0]):
+        R, L = _ql_decomposition(A[i])
+        L = L / L[2][2]
+        f_list.append(torch.stack((L[0][0], L[1][1])))
+        pp_list.append(torch.stack((L[2][0], L[2][1])))
+        R_list.append(R)
+    R = torch.stack(R_list).reshape(B, S, 3, 3)
+    focal = torch.stack(f_list).reshape(B, S, 2)
+    pp = torch.stack(pp_list).reshape(B, S, 2)
+
+    # Translation: confidence-weighted average of camray direction(s).
+    cf = confidence.flatten(0, 1).flatten(1, 2)
+    T = (camray.flatten(0, 1).flatten(1, 2)[..., 3:] * cf.unsqueeze(-1)).sum(dim=1)
+    T = T / cf.sum(dim=-1, keepdim=True)
+    T = T.reshape(B, S, 3)
+
+    # Match upstream output convention: focal -> 1/focal, pp + 1.
+    return R, T, 1.0 / focal, pp + 1.0
+
+
+def get_extrinsic_from_camray(
+    camray: torch.Tensor,  # (B, S, h, w, 6)
+    conf: torch.Tensor,  # (B, S, h, w, 1) or (B, S, h, w)
+    patch_size_y: int,
+    patch_size_x: int,
+) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    """Wrap a 4x4 extrinsic + per-view focal + principal-point output."""
+    if conf.ndim == 5 and conf.shape[-1] == 1:
+        conf = conf.squeeze(-1)
+    R, T, focal, pp = _camray_to_caminfo(camray, confidence=conf)
+    extr = torch.cat([R, T.unsqueeze(-1)], dim=-1)           # (B, S, 3, 4)
+    homo_row = torch.tensor([0, 0, 0, 1], dtype=R.dtype, device=R.device)
+    homo_row = homo_row.view(1, 1, 1, 4).expand(R.shape[0], R.shape[1], 1, 4)
+    extr = torch.cat([extr, homo_row], dim=-2)               # (B, S, 4, 4)
+    return extr, focal, pp

ComfyUI/comfy/ldm/depth_anything_3/reference_view_selector.py

@@ -0,0 +1,87 @@
+"""Reference-view selection for the multi-view path of Depth Anything 3."""
+
+from __future__ import annotations
+
+from typing import Literal
+
+import torch
+
+
+RefViewStrategy = Literal["first", "middle", "saddle_balanced", "saddle_sim_range"]
+
+
+# Per the upstream constants module: ``THRESH_FOR_REF_SELECTION = 3``.
+# Reference selection only runs when there are at least this many views.
+THRESH_FOR_REF_SELECTION: int = 3
+
+
+def select_reference_view(x: torch.Tensor, strategy: RefViewStrategy = "saddle_balanced") -> torch.Tensor:
+    """Pick a reference view index per batch element."""
+    B, S, _, _ = x.shape
+    if S <= 1:
+        return torch.zeros(B, dtype=torch.long, device=x.device)
+    if strategy == "first":
+        return torch.zeros(B, dtype=torch.long, device=x.device)
+    if strategy == "middle":
+        return torch.full((B,), S // 2, dtype=torch.long, device=x.device)
+
+    # Feature-based strategies: normalised cls/cam token per view.
+    img_class_feat = x[:, :, 0] / x[:, :, 0].norm(dim=-1, keepdim=True)  # (B,S,C)
+
+    if strategy == "saddle_balanced":
+        sim = torch.matmul(img_class_feat, img_class_feat.transpose(1, 2))  # (B,S,S)
+        sim_no_diag = sim - torch.eye(S, device=sim.device).unsqueeze(0)
+        sim_score = sim_no_diag.sum(dim=-1) / (S - 1)               # (B,S)
+        feat_norm = x[:, :, 0].norm(dim=-1)                          # (B,S)
+        feat_var = img_class_feat.var(dim=-1)                        # (B,S)
+
+        def _normalize(metric):
+            mn = metric.min(dim=1, keepdim=True).values
+            mx = metric.max(dim=1, keepdim=True).values
+            return (metric - mn) / (mx - mn + 1e-8)
+
+        sim_n, norm_n, var_n = _normalize(sim_score), _normalize(feat_norm), _normalize(feat_var)
+        balance = (sim_n - 0.5).abs() + (norm_n - 0.5).abs() + (var_n - 0.5).abs()
+        return balance.argmin(dim=1)
+
+    if strategy == "saddle_sim_range":
+        sim = torch.matmul(img_class_feat, img_class_feat.transpose(1, 2))
+        sim_no_diag = sim - torch.eye(S, device=sim.device).unsqueeze(0)
+        sim_max = sim_no_diag.max(dim=-1).values
+        sim_min = sim_no_diag.min(dim=-1).values
+        return (sim_max - sim_min).argmax(dim=1)
+
+    raise ValueError(
+        f"Unknown reference view selection strategy: {strategy!r}. "
+        f"Must be one of: 'first', 'middle', 'saddle_balanced', 'saddle_sim_range'"
+    )
+
+
+def reorder_by_reference(x: torch.Tensor, b_idx: torch.Tensor) -> torch.Tensor:
+    """Reorder x so the reference view is at position 0 in axis S."""
+    B, S = x.shape[0], x.shape[1]
+    if S <= 1:
+        return x
+    positions = torch.arange(S, device=x.device).unsqueeze(0).expand(B, -1)
+    b_idx_exp = b_idx.unsqueeze(1)
+    reorder = torch.where(
+        (positions > 0) & (positions <= b_idx_exp),
+        positions - 1,
+        positions,
+    )
+    reorder[:, 0] = b_idx
+    batch = torch.arange(B, device=x.device).unsqueeze(1)
+    return x[batch, reorder]
+
+
+def restore_original_order(x: torch.Tensor, b_idx: torch.Tensor) -> torch.Tensor:
+    """Inverse of reorder_by_reference."""
+    B, S = x.shape[0], x.shape[1]
+    if S <= 1:
+        return x
+    target_positions = torch.arange(S, device=x.device).unsqueeze(0).expand(B, -1)
+    b_idx_exp = b_idx.unsqueeze(1)
+    restore = torch.where(target_positions < b_idx_exp, target_positions + 1, target_positions)
+    restore = torch.scatter(restore, dim=1, index=b_idx_exp, src=torch.zeros_like(b_idx_exp))
+    batch = torch.arange(B, device=x.device).unsqueeze(1)
+    return x[batch, restore]

ComfyUI/comfy/ldm/depth_anything_3/transform.py

@@ -0,0 +1,160 @@
+"""Geometry / camera transform helpers for Depth Anything 3."""
+
+from __future__ import annotations
+
+from typing import Tuple
+
+import torch
+import torch.nn.functional as F
+
+
+# -----------------------------------------------------------------------------
+# Affine 4x4 helpers
+# -----------------------------------------------------------------------------
+
+
+def as_homogeneous(ext: torch.Tensor) -> torch.Tensor:
+    """Promote (...,3,4) extrinsics to (...,4,4) homogeneous form. No-op when the input is already ``(...,4,4)``."""
+    if ext.shape[-2:] == (4, 4):
+        return ext
+    if ext.shape[-2:] == (3, 4):
+        ones = torch.zeros_like(ext[..., :1, :4])
+        ones[..., 0, 3] = 1.0
+        return torch.cat([ext, ones], dim=-2)
+    raise ValueError(f"Invalid affine shape: {ext.shape}")
+
+
+def affine_inverse(A: torch.Tensor) -> torch.Tensor:
+    """Inverse of an affine matrix ``[R|T; 0 0 0 1]``."""
+    R = A[..., :3, :3]
+    T = A[..., :3, 3:]
+    P = A[..., 3:, :]
+    return torch.cat([torch.cat([R.mT, -R.mT @ T], dim=-1), P], dim=-2)
+
+
+# -----------------------------------------------------------------------------
+# Quaternion <-> rotation matrix (xyzw / scalar-last)
+# -----------------------------------------------------------------------------
+
+
+def _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:
+    """sqrt(max(0, x)) with a zero subgradient where x == 0."""
+    ret = torch.zeros_like(x)
+    positive_mask = x > 0
+    if torch.is_grad_enabled():
+        ret[positive_mask] = torch.sqrt(x[positive_mask])
+    else:
+        ret = torch.where(positive_mask, torch.sqrt(x), ret)
+    return ret
+
+
+def standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:
+    """Force the real part of a unit quaternion (xyzw) to be non-negative."""
+    return torch.where(quaternions[..., 3:4] < 0, -quaternions, quaternions)
+
+
+def quat_to_mat(quaternions: torch.Tensor) -> torch.Tensor:
+    """Convert quaternions (xyzw) to (...,3,3) rotation matrices."""
+    i, j, k, r = torch.unbind(quaternions, -1)
+    two_s = 2.0 / (quaternions * quaternions).sum(-1)
+    o = torch.stack(
+        (
+            1 - two_s * (j * j + k * k),
+            two_s * (i * j - k * r),
+            two_s * (i * k + j * r),
+            two_s * (i * j + k * r),
+            1 - two_s * (i * i + k * k),
+            two_s * (j * k - i * r),
+            two_s * (i * k - j * r),
+            two_s * (j * k + i * r),
+            1 - two_s * (i * i + j * j),
+        ),
+        -1,
+    )
+    return o.reshape(quaternions.shape[:-1] + (3, 3))
+
+
+def mat_to_quat(matrix: torch.Tensor) -> torch.Tensor:
+    """Convert (...,3,3) rotation matrices to quaternions (xyzw)."""
+    if matrix.size(-1) != 3 or matrix.size(-2) != 3:
+        raise ValueError(f"Invalid rotation matrix shape {matrix.shape}.")
+
+    batch_dim = matrix.shape[:-2]
+    m00, m01, m02, m10, m11, m12, m20, m21, m22 = torch.unbind(
+        matrix.reshape(batch_dim + (9,)), dim=-1
+    )
+
+    q_abs = _sqrt_positive_part(
+        torch.stack(
+            [
+                1.0 + m00 + m11 + m22,
+                1.0 + m00 - m11 - m22,
+                1.0 - m00 + m11 - m22,
+                1.0 - m00 - m11 + m22,
+            ],
+            dim=-1,
+        )
+    )
+
+    quat_by_rijk = torch.stack(
+        [
+            torch.stack([q_abs[..., 0] ** 2, m21 - m12, m02 - m20, m10 - m01], dim=-1),
+            torch.stack([m21 - m12, q_abs[..., 1] ** 2, m10 + m01, m02 + m20], dim=-1),
+            torch.stack([m02 - m20, m10 + m01, q_abs[..., 2] ** 2, m12 + m21], dim=-1),
+            torch.stack([m10 - m01, m20 + m02, m21 + m12, q_abs[..., 3] ** 2], dim=-1),
+        ],
+        dim=-2,
+    )
+
+    flr = torch.tensor(0.1).to(dtype=q_abs.dtype, device=q_abs.device)
+    quat_candidates = quat_by_rijk / (2.0 * q_abs[..., None].max(flr))
+
+    out = quat_candidates[F.one_hot(q_abs.argmax(dim=-1), num_classes=4) > 0.5, :].reshape(
+        batch_dim + (4,)
+    )
+    # Reorder rijk -> xyzw (i.e. ijkr).
+    out = out[..., [1, 2, 3, 0]]
+    return standardize_quaternion(out)
+
+
+# -----------------------------------------------------------------------------
+# Pose-encoding <-> extrinsics + intrinsics
+# -----------------------------------------------------------------------------
+
+
+def extri_intri_to_pose_encoding(extrinsics: torch.Tensor, intrinsics: torch.Tensor, image_size_hw: Tuple[int, int]) -> torch.Tensor:
+    """Pack (extr, intr, image_size) into the 9-D pose-encoding vector.
+    extrinsics: camera-to-world (c2w) (B,S,4,4) matrices,
+    intrinsics: pixel-space (B,S,3,3) matrices,
+    image_size_hw: is a (H, W) pair.
+    """
+    R = extrinsics[..., :3, :3]
+    T = extrinsics[..., :3, 3]
+    quat = mat_to_quat(R)
+    H, W = image_size_hw
+    fov_h = 2 * torch.atan((H / 2) / intrinsics[..., 1, 1])
+    fov_w = 2 * torch.atan((W / 2) / intrinsics[..., 0, 0])
+    return torch.cat([T, quat, fov_h[..., None], fov_w[..., None]], dim=-1).float()
+
+
+def pose_encoding_to_extri_intri(pose_encoding: torch.Tensor, image_size_hw: Tuple[int, int]) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Inverse of extri_intri_to_pose_encoding."""
+    T = pose_encoding[..., :3]
+    quat = pose_encoding[..., 3:7]
+    fov_h = pose_encoding[..., 7]
+    fov_w = pose_encoding[..., 8]
+    # Normalize to unit quaternion. CameraDec outputs raw values; a near-zero
+    # quaternion causes two_s = 2/norm² → inf in quat_to_mat → NaN extrinsics.
+    quat = quat / quat.norm(dim=-1, keepdim=True).clamp(min=1e-6)
+    R = quat_to_mat(quat)
+    extrinsics = torch.cat([R, T[..., None]], dim=-1)
+    H, W = image_size_hw
+    fy = (H / 2.0) / torch.clamp(torch.tan(fov_h / 2.0), 1e-6)
+    fx = (W / 2.0) / torch.clamp(torch.tan(fov_w / 2.0), 1e-6)
+    intrinsics = torch.zeros(pose_encoding.shape[:2] + (3, 3), device=pose_encoding.device, dtype=pose_encoding.dtype)
+    intrinsics[..., 0, 0] = fx
+    intrinsics[..., 1, 1] = fy
+    intrinsics[..., 0, 2] = W / 2
+    intrinsics[..., 1, 2] = H / 2
+    intrinsics[..., 2, 2] = 1.0
+    return extrinsics, intrinsics

ComfyUI/comfy/ldm/ernie/model.py

@@ -0,0 +1,307 @@
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+import comfy.ops
+import comfy.quant_ops
+
+def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
+    assert dim % 2 == 0
+    if not comfy.model_management.supports_fp64(pos.device):
+        device = torch.device("cpu")
+    else:
+        device = pos.device
+
+    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=device) / dim
+    omega = 1.0 / (theta**scale)
+    out = torch.einsum("...n,d->...nd", pos.to(device), omega)
+    out = torch.stack([torch.cos(out), torch.sin(out)], dim=0)
+    return out.to(dtype=torch.float32, device=pos.device)
+
+class ErnieImageEmbedND3(nn.Module):
+    def __init__(self, dim: int, theta: int, axes_dim: tuple):
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+        self.axes_dim = list(axes_dim)
+
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        emb = torch.cat([rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(3)], dim=-1)
+        cos_ = emb[0]
+        sin_ = emb[1]
+        N = cos_.shape[-1]
+        half = N // 2
+        cos_top = cos_[..., :half].repeat_interleave(2, dim=-1)
+        sin_top = sin_[..., :half].repeat_interleave(2, dim=-1)
+        cos_bot = cos_[..., half:].repeat_interleave(2, dim=-1)
+        sin_bot = sin_[..., half:].repeat_interleave(2, dim=-1)
+        rot = torch.stack([cos_top, -sin_top, sin_bot, cos_bot], dim=-1)
+        return rot.reshape(*rot.shape[:-1], 2, 2).unsqueeze(2)
+
+class ErnieImagePatchEmbedDynamic(nn.Module):
+    def __init__(self, in_channels: int, embed_dim: int, patch_size: int, operations, device=None, dtype=None):
+        super().__init__()
+        self.patch_size = patch_size
+        self.proj = operations.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        batch_size, dim, height, width = x.shape
+        return x.reshape(batch_size, dim, height * width).transpose(1, 2).contiguous()
+
+class Timesteps(nn.Module):
+    def __init__(self, num_channels: int, flip_sin_to_cos: bool = False):
+        super().__init__()
+        self.num_channels = num_channels
+        self.flip_sin_to_cos = flip_sin_to_cos
+
+    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
+        half_dim = self.num_channels // 2
+        exponent = -math.log(10000) * torch.arange(half_dim, dtype=torch.float32, device=timesteps.device) / half_dim
+        emb = torch.exp(exponent)
+        emb = timesteps[:, None].float() * emb[None, :]
+        if self.flip_sin_to_cos:
+            emb = torch.cat([torch.cos(emb), torch.sin(emb)], dim=-1)
+        else:
+            emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
+        return emb
+
+class TimestepEmbedding(nn.Module):
+    def __init__(self, in_channels: int, time_embed_dim: int, operations, device=None, dtype=None):
+        super().__init__()
+        Linear = operations.Linear
+        self.linear_1 = Linear(in_channels, time_embed_dim, bias=True, device=device, dtype=dtype)
+        self.act = nn.SiLU()
+        self.linear_2 = Linear(time_embed_dim, time_embed_dim, bias=True, device=device, dtype=dtype)
+
+    def forward(self, sample: torch.Tensor) -> torch.Tensor:
+        sample = self.linear_1(sample)
+        sample = self.act(sample)
+        sample = self.linear_2(sample)
+        return sample
+
+class ErnieImageAttention(nn.Module):
+    def __init__(self, query_dim: int, heads: int, dim_head: int, eps: float = 1e-6, operations=None, device=None, dtype=None):
+        super().__init__()
+        self.heads = heads
+        self.head_dim = dim_head
+        self.inner_dim = heads * dim_head
+
+        Linear = operations.Linear
+        RMSNorm = operations.RMSNorm
+
+        self.to_q = Linear(query_dim, self.inner_dim, bias=False, device=device, dtype=dtype)
+        self.to_k = Linear(query_dim, self.inner_dim, bias=False, device=device, dtype=dtype)
+        self.to_v = Linear(query_dim, self.inner_dim, bias=False, device=device, dtype=dtype)
+
+        self.norm_q = RMSNorm(dim_head, eps=eps, elementwise_affine=True, device=device, dtype=dtype)
+        self.norm_k = RMSNorm(dim_head, eps=eps, elementwise_affine=True, device=device, dtype=dtype)
+
+        self.to_out = nn.ModuleList([Linear(self.inner_dim, query_dim, bias=False, device=device, dtype=dtype)])
+
+    def forward(self, x: torch.Tensor, attention_mask: torch.Tensor = None, image_rotary_emb: torch.Tensor = None) -> torch.Tensor:
+        B, S, _ = x.shape
+
+        q_flat = self.to_q(x)
+        k_flat = self.to_k(x)
+        v_flat = self.to_v(x)
+
+        query = q_flat.view(B, S, self.heads, self.head_dim)
+        key = k_flat.view(B, S, self.heads, self.head_dim)
+
+        if image_rotary_emb is not None and not comfy.model_management.in_training:
+            q_scale, _, q_offload_stream = comfy.ops.cast_bias_weight(self.norm_q, query, offloadable=True)
+            k_scale, _, k_offload_stream = comfy.ops.cast_bias_weight(self.norm_k, key, offloadable=True)
+            query, key = comfy.quant_ops.ck.rms_rope_split_half(query, key, image_rotary_emb, q_scale, k_scale, self.norm_q.eps)
+            comfy.ops.uncast_bias_weight(self.norm_q, q_scale, None, q_offload_stream)
+            comfy.ops.uncast_bias_weight(self.norm_k, k_scale, None, k_offload_stream)
+        else:
+            query = self.norm_q(query)
+            key = self.norm_k(key)
+            if image_rotary_emb is not None:
+                query, key = comfy.quant_ops.ck.apply_rope_split_half(query, key, image_rotary_emb)
+
+        q_flat = query.reshape(B, S, -1)
+        k_flat = key.reshape(B, S, -1)
+
+        hidden_states = optimized_attention(q_flat, k_flat, v_flat, self.heads, mask=attention_mask)
+
+        return self.to_out[0](hidden_states)
+
+class ErnieImageFeedForward(nn.Module):
+    def __init__(self, hidden_size: int, ffn_hidden_size: int, operations, device=None, dtype=None):
+        super().__init__()
+        Linear = operations.Linear
+        self.gate_proj = Linear(hidden_size, ffn_hidden_size, bias=False, device=device, dtype=dtype)
+        self.up_proj = Linear(hidden_size, ffn_hidden_size, bias=False, device=device, dtype=dtype)
+        self.linear_fc2 = Linear(ffn_hidden_size, hidden_size, bias=False, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.linear_fc2(self.up_proj(x) * F.gelu(self.gate_proj(x)))
+
+class ErnieImageSharedAdaLNBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, ffn_hidden_size: int, eps: float = 1e-6, operations=None, device=None, dtype=None):
+        super().__init__()
+        RMSNorm = operations.RMSNorm
+
+        self.adaLN_sa_ln = RMSNorm(hidden_size, eps=eps, device=device, dtype=dtype)
+        self.self_attention = ErnieImageAttention(
+            query_dim=hidden_size,
+            dim_head=hidden_size // num_heads,
+            heads=num_heads,
+            eps=eps,
+            operations=operations,
+            device=device,
+            dtype=dtype
+        )
+        self.adaLN_mlp_ln = RMSNorm(hidden_size, eps=eps, device=device, dtype=dtype)
+        self.mlp = ErnieImageFeedForward(hidden_size, ffn_hidden_size, operations=operations, device=device, dtype=dtype)
+
+    def forward(self, x, rotary_pos_emb, temb, attention_mask=None):
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = temb
+
+        residual = x
+        x_norm = self.adaLN_sa_ln(x)
+        x_norm = x_norm * (1 + scale_msa) + shift_msa
+
+        attn_out = self.self_attention(x_norm, attention_mask=attention_mask, image_rotary_emb=rotary_pos_emb)
+        x = residual + gate_msa * attn_out
+
+        residual = x
+        x_norm = self.adaLN_mlp_ln(x)
+        x_norm = x_norm * (1 + scale_mlp) + shift_mlp
+
+        return residual + gate_mlp * self.mlp(x_norm)
+
+class ErnieImageAdaLNContinuous(nn.Module):
+    def __init__(self, hidden_size: int, eps: float = 1e-6, operations=None, device=None, dtype=None):
+        super().__init__()
+        LayerNorm = operations.LayerNorm
+        Linear = operations.Linear
+        self.norm = LayerNorm(hidden_size, elementwise_affine=False, eps=eps, device=device, dtype=dtype)
+        self.linear = Linear(hidden_size, hidden_size * 2, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, conditioning: torch.Tensor) -> torch.Tensor:
+        scale, shift = self.linear(conditioning).chunk(2, dim=-1)
+        x = self.norm(x)
+        x = torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1))
+        return x
+
+class ErnieImageModel(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int = 4096,
+        num_attention_heads: int = 32,
+        num_layers: int = 36,
+        ffn_hidden_size: int = 12288,
+        in_channels: int = 128,
+        out_channels: int = 128,
+        patch_size: int = 1,
+        text_in_dim: int = 3072,
+        rope_theta: int = 256,
+        rope_axes_dim: tuple = (32, 48, 48),
+        eps: float = 1e-6,
+        qk_layernorm: bool = True,
+        device=None,
+        dtype=None,
+        operations=None,
+        **kwargs
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.hidden_size = hidden_size
+        self.num_heads = num_attention_heads
+        self.head_dim = hidden_size // num_attention_heads
+        self.patch_size = patch_size
+        self.out_channels = out_channels
+
+        Linear = operations.Linear
+
+        self.x_embedder = ErnieImagePatchEmbedDynamic(in_channels, hidden_size, patch_size, operations, device, dtype)
+        self.text_proj = Linear(text_in_dim, hidden_size, bias=False, device=device, dtype=dtype) if text_in_dim != hidden_size else None
+
+        self.time_proj = Timesteps(hidden_size, flip_sin_to_cos=False)
+        self.time_embedding = TimestepEmbedding(hidden_size, hidden_size, operations, device, dtype)
+
+        self.pos_embed = ErnieImageEmbedND3(dim=self.head_dim, theta=rope_theta, axes_dim=rope_axes_dim)
+
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            Linear(hidden_size, 6 * hidden_size, device=device, dtype=dtype)
+        )
+
+        self.layers = nn.ModuleList([
+            ErnieImageSharedAdaLNBlock(hidden_size, num_attention_heads, ffn_hidden_size, eps, operations, device, dtype)
+            for _ in range(num_layers)
+        ])
+
+        self.final_norm = ErnieImageAdaLNContinuous(hidden_size, eps, operations, device, dtype)
+        self.final_linear = Linear(hidden_size, patch_size * patch_size * out_channels, device=device, dtype=dtype)
+
+    def forward(self, x, timesteps, context, **kwargs):
+        device, dtype = x.device, x.dtype
+        B, C, H, W = x.shape
+        p, Hp, Wp = self.patch_size, H // self.patch_size, W // self.patch_size
+        N_img = Hp * Wp
+
+        img_bsh = self.x_embedder(x)
+
+        text_bth = context
+        if self.text_proj is not None and text_bth.numel() > 0:
+            text_bth = self.text_proj(text_bth)
+        Tmax = text_bth.shape[1]
+
+        hidden_states = torch.cat([img_bsh, text_bth], dim=1)
+
+        text_ids = torch.zeros((B, Tmax, 3), device=device, dtype=torch.float32)
+        text_ids[:, :, 0] = torch.linspace(0, Tmax - 1, steps=Tmax, device=x.device, dtype=torch.float32)
+        index = float(Tmax)
+
+        transformer_options = kwargs.get("transformer_options", {})
+        rope_options = transformer_options.get("rope_options", None)
+
+        h_len, w_len = float(Hp), float(Wp)
+        h_offset, w_offset = 0.0, 0.0
+
+        if rope_options is not None:
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+            index += rope_options.get("shift_t", 0.0)
+            h_offset += rope_options.get("shift_y", 0.0)
+            w_offset += rope_options.get("shift_x", 0.0)
+
+        image_ids = torch.zeros((Hp, Wp, 3), device=device, dtype=torch.float32)
+        image_ids[:, :, 0] = image_ids[:, :, 1] + index
+        image_ids[:, :, 1] = image_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=Hp, device=device, dtype=torch.float32).unsqueeze(1)
+        image_ids[:, :, 2] = image_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=Wp, device=device, dtype=torch.float32).unsqueeze(0)
+
+        image_ids = image_ids.view(1, N_img, 3).expand(B, -1, -1)
+
+        rotary_pos_emb = self.pos_embed(torch.cat([image_ids, text_ids], dim=1))
+        del image_ids, text_ids
+
+        sample = self.time_proj(timesteps).to(dtype)
+        c = self.time_embedding(sample)
+
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = [
+            t.unsqueeze(1).contiguous() for t in self.adaLN_modulation(c).chunk(6, dim=-1)
+        ]
+
+        temb = [shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp]
+        for layer in self.layers:
+            hidden_states = layer(hidden_states, rotary_pos_emb, temb)
+
+        hidden_states = self.final_norm(hidden_states, c).type_as(hidden_states)
+
+        patches = self.final_linear(hidden_states)[:, :N_img, :]
+        output = (
+            patches.view(B, Hp, Wp, p, p, self.out_channels)
+            .permute(0, 5, 1, 3, 2, 4)
+            .contiguous()
+            .view(B, self.out_channels, H, W)
+        )
+
+        return output

