""" Lightning DiT's codes are built from original DiT & SiT. (https://github.com/facebookresearch/DiT; https://github.com/willisma/SiT) It demonstrates that a advanced DiT together with advanced diffusion skills could also achieve a very promising result with 1.35 FID on ImageNet 256 generation. Enjoy everyone, DiT strikes back! by Maple (Jingfeng Yao) from HUST-VL """ import os import math import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from torch.utils.checkpoint import checkpoint from timm.models.vision_transformer import PatchEmbed, Mlp from .swiglu_ffn import SwiGLUFFN from .pos_embed import VisionRotaryEmbeddingFast from .rmsnorm import RMSNorm class MultiHeadCrossAttention(nn.Module): def __init__(self, d_model, num_heads, attn_drop=0.0, proj_drop=0.0, qk_norm=False, fused_attn: bool = True, **block_kwargs): super().__init__() assert d_model % num_heads == 0, "d_model must be divisible by num_heads" self.d_model = d_model self.num_heads = num_heads self.head_dim = d_model // num_heads self.scale = self.head_dim ** -0.5 self.fused_attn = fused_attn self.q_linear = nn.Linear(d_model, d_model) self.k_linear = nn.Linear(d_model, d_model) self.v_linear = nn.Linear(d_model, d_model) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(d_model, d_model) self.proj_drop = nn.Dropout(proj_drop) if qk_norm: self.q_norm = RMSNorm(self.head_dim) self.k_norm = RMSNorm(self.head_dim) else: self.q_norm = nn.Identity() self.k_norm = nn.Identity() def forward(self, x, cond, mask=None): B, N, C = x.shape B_cond, N_cond, _ = cond.shape q = self.q_linear(x) k = self.k_linear(cond) v = self.v_linear(cond) q = q.view(B, N, self.num_heads, self.head_dim).permute(0, 2, 1, 3) k = k.view(B_cond, N_cond, self.num_heads, self.head_dim).permute(0, 2, 1, 3) v = v.view(B_cond, N_cond, self.num_heads, self.head_dim).permute(0, 2, 1, 3) q = self.q_norm(q) k = self.k_norm(k) if self.fused_attn: x = F.scaled_dot_product_attention( q, k, v, attn_mask=None, dropout_p=self.attn_drop.p if self.training else 0.0 ) else: q = q * self.scale attn = q @ k.transpose(-2, -1) attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = attn @ v x = x.permute(0, 2, 1, 3).contiguous().view(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x @torch.compile def modulate_adasin(x, shift, scale): if shift is None: return x * (1 + scale.unsqueeze(1)) return x * (1 + scale) + shift @torch.compile def modulate(x, shift, scale): if shift is None: return x * (1 + scale.unsqueeze(1)) return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1) class Attention(nn.Module): """ Attention module of LightningDiT. """ def __init__( self, dim: int, num_heads: int = 8, qkv_bias: bool = False, qk_norm: bool = False, attn_drop: float = 0., proj_drop: float = 0., norm_layer: nn.Module = nn.LayerNorm, fused_attn: bool = True, use_rmsnorm: bool = False, ) -> None: super().__init__() assert dim % num_heads == 0, 'dim should be divisible by num_heads' self.num_heads = num_heads self.head_dim = dim // num_heads self.scale = self.head_dim ** -0.5 self.fused_attn = fused_attn if use_rmsnorm: norm_layer = RMSNorm self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity() self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity() self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x: torch.Tensor, rope=None) -> torch.Tensor: B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4) q, k, v = qkv.unbind(0) q, k = self.q_norm(q), self.k_norm(k) if rope is not None: q = rope(q) k = rope(k) if self.fused_attn: x = F.scaled_dot_product_attention( q, k, v, dropout_p=self.attn_drop.p if self.training else 0., ) else: q = q * self.scale attn = q @ k.transpose(-2, -1) attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = attn @ v x = x.transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class TimestepEmbedder(nn.Module): """ Embeds scalar timesteps into vector representations. Same as DiT. """ def __init__(self, hidden_size: int, frequency_embedding_size: int = 256) -> None: super().