# Copyright 2024 NVIDIA CORPORATION & AFFILIATES # # 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. # # SPDX-License-Identifier: Apache-2.0 from copy import deepcopy from typing import Optional, Tuple # This file is modified from https://github.com/PixArt-alpha/PixArt-sigma import torch import torch.nn as nn from timm.models.layers import DropPath from torch.nn import Linear, Module, init from diffusion.model.builder import MODELS from diffusion.model.nets.basic_modules import DWMlp, GLUMBConv, MBConvPreGLU, Mlp from diffusion.model.nets.fastlinear.modules import TritonLiteMLA, TritonMBConvPreGLU from diffusion.model.nets.sana import Sana, get_2d_sincos_pos_embed from diffusion.model.nets.sana_blocks import ( Attention, CaptionEmbedder, FlashAttention, LiteLA, MultiHeadCrossAttention, PatchEmbedMS, T2IFinalLayer, t2i_modulate, ) from diffusion.model.nets.sana_multi_scale import SanaMS, SanaMSBlock from diffusion.model.utils import auto_grad_checkpoint, to_2tuple ############################################################################# # Core Sana Model # ################################################################################# class ControlSanaMSBlock(Module): def __init__(self, base_block: SanaMSBlock, block_index: int) -> None: super().__init__() self.copied_block = deepcopy(base_block) self.block_index = block_index self.hidden_size = hidden_size = base_block.hidden_size if self.block_index == 0: self.before_proj = Linear(hidden_size, hidden_size) self.after_proj = Linear(hidden_size, hidden_size) def initialize_all_and_copy_from_base(self, base_block): for name, param in self.named_parameters(): param.requires_grad_(True) self.copied_block.load_state_dict(base_block.state_dict()) self.train() if self.block_index == 0: init.zeros_(self.before_proj.weight) init.zeros_(self.before_proj.bias) init.zeros_(self.after_proj.weight) init.zeros_(self.after_proj.bias) def forward(self, x, y, t, control_signal, mask=None, HW=None): if self.block_index == 0: # the first block control_signal = self.before_proj(control_signal) control_signal = self.copied_block(x + control_signal, y, t, mask, HW) control_signal_skip = self.after_proj(control_signal) else: # load from previous control_signal and produce the control_signal for skip connection control_signal = self.copied_block(control_signal, y, t, mask, HW) control_signal_skip = self.after_proj(control_signal) return control_signal, control_signal_skip @MODELS.register_module() class SanaMSControlNet(SanaMS): """ Diffusion model with a Transformer backbone. """ def __init__( self, input_size=32, patch_size=2, in_channels=4, hidden_size=1152, depth=28, num_heads=16, mlp_ratio=4.0, class_dropout_prob=0.1, learn_sigma=True, pred_sigma=True, drop_path: float = 0.0, caption_channels=4096, pe_interpolation=1.0, config=None, model_max_length=300, qk_norm=False, y_norm=False, norm_eps=1e-5, attn_type="flash", ffn_type="mlp", use_pe=True, y_norm_scale_factor=1.0, patch_embed_kernel=None, mlp_acts=("silu", "gelu", None), linear_head_dim=32, copy_blocks_num=7, **kwargs, ): super().__init__( input_size=input_size, patch_size=patch_size, in_channels=in_channels, hidden_size=hidden_size, depth=depth, num_heads=num_heads, mlp_ratio=mlp_ratio, class_dropout_prob=class_dropout_prob, learn_sigma=learn_sigma, pred_sigma=pred_sigma, drop_path=drop_path, caption_channels=caption_channels, pe_interpolation=pe_interpolation, config=config, model_max_length=model_max_length, qk_norm=qk_norm, y_norm=y_norm, norm_eps=norm_eps, attn_type=attn_type, ffn_type=ffn_type, use_pe=use_pe, y_norm_scale_factor=y_norm_scale_factor, patch_embed_kernel=patch_embed_kernel, mlp_acts=mlp_acts, linear_head_dim=linear_head_dim, **kwargs, ) self.h = self.w = 0 approx_gelu = lambda: nn.GELU(approximate="tanh") self.t_block = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True)) self.pos_embed_ms = None kernel_size = patch_embed_kernel or patch_size self.x_embedder = PatchEmbedMS(patch_size, in_channels, hidden_size, kernel_size=kernel_size, bias=True) self.y_embedder = CaptionEmbedder( in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob, act_layer=approx_gelu, token_num=model_max_length, ) drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)] # stochastic depth decay rule self.blocks = nn.ModuleList( [ SanaMSBlock( hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i], input_size=(input_size // patch_size, input_size // patch_size), qk_norm=qk_norm, attn_type=attn_type, ffn_type=ffn_type, mlp_acts=mlp_acts, linear_head_dim=linear_head_dim, ) for