import numpy as np import torch import torch.distributed as dist import torch.nn as nn from einops import rearrange from .configuration_stdit2 import STDiT2Config from .layers import ( STDiT2Block, CaptionEmbedder, PatchEmbed3D, T2IFinalLayer, TimestepEmbedder, SizeEmbedder, PositionEmbedding2D ) from rotary_embedding_torch import RotaryEmbedding from .utils import ( get_2d_sincos_pos_embed, approx_gelu ) from transformers import PreTrainedModel class STDiT2(PreTrainedModel): config_class = STDiT2Config def __init__( self, config: STDiT2Config ): super().__init__(config) self.pred_sigma = config.pred_sigma self.in_channels = config.in_channels self.out_channels = config.in_channels * 2 if config.pred_sigma else config.in_channels self.hidden_size = config.hidden_size self.num_heads = config.num_heads self.no_temporal_pos_emb = config.no_temporal_pos_emb self.depth = config.depth self.mlp_ratio = config.mlp_ratio self.enable_flash_attn = config.enable_flash_attn self.enable_layernorm_kernel = config.enable_layernorm_kernel self.enable_sequence_parallelism = config.enable_sequence_parallelism # support dynamic input self.patch_size = config.patch_size self.input_size = config.input_size self.input_sq_size = config.input_sq_size self.pos_embed = PositionEmbedding2D(config.hidden_size) self.x_embedder = PatchEmbed3D(config.patch_size, config.in_channels, config.hidden_size) self.t_embedder = TimestepEmbedder(config.hidden_size) self.t_block = nn.Sequential(nn.SiLU(), nn.Linear(config.hidden_size, 6 * config.hidden_size, bias=True)) self.t_block_temp = nn.Sequential(nn.SiLU(), nn.Linear(config.hidden_size, 3 * config.hidden_size, bias=True)) # new self.y_embedder = CaptionEmbedder( in_channels=config.caption_channels, hidden_size=config.hidden_size, uncond_prob=config.class_dropout_prob, act_layer=approx_gelu, token_num=config.model_max_length, ) drop_path = [x.item() for x in torch.linspace(0, config.drop_path, config.depth)] self.rope = RotaryEmbedding(dim=self.hidden_size // self.num_heads) # new self.blocks = nn.ModuleList( [ STDiT2Block( self.hidden_size, self.num_heads, mlp_ratio=self.mlp_ratio, drop_path=drop_path[i], enable_flash_attn=self.enable_flash_attn, enable_layernorm_kernel=self.enable_layernorm_kernel, enable_sequence_parallelism=self.enable_sequence_parallelism, rope=self.rope.rotate_queries_or_keys, qk_norm=config.qk_norm, ) for i in range(self.depth) ] ) self.final_layer = T2IFinalLayer(config.hidden_size, np.prod(self.patch_size), self.out_channels) # multi_res assert self.hidden_size % 3 == 0, "hidden_size must be divisible by 3" self.csize_embedder = SizeEmbedder(self.hidden_size // 3) self.ar_embedder = SizeEmbedder(self.hidden_size // 3) self.fl_embedder = SizeEmbedder(self.hidden_size) # new self.fps_embedder = SizeEmbedder(self.hidden_size) # new # init model self.initialize_weights() self.initialize_temporal() if config.freeze is not None: assert config.freeze in ["not_temporal", "text"] if config.freeze == "not_temporal": self.freeze_not_temporal() elif config.freeze == "text": self.freeze_text() # sequence parallel related configs if self.enable_sequence_parallelism: self.sp_rank = dist.get_rank(get_sequence_parallel_group()) else: self.sp_rank = None def get_dynamic_size(self, x): _, _, T, H, W = x.size() if T % self.patch_size[0] != 0: T += self.patch_size[0] - T % self.patch_size[0] if H % self.patch_size[1] != 0: H += self.patch_size[1] - H % self.patch_size[1] if W % self.patch_size[2] != 0: W += self.patch_size[2] - W % self.patch_size[2] T = T // self.patch_size[0] H = H // self.patch_size[1] W = W // self.patch_size[2] return (T, H, W) def forward( self, x, timestep, y, mask=None, x_mask=None, num_frames=None, height=None, width=None, ar=None, fps=None ): """ Forward pass of STDiT. Args: x (torch.Tensor): latent representation of video; of shape [B, C, T, H, W] timestep (torch.Tensor): diffusion time steps; of shape [B] y (torch.Tensor): representation of prompts; of shape [B, 1, N_token, C] mask (torch.Tensor): mask for selecting prompt tokens; of shape [B, N_token] Returns: x (torch.Tensor): output latent representation; of shape [B, C, T, H, W] """ B = x.shape[0] x = x.to(self.final_layer.linear.weight.dtype) timestep = timestep.to(self.final_layer.linear.weight.dtype) y = y.to(self.final_layer.linear.weight.dtype) # === process data info === # 1. get dynamic size hw = torch.cat([height[:, None], width[:, None]], dim=1) rs = (height[0].item() * width[0].item()) ** 0.5 csize = self.csize_embedder(hw, B) # 2. get aspect ratio