Upload 6 files
Browse files- models/___init__.py +0 -0
- models/configuration_chatglm.py +61 -0
- models/controlnet.py +887 -0
- models/modeling_chatglm.py +1298 -0
- models/tokenization_chatglm.py +300 -0
- models/unet_2d_condition.py +1318 -0
models/___init__.py
ADDED
File without changes
|
models/configuration_chatglm.py
ADDED
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import PretrainedConfig
|
2 |
+
|
3 |
+
|
4 |
+
class ChatGLMConfig(PretrainedConfig):
|
5 |
+
model_type = "chatglm"
|
6 |
+
def __init__(
|
7 |
+
self,
|
8 |
+
num_layers=28,
|
9 |
+
padded_vocab_size=65024,
|
10 |
+
hidden_size=4096,
|
11 |
+
ffn_hidden_size=13696,
|
12 |
+
kv_channels=128,
|
13 |
+
num_attention_heads=32,
|
14 |
+
seq_length=2048,
|
15 |
+
hidden_dropout=0.0,
|
16 |
+
classifier_dropout=None,
|
17 |
+
attention_dropout=0.0,
|
18 |
+
layernorm_epsilon=1e-5,
|
19 |
+
rmsnorm=True,
|
20 |
+
apply_residual_connection_post_layernorm=False,
|
21 |
+
post_layer_norm=True,
|
22 |
+
add_bias_linear=False,
|
23 |
+
add_qkv_bias=False,
|
24 |
+
bias_dropout_fusion=True,
|
25 |
+
multi_query_attention=False,
|
26 |
+
multi_query_group_num=1,
|
27 |
+
apply_query_key_layer_scaling=True,
|
28 |
+
attention_softmax_in_fp32=True,
|
29 |
+
fp32_residual_connection=False,
|
30 |
+
quantization_bit=0,
|
31 |
+
pre_seq_len=None,
|
32 |
+
prefix_projection=False,
|
33 |
+
**kwargs
|
34 |
+
):
|
35 |
+
self.num_layers = num_layers
|
36 |
+
self.vocab_size = padded_vocab_size
|
37 |
+
self.padded_vocab_size = padded_vocab_size
|
38 |
+
self.hidden_size = hidden_size
|
39 |
+
self.ffn_hidden_size = ffn_hidden_size
|
40 |
+
self.kv_channels = kv_channels
|
41 |
+
self.num_attention_heads = num_attention_heads
|
42 |
+
self.seq_length = seq_length
|
43 |
+
self.hidden_dropout = hidden_dropout
|
44 |
+
self.classifier_dropout = classifier_dropout
|
45 |
+
self.attention_dropout = attention_dropout
|
46 |
+
self.layernorm_epsilon = layernorm_epsilon
|
47 |
+
self.rmsnorm = rmsnorm
|
48 |
+
self.apply_residual_connection_post_layernorm = apply_residual_connection_post_layernorm
|
49 |
+
self.post_layer_norm = post_layer_norm
|
50 |
+
self.add_bias_linear = add_bias_linear
|
51 |
+
self.add_qkv_bias = add_qkv_bias
|
52 |
+
self.bias_dropout_fusion = bias_dropout_fusion
|
53 |
+
self.multi_query_attention = multi_query_attention
|
54 |
+
self.multi_query_group_num = multi_query_group_num
|
55 |
+
self.apply_query_key_layer_scaling = apply_query_key_layer_scaling
|
56 |
+
self.attention_softmax_in_fp32 = attention_softmax_in_fp32
|
57 |
+
self.fp32_residual_connection = fp32_residual_connection
|
58 |
+
self.quantization_bit = quantization_bit
|
59 |
+
self.pre_seq_len = pre_seq_len
|
60 |
+
self.prefix_projection = prefix_projection
|
61 |
+
super().__init__(**kwargs)
|
models/controlnet.py
ADDED
@@ -0,0 +1,887 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2024 The HuggingFace Team. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
from dataclasses import dataclass
|
15 |
+
from typing import Any, Dict, List, Optional, Tuple, Union
|
16 |
+
|
17 |
+
import torch
|
18 |
+
from torch import nn
|
19 |
+
from torch.nn import functional as F
|
20 |
+
|
21 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
22 |
+
from diffusers.loaders.single_file_model import FromOriginalModelMixin
|
23 |
+
from diffusers.utils import BaseOutput, logging
|
24 |
+
from diffusers.models.attention_processor import (
|
25 |
+
ADDED_KV_ATTENTION_PROCESSORS,
|
26 |
+
CROSS_ATTENTION_PROCESSORS,
|
27 |
+
AttentionProcessor,
|
28 |
+
AttnAddedKVProcessor,
|
29 |
+
AttnProcessor,
|
30 |
+
)
|
31 |
+
from diffusers.models.embeddings import TextImageProjection, TextImageTimeEmbedding, TextTimeEmbedding, TimestepEmbedding, Timesteps
|
32 |
+
from diffusers.models.modeling_utils import ModelMixin
|
33 |
+
|
34 |
+
try:
|
35 |
+
from diffusers.unets.unet_2d_blocks import (
|
36 |
+
CrossAttnDownBlock2D,
|
37 |
+
DownBlock2D,
|
38 |
+
UNetMidBlock2D,
|
39 |
+
UNetMidBlock2DCrossAttn,
|
40 |
+
get_down_block,
|
41 |
+
)
|
42 |
+
from diffusers.unets.unet_2d_condition import UNet2DConditionModel
|
43 |
+
except:
|
44 |
+
from diffusers.models.unets.unet_2d_blocks import (
|
45 |
+
CrossAttnDownBlock2D,
|
46 |
+
DownBlock2D,
|
47 |
+
UNetMidBlock2D,
|
48 |
+
UNetMidBlock2DCrossAttn,
|
49 |
+
get_down_block,
|
50 |
+
)
|
51 |
+
from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel
|
52 |
+
|
53 |
+
|
54 |
+
|
55 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
56 |
+
|
57 |
+
|
58 |
+
@dataclass
|
59 |
+
class ControlNetOutput(BaseOutput):
|
60 |
+
"""
|
61 |
+
The output of [`ControlNetModel`].
|
62 |
+
|
63 |
+
Args:
|
64 |
+
down_block_res_samples (`tuple[torch.Tensor]`):
|
65 |
+
A tuple of downsample activations at different resolutions for each downsampling block. Each tensor should
|
66 |
+
be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
|
67 |
+
used to condition the original UNet's downsampling activations.
|
68 |
+
mid_down_block_re_sample (`torch.Tensor`):
|
69 |
+
The activation of the middle block (the lowest sample resolution). Each tensor should be of shape
|
70 |
+
`(batch_size, channel * lowest_resolution, height // lowest_resolution, width // lowest_resolution)`.
|
71 |
+
Output can be used to condition the original UNet's middle block activation.
|
72 |
+
"""
|
73 |
+
|
74 |
+
down_block_res_samples: Tuple[torch.Tensor]
|
75 |
+
mid_block_res_sample: torch.Tensor
|
76 |
+
|
77 |
+
|
78 |
+
class ControlNetConditioningEmbedding(nn.Module):
|
79 |
+
"""
|
80 |
+
Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
|
81 |
+
[11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
|
82 |
+
training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
|
83 |
+
convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
|
84 |
+
(activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
|
85 |
+
model) to encode image-space conditions ... into feature maps ..."
|
86 |
+
"""
|
87 |
+
|
88 |
+
def __init__(
|
89 |
+
self,
|
90 |
+
conditioning_embedding_channels: int,
|
91 |
+
conditioning_channels: int = 3,
|
92 |
+
block_out_channels: Tuple[int, ...] = (16, 32, 96, 256),
|
93 |
+
):
|
94 |
+
super().__init__()
|
95 |
+
|
96 |
+
self.conv_in = nn.Conv2d(conditioning_channels, block_out_channels[0], kernel_size=3, padding=1)
|
97 |
+
|
98 |
+
self.blocks = nn.ModuleList([])
|
99 |
+
|
100 |
+
for i in range(len(block_out_channels) - 1):
|
101 |
+
channel_in = block_out_channels[i]
|
102 |
+
channel_out = block_out_channels[i + 1]
|
103 |
+
self.blocks.append(nn.Conv2d(channel_in, channel_in, kernel_size=3, padding=1))
|
104 |
+
self.blocks.append(nn.Conv2d(channel_in, channel_out, kernel_size=3, padding=1, stride=2))
|
105 |
+
|
106 |
+
self.conv_out = zero_module(
|
107 |
+
nn.Conv2d(block_out_channels[-1], conditioning_embedding_channels, kernel_size=3, padding=1)
|
108 |
+
)
|
109 |
+
|
110 |
+
def forward(self, conditioning):
|
111 |
+
embedding = self.conv_in(conditioning)
|
112 |
+
embedding = F.silu(embedding)
|
113 |
+
|
114 |
+
for block in self.blocks:
|
115 |
+
embedding = block(embedding)
|
116 |
+
embedding = F.silu(embedding)
|
117 |
+
|
118 |
+
embedding = self.conv_out(embedding)
|
119 |
+
|
120 |
+
return embedding
|
121 |
+
|
122 |
+
|
123 |
+
class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
|
124 |
+
"""
|
125 |
+
A ControlNet model.
|
126 |
+
|
127 |
+
Args:
|
128 |
+
in_channels (`int`, defaults to 4):
|
129 |
+
The number of channels in the input sample.
|
130 |
+
flip_sin_to_cos (`bool`, defaults to `True`):
|
131 |
+
Whether to flip the sin to cos in the time embedding.
|
132 |
+
freq_shift (`int`, defaults to 0):
|
133 |
+
The frequency shift to apply to the time embedding.
|
134 |
+
down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
|
135 |
+
The tuple of downsample blocks to use.
|
136 |
+
only_cross_attention (`Union[bool, Tuple[bool]]`, defaults to `False`):
|
137 |
+
block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`):
|
138 |
+
The tuple of output channels for each block.
|
139 |
+
layers_per_block (`int`, defaults to 2):
|
140 |
+
The number of layers per block.
|
141 |
+
downsample_padding (`int`, defaults to 1):
|
142 |
+
The padding to use for the downsampling convolution.
|
143 |
+
mid_block_scale_factor (`float`, defaults to 1):
|
144 |
+
The scale factor to use for the mid block.
|
145 |
+
act_fn (`str`, defaults to "silu"):
|
146 |
+
The activation function to use.
|
147 |
+
norm_num_groups (`int`, *optional*, defaults to 32):
|
148 |
+
The number of groups to use for the normalization. If None, normalization and activation layers is skipped
|
149 |
+
in post-processing.
|
150 |
+
norm_eps (`float`, defaults to 1e-5):
|
151 |
+
The epsilon to use for the normalization.
|
152 |
+
cross_attention_dim (`int`, defaults to 1280):
|
153 |
+
The dimension of the cross attention features.
|
154 |
+
transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
|
155 |
+
The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
|
156 |
+
[`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
|
157 |
+
[`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
|
158 |
+
encoder_hid_dim (`int`, *optional*, defaults to None):
|
159 |
+
If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
|
160 |
+
dimension to `cross_attention_dim`.
|
161 |
+
encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
|
162 |
+
If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
|
163 |
+
embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
|
164 |
+
attention_head_dim (`Union[int, Tuple[int]]`, defaults to 8):
|
165 |
+
The dimension of the attention heads.
|
166 |
+
use_linear_projection (`bool`, defaults to `False`):
|
167 |
+
class_embed_type (`str`, *optional*, defaults to `None`):
|
168 |
+
The type of class embedding to use which is ultimately summed with the time embeddings. Choose from None,
|
169 |
+
`"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
|
170 |
+
addition_embed_type (`str`, *optional*, defaults to `None`):
|
171 |
+
Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
|
172 |
+
"text". "text" will use the `TextTimeEmbedding` layer.
|
173 |
+
num_class_embeds (`int`, *optional*, defaults to 0):
|
174 |
+
Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
|
175 |
+
class conditioning with `class_embed_type` equal to `None`.
|
176 |
+
upcast_attention (`bool`, defaults to `False`):
|
177 |
+
resnet_time_scale_shift (`str`, defaults to `"default"`):
|
178 |
+
Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`.
|
179 |
+
projection_class_embeddings_input_dim (`int`, *optional*, defaults to `None`):
|
180 |
+
The dimension of the `class_labels` input when `class_embed_type="projection"`. Required when
|
181 |
+
`class_embed_type="projection"`.
|
182 |
+
controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
|
183 |
+
The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
|
184 |
+
conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(16, 32, 96, 256)`):
|
185 |
+
The tuple of output channel for each block in the `conditioning_embedding` layer.
|
186 |
+
global_pool_conditions (`bool`, defaults to `False`):
|
187 |
+
TODO(Patrick) - unused parameter.
|
188 |
+
addition_embed_type_num_heads (`int`, defaults to 64):
|
189 |
+
The number of heads to use for the `TextTimeEmbedding` layer.