ComfyUI/comfy/ldm/flux/controlnet.py

@@ -0,0 +1,208 @@
+#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", []))

ComfyUI/comfy/ldm/flux/layers.py

@@ -0,0 +1,372 @@
+import math
+from dataclasses import dataclass
+
+import torch
+from torch import Tensor, nn
+
+from .math import attention, rope
+
+# Fix import for some custom nodes, TODO: delete eventually.
+RMSNorm = None
+
+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, bias=True, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.in_layer = operations.Linear(in_dim, hidden_dim, bias=bias, dtype=dtype, device=device)
+        self.silu = nn.SiLU()
+        self.out_layer = operations.Linear(hidden_dim, hidden_dim, bias=bias, dtype=dtype, device=device)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.out_layer(self.silu(self.in_layer(x)))
+
+class YakMLP(nn.Module):
+    def __init__(self, hidden_size: int, intermediate_size: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.gate_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=True, dtype=dtype, device=device)
+        self.up_proj = operations.Linear(self.hidden_size, self.intermediate_size, bias=True, dtype=dtype, device=device)
+        self.down_proj = operations.Linear(self.intermediate_size, self.hidden_size, bias=True, dtype=dtype, device=device)
+        self.act_fn = nn.SiLU()
+
+    def forward(self, x: Tensor) -> Tensor:
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+def build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=False, yak_mlp=False, dtype=None, device=None, operations=None):
+    if yak_mlp:
+        return YakMLP(hidden_size, mlp_hidden_dim, dtype=dtype, device=device, operations=operations)
+    if mlp_silu_act:
+        return nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim * 2, bias=False, dtype=dtype, device=device),
+            SiLUActivation(),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=False, dtype=dtype, device=device),
+        )
+    else:
+        return 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),
+        )
+
+
+class QKNorm(torch.nn.Module):
+    def __init__(self, dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.query_norm = operations.RMSNorm(dim, dtype=dtype, device=device)
+        self.key_norm = operations.RMSNorm(dim, dtype=dtype, device=device)
+
+    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, proj_bias: bool = True, 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, bias=proj_bias, dtype=dtype, device=device)
+
+
+@dataclass
+class ModulationOut:
+    shift: Tensor
+    scale: Tensor
+    gate: Tensor
+
+
+class Modulation(nn.Module):
+    def __init__(self, dim: int, double: bool, bias=True, 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=bias, 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]:d[2] + 1]
+            if m_add is not None:
+                tensor[:, d[0]:d[1]] += m_add[:, d[2]:d[2] + 1]
+        return tensor
+
+
+class SiLUActivation(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.gate_fn = nn.SiLU()
+
+    def forward(self, x: Tensor) -> Tensor:
+        x1, x2 = x.chunk(2, dim=-1)
+        return self.gate_fn(x1) * x2
+
+
+class DoubleStreamBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=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.modulation = modulation
+
+        if self.modulation:
+            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, proj_bias=proj_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 = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations)
+
+        if self.modulation:
+            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, proj_bias=proj_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 = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}):
+        if self.modulation:
+            img_mod1, img_mod2 = self.img_mod(vec)
+            txt_mod1, txt_mod2 = self.txt_mod(vec)
+        else:
+            (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
+
+        transformer_patches = transformer_options.get("patches", {})
+        extra_options = transformer_options.copy()
+
+        # 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)
+        del 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)
+        del img_qkv
+        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)
+        del 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)
+        del txt_qkv
+        txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
+
+        q = torch.cat((txt_q, img_q), dim=2)
+        del txt_q, img_q
+        k = torch.cat((txt_k, img_k), dim=2)
+        del txt_k, img_k
+        v = torch.cat((txt_v, img_v), dim=2)
+        del txt_v, img_v
+
+        extra_options["img_slice"] = [txt.shape[1], q.shape[2]]
+        if "attn1_patch" in transformer_patches:
+            patch = transformer_patches["attn1_patch"]
+            for p in patch:
+                out = p(q, k, v, pe=pe, attn_mask=attn_mask, extra_options=extra_options)
+                q, k, v, pe, attn_mask = out.get("q", q), out.get("k", k), out.get("v", v), out.get("pe", pe), out.get("attn_mask", attn_mask)
+
+        # run actual attention
+        attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
+        del q, k, v
+
+        if "attn1_output_patch" in transformer_patches:
+            patch = transformer_patches["attn1_output_patch"]
+            for p in patch:
+                attn = p(attn, extra_options)
+
+        txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
+
+        # calculate the img bloks
+        img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
+        del img_attn
+        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)
+        del txt_attn
+        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,
+        modulation=True,
+        mlp_silu_act=False,
+        bias=True,
+        yak_mlp=False,
+        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)
+
+        self.mlp_hidden_dim_first = self.mlp_hidden_dim
+        self.yak_mlp = yak_mlp
+        if mlp_silu_act:
+            self.mlp_hidden_dim_first = int(hidden_size * mlp_ratio * 2)
+            self.mlp_act = SiLUActivation()
+        else:
+            self.mlp_act = nn.GELU(approximate="tanh")
+
+        if self.yak_mlp:
+            self.mlp_hidden_dim_first *= 2
+            self.mlp_act = nn.SiLU()
+
+        # qkv and mlp_in
+        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim_first, bias=bias, dtype=dtype, device=device)
+        # proj and mlp_out
+        self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, bias=bias, 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)
+
+        if modulation:
+            self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
+        else:
+            self.modulation = None
+
+    def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims=None, transformer_options={}) -> Tensor:
+        if self.modulation:
+            mod, _ = self.modulation(vec)
+        else:
+            mod = vec
+
+        transformer_patches = transformer_options.get("patches", {})
+        extra_options = transformer_options.copy()
+
+        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_first], dim=-1)
+
+        q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        del qkv
+        q, k = self.norm(q, k, v)
+
+        if "attn1_patch" in transformer_patches:
+            patch = transformer_patches["attn1_patch"]
+            for p in patch:
+                out = p(q, k, v, pe=pe, attn_mask=attn_mask, extra_options=extra_options)
+                q, k, v, pe, attn_mask = out.get("q", q), out.get("k", k), out.get("v", v), out.get("pe", pe), out.get("attn_mask", attn_mask)
+
+        # compute attention
+        attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
+        del q, k, v
+
+        if "attn1_output_patch" in transformer_patches:
+            patch = transformer_patches["attn1_output_patch"]
+            for p in patch:
+                attn = p(attn, extra_options)
+
+        # compute activation in mlp stream, cat again and run second linear layer
+        if self.yak_mlp:
+            mlp = self.mlp_act(mlp[..., self.mlp_hidden_dim_first // 2:]) * mlp[..., :self.mlp_hidden_dim_first // 2]
+        else:
+            mlp = self.mlp_act(mlp)
+        output = self.linear2(torch.cat((attn, 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, bias=True, 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=bias, dtype=dtype, device=device)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=bias, 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

ComfyUI/comfy/ldm/flux/math.py

@@ -0,0 +1,58 @@
+import torch
+from einops import rearrange
+from torch import Tensor
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+import comfy.quant_ops
+
+
+def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transformer_options={}) -> Tensor:
+    if pe is not None:
+        q, k = apply_rope(q, k, pe)
+    heads = q.shape[1]
+    x = optimized_attention(q, k, v, heads, skip_reshape=True, mask=mask, transformer_options=transformer_options)
+    return x
+
+def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
+    assert dim % 2 == 0
+    if not comfy.model_management.supports_fp64(pos.device):
+        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_rope1(x: Tensor, freqs_cis: Tensor):
+    x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
+    if x_.shape[2] != 1 and freqs_cis.shape[2] != 1 and x_.shape[2] != freqs_cis.shape[2]:
+        freqs_cis = freqs_cis[:, :, :x_.shape[2]]
+
+    x_out = freqs_cis[..., 0] * x_[..., 0]
+    x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
+
+    return x_out.reshape(*x.shape).type_as(x)
+
+
+def _apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
+    return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+
+
+def apply_rope(xq, xk, freqs_cis):
+    if comfy.model_management.in_training:
+        return _apply_rope(xq, xk, freqs_cis)
+    else:
+        return comfy.quant_ops.ck.apply_rope(xq, xk, freqs_cis)
+
+
+def apply_rope1(x, freqs_cis):
+    if comfy.model_management.in_training:
+        return _apply_rope1(x, freqs_cis)
+    else:
+        return comfy.quant_ops.ck.apply_rope1(x, freqs_cis)

ComfyUI/comfy/ldm/flux/model.py

@@ -0,0 +1,408 @@
+#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,
+    Modulation,
+)
+
+@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
+    txt_ids_dims: list
+    global_modulation: bool = False
+    mlp_silu_act: bool = False
+    ops_bias: bool = True
+    default_ref_method: str = "offset"
+    ref_index_scale: float = 1.0
+    yak_mlp: bool = False
+    txt_norm: bool = False
+
+
+def invert_slices(slices, length):
+    sorted_slices = sorted(slices)
+    result = []
+    current = 0
+
+    for start, end in sorted_slices:
+        if current < start:
+            result.append((current, start))
+        current = max(current, end)
+
+    if current < length:
+        result.append((current, length))
+
+    return result
+
+
+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=params.ops_bias, dtype=dtype, device=device)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device, operations=operations)
+        if params.vec_in_dim is not None:
+            self.vector_in = MLPEmbedder(params.vec_in_dim, self.hidden_size, dtype=dtype, device=device, operations=operations)
+        else:
+            self.vector_in = None
+
+        self.guidance_in = (
+            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, bias=params.ops_bias, 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, bias=params.ops_bias, dtype=dtype, device=device)
+
+        if params.txt_norm:
+            self.txt_norm = operations.RMSNorm(params.context_in_dim, dtype=dtype, device=device)
+        else:
+            self.txt_norm = None
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                    modulation=params.global_modulation is False,
+                    mlp_silu_act=params.mlp_silu_act,
+                    proj_bias=params.ops_bias,
+                    yak_mlp=params.yak_mlp,
+                    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, modulation=params.global_modulation is False, mlp_silu_act=params.mlp_silu_act, bias=params.ops_bias, yak_mlp=params.yak_mlp, 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, bias=params.ops_bias, dtype=dtype, device=device, operations=operations)
+
+        if params.global_modulation:
+            self.double_stream_modulation_img = Modulation(
+                self.hidden_size,
+                double=True,
+                bias=False,
+                dtype=dtype, device=device, operations=operations
+            )
+            self.double_stream_modulation_txt = Modulation(
+                self.hidden_size,
+                double=True,
+                bias=False,
+                dtype=dtype, device=device, operations=operations
+            )
+            self.single_stream_modulation = Modulation(
+                self.hidden_size, double=False, bias=False, 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,
+        timestep_zero_index=None,
+        transformer_options={},
+        attn_mask: Tensor = None,
+    ) -> Tensor:
+
+        transformer_options = transformer_options.copy()
+        patches = transformer_options.get("patches", {})
+        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))
+
+        if self.vector_in is not None:
+            if y is None:
+                y = torch.zeros((img.shape[0], self.params.vec_in_dim), device=img.device, dtype=img.dtype)
+            vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
+
+        if self.txt_norm is not None:
+            txt = self.txt_norm(txt)
+        txt = self.txt_in(txt)
+
+        if "post_input" in patches:
+            for p in patches["post_input"]:
+                out = p({"img": img, "txt": txt, "img_ids": img_ids, "txt_ids": txt_ids, "transformer_options": transformer_options})
+                img = out["img"]
+                txt = out["txt"]
+                img_ids = out["img_ids"]
+                txt_ids = out["txt_ids"]
+
+        if img_ids is not None:
+            ids = torch.cat((txt_ids, img_ids), dim=1)
+            pe = self.pe_embedder(ids)
+        else:
+            pe = None
+
+        vec_orig = vec
+        txt_vec = vec
+        extra_kwargs = {}
+        if timestep_zero_index is not None:
+            modulation_dims = []
+            batch = vec.shape[0] // 2
+            vec_orig = vec_orig.reshape(2, batch, vec.shape[1]).movedim(0, 1)
+            invert = invert_slices(timestep_zero_index, img.shape[1])
+            for s in invert:
+                modulation_dims.append((s[0], s[1], 0))
+            for s in timestep_zero_index:
+                modulation_dims.append((s[0], s[1], 1))
+            extra_kwargs["modulation_dims_img"] = modulation_dims
+            txt_vec = vec[:batch]
+
+        if self.params.global_modulation:
+            vec = (self.double_stream_modulation_img(vec_orig), self.double_stream_modulation_txt(txt_vec))
+
+        blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.double_blocks)
+        transformer_options["block_type"] = "double"
+        for i, block in enumerate(self.double_blocks):
+            transformer_options["block_index"] = i
+            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"),
+                                                   transformer_options=args.get("transformer_options"),
+                                                   **extra_kwargs)
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": img,
+                                                           "txt": txt,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
+                                                          {"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,
+                                 transformer_options=transformer_options,
+                                 **extra_kwargs)
+
+            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)
+
+        if self.params.global_modulation:
+            vec, _ = self.single_stream_modulation(vec_orig)
+
+        extra_kwargs = {}
+        if timestep_zero_index is not None:
+            lambda a: 0 if a == 0 else a + txt.shape[1]
+            modulation_dims_combined = list(map(lambda x: (0 if x[0] == 0 else x[0] + txt.shape[1], x[1] + txt.shape[1], x[2]), modulation_dims))
+            extra_kwargs["modulation_dims"] = modulation_dims_combined
+
+        transformer_options["total_blocks"] = len(self.single_blocks)
+        transformer_options["block_type"] = "single"
+        transformer_options["img_slice"] = [txt.shape[1], img.shape[1]]
+        for i, block in enumerate(self.single_blocks):
+            transformer_options["block_index"] = i
+            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"),
+                                       transformer_options=args.get("transformer_options"),
+                                       **extra_kwargs)
+                    return out
+
+                out = blocks_replace[("single_block", i)]({"img": img,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
+                                                          {"original_block": block_wrap})
+                img = out["img"]
+            else:
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, transformer_options=transformer_options, **extra_kwargs)
+
+            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] :, ...]
+
+        extra_kwargs = {}
+        if timestep_zero_index is not None:
+            extra_kwargs["modulation_dims"] = modulation_dims
+
+        img = self.final_layer(img, vec_orig, **extra_kwargs)  # (N, T, patch_size ** 2 * out_channels)
+        return img
+
+    def process_img(self, x, index=0, h_offset=0, w_offset=0, transformer_options={}):
+        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)
+
+        steps_h = h_len
+        steps_w = w_len
+
+        rope_options = transformer_options.get("rope_options", None)
+        if rope_options is not None:
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+
+            index += rope_options.get("shift_t", 0.0)
+            h_offset += rope_options.get("shift_y", 0.0)
+            w_offset += rope_options.get("shift_x", 0.0)
+
+        img_ids = torch.zeros((steps_h, steps_w, len(self.params.axes_dim)), device=x.device, dtype=torch.float32)
+        img_ids[:, :, 0] = img_ids[:, :, 1] + index
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=torch.float32).unsqueeze(1)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=torch.float32).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, transformer_options=transformer_options)
+        img_tokens = img.shape[1]
+        timestep_zero_index = None
+        if ref_latents is not None:
+            ref_num_tokens = []
+            h = 0
+            w = 0
+            index = 0
+            ref_latents_method = kwargs.get("ref_latents_method", self.params.default_ref_method)
+            timestep_zero = ref_latents_method == "index_timestep_zero"
+            for ref in ref_latents:
+                if ref_latents_method in ("index", "index_timestep_zero"):
+                    index += self.params.ref_index_scale
+                    h_offset = 0
+                    w_offset = 0
+                elif ref_latents_method == "uxo":
+                    index = 0
+                    h_offset = h_len * patch_size + h
+                    w_offset = w_len * patch_size + w
+                    h += ref.shape[-2]
+                    w += ref.shape[-1]
+                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, transformer_options=transformer_options)
+                img = torch.cat([img, kontext], dim=1)
+                img_ids = torch.cat([img_ids, kontext_ids], dim=1)
+                ref_num_tokens.append(kontext.shape[1])
+            if timestep_zero:
+                if index > 0:
+                    timestep = torch.cat([timestep, timestep * 0], dim=0)
+                    timestep_zero_index = [[img_tokens, img_ids.shape[1]]]
+            transformer_options = transformer_options.copy()
+            transformer_options["reference_image_num_tokens"] = ref_num_tokens
+
+        txt_ids = torch.zeros((bs, context.shape[1], len(self.params.axes_dim)), device=x.device, dtype=torch.float32)
+
+        if len(self.params.txt_ids_dims) > 0:
+            for i in self.params.txt_ids_dims:
+                txt_ids[:, :, i] = torch.linspace(0, context.shape[1] - 1, steps=context.shape[1], device=x.device, dtype=torch.float32)
+
+        out = self.forward_orig(img, img_ids, context, txt_ids, timestep, y, guidance, control, timestep_zero_index=timestep_zero_index, transformer_options=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=self.patch_size, pw=self.patch_size)[:,:,:h_orig,:w_orig]

ComfyUI/comfy/ldm/flux/redux.py

@@ -0,0 +1,25 @@
+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

ComfyUI/comfy/ldm/genmo/joint_model/asymm_models_joint.py

@@ -0,0 +1,561 @@
+#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,
+        transformer_options={},
+        **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, transformer_options=transformer_options)
+
+        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,
+        transformer_options={},
+        **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,
+            transformer_options=transformer_options,
+            **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"],
+                                                    transformer_options=args["transformer_options"]
+                                                    )
+                    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, "transformer_options": transformer_options}, {"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,
+                    transformer_options=transformer_options,
+                )  # (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

ComfyUI/comfy/ldm/genmo/joint_model/layers.py

@@ -0,0 +1,153 @@
+#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

ComfyUI/comfy/ldm/genmo/joint_model/rope_mixed.py

@@ -0,0 +1,88 @@
+#original code from https://github.com/genmoai/models under apache 2.0 license
+
+# import functools
+import math
+
+import torch
+
+
+def centers(start: float, stop, num, dtype=None, device=None):
+    """linspace through bin centers.
+
+    Args:
+        start (float): Start of the range.
+        stop (float): End of the range.
+        num (int): Number of points.
+        dtype (torch.dtype): Data type of the points.
+        device (torch.device): Device of the points.
+
+    Returns:
+        centers (Tensor): Centers of the bins. Shape: (num,).
+    """
+    edges = torch.linspace(start, stop, num + 1, dtype=dtype, device=device)
+    return (edges[:-1] + edges[1:]) / 2
+
+
+# @functools.lru_cache(maxsize=1)
+def create_position_matrix(
+    T: int,
+    pH: int,
+    pW: int,
+    device: torch.device,
+    dtype: torch.dtype,
+    *,
+    target_area: float = 36864,
+):
+    """
+    Args:
+        T: int - Temporal dimension
+        pH: int - Height dimension after patchify
+        pW: int - Width dimension after patchify
+
+    Returns:
+        pos: [T * pH * pW, 3] - position matrix
+    """
+    # Create 1D tensors for each dimension
+    t = torch.arange(T, dtype=dtype)
+
+    # Positionally interpolate to area 36864.
+    # (3072x3072 frame with 16x16 patches = 192x192 latents).
+    # This automatically scales rope positions when the resolution changes.
+    # We use a large target area so the model is more sensitive
+    # to changes in the learned pos_frequencies matrix.
+    scale = math.sqrt(target_area / (pW * pH))
+    w = centers(-pW * scale / 2, pW * scale / 2, pW)
+    h = centers(-pH * scale / 2, pH * scale / 2, pH)
+
+    # Use meshgrid to create 3D grids
+    grid_t, grid_h, grid_w = torch.meshgrid(t, h, w, indexing="ij")
+
+    # Stack and reshape the grids.
+    pos = torch.stack([grid_t, grid_h, grid_w], dim=-1)  # [T, pH, pW, 3]
+    pos = pos.view(-1, 3)  # [T * pH * pW, 3]
+    pos = pos.to(dtype=dtype, device=device)
+
+    return pos
+
+
+def compute_mixed_rotation(
+    freqs: torch.Tensor,
+    pos: torch.Tensor,
+):
+    """
+    Project each 3-dim position into per-head, per-head-dim 1D frequencies.
+
+    Args:
+        freqs: [3, num_heads, num_freqs] - learned rotation frequency (for t, row, col) for each head position
+        pos: [N, 3] - position of each token
+        num_heads: int
+
+    Returns:
+        freqs_cos: [N, num_heads, num_freqs] - cosine components
+        freqs_sin: [N, num_heads, num_freqs] - sine components
+    """
+    assert freqs.ndim == 3
+    freqs_sum = torch.einsum("Nd,dhf->Nhf", pos.to(freqs), freqs)
+    freqs_cos = torch.cos(freqs_sum)
+    freqs_sin = torch.sin(freqs_sum)
+    return freqs_cos, freqs_sin

ComfyUI/comfy/ldm/genmo/joint_model/temporal_rope.py

@@ -0,0 +1,34 @@
+#original code from https://github.com/genmoai/models under apache 2.0 license
+
+# Based on Llama3 Implementation.
+import torch
+
+
+def apply_rotary_emb_qk_real(
+    xqk: torch.Tensor,
+    freqs_cos: torch.Tensor,
+    freqs_sin: torch.Tensor,
+) -> torch.Tensor:
+    """
+    Apply rotary embeddings to input tensors using the given frequency tensor without complex numbers.
+
+    Args:
+        xqk (torch.Tensor): Query and/or Key tensors to apply rotary embeddings. Shape: (B, S, *, num_heads, D)
+                            Can be either just query or just key, or both stacked along some batch or * dim.
+        freqs_cos (torch.Tensor): Precomputed cosine frequency tensor.
+        freqs_sin (torch.Tensor): Precomputed sine frequency tensor.
+
+    Returns:
+        torch.Tensor: The input tensor with rotary embeddings applied.
+    """
+    # Split the last dimension into even and odd parts
+    xqk_even = xqk[..., 0::2]
+    xqk_odd = xqk[..., 1::2]
+
+    # Apply rotation
+    cos_part = (xqk_even * freqs_cos - xqk_odd * freqs_sin).type_as(xqk)
+    sin_part = (xqk_even * freqs_sin + xqk_odd * freqs_cos).type_as(xqk)
+
+    # Interleave the results back into the original shape
+    out = torch.stack([cos_part, sin_part], dim=-1).flatten(-2)
+    return out

