uso/flux/modules/layers.py

# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates. All rights reserved.
# Copyright (c) 2024 Black Forest Labs and The XLabs-AI Team. All rights reserved.

# 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 dataclasses import dataclass

import torch
from einops import rearrange, repeat
from torch import Tensor, nn

from ..math import attention, rope


class EmbedND(nn.Module):
    def __init__(self, dim: int, theta: int, axes_dim: list[int]):
        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)
        / half
    ).to(t.device)

    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):
        super().__init__()
        self.in_layer = nn.Linear(in_dim, hidden_dim, bias=True)
        self.silu = nn.SiLU()
        self.out_layer = nn.Linear(hidden_dim, hidden_dim, bias=True)

    def forward(self, x: Tensor) -> Tensor:
        return self.out_layer(self.silu(self.in_layer(x)))


class RMSNorm(torch.nn.Module):
    def __init__(self, dim: int):
        super().__init__()
        self.scale = nn.Parameter(torch.ones(dim))

    def forward(self, x: Tensor):
        x_dtype = x.dtype
        x = x.float()
        rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + 1e-6)
        return ((x * rrms) * self.scale.float()).to(dtype=x_dtype)


class QKNorm(torch.nn.Module):
    def __init__(self, dim: int):
        super().__init__()
        self.query_norm = RMSNorm(dim)
        self.key_norm = RMSNorm(dim)

    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple[Tensor, Tensor]:
        q = self.query_norm(q)
        k = self.key_norm(k)
        return q.to(v), k.to(v)


class LoRALinearLayer(nn.Module):
    def __init__(
        self,
        in_features,
        out_features,
        rank=4,
        network_alpha=None,
        device=None,
        dtype=None,
    ):
        super().__init__()

        self.down = nn.Linear(in_features, rank, bias=False, device=device, dtype=dtype)
        self.up = nn.Linear(rank, out_features, bias=False, device=device, dtype=dtype)
        # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
        # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
        self.network_alpha = network_alpha
        self.rank = rank

        nn.init.normal_(self.down.weight, std=1 / rank)
        nn.init.zeros_(self.up.weight)

    def forward(self, hidden_states):
        orig_dtype = hidden_states.dtype
        dtype = self.down.weight.dtype

        down_hidden_states = self.down(hidden_states.to(dtype))
        up_hidden_states = self.up(down_hidden_states)

        if self.network_alpha is not None:
            up_hidden_states *= self.network_alpha / self.rank

        return up_hidden_states.to(orig_dtype)


class FLuxSelfAttnProcessor:
    def __call__(self, attn, x, pe, **attention_kwargs):
        qkv = attn.qkv(x)
        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
        q, k = attn.norm(q, k, v)
        x = attention(q, k, v, pe=pe)
        x = attn.proj(x)
        return x


class LoraFluxAttnProcessor(nn.Module):

    def __init__(self, dim: int, rank=4, network_alpha=None, lora_weight=1):
        super().__init__()
        self.qkv_lora = LoRALinearLayer(dim, dim * 3, rank, network_alpha)
        self.proj_lora = LoRALinearLayer(dim, dim, rank, network_alpha)
        self.lora_weight = lora_weight

    def __call__(self, attn, x, pe, **attention_kwargs):
        qkv = attn.qkv(x) + self.qkv_lora(x) * self.lora_weight
        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
        q, k = attn.norm(q, k, v)
        x = attention(q, k, v, pe=pe)
        x = attn.proj(x) + self.proj_lora(x) * self.lora_weight
        return x


class SelfAttention(nn.Module):
    def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.norm = QKNorm(head_dim)
        self.proj = nn.Linear(dim, dim)

    def forward():
        pass


@dataclass
class ModulationOut:
    shift: Tensor
    scale: Tensor
    gate: Tensor


class Modulation(nn.Module):
    def __init__(self, dim: int, double: bool):
        super().__init__()
        self.is_double = double
        self.multiplier = 6 if double else 3
        self.lin = nn.Linear(dim, self.multiplier * dim, bias=True)

    def forward(self, vec: Tensor) -> tuple[ModulationOut, ModulationOut | None]:
        out = self.lin(nn.functional.silu(vec))[:, None, :].chunk(
            self.multiplier, dim=-1
        )

        return (
            ModulationOut(*out[:3]),
            ModulationOut(*out[3:]) if self.is_double else None,
        )


class DoubleStreamBlockLoraProcessor(nn.Module):
    def __init__(self, dim: int, rank=4, network_alpha=None, lora_weight=1):
        super().__init__()
        self.qkv_lora1 = LoRALinearLayer(dim, dim * 3, rank, network_alpha)
        self.proj_lora1 = LoRALinearLayer(dim, dim, rank, network_alpha)
        self.qkv_lora2 = LoRALinearLayer(dim, dim * 3, rank, network_alpha)
        self.proj_lora2 = LoRALinearLayer(dim, dim, rank, network_alpha)
        self.lora_weight = lora_weight

    def forward(self, attn, img, txt, vec, pe, **attention_kwargs):
        img_mod1, img_mod2 = attn.img_mod(vec)
        txt_mod1, txt_mod2 = attn.txt_mod(vec)