|
190 |
+
"""
|
191 |
+
|
192 |
+
_supports_gradient_checkpointing = True
|
193 |
+
|
194 |
+
@register_to_config
|
195 |
+
def __init__(
|
196 |
+
self,
|
197 |
+
in_channels: int = 4,
|
198 |
+
conditioning_channels: int = 3,
|
199 |
+
flip_sin_to_cos: bool = True,
|
200 |
+
freq_shift: int = 0,
|
201 |
+
down_block_types: Tuple[str, ...] = (
|
202 |
+
"CrossAttnDownBlock2D",
|
203 |
+
"CrossAttnDownBlock2D",
|
204 |
+
"CrossAttnDownBlock2D",
|
205 |
+
"DownBlock2D",
|
206 |
+
),
|
207 |
+
mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
|
208 |
+
only_cross_attention: Union[bool, Tuple[bool]] = False,
|
209 |
+
block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
|
210 |
+
layers_per_block: int = 2,
|
211 |
+
downsample_padding: int = 1,
|
212 |
+
mid_block_scale_factor: float = 1,
|
213 |
+
act_fn: str = "silu",
|
214 |
+
norm_num_groups: Optional[int] = 32,
|
215 |
+
norm_eps: float = 1e-5,
|
216 |
+
cross_attention_dim: int = 1280,
|
217 |
+
transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
|
218 |
+
encoder_hid_dim: Optional[int] = None,
|
219 |
+
encoder_hid_dim_type: Optional[str] = None,
|
220 |
+
attention_head_dim: Union[int, Tuple[int, ...]] = 8,
|
221 |
+
num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None,
|
222 |
+
use_linear_projection: bool = False,
|
223 |
+
class_embed_type: Optional[str] = None,
|
224 |
+
addition_embed_type: Optional[str] = None,
|
225 |
+
addition_time_embed_dim: Optional[int] = None,
|
226 |
+
num_class_embeds: Optional[int] = None,
|
227 |
+
upcast_attention: bool = False,
|
228 |
+
resnet_time_scale_shift: str = "default",
|
229 |
+
projection_class_embeddings_input_dim: Optional[int] = None,
|
230 |
+
controlnet_conditioning_channel_order: str = "rgb",
|
231 |
+
conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
|
232 |
+
global_pool_conditions: bool = False,
|
233 |
+
addition_embed_type_num_heads: int = 64,
|
234 |
+
):
|
235 |
+
super().__init__()
|
236 |
+
|
237 |
+
# If `num_attention_heads` is not defined (which is the case for most models)
|
238 |
+
# it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
|
239 |
+
# The reason for this behavior is to correct for incorrectly named variables that were introduced
|
240 |
+
# when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
|
241 |
+
# Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
|
242 |
+
# which is why we correct for the naming here.
|
243 |
+
num_attention_heads = num_attention_heads or attention_head_dim
|
244 |
+
|
245 |
+
# Check inputs
|
246 |
+
if len(block_out_channels) != len(down_block_types):
|
247 |
+
raise ValueError(
|
248 |
+
f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
|
249 |
+
)
|
250 |
+
|
251 |
+
if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
|
252 |
+
raise ValueError(
|
253 |
+
f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
|
254 |
+
)
|
255 |
+
|
256 |
+
if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
|
257 |
+
raise ValueError(
|
258 |
+
f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
|
259 |
+
)
|
260 |
+
|
261 |
+
if isinstance(transformer_layers_per_block, int):
|
262 |
+
transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
|
263 |
+
|
264 |
+
# input
|
265 |
+
conv_in_kernel = 3
|
266 |
+
conv_in_padding = (conv_in_kernel - 1) // 2
|
267 |
+
self.conv_in = nn.Conv2d(
|
268 |
+
in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
|
269 |
+
)
|
270 |
+
|
271 |
+
# time
|
272 |
+
time_embed_dim = block_out_channels[0] * 4
|
273 |
+
self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
|
274 |
+
timestep_input_dim = block_out_channels[0]
|
275 |
+
self.time_embedding = TimestepEmbedding(
|
276 |
+
timestep_input_dim,
|
277 |
+
time_embed_dim,
|
278 |
+
act_fn=act_fn,
|
279 |
+
)
|
280 |
+
|
281 |
+
if encoder_hid_dim_type is None and encoder_hid_dim is not None:
|
282 |
+
encoder_hid_dim_type = "text_proj"
|
283 |
+
self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
|
284 |
+
logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
|
285 |
+
|
286 |
+
if encoder_hid_dim is None and encoder_hid_dim_type is not None:
|
287 |
+
raise ValueError(
|
288 |
+
f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
|
289 |
+
)
|
290 |
+
|
291 |
+
if encoder_hid_dim_type == "text_proj":
|
292 |
+
self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
|
293 |
+
elif encoder_hid_dim_type == "text_image_proj":
|
294 |
+
# image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
|
295 |
+
# they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
|
296 |
+
# case when `addition_embed_type == "text_image_proj"` (Kandinsky 2.1)`
|
297 |
+
self.encoder_hid_proj = TextImageProjection(
|
298 |
+
text_embed_dim=encoder_hid_dim,
|
299 |
+
image_embed_dim=cross_attention_dim,
|
300 |
+
cross_attention_dim=cross_attention_dim,
|
301 |
+
)
|
302 |
+
|
303 |
+
elif encoder_hid_dim_type is not None:
|
304 |
+
raise ValueError(
|
305 |
+
f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
|
306 |
+
)
|
307 |
+
else:
|
308 |
+
self.encoder_hid_proj = None
|
309 |
+
|
310 |
+
# class embedding
|
311 |
+
if class_embed_type is None and num_class_embeds is not None:
|
312 |
+
self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
|
313 |
+
elif class_embed_type == "timestep":
|
314 |
+
self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
|
315 |
+
elif class_embed_type == "identity":
|
316 |
+
self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
|
317 |
+
elif class_embed_type == "projection":
|
318 |
+
if projection_class_embeddings_input_dim is None:
|
319 |
+
raise ValueError(
|
320 |
+
"`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
|
321 |
+
)
|
322 |
+
# The projection `class_embed_type` is the same as the timestep `class_embed_type` except
|
323 |
+
# 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
|
324 |
+
# 2. it projects from an arbitrary input dimension.
|
325 |
+
#
|
326 |
+
# Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
|
327 |
+
# When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
|
328 |
+
# As a result, `TimestepEmbedding` can be passed arbitrary vectors.
|
329 |
+
self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
|
330 |
+
else:
|
331 |
+
self.class_embedding = None
|
332 |
+
|
333 |
+
if addition_embed_type == "text":
|
334 |
+
if encoder_hid_dim is not None:
|
335 |
+
text_time_embedding_from_dim = encoder_hid_dim
|
336 |
+
else:
|
337 |
+
text_time_embedding_from_dim = cross_attention_dim
|
338 |
+
|
339 |
+
self.add_embedding = TextTimeEmbedding(
|
340 |
+
text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
|
341 |
+
)
|
342 |
+
elif addition_embed_type == "text_image":
|
343 |
+
# text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
|
344 |
+
# they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
|
345 |
+
# case when `addition_embed_type == "text_image"` (Kandinsky 2.1)`
|
346 |
+
self.add_embedding = TextImageTimeEmbedding(
|
347 |
+
text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
|
348 |
+
)
|
349 |
+
elif addition_embed_type == "text_time":
|
350 |
+
self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
|
351 |
+
self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
|
352 |
+
|
353 |
+
elif addition_embed_type is not None:
|
354 |
+
raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
|
355 |
+
|
356 |
+
# control net conditioning embedding
|
357 |
+
self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
|
358 |
+
conditioning_embedding_channels=block_out_channels[0],
|
359 |
+
block_out_channels=conditioning_embedding_out_channels,
|
360 |
+
conditioning_channels=conditioning_channels,
|
361 |
+
)
|
362 |
+
|
363 |
+
self.down_blocks = nn.ModuleList([])
|
364 |
+
self.controlnet_down_blocks = nn.ModuleList([])
|
365 |
+
|
366 |
+
if isinstance(only_cross_attention, bool):
|
367 |
+
only_cross_attention = [only_cross_attention] * len(down_block_types)
|
368 |
+
|
369 |
+
if isinstance(attention_head_dim, int):
|
370 |
+
attention_head_dim = (attention_head_dim,) * len(down_block_types)
|
371 |
+
|
372 |
+
if isinstance(num_attention_heads, int):
|
373 |
+
num_attention_heads = (num_attention_heads,) * len(down_block_types)
|
374 |
+
|
375 |
+
# down
|
376 |
+
output_channel = block_out_channels[0]
|
377 |
+
|
378 |
+
controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
|
379 |
+
controlnet_block = zero_module(controlnet_block)
|
380 |
+
self.controlnet_down_blocks.append(controlnet_block)
|
381 |
+
|
382 |
+
for i, down_block_type in enumerate(down_block_types):
|
383 |
+
input_channel = output_channel
|
384 |
+
output_channel = block_out_channels[i]
|
385 |
+
is_final_block = i == len(block_out_channels) - 1
|
386 |
+
|
387 |
+
down_block = get_down_block(
|
388 |
+
down_block_type,
|
389 |
+
num_layers=layers_per_block,
|
390 |
+
transformer_layers_per_block=transformer_layers_per_block[i],
|
391 |
+
in_channels=input_channel,
|
392 |
+
out_channels=output_channel,
|
393 |
+
temb_channels=time_embed_dim,
|
394 |
+
add_downsample=not is_final_block,
|
395 |
+
resnet_eps=norm_eps,
|
396 |
+
resnet_act_fn=act_fn,
|
397 |
+
resnet_groups=norm_num_groups,
|
398 |
+
cross_attention_dim=cross_attention_dim,
|
399 |
+
num_attention_heads=num_attention_heads[i],
|
400 |
+
attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
|
401 |
+
downsample_padding=downsample_padding,
|
402 |
+
use_linear_projection=use_linear_projection,
|
403 |
+
only_cross_attention=only_cross_attention[i],
|
404 |
+
upcast_attention=upcast_attention,
|
405 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
406 |
+
)
|
407 |
+
self.down_blocks.append(down_block)
|
408 |
+
|
409 |
+
for _ in range(layers_per_block):
|
410 |
+
controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
|
411 |
+
controlnet_block = zero_module(controlnet_block)
|
412 |
+
self.controlnet_down_blocks.append(controlnet_block)
|
413 |
+
|
414 |
+
if not is_final_block:
|
415 |
+
controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
|
416 |
+
controlnet_block = zero_module(controlnet_block)
|
417 |
+
self.controlnet_down_blocks.append(controlnet_block)
|
418 |
+
|
419 |
+
# mid
|
420 |
+
mid_block_channel = block_out_channels[-1]
|
421 |
+
|
422 |
+
controlnet_block = nn.Conv2d(mid_block_channel, mid_block_channel, kernel_size=1)
|
423 |
+
controlnet_block = zero_module(controlnet_block)
|
424 |
+
self.controlnet_mid_block = controlnet_block
|
425 |
+
|
426 |
+
if mid_block_type == "UNetMidBlock2DCrossAttn":
|
427 |
+
self.mid_block = UNetMidBlock2DCrossAttn(
|
428 |
+
transformer_layers_per_block=transformer_layers_per_block[-1],
|
429 |
+
in_channels=mid_block_channel,
|
430 |
+
temb_channels=time_embed_dim,
|
431 |
+
resnet_eps=norm_eps,
|
432 |
+
resnet_act_fn=act_fn,
|
433 |
+
output_scale_factor=mid_block_scale_factor,
|
434 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
435 |
+
cross_attention_dim=cross_attention_dim,
|
436 |
+
num_attention_heads=num_attention_heads[-1],
|
437 |
+
resnet_groups=norm_num_groups,
|
438 |
+
use_linear_projection=use_linear_projection,
|
439 |
+
upcast_attention=upcast_attention,
|
440 |
+
)
|
441 |
+
elif mid_block_type == "UNetMidBlock2D":
|
442 |
+
self.mid_block = UNetMidBlock2D(
|
443 |
+
in_channels=block_out_channels[-1],
|
444 |
+
temb_channels=time_embed_dim,
|
445 |
+
num_layers=0,
|
446 |
+
resnet_eps=norm_eps,
|
447 |
+
resnet_act_fn=act_fn,
|
448 |
+
output_scale_factor=mid_block_scale_factor,
|
449 |
+
resnet_groups=norm_num_groups,
|
450 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
451 |
+
add_attention=False,
|
452 |
+
)
|
453 |
+
else:
|
454 |
+
raise ValueError(f"unknown mid_block_type : {mid_block_type}")
|
455 |
+
|
456 |
+
@classmethod
|
457 |
+
def from_unet(
|
458 |
+
cls,
|
459 |
+
unet: UNet2DConditionModel,
|
460 |
+
controlnet_conditioning_channel_order: str = "rgb",
|
461 |
+
conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
|
462 |
+
load_weights_from_unet: bool = True,
|
463 |
+
conditioning_channels: int = 3,
|
464 |
+
):
|
465 |
+
r"""
|
466 |
+
Instantiate a [`ControlNetModel`] from [`UNet2DConditionModel`].
|
467 |
+
|
468 |
+
Parameters:
|
469 |
+
unet (`UNet2DConditionModel`):
|
470 |
+
The UNet model weights to copy to the [`ControlNetModel`]. All configuration options are also copied
|
471 |
+
where applicable.
