Spaces:
Paused
Paused
File size: 10,835 Bytes
9d7cf7f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 | import inspect
import math
from typing import Callable, List, Optional, Tuple, Union
import torch
import torch.nn.functional as F
from diffusers.models.attention_processor import Attention
from diffusers.utils import logging
from diffusers.utils.import_utils import is_torch_npu_available, is_xformers_available
from diffusers.utils.torch_utils import is_torch_version, maybe_allow_in_graph
from torch import nn
try:
from flash_attn_interface import flash_attn_func
except Exception as e:
def flash_attn_func(q, k, v):
q = q.permute(0, 2, 1, 3) # (B, H, L, D)
k = k.permute(0, 2, 1, 3)
v = v.permute(0, 2, 1, 3)
if q.shape[1] != k.shape[1]:
repeat_factor = q.shape[1] // k.shape[1]
k = k.repeat_interleave(repeat_factor, dim=1)
v = v.repeat_interleave(repeat_factor, dim=1)
out = torch.nn.functional.scaled_dot_product_attention(q, k, v)
return out.permute(0, 2, 1, 3), None # (B, L, H, D)
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class Tripo2AttnProcessor2_0:
r"""
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). This is
used in the Tripo2DiT model. It applies a s normalization layer and rotary embedding on query and key vector.
"""
def __init__(self):
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 __call__(
self,
attn: Attention,
hidden_states: torch.Tensor,
encoder_hidden_states: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
temb: Optional[torch.Tensor] = None,
image_rotary_emb: Optional[torch.Tensor] = None,
) -> torch.Tensor:
from diffusers.models.embeddings import apply_rotary_emb
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:
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]
)
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)
# NOTE that tripo2 split heads first then split qkv or kv, like .view(..., attn.heads, 3, dim)
# instead of .view(..., 3, attn.heads, dim). So we need to re-split here.
if not attn.is_cross_attention:
qkv = torch.cat((query, key, value), dim=-1)
split_size = qkv.shape[-1] // attn.heads // 3
qkv = qkv.view(batch_size, -1, attn.heads, split_size * 3)
query, key, value = torch.split(qkv, split_size, dim=-1)
else:
kv = torch.cat((key, value), dim=-1)
split_size = kv.shape[-1] // attn.heads // 2
kv = kv.view(batch_size, -1, attn.heads, split_size * 2)
key, value = torch.split(kv, split_size, dim=-1)
head_dim = key.shape[-1]
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:
query = apply_rotary_emb(query, image_rotary_emb)
if not attn.is_cross_attention:
key = apply_rotary_emb(key, image_rotary_emb)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
# with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=True):
hidden_states = flash_attn_func(query.transpose(1, 2), key.transpose(1, 2), value.transpose(1, 2))
if type(hidden_states) == tuple:
hidden_states = hidden_states[0]
# 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 = hidden_states.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
return hidden_states
class FusedTripo2AttnProcessor2_0:
r"""
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0) with fused
projection layers. This is used in the HunyuanDiT model. It applies a s normalization layer and rotary embedding on
query and key vector.
"""
def __init__(self):
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError(
"FusedTripo2AttnProcessor2_0 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: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
temb: Optional[torch.Tensor] = None,
image_rotary_emb: Optional[torch.Tensor] = None,
) -> torch.Tensor:
from diffusers.models.embeddings import apply_rotary_emb
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:
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]
)
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(
1, 2
)
# NOTE that tripo2 split heads first, then split qkv
if encoder_hidden_states is None:
qkv = attn.to_qkv(hidden_states)
split_size = qkv.shape[-1] // attn.heads // 3
qkv = qkv.view(batch_size, -1, attn.heads, split_size * 3)
query, key, value = torch.split(qkv, split_size, dim=-1)
else:
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(
encoder_hidden_states
)
query = attn.to_q(hidden_states)
kv = attn.to_kv(encoder_hidden_states)
split_size = kv.shape[-1] // attn.heads // 2
kv = kv.view(batch_size, -1, attn.heads, split_size * 2)
key, value = torch.split(kv, split_size, dim=-1)
head_dim = key.shape[-1]
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:
query = apply_rotary_emb(query, image_rotary_emb)
if not attn.is_cross_attention:
key = apply_rotary_emb(key, image_rotary_emb)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
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
)
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
return hidden_states
|