        # prepare image for attention
        img_modulated = attn.img_norm1(img)
        img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
        img_qkv = (
            attn.img_attn.qkv(img_modulated)
            + self.qkv_lora1(img_modulated) * self.lora_weight
        )
        img_q, img_k, img_v = rearrange(
            img_qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads
        )
        img_q, img_k = attn.img_attn.norm(img_q, img_k, img_v)

        # prepare txt for attention
        txt_modulated = attn.txt_norm1(txt)
        txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
        txt_qkv = (
            attn.txt_attn.qkv(txt_modulated)
            + self.qkv_lora2(txt_modulated) * self.lora_weight
        )
        txt_q, txt_k, txt_v = rearrange(
            txt_qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads
        )
        txt_q, txt_k = attn.txt_attn.norm(txt_q, txt_k, txt_v)

        # run actual attention
        q = torch.cat((txt_q, img_q), dim=2)
        k = torch.cat((txt_k, img_k), dim=2)
        v = torch.cat((txt_v, img_v), dim=2)

        attn1 = attention(q, k, v, pe=pe)
        txt_attn, img_attn = attn1[:, : txt.shape[1]], attn1[:, txt.shape[1] :]

        # calculate the img bloks
        img = img + img_mod1.gate * (
            attn.img_attn.proj(img_attn) + self.proj_lora1(img_attn) * self.lora_weight
        )
        img = img + img_mod2.gate * attn.img_mlp(
            (1 + img_mod2.scale) * attn.img_norm2(img) + img_mod2.shift
        )

        # calculate the txt bloks
        txt = txt + txt_mod1.gate * (
            attn.txt_attn.proj(txt_attn) + self.proj_lora2(txt_attn) * self.lora_weight
        )
        txt = txt + txt_mod2.gate * attn.txt_mlp(
            (1 + txt_mod2.scale) * attn.txt_norm2(txt) + txt_mod2.shift
        )
        return img, txt


class DoubleStreamBlockProcessor:
    def __call__(self, attn, img, txt, vec, pe, **attention_kwargs):
        img_mod1, img_mod2 = attn.img_mod(vec)
        txt_mod1, txt_mod2 = attn.txt_mod(vec)

        # prepare image for attention
        img_modulated = attn.img_norm1(img)
        img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
        img_qkv = attn.img_attn.qkv(img_modulated)
        img_q, img_k, img_v = rearrange(
            img_qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads, D=attn.head_dim
        )
        img_q, img_k = attn.img_attn.norm(img_q, img_k, img_v)

        # prepare txt for attention
        txt_modulated = attn.txt_norm1(txt)
        txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
        txt_qkv = attn.txt_attn.qkv(txt_modulated)
        txt_q, txt_k, txt_v = rearrange(
            txt_qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads, D=attn.head_dim
        )
        txt_q, txt_k = attn.txt_attn.norm(txt_q, txt_k, txt_v)

        # run actual attention
        q = torch.cat((txt_q, img_q), dim=2)
        k = torch.cat((txt_k, img_k), dim=2)
        v = torch.cat((txt_v, img_v), dim=2)

        attn1 = attention(q, k, v, pe=pe)
        txt_attn, img_attn = attn1[:, : txt.shape[1]], attn1[:, txt.shape[1] :]

        # calculate the img bloks
        img = img + img_mod1.gate * attn.img_attn.proj(img_attn)
        img = img + img_mod2.gate * attn.img_mlp(
            (1 + img_mod2.scale) * attn.img_norm2(img) + img_mod2.shift
        )