|
472 |
+
"""
|
473 |
+
transformer_layers_per_block = (
|
474 |
+
unet.config.transformer_layers_per_block if "transformer_layers_per_block" in unet.config else 1
|
475 |
+
)
|
476 |
+
encoder_hid_dim = unet.config.encoder_hid_dim if "encoder_hid_dim" in unet.config else None
|
477 |
+
encoder_hid_dim_type = unet.config.encoder_hid_dim_type if "encoder_hid_dim_type" in unet.config else None
|
478 |
+
addition_embed_type = unet.config.addition_embed_type if "addition_embed_type" in unet.config else None
|
479 |
+
addition_time_embed_dim = (
|
480 |
+
unet.config.addition_time_embed_dim if "addition_time_embed_dim" in unet.config else None
|
481 |
+
)
|
482 |
+
|
483 |
+
controlnet = cls(
|
484 |
+
encoder_hid_dim=encoder_hid_dim,
|
485 |
+
encoder_hid_dim_type=encoder_hid_dim_type,
|
486 |
+
addition_embed_type=addition_embed_type,
|
487 |
+
addition_time_embed_dim=addition_time_embed_dim,
|
488 |
+
transformer_layers_per_block=transformer_layers_per_block,
|
489 |
+
in_channels=unet.config.in_channels,
|
490 |
+
flip_sin_to_cos=unet.config.flip_sin_to_cos,
|
491 |
+
freq_shift=unet.config.freq_shift,
|
492 |
+
down_block_types=unet.config.down_block_types,
|
493 |
+
only_cross_attention=unet.config.only_cross_attention,
|
494 |
+
block_out_channels=unet.config.block_out_channels,
|
495 |
+
layers_per_block=unet.config.layers_per_block,
|
496 |
+
downsample_padding=unet.config.downsample_padding,
|
497 |
+
mid_block_scale_factor=unet.config.mid_block_scale_factor,
|
498 |
+
act_fn=unet.config.act_fn,
|
499 |
+
norm_num_groups=unet.config.norm_num_groups,
|
500 |
+
norm_eps=unet.config.norm_eps,
|
501 |
+
cross_attention_dim=unet.config.cross_attention_dim,
|
502 |
+
attention_head_dim=unet.config.attention_head_dim,
|
503 |
+
num_attention_heads=unet.config.num_attention_heads,
|
504 |
+
use_linear_projection=unet.config.use_linear_projection,
|
505 |
+
class_embed_type=unet.config.class_embed_type,
|
506 |
+
num_class_embeds=unet.config.num_class_embeds,
|
507 |
+
upcast_attention=unet.config.upcast_attention,
|
508 |
+
resnet_time_scale_shift=unet.config.resnet_time_scale_shift,
|
509 |
+
projection_class_embeddings_input_dim=unet.config.projection_class_embeddings_input_dim,
|
510 |
+
mid_block_type=unet.config.mid_block_type,
|
511 |
+
controlnet_conditioning_channel_order=controlnet_conditioning_channel_order,
|
512 |
+
conditioning_embedding_out_channels=conditioning_embedding_out_channels,
|
513 |
+
conditioning_channels=conditioning_channels,
|
514 |
+
)
|
515 |
+
|
516 |
+
if load_weights_from_unet:
|
517 |
+
controlnet.conv_in.load_state_dict(unet.conv_in.state_dict())
|
518 |
+
controlnet.time_proj.load_state_dict(unet.time_proj.state_dict())
|
519 |
+
controlnet.time_embedding.load_state_dict(unet.time_embedding.state_dict())
|
520 |
+
|
521 |
+
if controlnet.class_embedding:
|
522 |
+
controlnet.class_embedding.load_state_dict(unet.class_embedding.state_dict())
|
523 |
+
|
524 |
+
if hasattr(controlnet, "add_embedding"):
|
525 |
+
controlnet.add_embedding.load_state_dict(unet.add_embedding.state_dict())
|
526 |
+
|
527 |
+
controlnet.down_blocks.load_state_dict(unet.down_blocks.state_dict())
|
528 |
+
controlnet.mid_block.load_state_dict(unet.mid_block.state_dict())
|
529 |
+
|
530 |
+
return controlnet
|
531 |
+
|
532 |
+
@property
|
533 |
+
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
|
534 |
+
def attn_processors(self) -> Dict[str, AttentionProcessor]:
|
535 |
+
r"""
|
536 |
+
Returns:
|
537 |
+
`dict` of attention processors: A dictionary containing all attention processors used in the model with
|
538 |
+
indexed by its weight name.
|
539 |
+
"""
|
540 |
+
# set recursively
|
541 |
+
processors = {}
|
542 |
+
|
543 |
+
def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
|
544 |
+
if hasattr(module, "get_processor"):
|
545 |
+
processors[f"{name}.processor"] = module.get_processor()
|
546 |
+
|
547 |
+
for sub_name, child in module.named_children():
|
548 |
+
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
|
549 |
+
|
550 |
+
return processors
|
551 |
+
|
552 |
+
for name, module in self.named_children():
|
553 |
+
fn_recursive_add_processors(name, module, processors)
|
554 |
+
|
555 |
+
return processors
|
556 |
+
|
557 |
+
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
|
558 |
+
def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
|
559 |
+
r"""
|
560 |
+
Sets the attention processor to use to compute attention.
|
561 |
+
|
562 |
+
Parameters:
|
563 |
+
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
|
564 |
+
The instantiated processor class or a dictionary of processor classes that will be set as the processor
|
565 |
+
for **all** `Attention` layers.
|
566 |
+
|
567 |
+
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
|
568 |
+
processor. This is strongly recommended when setting trainable attention processors.
|
569 |
+
|
570 |
+
"""
|
571 |
+
count = len(self.attn_processors.keys())
|
572 |
+
|
573 |
+
if isinstance(processor, dict) and len(processor) != count:
|
574 |
+
raise ValueError(
|
575 |
+
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
|
576 |
+
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
|
577 |
+
)
|
578 |
+
|
579 |
+
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
|
580 |
+
if hasattr(module, "set_processor"):
|
581 |
+
if not isinstance(processor, dict):
|
582 |
+
module.set_processor(processor)
|
583 |
+
else:
|
584 |
+
module.set_processor(processor.pop(f"{name}.processor"))
|
585 |
+
|
586 |
+
for sub_name, child in module.named_children():
|
587 |
+
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
|
588 |
+
|
589 |
+
for name, module in self.named_children():
|
590 |
+
fn_recursive_attn_processor(name, module, processor)
|
591 |
+
|
592 |
+
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
|
593 |
+
def set_default_attn_processor(self):
|
594 |
+
"""
|
595 |
+
Disables custom attention processors and sets the default attention implementation.
|
596 |
+
"""
|
597 |
+
if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
|
598 |
+
processor = AttnAddedKVProcessor()
|
599 |
+
elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
|
600 |
+
processor = AttnProcessor()
|
601 |
+
else:
|
602 |
+
raise ValueError(
|
603 |
+
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
|
604 |
+
)
|
605 |
+
|
606 |
+
self.set_attn_processor(processor)
|
607 |
+
|
608 |
+
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attention_slice
|
609 |
+
def set_attention_slice(self, slice_size: Union[str, int, List[int]]) -> None:
|
610 |
+
r"""
|
611 |
+
Enable sliced attention computation.
|
612 |
+
|
613 |
+
When this option is enabled, the attention module splits the input tensor in slices to compute attention in
|
614 |
+
several steps. This is useful for saving some memory in exchange for a small decrease in speed.
|
615 |
+
|
616 |
+
Args:
|
617 |
+
slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
|
618 |
+
When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
|
619 |
+
`"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
|
620 |
+
provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
|
621 |
+
must be a multiple of `slice_size`.
|
622 |
+
"""
|
623 |
+
sliceable_head_dims = []
|
624 |
+
|
625 |
+
def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
|
626 |
+
if hasattr(module, "set_attention_slice"):
|
627 |
+
sliceable_head_dims.append(module.sliceable_head_dim)
|
628 |
+
|
629 |
+
for child in module.children():
|
630 |
+
fn_recursive_retrieve_sliceable_dims(child)
|
631 |
+
|
632 |
+
# retrieve number of attention layers
|
633 |
+
for module in self.children():
|
634 |
+
fn_recursive_retrieve_sliceable_dims(module)
|
635 |
+
|
636 |
+
num_sliceable_layers = len(sliceable_head_dims)
|
637 |
+
|
638 |
+
if slice_size == "auto":
|
639 |
+
# half the attention head size is usually a good trade-off between
|
640 |
+
# speed and memory
|
641 |
+
slice_size = [dim // 2 for dim in sliceable_head_dims]
|
642 |
+
elif slice_size == "max":
|
643 |
+
# make smallest slice possible
|
644 |
+
slice_size = num_sliceable_layers * [1]
|
645 |
+
|
646 |
+
slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
|
647 |
+
|
648 |
+
if len(slice_size) != len(sliceable_head_dims):
|
649 |
+
raise ValueError(
|
650 |
+
f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
|
651 |
+
f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
|
652 |
+
)
|
653 |
+
|
654 |
+
for i in range(len(slice_size)):
|
655 |
+
size = slice_size[i]
|
656 |
+
dim = sliceable_head_dims[i]
|
657 |
+
if size is not None and size > dim:
|
658 |
+
raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
|
659 |
+
|
660 |
+
# Recursively walk through all the children.
|
661 |
+
# Any children which exposes the set_attention_slice method
|
662 |
+
# gets the message
|
663 |
+
def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
|
664 |
+
if hasattr(module, "set_attention_slice"):
|
665 |
+
module.set_attention_slice(slice_size.pop())
|
666 |
+
|
667 |
+
for child in module.children():
|
668 |
+
fn_recursive_set_attention_slice(child, slice_size)
|
669 |
+
|
670 |
+
reversed_slice_size = list(reversed(slice_size))
|
671 |
+
for module in self.children():
|
672 |
+
fn_recursive_set_attention_slice(module, reversed_slice_size)
|
673 |
+
|
674 |
+
def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
|
675 |
+
if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)):
|
676 |
+
module.gradient_checkpointing = value
|
677 |
+
|
678 |
+
def forward(
|
679 |
+
self,
|
680 |
+
sample: torch.Tensor,
|
681 |
+
timestep: Union[torch.Tensor, float, int],
|
682 |
+
encoder_hidden_states: torch.Tensor,
|
683 |
+
controlnet_cond: torch.Tensor,
|
684 |
+
conditioning_scale: float = 1.0,
|
685 |
+
class_labels: Optional[torch.Tensor] = None,
|
686 |
+
timestep_cond: Optional[torch.Tensor] = None,
|
687 |
+
attention_mask: Optional[torch.Tensor] = None,
|
688 |
+
added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
|
689 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
690 |
+
guess_mode: bool = False,
|
691 |
+
return_dict: bool = True,
|
692 |
+
) -> Union[ControlNetOutput, Tuple[Tuple[torch.Tensor, ...], torch.Tensor]]:
|
693 |
+
"""
|
694 |
+
The [`ControlNetModel`] forward method.
|
695 |
+
|
696 |
+
Args:
|
697 |
+
sample (`torch.Tensor`):
|
698 |
+
The noisy input tensor.
|
699 |
+
timestep (`Union[torch.Tensor, float, int]`):
|
700 |
+
The number of timesteps to denoise an input.
|
701 |
+
encoder_hidden_states (`torch.Tensor`):
|
702 |
+
The encoder hidden states.
|
703 |
+
controlnet_cond (`torch.Tensor`):
|
704 |
+
The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
|
705 |
+
conditioning_scale (`float`, defaults to `1.0`):
|
706 |
+
The scale factor for ControlNet outputs.
|
707 |
+
class_labels (`torch.Tensor`, *optional*, defaults to `None`):
|
708 |
+
Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
|
709 |
+
timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
|
710 |
+
Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
|
711 |
+
timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
|
712 |
+
embeddings.
|
713 |
+
attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
|
714 |
+
An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
|
715 |
+
is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
|
716 |
+
negative values to the attention scores corresponding to "discard" tokens.
|
717 |
+
added_cond_kwargs (`dict`):
|
718 |
+
Additional conditions for the Stable Diffusion XL UNet.
|
719 |
+
cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
|
720 |
+
A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
|
721 |
+
guess_mode (`bool`, defaults to `False`):
|
722 |
+
In this mode, the ControlNet encoder tries its best to recognize the input content of the input even if
|
723 |
+
you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
|
724 |
+
return_dict (`bool`, defaults to `True`):
|
725 |
+
Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.
|
726 |
+
|
727 |
+
Returns:
|
728 |
+
[`~models.controlnet.ControlNetOutput`] **or** `tuple`:
|
729 |
+
If `return_dict` is `True`, a [`~models.controlnet.ControlNetOutput`] is returned, otherwise a tuple is
|
730 |
+
returned where the first element is the sample tensor.
|
731 |
+
"""
|
732 |
+
# check channel order
|
733 |
+
channel_order = self.config.controlnet_conditioning_channel_order
|
734 |
+
|
735 |
+
if channel_order == "rgb":
|
736 |
+
# in rgb order by default
|
737 |
+
...
