Upload hunyuan3d-paintpbr-v2-1/unet/attn_processor.py with huggingface_hub
Browse files
hunyuan3d-paintpbr-v2-1/unet/attn_processor.py
ADDED
@@ -0,0 +1,839 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
|
2 |
+
# except for the third-party components listed below.
|
3 |
+
# Hunyuan 3D does not impose any additional limitations beyond what is outlined
|
4 |
+
# in the repsective licenses of these third-party components.
|
5 |
+
# Users must comply with all terms and conditions of original licenses of these third-party
|
6 |
+
# components and must ensure that the usage of the third party components adheres to
|
7 |
+
# all relevant laws and regulations.
|
8 |
+
|
9 |
+
# For avoidance of doubts, Hunyuan 3D means the large language models and
|
10 |
+
# their software and algorithms, including trained model weights, parameters (including
|
11 |
+
# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
|
12 |
+
# fine-tuning enabling code and other elements of the foregoing made publicly available
|
13 |
+
# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
|
14 |
+
|
15 |
+
import torch
|
16 |
+
import torch.nn as nn
|
17 |
+
import torch.nn.functional as F
|
18 |
+
from typing import Optional, Dict, Tuple, Union, Literal, List, Callable
|
19 |
+
from einops import rearrange
|
20 |
+
from diffusers.utils import deprecate
|
21 |
+
from diffusers.models.attention_processor import Attention, AttnProcessor
|
22 |
+
|
23 |
+
|
24 |
+
class AttnUtils:
|
25 |
+
"""
|
26 |
+
Shared utility functions for attention processing.
|
27 |
+
|
28 |
+
This class provides common operations used across different attention processors
|
29 |
+
to eliminate code duplication and improve maintainability.
|
30 |
+
"""
|
31 |
+
|
32 |
+
@staticmethod
|
33 |
+
def check_pytorch_compatibility():
|
34 |
+
"""
|
35 |
+
Check PyTorch compatibility for scaled_dot_product_attention.
|
36 |
+
|
37 |
+
Raises:
|
38 |
+
ImportError: If PyTorch version doesn't support scaled_dot_product_attention
|
39 |
+
"""
|
40 |
+
if not hasattr(F, "scaled_dot_product_attention"):
|
41 |
+
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
|
42 |
+
|
43 |
+
@staticmethod
|
44 |
+
def handle_deprecation_warning(args, kwargs):
|
45 |
+
"""
|
46 |
+
Handle deprecation warning for the 'scale' argument.
|
47 |
+
|
48 |
+
Args:
|
49 |
+
args: Positional arguments passed to attention processor
|
50 |
+
kwargs: Keyword arguments passed to attention processor
|
51 |
+
"""
|
52 |
+
if len(args) > 0 or kwargs.get("scale", None) is not None:
|
53 |
+
deprecation_message = (
|
54 |
+
"The `scale` argument is deprecated and will be ignored."
|
55 |
+
"Please remove it, as passing it will raise an error in the future."
|
56 |
+
"`scale` should directly be passed while calling the underlying pipeline component"
|
57 |
+
"i.e., via `cross_attention_kwargs`."
|
58 |
+
)
|
59 |
+
deprecate("scale", "1.0.0", deprecation_message)
|
60 |
+
|
61 |
+
@staticmethod
|
62 |
+
def prepare_hidden_states(
|
63 |
+
hidden_states, attn, temb, spatial_norm_attr="spatial_norm", group_norm_attr="group_norm"
|
64 |
+
):
|
65 |
+
"""
|
66 |
+
Common preprocessing of hidden states for attention computation.
|
67 |
+
|
68 |
+
Args:
|
69 |
+
hidden_states: Input hidden states tensor
|
70 |
+
attn: Attention module instance
|
71 |
+
temb: Optional temporal embedding tensor
|
72 |
+
spatial_norm_attr: Attribute name for spatial normalization
|
73 |
+
group_norm_attr: Attribute name for group normalization
|
74 |
+
|
75 |
+
Returns:
|
76 |
+
Tuple of (processed_hidden_states, residual, input_ndim, shape_info)
|
77 |
+
"""
|
78 |
+
residual = hidden_states
|
79 |
+
|
80 |
+
spatial_norm = getattr(attn, spatial_norm_attr, None)
|
81 |
+
if spatial_norm is not None:
|
82 |
+
hidden_states = spatial_norm(hidden_states, temb)
|
83 |
+
|
84 |
+
input_ndim = hidden_states.ndim
|
85 |
+
|
86 |
+
if input_ndim == 4:
|
87 |
+
batch_size, channel, height, width = hidden_states.shape
|
88 |
+
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
89 |
+
else:
|
90 |
+
batch_size, channel, height, width = None, None, None, None
|
91 |
+
|
92 |
+
group_norm = getattr(attn, group_norm_attr, None)
|
93 |
+
if group_norm is not None:
|
94 |
+
hidden_states = group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
|
95 |
+
|
96 |
+
return hidden_states, residual, input_ndim, (batch_size, channel, height, width)
|
97 |
+
|
98 |
+
@staticmethod
|
99 |
+
def prepare_attention_mask(attention_mask, attn, sequence_length, batch_size):
|
100 |
+
"""
|
101 |
+
Prepare attention mask for scaled_dot_product_attention.
