Open-Sora: Democratizing Efficient Video Production for All
Python
[feature] impl cached pos embedding (#1)
Hongxin Liu8d5278b99a30
diff --git a/opensora/models/layers/blocks.py b/opensora/models/layers/blocks.pyindex 24e6416..b8824ce 100644--- a/opensora/models/layers/blocks.py+++ b/opensora/models/layers/blocks.py@@ -9,7 +9,9 @@# MAE: https://github.com/facebookresearch/mae/blob/main/models_mae.py# --------------------------------------------------------+import functoolsimport math+from typing import Optionalimport numpy as npimport torch@@ -27,7 +29,9 @@ from opensora.acceleration.parallel_states import get_sequence_parallel_groupapprox_gelu = lambda: nn.GELU(approximate="tanh")-def get_layernorm(hidden_size: torch.Tensor, eps: float, affine: bool, use_kernel: bool):+def get_layernorm(+ hidden_size: torch.Tensor, eps: float, affine: bool, use_kernel: bool+):if use_kernel:try:from apex.normalization import FusedLayerNorm@@ -82,7 +86,9 @@ class PatchEmbed3D(nn.Module):self.in_chans = in_chansself.embed_dim = embed_dim- self.proj = nn.Conv3d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)+ 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:@@ -215,7 +221,9 @@ class SeqParallelAttention(Attention):)def forward(self, x: torch.Tensor) -> torch.Tensor:- B, N, C = x.shape # for sequence parallel here, the N is a local sequence length+ B, N, C = (+ x.shape+ ) # for sequence parallel here, the N is a local sequence lengthqkv = self.qkv(x)qkv_shape = (B, N, 3, self.num_heads, self.head_dim)@@ -228,9 +236,21 @@ class SeqParallelAttention(Attention):qkv = all_to_all(qkv, sp_group, scatter_dim=3, gather_dim=1)if self.enable_flashattn:- qkv_permute_shape = (2, 0, 1, 3, 4) # [3, B, N, NUM_HEAD_PER_DEVICE, HEAD_DIM]+ qkv_permute_shape = (+ 2,+ 0,+ 1,+ 3,+ 4,+ ) # [3, B, N, NUM_HEAD_PER_DEVICE, HEAD_DIM]else:- qkv_permute_shape = (2, 0, 3, 1, 4) # [3, B, NUM_HEAD_PER_DEVICE, N, HEAD_DIM]+ qkv_permute_shape = (+ 2,+ 0,+ 3,+ 1,+ 4,+ ) # [3, B, NUM_HEAD_PER_DEVICE, N, HEAD_DIM]qkv = qkv.permute(qkv_permute_shape)q, k, v = qkv.unbind(0)@@ -296,7 +316,9 @@ class MultiHeadCrossAttention(nn.Module):attn_bias = Noneif 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 = 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)@@ -312,7 +334,12 @@ class SeqParallelMultiHeadCrossAttention(MultiHeadCrossAttention):attn_drop=0.0,proj_drop=0.0,):- super().__init__(d_model=d_model, num_heads=num_heads, attn_drop=attn_drop, proj_drop=proj_drop)+ super().__init__(+ d_model=d_model,+ num_heads=num_heads,+ attn_drop=attn_drop,+ proj_drop=proj_drop,+ )def forward(self, x, cond, mask=None):# query/value: img tokens; key: condition; mask: if padding tokens@@ -330,8 +357,12 @@ class SeqParallelMultiHeadCrossAttention(MultiHeadCrossAttention):# apply all_to_all to gather sequence and split attention headsq = all_to_all(q, sp_group, scatter_dim=2, gather_dim=1)- k = split_forward_gather_backward(k, get_sequence_parallel_group(), dim=2, grad_scale="down")- v = split_forward_gather_backward(v, get_sequence_parallel_group(), dim=2, grad_scale="down")+ k = split_forward_gather_backward(+ k, get_sequence_parallel_group(), dim=2, grad_scale="down"+ )+ v = split_forward_gather_backward(+ v, get_sequence_parallel_group(), dim=2, grad_scale="down"+ )q = q.view(1, -1, self.num_heads // sp_size, self.head_dim)k = k.view(1, -1, self.num_heads // sp_size, self.head_dim)@@ -341,7 +372,9 @@ class SeqParallelMultiHeadCrossAttention(MultiHeadCrossAttention):attn_bias = Noneif 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 = xformers.ops.memory_efficient_attention(+ q, k, v, p=self.attn_drop.p, attn_bias=attn_bias+ )# apply all to all to gather back attention heads and scatter sequencex = x.view(B, -1, self.num_heads // sp_size, self.head_dim)@@ -363,7 +396,9 @@ class FinalLayer(nn.Module):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.