Upload models/lightningdit_rot_head.py with huggingface_hub

models/lightningdit_rot_head.py

@@ -0,0 +1,684 @@
+"""
+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 models.swiglu_ffn import SwiGLUFFN 
+from models.pos_embed import VisionRotaryEmbeddingFast
+from models.rmsnorm import RMSNorm
+from visualize_attention import visualize_attention_matrix
+
+# def rot_by_head(q,k,v,num_rot):
+#     # q, k, v: (B, num_heads, N, head_dim)
+#     # q1, q2, q3, q4 = torch.chunk(q, 4, dim=1)
+#     # k1, k2, k3, k4 = torch.chunk(k, 4, dim=1)
+#     # v1, v2, v3, v4 = torch.chunk(v, 4, dim=1)
+#     q_list = torch.chunk(q, num_rot, dim=1)
+#     k_list = torch.chunk(k, num_rot, dim=1)
+#     v_list = torch.chunk(v, num_rot, dim=1)
+#     q_rot_list = []
+#     k_rot_list = []
+#     v_rot_list = []
+#     c = int(4 / num_rot)
+#     for i, (q_i, k_i, v_i) in enumerate(zip(q_list, k_list, v_list)):
+#         if i != 0:
+#             B, num_heads_i, N, head_dim_i = q_i.shape
+#             H = W = int(N ** 0.5)
+#             q_i = q_i.reshape(B, num_heads_i, H, W, head_dim_i)
+#             k_i = k_i.reshape(B, num_heads_i, H, W, head_dim_i)
+#             v_i = v_i.reshape(B, num_heads_i, H, W, head_dim_i)
+#             k = c*i
+#             q_i_rot = torch.rot90(q_i, k=k, dims=(3, 2))
+#             k_i_rot = torch.rot90(k_i, k=k, dims=(3, 2))
+#             v_i_rot = torch.rot90(v_i, k=k, dims=(3, 2))
+#             q_rot_list.append(q_i_rot.reshape(B, num_heads_i, N, head_dim_i))
+#             k_rot_list.append(k_i_rot.reshape(B, num_heads_i, N, head_dim_i))
+#             v_rot_list.append(v_i_rot.reshape(B, num_heads_i, N, head_dim_i))
+#         else:
+#             q_rot_list.append(q_i)
+#             k_rot_list.append(k_i)
+#             v_rot_list.append(v_i)
+#     q_rot = torch.cat(q_rot_list, dim=1)
+#     k_rot = torch.cat(k_rot_list, dim=1)
+#     v_rot = torch.cat(v_rot_list, dim=1)
+
+#     return q_rot, k_rot, v_rot
+
+# def recover(x_rot, num_rot):
+#     x_rot_list = torch.chunk(x_rot, num_rot, dim=1)
+#     x_list = []
+#     c = int(4 / num_rot)
+#     for i, x_i in enumerate(x_rot_list):
+#         if i != 0:
+#             B, num_heads_i, N, head_dim_i = x_i.shape
+#             H = W = int(N ** 0.5)
+#             x_i = x_i.reshape(B, num_heads_i, H, W, head_dim_i)
+#             k = -c*i
+#             x_i_recov = torch.rot90(x_i, k=k, dims=(3, 2))
+#             x_list.append(x_i_recov.reshape(B, num_heads_i, N, head_dim_i))
+#         else:
+#             x_list.append(x_i)
+
+#     return torch.cat(x_list, dim=1)
+
+#----#
+# import torch
+@torch.compile
+def rot_by_head(q, k, v, num_rot):
+    # q, k, v: (B, num_heads, N, head_dim)
+    B, num_heads, N, head_dim = q.shape
+    H = W = int(N ** 0.5)
+    
+    # 1. 合并处理 q, k, v -> (3, B, num_heads, N, head_dim)
+    combined = torch.stack([q, k, v])
+    
+    # 2. 变形为 7 维以进行批量旋转
+    # 维度索引: 0:3, 1:B, 2:num_rot, 3:heads_per_group, 4:H, 5:W, 6:head_dim
+    heads_per_group = num_heads // num_rot
+    combined = combined.view(3, B, num_rot, heads_per_group, H, W, head_dim)
+
+    c = 4 // num_rot
+    results = []
+
+    for i in range(num_rot):
+        if i == 0:
+            results.append(combined[:, :, i]) # (3, B, heads_per_group, H, W, head_dim)
+        else:
+            k_rot = c * i
+            # k_rot=i
+            # 修正点：在 6 维张量中，H 和 W 分别是 index 3 和 4
+            # 原代码 dims=(3, 2) 对应 H, W 旋转，这里对应 (4, 3)
+            rotated = torch.rot90(combined[:, :, i], k=k_rot, dims=(5, 4)) # 这里的 5, 4 是相对于 (3, B, heads_per_group, H, W, head_dim) 的 H, W
