Upload STDiT

config.json

@@ -0,0 +1,39 @@
+{
+  "architectures": [
+    "STDiT"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_stdit.STDiTConfig",
+    "AutoModel": "modeling_stdit.STDiT"
+  },
+  "caption_channels": 4096,
+  "class_dropout_prob": 0.1,
+  "depth": 28,
+  "drop_path": 0.0,
+  "enable_flash_attn": false,
+  "enable_layernorm_kernel": false,
+  "enable_sequence_parallelism": false,
+  "freeze": null,
+  "hidden_size": 1152,
+  "in_channels": 4,
+  "input_size": [
+    16,
+    32,
+    32
+  ],
+  "mlp_ratio": 4.0,
+  "model_max_length": 120,
+  "model_type": "stdit",
+  "no_temporal_pos_emb": false,
+  "num_heads": 16,
+  "patch_size": [
+    1,
+    2,
+    2
+  ],
+  "pred_sigma": true,
+  "space_scale": 0.5,
+  "time_scale": 1.0,
+  "torch_dtype": "float32",
+  "transformers_version": "4.38.2"
+}

configuration_stdit.py

@@ -0,0 +1,52 @@
+import torch
+from transformers import PretrainedConfig
+from typing import List
+
+
+class STDiTConfig(PretrainedConfig):
+    
+    model_type = "stdit"
+
+    def __init__(
+        self,
+        input_size=(1, 32, 32),
+        in_channels=4,
+        patch_size=(1, 2, 2),
+        hidden_size=1152,
+        depth=28,
+        num_heads=16,
+        mlp_ratio=4.0,
+        class_dropout_prob=0.1,
+        pred_sigma=True,
+        drop_path=0.0,
+        no_temporal_pos_emb=False,
+        caption_channels=4096,
+        model_max_length=120,
+        space_scale=1.0,
+        time_scale=1.0,
+        freeze=None,
+        enable_flash_attn=False,
+        enable_layernorm_kernel=False,
+        enable_sequence_parallelism=False,
+        **kwargs,
+    ):
+        self.input_size = input_size
+        self.in_channels = in_channels
+        self.patch_size = patch_size
+        self.hidden_size = hidden_size
+        self.depth = depth
+        self.num_heads = num_heads
+        self.mlp_ratio = mlp_ratio
+        self.class_dropout_prob = class_dropout_prob
+        self.pred_sigma = pred_sigma
+        self.drop_path = drop_path
+        self.no_temporal_pos_emb = no_temporal_pos_emb
+        self.caption_channels = caption_channels
+        self.model_max_length = model_max_length
+        self.space_scale = space_scale
+        self.time_scale = time_scale
+        self.freeze = freeze
+        self.enable_flash_attn = enable_flash_attn
+        self.enable_layernorm_kernel = enable_layernorm_kernel
+        self.enable_sequence_parallelism = enable_sequence_parallelism
+        super().__init__(**kwargs)