        # calculate the txt bloks
        txt = txt + txt_mod1.gate * attn.txt_attn.proj(txt_attn)
        txt = txt + txt_mod2.gate * attn.txt_mlp(
            (1 + txt_mod2.scale) * attn.txt_norm2(txt) + txt_mod2.shift
        )
        return img, txt


class DoubleStreamBlock(nn.Module):
    def __init__(
        self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False
    ):
        super().__init__()
        mlp_hidden_dim = int(hidden_size * mlp_ratio)
        self.num_heads = num_heads
        self.hidden_size = hidden_size
        self.head_dim = hidden_size // num_heads

        self.img_mod = Modulation(hidden_size, double=True)
        self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.img_attn = SelfAttention(
            dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias
        )

        self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.img_mlp = nn.Sequential(
            nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
            nn.GELU(approximate="tanh"),
            nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
        )

        self.txt_mod = Modulation(hidden_size, double=True)
        self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.txt_attn = SelfAttention(
            dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias
        )

        self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.txt_mlp = nn.Sequential(
            nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
            nn.GELU(approximate="tanh"),
            nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
        )
        processor = DoubleStreamBlockProcessor()
        self.set_processor(processor)

    def set_processor(self, processor) -> None:
        self.processor = processor

    def get_processor(self):
        return self.processor

    def forward(
        self,
        img: Tensor,
        txt: Tensor,
        vec: Tensor,
        pe: Tensor,
        image_proj: Tensor = None,
        ip_scale: float = 1.0,
    ) -> tuple[Tensor, Tensor]:
        if image_proj is None:
            return self.processor(self, img, txt, vec, pe)
        else:
            return self.processor(self, img, txt, vec, pe, image_proj, ip_scale)


class SingleStreamBlockLoraProcessor(nn.Module):
    def __init__(
        self, dim: int, rank: int = 4, network_alpha=None, lora_weight: float = 1
    ):
        super().__init__()
        self.qkv_lora = LoRALinearLayer(dim, dim * 3, rank, network_alpha)
        self.proj_lora = LoRALinearLayer(15360, dim, rank, network_alpha)
        self.lora_weight = lora_weight

    def forward(self, attn: nn.Module, x: Tensor, vec: Tensor, pe: Tensor) -> Tensor:

        mod, _ = attn.modulation(vec)
        x_mod = (1 + mod.scale) * attn.pre_norm(x) + mod.shift
        qkv, mlp = torch.split(
            attn.linear1(x_mod), [3 * attn.hidden_size, attn.mlp_hidden_dim], dim=-1
        )
        qkv = qkv + self.qkv_lora(x_mod) * self.lora_weight

        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads)
        q, k = attn.norm(q, k, v)

        # compute attention
        attn_1 = attention(q, k, v, pe=pe)

        # compute activation in mlp stream, cat again and run second linear layer
        output = attn.linear2(torch.cat((attn_1, attn.mlp_act(mlp)), 2))
        output = (
            output
            + self.proj_lora(torch.cat((attn_1, attn.mlp_act(mlp)), 2))
            * self.lora_weight
        )
        output = x + mod.gate * output
        return output


class SingleStreamBlockProcessor:
    def __call__(
        self, attn: nn.Module, x: Tensor, vec: Tensor, pe: Tensor, **attention_kwargs
    ) -> Tensor:

        mod, _ = attn.modulation(vec)
        x_mod = (1 + mod.scale) * attn.pre_norm(x) + mod.shift
        qkv, mlp = torch.split(
            attn.linear1(x_mod), [3 * attn.hidden_size, attn.mlp_hidden_dim], dim=-1
        )

        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads)
        q, k = attn.norm(q, k, v)

        # compute attention
        attn_1 = attention(q, k, v, pe=pe)

        # compute activation in mlp stream, cat again and run second linear layer
        output = attn.linear2(torch.cat((attn_1, attn.mlp_act(mlp)), 2))
        output = x + mod.gate * output
        return output


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 = None,
    ):
        super().__init__()
        self.hidden_dim = hidden_size
        self.num_heads = num_heads
        self.head_dim = hidden_size // num_heads
        self.scale = qk_scale or self.head_dim**-0.5

        self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
        # qkv and mlp_in
        self.linear1 = nn.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim)
        # proj and mlp_out
        self.linear2 = nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size)

        self.norm = QKNorm(self.head_dim)

        self.hidden_size = hidden_size
        self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)

        self.mlp_act = nn.GELU(approximate="tanh")
        self.modulation = Modulation(hidden_size, double=False)

        processor = SingleStreamBlockProcessor()
        self.set_processor(processor)

    def set_processor(self, processor) -> None:
        self.processor = processor

    def get_processor(self):
        return self.processor

    def forward(
        self,
        x: Tensor,
        vec: Tensor,
        pe: Tensor,
        image_proj: Tensor | None = None,
        ip_scale: float = 1.0,
    ) -> Tensor:
        if image_proj is None:
            return self.processor(self, x, vec, pe)
        else:
            return self.processor(self, x, vec, pe, image_proj, ip_scale)


class LastLayer(nn.Module):
    def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
        super().__init__()
        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
        self.linear = nn.Linear(
            hidden_size, patch_size * patch_size * out_channels, bias=True
        )
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True)
        )

    def forward(self, x: Tensor, vec: Tensor) -> Tensor:
        shift, scale = self.adaLN_modulation(vec).chunk(2, dim=1)
        x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
        x = self.linear(x)
        return x


class SigLIPMultiFeatProjModel(torch.nn.Module):
    """
    SigLIP Multi-Feature Projection Model for processing style features from different layers 
    and projecting them into a unified hidden space.
    