ComfyUI/comfy/ldm/genmo/joint_model/utils.py

@@ -0,0 +1,102 @@
+#original code from https://github.com/genmoai/models under apache 2.0 license
+#adapted to ComfyUI
+
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def modulate(x, shift, scale):
+    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+
+def pool_tokens(x: torch.Tensor, mask: torch.Tensor, *, keepdim=False) -> torch.Tensor:
+    """
+    Pool tokens in x using mask.
+
+    NOTE: We assume x does not require gradients.
+
+    Args:
+        x: (B, L, D) tensor of tokens.
+        mask: (B, L) boolean tensor indicating which tokens are not padding.
+
+    Returns:
+        pooled: (B, D) tensor of pooled tokens.
+    """
+    assert x.size(1) == mask.size(1)  # Expected mask to have same length as tokens.
+    assert x.size(0) == mask.size(0)  # Expected mask to have same batch size as tokens.
+    mask = mask[:, :, None].to(dtype=x.dtype)
+    mask = mask / mask.sum(dim=1, keepdim=True).clamp(min=1)
+    pooled = (x * mask).sum(dim=1, keepdim=keepdim)
+    return pooled
+
+
+class AttentionPool(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        output_dim: int = None,
+        device: Optional[torch.device] = None,
+        dtype=None,
+        operations=None,
+    ):
+        """
+        Args:
+            spatial_dim (int): Number of tokens in sequence length.
+            embed_dim (int): Dimensionality of input tokens.
+            num_heads (int): Number of attention heads.
+            output_dim (int): Dimensionality of output tokens. Defaults to embed_dim.
+        """
+        super().__init__()
+        self.num_heads = num_heads
+        self.to_kv = operations.Linear(embed_dim, 2 * embed_dim, device=device, dtype=dtype)
+        self.to_q = operations.Linear(embed_dim, embed_dim, device=device, dtype=dtype)
+        self.to_out = operations.Linear(embed_dim, output_dim or embed_dim, device=device, dtype=dtype)
+
+    def forward(self, x, mask):
+        """
+        Args:
+            x (torch.Tensor): (B, L, D) tensor of input tokens.
+            mask (torch.Tensor): (B, L) boolean tensor indicating which tokens are not padding.
+
+        NOTE: We assume x does not require gradients.
+
+        Returns:
+            x (torch.Tensor): (B, D) tensor of pooled tokens.
+        """
+        D = x.size(2)
+
+        # Construct attention mask, shape: (B, 1, num_queries=1, num_keys=1+L).
+        attn_mask = mask[:, None, None, :].bool()  # (B, 1, 1, L).
+        attn_mask = F.pad(attn_mask, (1, 0), value=True)  # (B, 1, 1, 1+L).
+
+        # Average non-padding token features. These will be used as the query.
+        x_pool = pool_tokens(x, mask, keepdim=True)  # (B, 1, D)
+
+        # Concat pooled features to input sequence.
+        x = torch.cat([x_pool, x], dim=1)  # (B, L+1, D)
+
+        # Compute queries, keys, values. Only the mean token is used to create a query.
+        kv = self.to_kv(x)  # (B, L+1, 2 * D)
+        q = self.to_q(x[:, 0])  # (B, D)
+
+        # Extract heads.
+        head_dim = D // self.num_heads
+        kv = kv.unflatten(2, (2, self.num_heads, head_dim))  # (B, 1+L, 2, H, head_dim)
+        kv = kv.transpose(1, 3)  # (B, H, 2, 1+L, head_dim)
+        k, v = kv.unbind(2)  # (B, H, 1+L, head_dim)
+        q = q.unflatten(1, (self.num_heads, head_dim))  # (B, H, head_dim)
+        q = q.unsqueeze(2)  # (B, H, 1, head_dim)
+
+        # Compute attention.
+        x = F.scaled_dot_product_attention(
+            q, k, v, attn_mask=attn_mask, dropout_p=0.0
+        )  # (B, H, 1, head_dim)
+
+        # Concatenate heads and run output.
+        x = x.squeeze(2).flatten(1, 2)  # (B, D = H * head_dim)
+        x = self.to_out(x)
+        return x

ComfyUI/comfy/ldm/genmo/vae/model.py

@@ -0,0 +1,711 @@
+#original code from https://github.com/genmoai/models under apache 2.0 license
+#adapted to ComfyUI
+
+from typing import List, Optional, Tuple, Union
+from functools import partial
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+from comfy.ldm.modules.attention import optimized_attention
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+# import mochi_preview.dit.joint_model.context_parallel as cp
+# from mochi_preview.vae.cp_conv import cp_pass_frames, gather_all_frames
+
+
+def cast_tuple(t, length=1):
+    return t if isinstance(t, tuple) else ((t,) * length)
+
+
+class GroupNormSpatial(ops.GroupNorm):
+    """
+    GroupNorm applied per-frame.
+    """
+
+    def forward(self, x: torch.Tensor, *, chunk_size: int = 8):
+        B, C, T, H, W = x.shape
+        x = rearrange(x, "B C T H W -> (B T) C H W")
+        # Run group norm in chunks.
+        output = torch.empty_like(x)
+        for b in range(0, B * T, chunk_size):
+            output[b : b + chunk_size] = super().forward(x[b : b + chunk_size])
+        return rearrange(output, "(B T) C H W -> B C T H W", B=B, T=T)
+
+class PConv3d(ops.Conv3d):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size: Union[int, Tuple[int, int, int]],
+        stride: Union[int, Tuple[int, int, int]],
+        causal: bool = True,
+        context_parallel: bool = True,
+        **kwargs,
+    ):
+        self.causal = causal
+        self.context_parallel = context_parallel
+        kernel_size = cast_tuple(kernel_size, 3)
+        stride = cast_tuple(stride, 3)
+        height_pad = (kernel_size[1] - 1) // 2
+        width_pad = (kernel_size[2] - 1) // 2
+
+        super().__init__(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            dilation=(1, 1, 1),
+            padding=(0, height_pad, width_pad),
+            **kwargs,
+        )
+
+    def forward(self, x: torch.Tensor):
+        # Compute padding amounts.
+        context_size = self.kernel_size[0] - 1
+        if self.causal:
+            pad_front = context_size
+            pad_back = 0
+        else:
+            pad_front = context_size // 2
+            pad_back = context_size - pad_front
+
+        # Apply padding.
+        assert self.padding_mode == "replicate"  # DEBUG
+        mode = "constant" if self.padding_mode == "zeros" else self.padding_mode
+        x = F.pad(x, (0, 0, 0, 0, pad_front, pad_back), mode=mode)
+        return super().forward(x)
+
+
+class Conv1x1(ops.Linear):
+    """*1x1 Conv implemented with a linear layer."""
+
+    def __init__(self, in_features: int, out_features: int, *args, **kwargs):
+        super().__init__(in_features, out_features, *args, **kwargs)
+
+    def forward(self, x: torch.Tensor):
+        """Forward pass.
+
+        Args:
+            x: Input tensor. Shape: [B, C, *] or [B, *, C].
+
+        Returns:
+            x: Output tensor. Shape: [B, C', *] or [B, *, C'].
+        """
+        x = x.movedim(1, -1)
+        x = super().forward(x)
+        x = x.movedim(-1, 1)
+        return x
+
+
+class DepthToSpaceTime(nn.Module):
+    def __init__(
+        self,
+        temporal_expansion: int,
+        spatial_expansion: int,
+    ):
+        super().__init__()
+        self.temporal_expansion = temporal_expansion
+        self.spatial_expansion = spatial_expansion
+
+    # When printed, this module should show the temporal and spatial expansion factors.
+    def extra_repr(self):
+        return f"texp={self.temporal_expansion}, sexp={self.spatial_expansion}"
+
+    def forward(self, x: torch.Tensor):
+        """Forward pass.
+
+        Args:
+            x: Input tensor. Shape: [B, C, T, H, W].
+
+        Returns:
+            x: Rearranged tensor. Shape: [B, C/(st*s*s), T*st, H*s, W*s].
+        """
+        x = rearrange(
+            x,
+            "B (C st sh sw) T H W -> B C (T st) (H sh) (W sw)",
+            st=self.temporal_expansion,
+            sh=self.spatial_expansion,
+            sw=self.spatial_expansion,
+        )
+
+        # cp_rank, _ = cp.get_cp_rank_size()
+        if self.temporal_expansion > 1: # and cp_rank == 0:
+            # Drop the first self.temporal_expansion - 1 frames.
+            # This is because we always want the 3x3x3 conv filter to only apply
+            # to the first frame, and the first frame doesn't need to be repeated.
+            assert all(x.shape)
+            x = x[:, :, self.temporal_expansion - 1 :]
+            assert all(x.shape)
+
+        return x
+
+
+def norm_fn(
+    in_channels: int,
+    affine: bool = True,
+):
+    return GroupNormSpatial(affine=affine, num_groups=32, num_channels=in_channels)
+
+
+class ResBlock(nn.Module):
+    """Residual block that preserves the spatial dimensions."""
+
+    def __init__(
+        self,
+        channels: int,
+        *,
+        affine: bool = True,
+        attn_block: Optional[nn.Module] = None,
+        causal: bool = True,
+        prune_bottleneck: bool = False,
+        padding_mode: str,
+        bias: bool = True,
+    ):
+        super().__init__()
+        self.channels = channels
+
+        assert causal
+        self.stack = nn.Sequential(
+            norm_fn(channels, affine=affine),
+            nn.SiLU(inplace=True),
+            PConv3d(
+                in_channels=channels,
+                out_channels=channels // 2 if prune_bottleneck else channels,
+                kernel_size=(3, 3, 3),
+                stride=(1, 1, 1),
+                padding_mode=padding_mode,
+                bias=bias,
+                causal=causal,
+            ),
+            norm_fn(channels, affine=affine),
+            nn.SiLU(inplace=True),
+            PConv3d(
+                in_channels=channels // 2 if prune_bottleneck else channels,
+                out_channels=channels,
+                kernel_size=(3, 3, 3),
+                stride=(1, 1, 1),
+                padding_mode=padding_mode,
+                bias=bias,
+                causal=causal,
+            ),
+        )
+
+        self.attn_block = attn_block if attn_block else nn.Identity()
+
+    def forward(self, x: torch.Tensor):
+        """Forward pass.
+
+        Args:
+            x: Input tensor. Shape: [B, C, T, H, W].
+        """
+        residual = x
+        x = self.stack(x)
+        x = x + residual
+        del residual
+
+        return self.attn_block(x)
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        head_dim: int = 32,
+        qkv_bias: bool = False,
+        out_bias: bool = True,
+        qk_norm: bool = True,
+    ) -> None:
+        super().__init__()
+        self.head_dim = head_dim
+        self.num_heads = dim // head_dim
+        self.qk_norm = qk_norm
+
+        self.qkv = nn.Linear(dim, 3 * dim, bias=qkv_bias)
+        self.out = nn.Linear(dim, dim, bias=out_bias)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+    ) -> torch.Tensor:
+        """Compute temporal self-attention.
+
+        Args:
+            x: Input tensor. Shape: [B, C, T, H, W].
+            chunk_size: Chunk size for large tensors.
+
+        Returns:
+            x: Output tensor. Shape: [B, C, T, H, W].
+        """
+        B, _, T, H, W = x.shape
+
+        if T == 1:
+            # No attention for single frame.
+            x = x.movedim(1, -1)  # [B, C, T, H, W] -> [B, T, H, W, C]
+            qkv = self.qkv(x)
+            _, _, x = qkv.chunk(3, dim=-1)  # Throw away queries and keys.
+            x = self.out(x)
+            return x.movedim(-1, 1)  # [B, T, H, W, C] -> [B, C, T, H, W]
+
+        # 1D temporal attention.
+        x = rearrange(x, "B C t h w -> (B h w) t C")
+        qkv = self.qkv(x)
+
+        # Input: qkv with shape [B, t, 3 * num_heads * head_dim]
+        # Output: x with shape [B, num_heads, t, head_dim]
+        q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, self.head_dim).transpose(1, 3).unbind(2)
+
+        if self.qk_norm:
+            q = F.normalize(q, p=2, dim=-1)
+            k = F.normalize(k, p=2, dim=-1)
+
+        x = optimized_attention(q, k, v, self.num_heads, skip_reshape=True)
+
+        assert x.size(0) == q.size(0)
+
+        x = self.out(x)
+        x = rearrange(x, "(B h w) t C -> B C t h w", B=B, h=H, w=W)
+        return x
+
+
+class AttentionBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        **attn_kwargs,
+    ) -> None:
+        super().__init__()
+        self.norm = norm_fn(dim)
+        self.attn = Attention(dim, **attn_kwargs)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x + self.attn(self.norm(x))
+
+
+class CausalUpsampleBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        num_res_blocks: int,
+        *,
+        temporal_expansion: int = 2,
+        spatial_expansion: int = 2,
+        **block_kwargs,
+    ):
+        super().__init__()
+
+        blocks = []
+        for _ in range(num_res_blocks):
+            blocks.append(block_fn(in_channels, **block_kwargs))
+        self.blocks = nn.Sequential(*blocks)
+
+        self.temporal_expansion = temporal_expansion
+        self.spatial_expansion = spatial_expansion
+
+        # Change channels in the final convolution layer.
+        self.proj = Conv1x1(
+            in_channels,
+            out_channels * temporal_expansion * (spatial_expansion**2),
+        )
+
+        self.d2st = DepthToSpaceTime(
+            temporal_expansion=temporal_expansion, spatial_expansion=spatial_expansion
+        )
+
+    def forward(self, x):
+        x = self.blocks(x)
+        x = self.proj(x)
+        x = self.d2st(x)
+        return x
+
+
+def block_fn(channels, *, affine: bool = True, has_attention: bool = False, **block_kwargs):
+    attn_block = AttentionBlock(channels) if has_attention else None
+    return ResBlock(channels, affine=affine, attn_block=attn_block, **block_kwargs)
+
+
+class DownsampleBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        num_res_blocks,
+        *,
+        temporal_reduction=2,
+        spatial_reduction=2,
+        **block_kwargs,
+    ):
+        """
+        Downsample block for the VAE encoder.
+
+        Args:
+            in_channels: Number of input channels.
+            out_channels: Number of output channels.
+            num_res_blocks: Number of residual blocks.
+            temporal_reduction: Temporal reduction factor.
+            spatial_reduction: Spatial reduction factor.
+        """
+        super().__init__()
+        layers = []
+
+        # Change the channel count in the strided convolution.
+        # This lets the ResBlock have uniform channel count,
+        # as in ConvNeXt.
+        assert in_channels != out_channels
+        layers.append(
+            PConv3d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=(temporal_reduction, spatial_reduction, spatial_reduction),
+                stride=(temporal_reduction, spatial_reduction, spatial_reduction),
+                # First layer in each block always uses replicate padding
+                padding_mode="replicate",
+                bias=block_kwargs["bias"],
+            )
+        )
+
+        for _ in range(num_res_blocks):
+            layers.append(block_fn(out_channels, **block_kwargs))
+
+        self.layers = nn.Sequential(*layers)
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+def add_fourier_features(inputs: torch.Tensor, start=6, stop=8, step=1):
+    num_freqs = (stop - start) // step
+    assert inputs.ndim == 5
+    C = inputs.size(1)
+
+    # Create Base 2 Fourier features.
+    freqs = torch.arange(start, stop, step, dtype=inputs.dtype, device=inputs.device)
+    assert num_freqs == len(freqs)
+    w = torch.pow(2.0, freqs) * (2 * torch.pi)  # [num_freqs]
+    C = inputs.shape[1]
+    w = w.repeat(C)[None, :, None, None, None]  # [1, C * num_freqs, 1, 1, 1]
+
+    # Interleaved repeat of input channels to match w.
+    h = inputs.repeat_interleave(num_freqs, dim=1)  # [B, C * num_freqs, T, H, W]
+    # Scale channels by frequency.
+    h = w * h
+
+    return torch.cat(
+        [
+            inputs,
+            torch.sin(h),
+            torch.cos(h),
+        ],
+        dim=1,
+    )
+
+
+class FourierFeatures(nn.Module):
+    def __init__(self, start: int = 6, stop: int = 8, step: int = 1):
+        super().__init__()
+        self.start = start
+        self.stop = stop
+        self.step = step
+
+    def forward(self, inputs):
+        """Add Fourier features to inputs.
+
+        Args:
+            inputs: Input tensor. Shape: [B, C, T, H, W]
+
+        Returns:
+            h: Output tensor. Shape: [B, (1 + 2 * num_freqs) * C, T, H, W]
+        """
+        return add_fourier_features(inputs, self.start, self.stop, self.step)
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        out_channels: int = 3,
+        latent_dim: int,
+        base_channels: int,
+        channel_multipliers: List[int],
+        num_res_blocks: List[int],
+        temporal_expansions: Optional[List[int]] = None,
+        spatial_expansions: Optional[List[int]] = None,
+        has_attention: List[bool],
+        output_norm: bool = True,
+        nonlinearity: str = "silu",
+        output_nonlinearity: str = "silu",
+        causal: bool = True,
+        **block_kwargs,
+    ):
+        super().__init__()
+        self.input_channels = latent_dim
+        self.base_channels = base_channels
+        self.channel_multipliers = channel_multipliers
+        self.num_res_blocks = num_res_blocks
+        self.output_nonlinearity = output_nonlinearity
+        assert nonlinearity == "silu"
+        assert causal
+
+        ch = [mult * base_channels for mult in channel_multipliers]
+        self.num_up_blocks = len(ch) - 1
+        assert len(num_res_blocks) == self.num_up_blocks + 2
+
+        blocks = []
+
+        first_block = [
+            ops.Conv3d(latent_dim, ch[-1], kernel_size=(1, 1, 1))
+        ]  # Input layer.
+        # First set of blocks preserve channel count.
+        for _ in range(num_res_blocks[-1]):
+            first_block.append(
+                block_fn(
+                    ch[-1],
+                    has_attention=has_attention[-1],
+                    causal=causal,
+                    **block_kwargs,
+                )
+            )
+        blocks.append(nn.Sequential(*first_block))
+
+        assert len(temporal_expansions) == len(spatial_expansions) == self.num_up_blocks
+        assert len(num_res_blocks) == len(has_attention) == self.num_up_blocks + 2
+
+        upsample_block_fn = CausalUpsampleBlock
+
+        for i in range(self.num_up_blocks):
+            block = upsample_block_fn(
+                ch[-i - 1],
+                ch[-i - 2],
+                num_res_blocks=num_res_blocks[-i - 2],
+                has_attention=has_attention[-i - 2],
+                temporal_expansion=temporal_expansions[-i - 1],
+                spatial_expansion=spatial_expansions[-i - 1],
+                causal=causal,
+                **block_kwargs,
+            )
+            blocks.append(block)
+
+        assert not output_norm
+
+        # Last block. Preserve channel count.
+        last_block = []
+        for _ in range(num_res_blocks[0]):
+            last_block.append(
+                block_fn(
+                    ch[0], has_attention=has_attention[0], causal=causal, **block_kwargs
+                )
+            )
+        blocks.append(nn.Sequential(*last_block))
+
+        self.blocks = nn.ModuleList(blocks)
+        self.output_proj = Conv1x1(ch[0], out_channels)
+
+    def forward(self, x):
+        """Forward pass.
+
+        Args:
+            x: Latent tensor. Shape: [B, input_channels, t, h, w]. Scaled [-1, 1].
+
+        Returns:
+            x: Reconstructed video tensor. Shape: [B, C, T, H, W]. Scaled to [-1, 1].
+               T + 1 = (t - 1) * 4.
+               H = h * 16, W = w * 16.
+        """
+        for block in self.blocks:
+            x = block(x)
+
+        if self.output_nonlinearity == "silu":
+            x = F.silu(x, inplace=not self.training)
+        else:
+            assert (
+                not self.output_nonlinearity
+            )  # StyleGAN3 omits the to-RGB nonlinearity.
+
+        return self.output_proj(x).contiguous()
+
+class LatentDistribution:
+    def __init__(self, mean: torch.Tensor, logvar: torch.Tensor):
+        """Initialize latent distribution.
+
+        Args:
+            mean: Mean of the distribution. Shape: [B, C, T, H, W].
+            logvar: Logarithm of variance of the distribution. Shape: [B, C, T, H, W].
+        """
+        assert mean.shape == logvar.shape
+        self.mean = mean
+        self.logvar = logvar
+
+    def sample(self, temperature=1.0, generator: torch.Generator = None, noise=None):
+        if temperature == 0.0:
+            return self.mean
+
+        if noise is None:
+            noise = torch.randn(self.mean.shape, device=self.mean.device, dtype=self.mean.dtype, generator=generator)
+        else:
+            assert noise.device == self.mean.device
+            noise = noise.to(self.mean.dtype)
+
+        if temperature != 1.0:
+            raise NotImplementedError(f"Temperature {temperature} is not supported.")
+
+        # Just Gaussian sample with no scaling of variance.
+        return noise * torch.exp(self.logvar * 0.5) + self.mean
+
+    def mode(self):
+        return self.mean
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels: int,
+        base_channels: int,
+        channel_multipliers: List[int],
+        num_res_blocks: List[int],
+        latent_dim: int,
+        temporal_reductions: List[int],
+        spatial_reductions: List[int],
+        prune_bottlenecks: List[bool],
+        has_attentions: List[bool],
+        affine: bool = True,
+        bias: bool = True,
+        input_is_conv_1x1: bool = False,
+        padding_mode: str,
+    ):
+        super().__init__()
+        self.temporal_reductions = temporal_reductions
+        self.spatial_reductions = spatial_reductions
+        self.base_channels = base_channels
+        self.channel_multipliers = channel_multipliers
+        self.num_res_blocks = num_res_blocks
+        self.latent_dim = latent_dim
+
+        self.fourier_features = FourierFeatures()
+        ch = [mult * base_channels for mult in channel_multipliers]
+        num_down_blocks = len(ch) - 1
+        assert len(num_res_blocks) == num_down_blocks + 2
+
+        layers = (
+            [ops.Conv3d(in_channels, ch[0], kernel_size=(1, 1, 1), bias=True)]
+            if not input_is_conv_1x1
+            else [Conv1x1(in_channels, ch[0])]
+        )
+
+        assert len(prune_bottlenecks) == num_down_blocks + 2
+        assert len(has_attentions) == num_down_blocks + 2
+        block = partial(block_fn, padding_mode=padding_mode, affine=affine, bias=bias)
+
+        for _ in range(num_res_blocks[0]):
+            layers.append(block(ch[0], has_attention=has_attentions[0], prune_bottleneck=prune_bottlenecks[0]))
+        prune_bottlenecks = prune_bottlenecks[1:]
+        has_attentions = has_attentions[1:]
+
+        assert len(temporal_reductions) == len(spatial_reductions) == len(ch) - 1
+        for i in range(num_down_blocks):
+            layer = DownsampleBlock(
+                ch[i],
+                ch[i + 1],
+                num_res_blocks=num_res_blocks[i + 1],
+                temporal_reduction=temporal_reductions[i],
+                spatial_reduction=spatial_reductions[i],
+                prune_bottleneck=prune_bottlenecks[i],
+                has_attention=has_attentions[i],
+                affine=affine,
+                bias=bias,
+                padding_mode=padding_mode,
+            )
+
+            layers.append(layer)
+
+        # Additional blocks.
+        for _ in range(num_res_blocks[-1]):
+            layers.append(block(ch[-1], has_attention=has_attentions[-1], prune_bottleneck=prune_bottlenecks[-1]))
+
+        self.layers = nn.Sequential(*layers)
+
+        # Output layers.
+        self.output_norm = norm_fn(ch[-1])
+        self.output_proj = Conv1x1(ch[-1], 2 * latent_dim, bias=False)
+
+    @property
+    def temporal_downsample(self):
+        return math.prod(self.temporal_reductions)
+
+    @property
+    def spatial_downsample(self):
+        return math.prod(self.spatial_reductions)
+
+    def forward(self, x) -> LatentDistribution:
+        """Forward pass.
+
+        Args:
+            x: Input video tensor. Shape: [B, C, T, H, W]. Scaled to [-1, 1]
+
+        Returns:
+            means: Latent tensor. Shape: [B, latent_dim, t, h, w]. Scaled [-1, 1].
+                   h = H // 8, w = W // 8, t - 1 = (T - 1) // 6
+            logvar: Shape: [B, latent_dim, t, h, w].
+        """
+        assert x.ndim == 5, f"Expected 5D input, got {x.shape}"
+        x = self.fourier_features(x)
+
+        x = self.layers(x)
+
+        x = self.output_norm(x)
+        x = F.silu(x, inplace=True)
+        x = self.output_proj(x)
+
+        means, logvar = torch.chunk(x, 2, dim=1)
+
+        assert means.ndim == 5
+        assert logvar.shape == means.shape
+        assert means.size(1) == self.latent_dim
+
+        return LatentDistribution(means, logvar)
+
+
+class VideoVAE(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.encoder = Encoder(
+            in_channels=15,
+            base_channels=64,
+            channel_multipliers=[1, 2, 4, 6],
+            num_res_blocks=[3, 3, 4, 6, 3],
+            latent_dim=12,
+            temporal_reductions=[1, 2, 3],
+            spatial_reductions=[2, 2, 2],
+            prune_bottlenecks=[False, False, False, False, False],
+            has_attentions=[False, True, True, True, True],
+            affine=True,
+            bias=True,
+            input_is_conv_1x1=True,
+            padding_mode="replicate"
+        )
+        self.decoder = Decoder(
+            out_channels=3,
+            base_channels=128,
+            channel_multipliers=[1, 2, 4, 6],
+            temporal_expansions=[1, 2, 3],
+            spatial_expansions=[2, 2, 2],
+            num_res_blocks=[3, 3, 4, 6, 3],
+            latent_dim=12,
+            has_attention=[False, False, False, False, False],
+            padding_mode="replicate",
+            output_norm=False,
+            nonlinearity="silu",
+            output_nonlinearity="silu",
+            causal=True,
+        )
+
+    def encode(self, x):
+        return self.encoder(x).mode()
+
+    def decode(self, x):
+        return self.decoder(x)

