from einops import rearrange from torch import nn import torch import numpy as np from einops import rearrange, repeat from typing import Any, Dict, Optional, Tuple from diffusers.utils.torch_utils import maybe_allow_in_graph from typing import Any, Dict, Optional import torch import torch.nn.functional as F from torch import nn import diffusers from diffusers.utils import deprecate, logging from diffusers.utils.torch_utils import maybe_allow_in_graph from diffusers.models.attention import FeedForward, GatedSelfAttentionDense from diffusers.models.attention_processor import Attention as Attention_ from diffusers.models.embeddings import SinusoidalPositionalEmbedding from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm from .rope import PositionGetter3D, RoPE3D import re try: import torch_npu from opensora.npu_config import npu_config, set_run_dtype except: torch_npu = None npu_config = None set_run_dtype = None logger = logging.get_logger(__name__) def get_3d_sincos_pos_embed( embed_dim, grid_size, cls_token=False, extra_tokens=0, interpolation_scale=1.0, base_size=16, ): """ 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 isinstance(grid_size, int): # grid_size = (grid_size, grid_size) grid_t = np.arange(grid_size[0], dtype=np.float32) / (grid_size[0] / base_size[0]) / interpolation_scale[0] grid_h = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1] / base_size[1]) / interpolation_scale[1] grid_w = np.arange(grid_size[2], dtype=np.float32) / (grid_size[2] / base_size[2]) / interpolation_scale[2] grid = np.meshgrid(grid_w, grid_h, grid_t) # here w goes first grid = np.stack(grid, axis=0) grid = grid.reshape([3, 1, grid_size[2], grid_size[1], grid_size[0]]) pos_embed = get_3d_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_3d_sincos_pos_embed_from_grid(embed_dim, grid): if embed_dim % 3 != 0: raise ValueError("embed_dim must be divisible by 3") # use 1/3 of dimensions to encode grid_t/h/w emb_t = get_1d_sincos_pos_embed_from_grid(embed_dim // 3, grid[0]) # (T*H*W, D/3) emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 3, grid[1]) # (T*H*W, D/3) emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 3, grid[2]) # (T*H*W, D/3) emb = np.concatenate([emb_t, emb_h, emb_w], axis=1) # (T*H*W, D) return emb def get_2d_sincos_pos_embed( embed_dim, grid_size, cls_token=False, extra_tokens=0, interpolation_scale=1.0, base_size=16, ): """ 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 isinstance(grid_size, int): # grid_size = (grid_size, grid_size) grid_h = np.arange(grid_size[0], dtype=np.float32) / (grid_size[0] / base_size[0]) / interpolation_scale[0] grid_w = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1] / base_size[1]) / interpolation_scale[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): if embed_dim % 2 != 0: raise ValueError("embed_dim must be divisible by 2") # use 1/3 of dimensions to encode grid_t/h/w 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( embed_dim, grid_size, cls_token=False, extra_tokens=0, interpolation_scale=1.0, base_size=16, ): """ grid_size: int of the grid return: pos_embed: [grid_size, embed_dim] or [1+grid_size, embed_dim] (w/ or w/o cls_token) """ # if isinstance(grid_size, int): # grid_size = (grid_size, grid_size) grid = np.arange(grid_size, dtype=np.float32) / (grid_size / base_size) / interpolation_scale pos_embed = get_1d_sincos_pos_embed_from_grid(embed_dim, grid) # (H*W, D/2) 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_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) """ if embed_dim % 2 != 0: raise ValueError("embed_dim must be divisible by 2") 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 class FP32_Layernorm(nn.LayerNorm): def forward(self, inputs: torch.Tensor) -> torch.Tensor: origin_dtype = inputs.dtype return F.layer_norm(inputs.float(), self.normalized_shape, self.weight.float() if self.weight is not None else None, self.bias.float() if self.bias is not None else None, self.eps).to(origin_dtype) class FP32_SiLU(nn.SiLU): def forward(self, inputs: torch.Tensor) -> torch.Tensor: return torch.nn.functional.silu(inputs.float(), inplace=self.inplace).to(inputs.dtype) class FP32_GELU(nn.GELU): def forward(self, inputs: torch.Tensor) -> torch.Tensor: return torch.nn.functional.gelu(inputs.float(), approximate=self.approximate).to(inputs.dtype) class OverlapPatchEmbed3D(nn.Module): """2D Image to Patch Embedding but with 3D position embedding""" def __init__( self, num_frames=1, height=224, width=224, patch_size_t=1, patch_size=16, in_channels=3, embed_dim=768, layer_norm=False, flatten=True, bias=True, interpolation_scale=(1, 1), interpolation_scale_t=1, use_abs_pos=False, ): super().