nodes/kijai-ComfyUI-CogVideoXWrapper
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

support sageattn

Browse Source 
https://github.com/thu-ml/SageAttention
This commit is contained in:
Jukka Seppänen
2024-10-06 20:26:03 +03:00
parent 5e2e11dd0a
commit 49451c4a22
6 changed files with 835 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
cogvideox_fun/fun_pab_transformer_3d.py
+18 -4
View File
@@ -37,6 +37,14 @@ from ..videosys.modules.embeddings import apply_rotary_emb
from ..videosys.core.pab_mgr import enable_pab, if_broadcast_spatial
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
try:
from sageattention import sageattn
SAGEATTN_IS_AVAVILABLE = True
logger.info("Using sageattn")
except:
logger.info("sageattn not found, using sdpa")
SAGEATTN_IS_AVAVILABLE = False
class CogVideoXAttnProcessor2_0:
r"""
Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
@@ -106,9 +114,12 @@ class CogVideoXAttnProcessor2_0:
key[:, :, text_seq_length : emb_len + text_seq_length], image_rotary_emb
)
hidden_states = F.scaled_dot_product_attention(
if SAGEATTN_IS_AVAVILABLE:
hidden_states = sageattn(query, key, value, is_causal=False)
else:
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn_heads * head_dim)
@@ -178,9 +189,12 @@ class FusedCogVideoXAttnProcessor2_0:
if not attn.is_cross_attention:
key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
hidden_states = F.scaled_dot_product_attention(
if SAGEATTN_IS_AVAVILABLE:
hidden_states = sageattn(query, key, value, is_causal=False)
else:
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
cogvideox_fun/transformer_3d.py
+156 -2
View File
@@ -26,15 +26,169 @@ from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.utils import is_torch_version, logging
from diffusers.utils.torch_utils import maybe_allow_in_graph
from diffusers.models.attention import Attention, FeedForward
from diffusers.models.attention_processor import AttentionProcessor, CogVideoXAttnProcessor2_0, FusedCogVideoXAttnProcessor2_0
from diffusers.models.attention_processor import AttentionProcessor#, CogVideoXAttnProcessor2_0, FusedCogVideoXAttnProcessor2_0
from diffusers.models.embeddings import TimestepEmbedding, Timesteps, get_3d_sincos_pos_embed
from diffusers.models.modeling_outputs import Transformer2DModelOutput
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.normalization import AdaLayerNorm, CogVideoXLayerNormZero
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
try:
from sageattention import sageattn
SAGEATTN_IS_AVAVILABLE = True
logger.info("Using sageattn")
except:
logger.info("sageattn not found, using sdpa")
SAGEATTN_IS_AVAVILABLE = False
class CogVideoXAttnProcessor2_0:
r"""
Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
query and key vectors, but does not include spatial normalization.
"""
def __init__(self):
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("CogVideoXAttnProcessor requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
def __call__(
self,
attn: Attention,
hidden_states: torch.Tensor,
encoder_hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
image_rotary_emb: Optional[torch.Tensor] = None,
) -> torch.Tensor:
text_seq_length = encoder_hidden_states.size(1)
hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
query = attn.to_q(hidden_states)
key = attn.to_k(hidden_states)
value = attn.to_v(hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
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)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
if attn.norm_q is not None:
query = attn.norm_q(query)
if attn.norm_k is not None:
key = attn.norm_k(key)
# Apply RoPE if needed
if image_rotary_emb is not None:
from diffusers.models.embeddings import apply_rotary_emb
query[:, :, text_seq_length:] = apply_rotary_emb(query[:, :, text_seq_length:], image_rotary_emb)
if not attn.is_cross_attention:
key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
if SAGEATTN_IS_AVAVILABLE:
hidden_states = sageattn(query, key, value, is_causal=False)
else:
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
encoder_hidden_states, hidden_states = hidden_states.split(
[text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
)
return hidden_states, encoder_hidden_states
class FusedCogVideoXAttnProcessor2_0:
r"""
Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
query and key vectors, but does not include spatial normalization.
"""
def __init__(self):
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("CogVideoXAttnProcessor requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
def __call__(
self,
attn: Attention,
hidden_states: torch.Tensor,
encoder_hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
image_rotary_emb: Optional[torch.Tensor] = None,
) -> torch.Tensor:
text_seq_length = encoder_hidden_states.size(1)
hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
qkv = attn.to_qkv(hidden_states)
split_size = qkv.shape[-1] // 3
query, key, value = torch.split(qkv, split_size, dim=-1)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
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)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
if attn.norm_q is not None:
query = attn.norm_q(query)
if attn.norm_k is not None:
key = attn.norm_k(key)
# Apply RoPE if needed
if image_rotary_emb is not None:
from diffusers.models.embeddings import apply_rotary_emb
query[:, :, text_seq_length:] = apply_rotary_emb(query[:, :, text_seq_length:], image_rotary_emb)
if not attn.is_cross_attention:
key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
if SAGEATTN_IS_AVAVILABLE:
hidden_states = sageattn(query, key, value, is_causal=False)
else:
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
encoder_hidden_states, hidden_states = hidden_states.split(
[text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
)
return hidden_states, encoder_hidden_states
class CogVideoXPatchEmbed(nn.Module):
def __init__(
self,