Basic hunyuan dit implementation. (#4102)

* Let tokenizers return weights to be stored in the saved checkpoint.

* Basic hunyuan dit implementation.

* Fix some resolutions not working.

* Support hydit checkpoint save.

* Init with right dtype.

* Switch to optimized attention in pooler.

* Fix black images on hunyuan dit.

comfy/ldm/hydit/attn_layers.py

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

comfy/ldm/hydit/models.py

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

comfy/ldm/hydit/poolers.py

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

comfy/ldm/hydit/posemb_layers.py

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

comfy/model_base.py

@@ -7,6 +7,7 @@ from comfy.ldm.modules.encoders.noise_aug_modules import CLIPEmbeddingNoiseAugme
 from comfy.ldm.modules.diffusionmodules.upscaling import ImageConcatWithNoiseAugmentation
 from comfy.ldm.modules.diffusionmodules.mmdit import OpenAISignatureMMDITWrapper
 import comfy.ldm.aura.mmdit
+import comfy.ldm.hydit.models
 import comfy.ldm.audio.dit
 import comfy.ldm.audio.embedders
 import comfy.model_management
@@ -648,3 +649,35 @@ class StableAudio1(BaseModel):
             for l in s:
                 sd["{}{}".format(k, l)] = s[l]
         return sd
+
+class HunyuanDiT(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.V_PREDICTION, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hydit.models.HunYuanDiT)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out['text_embedding_mask'] = comfy.conds.CONDRegular(attention_mask)
+
+        conditioning_mt5xl = kwargs.get("conditioning_mt5xl", None)
+        if conditioning_mt5xl is not None:
+            out['encoder_hidden_states_t5'] = comfy.conds.CONDRegular(conditioning_mt5xl)
+
+        attention_mask_mt5xl = kwargs.get("attention_mask_mt5xl", None)
+        if attention_mask_mt5xl is not None:
+            out['text_embedding_mask_t5'] = comfy.conds.CONDRegular(attention_mask_mt5xl)
+
+        width = kwargs.get("width", 768)
+        height = kwargs.get("height", 768)
+        crop_w = kwargs.get("crop_w", 0)
+        crop_h = kwargs.get("crop_h", 0)
+        target_width = kwargs.get("target_width", width)
+        target_height = kwargs.get("target_height", height)
+
+        out['image_meta_size'] = comfy.conds.CONDRegular(torch.FloatTensor([[height, width, target_height, target_width, 0, 0]]))
+        return out

comfy/model_detection.py

@@ -115,6 +115,19 @@ def detect_unet_config(state_dict, key_prefix):
         unet_config["n_layers"] = double_layers + single_layers
         return unet_config
 
+    if '{}mlp_t5.0.weight'.format(key_prefix) in state_dict_keys: #Hunyuan DiT
+        unet_config = {}
+        unet_config["image_model"] = "hydit"
+        unet_config["depth"] = count_blocks(state_dict_keys, '{}blocks.'.format(key_prefix) + '{}.')
+        unet_config["hidden_size"] = state_dict['{}x_embedder.proj.weight'.format(key_prefix)].shape[0]
+        if unet_config["hidden_size"] == 1408 and unet_config["depth"] == 40: #DiT-g/2
+            unet_config["mlp_ratio"] = 4.3637
+        if state_dict['{}extra_embedder.0.weight'.format(key_prefix)].shape[1] == 3968:
+            unet_config["size_cond"] = True
+            unet_config["use_style_cond"] = True
+            unet_config["image_model"] = "hydit1"
+        return unet_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 

comfy/sd.py

@@ -22,6 +22,7 @@ from . import sdxl_clip
 import comfy.text_encoders.sd3_clip
 import comfy.text_encoders.sa_t5
 import comfy.text_encoders.aura_t5
+import comfy.text_encoders.hydit
 
 import comfy.model_patcher
 import comfy.lora
@@ -385,6 +386,7 @@ class CLIPType(Enum):
     STABLE_CASCADE = 2
     SD3 = 3
     STABLE_AUDIO = 4
+    HUNYUAN_DIT = 5
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION):
     clip_data = []
@@ -433,6 +435,9 @@ def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DI
         if clip_type == CLIPType.SD3:
             clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=True, clip_g=True, t5=False)
             clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
+        elif clip_type == CLIPType.HUNYUAN_DIT:
+            clip_target.clip = comfy.text_encoders.hydit.HyditModel
+            clip_target.tokenizer = comfy.text_encoders.hydit.HyditTokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer

comfy/supported_models.py

@@ -8,6 +8,7 @@ from . import sdxl_clip
 import comfy.text_encoders.sd3_clip
 import comfy.text_encoders.sa_t5
 import comfy.text_encoders.aura_t5
+import comfy.text_encoders.hydit
 
 from . import supported_models_base
 from . import latent_formats
@@ -580,6 +581,45 @@ class AuraFlow(supported_models_base.BASE):
     def clip_target(self, state_dict={}):
         return supported_models_base.ClipTarget(comfy.text_encoders.aura_t5.AuraT5Tokenizer, comfy.text_encoders.aura_t5.AuraT5Model)
 