__init__() # assert patch_size_t == 1 and patch_size == 1 self.use_abs_pos = use_abs_pos self.flatten = flatten self.layer_norm = layer_norm self.proj = nn.Conv3d( in_channels, embed_dim, kernel_size=3, padding=1, stride=(patch_size_t, patch_size, patch_size), bias=bias ) if layer_norm: self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6) else: self.norm = None self.patch_size_t = patch_size_t self.patch_size = patch_size # See: # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L161 self.height, self.width = height // patch_size, width // patch_size self.base_size = (height // patch_size, width // patch_size) self.interpolation_scale = (interpolation_scale[0], interpolation_scale[1]) pos_embed = get_2d_sincos_pos_embed( embed_dim, (self.height, self.width), base_size=self.base_size, interpolation_scale=self.interpolation_scale ) self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False) self.num_frames = (num_frames - 1) // patch_size_t + 1 if num_frames % 2 == 1 else num_frames // patch_size_t self.base_size_t = (num_frames - 1) // patch_size_t + 1 if num_frames % 2 == 1 else num_frames // patch_size_t self.interpolation_scale_t = interpolation_scale_t temp_pos_embed = get_1d_sincos_pos_embed(embed_dim, self.num_frames, base_size=self.base_size_t, interpolation_scale=self.interpolation_scale_t) self.register_buffer("temp_pos_embed", torch.from_numpy(temp_pos_embed).float().unsqueeze(0), persistent=False) # self.temp_embed_gate = nn.Parameter(torch.tensor([0.0])) def forward(self, latent, num_frames): b, _, _, _, _ = latent.shape video_latent, image_latent = None, None # b c 1 h w # assert latent.shape[-3] == 1 and num_frames == 1 num_frames = latent.shape[-3] // self.patch_size_t height, width = latent.shape[-2] // self.patch_size, latent.shape[-1] // self.patch_size # latent = rearrange(latent, 'b c t h w -> (b t) c h w') latent = self.proj(latent) if self.flatten: # latent = latent.flatten(2).transpose(1, 2) # BT C H W -> BT N C latent = rearrange(latent, 'b c t h w -> (b t) (h w) c ') if self.layer_norm: latent = self.norm(latent) if self.use_abs_pos: # Interpolate positional embeddings if needed. # (For PixArt-Alpha: https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L162C151-L162C160) if self.height != height or self.width != width: # raise NotImplementedError pos_embed = get_2d_sincos_pos_embed( embed_dim=self.pos_embed.shape[-1], grid_size=(height, width), base_size=self.base_size, interpolation_scale=self.interpolation_scale, ) pos_embed = torch.from_numpy(pos_embed) pos_embed = pos_embed.float().unsqueeze(0).to(latent.device) else: pos_embed = self.pos_embed if self.num_frames != num_frames: # import ipdb;ipdb.set_trace() # raise NotImplementedError temp_pos_embed = get_1d_sincos_pos_embed( embed_dim=self.temp_pos_embed.shape[-1], grid_size=num_frames, base_size=self.base_size_t, interpolation_scale=self.interpolation_scale_t, ) temp_pos_embed = torch.from_numpy(temp_pos_embed) temp_pos_embed = temp_pos_embed.float().unsqueeze(0).to(latent.device) else: temp_pos_embed = self.temp_pos_embed latent = (latent + pos_embed).to(latent.dtype) latent = rearrange(latent, '(b t) n c -> b t n c', b=b) video_latent = latent if self.use_abs_pos: # temp_pos_embed = temp_pos_embed.unsqueeze(2) * self.temp_embed_gate.tanh() temp_pos_embed = temp_pos_embed.unsqueeze(2) video_latent = (video_latent + temp_pos_embed).to(video_latent.dtype) video_latent = rearrange(video_latent, 'b t n c -> b (t n) c') return video_latent class PatchEmbed2D(nn.Module): """2D Image to Patch Embedding but with 3D position embedding""" def __init__( self, num_frames=1, height=224, width=224, patch_size_t=1, patch_size=16, in_channels=3, embed_dim=768, layer_norm=False, flatten=True, bias=True, interpolation_scale=(1, 1), interpolation_scale_t=1, use_abs_pos=True, ): super().__init__() # assert num_frames == 1 self.use_abs_pos = use_abs_pos self.flatten = flatten self.layer_norm = layer_norm self.proj = nn.Conv2d( in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=(patch_size, patch_size), bias=bias ) if layer_norm: self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6) else: self.norm = None self.patch_size_t = patch_size_t self.patch_size = patch_size # See: # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L161 self.height, self.width = height // patch_size, width // patch_size self.base_size = (height // patch_size, width // patch_size) self.interpolation_scale = (interpolation_scale[0], interpolation_scale[1]) pos_embed = get_2d_sincos_pos_embed( embed_dim, (self.height, self.width), base_size=self.base_size, interpolation_scale=self.interpolation_scale ) self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False) self.num_frames = (num_frames - 1) // patch_size_t + 1 if num_frames % 2 == 1 else num_frames // patch_size_t self.base_size_t = (num_frames - 1) // patch_size_t + 1 if num_frames % 2 == 1 else num_frames // patch_size_t self.interpolation_scale_t = interpolation_scale_t temp_pos_embed = get_1d_sincos_pos_embed(embed_dim, self.num_frames, base_size=self.base_size_t, interpolation_scale=self.interpolation_scale_t) self.register_buffer("temp_pos_embed", torch.from_numpy(temp_pos_embed).float().unsqueeze(0), persistent=False) # self.temp_embed_gate = nn.Parameter(torch.tensor([0.0])) def forward(self, latent, num_frames): b, _, _, _, _ = latent.shape video_latent, image_latent = None, None # b c 1 h w # assert latent.shape[-3] == 1 and num_frames == 1 height, width = latent.shape[-2] // self.patch_size, latent.shape[-1] // self.patch_size latent = rearrange(latent, 'b c t h w -> (b t) c h w') latent = self.proj(latent) if self.flatten: latent = latent.flatten(2).transpose(1, 2) # BT C H W -> BT N C if self.layer_norm: latent = self.norm(latent) if self.use_abs_pos: # Interpolate positional embeddings if needed. # (For PixArt-Alpha: https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L162C151-L162C160) if self.height != height or self.width != width: # raise NotImplementedError pos_embed = get_2d_sincos_pos_embed( embed_dim=self.pos_embed.shape[-1], grid_size=(height, width), base_size=self.base_size, interpolation_scale=self.interpolation_scale, ) pos_embed = torch.from_numpy(pos_embed) pos_embed = pos_embed.float().unsqueeze(0).to(latent.device) else: pos_embed = self.pos_embed if self.num_frames != num_frames: # import ipdb;ipdb.set_trace() # raise NotImplementedError if get_sequence_parallel_state(): if npu_config is not None: sp_size = hccl_info.world_size temp_pos_embed = get_1d_sincos_pos_embed( embed_dim=self.temp_pos_embed.shape[-1], grid_size=num_frames * sp_size, base_size=self.base_size_t, interpolation_scale=self.interpolation_scale_t, ) rank = hccl_info.rank % sp_size st_frame = rank * num_frames ed_frame = st_frame + num_frames temp_pos_embed = temp_pos_embed[st_frame: ed_frame] else: sp_size = nccl_info.world_size temp_pos_embed = get_1d_sincos_pos_embed( embed_dim=self.temp_pos_embed.shape[-1], grid_size=num_frames * sp_size, base_size=self.base_size_t, interpolation_scale=self.interpolation_scale_t, ) rank = nccl_info.rank % sp_size st_frame = rank * num_frames ed_frame = st_frame + num_frames temp_pos_embed = temp_pos_embed[st_frame: ed_frame] else: temp_pos_embed = get_1d_sincos_pos_embed( embed_dim=self.temp_pos_embed.shape[-1], grid_size=num_frames, base_size=self.base_size_t, interpolation_scale=self.interpolation_scale_t, ) temp_pos_embed = torch.from_numpy(temp_pos_embed) temp_pos_embed = temp_pos_embed.float().unsqueeze(0).to(latent.device) else: temp_pos_embed = self.temp_pos_embed latent = (latent + pos_embed).to(latent.dtype) latent = rearrange(latent, '(b t) n c -> b t n c', b=b) video_latent = latent if self.use_abs_pos: # temp_pos_embed = temp_pos_embed.unsqueeze(2) * self.temp_embed_gate.tanh() temp_pos_embed = temp_pos_embed.unsqueeze(2) video_latent = (video_latent + temp_pos_embed).to(video_latent.dtype) video_latent = rearrange(video_latent, 'b t n c -> b (t n) c') return video_latent class Attention(Attention_): def __init__(self, downsampler, attention_mode, use_rope, interpolation_scale_thw, **kwags): processor = AttnProcessor2_0(attention_mode=attention_mode, use_rope=use_rope, interpolation_scale_thw=interpolation_scale_thw) super().__init__(processor=processor, **kwags) self.downsampler = None # if downsampler: # downsampler k155_s122 # downsampler_ker_size = list(re.search(r'k(\d{2,3})', downsampler).group(1)) # 122 # down_factor = list(re.search(r's(\d{2,3})', downsampler).group(1)) # downsampler_ker_size = [int(i) for i in downsampler_ker_size] # downsampler_padding = [(i - 1) // 2 for i in downsampler_ker_size] # down_factor = [int(i) for i in down_factor] # if len(downsampler_ker_size) == 2: # self.downsampler = DownSampler2d(kwags['query_dim'], kwags['query_dim'], kernel_size=downsampler_ker_size, stride=1, # padding=downsampler_padding, groups=kwags['query_dim'], down_factor=down_factor, # down_shortcut=True) # elif len(downsampler_ker_size) == 3: # self.downsampler = DownSampler3d(kwags['query_dim'], kwags['query_dim'], kernel_size=downsampler_ker_size, stride=1, # padding=downsampler_padding, groups=kwags['query_dim'], down_factor=down_factor, # down_shortcut=True) # self.q_norm = nn.LayerNorm(kwags['dim_head'], elementwise_affine=True, eps=1e-6) # self.k_norm = nn.LayerNorm(kwags['dim_head'], elementwise_affine=True, eps=1e-6) class DownSampler3d(nn.Module): def __init__(self, *args, **kwargs): ''' Required kwargs: down_factor, downsampler''' super().