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))+ self.adaLN_modulation = nn.Sequential(+ nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True)+ )def forward(self, x, c):shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)@@ -381,7 +416,9 @@ class T2IFinalLayer(nn.Module):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.scale_shift_table = nn.Parameter(+ torch.randn(2, hidden_size) / hidden_size**0.5+ )self.out_channels = out_channelsself.d_t = d_tself.d_s = d_s@@ -404,7 +441,9 @@ class T2IFinalLayer(nn.Module):shift, scale = (self.scale_shift_table[None] + t[:, None]).chunk(2, dim=1)x = t2i_modulate(self.norm_final(x), shift, scale)if x_mask is not None:- shift_zero, scale_zero = (self.scale_shift_table[None] + t0[:, None]).chunk(2, dim=1)+ shift_zero, scale_zero = (self.scale_shift_table[None] + t0[:, None]).chunk(+ 2, dim=1+ )x_zero = t2i_modulate(self.norm_final(x), shift_zero, scale_zero)x = self.t_mask_select(x_mask, x, x_zero, T, S)x = self.linear(x)@@ -442,12 +481,18 @@ class TimestepEmbedder(nn.Module):"""# https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.pyhalf = dim // 2- freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half)+ 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)+ embedding = torch.cat(+ [embedding, torch.zeros_like(embedding[:, :1])], dim=-1+ )return embeddingdef forward(self, t, dtype):@@ -466,7 +511,9 @@ class LabelEmbedder(nn.Module):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.embedding_table = nn.Embedding(+ num_classes + use_cfg_embedding, hidden_size+ )self.num_classes = num_classesself.dropout_prob = dropout_prob@@ -494,7 +541,9 @@ class SizeEmbedder(TimestepEmbedder):"""def __init__(self, hidden_size, frequency_embedding_size=256):- super().__init__(hidden_size=hidden_size, frequency_embedding_size=frequency_embedding_size)+ super().__init__(+ hidden_size=hidden_size, frequency_embedding_size=frequency_embedding_size+ )self.mlp = nn.Sequential(nn.Linear(frequency_embedding_size, hidden_size, bias=True),nn.SiLU(),@@ -512,7 +561,9 @@ class SizeEmbedder(TimestepEmbedder):assert s.shape[0] == bsb, dims = s.shape[0], s.shape[1]s = rearrange(s, "b d -> (b d)")- s_freq = self.timestep_embedding(s, self.frequency_embedding_size).to(self.dtype)+ s_freq = self.timestep_embedding(s, self.frequency_embedding_size).to(+ self.dtype+ )s_emb = self.mlp(s_freq)s_emb = rearrange(s_emb, "(b d) d2 -> b (d d2)", b=b, d=dims, d2=self.outdim)return s_emb@@ -527,12 +578,26 @@ 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):+ 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+ 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.register_buffer("y_embedding", nn.Parameter(torch.randn(token_num, in_channels) / in_channels**0.5))self.uncond_prob = uncond_probdef token_drop(self, caption, force_drop_ids=None):@@ -556,13 +621,65 @@ class CaptionEmbedder(nn.Module):return caption+class PositionEmbedding2D(nn.Module):+ def __init__(+ self, dim: int, max_position_embeddings: int = 512, scale: float = 1.0+ ) -> None:+ super().