    Args:
        siglip_token_nums (int): Number of SigLIP tokens, default 257
        style_token_nums (int): Number of style tokens, default 256  
        siglip_token_dims (int): Dimension of SigLIP tokens, default 1536
        hidden_size (int): Hidden layer size, default 3072
        context_layer_norm (bool): Whether to use context layer normalization, default False
    """
    
    def __init__(
        self,
        siglip_token_nums: int = 257,
        style_token_nums: int = 256,
        siglip_token_dims: int = 1536,
        hidden_size: int = 3072,
        context_layer_norm: bool = False,
    ):
        super().__init__()
        
        # High-level feature processing (layer -2)
        self.high_embedding_linear = nn.Sequential(
            nn.Linear(siglip_token_nums, style_token_nums), 
            nn.SiLU()
        )
        self.high_layer_norm = (
            nn.LayerNorm(siglip_token_dims) if context_layer_norm else nn.Identity()
        )
        self.high_projection = nn.Linear(siglip_token_dims, hidden_size, bias=True)
        
        # Mid-level feature processing (layer -11)
        self.mid_embedding_linear = nn.Sequential(
            nn.Linear(siglip_token_nums, style_token_nums), 
            nn.SiLU()
        )
        self.mid_layer_norm = (
            nn.LayerNorm(siglip_token_dims) if context_layer_norm else nn.Identity()
        )
        self.mid_projection = nn.Linear(siglip_token_dims, hidden_size, bias=True)
        
        # Low-level feature processing (layer -20)
        self.low_embedding_linear = nn.Sequential(
            nn.Linear(siglip_token_nums, style_token_nums), 
            nn.SiLU()
        )
        self.low_layer_norm = (
            nn.LayerNorm(siglip_token_dims) if context_layer_norm else nn.Identity()
        )
        self.low_projection = nn.Linear(siglip_token_dims, hidden_size, bias=True)

    def forward(self, siglip_outputs):
        """
        Forward pass function
        
        Args:
            siglip_outputs: Output from SigLIP model, containing hidden_states
            
        Returns:
            torch.Tensor: Concatenated multi-layer features with shape [bs, 3*style_token_nums, hidden_size]
        """
        dtype = next(self.high_embedding_linear.parameters()).dtype
        
        # Process high-level features (layer -2)
        high_embedding = self._process_layer_features(
            siglip_outputs.hidden_states[-2],
            self.high_embedding_linear,
            self.high_layer_norm,
            self.high_projection,
            dtype
        )
        
        # Process mid-level features (layer -11)
        mid_embedding = self._process_layer_features(
            siglip_outputs.hidden_states[-11],
            self.mid_embedding_linear,
            self.mid_layer_norm,
            self.mid_projection,
            dtype
        )
        
        # Process low-level features (layer -20)
        low_embedding = self._process_layer_features(
            siglip_outputs.hidden_states[-20],
            self.low_embedding_linear,
            self.low_layer_norm,
            self.low_projection,
            dtype
        )
        
        # Concatenate features from all layers
        return torch.cat((high_embedding, mid_embedding, low_embedding), dim=1)
    
    def _process_layer_features(
        self, 
        hidden_states: torch.Tensor,
        embedding_linear: nn.Module,
        layer_norm: nn.Module,
        projection: nn.Module,
        dtype: torch.dtype
    ) -> torch.Tensor:
        """
        Helper function to process features from a single layer
        
        Args:
            hidden_states: Input hidden states [bs, seq_len, dim]
            embedding_linear: Embedding linear layer
            layer_norm: Layer normalization
            projection: Projection layer
            dtype: Target data type
            
        Returns:
            torch.Tensor: Processed features [bs, style_token_nums, hidden_size]
        """
        # Transform dimensions: [bs, seq_len, dim] -> [bs, dim, seq_len] -> [bs, dim, style_token_nums] -> [bs, style_token_nums, dim]
        embedding = embedding_linear(
            hidden_states.to(dtype).transpose(1, 2)
        ).transpose(1, 2)
        
        # Apply layer normalization
        embedding = layer_norm(embedding)
        
        # Project to target hidden space
        embedding = projection(embedding)
        
        return embedding