+            # 纠正：对于 combined[:, :, i] 来说，shape 是 (3, B, heads_per_group, H, W, head_dim)
+            # H 是 dim 3, W 是 dim 4。所以 dims 应为 (4, 3)
+            rotated = torch.rot90(combined[:, :, i], k=k_rot, dims=(4, 3))
+            results.append(rotated)
+            
+    # 3. 合并回原始形状
+    # cat 之后 shape: (3, B, num_heads, H, W, head_dim)
+    out = torch.cat(results, dim=2) 
+    out = out.reshape(3, B, num_heads, N, head_dim) # 使用 reshape 比 view 更稳健
+    
+    return out[0], out[1], out[2]
+
+@torch.compile
+def recover(x_rot, num_rot):
+    # x_rot: (B, num_heads, N, head_dim)
+    B, num_heads, N, head_dim = x_rot.shape
+    H = W = int(N ** 0.5)
+
+    heads_per_group = num_heads // num_rot
+    # 维度索引: 0:B, 1:num_rot, 2:heads_per_group, 3:H, 4:W, 5:head_dim
+    x_grouped = x_rot.view(B, num_rot, heads_per_group, H, W, head_dim)
+
+    c = 4 // num_rot
+    results = []
+
+    for i in range(num_rot):
+        if i == 0:
+            results.append(x_grouped[:, i])
+        else:
+            k_rot = -c * i
+            # 在 5 维张量 (B, heads_per_group, H, W, head_dim) 中，H 是 2, W 是 3
+            rotated = torch.rot90(x_grouped[:, i], k=k_rot, dims=(3, 2))
+            results.append(rotated)
+            
+    # 合并并还原
+    out = torch.cat(results, dim=1)
+    return out.reshape(B, num_heads, N, head_dim)
+#----#
+
+@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,
+        num_rot: int = 4
+    ) -> 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)
+
+        self.num_rot = num_rot
+        
+    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) #(B, self.num_heads, N, self.head_dim)
+
+        # divide into groups by heads
+        assert self.num_rot==4
+        # q_rot, k_rot, v_rot = rot_by_head(q, k, v, self.num_rot) #(B, self.num_heads, N, self.head_dim)
+        q_rot, k_rot, v_rot = rot_by_head(q, k, v, self.num_rot) #(B, self.num_heads, N, self.head_dim)
+
+        if self.fused_attn:
+            x = F.scaled_dot_product_attention(
+                q_rot, k_rot, v_rot,  
+                is_causal=True, 
+                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 = recover(x, self.num_rot) #(B, self.num_heads, N, self.head_dim)
+        x = recover(x, self.num_rot)
+
+        # #------Store attention map for visualization------#
+        # q_vis = q * self.scale
+        # attn_vis = q_vis @ k.transpose(-2, -1)
+        # attn_vis = attn_vis.softmax(dim=-1)
+        # print("**************")
+        # print("q.shape:", q_vis.shape) # (B, num_heads=16, N=256, head_dim=72)
+        # print("k.shape:", k.shape) # (B, num_heads=16, N=256, head_dim=72)
+        # print("attn_vis.shape:", attn_vis.shape) # (B, num_heads=16, N=256, N=256)
+        # print("**************")
+
+        
+        # Save attention weights as class attribute for visualization
+        # self.attn_weights = attn_vis.detach()
+        # #-------------------------------------------------#
+
+
+        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 LabelEmbedder(nn.Module):
+    """
+    Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
+    Same as DiT.
+    """
+    def __init__(self, num_classes, hidden_size, dropout_prob):
+        super().__init__()
+        use_cfg_embedding = dropout_prob > 0
+        self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size)
+        self.num_classes = num_classes
+        self.dropout_prob = dropout_prob
+
+    def token_drop(self, labels, force_drop_ids=None):
+        """
+        Drops labels to enable classifier-free guidance.