ComfyUI/comfy/ldm/hidream/model.py

@@ -0,0 +1,829 @@
+from typing import Optional, Tuple, List
+
+import torch
+import torch.nn as nn
+import einops
+from einops import repeat
+
+from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps
+import torch.nn.functional as F
+
+from comfy.ldm.flux.math import apply_rope, rope
+from comfy.ldm.flux.layers import LastLayer
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+import comfy.patcher_extension
+import comfy.ldm.common_dit
+
+
+# Copied from https://github.com/black-forest-labs/flux/blob/main/src/flux/modules/layers.py
+class EmbedND(nn.Module):
+    def __init__(self, theta: int, axes_dim: List[int]):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        n_axes = ids.shape[-1]
+        emb = torch.cat(
+            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
+            dim=-3,
+        )
+        return emb.unsqueeze(2)
+
+
+class PatchEmbed(nn.Module):
+    def __init__(
+        self,
+        patch_size=2,
+        in_channels=4,
+        out_channels=1024,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.out_channels = out_channels
+        self.proj = operations.Linear(in_channels * patch_size * patch_size, out_channels, bias=True, dtype=dtype, device=device)
+
+    def forward(self, latent):
+        latent = self.proj(latent)
+        return latent
+
+
+class PooledEmbed(nn.Module):
+    def __init__(self, text_emb_dim, hidden_size, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.pooled_embedder = TimestepEmbedding(in_channels=text_emb_dim, time_embed_dim=hidden_size, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, pooled_embed):
+        return self.pooled_embedder(pooled_embed)
+
+
+class TimestepEmbed(nn.Module):
+    def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=frequency_embedding_size, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=frequency_embedding_size, time_embed_dim=hidden_size, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, timesteps, wdtype):
+        t_emb = self.time_proj(timesteps).to(dtype=wdtype)
+        t_emb = self.timestep_embedder(t_emb)
+        return t_emb
+
+
+def attention(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, transformer_options={}):
+    return optimized_attention(query.view(query.shape[0], -1, query.shape[-1] * query.shape[-2]), key.view(key.shape[0], -1, key.shape[-1] * key.shape[-2]), value.view(value.shape[0], -1, value.shape[-1] * value.shape[-2]), query.shape[2], transformer_options=transformer_options)
+
+
+class HiDreamAttnProcessor_flashattn:
+    """Attention processor used typically in processing the SD3-like self-attention projections."""
+
+    def __call__(
+        self,
+        attn,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        rope: torch.FloatTensor = None,
+        transformer_options={},
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        dtype = image_tokens.dtype
+        batch_size = image_tokens.shape[0]
+
+        query_i = attn.q_rms_norm(attn.to_q(image_tokens)).to(dtype=dtype)
+        key_i = attn.k_rms_norm(attn.to_k(image_tokens)).to(dtype=dtype)
+        value_i = attn.to_v(image_tokens)
+
+        inner_dim = key_i.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query_i = query_i.view(batch_size, -1, attn.heads, head_dim)
+        key_i = key_i.view(batch_size, -1, attn.heads, head_dim)
+        value_i = value_i.view(batch_size, -1, attn.heads, head_dim)
+        if image_tokens_masks is not None:
+            key_i = key_i * image_tokens_masks.view(batch_size, -1, 1, 1)
+
+        if not attn.single:
+            query_t = attn.q_rms_norm_t(attn.to_q_t(text_tokens)).to(dtype=dtype)
+            key_t = attn.k_rms_norm_t(attn.to_k_t(text_tokens)).to(dtype=dtype)
+            value_t = attn.to_v_t(text_tokens)
+
+            query_t = query_t.view(batch_size, -1, attn.heads, head_dim)
+            key_t = key_t.view(batch_size, -1, attn.heads, head_dim)
+            value_t = value_t.view(batch_size, -1, attn.heads, head_dim)
+
+            num_image_tokens = query_i.shape[1]
+            num_text_tokens = query_t.shape[1]
+            query = torch.cat([query_i, query_t], dim=1)
+            key = torch.cat([key_i, key_t], dim=1)
+            value = torch.cat([value_i, value_t], dim=1)
+        else:
+            query = query_i
+            key = key_i
+            value = value_i
+
+        if query.shape[-1] == rope.shape[-3] * 2:
+            query, key = apply_rope(query, key, rope)
+        else:
+            query_1, query_2 = query.chunk(2, dim=-1)
+            key_1, key_2 = key.chunk(2, dim=-1)
+            query_1, key_1 = apply_rope(query_1, key_1, rope)
+            query = torch.cat([query_1, query_2], dim=-1)
+            key = torch.cat([key_1, key_2], dim=-1)
+
+        hidden_states = attention(query, key, value, transformer_options=transformer_options)
+
+        if not attn.single:
+            hidden_states_i, hidden_states_t = torch.split(hidden_states, [num_image_tokens, num_text_tokens], dim=1)
+            hidden_states_i = attn.to_out(hidden_states_i)
+            hidden_states_t = attn.to_out_t(hidden_states_t)
+            return hidden_states_i, hidden_states_t
+        else:
+            hidden_states = attn.to_out(hidden_states)
+            return hidden_states
+
+class HiDreamAttention(nn.Module):
+    def __init__(
+        self,
+        query_dim: int,
+        heads: int = 8,
+        dim_head: int = 64,
+        upcast_attention: bool = False,
+        upcast_softmax: bool = False,
+        scale_qk: bool = True,
+        eps: float = 1e-5,
+        processor = None,
+        out_dim: int = None,
+        single: bool = False,
+        dtype=None, device=None, operations=None
+    ):
+        # super(Attention, self).__init__()
+        super().__init__()
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.query_dim = query_dim
+        self.upcast_attention = upcast_attention
+        self.upcast_softmax = upcast_softmax
+        self.out_dim = out_dim if out_dim is not None else query_dim
+
+        self.scale_qk = scale_qk
+        self.scale = dim_head**-0.5 if self.scale_qk else 1.0
+
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+        self.sliceable_head_dim = heads
+        self.single = single
+
+        linear_cls = operations.Linear
+        self.linear_cls = linear_cls
+        self.to_q = linear_cls(query_dim, self.inner_dim, dtype=dtype, device=device)
+        self.to_k = linear_cls(self.inner_dim, self.inner_dim, dtype=dtype, device=device)
+        self.to_v = linear_cls(self.inner_dim, self.inner_dim, dtype=dtype, device=device)
+        self.to_out = linear_cls(self.inner_dim, self.out_dim, dtype=dtype, device=device)
+        self.q_rms_norm = operations.RMSNorm(self.inner_dim, eps, dtype=dtype, device=device)
+        self.k_rms_norm = operations.RMSNorm(self.inner_dim, eps, dtype=dtype, device=device)
+
+        if not single:
+            self.to_q_t = linear_cls(query_dim, self.inner_dim, dtype=dtype, device=device)
+            self.to_k_t = linear_cls(self.inner_dim, self.inner_dim, dtype=dtype, device=device)
+            self.to_v_t = linear_cls(self.inner_dim, self.inner_dim, dtype=dtype, device=device)
+            self.to_out_t = linear_cls(self.inner_dim, self.out_dim, dtype=dtype, device=device)
+            self.q_rms_norm_t = operations.RMSNorm(self.inner_dim, eps, dtype=dtype, device=device)
+            self.k_rms_norm_t = operations.RMSNorm(self.inner_dim, eps, dtype=dtype, device=device)
+
+        self.processor = processor
+
+    def forward(
+        self,
+        norm_image_tokens: torch.FloatTensor,
+        image_tokens_masks: torch.FloatTensor = None,
+        norm_text_tokens: torch.FloatTensor = None,
+        rope: torch.FloatTensor = None,
+        transformer_options={},
+    ) -> torch.Tensor:
+        return self.processor(
+            self,
+            image_tokens = norm_image_tokens,
+            image_tokens_masks = image_tokens_masks,
+            text_tokens = norm_text_tokens,
+            rope = rope,
+            transformer_options=transformer_options,
+        )
+
+
+class FeedForwardSwiGLU(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        hidden_dim = int(2 * hidden_dim / 3)
+        # custom dim factor multiplier
+        if ffn_dim_multiplier is not None:
+            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
+        hidden_dim = multiple_of * (
+            (hidden_dim + multiple_of - 1) // multiple_of
+        )
+
+        self.w1 = operations.Linear(dim, hidden_dim, bias=False, dtype=dtype, device=device)
+        self.w2 = operations.Linear(hidden_dim, dim, bias=False, dtype=dtype, device=device)
+        self.w3 = operations.Linear(dim, hidden_dim, bias=False, dtype=dtype, device=device)
+
+    def forward(self, x):
+        return self.w2(torch.nn.functional.silu(self.w1(x)) * self.w3(x))
+
+
+# Modified from https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/model.py
+class MoEGate(nn.Module):
+    def __init__(self, embed_dim, num_routed_experts=4, num_activated_experts=2, aux_loss_alpha=0.01, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.top_k = num_activated_experts
+        self.n_routed_experts = num_routed_experts
+
+        self.scoring_func = 'softmax'
+        self.alpha = aux_loss_alpha
+        self.seq_aux = False
+
+        # topk selection algorithm
+        self.norm_topk_prob = False
+        self.gating_dim = embed_dim
+        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim), dtype=dtype, device=device))
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        pass
+        # import torch.nn.init  as init
+        # init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+
+    def forward(self, hidden_states):
+        bsz, seq_len, h = hidden_states.shape
+
+        ### compute gating score
+        hidden_states = hidden_states.view(-1, h)
+        logits = F.linear(hidden_states, comfy.model_management.cast_to(self.weight, dtype=hidden_states.dtype, device=hidden_states.device), None)
+        if self.scoring_func == 'softmax':
+            scores = logits.softmax(dim=-1)
+        else:
+            raise NotImplementedError(f'insupportable scoring function for MoE gating: {self.scoring_func}')
+
+        ### select top-k experts
+        topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False)
+
+        ### norm gate to sum 1
+        if self.top_k > 1 and self.norm_topk_prob:
+            denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weight = topk_weight / denominator
+
+        aux_loss = None
+        return topk_idx, topk_weight, aux_loss
+
+
+# Modified from https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/model.py
+class MOEFeedForwardSwiGLU(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        num_routed_experts: int,
+        num_activated_experts: int,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.shared_experts = FeedForwardSwiGLU(dim, hidden_dim // 2, dtype=dtype, device=device, operations=operations)
+        self.experts = nn.ModuleList([FeedForwardSwiGLU(dim, hidden_dim, dtype=dtype, device=device, operations=operations) for i in range(num_routed_experts)])
+        self.gate = MoEGate(
+            embed_dim = dim,
+            num_routed_experts = num_routed_experts,
+            num_activated_experts = num_activated_experts,
+            dtype=dtype, device=device, operations=operations
+        )
+        self.num_activated_experts = num_activated_experts
+
+    def forward(self, x):
+        wtype = x.dtype
+        identity = x
+        orig_shape = x.shape
+        topk_idx, topk_weight, aux_loss = self.gate(x)
+        x = x.view(-1, x.shape[-1])
+        flat_topk_idx = topk_idx.view(-1)
+        if True:  # self.training: # TODO: check which branch performs faster
+            x = x.repeat_interleave(self.num_activated_experts, dim=0)
+            y = torch.empty_like(x, dtype=wtype)
+            for i, expert in enumerate(self.experts):
+                y[flat_topk_idx == i] = expert(x[flat_topk_idx == i]).to(dtype=wtype)
+            y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
+            y =  y.view(*orig_shape).to(dtype=wtype)
+            #y = AddAuxiliaryLoss.apply(y, aux_loss)
+        else:
+            y = self.moe_infer(x, flat_topk_idx, topk_weight.view(-1, 1)).view(*orig_shape)
+        y = y + self.shared_experts(identity)
+        return y
+
+    @torch.no_grad()
+    def moe_infer(self, x, flat_expert_indices, flat_expert_weights):
+        expert_cache = torch.zeros_like(x)
+        idxs = flat_expert_indices.argsort()
+        tokens_per_expert = flat_expert_indices.bincount().cpu().numpy().cumsum(0)
+        token_idxs = idxs // self.num_activated_experts
+        for i, end_idx in enumerate(tokens_per_expert):
+            start_idx = 0 if i == 0 else tokens_per_expert[i-1]
+            if start_idx == end_idx:
+                continue
+            expert = self.experts[i]
+            exp_token_idx = token_idxs[start_idx:end_idx]
+            expert_tokens = x[exp_token_idx]
+            expert_out = expert(expert_tokens)
+            expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]])
+
+            # for fp16 and other dtype
+            expert_cache = expert_cache.to(expert_out.dtype)
+            expert_cache.scatter_reduce_(0, exp_token_idx.view(-1, 1).repeat(1, x.shape[-1]), expert_out, reduce='sum')
+        return expert_cache
+
+
+class TextProjection(nn.Module):
+    def __init__(self, in_features, hidden_size, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.linear = operations.Linear(in_features=in_features, out_features=hidden_size, bias=False, dtype=dtype, device=device)
+
+    def forward(self, caption):
+        hidden_states = self.linear(caption)
+        return hidden_states
+
+
+class BlockType:
+    TransformerBlock = 1
+    SingleTransformerBlock = 2
+
+
+class HiDreamImageSingleTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(dim, 6 * dim, bias=True, dtype=dtype, device=device)
+        )
+
+        # 1. Attention
+        self.norm1_i = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        self.attn1 = HiDreamAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            processor = HiDreamAttnProcessor_flashattn(),
+            single = True,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        # 3. Feed-forward
+        self.norm3_i = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        if num_routed_experts > 0:
+            self.ff_i = MOEFeedForwardSwiGLU(
+                dim = dim,
+                hidden_dim = 4 * dim,
+                num_routed_experts = num_routed_experts,
+                num_activated_experts = num_activated_experts,
+                dtype=dtype, device=device, operations=operations
+            )
+        else:
+            self.ff_i = FeedForwardSwiGLU(dim = dim, hidden_dim = 4 * dim, dtype=dtype, device=device, operations=operations)
+
+    def forward(
+        self,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        adaln_input: Optional[torch.FloatTensor] = None,
+        rope: torch.FloatTensor = None,
+        transformer_options={},
+    ) -> torch.FloatTensor:
+        wtype = image_tokens.dtype
+        shift_msa_i, scale_msa_i, gate_msa_i, shift_mlp_i, scale_mlp_i, gate_mlp_i = \
+            self.adaLN_modulation(adaln_input)[:,None].chunk(6, dim=-1)
+
+        # 1. MM-Attention
+        norm_image_tokens = self.norm1_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_msa_i) + shift_msa_i
+        attn_output_i = self.attn1(
+            norm_image_tokens,
+            image_tokens_masks,
+            rope = rope,
+            transformer_options=transformer_options,
+        )
+        image_tokens = gate_msa_i * attn_output_i + image_tokens
+
+        # 2. Feed-forward
+        norm_image_tokens = self.norm3_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_mlp_i) + shift_mlp_i
+        ff_output_i = gate_mlp_i * self.ff_i(norm_image_tokens.to(dtype=wtype))
+        image_tokens = ff_output_i + image_tokens
+        return image_tokens
+
+
+class HiDreamImageTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(dim, 12 * dim, bias=True, dtype=dtype, device=device)
+        )
+        # nn.init.zeros_(self.adaLN_modulation[1].weight)
+        # nn.init.zeros_(self.adaLN_modulation[1].bias)
+
+        # 1. Attention
+        self.norm1_i = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        self.norm1_t = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        self.attn1 = HiDreamAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            processor = HiDreamAttnProcessor_flashattn(),
+            single = False,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        # 3. Feed-forward
+        self.norm3_i = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        if num_routed_experts > 0:
+            self.ff_i = MOEFeedForwardSwiGLU(
+                dim = dim,
+                hidden_dim = 4 * dim,
+                num_routed_experts = num_routed_experts,
+                num_activated_experts = num_activated_experts,
+                dtype=dtype, device=device, operations=operations
+            )
+        else:
+            self.ff_i = FeedForwardSwiGLU(dim = dim, hidden_dim = 4 * dim, dtype=dtype, device=device, operations=operations)
+        self.norm3_t = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False)
+        self.ff_t = FeedForwardSwiGLU(dim = dim, hidden_dim = 4 * dim, dtype=dtype, device=device, operations=operations)
+
+    def forward(
+        self,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        adaln_input: Optional[torch.FloatTensor] = None,
+        rope: torch.FloatTensor = None,
+        transformer_options={},
+    ) -> torch.FloatTensor:
+        wtype = image_tokens.dtype
+        shift_msa_i, scale_msa_i, gate_msa_i, shift_mlp_i, scale_mlp_i, gate_mlp_i, \
+        shift_msa_t, scale_msa_t, gate_msa_t, shift_mlp_t, scale_mlp_t, gate_mlp_t = \
+            self.adaLN_modulation(adaln_input)[:,None].chunk(12, dim=-1)
+
+        # 1. MM-Attention
+        norm_image_tokens = self.norm1_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_msa_i) + shift_msa_i
+        norm_text_tokens = self.norm1_t(text_tokens).to(dtype=wtype)
+        norm_text_tokens = norm_text_tokens * (1 + scale_msa_t) + shift_msa_t
+
+        attn_output_i, attn_output_t = self.attn1(
+            norm_image_tokens,
+            image_tokens_masks,
+            norm_text_tokens,
+            rope = rope,
+            transformer_options=transformer_options,
+        )
+
+        image_tokens = gate_msa_i * attn_output_i + image_tokens
+        text_tokens = gate_msa_t * attn_output_t + text_tokens
+
+        # 2. Feed-forward
+        norm_image_tokens = self.norm3_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_mlp_i) + shift_mlp_i
+        norm_text_tokens = self.norm3_t(text_tokens).to(dtype=wtype)
+        norm_text_tokens = norm_text_tokens * (1 + scale_mlp_t) + shift_mlp_t
+
+        ff_output_i = gate_mlp_i * self.ff_i(norm_image_tokens)
+        ff_output_t = gate_mlp_t * self.ff_t(norm_text_tokens)
+        image_tokens = ff_output_i + image_tokens
+        text_tokens = ff_output_t + text_tokens
+        return image_tokens, text_tokens
+
+
+class HiDreamImageBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        block_type: BlockType = BlockType.TransformerBlock,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        block_classes = {
+            BlockType.TransformerBlock: HiDreamImageTransformerBlock,
+            BlockType.SingleTransformerBlock: HiDreamImageSingleTransformerBlock,
+        }
+        self.block = block_classes[block_type](
+            dim,
+            num_attention_heads,
+            attention_head_dim,
+            num_routed_experts,
+            num_activated_experts,
+            dtype=dtype, device=device, operations=operations
+        )
+
+    def forward(
+        self,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        adaln_input: torch.FloatTensor = None,
+        rope: torch.FloatTensor = None,
+        transformer_options={},
+    ) -> torch.FloatTensor:
+        return self.block(
+            image_tokens,
+            image_tokens_masks,
+            text_tokens,
+            adaln_input,
+            rope,
+            transformer_options=transformer_options,
+        )
+
+
+class HiDreamImageTransformer2DModel(nn.Module):
+    def __init__(
+        self,
+        patch_size: Optional[int] = None,
+        in_channels: int = 64,
+        out_channels: Optional[int] = None,
+        num_layers: int = 16,
+        num_single_layers: int = 32,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 20,
+        caption_channels: List[int] = None,
+        text_emb_dim: int = 2048,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        axes_dims_rope: Tuple[int, int] = (32, 32),
+        max_resolution: Tuple[int, int] = (128, 128),
+        llama_layers: List[int] = None,
+        image_model=None,
+        dtype=None, device=None, operations=None
+    ):
+        self.patch_size = patch_size
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.num_layers = num_layers
+        self.num_single_layers = num_single_layers
+
+        self.gradient_checkpointing = False
+
+        super().__init__()
+        self.dtype = dtype
+        self.out_channels = out_channels or in_channels
+        self.inner_dim = self.num_attention_heads * self.attention_head_dim
+        self.llama_layers = llama_layers
+
+        self.t_embedder = TimestepEmbed(self.inner_dim, dtype=dtype, device=device, operations=operations)
+        self.p_embedder = PooledEmbed(text_emb_dim, self.inner_dim, dtype=dtype, device=device, operations=operations)
+        self.x_embedder = PatchEmbed(
+            patch_size = patch_size,
+            in_channels = in_channels,
+            out_channels = self.inner_dim,
+            dtype=dtype, device=device, operations=operations
+        )
+        self.pe_embedder = EmbedND(theta=10000, axes_dim=axes_dims_rope)
+
+        self.double_stream_blocks = nn.ModuleList(
+            [
+                HiDreamImageBlock(
+                    dim = self.inner_dim,
+                    num_attention_heads = self.num_attention_heads,
+                    attention_head_dim = self.attention_head_dim,
+                    num_routed_experts = num_routed_experts,
+                    num_activated_experts = num_activated_experts,
+                    block_type = BlockType.TransformerBlock,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for i in range(self.num_layers)
+            ]
+        )
+
+        self.single_stream_blocks = nn.ModuleList(
+            [
+                HiDreamImageBlock(
+                    dim = self.inner_dim,
+                    num_attention_heads = self.num_attention_heads,
+                    attention_head_dim = self.attention_head_dim,
+                    num_routed_experts = num_routed_experts,
+                    num_activated_experts = num_activated_experts,
+                    block_type = BlockType.SingleTransformerBlock,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for i in range(self.num_single_layers)
+            ]
+        )
+
+        self.final_layer = LastLayer(self.inner_dim, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations)
+
+        caption_channels = [caption_channels[1], ] * (num_layers + num_single_layers) + [caption_channels[0], ]
+        caption_projection = []
+        for caption_channel in caption_channels:
+            caption_projection.append(TextProjection(in_features=caption_channel, hidden_size=self.inner_dim, dtype=dtype, device=device, operations=operations))
+        self.caption_projection = nn.ModuleList(caption_projection)
+        self.max_seq = max_resolution[0] * max_resolution[1] // (patch_size * patch_size)
+
+    def expand_timesteps(self, timesteps, batch_size, device):
+        if not torch.is_tensor(timesteps):
+            is_mps = device.type == "mps"
+            if isinstance(timesteps, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(device)
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(batch_size)
+        return timesteps
+
+    def unpatchify(self, x: torch.Tensor, img_sizes: List[Tuple[int, int]]) -> List[torch.Tensor]:
+        x_arr = []
+        for i, img_size in enumerate(img_sizes):
+            pH, pW = img_size
+            x_arr.append(
+                einops.rearrange(x[i, :pH*pW].reshape(1, pH, pW, -1), 'B H W (p1 p2 C) -> B C (H p1) (W p2)',
+                    p1=self.patch_size, p2=self.patch_size)
+            )
+        x = torch.cat(x_arr, dim=0)
+        return x
+
+    def patchify(self, x, max_seq, img_sizes=None):
+        pz2 = self.patch_size * self.patch_size
+        if isinstance(x, torch.Tensor):
+            B = x.shape[0]
+            device = x.device
+            dtype = x.dtype
+        else:
+            B = len(x)
+            device = x[0].device
+            dtype = x[0].dtype
+        x_masks = torch.zeros((B, max_seq), dtype=dtype, device=device)
+
+        if img_sizes is not None:
+            for i, img_size in enumerate(img_sizes):
+                x_masks[i, 0:img_size[0] * img_size[1]] = 1
+            x = einops.rearrange(x, 'B C S p -> B S (p C)', p=pz2)
+        elif isinstance(x, torch.Tensor):
+            pH, pW = x.shape[-2] // self.patch_size, x.shape[-1] // self.patch_size
+            x = einops.rearrange(x, 'B C (H p1) (W p2) -> B (H W) (p1 p2 C)', p1=self.patch_size, p2=self.patch_size)
+            img_sizes = [[pH, pW]] * B
+            x_masks = None
+        else:
+            raise NotImplementedError
+        return x, x_masks, img_sizes
+
+    def forward(self,
+        x: torch.Tensor,
+        t: torch.Tensor,
+        y: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        encoder_hidden_states_llama3=None,
+        image_cond=None,
+        control = None,
+        transformer_options = {},
+    ):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, t, y, context, encoder_hidden_states_llama3, image_cond, control, transformer_options)
+
+    def _forward(
+        self,
+        x: torch.Tensor,
+        t: torch.Tensor,
+        y: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        encoder_hidden_states_llama3=None,
+        image_cond=None,
+        control = None,
+        transformer_options = {},
+    ) -> torch.Tensor:
+        bs, c, h, w = x.shape
+        if image_cond is not None:
+            x = torch.cat([x, image_cond], dim=-1)
+        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+        timesteps = t
+        pooled_embeds = y
+        T5_encoder_hidden_states = context
+
+        img_sizes = None
+
+        # spatial forward
+        batch_size = hidden_states.shape[0]
+        hidden_states_type = hidden_states.dtype
+
+        # 0. time
+        timesteps = self.expand_timesteps(timesteps, batch_size, hidden_states.device)
+        timesteps = self.t_embedder(timesteps, hidden_states_type)
+        p_embedder = self.p_embedder(pooled_embeds)
+        adaln_input = timesteps + p_embedder
+
+        hidden_states, image_tokens_masks, img_sizes = self.patchify(hidden_states, self.max_seq, img_sizes)
+        if image_tokens_masks is None:
+            pH, pW = img_sizes[0]
+            img_ids = torch.zeros(pH, pW, 3, device=hidden_states.device)
+            img_ids[..., 1] = img_ids[..., 1] + torch.arange(pH, device=hidden_states.device)[:, None]
+            img_ids[..., 2] = img_ids[..., 2] + torch.arange(pW, device=hidden_states.device)[None, :]
+            img_ids = repeat(img_ids, "h w c -> b (h w) c", b=batch_size)
+        hidden_states = self.x_embedder(hidden_states)
+
+        # T5_encoder_hidden_states = encoder_hidden_states[0]
+        encoder_hidden_states = encoder_hidden_states_llama3.movedim(1, 0)
+        encoder_hidden_states = [encoder_hidden_states[k] for k in self.llama_layers]
+
+        if self.caption_projection is not None:
+            new_encoder_hidden_states = []
+            for i, enc_hidden_state in enumerate(encoder_hidden_states):
+                enc_hidden_state = self.caption_projection[i](enc_hidden_state)
+                enc_hidden_state = enc_hidden_state.view(batch_size, -1, hidden_states.shape[-1])
+                new_encoder_hidden_states.append(enc_hidden_state)
+            encoder_hidden_states = new_encoder_hidden_states
+            T5_encoder_hidden_states = self.caption_projection[-1](T5_encoder_hidden_states)
+            T5_encoder_hidden_states = T5_encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+            encoder_hidden_states.append(T5_encoder_hidden_states)
+
+        txt_ids = torch.zeros(
+            batch_size,
+            encoder_hidden_states[-1].shape[1] + encoder_hidden_states[-2].shape[1] + encoder_hidden_states[0].shape[1],
+            3,
+            device=img_ids.device, dtype=img_ids.dtype
+        )
+        ids = torch.cat((img_ids, txt_ids), dim=1)
+        rope = self.pe_embedder(ids)
+
+        # 2. Blocks
+        block_id = 0
+        initial_encoder_hidden_states = torch.cat([encoder_hidden_states[-1], encoder_hidden_states[-2]], dim=1)
+        initial_encoder_hidden_states_seq_len = initial_encoder_hidden_states.shape[1]
+        for bid, block in enumerate(self.double_stream_blocks):
+            cur_llama31_encoder_hidden_states = encoder_hidden_states[block_id]
+            cur_encoder_hidden_states = torch.cat([initial_encoder_hidden_states, cur_llama31_encoder_hidden_states], dim=1)
+            hidden_states, initial_encoder_hidden_states = block(
+                image_tokens = hidden_states,
+                image_tokens_masks = image_tokens_masks,
+                text_tokens = cur_encoder_hidden_states,
+                adaln_input = adaln_input,
+                rope = rope,
+                transformer_options=transformer_options,
+            )
+            initial_encoder_hidden_states = initial_encoder_hidden_states[:, :initial_encoder_hidden_states_seq_len]
+            block_id += 1
+
+        image_tokens_seq_len = hidden_states.shape[1]
+        hidden_states = torch.cat([hidden_states, initial_encoder_hidden_states], dim=1)
+        hidden_states_seq_len = hidden_states.shape[1]
+        if image_tokens_masks is not None:
+            encoder_attention_mask_ones = torch.ones(
+                (batch_size, initial_encoder_hidden_states.shape[1] + cur_llama31_encoder_hidden_states.shape[1]),
+                device=image_tokens_masks.device, dtype=image_tokens_masks.dtype
+            )
+            image_tokens_masks = torch.cat([image_tokens_masks, encoder_attention_mask_ones], dim=1)
+
+        for bid, block in enumerate(self.single_stream_blocks):
+            cur_llama31_encoder_hidden_states = encoder_hidden_states[block_id]
+            hidden_states = torch.cat([hidden_states, cur_llama31_encoder_hidden_states], dim=1)
+            hidden_states = block(
+                image_tokens=hidden_states,
+                image_tokens_masks=image_tokens_masks,
+                text_tokens=None,
+                adaln_input=adaln_input,
+                rope=rope,
+                transformer_options=transformer_options,
+            )
+            hidden_states = hidden_states[:, :hidden_states_seq_len]
+            block_id += 1
+
+        hidden_states = hidden_states[:, :image_tokens_seq_len, ...]
+        output = self.final_layer(hidden_states, adaln_input)
+        output = self.unpatchify(output, img_sizes)
+        return -output[:, :, :h, :w]