ar = ar.unsqueeze(1) ar = self.ar_embedder(ar, B) data_info = torch.cat([csize, ar], dim=1) # 3. get number of frames fl = num_frames.unsqueeze(1) fps = fps.unsqueeze(1) fl = self.fl_embedder(fl, B) fl = fl + self.fps_embedder(fps, B) # === get dynamic shape size === _, _, Tx, Hx, Wx = x.size() T, H, W = self.get_dynamic_size(x) S = H * W scale = rs / self.input_sq_size base_size = round(S**0.5) pos_emb = self.pos_embed(x, H, W, scale=scale, base_size=base_size) # embedding x = self.x_embedder(x) # [B, N, C] x = rearrange(x, "B (T S) C -> B T S C", T=T, S=S) x = x + pos_emb x = rearrange(x, "B T S C -> B (T S) C") # shard over the sequence dim if sp is enabled if self.enable_sequence_parallelism: x = split_forward_gather_backward(x, get_sequence_parallel_group(), dim=1, grad_scale="down") # prepare adaIN t = self.t_embedder(timestep, dtype=x.dtype) # [B, C] t_spc = t + data_info # [B, C] t_tmp = t + fl # [B, C] t_spc_mlp = self.t_block(t_spc) # [B, 6*C] t_tmp_mlp = self.t_block_temp(t_tmp) # [B, 3*C] if x_mask is not None: t0_timestep = torch.zeros_like(timestep) t0 = self.t_embedder(t0_timestep, dtype=x.dtype) t0_spc = t0 + data_info t0_tmp = t0 + fl t0_spc_mlp = self.t_block(t0_spc) t0_tmp_mlp = self.t_block_temp(t0_tmp) else: t0_spc = None t0_tmp = None t0_spc_mlp = None t0_tmp_mlp = None # prepare y y = self.y_embedder(y, self.training) # [B, 1, N_token, C] 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]) # blocks for _, block in enumerate(self.blocks): x = block( x, y, t_spc_mlp, t_tmp_mlp, y_lens, x_mask, t0_spc_mlp, t0_tmp_mlp, T, S, ) if self.enable_sequence_parallelism: x = gather_forward_split_backward(x, get_sequence_parallel_group(), dim=1, grad_scale="up") # x.shape: [B, N, C] # final process x = self.final_layer(x, t, x_mask, t0_spc, T, S) # [B, N, C=T_p * H_p * W_p * C_out] x = self.unpatchify(x, T, H, W, Tx, Hx, Wx) # [B, C_out, T, H, W] # cast to float32 for better accuracy x = x.to(torch.float32) return x def unpatchify(self, x, N_t, N_h, N_w, R_t, R_h, R_w): """ Args: x (torch.Tensor): of shape [B, N, C] Return: x (torch.Tensor): of shape [B, C_out, T, H, W] """ # N_t, N_h, N_w = [self.input_size[i] // self.patch_size[i] for i in range(3)] T_p, H_p, W_p = self.patch_size x = rearrange( x, "B (N_t N_h N_w) (T_p H_p W_p C_out) -> B C_out (N_t T_p) (N_h H_p) (N_w W_p)", N_t=N_t, N_h=N_h, N_w=N_w, T_p=T_p, H_p=H_p, W_p=W_p, C_out=self.out_channels, ) # unpad x = x[:, :, :R_t, :R_h, :R_w] return x def unpatchify_old(self, x): c = self.out_channels t, h, w = [self.input_size[i] // self.patch_size[i] for i in range(3)] pt, ph, pw = self.patch_size x = x.reshape(shape=(x.shape[0], t, h, w, pt, ph, pw, c)) x = rearrange(x, "n t h w r p q c -> n c t r h p w q") imgs = x.reshape(shape=(x.shape[0], c, t * pt, h * ph, w * pw)) return imgs def get_spatial_pos_embed(self, H, W, scale=1.0, base_size=None): pos_embed = get_2d_sincos_pos_embed( self.hidden_size, (H, W), scale=scale, base_size=base_size, ) pos_embed = torch.from_numpy(pos_embed).float().unsqueeze(0).requires_grad_(False) return pos_embed def freeze_not_temporal(self): for n, p in self.named_parameters(): if "attn_temp" not in n: p.requires_grad = False def freeze_text(self): for n, p in self.named_parameters(): if "cross_attn" in n: p.requires_grad = False def initialize_temporal(self): for block in self.blocks: nn.init.constant_(block.attn_temp.proj.weight, 0) nn.init.constant_(block.attn_temp.proj.bias, 0) def initialize_weights(self): # Initialize transformer layers: def _basic_init(module): if isinstance(module, nn.Linear): torch.nn.init.xavier_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) self.apply(_basic_init) # Initialize patch_embed like nn.Linear (instead of nn.Conv2d): w = self.x_embedder.proj.weight.data nn.init.xavier_uniform_(w.view([w.shape[0], -1])) # Initialize timestep embedding MLP: nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02) nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02) nn.init.normal_(self.t_block[1].weight, std=0.02) nn.init.normal_(self.t_block_temp[1].weight, std=0.02) # Initialize caption embedding MLP: nn.init.normal_(self.y_embedder.y_proj.fc1.weight, std=0.02) nn.init.normal_(self.y_embedder.y_proj.fc2.weight, std=0.02) # Zero-out adaLN modulation layers in PixArt blocks: for block in self.blocks: nn.init.constant_(block.cross_attn.proj.weight, 0) nn.init.constant_(block.cross_attn.proj.bias, 0) # Zero-out output layers: nn.init.constant_(self.final_layer.linear.weight, 0) nn.init.constant_(self.final_layer.linear.bias, 0)