|
102 |
+
|
103 |
+
Args:
|
104 |
+
attention_mask: Input attention mask tensor or None
|
105 |
+
attn: Attention module instance
|
106 |
+
sequence_length: Length of the sequence
|
107 |
+
batch_size: Batch size
|
108 |
+
|
109 |
+
Returns:
|
110 |
+
Prepared attention mask tensor reshaped for multi-head attention
|
111 |
+
"""
|
112 |
+
if attention_mask is not None:
|
113 |
+
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
114 |
+
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
|
115 |
+
return attention_mask
|
116 |
+
|
117 |
+
@staticmethod
|
118 |
+
def reshape_qkv_for_attention(tensor, batch_size, attn_heads, head_dim):
|
119 |
+
"""
|
120 |
+
Reshape Q/K/V tensors for multi-head attention computation.
|
121 |
+
|
122 |
+
Args:
|
123 |
+
tensor: Input tensor to reshape
|
124 |
+
batch_size: Batch size
|
125 |
+
attn_heads: Number of attention heads
|
126 |
+
head_dim: Dimension per attention head
|
127 |
+
|
128 |
+
Returns:
|
129 |
+
Reshaped tensor with shape [batch_size, attn_heads, seq_len, head_dim]
|
130 |
+
"""
|
131 |
+
return tensor.view(batch_size, -1, attn_heads, head_dim).transpose(1, 2)
|
132 |
+
|
133 |
+
@staticmethod
|
134 |
+
def apply_norms(query, key, norm_q, norm_k):
|
135 |
+
"""
|
136 |
+
Apply Q/K normalization layers if available.
|
137 |
+
|
138 |
+
Args:
|
139 |
+
query: Query tensor
|
140 |
+
key: Key tensor
|
141 |
+
norm_q: Query normalization layer (optional)
|
142 |
+
norm_k: Key normalization layer (optional)
|
143 |
+
|
144 |
+
Returns:
|
145 |
+
Tuple of (normalized_query, normalized_key)
|
146 |
+
"""
|
147 |
+
if norm_q is not None:
|
148 |
+
query = norm_q(query)
|
149 |
+
if norm_k is not None:
|
150 |
+
key = norm_k(key)
|
151 |
+
return query, key
|
152 |
+
|
153 |
+
@staticmethod
|
154 |
+
def finalize_output(hidden_states, input_ndim, shape_info, attn, residual, to_out):
|
155 |
+
"""
|
156 |
+
Common output processing including projection, dropout, reshaping, and residual connection.
|
157 |
+
|
158 |
+
Args:
|
159 |
+
hidden_states: Processed hidden states from attention
|
160 |
+
input_ndim: Original input tensor dimensions
|
161 |
+
shape_info: Tuple containing original shape information
|
162 |
+
attn: Attention module instance
|
163 |
+
residual: Residual connection tensor
|
164 |
+
to_out: Output projection layers [linear, dropout]
|
165 |
+
|
166 |
+
Returns:
|
167 |
+
Final output tensor after all processing steps
|
168 |
+
"""
|
169 |
+
batch_size, channel, height, width = shape_info
|
170 |
+
|
171 |
+
# Apply output projection and dropout
|
172 |
+
hidden_states = to_out[0](hidden_states)
|
173 |
+
hidden_states = to_out[1](hidden_states)
|
174 |
+
|
175 |
+
# Reshape back if needed
|
176 |
+
if input_ndim == 4:
|
177 |
+
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
178 |
+
|
179 |
+
# Apply residual connection
|
180 |
+
if attn.residual_connection:
|
181 |
+
hidden_states = hidden_states + residual
|
182 |
+
|
183 |
+
# Apply rescaling
|
184 |
+
hidden_states = hidden_states / attn.rescale_output_factor
|
185 |
+
return hidden_states
|
186 |
+
|
187 |
+
|
188 |
+
# Base class for attention processors (eliminating initialization duplication)
|
189 |
+
class BaseAttnProcessor(nn.Module):
|
190 |
+
"""
|
191 |
+
Base class for attention processors with common initialization.
|
192 |
+
|
193 |
+
This base class provides shared parameter initialization and module registration
|
194 |
+
functionality to reduce code duplication across different attention processor types.
|
195 |
+
"""
|
196 |
+
|
197 |
+
def __init__(
|
198 |
+
self,
|
199 |
+
query_dim: int,
|
200 |
+
pbr_setting: List[str] = ["albedo", "mr"],
|
201 |
+
cross_attention_dim: Optional[int] = None,
|
202 |
+
heads: int = 8,
|
203 |
+
kv_heads: Optional[int] = None,
|
204 |
+
dim_head: int = 64,
|
205 |
+
dropout: float = 0.0,
|
206 |
+
bias: bool = False,
|
207 |
+
upcast_attention: bool = False,
|
208 |
+
upcast_softmax: bool = False,
|
209 |
+
cross_attention_norm: Optional[str] = None,
|
210 |
+
cross_attention_norm_num_groups: int = 32,
|
211 |
+
qk_norm: Optional[str] = None,
|
212 |
+
added_kv_proj_dim: Optional[int] = None,
|
213 |
+
added_proj_bias: Optional[bool] = True,
|
214 |
+
norm_num_groups: Optional[int] = None,
|
215 |
+
spatial_norm_dim: Optional[int] = None,
|
216 |
+
out_bias: bool = True,
|
217 |
+
scale_qk: bool = True,
|
218 |
+
only_cross_attention: bool = False,
|
219 |
+
eps: float = 1e-5,
|
220 |
+
rescale_output_factor: float = 1.0,
|
221 |
+
residual_connection: bool = False,
|
222 |
+
_from_deprecated_attn_block: bool = False,
|
223 |
+
processor: Optional["AttnProcessor"] = None,
|
224 |
+
out_dim: int = None,
|
225 |
+
out_context_dim: int = None,
|
226 |
+
context_pre_only=None,
|
227 |
+
pre_only=False,
|
228 |
+
elementwise_affine: bool = True,
|
229 |
+
is_causal: bool = False,
|
230 |
+
**kwargs,
|
231 |
+
):
|
232 |
+
"""
|
233 |
+
Initialize base attention processor with common parameters.