|
738 |
+
elif channel_order == "bgr":
|
739 |
+
controlnet_cond = torch.flip(controlnet_cond, dims=[1])
|
740 |
+
else:
|
741 |
+
raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}")
|
742 |
+
|
743 |
+
# prepare attention_mask
|
744 |
+
if attention_mask is not None:
|
745 |
+
attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
|
746 |
+
attention_mask = attention_mask.unsqueeze(1)
|
747 |
+
|
748 |
+
#Todo
|
749 |
+
if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
|
750 |
+
encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
|
751 |
+
|
752 |
+
# 1. time
|
753 |
+
timesteps = timestep
|
754 |
+
if not torch.is_tensor(timesteps):
|
755 |
+
# TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
|
756 |
+
# This would be a good case for the `match` statement (Python 3.10+)
|
757 |
+
is_mps = sample.device.type == "mps"
|
758 |
+
if isinstance(timestep, float):
|
759 |
+
dtype = torch.float32 if is_mps else torch.float64
|
760 |
+
else:
|
761 |
+
dtype = torch.int32 if is_mps else torch.int64
|
762 |
+
timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
|
763 |
+
elif len(timesteps.shape) == 0:
|
764 |
+
timesteps = timesteps[None].to(sample.device)
|
765 |
+
|
766 |
+
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
767 |
+
timesteps = timesteps.expand(sample.shape[0])
|
768 |
+
|
769 |
+
t_emb = self.time_proj(timesteps)
|
770 |
+
|
771 |
+
# timesteps does not contain any weights and will always return f32 tensors
|
772 |
+
# but time_embedding might actually be running in fp16. so we need to cast here.
|
773 |
+
# there might be better ways to encapsulate this.
|
774 |
+
t_emb = t_emb.to(dtype=sample.dtype)
|
775 |
+
|
776 |
+
emb = self.time_embedding(t_emb, timestep_cond)
|
777 |
+
aug_emb = None
|
778 |
+
|
779 |
+
if self.class_embedding is not None:
|
780 |
+
if class_labels is None:
|
781 |
+
raise ValueError("class_labels should be provided when num_class_embeds > 0")
|
782 |
+
|
783 |
+
if self.config.class_embed_type == "timestep":
|
784 |
+
class_labels = self.time_proj(class_labels)
|
785 |
+
|
786 |
+
class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
|
787 |
+
emb = emb + class_emb
|
788 |
+
|
789 |
+
if self.config.addition_embed_type is not None:
|
790 |
+
if self.config.addition_embed_type == "text":
|
791 |
+
aug_emb = self.add_embedding(encoder_hidden_states)
|
792 |
+
|
793 |
+
elif self.config.addition_embed_type == "text_time":
|
794 |
+
if "text_embeds" not in added_cond_kwargs:
|
795 |
+
raise ValueError(
|
796 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
|
797 |
+
)
|
798 |
+
text_embeds = added_cond_kwargs.get("text_embeds")
|
799 |
+
if "time_ids" not in added_cond_kwargs:
|
800 |
+
raise ValueError(
|
801 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
|
802 |
+
)
|
803 |
+
time_ids = added_cond_kwargs.get("time_ids")
|
804 |
+
time_embeds = self.add_time_proj(time_ids.flatten())
|
805 |
+
time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
|
806 |
+
|
807 |
+
add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
|
808 |
+
add_embeds = add_embeds.to(emb.dtype)
|
809 |
+
aug_emb = self.add_embedding(add_embeds)
|
810 |
+
|
811 |
+
emb = emb + aug_emb if aug_emb is not None else emb
|
812 |
+
|
813 |
+
# 2. pre-process
|
814 |
+
sample = self.conv_in(sample)
|
815 |
+
|
816 |
+
controlnet_cond = self.controlnet_cond_embedding(controlnet_cond)
|
817 |
+
sample = sample + controlnet_cond
|
818 |
+
|
819 |
+
# 3. down
|
820 |
+
down_block_res_samples = (sample,)
|
821 |
+
for downsample_block in self.down_blocks:
|
822 |
+
if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
|
823 |
+
sample, res_samples = downsample_block(
|
824 |
+
hidden_states=sample,
|
825 |
+
temb=emb,
|
826 |
+
encoder_hidden_states=encoder_hidden_states,
|
827 |
+
attention_mask=attention_mask,
|
828 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
829 |
+
)
|
830 |
+
else:
|
831 |
+
sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
|
832 |
+
|
833 |
+
down_block_res_samples += res_samples
|
834 |
+
|
835 |
+
# 4. mid
|
836 |
+
if self.mid_block is not None:
|
837 |
+
if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
|
838 |
+
sample = self.mid_block(
|
839 |
+
sample,
|
840 |
+
emb,
|
841 |
+
encoder_hidden_states=encoder_hidden_states,
|
842 |
+
attention_mask=attention_mask,
|
843 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
844 |
+
)
|
845 |
+
else:
|
846 |
+
sample = self.mid_block(sample, emb)
|
847 |
+
|
848 |
+
# 5. Control net blocks
|
849 |
+
|
850 |
+
controlnet_down_block_res_samples = ()
|
851 |
+
|
852 |
+
for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
|
853 |
+
down_block_res_sample = controlnet_block(down_block_res_sample)
|
854 |
+
controlnet_down_block_res_samples = controlnet_down_block_res_samples + (down_block_res_sample,)
|
855 |
+
|
856 |
+
down_block_res_samples = controlnet_down_block_res_samples
|
857 |
+
|
858 |
+
mid_block_res_sample = self.controlnet_mid_block(sample)
|
859 |
+
|
860 |
+
# 6. scaling
|
861 |
+
if guess_mode and not self.config.global_pool_conditions:
|
862 |
+
scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device) # 0.1 to 1.0
|
863 |
+
scales = scales * conditioning_scale
|
864 |
+
down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)]
|
865 |
+
mid_block_res_sample = mid_block_res_sample * scales[-1] # last one
|
866 |
+
else:
|
867 |
+
down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples]
|
868 |
+
mid_block_res_sample = mid_block_res_sample * conditioning_scale
|
869 |
+
|
870 |
+
if self.config.global_pool_conditions:
|
871 |
+
down_block_res_samples = [
|
872 |
+
torch.mean(sample, dim=(2, 3), keepdim=True) for sample in down_block_res_samples
|
873 |
+
]
|
874 |
+
mid_block_res_sample = torch.mean(mid_block_res_sample, dim=(2, 3), keepdim=True)
|
875 |
+
|
876 |
+
if not return_dict:
|
877 |
+
return (down_block_res_samples, mid_block_res_sample)
|
878 |
+
|
879 |
+
return ControlNetOutput(
|
880 |
+
down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
|
881 |
+
)
|
882 |
+
|
883 |
+
|
884 |
+
def zero_module(module):
|
885 |
+
for p in module.parameters():
|
886 |
+
nn.init.zeros_(p)
|
887 |
+
return module
|
models/modeling_chatglm.py
ADDED
@@ -0,0 +1,1298 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
""" PyTorch ChatGLM model. """
|
2 |
+
|
3 |
+
import math
|
4 |
+
import copy
|
5 |
+
import warnings
|
6 |
+
import re
|
7 |
+
import sys
|
8 |
+
|
9 |
+
import torch
|
10 |
+
import torch.utils.checkpoint
|
11 |
+
import torch.nn.functional as F
|
12 |
+
from torch import nn
|
13 |
+
from torch.nn import CrossEntropyLoss, LayerNorm
|
14 |
+
from torch.nn import CrossEntropyLoss, LayerNorm, MSELoss, BCEWithLogitsLoss
|
15 |
+
from torch.nn.utils import skip_init
|
16 |
+
from typing import Optional, Tuple, Union, List, Callable, Dict, Any
|
17 |
+
from copy import deepcopy
|
18 |
+
|
19 |
+
from transformers.modeling_outputs import (
|
20 |
+
BaseModelOutputWithPast,
|
21 |
+
CausalLMOutputWithPast,
|
22 |
+
SequenceClassifierOutputWithPast,
|
23 |
+
)
|
24 |
+
from transformers.modeling_utils import PreTrainedModel
|
25 |
+
from transformers.utils import logging
|
26 |
+
from transformers.generation.logits_process import LogitsProcessor
|
27 |
+
from transformers.generation.utils import LogitsProcessorList, StoppingCriteriaList, GenerationConfig, ModelOutput
|
28 |
+
|
29 |
+
try:
|
30 |
+
from .configuration_chatglm import ChatGLMConfig
|
31 |
+
except:
|
32 |
+
from configuration_chatglm import ChatGLMConfig
|
33 |
+
|
34 |
+
|
35 |
+
# flags required to enable jit fusion kernels
|
36 |
+
|
37 |
+
if sys.platform != 'darwin':
|
38 |
+
torch._C._jit_set_profiling_mode(False)
|
39 |
+
torch._C._jit_set_profiling_executor(False)
|
40 |
+
torch._C._jit_override_can_fuse_on_cpu(True)
|
41 |
+
torch._C._jit_override_can_fuse_on_gpu(True)
|
42 |
+
|
43 |
+
logger = logging.get_logger(__name__)
|
44 |
+
|
45 |
+
_CHECKPOINT_FOR_DOC = "THUDM/ChatGLM"
|
46 |
+
_CONFIG_FOR_DOC = "ChatGLM6BConfig"
|
47 |
+
|
48 |
+
CHATGLM_6B_PRETRAINED_MODEL_ARCHIVE_LIST = [
|
49 |
+
"THUDM/chatglm3-6b-base",
|
50 |
+
# See all ChatGLM models at https://huggingface.co/models?filter=chatglm
|
51 |
+
]
|
52 |
+
|
53 |
+
|
54 |
+
def default_init(cls, *args, **kwargs):
|
55 |
+
return cls(*args, **kwargs)
|
56 |
+
|
57 |
+
|
58 |
+
class InvalidScoreLogitsProcessor(LogitsProcessor):
|
59 |
+
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
|
60 |
+
if torch.isnan(scores).any() or torch.isinf(scores).any():
|
61 |
+
scores.zero_()
|
62 |
+
scores[..., 5] = 5e4
|
63 |
+
return scores
|
64 |
+
|
65 |
+
|
66 |
+
class PrefixEncoder(torch.nn.Module):
|
67 |
+
"""
|
68 |
+
The torch.nn model to encode the prefix
|
69 |
+
Input shape: (batch-size, prefix-length)
|
70 |
+
Output shape: (batch-size, prefix-length, 2*layers*hidden)
|
71 |
+
"""
|
72 |
+
|
73 |
+
def __init__(self, config: ChatGLMConfig):
|
74 |
+
super().__init__()
|
75 |
+
self.prefix_projection = config.prefix_projection
|
76 |
+
if self.prefix_projection:
|
77 |
+
# Use a two-layer MLP to encode the prefix
|
78 |
+
kv_size = config.num_layers * config.kv_channels * config.multi_query_group_num * 2
|
79 |
+
self.embedding = torch.nn.Embedding(config.pre_seq_len, kv_size)
|
80 |
+
self.trans = torch.nn.Sequential(
|
81 |
+
torch.nn.Linear(kv_size, config.hidden_size),
|
82 |
+
torch.nn.Tanh(),
|
83 |
+
torch.nn.Linear(config.hidden_size, kv_size)
|
84 |
+
)
|
85 |
+
else:
|
86 |
+
self.embedding = torch.nn.Embedding(config.pre_seq_len,
|
87 |
+
config.num_layers * config.kv_channels * config.multi_query_group_num * 2)
|
88 |
+
|
89 |
+
def forward(self, prefix: torch.Tensor):
|
90 |
+
if self.prefix_projection:
|
91 |
+
prefix_tokens = self.embedding(prefix)
|
92 |
+
past_key_values = self.trans(prefix_tokens)
|
93 |
+
else:
|
94 |
+
past_key_values = self.embedding(prefix)
|
95 |
+
return past_key_values
|
96 |
+
|
97 |
+
|
98 |
+
def split_tensor_along_last_dim(
|
99 |
+
tensor: torch.Tensor,
|
100 |
+
num_partitions: int,
|
101 |
+
contiguous_split_chunks: bool = False,
|
102 |
+
) -> List[torch.Tensor]:
|
103 |
+
"""Split a tensor along its last dimension.
|
104 |
+
|
105 |
+
Arguments:
|
106 |
+
tensor: input tensor.
|
107 |
+
num_partitions: number of partitions to split the tensor
|
108 |
+
contiguous_split_chunks: If True, make each chunk contiguous
|
109 |
+
in memory.