__init__() self.down_factor = kwargs.pop('down_factor') self.down_shortcut = kwargs.pop('down_shortcut') self.layer = nn.Conv3d(*args, **kwargs) def forward(self, x, attention_mask, t, h, w): b = x.shape[0] x = rearrange(x, 'b (t h w) d -> b d t h w', t=t, h=h, w=w) if npu_config is None: x = self.layer(x) + (x if self.down_shortcut else 0) else: x_dtype = x.dtype x = npu_config.run_conv3d(self.layer, x, x_dtype) + (x if self.down_shortcut else 0) self.t = t//self.down_factor[0] self.h = h//self.down_factor[1] self.w = w//self.down_factor[2] # import ipdb;ipdb.set_trace() x = rearrange(x, 'b d (t dt) (h dh) (w dw) -> (b dt dh dw) (t h w) d', t=t//self.down_factor[0], h=h//self.down_factor[1], w=w//self.down_factor[2], dt=self.down_factor[0], dh=self.down_factor[1], dw=self.down_factor[2]) attention_mask = rearrange(attention_mask, 'b 1 (t h w) -> b 1 t h w', t=t, h=h, w=w) attention_mask = rearrange(attention_mask, 'b 1 (t dt) (h dh) (w dw) -> (b dt dh dw) 1 (t h w)', t=t//self.down_factor[0], h=h//self.down_factor[1], w=w//self.down_factor[2], dt=self.down_factor[0], dh=self.down_factor[1], dw=self.down_factor[2]) return x, attention_mask def reverse(self, x, t, h, w): # import ipdb;ipdb.set_trace() x = rearrange(x, '(b dt dh dw) (t h w) d -> b (t dt h dh w dw) d', t=t, h=h, w=w, dt=self.down_factor[0], dh=self.down_factor[1], dw=self.down_factor[2]) return x class DownSampler2d(nn.Module): def __init__(self, *args, **kwargs): ''' Required kwargs: down_factor, downsampler''' super().__init__() self.down_factor = kwargs.pop('down_factor') self.down_shortcut = kwargs.pop('down_shortcut') self.layer = nn.Conv2d(*args, **kwargs) def forward(self, x, attention_mask, t, h, w): # import ipdb;ipdb.set_trace() b = x.shape[0] x = rearrange(x, 'b (t h w) d -> (b t) d h w', t=t, h=h, w=w) x = self.layer(x) + (x if self.down_shortcut else 0) self.t = 1 self.h = h//self.down_factor[0] self.w = w//self.down_factor[1] x = rearrange(x, 'b d (h dh) (w dw) -> (b dh dw) (h w) d', h=h//self.down_factor[0], w=w//self.down_factor[1], dh=self.down_factor[0], dw=self.down_factor[1]) attention_mask = rearrange(attention_mask, 'b 1 (t h w) -> (b t) 1 h w', h=h, w=w) attention_mask = rearrange(attention_mask, 'b 1 (h dh) (w dw) -> (b dh dw) 1 (h w)', h=h//self.down_factor[0], w=w//self.down_factor[1], dh=self.down_factor[0], dw=self.down_factor[1]) return x, attention_mask def reverse(self, x, t, h, w): x = rearrange(x, '(b t dh dw) (h w) d -> b (t h dh w dw) d', t=t, h=h, w=w, dh=self.down_factor[0], dw=self.down_factor[1]) return x class AttnProcessor2_0: r""" Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). """ def __init__(self, attention_mode='xformers', use_rope=False, interpolation_scale_thw=(1, 1, 1)): self.use_rope = use_rope self.interpolation_scale_thw = interpolation_scale_thw if self.use_rope: self._init_rope(interpolation_scale_thw) self.attention_mode = attention_mode if not hasattr(F, "scaled_dot_product_attention"): raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") def _init_rope(self, interpolation_scale_thw): self.rope = RoPE3D(interpolation_scale_thw=interpolation_scale_thw) self.position_getter = PositionGetter3D() def __call__( self, attn: Attention, hidden_states: torch.FloatTensor, encoder_hidden_states: Optional[torch.FloatTensor] = None, attention_mask: Optional[torch.FloatTensor] = None, temb: Optional[torch.FloatTensor] = None, frame: int = 8, height: int = 16, width: int = 16, *args, **kwargs, ) -> torch.FloatTensor: if len(args) > 0 or kwargs.get("scale", None) is not None: deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." deprecate("scale", "1.0.0", deprecation_message) if attn.downsampler is not None: hidden_states, attention_mask = attn.downsampler(hidden_states, attention_mask, t=frame, h=height, w=width) frame, height, width = attn.downsampler.t, attn.downsampler.h, attn.downsampler.w residual = hidden_states if attn.spatial_norm is not None: hidden_states = attn.spatial_norm(hidden_states, temb) input_ndim = hidden_states.ndim # if input_ndim == 4: # batch_size, channel, height, width = hidden_states.shape # hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) batch_size, sequence_length, _ = ( hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape ) if attention_mask is not None: if npu_config is None: attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) # scaled_dot_product_attention expects attention_mask shape to be # (batch, heads, source_length, target_length) attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) else: if npu_config.enable_FA: attention_mask = attention_mask.to(torch.bool) attention_mask = attention_mask.view(batch_size, 1, -1, attention_mask.shape[-1]) attention_mask = attention_mask.repeat(1, 1, hidden_states.shape[1], 1) if attn.group_norm is not None: hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) query = attn.to_q(hidden_states) if encoder_hidden_states is None: encoder_hidden_states = hidden_states elif attn.norm_cross: encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) key = attn.to_k(encoder_hidden_states) value = attn.to_v(encoder_hidden_states) inner_dim = key.shape[-1] head_dim = inner_dim // attn.heads if npu_config is not None and npu_config.on_npu: if npu_config.enable_FA and query.dtype == torch.float32: dtype = torch.bfloat16 else: dtype = None if self.use_rope: query = query.view(batch_size, -1, attn.heads, head_dim) key = key.view(batch_size, -1, attn.heads, head_dim) # require the shape of (batch_size x nheads x ntokens x dim) pos_thw = self.position_getter(batch_size, t=frame, h=height, w=width, device=query.device) query = self.rope(query, pos_thw) key = self.rope(key, pos_thw) query = query.view(batch_size, -1, attn.heads * head_dim) key = key.view(batch_size, -1, attn.heads * head_dim) with set_run_dtype(query, dtype): query, key, value = npu_config.set_current_run_dtype([query, key, value]) hidden_states = npu_config.run_attention(query, key, value, attention_mask, "BSH", head_dim, attn.heads) hidden_states = npu_config.restore_dtype(hidden_states) else: query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # qk norm # query = attn.q_norm(query) # key = attn.k_norm(key) if self.use_rope: # require the shape of (batch_size x nheads x ntokens x dim) pos_thw = self.position_getter(batch_size, t=frame, h=height, w=width, device=query.device) query = self.rope(query, pos_thw) key = self.rope(key, pos_thw) value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # 0, -10000 ->(bool) False, True ->(any) True ->(not) False # 0, 0 ->(bool) False, False ->(any) False ->(not) True if attention_mask is None or not torch.any(attention_mask.bool()): # 0 mean visible attention_mask = None # the output of sdp = (batch, num_heads, seq_len, head_dim) # TODO: add support for attn.scale when we move to Torch 2.1 if self.attention_mode == 'flash' and attention_mask is None: with torch.backends.cuda.sdp_kernel(enable_math=False, enable_flash=True, enable_mem_efficient=False): hidden_states = F.scaled_dot_product_attention( query, key, value, dropout_p=0.0, is_causal=False ) elif self.attention_mode == 'xformers' or (self.attention_mode == 'flash' and attention_mask is not None): with torch.backends.cuda.sdp_kernel(enable_math=False, enable_flash=False, enable_mem_efficient=True): hidden_states = F.scaled_dot_product_attention( query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False ) elif self.attention_mode == 'math': raise RuntimeError hidden_states = F.scaled_dot_product_attention( query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False ) else: raise NotImplementedError(f'Found attention_mode: {self.attention_mode}') hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) hidden_states = hidden_states.to(query.dtype) # linear proj hidden_states = attn.to_out[0](hidden_states) # dropout hidden_states = attn.to_out[1](hidden_states) if input_ndim == 4: hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) if attn.residual_connection: hidden_states = hidden_states + residual hidden_states = hidden_states / attn.rescale_output_factor if attn.downsampler is not None: hidden_states = attn.downsampler.reverse(hidden_states, t=frame, h=height, w=width) return hidden_states class PixelUnshuffle(nn.Module): def __init__(self, ratio, ratio_t=None): super().__init__() self.r = ratio self.r_t = ratio_t if ratio_t else 1 def forward(self, x): # if self.r_t is not None and self.r_t != 1: if x.ndim == 5: b, c, t, h, w = x.shape # import ipdb;ipdb.set_trace() assert t % self.r_t == 0 and h % self.r == 0 and w % self.r == 0 x = rearrange(x, 'b c (t r1) (h r2) (w r3) -> b (c r1 r2 r3) t h w', r1=self.r_t, r2=self.r, r3=self.r) else: b, c, h, w = x.shape assert h % self.r == 0 and w % self.r == 0 x = rearrange(x, 'b c (h r2) (w r3) -> b (c r2 r3) h w', r2=self.r, r3=self.r) return x class PixelShuffle(nn.Module): def __init__(self, ratio, ratio_t=None): super().__init__() self.r = ratio self.r_t = ratio_t if ratio_t else 1 def forward(self, x): if x.ndim == 5: b, c, t, h, w = x.shape # import ipdb;ipdb.set_trace() assert c % (self.r_t*self.r*self.r) == 0 x = rearrange(x, 'b (c r1 r2 r3) t h w -> b c (t r1) (h r2) (w r3)', r1=self.r_t, r2=self.r, r3=self.r) else: b, c, h, w = x.shape assert c % (self.r*self.r) == 0 x = rearrange(x, 'b (c r2 r3) h w -> b c (h r2) (w r3)', r2=self.r, r3=self.r) return x class Downsample3d(nn.Module): def __init__(self, n_feat): super(Downsample3d, self).__init__() self.body = nn.Sequential(nn.Conv3d(n_feat, n_feat // 4, kernel_size=3, stride=1, padding=1, bias=False), PixelUnshuffle(2, 2)) def forward(self, x, attention_mask, frames, height, width, pad_h=0, pad_w=0): x = rearrange(x, 'b (t h w) d -> b d t h w', t=frames, h=height, w=width) # x = F.pad(x, (0, pad_w, 0, pad_h, 0, 0), mode='reflect') x = F.pad(x, (0, pad_w, 0, pad_h, 0, 0)) if npu_config is None: x = self.body(x) else: x_dtype = x.dtype x = npu_config.run_conv3d(self.body, x, x_dtype) x = rearrange(x, 'b d t h w -> b (t h w) d') attention_mask = rearrange(attention_mask, 'b 1 (t h w) -> b 1 t h w', t=frames, h=height, w=width) attention_mask = F.pad(attention_mask, (0, pad_w, 0, pad_h, 0, 0)) attention_mask = F.max_pool3d(attention_mask, kernel_size=2, stride=2) attention_mask = rearrange(attention_mask, 'b 1 t h w -> b 1 (t h w)') attention_bias = (1 - attention_mask.bool().to(x.dtype)) * -10000.0 return x, attention_bias, attention_mask class Upsample3d(nn.Module): def __init__(self, n_feat): super(Upsample3d, self).