i in range(depth) ] ) self.final_layer = T2IFinalLayer(hidden_size, patch_size, self.out_channels) # define controlnet self.copy_blocks_num = copy_blocks_num self.controlnet = nn.ModuleList([ControlSanaMSBlock(self.blocks[i], i) for i in range(copy_blocks_num)]) def load_pretrain_and_initialize(self, model_path): missing, unexpected = self.load_state_dict(torch.load(model_path)["state_dict"], strict=False) self.initialize_all() return missing, unexpected def initialize_all(self): # freeze all the parameters for p in self.parameters(): p.requires_grad_(False) # self.eval() for i, block in enumerate(self.controlnet): block.initialize_all_and_copy_from_base(self.blocks[i]) def forward_controlnet(self, control_signal): if self.use_pe: control_signal = self.x_embedder(control_signal) + self.pos_embed_ms else: control_signal = self.x_embedder(control_signal) return control_signal def forward(self, x, timestep, control_signal, y, mask=None, data_info=None, **kwargs): """ Forward pass of Sana. x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images) t: (N,) tensor of diffusion timesteps y: (N, 1, 120, C) tensor of class labels """ bs = x.shape[0] x = x.to(self.dtype) timestep = timestep.to(self.dtype) y = y.to(self.dtype) self.h, self.w = x.shape[-2] // self.patch_size, x.shape[-1] // self.patch_size if self.use_pe: x = self.x_embedder(x) if self.pos_embed_ms is None or self.pos_embed_ms.shape[1:] != x.shape[1:]: self.pos_embed_ms = ( torch.from_numpy( get_2d_sincos_pos_embed( self.pos_embed.shape[-1], (self.h, self.w), pe_interpolation=self.pe_interpolation, base_size=self.base_size, ) ) .unsqueeze(0) .to(x.device) .to(self.dtype) ) x += self.pos_embed_ms # (N, T, D), where T = H * W / patch_size ** 2 else: x = self.x_embedder(x) # control signal branch control_signal = control_signal.to(self.dtype) control_signal = self.forward_controlnet(control_signal) t = self.t_embedder(timestep) # (N, D) t0 = self.t_block(t) y = self.y_embedder(y, self.training, mask=mask) # (N, D) if self.y_norm: y = self.attention_y_norm(y) if mask is not None: if mask.shape[0] != y.shape[0]: mask = mask.repeat(y.shape[0] // mask.shape[0], 1) mask = mask.squeeze(1).squeeze(1) y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1]) y_lens = mask.sum(dim=1).tolist() else: y_lens = [y.shape[2]] * y.shape[0] y = y.squeeze(1).view(1, -1, x.shape[-1]) x = auto_grad_checkpoint(self.blocks[0], x, y, t0, y_lens, (self.h, self.w), **kwargs) for i in range(1, self.copy_blocks_num + 1): control_signal, control_signal_skip = auto_grad_checkpoint( self.controlnet[i - 1], x, y, t0, control_signal, y_lens, (self.h, self.w), **kwargs ) x = auto_grad_checkpoint(self.blocks[i], x + control_signal_skip, y, t0, y_lens, (self.h, self.w), **kwargs) for i in range(self.copy_blocks_num + 1, len(self.blocks)): x = auto_grad_checkpoint(self.blocks[i], x, y, t0, y_lens, (self.h, self.w), **kwargs) x = self.final_layer(x, t) # (N, T, patch_size ** 2 * out_channels) x = self.unpatchify(x) # (N, out_channels, H, W) return x def __call__(self, *args, **kwargs): """ This method allows the object to be called like a function. It simply calls the forward method. """ return self.forward(*args, **kwargs) def forward_with_dpmsolver(self, x, timestep, y, data_info, **kwargs): """ dpm solver donnot need variance prediction """ control_signal = kwargs.pop("control_signal", None) assert control_signal is not None, "control_signal is required for dpm solver" assert control_signal.dim() == 4, "control_signal should be a 4D tensor" if x.shape[0] != control_signal.shape[0]: control_signal = control_signal.repeat(x.shape[0] // control_signal.shape[0], 1, 1, 1) assert control_signal.shape[0] == x.shape[0], "control_signal and x should have the same batch size" # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb model_out = self.forward(x, timestep, control_signal, y, data_info=data_info, **kwargs) return model_out.chunk(2, dim=1)[0] if self.pred_sigma else model_out ################################################################################# # Sana Multi-scale Configs # ################################################################################# @MODELS.register_module() def SanaMSControlNet_600M_P1_D28(**kwargs): return SanaMSControlNet(depth=28, hidden_size=1152, patch_size=1, num_heads=16, **kwargs) @MODELS.register_module() def SanaMSControlNet_1600M_P1_D20(**kwargs): return SanaMSControlNet(depth=20, hidden_size=2240, patch_size=1, num_heads=20, **kwargs)