Spaces:
Running on Zero
Running on Zero
File size: 9,821 Bytes
257bd28 | 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 | from typing import Optional
import torch
import torch.nn.functional as F
from torch.nn.attention import SDPBackend, sdpa_kernel
from diffusers.models.attention import Attention
class OneHOIAttnProcessor2_0:
"""
Attention processor used typically in processing the SD3-like self-attention projections.
Here, we use the `sdpa_kernel` context manager to enforce the use of CUDNN attention backends.
This is because FlashAttention did not support attention mask, while xformers efficient attention is too slow.
CUDNN attention is ~2x faster than xformers efficient attention, and it supports attention mask.
This processor is used in the `OneHOITransformer2DModel` to process the attention mechanism.
"""
def __init__(self):
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("OneHOIAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
def __call__(
self,
attn: Attention,
hidden_states: torch.FloatTensor,
encoder_hidden_states: torch.FloatTensor = None,
attention_mask: Optional[torch.FloatTensor] = None,
image_rotary_emb: Optional[torch.Tensor] = None,
) -> torch.FloatTensor:
batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
# `sample` projections.
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)
# the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
if encoder_hidden_states is not None:
# `context` projections.
encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
batch_size, -1, attn.heads, head_dim
).transpose(1, 2)
encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
batch_size, -1, attn.heads, head_dim
).transpose(1, 2)
encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
batch_size, -1, attn.heads, head_dim
).transpose(1, 2)
if attn.norm_added_q is not None:
encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
if attn.norm_added_k is not None:
encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
# attention
query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
if image_rotary_emb is not None:
from diffusers.models.embeddings import apply_rotary_emb
query = apply_rotary_emb(query, image_rotary_emb)
key = apply_rotary_emb(key, image_rotary_emb)
# from torch.autograd import profiler
# with profiler.profile(use_cuda=True) as prof:
with sdpa_kernel([SDPBackend.CUDNN_ATTENTION, SDPBackend.FLASH_ATTENTION, SDPBackend.EFFICIENT_ATTENTION]):
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
# print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=5))
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
if encoder_hidden_states is not None:
encoder_hidden_states, hidden_states = (
hidden_states[:, : encoder_hidden_states.shape[1]],
hidden_states[:, encoder_hidden_states.shape[1] :],
)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
return hidden_states, encoder_hidden_states
else:
return hidden_states
class OneHOIAttnProcessor2_0_Contiguous:
"""
Attention processor used typically in processing the SD3-like self-attention projections.
Differently from `OneHOIAttnProcessor2_0`, this set query to be contiguous, to allow compiler optimization.
Here, we use the `sdpa_kernel` context manager to enforce the use of CUDNN attention backends.
This is because FlashAttention did not support attention mask, while xformers efficient attention is too slow.
CUDNN attention is ~2x faster than xformers efficient attention, and it supports attention mask.
This processor is used in the `OneHOITransformer2DModel` to process the attention mechanism.
"""
def __init__(self):
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("FluxAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
def __call__(
self,
attn: Attention,
hidden_states: torch.FloatTensor,
encoder_hidden_states: torch.FloatTensor = None,
attention_mask: Optional[torch.FloatTensor] = None,
image_rotary_emb: Optional[torch.Tensor] = None,
) -> torch.FloatTensor:
batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
# `sample` projections.
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)
# the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
if encoder_hidden_states is not None:
# `context` projections.
encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
batch_size, -1, attn.heads, head_dim
).transpose(1, 2)
encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
batch_size, -1, attn.heads, head_dim
).transpose(1, 2)
encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
batch_size, -1, attn.heads, head_dim
).transpose(1, 2)
if attn.norm_added_q is not None:
encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
if attn.norm_added_k is not None:
encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
# attention
query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
if image_rotary_emb is not None:
from diffusers.models.embeddings import apply_rotary_emb
query = apply_rotary_emb(query, image_rotary_emb)
key = apply_rotary_emb(key, image_rotary_emb)
# from torch.autograd import profiler
# with profiler.profile(use_cuda=True) as prof:
query = query.contiguous()
with sdpa_kernel([SDPBackend.CUDNN_ATTENTION, SDPBackend.FLASH_ATTENTION, SDPBackend.EFFICIENT_ATTENTION]):
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
# print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=5))
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
if encoder_hidden_states is not None:
encoder_hidden_states, hidden_states = (
hidden_states[:, : encoder_hidden_states.shape[1]],
hidden_states[:, encoder_hidden_states.shape[1] :],
)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
return hidden_states, encoder_hidden_states
else:
return hidden_states
|