|
110 |
+
|
111 |
+
Returns:
|
112 |
+
A list of Tensors
|
113 |
+
"""
|
114 |
+
# Get the size and dimension.
|
115 |
+
last_dim = tensor.dim() - 1
|
116 |
+
last_dim_size = tensor.size()[last_dim] // num_partitions
|
117 |
+
# Split.
|
118 |
+
tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
|
119 |
+
# Note: torch.split does not create contiguous tensors by default.
|
120 |
+
if contiguous_split_chunks:
|
121 |
+
return tuple(chunk.contiguous() for chunk in tensor_list)
|
122 |
+
|
123 |
+
return tensor_list
|
124 |
+
|
125 |
+
|
126 |
+
class RotaryEmbedding(nn.Module):
|
127 |
+
def __init__(self, dim, original_impl=False, device=None, dtype=None):
|
128 |
+
super().__init__()
|
129 |
+
inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, device=device).to(dtype=dtype) / dim))
|
130 |
+
self.register_buffer("inv_freq", inv_freq)
|
131 |
+
self.dim = dim
|
132 |
+
self.original_impl = original_impl
|
133 |
+
|
134 |
+
def forward_impl(
|
135 |
+
self, seq_len: int, n_elem: int, dtype: torch.dtype, device: torch.device, base: int = 10000
|
136 |
+
):
|
137 |
+
"""Enhanced Transformer with Rotary Position Embedding.
|
138 |
+
|
139 |
+
Derived from: https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/labml_nn/
|
140 |
+
transformers/rope/__init__.py. MIT License:
|
141 |
+
https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/license.
|
142 |
+
"""
|
143 |
+
# $\Theta = {\theta_i = 10000^{\frac{2(i-1)}{d}}, i \in [1, 2, ..., \frac{d}{2}]}$
|
144 |
+
theta = 1.0 / (base ** (torch.arange(0, n_elem, 2, dtype=torch.float, device=device) / n_elem))
|
145 |
+
|
146 |
+
# Create position indexes `[0, 1, ..., seq_len - 1]`
|
147 |
+
seq_idx = torch.arange(seq_len, dtype=torch.float, device=device)
|
148 |
+
|
149 |
+
# Calculate the product of position index and $\theta_i$
|
150 |
+
idx_theta = torch.outer(seq_idx, theta).float()
|
151 |
+
|
152 |
+
cache = torch.stack([torch.cos(idx_theta), torch.sin(idx_theta)], dim=-1)
|
153 |
+
|
154 |
+
# this is to mimic the behaviour of complex32, else we will get different results
|
155 |
+
if dtype in (torch.float16, torch.bfloat16, torch.int8):
|
156 |
+
cache = cache.bfloat16() if dtype == torch.bfloat16 else cache.half()
|
157 |
+
return cache
|
158 |
+
|
159 |
+
def forward(self, max_seq_len, offset=0):
|
160 |
+
return self.forward_impl(
|
161 |
+
max_seq_len, self.dim, dtype=self.inv_freq.dtype, device=self.inv_freq.device
|
162 |
+
)
|
163 |
+
|
164 |
+
|
165 |
+
@torch.jit.script
|
166 |
+
def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
|
167 |
+
# x: [sq, b, np, hn]
|
168 |
+
sq, b, np, hn = x.size(0), x.size(1), x.size(2), x.size(3)
|
169 |
+
rot_dim = rope_cache.shape[-2] * 2
|
170 |
+
x, x_pass = x[..., :rot_dim], x[..., rot_dim:]
|
171 |
+
# truncate to support variable sizes
|
172 |
+
rope_cache = rope_cache[:sq]
|
173 |
+
xshaped = x.reshape(sq, -1, np, rot_dim // 2, 2)
|
174 |
+
rope_cache = rope_cache.view(sq, -1, 1, xshaped.size(3), 2)
|
175 |
+
x_out2 = torch.stack(
|
176 |
+
[
|
177 |
+
xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1],
|
178 |
+
xshaped[..., 1] * rope_cache[..., 0] + xshaped[..., 0] * rope_cache[..., 1],
|
179 |
+
],
|
180 |
+
-1,
|
181 |
+
)
|
182 |
+
x_out2 = x_out2.flatten(3)
|
183 |
+
return torch.cat((x_out2, x_pass), dim=-1)
|
184 |
+
|
185 |
+
|
186 |
+
class RMSNorm(torch.nn.Module):
|
187 |
+
def __init__(self, normalized_shape, eps=1e-5, device=None, dtype=None, **kwargs):
|
188 |
+
super().__init__()
|
189 |
+
self.weight = torch.nn.Parameter(torch.empty(normalized_shape, device=device, dtype=dtype))
|
190 |
+
self.eps = eps
|
191 |
+
|
192 |
+
def forward(self, hidden_states: torch.Tensor):
|
193 |
+
input_dtype = hidden_states.dtype
|
194 |
+
variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
|
195 |
+
hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
|
196 |
+
|
197 |
+
return (self.weight * hidden_states).to(input_dtype)
|
198 |
+
|
199 |
+
|
200 |
+
class CoreAttention(torch.nn.Module):
|
201 |
+
def __init__(self, config: ChatGLMConfig, layer_number):
|
202 |
+
super(CoreAttention, self).__init__()
|
203 |
+
|
204 |
+
self.apply_query_key_layer_scaling = config.apply_query_key_layer_scaling
|
205 |
+
self.attention_softmax_in_fp32 = config.attention_softmax_in_fp32
|
206 |
+
if self.apply_query_key_layer_scaling:
|
207 |
+
self.attention_softmax_in_fp32 = True
|
208 |
+
self.layer_number = max(1, layer_number)
|
209 |
+
|
210 |
+
projection_size = config.kv_channels * config.num_attention_heads
|
211 |
+
|
212 |
+
# Per attention head and per partition values.
|
213 |
+
self.hidden_size_per_partition = projection_size
|
214 |
+
self.hidden_size_per_attention_head = projection_size // config.num_attention_heads
|
215 |
+
self.num_attention_heads_per_partition = config.num_attention_heads
|
216 |
+
|
217 |
+
coeff = None
|
218 |
+
self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)
|
219 |
+
if self.apply_query_key_layer_scaling:
|
220 |
+
coeff = self.layer_number
|
221 |
+
self.norm_factor *= coeff
|
222 |
+
self.coeff = coeff
|
223 |
+
|
224 |
+
self.attention_dropout = torch.nn.Dropout(config.attention_dropout)
|
225 |
+
|
226 |
+
def forward(self, query_layer, key_layer, value_layer, attention_mask):
|
227 |
+
pytorch_major_version = int(torch.__version__.split('.')[0])
|
228 |
+
if pytorch_major_version >= 2:
|
229 |
+
query_layer, key_layer, value_layer = [k.permute(1, 2, 0, 3) for k in [query_layer, key_layer, value_layer]]
|
230 |
+
if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]:
|
231 |
+
context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
|
232 |
+
is_causal=True)
|
233 |
+
else:
|
234 |
+
if attention_mask is not None:
|
235 |
+
attention_mask = ~attention_mask
|
236 |
+
context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer,
|
237 |
+
attention_mask)
|
238 |
+
context_layer = context_layer.permute(2, 0, 1, 3)
|
239 |
+
new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
|
240 |
+
context_layer = context_layer.reshape(*new_context_layer_shape)
|
241 |
+
else:
|
242 |
+
# Raw attention scores
|
243 |
+
|
244 |
+
# [b, np, sq, sk]
|
245 |
+
output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0))
|
246 |
+
|
247 |
+
# [sq, b, np, hn] -> [sq, b * np, hn]
|
248 |
+
query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
|
249 |
+
# [sk, b, np, hn] -> [sk, b * np, hn]
|
250 |
+
key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1)
|
251 |
+
|
252 |
+
# preallocting input tensor: [b * np, sq, sk]
|
253 |
+
matmul_input_buffer = torch.empty(
|
254 |
+
output_size[0] * output_size[1], output_size[2], output_size[3], dtype=query_layer.dtype,
|
255 |
+
device=query_layer.device
|
256 |
+
)
|
257 |
+
|
258 |
+
# Raw attention scores. [b * np, sq, sk]
|
259 |
+
matmul_result = torch.baddbmm(
|
260 |
+
matmul_input_buffer,
|
261 |
+
query_layer.transpose(0, 1), # [b * np, sq, hn]
|
262 |
+
key_layer.transpose(0, 1).transpose(1, 2), # [b * np, hn, sk]
|
263 |
+
beta=0.0,
|
264 |
+
alpha=(1.0 / self.norm_factor),
|
265 |
+
)
|
266 |
+
|
267 |
+
# change view to [b, np, sq, sk]
|
268 |
+
attention_scores = matmul_result.view(*output_size)
|
269 |
+
|
270 |
+
# ===========================
|
271 |
+
# Attention probs and dropout
|
272 |
+
# ===========================
|
273 |
+
|
274 |
+
# attention scores and attention mask [b, np, sq, sk]
|
275 |
+
if self.attention_softmax_in_fp32:
|
276 |
+
attention_scores = attention_scores.float()
|
277 |
+
if self.coeff is not None:
|
278 |
+
attention_scores = attention_scores * self.coeff
|
279 |
+
if attention_mask is None and attention_scores.shape[2] == attention_scores.shape[3]:
|
280 |
+
attention_mask = torch.ones(output_size[0], 1, output_size[2], output_size[3],
|
281 |
+
device=attention_scores.device, dtype=torch.bool)
|
282 |
+
attention_mask.tril_()
|
283 |
+
attention_mask = ~attention_mask
|
284 |
+
if attention_mask is not None:
|
285 |
+
attention_scores = attention_scores.masked_fill(attention_mask, float("-inf"))
|
286 |
+
attention_probs = F.softmax(attention_scores, dim=-1)
|
287 |
+
attention_probs = attention_probs.type_as(value_layer)
|
288 |
+
|
289 |
+
# This is actually dropping out entire tokens to attend to, which might
|
290 |
+
# seem a bit unusual, but is taken from the original Transformer paper.
|
291 |
+
attention_probs = self.attention_dropout(attention_probs)
|
292 |
+
# =========================
|
293 |
+
# Context layer. [sq, b, hp]
|
294 |
+
# =========================
|
295 |
+
|
296 |
+
# value_layer -> context layer.
|
297 |
+
# [sk, b, np, hn] --> [b, np, sq, hn]
|
298 |
+
|
299 |
+
# context layer shape: [b, np, sq, hn]
|
300 |
+
output_size = (value_layer.size(1), value_layer.size(2), query_layer.size(0), value_layer.size(3))
|
301 |
+
# change view [sk, b * np, hn]
|
302 |
+
value_layer = value_layer.view(value_layer.size(0), output_size[0] * output_size[1], -1)
|
303 |
+
# change view [b * np, sq, sk]
|
304 |
+
attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
|
305 |
+
# matmul: [b * np, sq, hn]
|
306 |
+
context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))
|
307 |
+
# change view [b, np, sq, hn]
|
308 |
+
context_layer = context_layer.view(*output_size)
|
309 |
+
# [b, np, sq, hn] --> [sq, b, np, hn]
|
310 |
+
context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
|
311 |
+
# [sq, b, np, hn] --> [sq, b, hp]
|
312 |
+
new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
|
313 |
+
context_layer = context_layer.view(*new_context_layer_shape)
|
314 |
+
|
315 |
+
return context_layer
|
316 |
+
|
317 |
+
|
318 |
+
class SelfAttention(torch.nn.Module):
|
319 |
+
"""Parallel self-attention layer abstract class.
|
320 |
+
|
321 |
+
Self-attention layer takes input with size [s, b, h]
|
322 |
+
and returns output of the same size.
|
323 |
+
"""
|
324 |
+
|
325 |
+
def __init__(self, config: ChatGLMConfig, layer_number, device=None):
|
326 |
+
super(SelfAttention, self).__init__()
|
327 |
+
self.layer_number = max(1, layer_number)
|
328 |
+
|
329 |
+
self.projection_size = config.kv_channels * config.num_attention_heads
|
330 |
+
|
331 |
+
# Per attention head and per partition values.
|
332 |
+
self.hidden_size_per_attention_head = self.projection_size // config.num_attention_heads
|
333 |
+
self.num_attention_heads_per_partition = config.num_attention_heads
|
334 |
+
|
335 |
+
self.multi_query_attention = config.multi_query_attention
|
336 |
+
self.qkv_hidden_size = 3 * self.projection_size
|
337 |
+
if self.multi_query_attention:
|
338 |
+
self.num_multi_query_groups_per_partition = config.multi_query_group_num
|
339 |
+
self.qkv_hidden_size = (
|
340 |
+
self.projection_size + 2 * self.hidden_size_per_attention_head * config.multi_query_group_num
|
341 |
+
)
|
342 |
+
self.query_key_value = nn.Linear(config.hidden_size, self.qkv_hidden_size,
|
343 |
+
bias=config.add_bias_linear or config.add_qkv_bias,
|
344 |
+
device=device, **_config_to_kwargs(config)
|
345 |
+
)
|
346 |
+
|
347 |
+
self.core_attention = CoreAttention(config, self.layer_number)
|
348 |
+
|
349 |
+
# Output.
|
350 |
+
self.dense = nn.Linear(self.projection_size, config.hidden_size, bias=config.add_bias_linear,
|
351 |
+
device=device, **_config_to_kwargs(config)
|
352 |
+
)
|
353 |
+
|
354 |
+
def _allocate_memory(self, inference_max_sequence_len, batch_size, device=None, dtype=None):
|
355 |
+
if self.multi_query_attention:
|
356 |
+
num_attention_heads = self.num_multi_query_groups_per_partition
|
357 |
+
else:
|
358 |
+
num_attention_heads = self.num_attention_heads_per_partition
|
359 |
+
return torch.empty(
|
360 |
+
inference_max_sequence_len,
|
361 |
+
batch_size,
|
362 |
+
num_attention_heads,
|
363 |
+
self.hidden_size_per_attention_head,
|
364 |
+
dtype=dtype,
|
365 |
+
device=device,
|
366 |
+
)
|
367 |
+
|
368 |
+
def forward(
|
369 |
+
self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True
|
370 |
+
):