|
234 |
+
|
235 |
+
Args:
|
236 |
+
query_dim: Dimension of query features
|
237 |
+
pbr_setting: List of PBR material types to process (e.g., ["albedo", "mr"])
|
238 |
+
cross_attention_dim: Dimension of cross-attention features (optional)
|
239 |
+
heads: Number of attention heads
|
240 |
+
kv_heads: Number of key-value heads for grouped query attention (optional)
|
241 |
+
dim_head: Dimension per attention head
|
242 |
+
dropout: Dropout rate
|
243 |
+
bias: Whether to use bias in linear projections
|
244 |
+
upcast_attention: Whether to upcast attention computation to float32
|
245 |
+
upcast_softmax: Whether to upcast softmax computation to float32
|
246 |
+
cross_attention_norm: Type of cross-attention normalization (optional)
|
247 |
+
cross_attention_norm_num_groups: Number of groups for cross-attention norm
|
248 |
+
qk_norm: Type of query-key normalization (optional)
|
249 |
+
added_kv_proj_dim: Dimension for additional key-value projections (optional)
|
250 |
+
added_proj_bias: Whether to use bias in additional projections
|
251 |
+
norm_num_groups: Number of groups for normalization (optional)
|
252 |
+
spatial_norm_dim: Dimension for spatial normalization (optional)
|
253 |
+
out_bias: Whether to use bias in output projection
|
254 |
+
scale_qk: Whether to scale query-key products
|
255 |
+
only_cross_attention: Whether to only perform cross-attention
|
256 |
+
eps: Small epsilon value for numerical stability
|
257 |
+
rescale_output_factor: Factor to rescale output values
|
258 |
+
residual_connection: Whether to use residual connections
|
259 |
+
_from_deprecated_attn_block: Flag for deprecated attention blocks
|
260 |
+
processor: Optional attention processor instance
|
261 |
+
out_dim: Output dimension (optional)
|
262 |
+
out_context_dim: Output context dimension (optional)
|
263 |
+
context_pre_only: Whether to only process context in pre-processing
|
264 |
+
pre_only: Whether to only perform pre-processing
|
265 |
+
elementwise_affine: Whether to use element-wise affine transformations
|
266 |
+
is_causal: Whether to use causal attention masking
|
267 |
+
**kwargs: Additional keyword arguments
|
268 |
+
"""
|
269 |
+
super().__init__()
|
270 |
+
AttnUtils.check_pytorch_compatibility()
|
271 |
+
|
272 |
+
# Store common attributes
|
273 |
+
self.pbr_setting = pbr_setting
|
274 |
+
self.n_pbr_tokens = len(self.pbr_setting)
|
275 |
+
self.inner_dim = out_dim if out_dim is not None else dim_head * heads
|
276 |
+
self.inner_kv_dim = self.inner_dim if kv_heads is None else dim_head * kv_heads
|
277 |
+
self.query_dim = query_dim
|
278 |
+
self.use_bias = bias
|
279 |
+
self.is_cross_attention = cross_attention_dim is not None
|
280 |
+
self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
|
281 |
+
self.upcast_attention = upcast_attention
|
282 |
+
self.upcast_softmax = upcast_softmax
|
283 |
+
self.rescale_output_factor = rescale_output_factor
|
284 |
+
self.residual_connection = residual_connection
|
285 |
+
self.dropout = dropout
|
286 |
+
self.fused_projections = False
|
287 |
+
self.out_dim = out_dim if out_dim is not None else query_dim
|
288 |
+
self.out_context_dim = out_context_dim if out_context_dim is not None else query_dim
|
289 |
+
self.context_pre_only = context_pre_only
|
290 |
+
self.pre_only = pre_only
|
291 |
+
self.is_causal = is_causal
|
292 |
+
self._from_deprecated_attn_block = _from_deprecated_attn_block
|
293 |
+
self.scale_qk = scale_qk
|
294 |
+
self.scale = dim_head**-0.5 if self.scale_qk else 1.0
|
295 |
+
self.heads = out_dim // dim_head if out_dim is not None else heads
|
296 |
+
self.sliceable_head_dim = heads
|
297 |
+
self.added_kv_proj_dim = added_kv_proj_dim
|
298 |
+
self.only_cross_attention = only_cross_attention
|
299 |
+
self.added_proj_bias = added_proj_bias
|
300 |
+
|
301 |
+
# Validation
|
302 |
+
if self.added_kv_proj_dim is None and self.only_cross_attention:
|
303 |
+
raise ValueError(
|
304 |
+
"`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None."
|
305 |
+
"Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`."
|
306 |
+
)
|
307 |
+
|
308 |
+
def register_pbr_modules(self, module_types: List[str], **kwargs):
|
309 |
+
"""
|
310 |
+
Generic PBR module registration to eliminate code repetition.
|
311 |
+
|
312 |
+
Dynamically registers PyTorch modules for different PBR material types
|
313 |
+
based on the specified module types and PBR settings.