layers.py

@@ -0,0 +1,466 @@
+import math
+import collections.abc
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from einops import rearrange
+from itertools import repeat
+from functools import partial
+from .utils import get_layernorm, t2i_modulate, approx_gelu
+
+
+try:
+    import xformers
+    HAS_XFORMERS = True
+except:
+    HAS_XFORMERS = False
+
+# ============================
+# Attention
+# ============================
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = False,
+        qk_norm: bool = False,
+        attn_drop: float = 0.0,
+        proj_drop: float = 0.0,
+        norm_layer: nn.Module = nn.LayerNorm,
+        enable_flash_attn: bool = False,
+    ) -> None:
+        super().__init__()
+        assert dim % num_heads == 0, "dim should be divisible by num_heads"
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.scale = self.head_dim**-0.5
+        self.enable_flash_attn = enable_flash_attn
+
+        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) -> torch.Tensor:
+        B, N, C = x.shape
+        qkv = self.qkv(x)
+        qkv_shape = (B, N, 3, self.num_heads, self.head_dim)
+        if self.enable_flash_attn:
+            qkv_permute_shape = (2, 0, 1, 3, 4)
+        else:
+            qkv_permute_shape = (2, 0, 3, 1, 4)
+        qkv = qkv.view(qkv_shape).permute(qkv_permute_shape)
+        q, k, v = qkv.unbind(0)
+        q, k = self.q_norm(q), self.k_norm(k)
+        if self.enable_flash_attn:
+            from flash_attn import flash_attn_func
+
+            x = flash_attn_func(
+                q,
+                k,
+                v,
+                dropout_p=self.attn_drop.p if self.training else 0.0,
+                softmax_scale=self.scale,
+            )
+        else:
+            dtype = q.dtype
+            q = q * self.scale
+            attn = q @ k.transpose(-2, -1)  # translate attn to float32
+            attn = attn.to(torch.float32)
+            attn = attn.softmax(dim=-1)
+            attn = attn.to(dtype)  # cast back attn to original dtype
+            attn = self.attn_drop(attn)
+            x = attn @ v
+
+        x_output_shape = (B, N, C)
+        if not self.enable_flash_attn:
+            x = x.transpose(1, 2)
+        x = x.reshape(x_output_shape)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+# ============================
+# Caption Embedding
+# ============================
+class CaptionEmbedder(nn.Module):
+    """
+    Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
+    """
+
+    def __init__(self, in_channels, hidden_size, uncond_prob, act_layer=nn.GELU(approximate="tanh"), token_num=120):
+        super().__init__()
+        self.y_proj = Mlp(
+            in_features=in_channels, hidden_features=hidden_size, out_features=hidden_size, act_layer=act_layer, drop=0
+        )
+        self.register_buffer("y_embedding", nn.Parameter(torch.randn(token_num, in_channels) / in_channels**0.5))
+        self.uncond_prob = uncond_prob
+
+    def token_drop(self, caption, force_drop_ids=None):
+        """
+        Drops labels to enable classifier-free guidance.
+        """
+        if force_drop_ids is None:
+            drop_ids = torch.rand(caption.shape[0]).cuda() < self.uncond_prob
+        else:
+            drop_ids = force_drop_ids == 1
+        caption = torch.where(drop_ids[:, None, None, None], self.y_embedding, caption)
+        return caption
+
+    def forward(self, caption, train, force_drop_ids=None):
+        if train:
+            assert caption.shape[2:] == self.y_embedding.shape
+        use_dropout = self.uncond_prob > 0
+        if (train and use_dropout) or (force_drop_ids is not None):
+            caption = self.token_drop(caption, force_drop_ids)
+        caption = self.y_proj(caption)
+        return caption
+
+
+# ============================
+# Drop Path
+# ============================
+def drop_path(x, drop_prob: float = 0., training: bool = False, scale_by_keep: bool = True):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+    'survival rate' as the argument.
+
+    """
+    if drop_prob == 0. or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+    if keep_prob > 0.0 and scale_by_keep:
+        random_tensor.div_(keep_prob)
+    return x * random_tensor
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob: float = 0., scale_by_keep: bool = True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)
+
+    def extra_repr(self):