|
371 |
+
# hidden_states: [sq, b, h]
|
372 |
+
|
373 |
+
# =================================================
|
374 |
+
# Pre-allocate memory for key-values for inference.
|
375 |
+
# =================================================
|
376 |
+
# =====================
|
377 |
+
# Query, Key, and Value
|
378 |
+
# =====================
|
379 |
+
|
380 |
+
# Attention heads [sq, b, h] --> [sq, b, (np * 3 * hn)]
|
381 |
+
mixed_x_layer = self.query_key_value(hidden_states)
|
382 |
+
|
383 |
+
if self.multi_query_attention:
|
384 |
+
(query_layer, key_layer, value_layer) = mixed_x_layer.split(
|
385 |
+
[
|
386 |
+
self.num_attention_heads_per_partition * self.hidden_size_per_attention_head,
|
387 |
+
self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
|
388 |
+
self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
|
389 |
+
],
|
390 |
+
dim=-1,
|
391 |
+
)
|
392 |
+
query_layer = query_layer.view(
|
393 |
+
query_layer.size()[:-1] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
|
394 |
+
)
|
395 |
+
key_layer = key_layer.view(
|
396 |
+
key_layer.size()[:-1] + (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
|
397 |
+
)
|
398 |
+
value_layer = value_layer.view(
|
399 |
+
value_layer.size()[:-1]
|
400 |
+
+ (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
|
401 |
+
)
|
402 |
+
else:
|
403 |
+
new_tensor_shape = mixed_x_layer.size()[:-1] + \
|
404 |
+
(self.num_attention_heads_per_partition,
|
405 |
+
3 * self.hidden_size_per_attention_head)
|
406 |
+
mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
|
407 |
+
|
408 |
+
# [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn]
|
409 |
+
(query_layer, key_layer, value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3)
|
410 |
+
|
411 |
+
# apply relative positional encoding (rotary embedding)
|
412 |
+
if rotary_pos_emb is not None:
|
413 |
+
query_layer = apply_rotary_pos_emb(query_layer, rotary_pos_emb)
|
414 |
+
key_layer = apply_rotary_pos_emb(key_layer, rotary_pos_emb)
|
415 |
+
|
416 |
+
# adjust key and value for inference
|
417 |
+
if kv_cache is not None:
|
418 |
+
cache_k, cache_v = kv_cache
|
419 |
+
key_layer = torch.cat((cache_k, key_layer), dim=0)
|
420 |
+
value_layer = torch.cat((cache_v, value_layer), dim=0)
|
421 |
+
if use_cache:
|
422 |
+
kv_cache = (key_layer, value_layer)
|
423 |
+
else:
|
424 |
+
kv_cache = None
|
425 |
+
|
426 |
+
if self.multi_query_attention:
|
427 |
+
key_layer = key_layer.unsqueeze(-2)
|
428 |
+
key_layer = key_layer.expand(
|
429 |
+
-1, -1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1
|
430 |
+
)
|
431 |
+
key_layer = key_layer.contiguous().view(
|
432 |
+
key_layer.size()[:2] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
|
433 |
+
)
|
434 |
+
value_layer = value_layer.unsqueeze(-2)
|
435 |
+
value_layer = value_layer.expand(
|
436 |
+
-1, -1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1
|
437 |
+
)
|
438 |
+
value_layer = value_layer.contiguous().view(
|
439 |
+
value_layer.size()[:2] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
|
440 |
+
)
|
441 |
+
|
442 |
+
# ==================================
|
443 |
+
# core attention computation
|
444 |
+
# ==================================
|
445 |
+
|
446 |
+
context_layer = self.core_attention(query_layer, key_layer, value_layer, attention_mask)
|
447 |
+
|
448 |
+
# =================
|
449 |
+
# Output. [sq, b, h]
|
450 |
+
# =================
|
451 |
+
|
452 |
+
output = self.dense(context_layer)
|
453 |
+
|
454 |
+
return output, kv_cache
|
455 |
+
|
456 |
+
|
457 |
+
def _config_to_kwargs(args):
|
458 |
+
common_kwargs = {
|
459 |
+
"dtype": args.torch_dtype,
|
460 |
+
}
|
461 |
+
return common_kwargs
|
462 |
+
|
463 |
+
|
464 |
+
class MLP(torch.nn.Module):
|
465 |
+
"""MLP.
|
466 |
+
|
467 |
+
MLP will take the input with h hidden state, project it to 4*h
|
468 |
+
hidden dimension, perform nonlinear transformation, and project the
|
469 |
+
state back into h hidden dimension.
|
470 |
+
"""
|
471 |
+
|
472 |
+
def __init__(self, config: ChatGLMConfig, device=None):
|
473 |
+
super(MLP, self).__init__()
|
474 |
+
|
475 |
+
self.add_bias = config.add_bias_linear
|
476 |
+
|
477 |
+
# Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
|
478 |
+
self.dense_h_to_4h = nn.Linear(
|
479 |
+
config.hidden_size,
|
480 |
+
config.ffn_hidden_size * 2,
|
481 |
+
bias=self.add_bias,
|
482 |
+
device=device,
|
483 |
+
**_config_to_kwargs(config)
|
484 |
+
)
|
485 |
+
|
486 |
+
def swiglu(x):
|
487 |
+
x = torch.chunk(x, 2, dim=-1)
|
488 |
+
return F.silu(x[0]) * x[1]
|
489 |
+
|
490 |
+
self.activation_func = swiglu
|
491 |
+
|
492 |
+
# Project back to h.
|
493 |
+
self.dense_4h_to_h = nn.Linear(
|
494 |
+
config.ffn_hidden_size,
|
495 |
+
config.hidden_size,
|
496 |
+
bias=self.add_bias,
|
497 |
+
device=device,
|
498 |
+
**_config_to_kwargs(config)
|
499 |
+
)
|
500 |
+
|
501 |
+
def forward(self, hidden_states):
|
502 |
+
# [s, b, 4hp]
|
503 |
+
intermediate_parallel = self.dense_h_to_4h(hidden_states)
|
504 |
+
intermediate_parallel = self.activation_func(intermediate_parallel)
|
505 |
+
# [s, b, h]
|
506 |
+
output = self.dense_4h_to_h(intermediate_parallel)
|
507 |
+
return output
|
508 |
+
|
509 |
+
|
510 |
+
class GLMBlock(torch.nn.Module):
|
511 |
+
"""A single transformer layer.
|
512 |
+
|
513 |
+
Transformer layer takes input with size [s, b, h] and returns an
|
514 |
+
output of the same size.
|
515 |
+
"""
|
516 |
+
|
517 |
+
def __init__(self, config: ChatGLMConfig, layer_number, device=None):
|
518 |
+
super(GLMBlock, self).__init__()
|
519 |
+
self.layer_number = layer_number
|
520 |
+
|
521 |
+
self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm
|
522 |
+
|
523 |
+
self.fp32_residual_connection = config.fp32_residual_connection
|
524 |
+
|
525 |
+
LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
|
526 |
+
# Layernorm on the input data.
|
527 |
+
self.input_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
|
528 |
+
dtype=config.torch_dtype)
|
529 |
+
|
530 |
+
# Self attention.
|
531 |
+
self.self_attention = SelfAttention(config, layer_number, device=device)
|
532 |
+
self.hidden_dropout = config.hidden_dropout
|
533 |
+
|
534 |
+
# Layernorm on the attention output
|
535 |
+
self.post_attention_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
|
536 |
+
dtype=config.torch_dtype)
|
537 |
+
|
538 |
+
# MLP
|
539 |
+
self.mlp = MLP(config, device=device)
|
540 |
+
|
541 |
+
def forward(
|
542 |
+
self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True,
|
543 |
+
):
|
544 |
+
# hidden_states: [s, b, h]
|
545 |
+
|
546 |
+
# Layer norm at the beginning of the transformer layer.
|
547 |
+
layernorm_output = self.input_layernorm(hidden_states)
|
548 |
+
# Self attention.
|
549 |
+
attention_output, kv_cache = self.self_attention(
|
550 |
+
layernorm_output,
|
551 |
+
attention_mask,
|
552 |
+
rotary_pos_emb,
|
553 |
+
kv_cache=kv_cache,
|
554 |
+
use_cache=use_cache
|
555 |
+
)
|
556 |
+
|
557 |
+
# Residual connection.
|
558 |
+
if self.apply_residual_connection_post_layernorm:
|
559 |
+
residual = layernorm_output
|
560 |
+
else:
|
561 |
+
residual = hidden_states
|
562 |
+
|
563 |
+
layernorm_input = torch.nn.functional.dropout(attention_output, p=self.hidden_dropout, training=self.training)
|
564 |
+
layernorm_input = residual + layernorm_input
|
565 |
+
|
566 |
+
# Layer norm post the self attention.
|
567 |
+
layernorm_output = self.post_attention_layernorm(layernorm_input)
|
568 |
+
|
569 |
+
# MLP.
|
570 |
+
mlp_output = self.mlp(layernorm_output)
|
571 |
+
|
572 |
+
# Second residual connection.
|
573 |
+
if self.apply_residual_connection_post_layernorm:
|
574 |
+
residual = layernorm_output
|
575 |
+
else:
|
576 |
+
residual = layernorm_input
|
577 |
+
|
578 |
+
output = torch.nn.functional.dropout(mlp_output, p=self.hidden_dropout, training=self.training)
|
579 |
+
output = residual + output
|
580 |
+
|
581 |
+
return output, kv_cache
|
582 |
+
|
583 |
+
|
584 |
+
class GLMTransformer(torch.nn.Module):
|
585 |
+
"""Transformer class."""
|
586 |
+
|
587 |
+
def __init__(self, config: ChatGLMConfig, device=None):
|
588 |
+
super(GLMTransformer, self).__init__()
|
589 |
+
|
590 |
+
self.fp32_residual_connection = config.fp32_residual_connection
|
591 |
+
self.post_layer_norm = config.post_layer_norm
|
592 |
+
|
593 |
+
# Number of layers.
|
594 |
+
self.num_layers = config.num_layers
|
595 |
+
|
596 |
+
# Transformer layers.
|
597 |
+
def build_layer(layer_number):
|
598 |
+
return GLMBlock(config, layer_number, device=device)
|
599 |
+
|
600 |
+
self.layers = torch.nn.ModuleList([build_layer(i + 1) for i in range(self.num_layers)])
|
601 |
+
|
602 |
+
if self.post_layer_norm:
|
603 |
+
LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
|
604 |
+
# Final layer norm before output.
|
605 |
+
self.final_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
|
606 |
+
dtype=config.torch_dtype)
|
607 |
+
|
608 |
+
self.gradient_checkpointing = False
|
609 |
+
|
610 |
+
def _get_layer(self, layer_number):
|
611 |
+
return self.layers[layer_number]
|
612 |
+
|
613 |
+
def forward(
|
614 |
+
self, hidden_states, attention_mask, rotary_pos_emb, kv_caches=None,
|
615 |
+
use_cache: Optional[bool] = True,
|
616 |
+
output_hidden_states: Optional[bool] = False,
|
617 |
+
):
|
618 |
+
if not kv_caches:
|
619 |
+
kv_caches = [None for _ in range(self.num_layers)]
|
620 |
+
presents = () if use_cache else None
|
621 |
+
if self.gradient_checkpointing and self.training:
|
622 |
+
if use_cache:
|
623 |
+
logger.warning_once(
|
624 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
625 |
+
)
|
626 |
+
use_cache = False
|
627 |
+
|
628 |
+
all_self_attentions = None
|
629 |
+
all_hidden_states = () if output_hidden_states else None
|
630 |
+
for index in range(self.num_layers):
|
631 |
+
if output_hidden_states:
|
632 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
633 |
+
|
634 |
+
layer = self._get_layer(index)
|
635 |
+
if self.gradient_checkpointing and self.training:
|
636 |
+
layer_ret = torch.utils.checkpoint.checkpoint(
|
637 |
+
layer,
|
638 |
+
hidden_states,
|
639 |
+
attention_mask,
|
640 |
+
rotary_pos_emb,
|
641 |
+
kv_caches[index],
|
642 |
+
use_cache
|
643 |
+
)
|
644 |
+
else:
|
645 |
+
layer_ret = layer(
|
646 |
+
hidden_states,
|
647 |
+
attention_mask,
|
648 |
+
rotary_pos_emb,
|
649 |
+
kv_cache=kv_caches[index],
|
650 |
+
use_cache=use_cache
|
651 |
+
)
|
652 |
+
hidden_states, kv_cache = layer_ret
|
653 |
+
if use_cache:
|
654 |
+
presents = presents + (kv_cache,)
|
655 |
+
|
656 |
+
if output_hidden_states:
|
657 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
658 |
+
|
659 |
+
# Final layer norm.
|
660 |
+
if self.post_layer_norm:
|
661 |
+
hidden_states = self.final_layernorm(hidden_states)
|
662 |
+
|
663 |
+
return hidden_states, presents, all_hidden_states, all_self_attentions
|
664 |
+
|
665 |
+
|
666 |
+
class ChatGLMPreTrainedModel(PreTrainedModel):
|
667 |
+
"""
|
668 |
+
An abstract class to handle weights initialization and
|
669 |
+
a simple interface for downloading and loading pretrained models.