|
314 |
+
|
315 |
+
Args:
|
316 |
+
module_types: List of module types to register ("qkv", "v_only", "out", "add_kv")
|
317 |
+
**kwargs: Additional arguments for module configuration
|
318 |
+
"""
|
319 |
+
for pbr_token in self.pbr_setting:
|
320 |
+
if pbr_token == "albedo":
|
321 |
+
continue
|
322 |
+
|
323 |
+
for module_type in module_types:
|
324 |
+
if module_type == "qkv":
|
325 |
+
self.register_module(
|
326 |
+
f"to_q_{pbr_token}", nn.Linear(self.query_dim, self.inner_dim, bias=self.use_bias)
|
327 |
+
)
|
328 |
+
self.register_module(
|
329 |
+
f"to_k_{pbr_token}", nn.Linear(self.cross_attention_dim, self.inner_dim, bias=self.use_bias)
|
330 |
+
)
|
331 |
+
self.register_module(
|
332 |
+
f"to_v_{pbr_token}", nn.Linear(self.cross_attention_dim, self.inner_dim, bias=self.use_bias)
|
333 |
+
)
|
334 |
+
elif module_type == "v_only":
|
335 |
+
self.register_module(
|
336 |
+
f"to_v_{pbr_token}", nn.Linear(self.cross_attention_dim, self.inner_dim, bias=self.use_bias)
|
337 |
+
)
|
338 |
+
elif module_type == "out":
|
339 |
+
if not self.pre_only:
|
340 |
+
self.register_module(
|
341 |
+
f"to_out_{pbr_token}",
|
342 |
+
nn.ModuleList(
|
343 |
+
[
|
344 |
+
nn.Linear(self.inner_dim, self.out_dim, bias=kwargs.get("out_bias", True)),
|
345 |
+
nn.Dropout(self.dropout),
|
346 |
+
]
|
347 |
+
),
|
348 |
+
)
|
349 |
+
else:
|
350 |
+
self.register_module(f"to_out_{pbr_token}", None)
|
351 |
+
elif module_type == "add_kv":
|
352 |
+
if self.added_kv_proj_dim is not None:
|
353 |
+
self.register_module(
|
354 |
+
f"add_k_proj_{pbr_token}",
|
355 |
+
nn.Linear(self.added_kv_proj_dim, self.inner_kv_dim, bias=self.added_proj_bias),
|
356 |
+
)
|
357 |
+
self.register_module(
|
358 |
+
f"add_v_proj_{pbr_token}",
|
359 |
+
nn.Linear(self.added_kv_proj_dim, self.inner_kv_dim, bias=self.added_proj_bias),
|
360 |
+
)
|
361 |
+
else:
|
362 |
+
self.register_module(f"add_k_proj_{pbr_token}", None)
|
363 |
+
self.register_module(f"add_v_proj_{pbr_token}", None)
|
364 |
+
|
365 |
+
|
366 |
+
# Rotary Position Embedding utilities (specialized for PoseRoPE)
|
367 |
+
class RotaryEmbedding:
|
368 |
+
"""
|
369 |
+
Rotary position embedding utilities for 3D spatial attention.
|
370 |
+
|
371 |
+
Provides functions to compute and apply rotary position embeddings (RoPE)
|
372 |
+
for 1D, 3D spatial coordinates used in 3D-aware attention mechanisms.
|
373 |
+
"""
|
374 |
+
|
375 |
+
@staticmethod
|
376 |
+
def get_1d_rotary_pos_embed(dim: int, pos: torch.Tensor, theta: float = 10000.0, linear_factor=1.0, ntk_factor=1.0):
|
377 |
+
"""
|
378 |
+
Compute 1D rotary position embeddings.
|
379 |
+
|
380 |
+
Args:
|
381 |
+
dim: Embedding dimension (must be even)
|
382 |
+
pos: Position tensor
|
383 |
+
theta: Base frequency for rotary embeddings
|
384 |
+
linear_factor: Linear scaling factor
|
385 |
+
ntk_factor: NTK (Neural Tangent Kernel) scaling factor
|
386 |
+
|
387 |
+
Returns:
|
388 |
+
Tuple of (cos_embeddings, sin_embeddings)
|
389 |
+
"""
|
390 |
+
assert dim % 2 == 0
|
391 |
+
theta = theta * ntk_factor
|
392 |
+
freqs = (
|
393 |
+
1.0
|
394 |
+
/ (theta ** (torch.arange(0, dim, 2, dtype=pos.dtype, device=pos.device)[: (dim // 2)] / dim))
|
395 |
+
/ linear_factor
|
396 |
+
)
|
397 |
+
freqs = torch.outer(pos, freqs)
|
398 |
+
freqs_cos = freqs.cos().repeat_interleave(2, dim=1).float()
|
399 |
+
freqs_sin = freqs.sin().repeat_interleave(2, dim=1).float()
|
400 |
+
return freqs_cos, freqs_sin
|
401 |
+
|
402 |
+
@staticmethod
|
403 |
+
def get_3d_rotary_pos_embed(position, embed_dim, voxel_resolution, theta: int = 10000):
|
404 |
+
"""
|
405 |
+
Compute 3D rotary position embeddings for spatial coordinates.