ComfyUI/comfy/ldm/hidream_o1/attention.py

@@ -0,0 +1,41 @@
+"""HiDream-O1 two-pass attention: tokens [0, ar_len) are causal, [ar_len, T)
+attend full K/V. Splitting Q at the boundary avoids the (B, 1, T, T) additive
+mask the general-purpose path would build (~500 MB at T~16K) and lets the
+gen half hit the user's preferred backend via optimized_attention.
+"""
+
+import torch
+
+import comfy.ops
+from comfy.ldm.modules.attention import optimized_attention
+
+
+def make_two_pass_attention(ar_len: int, transformer_options=None):
+    """Build a two-pass attention callable. AR pass uses SDPA-causal directly, gen pass routes through optimized_attention.
+    The AR pass goes through SDPA directand bypasses wrappers, it is only ~1% of T at typical edit sizes.
+    """
+
+    def two_pass_attention(q, k, v, heads, enable_gqa=False, **kwargs):
+        B, H, T, D = q.shape
+
+        if T < k.shape[2]: # KV-cache hot path: Q is shorter than K/V (cached AR prefix is in K/V only), all fresh Q positions are in the gen region, single full-attention call
+            out = optimized_attention(q, k, v, heads, mask=None, skip_reshape=True, skip_output_reshape=True, transformer_options=transformer_options, enable_gqa=enable_gqa)
+        elif ar_len >= T:
+            out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=True, enable_gqa=enable_gqa)
+        elif ar_len <= 0:
+            out = optimized_attention(q, k, v, heads, mask=None, skip_reshape=True, skip_output_reshape=True, transformer_options=transformer_options, enable_gqa=enable_gqa)
+        else:
+            out_ar = comfy.ops.scaled_dot_product_attention(
+                q[:, :, :ar_len], k[:, :, :ar_len], v[:, :, :ar_len],
+                attn_mask=None, dropout_p=0.0, is_causal=True, enable_gqa=enable_gqa,
+            )
+            out_gen = optimized_attention(
+                q[:, :, ar_len:], k, v, heads,
+                mask=None, skip_reshape=True, skip_output_reshape=True,
+                transformer_options=transformer_options, enable_gqa=enable_gqa,
+            )
+            out = torch.cat([out_ar, out_gen], dim=2)
+
+        return out.transpose(1, 2).reshape(B, T, H * D)
+
+    return two_pass_attention

ComfyUI/comfy/ldm/hidream_o1/conditioning.py

@@ -0,0 +1,230 @@
+"""HiDream-O1 conditioning prep — ref-image dual path + extra_conds assembly.
+
+Each ref image goes through two paths: a 32x32 patchified stream concatenated
+to the noised target, and a Qwen3-VL ViT path producing tokens that scatter
+into input_ids at <|image_pad|> positions.
+"""
+
+from typing import List
+
+import torch
+
+import comfy.utils
+from comfy.text_encoders.qwen_vl import process_qwen2vl_images
+
+from .utils import (PATCH_SIZE, calculate_dimensions, cond_image_size, ref_max_size, resize_tensor)
+
+# Qwen3-VL ViT preprocessing constants (preprocessor_config.json).
+VIT_PATCH = 16
+VIT_MERGE = 2
+VIT_IMAGE_MEAN = [0.5, 0.5, 0.5]
+VIT_IMAGE_STD = [0.5, 0.5, 0.5]
+
+
+def prepare_ref_images(
+    ref_images: List[torch.Tensor],
+    target_h: int,
+    target_w: int,
+    device: torch.device,
+    dtype: torch.dtype,
+):
+    """Build the dual-path tensors for K reference images at (target_h, target_w).
+
+    Returns None for K=0, else a dict with ref_patches, ref_pixel_values,
+    ref_image_grid_thw, per_ref_vit_tokens, per_ref_patch_grids.
+    """
+    K = len(ref_images)
+    if K == 0:
+        return None
+    max_size = ref_max_size(max(target_h, target_w), K)
+    cis = cond_image_size(K)
+
+    refs_t = [img[0].clamp(0, 1).permute(2, 0, 1).unsqueeze(0).contiguous().float() for img in ref_images]
+    refs_t = [resize_tensor(t, max_size, PATCH_SIZE) for t in refs_t]
+
+    # 32-patch path.
+    ref_patches_per = []
+    per_ref_patch_grids = []
+    for t in refs_t:
+        t_norm = (t.squeeze(0) - 0.5) / 0.5  # (3, H, W) in [-1, 1]
+        h_p, w_p = t_norm.shape[-2] // PATCH_SIZE, t_norm.shape[-1] // PATCH_SIZE
+        per_ref_patch_grids.append((h_p, w_p))
+        patches = (
+            t_norm.reshape(3, h_p, PATCH_SIZE, w_p, PATCH_SIZE)
+            .permute(1, 3, 0, 2, 4)
+            .reshape(h_p * w_p, 3 * PATCH_SIZE * PATCH_SIZE)
+        )
+        ref_patches_per.append(patches)
+    ref_patches = torch.cat(ref_patches_per, dim=0).unsqueeze(0).to(device=device, dtype=dtype)
+
+    # ViT path.
+    refs_vlm_t = []
+    for t in refs_t:
+        _, _, h, w = t.shape
+        cond_w, cond_h = calculate_dimensions(cis, w / h)
+        cond_w = max(cond_w, VIT_PATCH * VIT_MERGE)
+        cond_h = max(cond_h, VIT_PATCH * VIT_MERGE)
+        refs_vlm_t.append(comfy.utils.common_upscale(t, cond_w, cond_h, "lanczos", "disabled"))
+
+    pv_list, grid_list, per_ref_vit_tokens = [], [], []
+    for t_v in refs_vlm_t:
+        pv, grid_thw = process_qwen2vl_images(
+            t_v.permute(0, 2, 3, 1),
+            min_pixels=0, max_pixels=10**12,
+            patch_size=VIT_PATCH, merge_size=VIT_MERGE,
+            image_mean=VIT_IMAGE_MEAN, image_std=VIT_IMAGE_STD,
+        )
+        grid_thw = grid_thw[0]
+        pv_list.append(pv.to(device=device, dtype=dtype))
+        grid_list.append(grid_thw.to(device=device))
+        # Post-merge token count = number of <|image_pad|> tokens this image expands to in input_ids.
+        gh, gw = int(grid_thw[1].item()), int(grid_thw[2].item())
+        per_ref_vit_tokens.append((gh // VIT_MERGE) * (gw // VIT_MERGE))
+
+    return {
+        "ref_patches": ref_patches,
+        "ref_pixel_values": torch.cat(pv_list, dim=0),
+        "ref_image_grid_thw": torch.stack(grid_list, dim=0),
+        "per_ref_vit_tokens": per_ref_vit_tokens,
+        "per_ref_patch_grids": per_ref_patch_grids,
+    }
+
+
+def build_ref_input_ids(
+    text_input_ids: torch.Tensor,
+    per_ref_vit_tokens: List[int],
+    image_token_id: int,
+    vision_start_id: int,
+    vision_end_id: int,
+):
+    """Splice [vision_start, image_pad*N, vision_end] blocks into input_ids
+    after the [im_start, user, \\n] prefix (matches original chat template).
+    """
+    ids = text_input_ids[0].tolist()
+    inserted = []
+    for n_pad in per_ref_vit_tokens:
+        inserted.extend([vision_start_id] + [image_token_id] * n_pad + [vision_end_id])
+    new_ids = ids[:3] + inserted + ids[3:]  # 3 = len([im_start, user, \n])
+    return torch.tensor([new_ids], dtype=text_input_ids.dtype, device=text_input_ids.device)
+
+
+def build_extra_conds(
+    text_input_ids: torch.Tensor,
+    noise: torch.Tensor,
+    ref_images: List[torch.Tensor] = None,
+    target_patch_size: int = 32,
+):
+    """Assemble all conditioning tensors for HiDreamO1Transformer.forward:
+    input_ids (with ref-vision tokens spliced in for the edit/IP path),
+    position_ids (MRoPE), token_types, vinput_mask, plus the ref
+    dual-path tensors when refs are provided.
+    """
+    from .utils import get_rope_index_fix_point
+    from comfy.text_encoders.hidream_o1 import (
+        IMAGE_TOKEN_ID, VISION_START_ID, VISION_END_ID,
+    )
+
+    if text_input_ids.dim() == 1:
+        text_input_ids = text_input_ids.unsqueeze(0)
+    text_input_ids = text_input_ids.long().to(noise.device)
+    B = noise.shape[0]
+    if text_input_ids.shape[0] == 1 and B > 1:
+        text_input_ids = text_input_ids.expand(B, -1)
+
+    H, W = noise.shape[-2], noise.shape[-1]
+    h_p, w_p = H // target_patch_size, W // target_patch_size
+    image_len = h_p * w_p
+    image_grid_thw_tgt = torch.tensor(
+        [[1, h_p, w_p]], dtype=torch.long, device=text_input_ids.device,
+    )
+
+    out = {}
+    if ref_images:
+        ref = prepare_ref_images(ref_images, H, W, device=noise.device, dtype=noise.dtype)
+        text_input_ids = build_ref_input_ids(
+            text_input_ids, ref["per_ref_vit_tokens"],
+            IMAGE_TOKEN_ID, VISION_START_ID, VISION_END_ID,
+        )
+        new_txt_len = text_input_ids.shape[1]
+
+        # Each ref's patchified stream gets a [vision_start, image_pad*N-1]
+        # block in the position-id stream after the noised target.
+        ref_grid_lengths = [hp * wp for (hp, wp) in ref["per_ref_patch_grids"]]
+        tgt_vision = torch.full((1, image_len), IMAGE_TOKEN_ID,
+                                dtype=text_input_ids.dtype, device=text_input_ids.device)
+        tgt_vision[:, 0] = VISION_START_ID
+        ref_vision_blocks = []
+        for rl in ref_grid_lengths:
+            blk = torch.full((1, rl), IMAGE_TOKEN_ID,
+                             dtype=text_input_ids.dtype, device=text_input_ids.device)
+            blk[:, 0] = VISION_START_ID
+            ref_vision_blocks.append(blk)
+        ref_vision_cat = torch.cat([tgt_vision] + ref_vision_blocks, dim=1)
+        input_ids_pad = torch.cat([text_input_ids, ref_vision_cat], dim=-1)
+        total_ref_patches_len = sum(ref_grid_lengths)
+        total_len = new_txt_len + image_len + total_ref_patches_len
+
+        # K (ViT, post-merge) + 1 (target) + K (ref-patches) image grids.
+        K = len(ref_images)
+        igthw_cond = ref["ref_image_grid_thw"].clone()
+        igthw_cond[:, 1] //= 2
+        igthw_cond[:, 2] //= 2
+        image_grid_thw_ref = torch.tensor(
+            [[1, hp, wp] for (hp, wp) in ref["per_ref_patch_grids"]],
+            dtype=torch.long, device=text_input_ids.device,
+        )
+        igthw_all = torch.cat([
+            igthw_cond.to(text_input_ids.device),
+            image_grid_thw_tgt,
+            image_grid_thw_ref,
+        ], dim=0)
+        position_ids, _ = get_rope_index_fix_point(
+            spatial_merge_size=1,
+            image_token_id=IMAGE_TOKEN_ID,
+            vision_start_token_id=VISION_START_ID,
+            input_ids=input_ids_pad, image_grid_thw=igthw_all,
+            attention_mask=None,
+            skip_vision_start_token=[0] * K + [1] + [1] * K,
+            fix_point=4096,
+        )
+
+        # tms + target_image + ref_patches are all gen.
+        tms_pos = new_txt_len - 1
+        ar_len = tms_pos
+        token_types = torch.zeros(B, total_len, dtype=torch.long, device=noise.device)
+        token_types[:, tms_pos:] = 1
+        vinput_mask = torch.zeros(B, total_len, dtype=torch.bool, device=noise.device)
+        vinput_mask[:, new_txt_len:] = True
+
+        # Leading batch dim sidesteps CONDRegular.process_cond's repeat_to_batch_size truncation
+        out["ref_pixel_values"] = ref["ref_pixel_values"].unsqueeze(0)
+        out["ref_image_grid_thw"] = ref["ref_image_grid_thw"].unsqueeze(0)
+        out["ref_patches"] = ref["ref_patches"]
+    else:
+        # T2I: text + noised target only, vision_start replaces the first image token
+        txt_len = text_input_ids.shape[1]
+        total_len = txt_len + image_len
+        vision_tokens = torch.full((B, image_len), IMAGE_TOKEN_ID,
+                                   dtype=text_input_ids.dtype, device=text_input_ids.device)
+        vision_tokens[:, 0] = VISION_START_ID
+        input_ids_pad = torch.cat([text_input_ids, vision_tokens], dim=-1)
+        position_ids, _ = get_rope_index_fix_point(
+            spatial_merge_size=1,
+            image_token_id=IMAGE_TOKEN_ID,
+            vision_start_token_id=VISION_START_ID,
+            input_ids=input_ids_pad, image_grid_thw=image_grid_thw_tgt,
+            attention_mask=None,
+            skip_vision_start_token=[1],
+        )
+        ar_len = txt_len - 1
+        token_types = torch.zeros(B, total_len, dtype=torch.long, device=noise.device)
+        token_types[:, ar_len:] = 1
+        vinput_mask = torch.zeros(B, total_len, dtype=torch.bool, device=noise.device)
+        vinput_mask[:, txt_len:] = True
+
+    out["input_ids"] = text_input_ids
+    out["position_ids"] = position_ids[:, 0].unsqueeze(0) # Collapse position_ids batch and add a leading dim so CONDRegular's batch-resize doesn't truncate the 3-axis MRoPE dim
+    out["token_types"] = token_types
+    out["vinput_mask"] = vinput_mask
+    out["ar_len"] = ar_len
+    return out