|
670 |
+
"""
|
671 |
+
|
672 |
+
is_parallelizable = False
|
673 |
+
supports_gradient_checkpointing = True
|
674 |
+
config_class = ChatGLMConfig
|
675 |
+
base_model_prefix = "transformer"
|
676 |
+
_no_split_modules = ["GLMBlock"]
|
677 |
+
|
678 |
+
def _init_weights(self, module: nn.Module):
|
679 |
+
"""Initialize the weights."""
|
680 |
+
return
|
681 |
+
|
682 |
+
def get_masks(self, input_ids, past_key_values, padding_mask=None):
|
683 |
+
batch_size, seq_length = input_ids.shape
|
684 |
+
full_attention_mask = torch.ones(batch_size, seq_length, seq_length, device=input_ids.device)
|
685 |
+
full_attention_mask.tril_()
|
686 |
+
past_length = 0
|
687 |
+
if past_key_values:
|
688 |
+
past_length = past_key_values[0][0].shape[0]
|
689 |
+
if past_length:
|
690 |
+
full_attention_mask = torch.cat((torch.ones(batch_size, seq_length, past_length,
|
691 |
+
device=input_ids.device), full_attention_mask), dim=-1)
|
692 |
+
if padding_mask is not None:
|
693 |
+
full_attention_mask = full_attention_mask * padding_mask.unsqueeze(1)
|
694 |
+
if not past_length and padding_mask is not None:
|
695 |
+
full_attention_mask -= padding_mask.unsqueeze(-1) - 1
|
696 |
+
full_attention_mask = (full_attention_mask < 0.5).bool()
|
697 |
+
full_attention_mask.unsqueeze_(1)
|
698 |
+
return full_attention_mask
|
699 |
+
|
700 |
+
def get_position_ids(self, input_ids, device):
|
701 |
+
batch_size, seq_length = input_ids.shape
|
702 |
+
position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
|
703 |
+
return position_ids
|
704 |
+
|
705 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
706 |
+
if isinstance(module, GLMTransformer):
|
707 |
+
module.gradient_checkpointing = value
|
708 |
+
|
709 |
+
|
710 |
+
class Embedding(torch.nn.Module):
|
711 |
+
"""Language model embeddings."""
|
712 |
+
|
713 |
+
def __init__(self, config: ChatGLMConfig, device=None):
|
714 |
+
super(Embedding, self).__init__()
|
715 |
+
|
716 |
+
self.hidden_size = config.hidden_size
|
717 |
+
# Word embeddings (parallel).
|
718 |
+
self.word_embeddings = nn.Embedding(
|
719 |
+
config.padded_vocab_size,
|
720 |
+
self.hidden_size,
|
721 |
+
dtype=config.torch_dtype,
|
722 |
+
device=device
|
723 |
+
)
|
724 |
+
self.fp32_residual_connection = config.fp32_residual_connection
|
725 |
+
|
726 |
+
def forward(self, input_ids):
|
727 |
+
# Embeddings.
|
728 |
+
words_embeddings = self.word_embeddings(input_ids)
|
729 |
+
embeddings = words_embeddings
|
730 |
+
# Data format change to avoid explicit tranposes : [b s h] --> [s b h].
|
731 |
+
embeddings = embeddings.transpose(0, 1).contiguous()
|
732 |
+
# If the input flag for fp32 residual connection is set, convert for float.
|
733 |
+
if self.fp32_residual_connection:
|
734 |
+
embeddings = embeddings.float()
|
735 |
+
return embeddings
|
736 |
+
|
737 |
+
|
738 |
+
class ChatGLMModel(ChatGLMPreTrainedModel):
|
739 |
+
def __init__(self, config: ChatGLMConfig, device=None, empty_init=True):
|
740 |
+
super().__init__(config)
|
741 |
+
if empty_init:
|
742 |
+
init_method = skip_init
|
743 |
+
else:
|
744 |
+
init_method = default_init
|
745 |
+
init_kwargs = {}
|
746 |
+
if device is not None:
|
747 |
+
init_kwargs["device"] = device
|
748 |
+
self.embedding = init_method(Embedding, config, **init_kwargs)
|
749 |
+
self.num_layers = config.num_layers
|
750 |
+
self.multi_query_group_num = config.multi_query_group_num
|
751 |
+
self.kv_channels = config.kv_channels
|
752 |
+
|
753 |
+
# Rotary positional embeddings
|
754 |
+
self.seq_length = config.seq_length
|
755 |
+
rotary_dim = (
|
756 |
+
config.hidden_size // config.num_attention_heads if config.kv_channels is None else config.kv_channels
|
757 |
+
)
|
758 |
+
|
759 |
+
self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, original_impl=config.original_rope, device=device,
|
760 |
+
dtype=config.torch_dtype)
|
761 |
+
self.encoder = init_method(GLMTransformer, config, **init_kwargs)
|
762 |
+
self.output_layer = init_method(nn.Linear, config.hidden_size, config.padded_vocab_size, bias=False,
|
763 |
+
dtype=config.torch_dtype, **init_kwargs)
|
764 |
+
self.pre_seq_len = config.pre_seq_len
|
765 |
+
self.prefix_projection = config.prefix_projection
|
766 |
+
if self.pre_seq_len is not None:
|
767 |
+
for param in self.parameters():
|
768 |
+
param.requires_grad = False
|
769 |
+
self.prefix_tokens = torch.arange(self.pre_seq_len).long()
|
770 |
+
self.prefix_encoder = PrefixEncoder(config)
|
771 |
+
self.dropout = torch.nn.Dropout(0.1)
|
772 |
+
|
773 |
+
def get_input_embeddings(self):
|
774 |
+
return self.embedding.word_embeddings
|
775 |
+
|
776 |
+
def get_prompt(self, batch_size, device, dtype=torch.half):
|
777 |
+
prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(device)
|
778 |
+
past_key_values = self.prefix_encoder(prefix_tokens).type(dtype)
|
779 |
+
past_key_values = past_key_values.view(
|
780 |
+
batch_size,
|
781 |
+
self.pre_seq_len,
|
782 |
+
self.num_layers * 2,
|
783 |
+
self.multi_query_group_num,
|
784 |
+
self.kv_channels
|
785 |
+
)
|
786 |
+
# seq_len, b, nh, hidden_size
|
787 |
+
past_key_values = self.dropout(past_key_values)
|
788 |
+
past_key_values = past_key_values.permute([2, 1, 0, 3, 4]).split(2)
|
789 |
+
return past_key_values
|
790 |
+
|
791 |
+
def forward(
|
792 |
+
self,
|
793 |
+
input_ids,
|
794 |
+
position_ids: Optional[torch.Tensor] = None,
|
795 |
+
attention_mask: Optional[torch.BoolTensor] = None,
|
796 |
+
full_attention_mask: Optional[torch.BoolTensor] = None,
|
797 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
|
798 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
799 |
+
use_cache: Optional[bool] = None,
|
800 |
+
output_hidden_states: Optional[bool] = None,
|
801 |
+
return_dict: Optional[bool] = None,
|
802 |
+
):
|
803 |
+
output_hidden_states = (
|
804 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
805 |
+
)
|
806 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
807 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
808 |
+
|
809 |
+
batch_size, seq_length = input_ids.shape
|
810 |
+
|
811 |
+
if inputs_embeds is None:
|
812 |
+
inputs_embeds = self.embedding(input_ids)
|
813 |
+
|
814 |
+
if self.pre_seq_len is not None:
|
815 |
+
if past_key_values is None:
|
816 |
+
past_key_values = self.get_prompt(batch_size=batch_size, device=input_ids.device,
|
817 |
+
dtype=inputs_embeds.dtype)
|
818 |
+
if attention_mask is not None:
|
819 |
+
attention_mask = torch.cat([attention_mask.new_ones((batch_size, self.pre_seq_len)),
|
820 |
+
attention_mask], dim=-1)
|
821 |
+
|
822 |
+
if full_attention_mask is None:
|
823 |
+
if (attention_mask is not None and not attention_mask.all()) or (past_key_values and seq_length != 1):
|
824 |
+
full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask)
|
825 |
+
|
826 |
+
# Rotary positional embeddings
|
827 |
+
rotary_pos_emb = self.rotary_pos_emb(self.seq_length)
|
828 |
+
if position_ids is not None:
|
829 |
+
rotary_pos_emb = rotary_pos_emb[position_ids]
|
830 |
+
else:
|
831 |
+
rotary_pos_emb = rotary_pos_emb[None, :seq_length]
|
832 |
+
rotary_pos_emb = rotary_pos_emb.transpose(0, 1).contiguous()
|
833 |
+
|
834 |
+
# Run encoder.
|
835 |
+
hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder(
|
836 |
+
inputs_embeds, full_attention_mask, rotary_pos_emb=rotary_pos_emb,
|
837 |
+
kv_caches=past_key_values, use_cache=use_cache, output_hidden_states=output_hidden_states
|
838 |
+
)
|
839 |
+
|
840 |
+
if not return_dict:
|
841 |
+
return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
|
842 |
+
|
843 |
+
return BaseModelOutputWithPast(
|
844 |
+
last_hidden_state=hidden_states,
|
845 |
+
past_key_values=presents,
|
846 |
+
hidden_states=all_hidden_states,
|
847 |
+
attentions=all_self_attentions,
|
848 |
+
)
|
849 |
+
|
850 |
+
def quantize(self, weight_bit_width: int):
|
851 |
+
from .quantization import quantize
|
852 |
+
quantize(self.encoder, weight_bit_width)
|
853 |
+
return self
|
854 |
+
|
855 |
+
|
856 |
+
class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
|
857 |
+
def __init__(self, config: ChatGLMConfig, empty_init=True, device=None):
|
858 |
+
super().__init__(config)
|
859 |
+
|
860 |
+
self.max_sequence_length = config.max_length
|
861 |
+
self.transformer = ChatGLMModel(config, empty_init=empty_init, device=device)
|
862 |
+
self.config = config
|
863 |
+
self.quantized = False
|
864 |
+
|
865 |
+
if self.config.quantization_bit:
|
866 |
+
self.quantize(self.config.quantization_bit, empty_init=True)
|
867 |
+
|
868 |
+
def _update_model_kwargs_for_generation(
|
869 |
+
self,
|
870 |
+
outputs: ModelOutput,
|
871 |
+
model_kwargs: Dict[str, Any],
|
872 |
+
is_encoder_decoder: bool = False,
|
873 |
+
standardize_cache_format: bool = False,
|
874 |
+
) -> Dict[str, Any]:
|
875 |
+
# update past_key_values
|
876 |
+
model_kwargs["past_key_values"] = self._extract_past_from_model_output(
|
877 |
+
outputs, standardize_cache_format=standardize_cache_format
|
878 |
+
)
|
879 |
+
|
880 |
+
# update attention mask
|
881 |
+
if "attention_mask" in model_kwargs:
|
882 |
+
attention_mask = model_kwargs["attention_mask"]
|
883 |
+
model_kwargs["attention_mask"] = torch.cat(
|
884 |
+
[attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
|
885 |
+
)
|
886 |
+
|
887 |
+
# update position ids
|
888 |
+
if "position_ids" in model_kwargs:
|
889 |
+
position_ids = model_kwargs["position_ids"]
|
890 |
+
new_position_id = position_ids[..., -1:].clone()
|
891 |
+
new_position_id += 1
|
892 |
+
model_kwargs["position_ids"] = torch.cat(
|
893 |
+
[position_ids, new_position_id], dim=-1
|
894 |
+
)
|
895 |
+
|
896 |
+
model_kwargs["is_first_forward"] = False
|
897 |
+
return model_kwargs
|
898 |
+
|
899 |
+
def prepare_inputs_for_generation(
|
900 |
+
self,
|
901 |
+
input_ids: torch.LongTensor,
|
902 |
+
past_key_values: Optional[torch.Tensor] = None,
|
903 |
+
attention_mask: Optional[torch.Tensor] = None,
|
904 |
+
position_ids: Optional[torch.Tensor] = None,
|
905 |
+
use_cache: Optional[bool] = None,
|
906 |
+
is_first_forward: bool = True,
|
907 |
+
**kwargs
|
908 |
+
) -> dict:
|
909 |
+
# only last token for input_ids if past is not None
|
910 |
+
if position_ids is None:
|
911 |
+
position_ids = self.get_position_ids(input_ids, device=input_ids.device)
|
912 |
+
if not is_first_forward:
|
913 |
+
if past_key_values is not None:
|
914 |
+
position_ids = position_ids[..., -1:]
|
915 |
+
input_ids = input_ids[:, -1:]
|
916 |
+
return {
|
917 |
+
"input_ids": input_ids,
|
918 |
+
"past_key_values": past_key_values,
|
919 |
+
"position_ids": position_ids,
|
920 |
+
"attention_mask": attention_mask,
|
921 |
+
"return_last_logit": True,
|
922 |
+
"use_cache": use_cache
|
923 |
+
}
|
924 |
+
|
925 |
+
def forward(
|
926 |
+
self,
|
927 |
+
input_ids: Optional[torch.Tensor] = None,
|
928 |
+
position_ids: Optional[torch.Tensor] = None,
|
929 |
+
attention_mask: Optional[torch.Tensor] = None,
|
930 |
+
past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
|
931 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
932 |
+
labels: Optional[torch.Tensor] = None,
|
933 |
+
use_cache: Optional[bool] = None,
|
934 |
+
output_attentions: Optional[bool] = None,
|
935 |
+
output_hidden_states: Optional[bool] = None,
|
936 |
+
return_dict: Optional[bool] = None,
|
937 |
+
return_last_logit: Optional[bool] = False,
|
938 |
+
):
|
939 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
940 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
941 |
+
|
942 |
+
transformer_outputs = self.transformer(
|
943 |
+
input_ids=input_ids,
|
944 |
+
position_ids=position_ids,
|
945 |
+
attention_mask=attention_mask,
|
946 |
+
past_key_values=past_key_values,
|
947 |
+
inputs_embeds=inputs_embeds,
|
948 |
+
use_cache=use_cache,
|
949 |
+
output_hidden_states=output_hidden_states,
|
950 |
+
return_dict=return_dict,
|
951 |
+
)
|
952 |
+
|
953 |
+
hidden_states = transformer_outputs[0]
|
954 |
+
if return_last_logit:
|
955 |
+
hidden_states = hidden_states[-1:]
|
956 |
+
lm_logits = self.transformer.output_layer(hidden_states)
|
957 |
+
lm_logits = lm_logits.transpose(0, 1).contiguous()
|
958 |
+
|
959 |
+
loss = None
|
960 |
+
if labels is not None:
|
961 |
+
lm_logits = lm_logits.to(torch.float32)
|
962 |
+
|
963 |
+
# Shift so that tokens < n predict n
|
964 |
+
shift_logits = lm_logits[..., :-1, :].contiguous()
|
965 |
+
shift_labels = labels[..., 1:].contiguous()
|
966 |
+
# Flatten the tokens
|
967 |
+
loss_fct = CrossEntropyLoss(ignore_index=-100)
|
968 |
+
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
969 |
+
|
970 |
+
lm_logits = lm_logits.to(hidden_states.dtype)
|
971 |
+
loss = loss.to(hidden_states.dtype)
|
972 |
+
|
973 |
+
if not return_dict:
|
974 |
+
output = (lm_logits,) + transformer_outputs[1:]
|
975 |
+
return ((loss,) + output) if loss is not None else output
|
976 |
+
|
977 |
+
return CausalLMOutputWithPast(
|
978 |
+
loss=loss,
|
979 |
+
logits=lm_logits,
|
980 |
+
past_key_values=transformer_outputs.past_key_values,
|
981 |
+
hidden_states=transformer_outputs.hidden_states,
|
982 |
+
attentions=transformer_outputs.attentions,
|
983 |
+
)
|
984 |
+
|
985 |
+
@staticmethod
|
986 |
+
def _reorder_cache(
|
987 |
+
past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor
|
988 |
+
) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]:
|
989 |
+
"""
|
990 |
+
This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
|
991 |
+
[`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
|
992 |
+
beam_idx at every generation step.