-models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow]
+class HunyuanDiT(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "hydit",
+    }
+
+    unet_extra_config = {
+        "attn_precision": torch.float32,
+    }
+
+    sampling_settings = {
+        "linear_start": 0.00085,
+        "linear_end": 0.018,
+    }
+
+    latent_format = latent_formats.SDXL
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.HunyuanDiT(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return supported_models_base.ClipTarget(comfy.text_encoders.hydit.HyditTokenizer, comfy.text_encoders.hydit.HyditModel)
+
+class HunyuanDiT1(HunyuanDiT):
+    unet_config = {
+        "image_model": "hydit1",
+    }
+
+    unet_extra_config = {}
+
+    sampling_settings = {
+        "linear_start" : 0.00085,
+        "linear_end" : 0.03,
+    }
+
+
+models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, HunyuanDiT, HunyuanDiT1]
 
 models += [SVD_img2vid]

comfy/text_encoders/hydit.py

+from comfy import sd1_clip
+from transformers import T5TokenizerFast, BertTokenizer, BertModel, modeling_utils, BertConfig
+from .spiece_tokenizer import SPieceTokenizer
+import comfy.text_encoders.t5
+import os
+
+import torch
+import contextlib
+
+@contextlib.contextmanager
+def use_comfy_ops(ops, device=None, dtype=None):
+    old_torch_nn_linear = torch.nn.Linear
+    force_device = device
+    force_dtype = dtype
+    def linear_with_dtype(in_features: int, out_features: int, bias: bool = True, device=None, dtype=None):
+        if force_device is not None:
+            device = force_device
+        if force_dtype is not None:
+            dtype = force_dtype
+        return ops.Linear(in_features, out_features, bias=bias, device=device, dtype=dtype)
+
+    torch.nn.Linear = linear_with_dtype
+    try:
+        yield
+    finally:
+        torch.nn.Linear = old_torch_nn_linear
+
+
+class RobertaWrapper(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = BertConfig(**config_dict)
+        with use_comfy_ops(operations, device, dtype):
+            with modeling_utils.no_init_weights():
+                self.bert = BertModel(config, add_pooling_layer=False)
+
+        self.num_layers = config.num_hidden_layers
+
+    def get_input_embeddings(self):
+        return self.bert.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        return self.bert.set_input_embeddings(value)
+
+    def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True):
+        intermediate = None
+        out = self.bert(input_ids=input_tokens, output_hidden_states=intermediate_output is not None, attention_mask=attention_mask)
+        return out.last_hidden_state, intermediate, out.pooler_output
+
+class HyditBertModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "hydit_clip.json")
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 101, "end": 102, "pad": 0}, model_class=RobertaWrapper, enable_attention_masks=True, return_attention_masks=True)
+
+class HyditBertTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "hydit_clip_tokenizer")
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=1024, embedding_key='chinese_roberta', tokenizer_class=BertTokenizer)
+
+
+class MT5XLModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None):
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "mt5_config_xl.json")
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, return_attention_masks=True)
+
+class MT5XLTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        #tokenizer_path = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "mt5_tokenizer"), "spiece.model")
+        tokenizer = tokenizer_data.get("spiece_model", None)
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=2048, embedding_key='mt5xl', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256)
+
+    def state_dict(self):
+        return {"spiece_model": self.tokenizer.serialize_model()}
+
+class HyditTokenizer:
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        mt5_tokenizer_data = tokenizer_data.get("mt5xl.spiece_model", None)
+        self.hydit_clip = HyditBertTokenizer(embedding_directory=embedding_directory)
+        self.mt5xl = MT5XLTokenizer(tokenizer_data={"spiece_model": mt5_tokenizer_data}, embedding_directory=embedding_directory)
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False):
+        out = {}
+        out["hydit_clip"] = self.hydit_clip.tokenize_with_weights(text, return_word_ids)
+        out["mt5xl"] = self.mt5xl.tokenize_with_weights(text, return_word_ids)
+        return out
+
+    def untokenize(self, token_weight_pair):
+        return self.hydit_clip.untokenize(token_weight_pair)
+
+    def state_dict(self):
+        return {"mt5xl.spiece_model": self.mt5xl.state_dict()["spiece_model"]}
+
+class HyditModel(torch.nn.Module):
+    def __init__(self, device="cpu", dtype=None):
+        super().__init__()
+        self.hydit_clip = HyditBertModel()
+        self.mt5xl = MT5XLModel()
+
+        self.dtypes = set()
+        if dtype is not None:
+            self.dtypes.add(dtype)
+
+    def encode_token_weights(self, token_weight_pairs):
+        hydit_out = self.hydit_clip.encode_token_weights(token_weight_pairs["hydit_clip"])
+        mt5_out = self.mt5xl.encode_token_weights(token_weight_pairs["mt5xl"])
+        return hydit_out[0], hydit_out[1], {"attention_mask": hydit_out[2]["attention_mask"], "conditioning_mt5xl": mt5_out[0], "attention_mask_mt5xl": mt5_out[2]["attention_mask"]}
+
+    def load_sd(self, sd):
+        if "bert.encoder.layer.0.attention.self.query.weight" in sd:
+            return self.hydit_clip.load_sd(sd)
+        else:
+            return self.mt5xl.load_sd(sd)
+
+    def set_clip_options(self, options):
+        self.hydit_clip.set_clip_options(options)
+        self.mt5xl.set_clip_options(options)
+
+    def reset_clip_options(self):
+        self.hydit_clip.reset_clip_options()
+        self.mt5xl.reset_clip_options()