ComfyUI/comfy/ldm/hidream_o1/model.py

@@ -0,0 +1,306 @@
+"""HiDream-O1-Image transformer.
+
+Pixel-space DiT built on Qwen3-VL: the vision tower (Qwen35VisionModel)
+encodes ref images, the Qwen3-VL-8B decoder (Llama2_ with interleaved MRoPE)
+processes a unified text+image sequence, and 32x32 patch embed/unembed
+shims map raw RGB in and out of LLM hidden space. The Qwen3-VL deepstack
+mergers go unused — their weights are dropped at load.
+"""
+
+from dataclasses import dataclass, field
+from typing import List, Optional
+
+import einops
+import torch
+import torch.nn as nn
+
+import comfy.patcher_extension
+from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder
+from comfy.text_encoders.llama import Llama2_
+from comfy.text_encoders.qwen35 import Qwen35VisionModel
+
+from .attention import make_two_pass_attention
+
+
+IMAGE_TOKEN_ID = 151655   # Qwen3-VL <|image_pad|>
+TMS_TOKEN_ID = 151673     # HiDream-O1 <|tms_token|>
+PATCH_SIZE = 32
+
+
+@dataclass
+class HiDreamO1TextConfig:
+    """Qwen3-VL-8B text-decoder dims (matches public Qwen3-VL-8B-Instruct)."""
+    vocab_size: int = 151936
+    hidden_size: int = 4096
+    intermediate_size: int = 12288
+    num_hidden_layers: int = 36
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 8
+    head_dim: int = 128
+    max_position_embeddings: int = 128000
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 5000000.0
+    rope_scale: Optional[float] = None
+    rope_dims: List[int] = field(default_factory=lambda: [24, 20, 20])
+    interleaved_mrope: bool = True
+    transformer_type: str = "llama"
+    rms_norm_add: bool = False
+    mlp_activation: str = "silu"
+    qkv_bias: bool = False
+    q_norm: str = "gemma3"
+    k_norm: str = "gemma3"
+    final_norm: bool = True
+    lm_head: bool = False
+    stop_tokens: List[int] = field(default_factory=lambda: [151643, 151645])
+
+
+QWEN3VL_VISION_DEFAULTS = dict(
+    hidden_size=1152,
+    num_heads=16,
+    intermediate_size=4304,
+    depth=27,
+    patch_size=16,
+    temporal_patch_size=2,
+    in_channels=3,
+    spatial_merge_size=2,
+    num_position_embeddings=2304,
+    deepstack_visual_indexes=(8, 16, 24),
+    out_hidden_size=4096,  # final merger projects directly into LLM hidden
+)
+
+
+class BottleneckPatchEmbed(nn.Module):
+    # 3072 -> 1024 -> 4096 (raw 32x32 RGB patch -> bottleneck -> LLM hidden).
+    def __init__(self, patch_size=32, in_chans=3, pca_dim=1024, embed_dim=4096, bias=True, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.proj1 = ops.Linear(patch_size * patch_size * in_chans, pca_dim, bias=False, device=device, dtype=dtype)
+        self.proj2 = ops.Linear(pca_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+
+    def forward(self, x):
+        return self.proj2(self.proj1(x))
+
+
+class FinalLayer(nn.Module):
+    # 4096 -> 3072 (LLM hidden -> flat pixel patch).
+    def __init__(self, hidden_size, patch_size=32, out_channels=3, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.linear = ops.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        return self.linear(x)
+
+
+class HiDreamO1Transformer(nn.Module):
+    """HiDream-O1 unified pixel-level transformer."""
+
+    def __init__(self, image_model=None, dtype=None, device=None, operations=None,
+                 text_config_overrides=None, vision_config_overrides=None, **kwargs):
+        super().__init__()
+        self.dtype = dtype
+
+        text_cfg = HiDreamO1TextConfig(**(text_config_overrides or {}))
+        vision_cfg = dict(QWEN3VL_VISION_DEFAULTS)
+        if vision_config_overrides:
+            vision_cfg.update(vision_config_overrides)
+        vision_cfg["out_hidden_size"] = text_cfg.hidden_size
+
+        self.text_config = text_cfg
+        self.vision_config = vision_cfg
+        self.hidden_size = text_cfg.hidden_size
+        self.patch_size = PATCH_SIZE
+        self.in_channels = 3
+        self.tms_token_id = TMS_TOKEN_ID
+
+        self.visual = Qwen35VisionModel(vision_cfg, device=device, dtype=dtype, ops=operations)
+        self.language_model = Llama2_(text_cfg, device=device, dtype=dtype, ops=operations)
+        self.t_embedder1 = TimestepEmbedder(
+            text_cfg.hidden_size, device=device, dtype=dtype, operations=operations,
+        )
+        self.x_embedder = BottleneckPatchEmbed(
+            patch_size=self.patch_size, in_chans=self.in_channels,
+            pca_dim=text_cfg.hidden_size // 4, embed_dim=text_cfg.hidden_size,
+            bias=True, device=device, dtype=dtype, ops=operations,
+        )
+        self.final_layer2 = FinalLayer(
+            text_cfg.hidden_size, patch_size=self.patch_size,
+            out_channels=self.in_channels, device=device, dtype=dtype, ops=operations,
+        )
+
+        self._visual_cache = None
+        self._kv_cache_entries = []
+
+    def clear_kv_cache(self):
+        self._kv_cache_entries = []
+        self._visual_cache = None
+
+    def forward(self, x, timesteps, context=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timesteps, context, transformer_options, **kwargs)
+
+    def _forward(self, x, timesteps, context=None, transformer_options={}, input_ids=None, attention_mask=None, position_ids=None,
+                 vinput_mask=None, ar_len=None, ref_pixel_values=None, ref_image_grid_thw=None, ref_patches=None, **kwargs):
+        """Returns flow-match velocity (x - x_pred) / sigma"""
+
+        if input_ids is None or position_ids is None:
+            raise ValueError("HiDreamO1Transformer requires input_ids and position_ids in conditioning")
+
+        B, _, H, W = x.shape
+        h_p, w_p = H // self.patch_size, W // self.patch_size
+        tgt_image_len = h_p * w_p
+
+        z = einops.rearrange(
+            x, 'B C (H p1) (W p2) -> B (H W) (C p1 p2)',
+            p1=self.patch_size, p2=self.patch_size,
+        )
+        vinputs = torch.cat([z, ref_patches.to(z.dtype)], dim=1) if ref_patches is not None else z
+
+        inputs_embeds = self.language_model.embed_tokens(input_ids).to(x.dtype)
+
+        if ref_pixel_values is not None and ref_image_grid_thw is not None:
+            # ViT output is constant across sampling steps within a generation
+            # identity-key by the input tensor so refs don't recompute every step.
+            cached = self._visual_cache
+            if cached is not None and cached[0] is ref_pixel_values:
+                image_embeds = cached[1]
+            else:
+                ref_pv = ref_pixel_values.to(inputs_embeds.device)
+                ref_grid = ref_image_grid_thw.to(inputs_embeds.device).long()
+                # extra_conds wraps with a leading batch dim; refs are model-level so [0] always recovers them.
+                if ref_pv.dim() == 3:
+                    ref_pv = ref_pv[0]
+                if ref_grid.dim() == 3:
+                    ref_grid = ref_grid[0]
+                image_embeds = self.visual(ref_pv, ref_grid).to(inputs_embeds.dtype)
+                self._visual_cache = (ref_pixel_values, image_embeds)
+            # image_pad positions identical across batch (input_ids shared cond/uncond).
+            image_idx = (input_ids[0] == IMAGE_TOKEN_ID).nonzero(as_tuple=True)[0]
+            if image_idx.shape[0] != image_embeds.shape[0]:
+                raise ValueError(
+                    f"Image-token count {image_idx.shape[0]} != ViT output count "
+                    f"{image_embeds.shape[0]}; check tokenizer/processor alignment."
+                )
+            inputs_embeds[:, image_idx] = image_embeds.unsqueeze(0).expand(B, -1, -1)
+
+        sigma = timesteps.float() / 1000.0
+        t_pixeldit = 1.0 - sigma
+        t_emb = self.t_embedder1(t_pixeldit * 1000, inputs_embeds.dtype)
+        tms_mask_3d = (input_ids == self.tms_token_id).unsqueeze(-1).expand_as(inputs_embeds)
+        inputs_embeds = torch.where(tms_mask_3d, t_emb.unsqueeze(1).expand_as(inputs_embeds), inputs_embeds)
+
+        vinputs_embedded = self.x_embedder(vinputs.to(inputs_embeds.dtype))
+        inputs_embeds = torch.cat([inputs_embeds, vinputs_embedded], dim=1)
+
+        # extra_conds stores position_ids as (1, 3, T); process_cond repeats dim 0 to B. Take row 0.
+        freqs_cis = self.language_model.compute_freqs_cis(position_ids[0].to(x.device), x.device)
+        freqs_cis = tuple(t.to(x.dtype) for t in freqs_cis)
+
+        two_pass_attn = make_two_pass_attention(ar_len, transformer_options=transformer_options)
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.language_model.layers)
+        transformer_options["block_type"] = "double"
+
+        # Cache prefix K/V across steps. Key includes input_ids (prompt), ref_id
+        # (refs scatter into inputs_embeds), and position_ids (RoPE baked into cached K).
+        can_cache = not blocks_replace and ar_len > 0
+        cache_len = ar_len if can_cache else 0
+        ref_id = id(ref_pixel_values) if ref_pixel_values is not None else None
+        pos_ids_key = position_ids[..., :cache_len] if can_cache else position_ids
+        cache_entries = self._kv_cache_entries
+        # Drop stale entries from a previous device (model was unloaded and reloaded).
+        if cache_entries and cache_entries[0]["input_ids"].device != input_ids.device:
+            cache_entries = []
+            self._kv_cache_entries = []
+        kv_cache = None
+        if can_cache:
+            for entry in cache_entries:
+                ck = entry["input_ids"]
+                ep = entry["position_ids"]
+                if (entry["cache_len"] == cache_len
+                        and ck.shape == input_ids.shape and torch.equal(ck, input_ids)
+                        and entry["ref_id"] == ref_id
+                        and ep.shape == pos_ids_key.shape and torch.equal(ep, pos_ids_key)):
+                    kv_cache = entry
+                    break
+
+        if kv_cache is not None:
+            # Hot path: project Q/K/V only for fresh positions; past_key_value prepends cached AR K/V.
+            hidden_states = inputs_embeds[:, cache_len:]
+            sliced_freqs = tuple(t[..., cache_len:, :] for t in freqs_cis)
+            for i, layer in enumerate(self.language_model.layers):
+                transformer_options["block_index"] = i
+                K_i, V_i = kv_cache["kv"][i]
+                hidden_states, _ = layer(
+                    x=hidden_states, attention_mask=None, freqs_cis=sliced_freqs, optimized_attention=two_pass_attn,
+                    past_key_value=(K_i, V_i, cache_len),
+                )
+        else:
+            # Cold path: run full sequence; if cacheable, snapshot K/V at AR positions.
+            snapshots = [] if can_cache else None
+            past_kv_cold = () if can_cache else None
+            hidden_states = inputs_embeds
+            for i, layer in enumerate(self.language_model.layers):
+                transformer_options["block_index"] = i
+                if ("double_block", i) in blocks_replace:
+                    def block_wrap(args, _layer=layer):
+                        out = {}
+                        out["x"], _ = _layer(
+                            x=args["x"], attention_mask=args.get("attention_mask"),
+                            freqs_cis=args["freqs_cis"], optimized_attention=args["optimized_attention"],
+                            past_key_value=None,
+                        )
+                        return out
+                    out = blocks_replace[("double_block", i)](
+                        {"x": hidden_states, "attention_mask": None,
+                         "freqs_cis": freqs_cis, "optimized_attention": two_pass_attn,
+                         "transformer_options": transformer_options},
+                        {"original_block": block_wrap},
+                    )
+                    hidden_states = out["x"]
+                else:
+                    hidden_states, present_kv = layer(
+                        x=hidden_states, attention_mask=None,
+                        freqs_cis=freqs_cis, optimized_attention=two_pass_attn,
+                        past_key_value=past_kv_cold,
+                    )
+                    if snapshots is not None:
+                        K, V, _ = present_kv
+                        snapshots.append((K[:, :, :cache_len].contiguous(),
+                                          V[:, :, :cache_len].contiguous()))
+            if snapshots is not None:
+                # Cap at 2 entries (cond + uncond). Multi-cond workflows LRU-evict.
+                new_entry = {
+                    "input_ids": input_ids.clone(),
+                    "cache_len": cache_len,
+                    "kv": snapshots,
+                    "ref_id": ref_id,
+                    "position_ids": pos_ids_key.clone(),
+                }
+                self._kv_cache_entries = (cache_entries + [new_entry])[-2:]
+
+        if self.language_model.norm is not None:
+            hidden_states = self.language_model.norm(hidden_states)
+
+        # Slice target-image positions before the final projection so the Linear only runs on tgt_image_len tokens.
+        # In the hot path hidden_states starts at original position cache_len, so masks/indices shift by cache_len.
+        sliced_offset = cache_len if kv_cache is not None else 0
+        if vinput_mask is not None:
+            vmask = vinput_mask.to(x.device).bool()
+            if sliced_offset > 0:
+                vmask = vmask[:, sliced_offset:]
+            target_hidden = hidden_states[vmask].view(B, -1, hidden_states.shape[-1])[:, :tgt_image_len]
+        else:
+            txt_seq_len = input_ids.shape[1]
+            start = txt_seq_len - sliced_offset
+            target_hidden = hidden_states[:, start:start + tgt_image_len]
+        x_pred_tgt = self.final_layer2(target_hidden)
+
+        # fp32 final subtraction, bf16 here noticeably degrades samples.
+        x_pred_img = einops.rearrange(
+            x_pred_tgt, 'B (H W) (C p1 p2) -> B C (H p1) (W p2)',
+            H=h_p, W=w_p, p1=self.patch_size, p2=self.patch_size,
+        )
+        return (x.float() - x_pred_img.float()) / sigma.view(B, 1, 1, 1).clamp_min(1e-3)

ComfyUI/comfy/ldm/hidream_o1/utils.py

@@ -0,0 +1,173 @@
+"""HiDream-O1 input-prep helpers: image/resolution math and unified-sequence
+RoPE position-id assembly. The fix_point offset in get_rope_index_fix_point
+lets the target image and patchified ref images share spatial RoPE positions
+despite living at different sequence indices — same 2D image plane.
+"""
+
+import math
+from typing import Optional
+
+import torch
+
+
+PATCH_SIZE = 32
+CONDITION_IMAGE_SIZE = 384  # ViT-side base size for ref images
+
+
+def resize_tensor(img_t, image_size, patch_size=16):
+    """img_t: (1, 3, H, W) float [0, 1]. Fit to image_size**2 area, patch-aligned, center-cropped."""
+
+    while min(img_t.shape[-2], img_t.shape[-1]) >= 2 * image_size: # Pre-halves with 2x2 box averaging while the image is still very large
+        img_t = torch.nn.functional.avg_pool2d(img_t, kernel_size=2, stride=2)
+
+    _, _, height, width = img_t.shape
+    m = patch_size
+    s_max = image_size * image_size
+    scale = math.sqrt(s_max / (width * height))
+
+    candidates = [
+        (round(width * scale) // m * m, round(height * scale) // m * m),
+        (round(width * scale) // m * m, math.floor(height * scale) // m * m),
+        (math.floor(width * scale) // m * m, round(height * scale) // m * m),
+        (math.floor(width * scale) // m * m, math.floor(height * scale) // m * m),
+    ]
+    candidates = sorted(candidates, key=lambda x: x[0] * x[1], reverse=True)
+    new_size = candidates[-1]
+    for c in candidates:
+        if c[0] * c[1] <= s_max:
+            new_size = c
+            break
+
+    new_w, new_h = new_size
+    s1 = width / new_w
+    s2 = height / new_h
+    if s1 < s2:
+        resize_w, resize_h = new_w, round(height / s1)
+    else:
+        resize_w, resize_h = round(width / s2), new_h
+    img_t = torch.nn.functional.interpolate(img_t, size=(resize_h, resize_w), mode="bicubic")
+    top = (resize_h - new_h) // 2
+    left = (resize_w - new_w) // 2
+    return img_t[..., top:top + new_h, left:left + new_w]
+
+
+def calculate_dimensions(max_size, ratio):
+    """(W, H) for an aspect ratio fitting in max_size**2 area, 32-aligned."""
+    width = math.sqrt(max_size * max_size * ratio)
+    height = width / ratio
+    width = int(width / 32) * 32
+    height = int(height / 32) * 32
+    return width, height
+
+
+def ref_max_size(target_max_dim, k):
+    """K-dependent ref-image max dim before patchifying."""
+    if k == 1:
+        return target_max_dim
+    if k == 2:
+        return target_max_dim * 48 // 64
+    if k <= 4:
+        return target_max_dim // 2
+    if k <= 8:
+        return target_max_dim * 24 // 64
+    return target_max_dim // 4
+
+
+def cond_image_size(k):
+    """K-dependent ViT-side image size."""
+    if k <= 4:
+        return CONDITION_IMAGE_SIZE
+    if k <= 8:
+        return CONDITION_IMAGE_SIZE * 48 // 64
+    return CONDITION_IMAGE_SIZE // 2
+
+
+def get_rope_index_fix_point(
+    spatial_merge_size: int,
+    image_token_id: int,
+    vision_start_token_id: int,
+    input_ids: Optional[torch.LongTensor] = None,
+    image_grid_thw: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    skip_vision_start_token=None,
+    fix_point: int = 4096,
+):
+    mrope_position_deltas = []
+    if input_ids is not None and image_grid_thw is not None:
+        total_input_ids = input_ids
+        if attention_mask is None:
+            attention_mask = torch.ones_like(total_input_ids)
+        position_ids = torch.ones(
+            3, input_ids.shape[0], input_ids.shape[1],
+            dtype=input_ids.dtype, device=input_ids.device,
+        )
+        attention_mask = attention_mask.to(total_input_ids.device)
+        for i, input_ids_b in enumerate(total_input_ids):
+            fp = fix_point
+            image_index = 0
+            input_ids_b = input_ids_b[attention_mask[i] == 1]
+            vision_start_indices = torch.argwhere(input_ids_b == vision_start_token_id).squeeze(1)
+            vision_tokens = input_ids_b[vision_start_indices + 1]
+            image_nums = (vision_tokens == image_token_id).sum()
+            input_tokens = input_ids_b.tolist()
+            llm_pos_ids_list = []
+            st = 0
+            remain_images = image_nums
+            for _ in range(image_nums):
+                if image_token_id in input_tokens and remain_images > 0:
+                    ed = input_tokens.index(image_token_id, st)
+                else:
+                    ed = len(input_tokens) + 1
+                t = image_grid_thw[image_index][0]
+                h = image_grid_thw[image_index][1]
+                w = image_grid_thw[image_index][2]
+                image_index += 1
+                remain_images -= 1
+                llm_grid_t = t.item()
+                llm_grid_h = h.item() // spatial_merge_size
+                llm_grid_w = w.item() // spatial_merge_size
+                text_len = ed - st
+                text_len -= skip_vision_start_token[image_index - 1]
+                text_len = max(0, text_len)
+                st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
+                llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)
+
+                t_index = torch.arange(llm_grid_t).view(-1, 1).expand(-1, llm_grid_h * llm_grid_w).flatten()
+                h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten()
+                w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten()
+
+                if skip_vision_start_token[image_index - 1]:
+                    if fp > 0:
+                        fp = fp - st_idx
+                    llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + fp + st_idx)
+                    fp = 0
+                else:
+                    llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + text_len + st_idx)
+                st = ed + llm_grid_t * llm_grid_h * llm_grid_w
+
+            if st < len(input_tokens):
+                st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
+                text_len = len(input_tokens) - st
+                llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)
+
+            llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
+            position_ids[..., i, attention_mask[i] == 1] = llm_positions.to(position_ids.device)
+            mrope_position_deltas.append(llm_positions.max() + 1 - len(total_input_ids[i]))
+        mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1)
+        return position_ids, mrope_position_deltas
+
+    if attention_mask is not None:
+        position_ids = attention_mask.long().cumsum(-1) - 1
+        position_ids.masked_fill_(attention_mask == 0, 1)
+        position_ids = position_ids.unsqueeze(0).expand(3, -1, -1).to(attention_mask.device)
+        max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0]
+        mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
+    else:
+        position_ids = (
+            torch.arange(input_ids.shape[1], device=input_ids.device)
+            .view(1, 1, -1).expand(3, input_ids.shape[0], -1)
+        )
+        mrope_position_deltas = torch.zeros(
+            [input_ids.shape[0], 1], device=input_ids.device, dtype=input_ids.dtype,
+        )
+    return position_ids, mrope_position_deltas

ComfyUI/comfy/ldm/hunyuan3d/model.py

@@ -0,0 +1,148 @@
+import torch
+from torch import nn
+from comfy.ldm.flux.layers import (
+    DoubleStreamBlock,
+    LastLayer,
+    MLPEmbedder,
+    SingleStreamBlock,
+    timestep_embedding,
+)
+import comfy.patcher_extension
+
+
+class Hunyuan3Dv2(nn.Module):
+    def __init__(
+        self,
+        in_channels=64,
+        context_in_dim=1536,
+        hidden_size=1024,
+        mlp_ratio=4.0,
+        num_heads=16,
+        depth=16,
+        depth_single_blocks=32,
+        qkv_bias=True,
+        guidance_embed=False,
+        image_model=None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.dtype = dtype
+
+        if hidden_size % num_heads != 0:
+            raise ValueError(
+                f"Hidden size {hidden_size} must be divisible by num_heads {num_heads}"
+            )
+
+        self.max_period = 1000  # While reimplementing the model I noticed that they messed up. This 1000 value was meant to be the time_factor but they set the max_period instead
+        self.latent_in = operations.Linear(in_channels, hidden_size, bias=True, dtype=dtype, device=device)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=hidden_size, dtype=dtype, device=device, operations=operations)
+        self.guidance_in = (
+            MLPEmbedder(in_dim=256, hidden_dim=hidden_size, dtype=dtype, device=device, operations=operations) if guidance_embed else None
+        )
+        self.cond_in = operations.Linear(context_in_dim, hidden_size, dtype=dtype, device=device)
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(depth)
+            ]
+        )
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(depth_single_blocks)
+            ]
+        )
+        self.final_layer = LastLayer(hidden_size, 1, in_channels, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, x, timestep, context, guidance=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, guidance, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, guidance=None, transformer_options={}, **kwargs):
+        x = x.movedim(-1, -2)
+        timestep = 1.0 - timestep
+        txt = context
+        img = self.latent_in(x)
+
+        vec = self.time_in(timestep_embedding(timestep, 256, self.max_period).to(dtype=img.dtype))
+        if self.guidance_in is not None:
+            if guidance is not None:
+                vec = vec + self.guidance_in(timestep_embedding(guidance, 256, self.max_period).to(img.dtype))
+
+        txt = self.cond_in(txt)
+        pe = None
+        attn_mask = None
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.double_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"], out["txt"] = block(img=args["img"],
+                                                   txt=args["txt"],
+                                                   vec=args["vec"],
+                                                   pe=args["pe"],
+                                                   attn_mask=args.get("attn_mask"),
+                                                   transformer_options=args["transformer_options"])
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": img,
+                                                           "txt": txt,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
+                                                          {"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,
+                                 transformer_options=transformer_options)
+
+        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"),
+                                       transformer_options=args["transformer_options"])
+                    return out
+
+                out = blocks_replace[("single_block", i)]({"img": img,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask,
+                                                           "transformer_options": transformer_options},
+                                                          {"original_block": block_wrap})
+                img = out["img"]
+            else:
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, transformer_options=transformer_options)
+
+        img = img[:, txt.shape[1]:, ...]
+        img = self.final_layer(img, vec)
+        return img.movedim(-2, -1) * (-1.0)