|
406 |
+
|
407 |
+
Args:
|
408 |
+
position: 3D position tensor with shape [..., 3]
|
409 |
+
embed_dim: Embedding dimension
|
410 |
+
voxel_resolution: Resolution of the voxel grid
|
411 |
+
theta: Base frequency for rotary embeddings
|
412 |
+
|
413 |
+
Returns:
|
414 |
+
Tuple of (cos_embeddings, sin_embeddings) for 3D positions
|
415 |
+
"""
|
416 |
+
assert position.shape[-1] == 3
|
417 |
+
dim_xy = embed_dim // 8 * 3
|
418 |
+
dim_z = embed_dim // 8 * 2
|
419 |
+
|
420 |
+
grid = torch.arange(voxel_resolution, dtype=torch.float32, device=position.device)
|
421 |
+
freqs_xy = RotaryEmbedding.get_1d_rotary_pos_embed(dim_xy, grid, theta=theta)
|
422 |
+
freqs_z = RotaryEmbedding.get_1d_rotary_pos_embed(dim_z, grid, theta=theta)
|
423 |
+
|
424 |
+
xy_cos, xy_sin = freqs_xy
|
425 |
+
z_cos, z_sin = freqs_z
|
426 |
+
|
427 |
+
embed_flattn = position.view(-1, position.shape[-1])
|
428 |
+
x_cos = xy_cos[embed_flattn[:, 0], :]
|
429 |
+
x_sin = xy_sin[embed_flattn[:, 0], :]
|
430 |
+
y_cos = xy_cos[embed_flattn[:, 1], :]
|
431 |
+
y_sin = xy_sin[embed_flattn[:, 1], :]
|
432 |
+
z_cos = z_cos[embed_flattn[:, 2], :]
|
433 |
+
z_sin = z_sin[embed_flattn[:, 2], :]
|
434 |
+
|
435 |
+
cos = torch.cat((x_cos, y_cos, z_cos), dim=-1)
|
436 |
+
sin = torch.cat((x_sin, y_sin, z_sin), dim=-1)
|
437 |
+
|
438 |
+
cos = cos.view(*position.shape[:-1], embed_dim)
|
439 |
+
sin = sin.view(*position.shape[:-1], embed_dim)
|
440 |
+
return cos, sin
|
441 |
+
|
442 |
+
@staticmethod
|
443 |
+
def apply_rotary_emb(x: torch.Tensor, freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]]):
|
444 |
+
"""
|
445 |
+
Apply rotary position embeddings to input tensor.
|
446 |
+
|
447 |
+
Args:
|
448 |
+
x: Input tensor to apply rotary embeddings to
|
449 |
+
freqs_cis: Tuple of (cos_embeddings, sin_embeddings) or single tensor
|
450 |
+
|
451 |
+
Returns:
|
452 |
+
Tensor with rotary position embeddings applied
|
453 |
+
"""
|
454 |
+
cos, sin = freqs_cis
|
455 |
+
cos, sin = cos.to(x.device), sin.to(x.device)
|
456 |
+
cos = cos.unsqueeze(1)
|
457 |
+
sin = sin.unsqueeze(1)
|
458 |
+
|
459 |
+
x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)
|
460 |
+
x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
|
461 |
+
|
462 |
+
out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
|
463 |
+
return out
|
464 |
+
|
465 |
+
|
466 |
+
# Core attention processing logic (eliminating major duplication)
|
467 |
+
class AttnCore:
|
468 |
+
"""
|
469 |
+
Core attention processing logic shared across processors.
|
470 |
+
|
471 |
+
This class provides the fundamental attention computation pipeline
|
472 |
+
that can be reused across different attention processor implementations.
|
473 |
+
"""
|
474 |
+
|
475 |
+
@staticmethod
|
476 |
+
def process_attention_base(
|
477 |
+
attn: Attention,
|
478 |
+
hidden_states: torch.Tensor,
|
479 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
480 |
+
attention_mask: Optional[torch.Tensor] = None,
|
481 |
+
temb: Optional[torch.Tensor] = None,
|
482 |
+
get_qkv_fn: Callable = None,
|
483 |
+
apply_rope_fn: Optional[Callable] = None,
|
484 |
+
**kwargs,
|
485 |
+
):
|
486 |
+
"""
|
487 |
+
Generic attention processing core shared across different processors.
|
488 |
+
|
489 |
+
This function implements the common attention computation pipeline including:
|
490 |
+
1. Hidden state preprocessing
|
491 |
+
2. Attention mask preparation
|
492 |
+
3. Q/K/V computation via provided function
|
493 |
+
4. Tensor reshaping for multi-head attention
|
494 |
+
5. Optional normalization and RoPE application
|
495 |
+
6. Scaled dot-product attention computation
|
496 |
+
|
497 |
+
Args:
|
498 |
+
attn: Attention module instance
|
499 |
+
hidden_states: Input hidden states tensor
|
500 |
+
encoder_hidden_states: Optional encoder hidden states for cross-attention
|
501 |
+
attention_mask: Optional attention mask tensor
|
502 |
+
temb: Optional temporal embedding tensor
|
503 |
+
get_qkv_fn: Function to compute Q, K, V tensors
|
504 |
+
apply_rope_fn: Optional function to apply rotary position embeddings
|
505 |
+
**kwargs: Additional keyword arguments passed to subfunctions
|
506 |
+
|
507 |
+
Returns:
|
508 |
+
Tuple containing (attention_output, residual, input_ndim, shape_info,
|
509 |
+
batch_size, num_heads, head_dim)
|
510 |
+
"""
|
511 |
+
# Prepare hidden states
|
512 |
+
hidden_states, residual, input_ndim, shape_info = AttnUtils.prepare_hidden_states(hidden_states, attn, temb)
|
513 |
+
|
514 |
+
batch_size, sequence_length, _ = (
|
515 |
+
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
516 |
+
)
|
517 |
+
|
518 |
+
# Prepare attention mask
|
519 |
+
attention_mask = AttnUtils.prepare_attention_mask(attention_mask, attn, sequence_length, batch_size)
|
520 |
+
|
521 |
+
# Get Q, K, V
|
522 |
+
if encoder_hidden_states is None:
|
523 |
+
encoder_hidden_states = hidden_states
|
524 |
+
elif attn.norm_cross:
|
525 |
+
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
|
526 |
+
|
527 |
+
query, key, value = get_qkv_fn(attn, hidden_states, encoder_hidden_states, **kwargs)
|
528 |
+
|
529 |
+
# Reshape for attention
|
530 |
+
inner_dim = key.shape[-1]
|
531 |
+
head_dim = inner_dim // attn.heads
|
532 |
+
|
533 |
+
query = AttnUtils.reshape_qkv_for_attention(query, batch_size, attn.heads, head_dim)
|
534 |
+
key = AttnUtils.reshape_qkv_for_attention(key, batch_size, attn.heads, head_dim)
|
535 |
+
value = AttnUtils.reshape_qkv_for_attention(value, batch_size, attn.heads, value.shape[-1] // attn.heads)
|
536 |
+
|
537 |
+
# Apply normalization
|
538 |
+
query, key = AttnUtils.apply_norms(query, key, getattr(attn, "norm_q", None), getattr(attn, "norm_k", None))
|
539 |
+
|
540 |
+
# Apply RoPE if provided
|
541 |
+
if apply_rope_fn is not None:
|
542 |
+
query, key = apply_rope_fn(query, key, head_dim, **kwargs)
|
543 |
+
|
544 |
+
# Compute attention
|
545 |
+
hidden_states = F.scaled_dot_product_attention(
|
546 |
+
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
|
547 |
+
)
|
548 |
+
|
549 |
+
return hidden_states, residual, input_ndim, shape_info, batch_size, attn.heads, head_dim
|
550 |
+
|
551 |
+
|
552 |
+
# Specific processor implementations (minimal unique code)
|
553 |
+
class PoseRoPEAttnProcessor2_0:
|
554 |
+
"""
|
555 |
+
Attention processor with Rotary Position Encoding (RoPE) for 3D spatial awareness.