+        return f'drop_prob={round(self.drop_prob,3):0.3f}'
+
+
+# ============================
+# MHA
+# ============================
+class MultiHeadCrossAttention(nn.Module):
+    def __init__(self, d_model, num_heads, attn_drop=0.0, proj_drop=0.0):
+        super(MultiHeadCrossAttention, self).__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.q_linear = nn.Linear(d_model, d_model)
+        self.kv_linear = nn.Linear(d_model, d_model * 2)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(d_model, d_model)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x, cond, mask=None):
+        # query/value: img tokens; key: condition; mask: if padding tokens
+        B, N, C = x.shape
+
+        q = self.q_linear(x).view(1, -1, self.num_heads, self.head_dim)
+        kv = self.kv_linear(cond).view(1, -1, 2, self.num_heads, self.head_dim)
+        k, v = kv.unbind(2)
+
+        attn_bias = None
+        assert HAS_XFORMERS, "Please install xformers to use this module."
+        if mask is not None:
+            attn_bias = xformers.ops.fmha.BlockDiagonalMask.from_seqlens([N] * B, mask)
+        x = xformers.ops.memory_efficient_attention(q, k, v, p=self.attn_drop.p, attn_bias=attn_bias)
+
+        x = x.view(B, -1, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+# ============================
+# MLP
+# ============================
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
+            return tuple(x)
+        return tuple(repeat(x, n))
+    return parse
+
+to_2tuple = _ntuple(2)
+
+class Mlp(nn.Module):
+    """ MLP as used in Vision Transformer, MLP-Mixer and related networks
+    """
+    def __init__(
+            self,
+            in_features,
+            hidden_features=None,
+            out_features=None,
+            act_layer=nn.GELU,
+            norm_layer=None,
+            bias=True,
+            drop=0.,
+            use_conv=False,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        bias = to_2tuple(bias)
+        drop_probs = to_2tuple(drop)
+        linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear
+
+        self.fc1 = linear_layer(in_features, hidden_features, bias=bias[0])
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs[0])
+        self.norm = norm_layer(hidden_features) if norm_layer is not None else nn.Identity()
+        self.fc2 = linear_layer(hidden_features, out_features, bias=bias[1])
+        self.drop2 = nn.Dropout(drop_probs[1])
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop1(x)
+        x = self.norm(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x
+
+
+# ============================
+# Patch Embedding
+# ============================
+class PatchEmbed3D(nn.Module):
+    """Video to Patch Embedding.
+
+    Args:
+        patch_size (int): Patch token size. Default: (2,4,4).
+        in_chans (int): Number of input video channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(
+        self,
+        patch_size=(2, 4, 4),
+        in_chans=3,
+        embed_dim=96,
+        norm_layer=None,
+        flatten=True,
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.flatten = flatten
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv3d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        """Forward function."""
+        # padding
+        _, _, D, H, W = x.size()
+        if W % self.patch_size[2] != 0:
+            x = F.pad(x, (0, self.patch_size[2] - W % self.patch_size[2]))
+        if H % self.patch_size[1] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[1] - H % self.patch_size[1]))
+        if D % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, 0, 0, self.patch_size[0] - D % self.patch_size[0]))
+
+        x = self.proj(x)  # (B C T H W)
+        if self.norm is not None:
+            D, Wh, Ww = x.size(2), x.size(3), x.size(4)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, D, Wh, Ww)
+        if self.flatten:
+            x = x.flatten(2).transpose(1, 2)  # BCTHW -> BNC
+        return x
+
+
+# ============================
+# T2I
+# ============================
+def t2i_modulate(x, shift, scale):
+    return x * (1 + scale) + shift
+
+class T2IFinalLayer(nn.Module):
+    """
+    The final layer of PixArt.
+    """
+
+    def __init__(self, hidden_size, num_patch, out_channels):
+        super().__init__()
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = nn.Linear(hidden_size, num_patch * out_channels, bias=True)