ComfyUI/comfy/ldm/hunyuan3d/vae.py

@@ -0,0 +1,988 @@
+# Original: https://github.com/Tencent/Hunyuan3D-2/blob/main/hy3dgen/shapegen/models/autoencoders/model.py
+# Since the header on their VAE source file was a bit confusing we asked for permission to use this code from tencent under the GPL license used in ComfyUI.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+import math
+from tqdm import tqdm
+
+from typing import Optional
+
+import logging
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+def fps(src: torch.Tensor, batch: torch.Tensor, sampling_ratio: float, start_random: bool = True):
+
+    # manually create the pointer vector
+    assert src.size(0) == batch.numel()
+
+    batch_size = int(batch.max()) + 1
+    deg = src.new_zeros(batch_size, dtype = torch.long)
+
+    deg.scatter_add_(0, batch, torch.ones_like(batch))
+
+    ptr_vec = deg.new_zeros(batch_size + 1)
+    torch.cumsum(deg, 0, out=ptr_vec[1:])
+
+    #return fps_sampling(src, ptr_vec, ratio)
+    sampled_indicies = []
+
+    for b in range(batch_size):
+        # start and the end of each batch
+        start, end = ptr_vec[b].item(), ptr_vec[b + 1].item()
+        # points from the point cloud
+        points = src[start:end]
+
+        num_points = points.size(0)
+        num_samples = max(1, math.ceil(num_points * sampling_ratio))
+
+        selected = torch.zeros(num_samples, device = src.device, dtype = torch.long)
+        distances = torch.full((num_points,), float("inf"), device = src.device)
+
+        # select a random start point
+        if start_random:
+            farthest = torch.randint(0, num_points, (1,), device = src.device)
+        else:
+            farthest = torch.tensor([0], device = src.device, dtype = torch.long)
+
+        for i in range(num_samples):
+            selected[i] = farthest
+            centroid = points[farthest].squeeze(0)
+            dist = torch.norm(points - centroid, dim = 1) # compute euclidean distance
+            distances = torch.minimum(distances, dist)
+            farthest = torch.argmax(distances)
+
+        sampled_indicies.append(torch.arange(start, end)[selected])
+
+    return torch.cat(sampled_indicies, dim = 0)
+class PointCrossAttention(nn.Module):
+    def __init__(self,
+        num_latents: int,
+        downsample_ratio: float,
+        pc_size: int,
+        pc_sharpedge_size: int,
+        point_feats: int,
+        width: int,
+        heads: int,
+        layers: int,
+        fourier_embedder,
+        normal_pe: bool = False,
+        qkv_bias: bool = False,
+        use_ln_post: bool = True,
+        qk_norm: bool = True):
+
+        super().__init__()
+
+        self.fourier_embedder = fourier_embedder
+
+        self.pc_size = pc_size
+        self.normal_pe = normal_pe
+        self.downsample_ratio = downsample_ratio
+        self.pc_sharpedge_size = pc_sharpedge_size
+        self.num_latents = num_latents
+        self.point_feats = point_feats
+
+        self.input_proj = nn.Linear(self.fourier_embedder.out_dim + point_feats, width)
+
+        self.cross_attn = ResidualCrossAttentionBlock(
+            width = width,
+            heads = heads,
+            qkv_bias = qkv_bias,
+            qk_norm = qk_norm
+        )
+
+        self.self_attn = None
+        if layers > 0:
+            self.self_attn = Transformer(
+                width = width,
+                heads = heads,
+                qkv_bias = qkv_bias,
+                qk_norm = qk_norm,
+                layers = layers
+            )
+
+        if use_ln_post:
+            self.ln_post = nn.LayerNorm(width)
+        else:
+            self.ln_post = None
+
+    def sample_points_and_latents(self, point_cloud: torch.Tensor, features: torch.Tensor):
+
+        """
+        Subsample points randomly from the point cloud (input_pc)
+        Further sample the subsampled points to get query_pc
+        take the fourier embeddings for both input and query pc
+
+        Mental Note: FPS-sampled points (query_pc) act as latent tokens that attend to and learn from the broader context in input_pc.
+        Goal: get a smaller represenation (query_pc) to represent the entire scence structure by learning from a broader subset (input_pc).
+        More computationally efficient.
+
+        Features are additional information for each point in the cloud
+        """
+
+        B, _, D = point_cloud.shape
+
+        num_latents = int(self.num_latents)
+
+        num_random_query = self.pc_size / (self.pc_size + self.pc_sharpedge_size) * num_latents
+        num_sharpedge_query = num_latents - num_random_query
+
+        # Split random and sharpedge surface points
+        random_pc, sharpedge_pc = torch.split(point_cloud, [self.pc_size, self.pc_sharpedge_size], dim=1)
+
+        # assert statements
+        assert random_pc.shape[1] <= self.pc_size, "Random surface points size must be less than or equal to pc_size"
+        assert sharpedge_pc.shape[1] <= self.pc_sharpedge_size, "Sharpedge surface points size must be less than or equal to pc_sharpedge_size"
+
+        input_random_pc_size = int(num_random_query * self.downsample_ratio)
+        random_query_pc, random_input_pc, random_idx_pc, random_idx_query = \
+            self.subsample(pc = random_pc, num_query = num_random_query, input_pc_size = input_random_pc_size)
+
+        input_sharpedge_pc_size = int(num_sharpedge_query * self.downsample_ratio)
+
+        if input_sharpedge_pc_size == 0:
+            sharpedge_input_pc = torch.zeros(B, 0, D, dtype = random_input_pc.dtype).to(point_cloud.device)
+            sharpedge_query_pc = torch.zeros(B, 0, D, dtype= random_query_pc.dtype).to(point_cloud.device)
+
+        else:
+            sharpedge_query_pc, sharpedge_input_pc, sharpedge_idx_pc, sharpedge_idx_query = \
+            self.subsample(pc = sharpedge_pc, num_query = num_sharpedge_query, input_pc_size = input_sharpedge_pc_size)
+
+        # concat the random and sharpedges
+        query_pc = torch.cat([random_query_pc, sharpedge_query_pc], dim = 1)
+        input_pc = torch.cat([random_input_pc, sharpedge_input_pc], dim = 1)
+
+        query = self.fourier_embedder(query_pc)
+        data = self.fourier_embedder(input_pc)
+
+        if self.point_feats > 0:
+            random_surface_features, sharpedge_surface_features = torch.split(features, [self.pc_size, self.pc_sharpedge_size], dim = 1)
+
+            input_random_surface_features, query_random_features = \
+                self.handle_features(features = random_surface_features, idx_pc = random_idx_pc, batch_size = B,
+                                     input_pc_size = input_random_pc_size, idx_query = random_idx_query)
+
+            if input_sharpedge_pc_size == 0:
+                input_sharpedge_surface_features = torch.zeros(B, 0, self.point_feats,
+                                                               dtype = input_random_surface_features.dtype, device = point_cloud.device)
+
+                query_sharpedge_features = torch.zeros(B, 0, self.point_feats,
+                                                       dtype = query_random_features.dtype, device = point_cloud.device)
+            else:
+
+                input_sharpedge_surface_features, query_sharpedge_features = \
+                    self.handle_features(idx_pc = sharpedge_idx_pc, features = sharpedge_surface_features,
+                                         batch_size = B, idx_query = sharpedge_idx_query, input_pc_size = input_sharpedge_pc_size)
+
+            query_features = torch.cat([query_random_features, query_sharpedge_features], dim = 1)
+            input_features = torch.cat([input_random_surface_features, input_sharpedge_surface_features], dim = 1)
+
+            if self.normal_pe:
+                # apply the fourier embeddings on the first 3 dims (xyz)
+                input_features_pe = self.fourier_embedder(input_features[..., :3])
+                query_features_pe = self.fourier_embedder(query_features[..., :3])
+                # replace the first 3 dims with the new PE ones
+                input_features = torch.cat([input_features_pe, input_features[..., :3]], dim = -1)
+                query_features = torch.cat([query_features_pe, query_features[..., :3]], dim = -1)
+
+            # concat at the channels dim
+            query = torch.cat([query, query_features], dim = -1)
+            data = torch.cat([data, input_features], dim = -1)
+
+        # don't return pc_info to avoid unnecessary memory usuage
+        return query.view(B, -1, query.shape[-1]), data.view(B, -1, data.shape[-1])
+
+    def forward(self, point_cloud: torch.Tensor, features: torch.Tensor):
+
+        query, data = self.sample_points_and_latents(point_cloud = point_cloud, features = features)
+
+        # apply projections
+        query = self.input_proj(query)
+        data = self.input_proj(data)
+
+        # apply cross attention between query and data
+        latents = self.cross_attn(query, data)
+
+        if self.self_attn is not None:
+            latents = self.self_attn(latents)
+
+        if self.ln_post is not None:
+            latents = self.ln_post(latents)
+
+        return latents
+
+
+    def subsample(self, pc, num_query, input_pc_size: int):
+
+        """
+        num_query: number of points to keep after FPS
+        input_pc_size: number of points to select before FPS
+        """
+
+        B, _, D = pc.shape
+        query_ratio = num_query / input_pc_size
+
+        # random subsampling of points inside the point cloud
+        idx_pc = torch.randperm(pc.shape[1], device = pc.device)[:input_pc_size]
+        input_pc = pc[:, idx_pc, :]
+
+        # flatten to allow applying fps across the whole batch
+        flattent_input_pc = input_pc.view(B * input_pc_size, D)
+
+        # construct a batch_down tensor to tell fps
+        # which points belong to which batch
+        N_down = int(flattent_input_pc.shape[0] / B)
+        batch_down = torch.arange(B).to(pc.device)
+        batch_down = torch.repeat_interleave(batch_down, N_down)
+
+        idx_query = fps(flattent_input_pc, batch_down, sampling_ratio = query_ratio)
+        query_pc = flattent_input_pc[idx_query].view(B, -1, D)
+
+        return query_pc, input_pc, idx_pc, idx_query
+
+    def handle_features(self, features, idx_pc, input_pc_size, batch_size: int, idx_query):
+
+        B = batch_size
+
+        input_surface_features = features[:, idx_pc, :]
+        flattent_input_features = input_surface_features.view(B * input_pc_size, -1)
+        query_features = flattent_input_features[idx_query].view(B, -1,
+                                                                 flattent_input_features.shape[-1])
+
+        return input_surface_features, query_features
+
+def normalize_mesh(mesh, scale = 0.9999):
+    """Normalize mesh to fit in [-scale, scale]. Translate mesh so its center is [0,0,0]"""
+
+    bbox = mesh.bounds
+    center = (bbox[1] + bbox[0]) / 2
+
+    max_extent = (bbox[1] - bbox[0]).max()
+    mesh.apply_translation(-center)
+    mesh.apply_scale((2 * scale) / max_extent)
+
+    return mesh
+
+def sample_pointcloud(mesh, num = 200000):
+    """ Uniformly sample points from the surface of the mesh """
+
+    points, face_idx = mesh.sample(num, return_index = True)
+    normals = mesh.face_normals[face_idx]
+    return torch.from_numpy(points.astype(np.float32)), torch.from_numpy(normals.astype(np.float32))
+
+def detect_sharp_edges(mesh, threshold=0.985):
+    """Return edge indices (a, b) that lie on sharp boundaries of the mesh."""
+
+    V, F = mesh.vertices, mesh.faces
+    VN, FN = mesh.vertex_normals, mesh.face_normals
+
+    sharp_mask = np.ones(V.shape[0])
+    for i in range(3):
+        indices = F[:, i]
+        alignment = np.einsum('ij,ij->i', VN[indices], FN)
+        dot_stack = np.stack((sharp_mask[indices], alignment), axis=-1)
+        sharp_mask[indices] = np.min(dot_stack, axis=-1)
+
+    edge_a = np.concatenate([F[:, 0], F[:, 1], F[:, 2]])
+    edge_b = np.concatenate([F[:, 1], F[:, 2], F[:, 0]])
+    sharp_edges = (sharp_mask[edge_a] < threshold) & (sharp_mask[edge_b] < threshold)
+
+    return edge_a[sharp_edges], edge_b[sharp_edges]
+
+
+def sharp_sample_pointcloud(mesh, num = 16384):
+    """ Sample points preferentially from sharp edges in the mesh. """
+
+    edge_a, edge_b = detect_sharp_edges(mesh)
+    V, VN = mesh.vertices, mesh.vertex_normals
+
+    va, vb = V[edge_a], V[edge_b]
+    na, nb = VN[edge_a], VN[edge_b]
+
+    edge_lengths = np.linalg.norm(vb - va, axis=-1)
+    weights = edge_lengths / edge_lengths.sum()
+
+    indices = np.searchsorted(np.cumsum(weights), np.random.rand(num))
+    t = np.random.rand(num, 1)
+
+    samples = t * va[indices] + (1 - t) * vb[indices]
+    normals = t * na[indices] + (1 - t) * nb[indices]
+
+    return samples.astype(np.float32), normals.astype(np.float32)
+
+def load_surface_sharpedge(mesh, num_points=4096, num_sharp_points=4096, sharpedge_flag = True, device = "cuda"):
+    """Load a surface with optional sharp-edge annotations from a trimesh mesh."""
+
+    import trimesh
+
+    try:
+        mesh_full = trimesh.util.concatenate(mesh.dump())
+    except Exception:
+        mesh_full = trimesh.util.concatenate(mesh)
+
+    mesh_full = normalize_mesh(mesh_full)
+
+    faces = mesh_full.faces
+    vertices = mesh_full.vertices
+    origin_face_count = faces.shape[0]
+
+    mesh_surface = trimesh.Trimesh(vertices=vertices, faces=faces[:origin_face_count])
+    mesh_fill = trimesh.Trimesh(vertices=vertices, faces=faces[origin_face_count:])
+
+    area_surface = mesh_surface.area
+    area_fill = mesh_fill.area
+    total_area = area_surface + area_fill
+
+    sample_num = 499712 // 2
+    fill_ratio = area_fill / total_area if total_area > 0 else 0
+
+    num_fill = int(sample_num * fill_ratio)
+    num_surface = sample_num - num_fill
+
+    surf_pts, surf_normals = sample_pointcloud(mesh_surface, num_surface)
+    fill_pts, fill_normals = (torch.zeros(0, 3), torch.zeros(0, 3)) if num_fill == 0 else sample_pointcloud(mesh_fill, num_fill)
+
+    sharp_pts, sharp_normals = sharp_sample_pointcloud(mesh_surface, sample_num)
+
+    def assemble_tensor(points, normals, label=None):
+
+        data = torch.cat([points, normals], dim=1).half().to(device)
+
+        if label is not None:
+            label_tensor = torch.full((data.shape[0], 1), float(label), dtype=torch.float16).to(device)
+            data = torch.cat([data, label_tensor], dim=1)
+
+        return data
+
+    surface = assemble_tensor(torch.cat([surf_pts.to(device), fill_pts.to(device)], dim=0),
+                              torch.cat([surf_normals.to(device), fill_normals.to(device)], dim=0),
+                              label = 0 if sharpedge_flag else None)
+
+    sharp_surface = assemble_tensor(torch.from_numpy(sharp_pts), torch.from_numpy(sharp_normals),
+                                    label = 1 if sharpedge_flag else None)
+
+    rng = np.random.default_rng()
+
+    surface = surface[rng.choice(surface.shape[0], num_points, replace = False)]
+    sharp_surface = sharp_surface[rng.choice(sharp_surface.shape[0], num_sharp_points, replace = False)]
+
+    full = torch.cat([surface, sharp_surface], dim = 0).unsqueeze(0)
+
+    return full
+
+class SharpEdgeSurfaceLoader:
+    """ Load mesh surface and sharp edge samples. """
+
+    def __init__(self, num_uniform_points = 8192, num_sharp_points = 8192):
+
+        self.num_uniform_points = num_uniform_points
+        self.num_sharp_points = num_sharp_points
+        self.total_points = num_uniform_points + num_sharp_points
+
+    def __call__(self, mesh_input, device = "cuda"):
+        mesh = self._load_mesh(mesh_input)
+        return load_surface_sharpedge(mesh, self.num_uniform_points, self.num_sharp_points, device = device)
+
+    @staticmethod
+    def _load_mesh(mesh_input):
+        import trimesh
+
+        if isinstance(mesh_input, str):
+            mesh = trimesh.load(mesh_input, force="mesh", merge_primitives = True)
+        else:
+            mesh = mesh_input
+
+        if isinstance(mesh, trimesh.Scene):
+            combined = None
+            for obj in mesh.geometry.values():
+                combined = obj if combined is None else combined + obj
+            return combined
+
+        return mesh
+
+class DiagonalGaussianDistribution:
+    def __init__(self, params: torch.Tensor, feature_dim: int = -1):
+
+        # divide quant channels (8) into mean and log variance
+        self.mean, self.logvar = torch.chunk(params, 2, dim = feature_dim)
+
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.std = torch.exp(0.5 * self.logvar)
+
+    def sample(self):
+
+        eps = torch.randn_like(self.std)
+        z = self.mean + eps * self.std
+
+        return z
+
+################################################
+# Volume Decoder
+################################################
+
+class VanillaVolumeDecoder():
+    @torch.no_grad()
+    def __call__(self, latents: torch.Tensor, geo_decoder: callable, octree_resolution: int, bounds = 1.01,
+                 num_chunks: int = 10_000, enable_pbar: bool = True, **kwargs):
+
+        if isinstance(bounds, float):
+            bounds = [-bounds, -bounds, -bounds, bounds, bounds, bounds]
+
+        bbox_min, bbox_max = torch.tensor(bounds[:3]), torch.tensor(bounds[3:])
+
+        x = torch.linspace(bbox_min[0], bbox_max[0], int(octree_resolution) + 1, dtype = torch.float32)
+        y = torch.linspace(bbox_min[1], bbox_max[1], int(octree_resolution) + 1, dtype = torch.float32)
+        z = torch.linspace(bbox_min[2], bbox_max[2], int(octree_resolution) + 1, dtype = torch.float32)
+
+        [xs, ys, zs] = torch.meshgrid(x, y, z, indexing = "ij")
+        xyz = torch.stack((xs, ys, zs), axis=-1).to(latents.device, dtype = latents.dtype).contiguous().reshape(-1, 3)
+        grid_size = [int(octree_resolution) + 1, int(octree_resolution) + 1, int(octree_resolution) + 1]
+
+        batch_logits = []
+        for start in tqdm(range(0, xyz.shape[0], num_chunks), desc="Volume Decoding",
+                          disable=not enable_pbar):
+
+            chunk_queries = xyz[start: start + num_chunks, :]
+            chunk_queries = chunk_queries.unsqueeze(0).repeat(latents.shape[0], 1, 1)
+            logits = geo_decoder(queries = chunk_queries, latents = latents)
+            batch_logits.append(logits)
+
+        grid_logits = torch.cat(batch_logits, dim = 1)
+        grid_logits = grid_logits.view((latents.shape[0], *grid_size)).float()
+
+        return grid_logits
+
+class FourierEmbedder(nn.Module):
+    """The sin/cosine positional embedding. Given an input tensor `x` of shape [n_batch, ..., c_dim], it converts
+    each feature dimension of `x[..., i]` into:
+        [
+            sin(x[..., i]),
+            sin(f_1*x[..., i]),
+            sin(f_2*x[..., i]),
+            ...
+            sin(f_N * x[..., i]),
+            cos(x[..., i]),
+            cos(f_1*x[..., i]),
+            cos(f_2*x[..., i]),
+            ...
+            cos(f_N * x[..., i]),
+            x[..., i]     # only present if include_input is True.
+        ], here f_i is the frequency.
+
+    Denote the space is [0 / num_freqs, 1 / num_freqs, 2 / num_freqs, 3 / num_freqs, ..., (num_freqs - 1) / num_freqs].
+    If logspace is True, then the frequency f_i is [2^(0 / num_freqs), ..., 2^(i / num_freqs), ...];
+    Otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)].
+
+    Args:
+        num_freqs (int): the number of frequencies, default is 6;
+        logspace (bool): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+            otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)];
+        input_dim (int): the input dimension, default is 3;
+        include_input (bool): include the input tensor or not, default is True.
+
+    Attributes:
+        frequencies (torch.Tensor): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+                otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1);
+
+        out_dim (int): the embedding size, if include_input is True, it is input_dim * (num_freqs * 2 + 1),
+            otherwise, it is input_dim * num_freqs * 2.
+
+    """
+
+    def __init__(self,
+                 num_freqs: int = 6,
+                 logspace: bool = True,
+                 input_dim: int = 3,
+                 include_input: bool = True,
+                 include_pi: bool = True) -> None:
+
+        """The initialization"""
+
+        super().__init__()
+
+        if logspace:
+            frequencies = 2.0 ** torch.arange(
+                num_freqs,
+                dtype=torch.float32
+            )
+        else:
+            frequencies = torch.linspace(
+                1.0,
+                2.0 ** (num_freqs - 1),
+                num_freqs,
+                dtype=torch.float32
+            )
+
+        if include_pi:
+            frequencies *= torch.pi
+
+        self.register_buffer("frequencies", frequencies, persistent=False)
+        self.include_input = include_input
+        self.num_freqs = num_freqs
+
+        self.out_dim = self.get_dims(input_dim)
+
+    def get_dims(self, input_dim):
+        temp = 1 if self.include_input or self.num_freqs == 0 else 0
+        out_dim = input_dim * (self.num_freqs * 2 + temp)
+
+        return out_dim
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """ Forward process.
+
+        Args:
+            x: tensor of shape [..., dim]
+
+        Returns:
+            embedding: an embedding of `x` of shape [..., dim * (num_freqs * 2 + temp)]
+                where temp is 1 if include_input is True and 0 otherwise.
+        """
+
+        if self.num_freqs > 0:
+            embed = (x[..., None].contiguous() * self.frequencies.to(device=x.device, dtype=x.dtype)).view(*x.shape[:-1], -1)
+            if self.include_input:
+                return torch.cat((x, embed.sin(), embed.cos()), dim=-1)
+            else:
+                return torch.cat((embed.sin(), embed.cos()), dim=-1)
+        else:
+            return x
+
+class CrossAttentionProcessor:
+    def __call__(self, attn, q, k, v):
+        out = comfy.ops.scaled_dot_product_attention(q, k, v)
+        return out
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+
+    def __init__(self, drop_prob: float = 0., scale_by_keep: bool = True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+        This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+        the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+        See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+        changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+        'survival rate' as the argument.
+
+        """
+        if self.drop_prob == 0. or not self.training:
+            return x
+        keep_prob = 1 - self.drop_prob
+        shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+        random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+        if keep_prob > 0.0 and self.scale_by_keep:
+            random_tensor.div_(keep_prob)
+        return x * random_tensor
+
+    def extra_repr(self):
+        return f'drop_prob={round(self.drop_prob, 3):0.3f}'
+
+
+class MLP(nn.Module):
+    def __init__(
+        self, *,
+        width: int,
+        expand_ratio: int = 4,
+        output_width: int = None,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.c_fc = ops.Linear(width, width * expand_ratio)
+        self.c_proj = ops.Linear(width * expand_ratio, output_width if output_width is not None else width)
+        self.gelu = nn.GELU()
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
+
+    def forward(self, x):
+        return self.drop_path(self.c_proj(self.gelu(self.c_fc(x))))
+
+class QKVMultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        heads: int,
+        n_data = None,
+        width=None,
+        qk_norm=False,
+        norm_layer=ops.LayerNorm
+    ):
+        super().__init__()
+        self.heads = heads
+        self.n_data = n_data
+        self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+
+    def forward(self, q, kv):
+
+        _, n_ctx, _ = q.shape
+        bs, n_data, width = kv.shape
+
+        attn_ch = width // self.heads // 2
+        q = q.view(bs, n_ctx, self.heads, -1)
+
+        kv = kv.view(bs, n_data, self.heads, -1)
+        k, v = torch.split(kv, attn_ch, dim=-1)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        q, k, v = [t.permute(0, 2, 1, 3) for t in (q, k, v)]
+        out = F.scaled_dot_product_attention(q, k, v)
+
+        out = out.transpose(1, 2).reshape(bs, n_ctx, -1)
+
+        return out
+
+class MultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        data_width: Optional[int] = None,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        kv_cache: bool = False,
+    ):
+        super().__init__()
+        self.width = width
+        self.heads = heads
+        self.data_width = width if data_width is None else data_width
+        self.c_q = ops.Linear(width, width, bias=qkv_bias)
+        self.c_kv = ops.Linear(self.data_width, width * 2, bias=qkv_bias)
+        self.c_proj = ops.Linear(width, width)
+        self.attention = QKVMultiheadCrossAttention(
+            heads=heads,
+            width=width,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.kv_cache = kv_cache
+        self.data = None
+
+    def forward(self, x, data):
+        x = self.c_q(x)
+        if self.kv_cache:
+            if self.data is None:
+                self.data = self.c_kv(data)
+                logging.info('Save kv cache,this should be called only once for one mesh')
+            data = self.data
+        else:
+            data = self.c_kv(data)
+        x = self.attention(x, data)
+        x = self.c_proj(x)
+        return x
+
+class ResidualCrossAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        mlp_expand_ratio: int = 4,
+        data_width: Optional[int] = None,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False
+    ):
+        super().__init__()
+
+        if data_width is None:
+            data_width = width
+
+        self.attn = MultiheadCrossAttention(
+            width=width,
+            heads=heads,
+            data_width=data_width,
+            qkv_bias=qkv_bias,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.ln_2 = norm_layer(data_width, elementwise_affine=True, eps=1e-6)
+        self.ln_3 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.mlp = MLP(width=width, expand_ratio=mlp_expand_ratio)
+
+    def forward(self, x: torch.Tensor, data: torch.Tensor):
+        x = x + self.attn(self.ln_1(x), self.ln_2(data))
+        x = x + self.mlp(self.ln_3(x))
+        return x
+
+
+class QKVMultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        heads: int,
+        width=None,
+        qk_norm=False,
+        norm_layer=ops.LayerNorm
+    ):
+        super().__init__()
+        self.heads = heads
+        self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+
+    def forward(self, qkv):
+        bs, n_ctx, width = qkv.shape
+        attn_ch = width // self.heads // 3
+        qkv = qkv.view(bs, n_ctx, self.heads, -1)
+        q, k, v = torch.split(qkv, attn_ch, dim=-1)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        q, k, v = [t.permute(0, 2, 1, 3) for t in (q, k, v)]
+        out = F.scaled_dot_product_attention(q, k, v).transpose(1, 2).reshape(bs, n_ctx, -1)
+        return out
+
+
+class MultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+
+        self.c_qkv = ops.Linear(width, width * 3, bias=qkv_bias)
+        self.c_proj = ops.Linear(width, width)
+        self.attention = QKVMultiheadAttention(
+            heads=heads,
+            width=width,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
+
+    def forward(self, x):
+        x = self.c_qkv(x)
+        x = self.attention(x)
+        x = self.drop_path(self.c_proj(x))
+        return x
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0,
+    ):
+        super().__init__()
+        self.attn = MultiheadAttention(
+            width=width,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm,
+            drop_path_rate=drop_path_rate
+        )
+        self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.mlp = MLP(width=width, drop_path_rate=drop_path_rate)
+        self.ln_2 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+
+    def forward(self, x: torch.Tensor):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        layers: int,
+        heads: int,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(
+                    width=width,
+                    heads=heads,
+                    qkv_bias=qkv_bias,
+                    norm_layer=norm_layer,
+                    qk_norm=qk_norm,
+                    drop_path_rate=drop_path_rate
+                )
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor):
+        for block in self.resblocks:
+            x = block(x)
+        return x
+
+
+class CrossAttentionDecoder(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        out_channels: int,
+        fourier_embedder: FourierEmbedder,
+        width: int,
+        heads: int,
+        mlp_expand_ratio: int = 4,
+        downsample_ratio: int = 1,
+        enable_ln_post: bool = True,
+        qkv_bias: bool = True,
+        qk_norm: bool = False,
+        label_type: str = "binary"
+    ):
+        super().__init__()
+
+        self.enable_ln_post = enable_ln_post
+        self.fourier_embedder = fourier_embedder
+        self.downsample_ratio = downsample_ratio
+        self.query_proj = ops.Linear(self.fourier_embedder.out_dim, width)
+        if self.downsample_ratio != 1:
+            self.latents_proj = ops.Linear(width * downsample_ratio, width)
+        if not self.enable_ln_post:
+            qk_norm = False
+        self.cross_attn_decoder = ResidualCrossAttentionBlock(
+            width=width,
+            mlp_expand_ratio=mlp_expand_ratio,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm
+        )
+
+        if self.enable_ln_post:
+            self.ln_post = ops.LayerNorm(width)
+        self.output_proj = ops.Linear(width, out_channels)
+        self.label_type = label_type
+        self.count = 0
+
+    def forward(self, queries=None, query_embeddings=None, latents=None):
+        if query_embeddings is None:
+            query_embeddings = self.query_proj(self.fourier_embedder(queries).to(latents.dtype))
+        self.count += query_embeddings.shape[1]
+        if self.downsample_ratio != 1:
+            latents = self.latents_proj(latents)
+        x = self.cross_attn_decoder(query_embeddings, latents)
+        if self.enable_ln_post:
+            x = self.ln_post(x)
+        occ = self.output_proj(x)
+        return occ
+
+
+class ShapeVAE(nn.Module):
+    def __init__(
+            self,
+            *,
+            num_latents: int = 4096,
+            embed_dim: int = 64,
+            width: int = 1024,
+            heads: int = 16,
+            num_decoder_layers: int = 16,
+            num_encoder_layers: int = 8,
+            pc_size: int = 81920,
+            pc_sharpedge_size: int = 0,
+            point_feats: int = 4,
+            downsample_ratio: int = 20,
+            geo_decoder_downsample_ratio: int = 1,
+            geo_decoder_mlp_expand_ratio: int = 4,
+            geo_decoder_ln_post: bool = True,
+            num_freqs: int = 8,
+            qkv_bias: bool = False,
+            qk_norm: bool = True,
+            drop_path_rate: float = 0.0,
+            include_pi: bool = False,
+            scale_factor: float = 1.0039506158752403,
+            label_type: str = "binary",
+    ):
+        super().__init__()
+        self.geo_decoder_ln_post = geo_decoder_ln_post
+
+        self.fourier_embedder = FourierEmbedder(num_freqs=num_freqs, include_pi=include_pi)
+
+        self.encoder = PointCrossAttention(layers = num_encoder_layers,
+                                    num_latents = num_latents,
+                                    downsample_ratio = downsample_ratio,
+                                    heads = heads,
+                                    pc_size = pc_size,
+                                    width = width,
+                                    point_feats = point_feats,
+                                    fourier_embedder = self.fourier_embedder,
+                                    pc_sharpedge_size = pc_sharpedge_size)
+
+        self.post_kl = ops.Linear(embed_dim, width)
+
+        self.transformer = Transformer(
+            width=width,
+            layers=num_decoder_layers,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm,
+            drop_path_rate=drop_path_rate
+        )
+
+        self.geo_decoder = CrossAttentionDecoder(
+            fourier_embedder=self.fourier_embedder,
+            out_channels=1,
+            mlp_expand_ratio=geo_decoder_mlp_expand_ratio,
+            downsample_ratio=geo_decoder_downsample_ratio,
+            enable_ln_post=self.geo_decoder_ln_post,
+            width=width // geo_decoder_downsample_ratio,
+            heads=heads // geo_decoder_downsample_ratio,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm,
+            label_type=label_type,
+        )
+
+        self.volume_decoder = VanillaVolumeDecoder()
+        self.scale_factor = scale_factor
+
+    def decode(self, latents, **kwargs):
+        latents = self.post_kl(latents.movedim(-2, -1))
+        latents = self.transformer(latents)
+
+        bounds = kwargs.get("bounds", 1.01)
+        num_chunks = kwargs.get("num_chunks", 8000)
+        octree_resolution = kwargs.get("octree_resolution", 256)
+        enable_pbar = kwargs.get("enable_pbar", True)
+
+        grid_logits = self.volume_decoder(latents, self.geo_decoder, bounds=bounds, num_chunks=num_chunks, octree_resolution=octree_resolution, enable_pbar=enable_pbar)
+        return grid_logits.movedim(-2, -1)
+
+    def encode(self, surface):
+
+        pc, feats = surface[:, :, :3], surface[:, :, 3:]
+        latents = self.encoder(pc, feats)
+
+        moments = self.pre_kl(latents)
+        posterior = DiagonalGaussianDistribution(moments, feature_dim = -1)
+
+        latents = posterior.sample()
+
+        return latents