__init__() self.body = nn.Sequential(nn.Conv3d(n_feat, n_feat * 4, kernel_size=3, stride=1, padding=1, bias=False), PixelShuffle(2, 2)) def forward(self, x, attention_mask, frames, height, width, pad_h=0, pad_w=0): x = rearrange(x, 'b (t h w) d -> b d t h w', t=frames, h=height, w=width) if npu_config is None: x = self.body(x) else: x_dtype = x.dtype x = npu_config.run_conv3d(self.body, x, x_dtype) x = x[:, :, :, :height*2-pad_h, :width*2-pad_w] x = rearrange(x, 'b d t h w -> b (t h w) d') attention_mask = rearrange(attention_mask, 'b 1 (t h w) -> b 1 t h w', t=frames, h=height, w=width) attention_mask = attention_mask.repeat_interleave(2, -1).repeat_interleave(2, -2).repeat_interleave(2, -3) attention_mask = attention_mask[:, :, :, :height*2-pad_h, :width*2-pad_w] attention_mask = rearrange(attention_mask, 'b 1 t h w -> b 1 (t h w)') attention_bias = (1 - attention_mask.bool().to(x.dtype)) * -10000.0 return x, attention_bias, attention_mask class Downsample2d(nn.Module): def __init__(self, n_feat, is_video_model=False): super(Downsample2d, self).__init__() self.body = nn.Conv2d(n_feat, 2*n_feat, kernel_size=3, stride=2, padding=1, bias=False) def forward(self, x, attention_mask, frames, height, width, pad_h=0, pad_w=0): # import ipdb;ipdb.set_trace() x = rearrange(x, 'b (t h w) d -> (b t) d h w', t=frames, h=height, w=width) if npu_config is None: x = F.pad(x, (0, pad_w, 0, pad_h)) else: x = npu_config.run_pad_2d(F.pad, x, pad=(0, pad_w, 0, pad_h)) x = self.body(x) x = rearrange(x, '(b t) d h w -> b (t h w) d', t=frames) attention_mask = rearrange(attention_mask, 'b 1 (t h w) -> (b t) 1 h w', t=frames, h=height, w=width) attention_mask = F.pad(attention_mask, (0, pad_w, 0, pad_h)) attention_mask = F.max_pool2d(attention_mask.float(), kernel_size=2, stride=2) attention_mask = rearrange(attention_mask, '(b t) 1 h w -> b 1 (t h w)', t=frames) attention_bias = (1 - attention_mask.bool().to(x.dtype)) * -10000.0 return x, attention_bias, attention_mask class Upsample2d(nn.Module): def __init__(self, n_feat, is_video_model=False): super(Upsample2d, self).__init__() self.body = nn.Conv2d(n_feat, n_feat // 2, kernel_size=3, stride=1, padding=1, bias=False) def forward(self, x, attention_mask, frames, height, width, pad_h=0, pad_w=0): x = rearrange(x, 'b (t h w) d -> (b t) d h w', t=frames, h=height, w=width) x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") x = self.body(x) x = x[:, :, :height*2-pad_h, :width*2-pad_w] x = rearrange(x, '(b t) d h w -> b (t h w) d', t=frames) attention_mask = rearrange(attention_mask, 'b 1 (t h w) -> (b t) 1 h w', t=frames, h=height, w=width) attention_mask = attention_mask.repeat_interleave(2, -1).repeat_interleave(2, -2) attention_mask = attention_mask[:, :, :height*2-pad_h, :width*2-pad_w] attention_mask = rearrange(attention_mask, '(b t) 1 h w -> b 1 (t h w)', t=frames) attention_bias = (1 - attention_mask.bool().to(x.dtype)) * -10000.0 return x, attention_bias, attention_mask class FeedForward_Conv2d(nn.Module): def __init__(self, downsampler, dim, hidden_features, bias=True): super(FeedForward_Conv2d, self).__init__() self.bias = bias self.project_in = nn.Linear(dim, hidden_features, bias=bias) self.act = nn.GELU() self.dwconv = nn.ModuleList([ nn.Conv2d(hidden_features, hidden_features, kernel_size=(5, 5), stride=1, padding=(2, 2), dilation=1, groups=hidden_features, bias=bias), nn.Conv2d(hidden_features, hidden_features, kernel_size=(3, 3), stride=1, padding=(1, 1), dilation=1, groups=hidden_features, bias=bias), nn.Conv2d(hidden_features, hidden_features, kernel_size=(1, 1), stride=1, padding=(0, 0), dilation=1, groups=hidden_features, bias=bias) ]) self.project_out = nn.Linear(hidden_features, dim, bias=bias) def forward(self, x, t, h, w): # import ipdb;ipdb.set_trace() x = self.project_in(x) x = rearrange(x, 'b (t h w) d -> (b t) d h w', t=t, h=h, w=w) x = self.act(x) out = x for module in self.dwconv: out = out + module(x) out = rearrange(out, '(b t) d h w -> b (t h w) d', t=t, h=h, w=w) x = self.project_out(out) return x @maybe_allow_in_graph class BasicTransformerBlock(nn.Module): r""" A basic Transformer block. Parameters: dim (`int`): The number of channels in the input and output. num_attention_heads (`int`): The number of heads to use for multi-head attention. attention_head_dim (`int`): The number of channels in each head. dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. num_embeds_ada_norm (: obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`. attention_bias (: obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter. only_cross_attention (`bool`, *optional*): Whether to use only cross-attention layers. In this case two cross attention layers are used. double_self_attention (`bool`, *optional*): Whether to use two self-attention layers. In this case no cross attention layers are used. upcast_attention (`bool`, *optional*): Whether to upcast the attention computation to float32. This is useful for mixed precision training. norm_elementwise_affine (`bool`, *optional*, defaults to `True`): Whether to use learnable elementwise affine parameters for normalization. norm_type (`str`, *optional*, defaults to `"layer_norm"`): The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`. final_dropout (`bool` *optional*, defaults to False): Whether to apply a final dropout after the last feed-forward layer. attention_type (`str`, *optional*, defaults to `"default"`): The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`. positional_embeddings (`str`, *optional*, defaults to `None`): The type of positional embeddings to apply to. num_positional_embeddings (`int`, *optional*, defaults to `None`): The maximum number of positional embeddings to apply. """ def __init__( self, dim: int, num_attention_heads: int, attention_head_dim: int, dropout=0.0, cross_attention_dim: Optional[int] = None, activation_fn: str = "geglu", num_embeds_ada_norm: Optional[int] = None, attention_bias: bool = False, only_cross_attention: bool = False, double_self_attention: bool = False, upcast_attention: bool = False, norm_elementwise_affine: bool = True, norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen' norm_eps: float = 1e-5, final_dropout: bool = False, attention_type: str = "default", positional_embeddings: Optional[str] = None, num_positional_embeddings: Optional[int] = None, ada_norm_continous_conditioning_embedding_dim: Optional[int] = None, ada_norm_bias: Optional[int] = None, ff_inner_dim: Optional[int] = None, ff_bias: bool = True, attention_out_bias: bool = True, attention_mode: str = "xformers", downsampler: str = None, mlp_ratio: int = 4, use_rope: bool = False, interpolation_scale_thw: Tuple[int] = (1, 1, 1), ): super().__init__() self.only_cross_attention = only_cross_attention # We keep these boolean flags for backward-compatibility. self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero" self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm" self.use_ada_layer_norm_single = norm_type == "ada_norm_single" self.use_layer_norm = norm_type == "layer_norm" self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous" if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None: raise ValueError( f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to" f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}." ) self.norm_type = norm_type self.num_embeds_ada_norm = num_embeds_ada_norm if positional_embeddings and (num_positional_embeddings is None): raise ValueError( "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined." ) if positional_embeddings == "sinusoidal": self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings) else: self.pos_embed = None # Define 3 blocks. Each block has its own normalization layer. # 1. Self-Attn if norm_type == "ada_norm": self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) elif norm_type == "ada_norm_zero": self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm) elif norm_type == "ada_norm_continuous": self.norm1 = AdaLayerNormContinuous( dim, ada_norm_continous_conditioning_embedding_dim, norm_elementwise_affine, norm_eps, ada_norm_bias, "rms_norm", ) else: self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps) self.attn1 = Attention( query_dim=dim, heads=num_attention_heads, dim_head=attention_head_dim, dropout=dropout, bias=attention_bias, cross_attention_dim=cross_attention_dim if only_cross_attention else None, upcast_attention=upcast_attention, out_bias=attention_out_bias, downsampler=downsampler, attention_mode=attention_mode, use_rope=use_rope, interpolation_scale_thw=interpolation_scale_thw, ) # 2. Cross-Attn if cross_attention_dim is not None or double_self_attention: # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during # the second cross attention block. if norm_type == "ada_norm": self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) elif norm_type == "ada_norm_continuous": self.norm2 = AdaLayerNormContinuous( dim, ada_norm_continous_conditioning_embedding_dim, norm_elementwise_affine, norm_eps, ada_norm_bias, "rms_norm", ) else: self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) self.attn2 = Attention( query_dim=dim, cross_attention_dim=cross_attention_dim if not double_self_attention else None, heads=num_attention_heads, dim_head=attention_head_dim, dropout=dropout, bias=attention_bias, upcast_attention=upcast_attention, out_bias=attention_out_bias, downsampler=None, attention_mode=attention_mode, use_rope=False, interpolation_scale_thw=interpolation_scale_thw, ) # is self-attn if encoder_hidden_states is none else: self.norm2 = None self.attn2 = None # 3. Feed-forward if norm_type == "ada_norm_continuous": self.norm3 = AdaLayerNormContinuous( dim, ada_norm_continous_conditioning_embedding_dim, norm_elementwise_affine, norm_eps, ada_norm_bias, "layer_norm", ) elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm", "ada_norm_continuous"]: self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) elif norm_type == "layer_norm_i2vgen": self.norm3 = None self.ff = FeedForward( dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout, inner_dim=ff_inner_dim, bias=ff_bias, ) # 