+        """
+        if force_drop_ids is None:
+            drop_ids = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
+        else:
+            drop_ids = force_drop_ids == 1
+        labels = torch.where(drop_ids, self.num_classes, labels)
+        return labels
+
+    @torch.compile
+    def forward(self, labels, train, force_drop_ids=None):
+        use_dropout = self.dropout_prob > 0
+        if (train and use_dropout) or (force_drop_ids is not None):
+            labels = self.token_drop(labels, force_drop_ids)
+        embeddings = self.embedding_table(labels)
+        return embeddings
+
+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,
+        num_rot=4,
+        **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,
+            num_rot=num_rot,
+            **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
+            )
+            
+        # Initialize AdaLN modulation
+        if wo_shift:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(hidden_size, 4 * hidden_size, bias=True)
+            )
+        else:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(hidden_size, 6 * hidden_size, bias=True)
+            )
+        self.wo_shift = wo_shift
+
+    @torch.compile
+    def forward(self, x, c, feat_rope=None):
+        if self.wo_shift:
+            scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(4, dim=1)
+            shift_msa = None
+            shift_mlp = None
+        else:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=1)
+            
+        x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa), rope=feat_rope)
+        x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(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,
+        hidden_size=1152,
+        depth=28,
+        num_heads=16,
+        mlp_ratio=4.0,
+        class_dropout_prob=0.1,
+        num_classes=1000,
+        learn_sigma=False,
+        use_qknorm=False,
+        use_swiglu=False,
+        use_rope=False,
+        use_rmsnorm=False,
+        wo_shift=False,
+        num_rot=4,
+        use_checkpoint=False,
+    ):
+        super().__init__()
+        self.learn_sigma = learn_sigma
+        self.in_channels = in_channels
+        self.out_channels = in_channels if not learn_sigma else in_channels * 2
+        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, in_channels, hidden_size, bias=True)
+        self.t_embedder = TimestepEmbedder(hidden_size)
+        self.y_embedder = LabelEmbedder(num_classes, hidden_size, class_dropout_prob)
+        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.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,
+                     num_rot=num_rot,
+                     ) for _ in range(depth)
+        ])
+        self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels, use_rmsnorm=use_rmsnorm)
+        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)
+
+        # Zero-out adaLN modulation layers in LightningDiT blocks:
+        for block in self.blocks:
+            nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
+            nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
+
+        # 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 forward(self, x, t=None, y=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
+        y: (N,) tensor of class labels
+        use_checkpoint: boolean to toggle checkpointing
+        """
+
+        use_checkpoint = self.use_checkpoint
+        # print("x.shape:", x.shape) # (N, C, H, W)
+        x = self.x_embedder(x) + self.pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
+        t = self.t_embedder(t)                   # (N, D)
+        y = self.y_embedder(y, self.training)    # (N, D)
+        c = t + y                                # (N, D)
+
+        for block in self.blocks:
+            if use_checkpoint:
+                x = checkpoint(block, x, c, self.feat_rope, use_reentrant=True)
+            else:
+                x = block(x, c, 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)
+
+        if self.learn_sigma:
+            x, _ = x.chunk(2, dim=1)
+        return x
+
+    def forward_with_cfg(self, x, t, y, cfg_scale, cfg_interval=None, cfg_interval_start=None):
+        """
+        Forward pass of LightningDiT, but also batches the unconditional forward pass for classifier-free guidance.
+        """
+        # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
+        half = x[: len(x) // 2]
+        combined = torch.cat([half, half], dim=0)
+        model_out = self.forward(combined, t, y)
+        # For exact reproducibility reasons, we apply classifier-free guidance on only
+        # three channels by default. The standard approach to cfg applies it to all channels.
+        # This can be done by uncommenting the following line and commenting-out the line following that.
+        # eps, rest = model_out[:, :self.in_channels], model_out[:, self.in_channels:]
+        eps, rest = model_out[:, :3], model_out[:, 3:]
+        cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
+        half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps)
+        
+        if cfg_interval is True:
+            timestep = t[0]
+            if timestep < cfg_interval_start:
+                half_eps = cond_eps
+
+        eps = torch.cat([half_eps, half_eps], dim=0)
+        return torch.cat([eps, rest], dim=1)
+
+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
+
+
+#################################################################################
+#                             LightningDiT Configs                              #
+#################################################################################
+
+def LightningDiT_XL_1(**kwargs):
+    return LightningDiT(depth=28, hidden_size=1152, patch_size=1, num_heads=16, **kwargs)
+
+def LightningDiT_XL_2(**kwargs):
+    return LightningDiT(depth=28, hidden_size=1152, patch_size=2, num_heads=16, **kwargs)
+
+def LightningDiT_L_2(**kwargs):
+    return LightningDiT(depth=24, hidden_size=1024, patch_size=2, num_heads=16, **kwargs)
+
+def LightningDiT_B_1(**kwargs):
+    return LightningDiT(depth=12, hidden_size=768, patch_size=1, num_heads=12, **kwargs)
+
+def LightningDiT_B_2(**kwargs):
+    return LightningDiT(depth=12, hidden_size=768, patch_size=2, num_heads=12, **kwargs)
+
+def LightningDiT_1p0B_1(**kwargs):
+    return LightningDiT(depth=24, hidden_size=1536, patch_size=1, num_heads=24, **kwargs)
+
+def LightningDiT_1p0B_2(**kwargs):
+    return LightningDiT(depth=24, hidden_size=1536, patch_size=2, num_heads=24, **kwargs)
+
+def LightningDiT_1p6B_1(**kwargs):
+    return LightningDiT(depth=28, hidden_size=1792, patch_size=1, num_heads=28, **kwargs)
+
+def LightningDiT_1p6B_2(**kwargs):
+    return LightningDiT(depth=28, hidden_size=1792, patch_size=2, num_heads=28, **kwargs)
+
+LightningDiT_models = {
+    'LightningDiT-B/1': LightningDiT_B_1, 'LightningDiT-B/2': LightningDiT_B_2,
+    'LightningDiT-L/2': LightningDiT_L_2,
+    'LightningDiT-XL/1': LightningDiT_XL_1, 'LightningDiT-XL/2': LightningDiT_XL_2,
+    'LightningDiT-1p0B/1': LightningDiT_1p0B_1, 'LightningDiT-1p0B/2': LightningDiT_1p0B_2,
+    'LightningDiT-1p6B/1': LightningDiT_1p6B_1, 'LightningDiT-1p6B/2': LightningDiT_1p6B_2,
+}