__init__() self.frequency_embedding_size = frequency_embedding_size self.mlp = nn.Sequential( nn.Linear(frequency_embedding_size, hidden_size, bias=True), nn.SiLU(), nn.Linear(hidden_size, hidden_size, bias=True), ) @staticmethod def timestep_embedding(t: torch.Tensor, dim: int, max_period: int = 10000) -> torch.Tensor: """ Create sinusoidal timestep embeddings. Args: t: A 1-D Tensor of N indices, one per batch element. These may be fractional. dim: The dimension of the output. max_period: Controls the minimum frequency of the embeddings. Returns: An (N, D) Tensor of positional embeddings. """ # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py half = dim // 2 freqs = torch.exp( -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half ).to(device=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) return embedding # @torch.compile def forward(self, t: torch.Tensor) -> torch.Tensor: t_freq = self.timestep_embedding(t, self.frequency_embedding_size) t_emb = self.mlp(t_freq) return t_emb class LightningDiTBlock(nn.Module): """ Lightning DiT Block. We add features including: - ROPE - QKNorm - RMSNorm - SwiGLU - No shift AdaLN. Not all of them are used in the final model, please refer to the paper for more details. """ def __init__( self, hidden_size, num_heads, mlp_ratio=4.0, use_qknorm=False, use_swiglu=False, use_rmsnorm=False, wo_shift=False, z_dims=None, num_fused_layers=1, encdim_ratio=2, **block_kwargs ): super().__init__() # Initialize normalization layers if not use_rmsnorm: self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) else: self.norm1 = RMSNorm(hidden_size) self.norm2 = RMSNorm(hidden_size) # Initialize attention layer self.attn = Attention( hidden_size, num_heads=num_heads, qkv_bias=True, qk_norm=use_qknorm, use_rmsnorm=use_rmsnorm, **block_kwargs ) # Initialize MLP layer mlp_hidden_dim = int(hidden_size * mlp_ratio) approx_gelu = lambda: nn.GELU(approximate="tanh") if use_swiglu: # here we did not use SwiGLU from xformers because it is not compatible with torch.compile for now. self.mlp = SwiGLUFFN(hidden_size, int(2/3 * mlp_hidden_dim)) else: self.mlp = Mlp( in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0 ) self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size**0.5) self.z_dims = z_dims if self.z_dims is not None: # self.num_fused_layers = num_fused_layers # self.layer_weights = nn.Parameter(torch.zeros(num_fused_layers)) # approx_gelu = lambda: nn.GELU(approximate="tanh") # self.layer_norm = nn.LayerNorm(self.z_dims) # self.mlp_ca = Mlp( # in_features=self.z_dims, # hidden_features=hidden_size * encdim_ratio, # out_features=hidden_size, # act_layer=approx_gelu, # drop=0 # ) self.cross_attn = MultiHeadCrossAttention(d_model=hidden_size, num_heads=num_heads, qk_norm=use_qknorm) # self.cross_attn = MultiHeadCrossAttention(d_model=hidden_size, num_heads=num_heads, qk_norm=use_qknorm, **block_kwargs) @torch.compile def forward(self, x, c, z=None, feat_rope=None): B, N, C = x.shape shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( self.scale_shift_table[None] + c.reshape(B, 6, -1) ).chunk(6, dim=1) x = x + gate_msa * self.attn(modulate_adasin(self.norm1(x), shift_msa, scale_msa), rope=feat_rope) if self.z_dims is not None: x = x + self.cross_attn(x, z) x = x + gate_mlp * self.mlp(modulate_adasin(self.norm2(x), shift_mlp, scale_mlp)) return x class FinalLayer(nn.Module): """ The final layer of LightningDiT. """ def __init__(self, hidden_size, patch_size, out_channels, use_rmsnorm=False): super().