|
556 |
+
|
557 |
+
This processor extends standard attention with 3D rotary position embeddings
|
558 |
+
to provide spatial awareness for 3D scene understanding tasks.
|
559 |
+
"""
|
560 |
+
|
561 |
+
def __init__(self):
|
562 |
+
"""Initialize the RoPE attention processor."""
|
563 |
+
AttnUtils.check_pytorch_compatibility()
|
564 |
+
|
565 |
+
def __call__(
|
566 |
+
self,
|
567 |
+
attn: Attention,
|
568 |
+
hidden_states: torch.Tensor,
|
569 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
570 |
+
attention_mask: Optional[torch.Tensor] = None,
|
571 |
+
position_indices: Dict = None,
|
572 |
+
temb: Optional[torch.Tensor] = None,
|
573 |
+
n_pbrs=1,
|
574 |
+
*args,
|
575 |
+
**kwargs,
|
576 |
+
) -> torch.Tensor:
|
577 |
+
"""
|
578 |
+
Apply RoPE-enhanced attention computation.
|
579 |
+
|
580 |
+
Args:
|
581 |
+
attn: Attention module instance
|
582 |
+
hidden_states: Input hidden states tensor
|
583 |
+
encoder_hidden_states: Optional encoder hidden states for cross-attention
|
584 |
+
attention_mask: Optional attention mask tensor
|
585 |
+
position_indices: Dictionary containing 3D position information for RoPE
|
586 |
+
temb: Optional temporal embedding tensor
|
587 |
+
n_pbrs: Number of PBR material types
|
588 |
+
*args: Additional positional arguments
|
589 |
+
**kwargs: Additional keyword arguments
|
590 |
+
|
591 |
+
Returns:
|
592 |
+
Attention output tensor with applied rotary position encodings
|
593 |
+
"""
|
594 |
+
AttnUtils.handle_deprecation_warning(args, kwargs)
|
595 |
+
|
596 |
+
def get_qkv(attn, hidden_states, encoder_hidden_states, **kwargs):
|
597 |
+
return attn.to_q(hidden_states), attn.to_k(encoder_hidden_states), attn.to_v(encoder_hidden_states)
|
598 |
+
|
599 |
+
def apply_rope(query, key, head_dim, **kwargs):
|
600 |
+
if position_indices is not None:
|
601 |
+
if head_dim in position_indices:
|
602 |
+
image_rotary_emb = position_indices[head_dim]
|
603 |
+
else:
|
604 |
+
image_rotary_emb = RotaryEmbedding.get_3d_rotary_pos_embed(
|
605 |
+
rearrange(
|
606 |
+
position_indices["voxel_indices"].unsqueeze(1).repeat(1, n_pbrs, 1, 1),
|
607 |
+
"b n_pbrs l c -> (b n_pbrs) l c",
|
608 |
+
),
|
609 |
+
head_dim,
|
610 |
+
voxel_resolution=position_indices["voxel_resolution"],
|
611 |
+
)
|
612 |
+
position_indices[head_dim] = image_rotary_emb
|
613 |
+
|
614 |
+
query = RotaryEmbedding.apply_rotary_emb(query, image_rotary_emb)
|
615 |
+
key = RotaryEmbedding.apply_rotary_emb(key, image_rotary_emb)
|
616 |
+
return query, key
|
617 |
+
|
618 |
+
# Core attention processing
|
619 |
+
hidden_states, residual, input_ndim, shape_info, batch_size, heads, head_dim = AttnCore.process_attention_base(
|
620 |
+
attn,
|
621 |
+
hidden_states,
|
622 |
+
encoder_hidden_states,
|
623 |
+
attention_mask,
|
624 |
+
temb,
|
625 |
+
get_qkv_fn=get_qkv,
|
626 |
+
apply_rope_fn=apply_rope,
|
627 |
+
position_indices=position_indices,
|
628 |
+
n_pbrs=n_pbrs,
|
629 |
+
)
|
630 |
+
|
631 |
+
# Finalize output
|
632 |
+
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, heads * head_dim)
|
633 |
+
hidden_states = hidden_states.to(hidden_states.dtype)
|
634 |
+
|
635 |
+
return AttnUtils.finalize_output(hidden_states, input_ndim, shape_info, attn, residual, attn.to_out)
|
636 |
+
|
637 |
+
|
638 |
+
class SelfAttnProcessor2_0(BaseAttnProcessor):
|
639 |
+
"""
|
640 |
+
Self-attention processor with PBR (Physically Based Rendering) material support.