+        self.scale_shift_table = nn.Parameter(torch.randn(2, hidden_size) / hidden_size**0.5)
+        self.out_channels = out_channels
+
+    def forward(self, x, t):
+        shift, scale = (self.scale_shift_table[None] + t[:, None]).chunk(2, dim=1)
+        x = t2i_modulate(self.norm_final(x), shift, scale)
+        x = self.linear(x)
+        return x
+
+
+# ============================
+# Time Step Embedding
+# ============================
+class TimestepEmbedder(nn.Module):
+    """
+    Embeds scalar timesteps into vector representations.
+    """
+
+    def __init__(self, hidden_size, frequency_embedding_size=256):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
+            nn.SiLU(),
+            nn.Linear(hidden_size, hidden_size, bias=True),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+
+    @staticmethod
+    def timestep_embedding(t, dim, max_period=10000):
+        """
+        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.
+        """
+        # 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)
+        freqs = freqs.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
+
+    def forward(self, t, dtype):
+        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
+        if t_freq.dtype != dtype:
+            t_freq = t_freq.to(dtype)
+        t_emb = self.mlp(t_freq)
+        return t_emb
+
+
+# ============================
+# STDiT Block
+# ============================
+class STDiTBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        num_heads,
+        d_s=None,
+        d_t=None,
+        mlp_ratio=4.0,
+        drop_path=0.0,
+        enable_flash_attn=False,
+        enable_layernorm_kernel=False,
+        enable_sequence_parallelism=False,
+    ):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.enable_flash_attn = enable_flash_attn
+        self._enable_sequence_parallelism = enable_sequence_parallelism
+
+        if enable_sequence_parallelism:
+            self.attn_cls = SeqParallelAttention
+            self.mha_cls = SeqParallelMultiHeadCrossAttention
+        else:
+            self.attn_cls = Attention
+            self.mha_cls = MultiHeadCrossAttention
+
+        self.norm1 = get_layernorm(hidden_size, eps=1e-6, affine=False, use_kernel=enable_layernorm_kernel)
+        self.attn = self.attn_cls(
+            hidden_size,
+            num_heads=num_heads,
+            qkv_bias=True,
+            enable_flash_attn=enable_flash_attn,
+        )
+        self.cross_attn = self.mha_cls(hidden_size, num_heads)
+        self.norm2 = get_layernorm(hidden_size, eps=1e-6, affine=False, use_kernel=enable_layernorm_kernel)
+        self.mlp = Mlp(
+            in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu, drop=0
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+        self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size**0.5)
+
+        # temporal attention
+        self.d_s = d_s
+        self.d_t = d_t
+
+        if self._enable_sequence_parallelism:
+            sp_size = dist.get_world_size(get_sequence_parallel_group())
+            # make sure d_t is divisible by sp_size
+            assert d_t % sp_size == 0
+            self.d_t = d_t // sp_size
+
+        self.attn_temp = self.attn_cls(
+            hidden_size,
+            num_heads=num_heads,
+            qkv_bias=True,
+            enable_flash_attn=self.enable_flash_attn,
+        )
+
+    def forward(self, x, y, t, mask=None, tpe=None):
+        B, N, C = x.shape
+
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+            self.scale_shift_table[None] + t.reshape(B, 6, -1)
+        ).chunk(6, dim=1)
+        x_m = t2i_modulate(self.norm1(x), shift_msa, scale_msa)
+
+        # spatial branch
+        x_s = rearrange(x_m, "B (T S) C -> (B T) S C", T=self.d_t, S=self.d_s)
+        x_s = self.attn(x_s)
+        x_s = rearrange(x_s, "(B T) S C -> B (T S) C", T=self.d_t, S=self.d_s)
+        x = x + self.drop_path(gate_msa * x_s)
+
+        # temporal branch
+        x_t = rearrange(x, "B (T S) C -> (B S) T C", T=self.d_t, S=self.d_s)
+        if tpe is not None:
+            x_t = x_t + tpe
+        x_t = self.attn_temp(x_t)
+        x_t = rearrange(x_t, "(B S) T C -> B (T S) C", T=self.d_t, S=self.d_s)
+        x = x + self.drop_path(gate_msa * x_t)
+
+        # cross attn
+        x = x + self.cross_attn(x, y, mask)
+
+        # mlp
+        x = x + self.drop_path(gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))
+
+        return x