__init__() if not use_rmsnorm: self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) else: self.norm_final = RMSNorm(hidden_size) 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) ) @torch.compile 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 LightningDiT(nn.Module): """ Diffusion model with a Transformer backbone. """ def __init__( self, input_size=32, patch_size=2, in_channels=32, out_channels=32, hidden_size=1152, depth=28, num_heads=16, mlp_ratio=4.0, learn_sigma=False, use_qknorm=False, use_swiglu=False, use_rope=False, use_rmsnorm=False, wo_shift=False, use_checkpoint=False, z_dims=None, encdim_ratio=2, num_fused_layers=1, auxiliary_time_cond=False, ): super().__init__() self.learn_sigma = learn_sigma self.in_channels = in_channels # lq + zt self.out_channels = out_channels self.patch_size = patch_size self.num_heads = num_heads self.use_rope = use_rope self.use_rmsnorm = use_rmsnorm self.depth = depth self.hidden_size = hidden_size self.use_checkpoint = use_checkpoint self.x_embedder = PatchEmbed(input_size, patch_size, self.in_channels, hidden_size, bias=True, strict_img_size=False) self.t_embedder = TimestepEmbedder(hidden_size) num_patches = self.x_embedder.num_patches # Will use fixed sin-cos embedding: # self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, hidden_size), requires_grad=False) # use rotary position encoding, borrow from EVA if self.use_rope: half_head_dim = hidden_size // num_heads // 2 hw_seq_len = input_size // patch_size self.feat_rope = VisionRotaryEmbeddingFast( dim=half_head_dim, pt_seq_len=hw_seq_len, ) else: self.feat_rope = None self.t_block = nn.Sequential( nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True) ) self.blocks = nn.ModuleList([ LightningDiTBlock(hidden_size, num_heads, mlp_ratio=mlp_ratio, use_qknorm=use_qknorm, use_swiglu=use_swiglu, use_rmsnorm=use_rmsnorm, wo_shift=wo_shift, z_dims=z_dims, num_fused_layers=num_fused_layers, encdim_ratio=encdim_ratio, ) for _ in range(depth) ]) self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels, use_rmsnorm=use_rmsnorm) self.z_dims = z_dims if self.z_dims is not None: self.num_fused_layers = num_fused_layers # self.layer_weights = nn.Parameter(torch.zeros(num_fused_layers)) approx_gelu = lambda: nn.GELU(approximate="tanh") self.layer_norm = nn.LayerNorm(self.z_dims) self.mlp_ca = Mlp( in_features=self.z_dims, hidden_features=hidden_size * encdim_ratio, out_features=hidden_size, act_layer=approx_gelu, drop=0 ) self.auxiliary_time_cond = auxiliary_time_cond if auxiliary_time_cond: self.r_embedder = TimestepEmbedder(hidden_size) else: self.r_embedder = None self.initialize_weights() 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 (and freeze) pos_embed by sin-cos embedding: # pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.x_embedder.num_patches ** 0.5)) # self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0)) # 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])) nn.init.constant_(self.x_embedder.proj.bias, 0) # Initialize label embedding table: # nn.init.normal_(self.y_embedder.embedding_table.weight, std=0.02) # 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) if self.auxiliary_time_cond: nn.init.normal_(self.r_embedder.mlp[0].weight, std=0.02) nn.init.normal_(self.r_embedder.mlp[2].weight, std=0.02) nn.init.normal_(self.t_block[1].weight, std=0.02) # Zero-out output layers: nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0) nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0) nn.init.constant_(self.final_layer.linear.weight, 0) nn.init.constant_(self.final_layer.linear.bias, 0) def unpatchify(self, x): """ x: (N, T, patch_size**2 * C) imgs: (N, H, W, C) """ c = self.out_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, h * p)) return imgs def enable_fused_attn(self): for block in self.blocks: block.attn.fused_attn = True block.cross_attn.fused_attn = True def disable_fused_attn(self): for block in self.blocks: block.attn.fused_attn = False block.cross_attn.fused_attn = False def forward(self, x, t, r=None, z=None): """ Forward pass of LightningDiT. x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images) t: (N,) tensor of diffusion timesteps r: (N,) tensor of auxiliary