ComfyUI/comfy/ldm/hunyuan3dv2_1/hunyuandit.py

@@ -0,0 +1,668 @@
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+
+class GELU(nn.Module):
+
+    def __init__(self, dim_in: int, dim_out: int, operations, device, dtype):
+        super().__init__()
+        self.proj = operations.Linear(dim_in, dim_out, device = device, dtype = dtype)
+
+    def gelu(self, gate: torch.Tensor) -> torch.Tensor:
+
+        if gate.device.type == "mps":
+            return F.gelu(gate.to(dtype = torch.float32)).to(dtype = gate.dtype)
+
+        return F.gelu(gate)
+
+    def forward(self, hidden_states):
+
+        hidden_states = self.proj(hidden_states)
+        hidden_states = self.gelu(hidden_states)
+
+        return hidden_states
+
+class FeedForward(nn.Module):
+
+    def __init__(self, dim: int, dim_out = None, mult: int = 4,
+                dropout: float = 0.0, inner_dim = None, operations = None, device = None, dtype = None):
+
+        super().__init__()
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
+
+        dim_out = dim_out if dim_out is not None else dim
+
+        act_fn = GELU(dim, inner_dim, operations = operations, device = device, dtype = dtype)
+
+        self.net = nn.ModuleList([])
+        self.net.append(act_fn)
+
+        self.net.append(nn.Dropout(dropout))
+        self.net.append(operations.Linear(inner_dim, dim_out, device = device, dtype = dtype))
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states
+
+class AddAuxLoss(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, x, loss):
+        # do nothing in forward (no computation)
+        ctx.requires_aux_loss = loss.requires_grad
+        ctx.dtype = loss.dtype
+
+        return x
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        # add the aux loss gradients
+        grad_loss = None
+        # put the aux grad the same as the main grad loss
+        # aux grad contributes equally
+        if ctx.requires_aux_loss:
+            grad_loss = torch.ones(1, dtype = ctx.dtype, device = grad_output.device)
+
+        return grad_output, grad_loss
+
+class MoEGate(nn.Module):
+
+    def __init__(self, embed_dim, num_experts=16, num_experts_per_tok=2, aux_loss_alpha=0.01, device = None, dtype = None):
+
+        super().__init__()
+        self.top_k = num_experts_per_tok
+        self.n_routed_experts = num_experts
+
+        self.alpha = aux_loss_alpha
+
+        self.gating_dim = embed_dim
+        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim), device = device, dtype = dtype))
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+
+        # flatten hidden states
+        hidden_states = hidden_states.view(-1, hidden_states.size(-1))
+
+        # get logits and pass it to softmax
+        logits = F.linear(hidden_states, comfy.model_management.cast_to(self.weight, dtype=hidden_states.dtype, device=hidden_states.device), bias = None)
+        scores = logits.softmax(dim = -1)
+
+        topk_weight, topk_idx = torch.topk(scores, k = self.top_k, dim = -1, sorted = False)
+
+        if self.training and self.alpha > 0.0:
+            scores_for_aux = scores
+
+            # used bincount instead of one hot encoding
+            counts = torch.bincount(topk_idx.view(-1), minlength = self.n_routed_experts).float()
+            ce = counts / topk_idx.numel()  # normalized expert usage
+
+            # mean expert score
+            Pi = scores_for_aux.mean(0)
+
+            # expert balance loss
+            aux_loss = (Pi * ce * self.n_routed_experts).sum() * self.alpha
+        else:
+            aux_loss = None
+
+        return topk_idx, topk_weight, aux_loss
+
+class MoEBlock(nn.Module):
+    def __init__(self, dim, num_experts: int = 6, moe_top_k: int = 2, dropout: float = 0.0,
+                 ff_inner_dim: int = None, operations = None, device = None, dtype = None):
+        super().__init__()
+
+        self.moe_top_k = moe_top_k
+        self.num_experts = num_experts
+
+        self.experts = nn.ModuleList([
+            FeedForward(dim, dropout = dropout, inner_dim = ff_inner_dim, operations = operations, device = device, dtype = dtype)
+            for _ in range(num_experts)
+        ])
+
+        self.gate = MoEGate(dim, num_experts = num_experts, num_experts_per_tok = moe_top_k, device = device, dtype = dtype)
+        self.shared_experts = FeedForward(dim, dropout = dropout, inner_dim = ff_inner_dim, operations = operations, device = device, dtype = dtype)
+
+    def forward(self, hidden_states) -> torch.Tensor:
+
+        identity = hidden_states
+        orig_shape = hidden_states.shape
+        topk_idx, topk_weight, aux_loss = self.gate(hidden_states)
+
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        flat_topk_idx = topk_idx.view(-1)
+
+        if self.training:
+
+            hidden_states = hidden_states.repeat_interleave(self.moe_top_k, dim = 0)
+            y = torch.empty_like(hidden_states, dtype = hidden_states.dtype)
+
+            for i, expert in enumerate(self.experts):
+                tmp = expert(hidden_states[flat_topk_idx == i])
+                y[flat_topk_idx == i] = tmp.to(hidden_states.dtype)
+
+            y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim = 1)
+            y =  y.view(*orig_shape)
+
+            y = AddAuxLoss.apply(y, aux_loss)
+        else:
+            y = self.moe_infer(hidden_states, flat_expert_indices = flat_topk_idx,flat_expert_weights = topk_weight.view(-1, 1)).view(*orig_shape)
+
+        y = y + self.shared_experts(identity)
+
+        return y
+
+    @torch.no_grad()
+    def moe_infer(self, x, flat_expert_indices, flat_expert_weights):
+
+        expert_cache = torch.zeros_like(x)
+        idxs = flat_expert_indices.argsort()
+
+        # no need for .numpy().cpu() here
+        tokens_per_expert = flat_expert_indices.bincount().cumsum(0)
+        token_idxs = idxs // self.moe_top_k
+
+        for i, end_idx in enumerate(tokens_per_expert):
+
+            start_idx = 0 if i == 0 else tokens_per_expert[i-1]
+
+            if start_idx == end_idx:
+                continue
+
+            expert = self.experts[i]
+            exp_token_idx = token_idxs[start_idx:end_idx]
+
+            expert_tokens = x[exp_token_idx]
+            expert_out = expert(expert_tokens)
+
+            expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]])
+
+            # use index_add_ with a 1-D index tensor directly avoids building a large [N, D] index map and extra memcopy required by scatter_reduce_
+            # + avoid dtype conversion
+            expert_cache.index_add_(0, exp_token_idx, expert_out)
+
+        return expert_cache
+
+class Timesteps(nn.Module):
+    def __init__(self, num_channels: int, downscale_freq_shift: float = 0.0,
+                 scale: float = 1.0, max_period: int = 10000):
+        super().__init__()
+
+        self.num_channels = num_channels
+        half_dim = num_channels // 2
+
+        # precompute the “inv_freq” vector once
+        exponent = -math.log(max_period) * torch.arange(
+            half_dim, dtype=torch.float32
+        ) / (half_dim - downscale_freq_shift)
+
+        inv_freq = torch.exp(exponent)
+
+        # pad
+        if num_channels % 2 == 1:
+            # we’ll pad a zero at the end of the cos-half
+            inv_freq = torch.cat([inv_freq, inv_freq.new_zeros(1)])
+
+        # register to buffer so it moves with the device
+        self.register_buffer("inv_freq", inv_freq, persistent = False)
+        self.scale = scale
+
+    def forward(self, timesteps: torch.Tensor):
+
+        x = timesteps.float().unsqueeze(1) * self.inv_freq.to(timesteps.device).unsqueeze(0)
+
+
+        # fused CUDA kernels for sin and cos
+        sin_emb = x.sin()
+        cos_emb = x.cos()
+
+        emb = torch.cat([sin_emb, cos_emb], dim = 1)
+
+        # scale factor
+        if self.scale != 1.0:
+            emb = emb * self.scale
+
+        # If we padded inv_freq for odd, emb is already wide enough; otherwise:
+        if emb.shape[1] > self.num_channels:
+            emb = emb[:, :self.num_channels]
+
+        return emb
+
+class TimestepEmbedder(nn.Module):
+    def __init__(self, hidden_size, frequency_embedding_size = 256, cond_proj_dim = None, operations = None, device = None, dtype = None):
+        super().__init__()
+
+        self.mlp = nn.Sequential(
+            operations.Linear(hidden_size, frequency_embedding_size, bias=True, device = device, dtype = dtype),
+            nn.GELU(),
+            operations.Linear(frequency_embedding_size, hidden_size, bias=True, device = device, dtype = dtype),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+
+        if cond_proj_dim is not None:
+            self.cond_proj = operations.Linear(cond_proj_dim, frequency_embedding_size, bias=False, device = device, dtype = dtype)
+
+        self.time_embed = Timesteps(hidden_size)
+
+    def forward(self, timesteps, condition):
+
+        timestep_embed = self.time_embed(timesteps).type(self.mlp[0].weight.dtype)
+
+        if condition is not None:
+            cond_embed = self.cond_proj(condition)
+            timestep_embed = timestep_embed + cond_embed
+
+        time_conditioned = self.mlp(timestep_embed)
+
+        # for broadcasting with image tokens
+        return time_conditioned.unsqueeze(1)
+
+class MLP(nn.Module):
+    def __init__(self, *, width: int, operations = None, device = None, dtype = None):
+        super().__init__()
+        self.width = width
+        self.fc1 = operations.Linear(width, width * 4, device = device, dtype = dtype)
+        self.fc2 = operations.Linear(width * 4, width, device = device, dtype = dtype)
+        self.gelu = nn.GELU()
+
+    def forward(self, x):
+        return self.fc2(self.gelu(self.fc1(x)))
+
+class CrossAttention(nn.Module):
+    def __init__(
+        self,
+        qdim,
+        kdim,
+        num_heads,
+        qkv_bias=True,
+        qk_norm=False,
+        norm_layer=nn.LayerNorm,
+        use_fp16: bool = False,
+        operations = None,
+        dtype = None,
+        device = None,
+        **kwargs,
+    ):
+        super().__init__()
+        self.qdim = qdim
+        self.kdim = kdim
+
+        self.num_heads = num_heads
+        self.head_dim = self.qdim // num_heads
+
+        self.scale = self.head_dim ** -0.5
+
+        self.to_q = operations.Linear(qdim, qdim, bias=qkv_bias, device = device, dtype = dtype)
+        self.to_k = operations.Linear(kdim, qdim, bias=qkv_bias, device = device, dtype = dtype)
+        self.to_v = operations.Linear(kdim, qdim, bias=qkv_bias, device = device, dtype = dtype)
+
+        if use_fp16:
+            eps = 1.0 / 65504
+        else:
+            eps = 1e-6
+
+        if norm_layer == nn.LayerNorm:
+            norm_layer = operations.LayerNorm
+        else:
+            norm_layer = operations.RMSNorm
+
+        self.q_norm = norm_layer(self.head_dim, elementwise_affine=True, eps = eps, device = device, dtype = dtype) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(self.head_dim, elementwise_affine=True, eps = eps, device = device, dtype = dtype) if qk_norm else nn.Identity()
+        self.out_proj = operations.Linear(qdim, qdim, bias=True, device = device, dtype = dtype)
+
+    def forward(self, x, y):
+
+        b, s1, _ = x.shape
+        _, s2, _ = y.shape
+
+        y = y.to(next(self.to_k.parameters()).dtype)
+
+        q = self.to_q(x)
+        k = self.to_k(y)
+        v = self.to_v(y)
+
+        kv = torch.cat((k, v), dim=-1)
+        split_size = kv.shape[-1] // self.num_heads // 2
+
+        kv = kv.view(b, -1, self.num_heads, split_size * 2)
+        k, v = torch.split(kv, split_size, dim=-1)
+
+        q = q.view(b, s1, self.num_heads, self.head_dim)
+        k = k.view(b, s2, self.num_heads, self.head_dim)
+        v = v.reshape(b, s2, self.num_heads * self.head_dim)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        x = optimized_attention(
+            q.reshape(b, s1, self.num_heads * self.head_dim),
+            k.reshape(b, s2, self.num_heads * self.head_dim),
+            v,
+            heads=self.num_heads,
+            low_precision_attention=False,
+        )
+
+        out = self.out_proj(x)
+
+        return out
+
+class Attention(nn.Module):
+
+    def __init__(
+        self,
+        dim,
+        num_heads,
+        qkv_bias = True,
+        qk_norm = False,
+        norm_layer = nn.LayerNorm,
+        use_fp16: bool = False,
+        operations = None,
+        device = None,
+        dtype = None
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = self.dim // num_heads
+        self.scale = self.head_dim ** -0.5
+
+        self.to_q = operations.Linear(dim, dim, bias = qkv_bias, device = device, dtype = dtype)
+        self.to_k = operations.Linear(dim, dim, bias = qkv_bias, device = device, dtype = dtype)
+        self.to_v = operations.Linear(dim, dim, bias = qkv_bias, device = device, dtype = dtype)
+
+        if use_fp16:
+            eps = 1.0 / 65504
+        else:
+            eps = 1e-6
+
+        if norm_layer == nn.LayerNorm:
+            norm_layer = operations.LayerNorm
+        else:
+            norm_layer = operations.RMSNorm
+
+        self.q_norm = norm_layer(self.head_dim, elementwise_affine=True, eps = eps, device = device, dtype = dtype) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(self.head_dim, elementwise_affine=True, eps = eps, device = device, dtype = dtype) if qk_norm else nn.Identity()
+        self.out_proj = operations.Linear(dim, dim, device = device, dtype = dtype)
+
+    def forward(self, x):
+        B, N, _ = x.shape
+
+        query = self.to_q(x)
+        key = self.to_k(x)
+        value = self.to_v(x)
+
+        qkv_combined = torch.cat((query, key, value), dim=-1)
+        split_size = qkv_combined.shape[-1] // self.num_heads // 3
+
+        qkv = qkv_combined.view(B, -1, self.num_heads, split_size * 3)
+        query, key, value = torch.split(qkv, split_size, dim=-1)
+
+        query = query.reshape(B, N, self.num_heads, self.head_dim)
+        key = key.reshape(B, N, self.num_heads, self.head_dim)
+        value = value.reshape(B, N, self.num_heads * self.head_dim)
+
+        query = self.q_norm(query)
+        key = self.k_norm(key)
+
+        x = optimized_attention(
+            query.reshape(B, N, self.num_heads * self.head_dim),
+            key.reshape(B, N, self.num_heads * self.head_dim),
+            value,
+            heads=self.num_heads,
+            low_precision_attention=False,
+        )
+
+        x = self.out_proj(x)
+        return x
+
+class HunYuanDiTBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        c_emb_size,
+        num_heads,
+        text_states_dim=1024,
+        qk_norm=False,
+        norm_layer=nn.LayerNorm,
+        qk_norm_layer=True,
+        qkv_bias=True,
+        skip_connection=True,
+        timested_modulate=False,
+        use_moe: bool = False,
+        num_experts: int = 8,
+        moe_top_k: int = 2,
+        use_fp16: bool = False,
+        operations = None,
+        device = None, dtype = None
+    ):
+        super().__init__()
+
+        # eps can't be 1e-6 in fp16 mode because of numerical stability issues
+        if use_fp16:
+            eps = 1.0 / 65504
+        else:
+            eps = 1e-6
+
+        self.norm1 = norm_layer(hidden_size, elementwise_affine = True, eps = eps, device = device, dtype = dtype)
+
+        self.attn1 = Attention(hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, qk_norm=qk_norm,
+                               norm_layer=qk_norm_layer, use_fp16 = use_fp16, device = device, dtype = dtype, operations = operations)
+
+        self.norm2 = norm_layer(hidden_size, elementwise_affine = True, eps = eps, device = device, dtype = dtype)
+
+        self.timested_modulate = timested_modulate
+        if self.timested_modulate:
+            self.default_modulation = nn.Sequential(
+                nn.SiLU(),
+                operations.Linear(c_emb_size, hidden_size, bias=True, device = device, dtype = dtype)
+            )
+
+        self.attn2 = CrossAttention(hidden_size, text_states_dim, num_heads=num_heads, qkv_bias=qkv_bias,
+                                    qk_norm=qk_norm, norm_layer=qk_norm_layer, use_fp16 = use_fp16,
+                                    device = device, dtype = dtype, operations = operations)
+
+        self.norm3 = norm_layer(hidden_size, elementwise_affine = True, eps = eps, device = device, dtype = dtype)
+
+        if skip_connection:
+            self.skip_norm = norm_layer(hidden_size, elementwise_affine = True, eps = eps, device = device, dtype = dtype)
+            self.skip_linear = operations.Linear(2 * hidden_size, hidden_size, device = device, dtype = dtype)
+        else:
+            self.skip_linear = None
+
+        self.use_moe = use_moe
+
+        if self.use_moe:
+            self.moe = MoEBlock(
+                hidden_size,
+                num_experts = num_experts,
+                moe_top_k = moe_top_k,
+                dropout = 0.0,
+                ff_inner_dim = int(hidden_size * 4.0),
+                device = device, dtype = dtype,
+                operations = operations
+            )
+        else:
+            self.mlp = MLP(width=hidden_size, operations=operations, device = device, dtype = dtype)
+
+    def forward(self, hidden_states, conditioning=None, text_states=None, skip_tensor=None):
+
+        if self.skip_linear is not None:
+            combined = torch.cat([skip_tensor, hidden_states], dim=-1)
+            hidden_states = self.skip_linear(combined)
+            hidden_states = self.skip_norm(hidden_states)
+
+        # self attention
+        if self.timested_modulate:
+            modulation_shift = self.default_modulation(conditioning).unsqueeze(dim=1)
+            hidden_states = hidden_states + modulation_shift
+
+        self_attn_out = self.attn1(self.norm1(hidden_states))
+        hidden_states = hidden_states + self_attn_out
+
+        # cross attention
+        hidden_states = hidden_states + self.attn2(self.norm2(hidden_states), text_states)
+
+        # MLP Layer
+        mlp_input = self.norm3(hidden_states)
+
+        if self.use_moe:
+            hidden_states = hidden_states + self.moe(mlp_input)
+        else:
+            hidden_states = hidden_states + self.mlp(mlp_input)
+
+        return hidden_states
+
+class FinalLayer(nn.Module):
+
+    def __init__(self, final_hidden_size, out_channels, operations, use_fp16: bool = False, device = None, dtype = None):
+        super().__init__()
+
+        if use_fp16:
+            eps = 1.0 / 65504
+        else:
+            eps = 1e-6
+
+        self.norm_final = operations.LayerNorm(final_hidden_size, elementwise_affine = True, eps = eps, device = device, dtype = dtype)
+        self.linear = operations.Linear(final_hidden_size, out_channels, bias = True, device = device, dtype = dtype)
+
+    def forward(self, x):
+        x = self.norm_final(x)
+        x = x[:, 1:]
+        x = self.linear(x)
+        return x
+
+class HunYuanDiTPlain(nn.Module):
+
+    # init with the defaults values from https://huggingface.co/tencent/Hunyuan3D-2.1/blob/main/hunyuan3d-dit-v2-1/config.yaml
+    def __init__(
+        self,
+        in_channels: int = 64,
+        hidden_size: int = 2048,
+        context_dim: int = 1024,
+        depth: int = 21,
+        num_heads: int = 16,
+        qk_norm: bool = True,
+        qkv_bias: bool = False,
+        num_moe_layers: int = 6,
+        guidance_cond_proj_dim = 2048,
+        norm_type = 'layer',
+        num_experts: int = 8,
+        moe_top_k: int = 2,
+        use_fp16: bool = False,
+        dtype = None,
+        device = None,
+        operations = None,
+        **kwargs
+        ):
+
+        self.dtype = dtype
+
+        super().__init__()
+
+        self.depth = depth
+
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+
+        norm = operations.LayerNorm if norm_type == 'layer' else operations.RMSNorm
+        qk_norm = operations.RMSNorm
+
+        self.context_dim = context_dim
+        self.guidance_cond_proj_dim = guidance_cond_proj_dim
+
+        self.x_embedder = operations.Linear(in_channels, hidden_size, bias = True, device = device, dtype = dtype)
+        self.t_embedder = TimestepEmbedder(hidden_size, hidden_size * 4, cond_proj_dim = guidance_cond_proj_dim, device = device, dtype = dtype, operations = operations)
+
+
+        # HUnYuanDiT Blocks
+        self.blocks = nn.ModuleList([
+            HunYuanDiTBlock(hidden_size=hidden_size,
+                            c_emb_size=hidden_size,
+                            num_heads=num_heads,
+                            text_states_dim=context_dim,
+                            qk_norm=qk_norm,
+                            norm_layer = norm,
+                            qk_norm_layer = qk_norm,
+                            skip_connection=layer > depth // 2,
+                            qkv_bias=qkv_bias,
+                            use_moe=True if depth - layer <= num_moe_layers else False,
+                            num_experts=num_experts,
+                            moe_top_k=moe_top_k,
+                            use_fp16 = use_fp16,
+                            device = device, dtype = dtype, operations = operations)
+            for layer in range(depth)
+        ])
+
+        self.depth = depth
+
+        self.final_layer = FinalLayer(hidden_size, self.out_channels, use_fp16 = use_fp16, operations = operations, device = device, dtype = dtype)
+
+    def forward(self, x, t, context, transformer_options = {}, **kwargs):
+
+        x = x.movedim(-1, -2)
+
+        swap_cfg_halves = context.shape[0] >= 2
+
+        if swap_cfg_halves:
+            first_half, second_half = context.chunk(2, dim = 0)
+            context = torch.cat([second_half, first_half], dim = 0)
+
+        main_condition = context
+
+        t = 1.0 - t
+
+        time_embedded = self.t_embedder(t, condition = kwargs.get('guidance_cond'))
+
+        x = x.to(dtype = next(self.x_embedder.parameters()).dtype)
+        x_embedded = self.x_embedder(x)
+
+        combined = torch.cat([time_embedded, x_embedded], dim=1)
+
+        def block_wrap(args):
+            return block(
+                args["x"],
+                args["t"],
+                args["cond"],
+                skip_tensor=args.get("skip"),)
+
+        skip_stack = []
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for idx, block in enumerate(self.blocks):
+            if idx <= self.depth // 2:
+                skip_input = None
+            else:
+                skip_input = skip_stack.pop()
+
+            if ("block", idx) in blocks_replace:
+
+                combined = blocks_replace[("block", idx)](
+                    {
+                        "x": combined,
+                        "t": time_embedded,
+                        "cond": main_condition,
+                        "skip": skip_input,
+                    },
+                    {"original_block": block_wrap},
+                )
+            else:
+                combined = block(combined, time_embedded, main_condition, skip_tensor=skip_input)
+
+            if idx < self.depth // 2:
+                skip_stack.append(combined)
+
+        output = self.final_layer(combined)
+        output =  output.movedim(-2, -1) * (-1.0)
+
+        if swap_cfg_halves:
+            first_half, second_half = output.chunk(2, dim = 0)
+            output = torch.cat([second_half, first_half], dim = 0)
+
+        return output