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c357c5382353a0862f99e6a06a888f74b2f33b2e614f2084636a106037dfe68d
+size 3041758968

modeling_stdit.py

@@ -0,0 +1,260 @@
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+from einops import rearrange
+from .configuration_stdit import STDiTConfig
+from .layers import (
+    STDiTBlock,
+    CaptionEmbedder,
+    PatchEmbed3D,
+    T2IFinalLayer,
+    TimestepEmbedder,
+)
+from .utils import (
+    approx_gelu,
+    get_1d_sincos_pos_embed,
+    get_2d_sincos_pos_embed,
+)
+from transformers import PreTrainedModel
+
+
+class STDiT(PreTrainedModel):
+
+    config_class = STDiTConfig
+
+    def __init__(
+        self,
+        config
+    ):
+        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.patch_size = config.patch_size
+        self.input_size = config.input_size
+        num_patches = np.prod([config.input_size[i] // config.patch_size[i] for i in range(3)])
+        self.num_patches = num_patches
+        self.num_temporal = config.input_size[0] // config.patch_size[0]
+        self.num_spatial = num_patches // self.num_temporal
+        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.space_scale = config.space_scale
+        self.time_scale = config.time_scale
+
+        self.register_buffer("pos_embed", self.get_spatial_pos_embed())
+        self.register_buffer("pos_embed_temporal", self.get_temporal_pos_embed())
+
+        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.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.blocks = nn.ModuleList(
+            [
+                STDiTBlock(
+                    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=config.enable_sequence_parallelism,
+                    d_t=self.num_temporal,
+                    d_s=self.num_spatial,
+                )
+                for i in range(self.depth)
+            ]
+        )
+        self.final_layer = T2IFinalLayer(config.hidden_size, np.prod(self.patch_size), self.out_channels)
+
+        # 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
+        self.enable_sequence_parallelism = config.enable_sequence_parallelism
+        if config.enable_sequence_parallelism:
+            self.sp_rank = dist.get_rank(get_sequence_parallel_group())
+        else:
+            self.sp_rank = None
+
+    def forward(self, x, timestep, y, mask=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]
+        """
+        # embedding
+        x = self.x_embedder(x)  # [B, N, C]
+        x = rearrange(x, "B (T S) C -> B T S C", T=self.num_temporal, S=self.num_spatial)
+        x = x + self.pos_embed
+        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")
+
+        t = self.t_embedder(timestep, dtype=x.dtype)  # [B, C]
+        t0 = self.t_block(t)  # [B, C]
+        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 i, block in enumerate(self.blocks):
+            if i == 0:
+                if self.enable_sequence_parallelism:
+                    tpe = torch.chunk(
+                        self.pos_embed_temporal, dist.get_world_size(get_sequence_parallel_group()), dim=1
+                    )[self.sp_rank].contiguous()
+                else:
+                    tpe = self.pos_embed_temporal
+            else:
+                tpe = None
+            x = auto_grad_checkpoint(block, x, y, t0, y_lens, tpe)
+
+        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)  # [B, N, C=T_p * H_p * W_p * C_out]
+        x = self.unpatchify(x)  # [B, C_out, T, H, W]
+
+        # cast to float32 for better accuracy
+        x = x.to(torch.float32)
+        return x
+
+    def unpatchify(self, x):
+        """
+        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,
+        )
+        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, grid_size=None):
+        if grid_size is None:
+            grid_size = self.input_size[1:]
+        pos_embed = get_2d_sincos_pos_embed(
+            self.hidden_size,
+            (grid_size[0] // self.patch_size[1], grid_size[1] // self.patch_size[2]),
+            scale=self.space_scale,
+        )
+        pos_embed = torch.from_numpy(pos_embed).float().unsqueeze(0).requires_grad_(False)
+        return pos_embed
+
+    def get_temporal_pos_embed(self):
+        pos_embed = get_1d_sincos_pos_embed(
+            self.hidden_size,
+            self.input_size[0] // self.patch_size[0],
+            scale=self.time_scale,
+        )
+        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)
+
+        # 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)

utils.py

@@ -0,0 +1,80 @@
+import torch
+import torch.nn as nn
+import numpy as np
+
+approx_gelu = lambda: nn.GELU(approximate="tanh")
+
+def get_layernorm(hidden_size: torch.Tensor, eps: float, affine: bool, use_kernel: bool):
+    if use_kernel:
+        try:
+            from apex.normalization import FusedLayerNorm
+
+            return FusedLayerNorm(hidden_size, elementwise_affine=affine, eps=eps)
+        except ImportError:
+            raise RuntimeError("FusedLayerNorm not available. Please install apex.")
+    else:
+        return nn.LayerNorm(hidden_size, eps, elementwise_affine=affine)
+
+def get_1d_sincos_pos_embed(embed_dim, length, scale=1.0):
+    pos = np.arange(0, length)[..., None] / scale
+    return get_1d_sincos_pos_embed_from_grid(embed_dim, pos)
+
+
+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 / 2.0
+    omega = 1.0 / 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
+
+
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0, scale=1.0, base_size=None):
+    """
+    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)
+    """
+    if not isinstance(grid_size, tuple):
+        grid_size = (grid_size, grid_size)
+
+    grid_h = np.arange(grid_size[0], dtype=np.float32) / scale
+    grid_w = np.arange(grid_size[1], dtype=np.float32) / scale
+    if base_size is not None:
+        grid_h *= base_size / grid_size[0]
+        grid_w *= base_size / grid_size[1]
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    grid = grid.reshape([2, 1, grid_size[1], grid_size[0]])
+    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 t2i_modulate(x, shift, scale):
+    return x * (1 + scale) + shift