|
641 |
+
|
642 |
+
This processor handles multiple PBR material types (e.g., albedo, metallic-roughness)
|
643 |
+
with separate attention computation paths for each material type.
|
644 |
+
"""
|
645 |
+
|
646 |
+
def __init__(self, **kwargs):
|
647 |
+
"""
|
648 |
+
Initialize self-attention processor with PBR support.
|
649 |
+
|
650 |
+
Args:
|
651 |
+
**kwargs: Arguments passed to BaseAttnProcessor initialization
|
652 |
+
"""
|
653 |
+
super().__init__(**kwargs)
|
654 |
+
self.register_pbr_modules(["qkv", "out", "add_kv"], **kwargs)
|
655 |
+
|
656 |
+
def process_single(
|
657 |
+
self,
|
658 |
+
attn: Attention,
|
659 |
+
hidden_states: torch.Tensor,
|
660 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
661 |
+
attention_mask: Optional[torch.Tensor] = None,
|
662 |
+
temb: Optional[torch.Tensor] = None,
|
663 |
+
token: Literal["albedo", "mr"] = "albedo",
|
664 |
+
multiple_devices=False,
|
665 |
+
*args,
|
666 |
+
**kwargs,
|
667 |
+
):
|
668 |
+
"""
|
669 |
+
Process attention for a single PBR material type.
|
670 |
+
|
671 |
+
Args:
|
672 |
+
attn: Attention module instance
|
673 |
+
hidden_states: Input hidden states tensor
|
674 |
+
encoder_hidden_states: Optional encoder hidden states for cross-attention
|
675 |
+
attention_mask: Optional attention mask tensor
|
676 |
+
temb: Optional temporal embedding tensor
|
677 |
+
token: PBR material type to process ("albedo", "mr", etc.)
|
678 |
+
multiple_devices: Whether to use multiple GPU devices
|
679 |
+
*args: Additional positional arguments
|
680 |
+
**kwargs: Additional keyword arguments
|
681 |
+
|
682 |
+
Returns:
|
683 |
+
Processed attention output for the specified PBR material type
|
684 |
+
"""
|
685 |
+
target = attn if token == "albedo" else attn.processor
|
686 |
+
token_suffix = "" if token == "albedo" else "_" + token
|
687 |
+
|
688 |
+
# Device management (if needed)
|
689 |
+
if multiple_devices:
|
690 |
+
device = torch.device("cuda:0") if token == "albedo" else torch.device("cuda:1")
|
691 |
+
for attr in [f"to_q{token_suffix}", f"to_k{token_suffix}", f"to_v{token_suffix}", f"to_out{token_suffix}"]:
|
692 |
+
getattr(target, attr).to(device)
|
693 |
+
|
694 |
+
def get_qkv(attn, hidden_states, encoder_hidden_states, **kwargs):
|
695 |
+
return (
|
696 |
+
getattr(target, f"to_q{token_suffix}")(hidden_states),
|
697 |
+
getattr(target, f"to_k{token_suffix}")(encoder_hidden_states),
|
698 |
+
getattr(target, f"to_v{token_suffix}")(encoder_hidden_states),
|
699 |
+
)
|
700 |
+
|
701 |
+
# Core processing using shared logic
|
702 |
+
hidden_states, residual, input_ndim, shape_info, batch_size, heads, head_dim = AttnCore.process_attention_base(
|
703 |
+
attn, hidden_states, encoder_hidden_states, attention_mask, temb, get_qkv_fn=get_qkv
|
704 |
+
)
|
705 |
+
|
706 |
+
# Finalize
|
707 |
+
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, heads * head_dim)
|
708 |
+
hidden_states = hidden_states.to(hidden_states.dtype)
|
709 |
+
|
710 |
+
return AttnUtils.finalize_output(
|
711 |
+
hidden_states, input_ndim, shape_info, attn, residual, getattr(target, f"to_out{token_suffix}")
|
712 |
+
)
|
713 |
+
|
714 |
+
def __call__(
|
715 |
+
self,
|
716 |
+
attn: Attention,
|
717 |
+
hidden_states: torch.Tensor,
|
718 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
719 |
+
attention_mask: Optional[torch.Tensor] = None,
|
720 |
+
temb: Optional[torch.Tensor] = None,
|
721 |
+
*args,
|
722 |
+
**kwargs,
|
723 |
+
) -> torch.Tensor:
|
724 |
+
"""
|
725 |
+
Apply self-attention with PBR material processing.
|
726 |
+
|
727 |
+
Processes multiple PBR material types sequentially, applying attention
|
728 |
+
computation for each material type separately and combining results.