comfy/text_encoders/hydit_clip.json

+{
+  "_name_or_path": "hfl/chinese-roberta-wwm-ext-large",
+  "architectures": [
+    "BertModel"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "directionality": "bidi",
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 24,
+  "output_past": true,
+  "pad_token_id": 0,
+  "pooler_fc_size": 768,
+  "pooler_num_attention_heads": 12,
+  "pooler_num_fc_layers": 3,
+  "pooler_size_per_head": 128,
+  "pooler_type": "first_token_transform",
+  "position_embedding_type": "absolute",
+  "torch_dtype": "float32",
+  "transformers_version": "4.22.1",
+  "type_vocab_size": 2,
+  "use_cache": true,
+  "vocab_size": 47020
+}
+

comfy/text_encoders/hydit_clip_tokenizer/special_tokens_map.json

+{
+  "cls_token": "[CLS]",
+  "mask_token": "[MASK]",
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "unk_token": "[UNK]"
+}

comfy/text_encoders/hydit_clip_tokenizer/tokenizer_config.json

+{
+  "cls_token": "[CLS]",
+  "do_basic_tokenize": true,
+  "do_lower_case": true,
+  "mask_token": "[MASK]",
+  "name_or_path": "hfl/chinese-roberta-wwm-ext",
+  "never_split": null,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "special_tokens_map_file": "/home/chenweifeng/.cache/huggingface/hub/models--hfl--chinese-roberta-wwm-ext/snapshots/5c58d0b8ec1d9014354d691c538661bf00bfdb44/special_tokens_map.json",
+  "strip_accents": null,
+  "tokenize_chinese_chars": true,
+  "tokenizer_class": "BertTokenizer",
+  "unk_token": "[UNK]",
+  "model_max_length": 77
+}

comfy/text_encoders/mt5_config_xl.json

+{
+  "d_ff": 5120,
+  "d_kv": 64,
+  "d_model": 2048,
+  "decoder_start_token_id": 0,
+  "dropout_rate": 0.1,
+  "eos_token_id": 1,
+  "dense_act_fn": "gelu_pytorch_tanh",
+  "initializer_factor": 1.0,
+  "is_encoder_decoder": true,
+  "is_gated_act": true,
+  "layer_norm_epsilon": 1e-06,
+  "model_type": "mt5",
+  "num_decoder_layers": 24,
+  "num_heads": 32,
+  "num_layers": 24,
+  "output_past": true,
+  "pad_token_id": 0,
+  "relative_attention_num_buckets": 32,
+  "tie_word_embeddings": false,
+  "vocab_size": 250112
+}

comfy/text_encoders/spiece_tokenizer.py

@@ -27,3 +27,6 @@ class SPieceTokenizer:
     def __call__(self, string):
         out = self.tokenizer.encode(string)
         return {"input_ids": out}
+
+    def serialize_model(self):
+        return torch.ByteTensor(list(self.tokenizer.serialized_model_proto()))