__init__()+ self.dim = dim+ self.max_position_embeddings = max_position_embeddings+ self.scale = scale+ assert dim % 4 == 0, "dim must be divisible by 4"+ half_dim = dim // 2+ inv_freq = 1.0 / (10000 ** (torch.arange(0, half_dim, 2).float() / half_dim))+ self.register_buffer("inv_freq", inv_freq, persistent=False)++ def _get_sin_cos_emb(self, t: torch.Tensor):+ out = torch.einsum("i,d->id", t, self.inv_freq)+ emb_cos = torch.cos(out)+ emb_sin = torch.sin(out)+ return torch.cat((emb_sin, emb_cos), dim=-1)++ @functools.lru_cache(maxsize=128)+ def _get_cached_emb(+ self,+ device: torch.device,+ dtype: torch.dtype,+ h: int,+ w: int,+ base_size: Optional[int] = None,+ ):+ grid_h = torch.arange(h, device=device) / self.scale+ grid_w = torch.arange(w, device=device) / self.scale+ if base_size is not None:+ grid_h *= base_size / h+ grid_w *= base_size / w+ grid_h, grid_w = torch.meshgrid(+ grid_w,+ grid_h,+ indexing="ij",+ ) # here w goes first+ grid_h = grid_h.t().reshape(-1)+ grid_w = grid_w.t().reshape(-1)+ emb_h = self._get_sin_cos_emb(grid_h)+ emb_w = self._get_sin_cos_emb(grid_w)+ return torch.concat([emb_h, emb_w], dim=-1).unsqueeze(0).to(dtype)++ def forward(+ self, x: torch.Tensor, h: int, w: int, base_size: Optional[int] = None+ ) -> torch.Tensor:+ return self._get_cached_emb(x.device, x.dtype, h, w, base_size)++# ===============================================# Sine/Cosine Positional Embedding Functions# ===============================================# https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py-def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0, scale=1.0, base_size=None):+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 widthreturn:@@ -582,7 +699,9 @@ def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=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)+ pos_embed = np.concatenate(+ [np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0+ )return pos_embeddiff --git a/tests/test_pos_emb.py b/tests/test_pos_emb.pynew file mode 100644index 0000000..c409acb--- /dev/null+++ b/tests/test_pos_emb.py@@ -0,0 +1,45 @@+import pytest+import torch++from opensora.models.layers.blocks import PositionEmbedding2D, get_2d_sincos_pos_embed++D = 8+SCALE = 2.0+from torch.testing import assert_close+++def get_spatial_pos_embed(x, hidden_size, h, w, scale, base_size=None):+ pos_embed = get_2d_sincos_pos_embed(+ hidden_size,+ (h, w),+ scale=scale,+ base_size=base_size,+ )+ pos_embed = torch.from_numpy(pos_embed).float().unsqueeze(0).requires_grad_(False)+ return pos_embed.to(device=x.device, dtype=x.dtype)++[email protected]("dtype", [torch.float, torch.float16])[email protected]("device", ["cpu", "cuda"])+def test_pos_emb(dtype, device):+ # just a placeholder to get the device and dtype+ x = torch.empty(1, dtype=dtype, device=device)+ pos_embedder = PositionEmbedding2D(+ D,+ max_position_embeddings=8,+ scale=SCALE,+ ).to(device=device, dtype=dtype)+ output = pos_embedder(x, 8, 7)+ target = get_spatial_pos_embed(x, D, 8, 7, SCALE)+ assert_close(output, target)+ output = pos_embedder(x, 15, 16)+ target = get_spatial_pos_embed(x, D, 15, 16, SCALE)+ assert_close(output, target)+ output = pos_embedder(x, 30, 20, base_size=2)+ target = get_spatial_pos_embed(x, D, 30, 20, SCALE, base_size=2)+ assert_close(output, target)+ # test cache+ output = pos_embedder(x, 30, 20, base_size=2)+ target = get_spatial_pos_embed(x, D, 30, 20, SCALE, base_size=2)+ assert_close(output, target)+ assert pos_embedder._get_cached_emb.cache_info().hits >= 1