timesteps (shortcut endpoint) """ use_checkpoint = self.use_checkpoint x = self.x_embedder(x) t_raw = t # (N,) save raw timestep t = self.t_embedder(t) # (N, D) r = self.r_embedder(r)*(t_raw-r).unsqueeze(-1) if self.r_embedder is not None else 0 # (N, D) c = t + r # (N, D) c0 = self.t_block(c) if self.z_dims is not None: z = z[0] z = self.layer_norm(z) z = self.mlp_ca(z) for block in self.blocks: if use_checkpoint: x = checkpoint(block, x, c0, z, self.feat_rope, use_reentrant=True) else: x = block(x, c0, z, self.feat_rope) x = self.final_layer(x, c) # (N, T, patch_size ** 2 * out_channels) x = self.unpatchify(x) # (N, out_channels, H, W) return x def _get_dynamic_rope(self, hw_seq_len, device, dtype): """ Dynamically generate RoPE for variable-size inputs. Args: hw_seq_len: patch grid side length (H // patch_size) device: target device dtype: target dtype Returns: A temporary RoPE forward callable. """ if not self.use_rope: return None half_head_dim = self.hidden_size // self.num_heads // 2 pt_seq_len = self.x_embedder.img_size[0] // self.patch_size from einops import repeat theta = 10000 freqs = 1. / (theta ** (torch.arange(0, half_head_dim, 2, device=device)[:(half_head_dim // 2)].float() / half_head_dim)) t = torch.arange(hw_seq_len, device=device).float() / hw_seq_len * pt_seq_len freqs = torch.einsum('..., f -> ... f', t, freqs) freqs = repeat(freqs, '... n -> ... (n r)', r=2) from .pos_embed import broadcat freqs = broadcat((freqs[:, None, :], freqs[None, :, :]), dim=-1) freqs_cos = freqs.cos().view(-1, freqs.shape[-1]).to(dtype) freqs_sin = freqs.sin().view(-1, freqs.shape[-1]).to(dtype) def dynamic_rope_fn(t_input): from .pos_embed import rotate_half return t_input * freqs_cos + rotate_half(t_input) * freqs_sin return dynamic_rope_fn def forward_flexible(self, x, t, r=None, z=None): """ Forward pass that supports variable input sizes. Unlike the standard forward, this dynamically generates RoPE to accommodate the current input resolution. """ use_checkpoint = self.use_checkpoint N, C, H, W = x.shape assert H == W, "forward_flexible currently only supports square inputs" current_hw_seq_len = H // self.patch_size x = self.x_embedder(x) # (N, T, D) t_raw = t t = self.t_embedder(t) # (N, D) r = self.r_embedder(r)*(t_raw-r).unsqueeze(-1) if self.r_embedder is not None else 0 # (N, D) c = t + r # (N, D) c0 = self.t_block(c) if self.z_dims is not None and z is not None: z = z[0] z = self.layer_norm(z) z = self.mlp_ca(z) if self.use_rope: train_hw_seq_len = self.x_embedder.img_size[0] // self.patch_size if current_hw_seq_len != train_hw_seq_len: feat_rope = self._get_dynamic_rope(current_hw_seq_len, x.device, x.dtype) else: feat_rope = self.feat_rope else: feat_rope = None for block in self.blocks: if use_checkpoint: x = checkpoint(block, x, c0, z, feat_rope, use_reentrant=True) else: x = block(x, c0, z, feat_rope) x = self.final_layer(x, c) # (N, T, patch_size ** 2 * out_channels) x = self.unpatchify(x) # (N, out_channels, H, W) return x def interpolate_pos_embed_2d(pos_embed, new_size, old_size): """ Interpolate 2D positional embeddings via bicubic resize. pos_embed: [1, old_H*old_W, D] new_size: (new_H, new_W) old_size: (old_H, old_W) """ if new_size == old_size: return pos_embed old_H, old_W = old_size new_H, new_W = new_size D = pos_embed.shape[-1] pos_embed_2d = pos_embed.reshape(1, old_H, old_W, D).permute(0, 3, 1, 2) pos_embed_new = F.interpolate( pos_embed_2d, size=(new_H, new_W), mode='bicubic', align_corners=False ) pos_embed_new = pos_embed_new.permute(0, 2, 3, 1).reshape(1, new_H * new_W, D) return pos_embed_new def get_2d_sincos_pos_embed(embed_dim, grid_size, 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_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) grid = grid.reshape([2, 1, grid_size, grid_size]) 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 (M,) 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