4. Fuser if attention_type == "gated" or attention_type == "gated-text-image": self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim) # 5. Scale-shift for PixArt-Alpha. if norm_type == "ada_norm_single": self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5) # let chunk size default to None self._chunk_size = None self._chunk_dim = 0 def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0): # Sets chunk feed-forward self._chunk_size = chunk_size self._chunk_dim = dim def forward( self, hidden_states: torch.FloatTensor, attention_mask: Optional[torch.FloatTensor] = None, encoder_hidden_states: Optional[torch.FloatTensor] = None, encoder_attention_mask: Optional[torch.FloatTensor] = None, timestep: Optional[torch.LongTensor] = None, cross_attention_kwargs: Dict[str, Any] = None, class_labels: Optional[torch.LongTensor] = None, frame: int = None, height: int = None, width: int = None, added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, ) -> torch.FloatTensor: if cross_attention_kwargs is not None: if cross_attention_kwargs.get("scale", None) is not None: logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.") # Notice that normalization is always applied before the real computation in the following blocks. # 0. Self-Attention batch_size = hidden_states.shape[0] if self.norm_type == "ada_norm": norm_hidden_states = self.norm1(hidden_states, timestep) elif self.norm_type == "ada_norm_zero": norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1( hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype ) elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]: norm_hidden_states = self.norm1(hidden_states) elif self.norm_type == "ada_norm_continuous": norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"]) elif self.norm_type == "ada_norm_single": shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1) ).chunk(6, dim=1) norm_hidden_states = self.norm1(hidden_states) norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa norm_hidden_states = norm_hidden_states.squeeze(1) else: raise ValueError("Incorrect norm used") if self.pos_embed is not None: norm_hidden_states = self.pos_embed(norm_hidden_states) # 1. Prepare GLIGEN inputs cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} gligen_kwargs = cross_attention_kwargs.pop("gligen", None) # 0, -10000 ->(bool) False, True ->(any) True ->(not) False # 0, 0 ->(bool) False, False ->(any) False ->(not) True # assert attention_mask.bool().float().sum() / attention_mask.bool().float().numel() <= 1/16, 'must ~all visible' attn_output = self.attn1( norm_hidden_states, encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, attention_mask=attention_mask, frame=frame, height=height, width=width, **cross_attention_kwargs, ) if self.norm_type == "ada_norm_zero": attn_output = gate_msa.unsqueeze(1) * attn_output elif self.norm_type == "ada_norm_single": attn_output = gate_msa * attn_output hidden_states = attn_output + hidden_states if hidden_states.ndim == 4: hidden_states = hidden_states.squeeze(1) # 1.2 GLIGEN Control if gligen_kwargs is not None: hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"]) # 3. Cross-Attention if self.attn2 is not None: if self.norm_type == "ada_norm": norm_hidden_states = self.norm2(hidden_states, timestep) elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]: norm_hidden_states = self.norm2(hidden_states) elif self.norm_type == "ada_norm_single": # For PixArt norm2 isn't applied here: # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103 norm_hidden_states = hidden_states elif self.norm_type == "ada_norm_continuous": norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"]) else: raise ValueError("Incorrect norm") if self.pos_embed is not None and self.norm_type != "ada_norm_single": norm_hidden_states = self.pos_embed(norm_hidden_states) attn_output = self.attn2( norm_hidden_states, encoder_hidden_states=encoder_hidden_states, attention_mask=encoder_attention_mask, **cross_attention_kwargs, ) hidden_states = attn_output + hidden_states # 4. Feed-forward # i2vgen doesn't have this norm 🤷‍♂️ if self.norm_type == "ada_norm_continuous": norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"]) elif not self.norm_type == "ada_norm_single": norm_hidden_states = self.norm3(hidden_states) if self.norm_type == "ada_norm_zero": norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] if self.norm_type == "ada_norm_single": norm_hidden_states = self.norm2(hidden_states) norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp # if self.downsampler: # ff_output = self.ff(norm_hidden_states, t=frame, h=height, w=width) # else: ff_output = self.ff(norm_hidden_states) if self.norm_type == "ada_norm_zero": ff_output = gate_mlp.unsqueeze(1) * ff_output elif self.norm_type == "ada_norm_single": ff_output = gate_mlp * ff_output hidden_states = ff_output + hidden_states if hidden_states.ndim == 4: hidden_states = hidden_states.squeeze(1) return hidden_states