|
729 |
+
|
730 |
+
Args:
|
731 |
+
attn: Attention module instance
|
732 |
+
hidden_states: Input hidden states tensor with PBR dimension
|
733 |
+
encoder_hidden_states: Optional encoder hidden states for cross-attention
|
734 |
+
attention_mask: Optional attention mask tensor
|
735 |
+
temb: Optional temporal embedding tensor
|
736 |
+
*args: Additional positional arguments
|
737 |
+
**kwargs: Additional keyword arguments
|
738 |
+
|
739 |
+
Returns:
|
740 |
+
Combined attention output for all PBR material types
|
741 |
+
"""
|
742 |
+
AttnUtils.handle_deprecation_warning(args, kwargs)
|
743 |
+
|
744 |
+
B = hidden_states.size(0)
|
745 |
+
pbr_hidden_states = torch.split(hidden_states, 1, dim=1)
|
746 |
+
|
747 |
+
# Process each PBR setting
|
748 |
+
results = []
|
749 |
+
for token, pbr_hs in zip(self.pbr_setting, pbr_hidden_states):
|
750 |
+
processed_hs = rearrange(pbr_hs, "b n_pbrs n l c -> (b n_pbrs n) l c").to("cuda:0")
|
751 |
+
result = self.process_single(attn, processed_hs, None, attention_mask, temb, token, False)
|
752 |
+
results.append(result)
|
753 |
+
|
754 |
+
outputs = [rearrange(result, "(b n_pbrs n) l c -> b n_pbrs n l c", b=B, n_pbrs=1) for result in results]
|
755 |
+
return torch.cat(outputs, dim=1)
|
756 |
+
|
757 |
+
|
758 |
+
class RefAttnProcessor2_0(BaseAttnProcessor):
|
759 |
+
"""
|
760 |
+
Reference attention processor with shared value computation across PBR materials.
|
761 |
+
|
762 |
+
This processor computes query and key once, but uses separate value projections
|
763 |
+
for different PBR material types, enabling efficient multi-material processing.
|
764 |
+
"""
|
765 |
+
|
766 |
+
def __init__(self, **kwargs):
|
767 |
+
"""
|
768 |
+
Initialize reference attention processor.
|
769 |
+
|
770 |
+
Args:
|
771 |
+
**kwargs: Arguments passed to BaseAttnProcessor initialization
|
772 |
+
"""
|
773 |
+
super().__init__(**kwargs)
|
774 |
+
self.pbr_settings = self.pbr_setting # Alias for compatibility
|
775 |
+
self.register_pbr_modules(["v_only", "out"], **kwargs)
|
776 |
+
|
777 |
+
def __call__(
|
778 |
+
self,
|
779 |
+
attn: Attention,
|
780 |
+
hidden_states: torch.Tensor,
|
781 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
782 |
+
attention_mask: Optional[torch.Tensor] = None,
|
783 |
+
temb: Optional[torch.Tensor] = None,
|
784 |
+
*args,
|
785 |
+
**kwargs,
|
786 |
+
) -> torch.Tensor:
|
787 |
+
"""
|
788 |
+
Apply reference attention with shared Q/K and separate V projections.
|
789 |
+
|
790 |
+
This method computes query and key tensors once and reuses them across
|
791 |
+
all PBR material types, while using separate value projections for each
|
792 |
+
material type to maintain material-specific information.
|
793 |
+
|
794 |
+
Args:
|
795 |
+
attn: Attention module instance
|
796 |
+
hidden_states: Input hidden states tensor
|
797 |
+
encoder_hidden_states: Optional encoder hidden states for cross-attention
|
798 |
+
attention_mask: Optional attention mask tensor
|
799 |
+
temb: Optional temporal embedding tensor
|
800 |
+
*args: Additional positional arguments
|
801 |
+
**kwargs: Additional keyword arguments
|
802 |
+
|
803 |
+
Returns:
|
804 |
+
Stacked attention output for all PBR material types
|
805 |
+
"""
|
806 |
+
AttnUtils.handle_deprecation_warning(args, kwargs)
|
807 |
+
|
808 |
+
def get_qkv(attn, hidden_states, encoder_hidden_states, **kwargs):
|
809 |
+
query = attn.to_q(hidden_states)
|
810 |
+
key = attn.to_k(encoder_hidden_states)
|
811 |
+
|
812 |
+
# Concatenate values from all PBR settings
|
813 |
+
value_list = [attn.to_v(encoder_hidden_states)]
|
814 |
+
for token in ["_" + token for token in self.pbr_settings if token != "albedo"]:
|
815 |
+
value_list.append(getattr(attn.processor, f"to_v{token}")(encoder_hidden_states))
|
816 |
+
value = torch.cat(value_list, dim=-1)
|
817 |
+
|
818 |
+
return query, key, value
|
819 |
+
|
820 |
+
# Core processing
|
821 |
+
hidden_states, residual, input_ndim, shape_info, batch_size, heads, head_dim = AttnCore.process_attention_base(
|
822 |
+
attn, hidden_states, encoder_hidden_states, attention_mask, temb, get_qkv_fn=get_qkv
|
823 |
+
)
|
824 |
+
|
825 |
+
# Split and process each PBR setting output
|
826 |
+
hidden_states_list = torch.split(hidden_states, head_dim, dim=-1)
|
827 |
+
output_hidden_states_list = []
|
828 |
+
|
829 |
+
for i, hs in enumerate(hidden_states_list):
|
830 |
+
hs = hs.transpose(1, 2).reshape(batch_size, -1, heads * head_dim).to(hs.dtype)
|
831 |
+
token_suffix = "_" + self.pbr_settings[i] if self.pbr_settings[i] != "albedo" else ""
|
832 |
+
target = attn if self.pbr_settings[i] == "albedo" else attn.processor
|
833 |
+
|
834 |
+
hs = AttnUtils.finalize_output(
|
835 |
+
hs, input_ndim, shape_info, attn, residual, getattr(target, f"to_out{token_suffix}")
|
836 |
+
)
|
837 |
+
output_hidden_states_list.append(hs)
|
838 |
+
|
839 |
+
return torch.stack(output_hidden_states_list, dim=1)
|