Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

[hotfix] fix attn impl (#17)

Hongxin Liubdb9b7ccbfa9
diff --git a/opensora/models/layers/blocks.py b/opensora/models/layers/blocks.py
index f7c67b1..0a9ebb1 100644
--- a/opensora/models/layers/blocks.py
+++ b/opensora/models/layers/blocks.py
@@ -23,13 +23,18 @@ import xformers.ops
from einops import rearrange
from timm.models.vision_transformer import Mlp
-from opensora.acceleration.communications import all_to_all, split_forward_gather_backward
+from opensora.acceleration.communications import (
+ all_to_all,
+ split_forward_gather_backward,
+)
from opensora.acceleration.parallel_states import get_sequence_parallel_group
approx_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
@@ -84,7 +89,9 @@ class PatchEmbed3D(nn.Module):
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)
+ 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:
@@ -147,6 +154,8 @@ class Attention(nn.Module):
def forward(self, x: torch.Tensor) -> torch.Tensor:
B, N, C = x.shape
+ # flash attn is not memory efficient for small sequences, this is empirical
+ enable_flashattn = self.enable_flashattn and (N > B)
qkv = self.qkv(x)
qkv_shape = (B, N, 3, self.num_heads, self.head_dim)
@@ -155,14 +164,14 @@ class Attention(nn.Module):
if self.rope:
q = self.rotary_emb(q)
k = self.rotary_emb(k)
- if self.enable_flashattn:
+ if enable_flashattn:
# (B, #heads, N, #dim) -> (B, N, #heads, #dim)
q = q.permute(0, 2, 1, 3)
k = k.permute(0, 2, 1, 3)
v = v.permute(0, 2, 1, 3)
q, k = self.q_norm(q), self.k_norm(k)
- if self.enable_flashattn:
+ if enable_flashattn:
from flash_attn import flash_attn_func
x = flash_attn_func(
@@ -183,7 +192,7 @@ class Attention(nn.Module):
x = attn @ v
x_output_shape = (B, N, C)
- if not self.enable_flashattn:
+ if not enable_flashattn:
x = x.transpose(1, 2)
x = x.reshape(x_output_shape)
x = self.proj(x)
@@ -217,7 +226,9 @@ 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 length
qkv = self.qkv(x)
qkv_shape = (B, N, 3, self.num_heads, self.head_dim)
@@ -310,7 +321,9 @@ class MultiHeadCrossAttention(nn.Module):
attn_bias = None
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 = 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)
@@ -349,8 +362,12 @@ class SeqParallelMultiHeadCrossAttention(MultiHeadCrossAttention):
# apply all_to_all to gather sequence and split attention heads
q = 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)
@@ -360,7 +377,9 @@ class SeqParallelMultiHeadCrossAttention(MultiHeadCrossAttention):
attn_bias = None
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 = 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 sequence
x = x.view(B, -1, self.num_heads // sp_size, self.head_dim)
@@ -382,7 +401,9 @@ 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)
@@ -400,7 +421,9 @@ 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_channels
self.d_t = d_t
self.d_s = d_s
@@ -423,7 +446,9 @@ 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)
@@ -461,12 +486,18 @@ class TimestepEmbedder(nn.Module):
"""
# 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 = 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 embedding
def forward(self, t, dtype):
@@ -485,7 +516,9 @@ 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_classes
self.dropout_prob = dropout_prob
@@ -513,7 +546,9 @@ 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(),
@@ -531,7 +566,9 @@ class SizeEmbedder(TimestepEmbedder):
assert s.shape[0] == bs
b, 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
@@ -631,7 +668,12 @@ class PositionEmbedding2D(nn.Module):
return torch.concat([emb_h, emb_w], dim=-1).unsqueeze(0).to(dtype)
def forward(
- self, x: torch.Tensor, h: int, w: int, scale: Optional[float] = 1.0, base_size: Optional[int] = None
+ self,
+ x: torch.Tensor,
+ h: int,
+ w: int,
+ scale: Optional[float] = 1.0,
+ base_size: Optional[int] = None,
) -> torch.Tensor:
return self._get_cached_emb(x.device, x.dtype, h, w, scale, base_size)
@@ -642,7 +684,9 @@ class PositionEmbedding2D(nn.Module):
# 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 width
return:
@@ -662,7 +706,9 @@ 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_embed
diff --git a/tests/test_attn.py b/tests/test_attn.py
new file mode 100644
index 0000000..22444a1
--- /dev/null
+++ b/tests/test_attn.py
@@ -0,0 +1,36 @@
+import torch
+from colossalai.accelerator import get_accelerator
+from colossalai.utils import get_current_device
+from rotary_embedding_torch import RotaryEmbedding
+
+from opensora.models.layers.blocks import Attention
+
+# B, S, H = 7488, 1, 1152
+# B, S, H = 32, 234, 1152
+B, S, H = 128, 32, 1152
+N, D = 16, 72
+
+
+def run_attn(enable_flashattn: bool):
+ get_accelerator().reset_peak_memory_stats()
+ rope = RotaryEmbedding(D).to(device=get_current_device(), dtype=torch.bfloat16)
+ attn = Attention(
+ H,
+ N,
+ qkv_bias=True,
+ rope=rope.rotate_queries_or_keys,
+ enable_flashattn=enable_flashattn,
+ ).to(device=get_current_device(), dtype=torch.bfloat16)
+ x = torch.randn(
+ B, S, H, device=get_current_device(), dtype=torch.bfloat16
+ ).requires_grad_()
+ y = attn(x)
+ y.mean().backward()
+ print(f"Peak memory: {get_accelerator().max_memory_allocated() / 1024**2:.2f} MB")
+
+
+if __name__ == "__main__":
+ print("Use flashattn")
+ run_attn(True)
+ print("No flashattn")
+ run_attn(False)