|
993 |
+
|
994 |
+
Output shares the same memory storage as `past`.
|
995 |
+
"""
|
996 |
+
return tuple(
|
997 |
+
(
|
998 |
+
layer_past[0].index_select(1, beam_idx.to(layer_past[0].device)),
|
999 |
+
layer_past[1].index_select(1, beam_idx.to(layer_past[1].device)),
|
1000 |
+
)
|
1001 |
+
for layer_past in past
|
1002 |
+
)
|
1003 |
+
|
1004 |
+
def process_response(self, output, history):
|
1005 |
+
content = ""
|
1006 |
+
history = deepcopy(history)
|
1007 |
+
for response in output.split("<|assistant|>"):
|
1008 |
+
metadata, content = response.split("\n", maxsplit=1)
|
1009 |
+
if not metadata.strip():
|
1010 |
+
content = content.strip()
|
1011 |
+
history.append({"role": "assistant", "metadata": metadata, "content": content})
|
1012 |
+
content = content.replace("[[训练时间]]", "2023年")
|
1013 |
+
else:
|
1014 |
+
history.append({"role": "assistant", "metadata": metadata, "content": content})
|
1015 |
+
if history[0]["role"] == "system" and "tools" in history[0]:
|
1016 |
+
content = "\n".join(content.split("\n")[1:-1])
|
1017 |
+
def tool_call(**kwargs):
|
1018 |
+
return kwargs
|
1019 |
+
parameters = eval(content)
|
1020 |
+
content = {"name": metadata.strip(), "parameters": parameters}
|
1021 |
+
else:
|
1022 |
+
content = {"name": metadata.strip(), "content": content}
|
1023 |
+
return content, history
|
1024 |
+
|
1025 |
+
@torch.inference_mode()
|
1026 |
+
def chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, role: str = "user",
|
1027 |
+
max_length: int = 8192, num_beams=1, do_sample=True, top_p=0.8, temperature=0.8, logits_processor=None,
|
1028 |
+
**kwargs):
|
1029 |
+
if history is None:
|
1030 |
+
history = []
|
1031 |
+
if logits_processor is None:
|
1032 |
+
logits_processor = LogitsProcessorList()
|
1033 |
+
logits_processor.append(InvalidScoreLogitsProcessor())
|
1034 |
+
gen_kwargs = {"max_length": max_length, "num_beams": num_beams, "do_sample": do_sample, "top_p": top_p,
|
1035 |
+
"temperature": temperature, "logits_processor": logits_processor, **kwargs}
|
1036 |
+
inputs = tokenizer.build_chat_input(query, history=history, role=role)
|
1037 |
+
inputs = inputs.to(self.device)
|
1038 |
+
eos_token_id = [tokenizer.eos_token_id, tokenizer.get_command("<|user|>"),
|
1039 |
+
tokenizer.get_command("<|observation|>")]
|
1040 |
+
outputs = self.generate(**inputs, **gen_kwargs, eos_token_id=eos_token_id)
|
1041 |
+
outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):-1]
|
1042 |
+
response = tokenizer.decode(outputs)
|
1043 |
+
history.append({"role": role, "content": query})
|
1044 |
+
response, history = self.process_response(response, history)
|
1045 |
+
return response, history
|
1046 |
+
|
1047 |
+
@torch.inference_mode()
|
1048 |
+
def stream_chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, role: str = "user",
|
1049 |
+
past_key_values=None,max_length: int = 8192, do_sample=True, top_p=0.8, temperature=0.8,
|
1050 |
+
logits_processor=None, return_past_key_values=False, **kwargs):
|
1051 |
+
if history is None:
|
1052 |
+
history = []
|
1053 |
+
if logits_processor is None:
|
1054 |
+
logits_processor = LogitsProcessorList()
|
1055 |
+
logits_processor.append(InvalidScoreLogitsProcessor())
|
1056 |
+
eos_token_id = [tokenizer.eos_token_id, tokenizer.get_command("<|user|>"),
|
1057 |
+
tokenizer.get_command("<|observation|>")]
|
1058 |
+
gen_kwargs = {"max_length": max_length, "do_sample": do_sample, "top_p": top_p,
|
1059 |
+
"temperature": temperature, "logits_processor": logits_processor, **kwargs}
|
1060 |
+
if past_key_values is None:
|
1061 |
+
inputs = tokenizer.build_chat_input(query, history=history, role=role)
|
1062 |
+
else:
|
1063 |
+
inputs = tokenizer.build_chat_input(query, role=role)
|
1064 |
+
inputs = inputs.to(self.device)
|
1065 |
+
if past_key_values is not None:
|
1066 |
+
past_length = past_key_values[0][0].shape[0]
|
1067 |
+
if self.transformer.pre_seq_len is not None:
|
1068 |
+
past_length -= self.transformer.pre_seq_len
|
1069 |
+
inputs.position_ids += past_length
|
1070 |
+
attention_mask = inputs.attention_mask
|
1071 |
+
attention_mask = torch.cat((attention_mask.new_ones(1, past_length), attention_mask), dim=1)
|
1072 |
+
inputs['attention_mask'] = attention_mask
|
1073 |
+
history.append({"role": role, "content": query})
|
1074 |
+
for outputs in self.stream_generate(**inputs, past_key_values=past_key_values,
|
1075 |
+
eos_token_id=eos_token_id, return_past_key_values=return_past_key_values,
|
1076 |
+
**gen_kwargs):
|
1077 |
+
if return_past_key_values:
|
1078 |
+
outputs, past_key_values = outputs
|
1079 |
+
outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):-1]
|
1080 |
+
response = tokenizer.decode(outputs)
|
1081 |
+
if response and response[-1] != "�":
|
1082 |
+
response, new_history = self.process_response(response, history)
|
1083 |
+
if return_past_key_values:
|
1084 |
+
yield response, new_history, past_key_values
|
1085 |
+
else:
|
1086 |
+
yield response, new_history
|
1087 |
+
|
1088 |
+
@torch.inference_mode()
|
1089 |
+
def stream_generate(
|
1090 |
+
self,
|
1091 |
+
input_ids,
|
1092 |
+
generation_config: Optional[GenerationConfig] = None,
|
1093 |
+
logits_processor: Optional[LogitsProcessorList] = None,
|
1094 |
+
stopping_criteria: Optional[StoppingCriteriaList] = None,
|
1095 |
+
prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
|
1096 |
+
return_past_key_values=False,
|
1097 |
+
**kwargs,
|
1098 |
+
):
|
1099 |
+
batch_size, input_ids_seq_length = input_ids.shape[0], input_ids.shape[-1]
|
1100 |
+
|
1101 |
+
if generation_config is None:
|
1102 |
+
generation_config = self.generation_config
|
1103 |
+
generation_config = copy.deepcopy(generation_config)
|
1104 |
+
model_kwargs = generation_config.update(**kwargs)
|
1105 |
+
model_kwargs["use_cache"] = generation_config.use_cache
|
1106 |
+
bos_token_id, eos_token_id = generation_config.bos_token_id, generation_config.eos_token_id
|
1107 |
+
|
1108 |
+
if isinstance(eos_token_id, int):
|
1109 |
+
eos_token_id = [eos_token_id]
|
1110 |
+
eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None
|
1111 |
+
|
1112 |
+
has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None
|
1113 |
+
if has_default_max_length and generation_config.max_new_tokens is None:
|
1114 |
+
warnings.warn(
|
1115 |
+
f"Using `max_length`'s default ({generation_config.max_length}) to control the generation length. "
|
1116 |
+
"This behaviour is deprecated and will be removed from the config in v5 of Transformers -- we"
|
1117 |
+
" recommend using `max_new_tokens` to control the maximum length of the generation.",
|
1118 |
+
UserWarning,
|
1119 |
+
)
|
1120 |
+
elif generation_config.max_new_tokens is not None:
|
1121 |
+
generation_config.max_length = generation_config.max_new_tokens + input_ids_seq_length
|
1122 |
+
if not has_default_max_length:
|
1123 |
+
logger.warn(
|
1124 |
+
f"Both `max_new_tokens` (={generation_config.max_new_tokens}) and `max_length`(="
|
1125 |
+
f"{generation_config.max_length}) seem to have been set. `max_new_tokens` will take precedence. "
|
1126 |
+
"Please refer to the documentation for more information. "
|
1127 |
+
"(https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)",
|
1128 |
+
UserWarning,
|
1129 |
+
)
|
1130 |
+
|
1131 |
+
if input_ids_seq_length >= generation_config.max_length:
|
1132 |
+
input_ids_string = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids"
|
1133 |
+
logger.warning(
|
1134 |
+
f"Input length of {input_ids_string} is {input_ids_seq_length}, but `max_length` is set to"
|
1135 |
+
f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider"
|
1136 |
+
" increasing `max_new_tokens`."
|
1137 |
+
)
|
1138 |
+
|
1139 |
+
# 2. Set generation parameters if not already defined
|
1140 |
+
logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
|
1141 |
+
stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
|
1142 |
+
|
1143 |
+
logits_processor = self._get_logits_processor(
|
1144 |
+
generation_config=generation_config,
|
1145 |
+
input_ids_seq_length=input_ids_seq_length,
|
1146 |
+
encoder_input_ids=input_ids,
|
1147 |
+
prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
|
1148 |
+
logits_processor=logits_processor,
|
1149 |
+
)
|
1150 |
+
|
1151 |
+
stopping_criteria = self._get_stopping_criteria(
|
1152 |
+
generation_config=generation_config, stopping_criteria=stopping_criteria
|
1153 |
+
)
|
1154 |
+
logits_warper = self._get_logits_warper(generation_config)
|
1155 |
+
|
1156 |
+
unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1)
|
1157 |
+
scores = None
|
1158 |
+
while True:
|
1159 |
+
model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
|
1160 |
+
# forward pass to get next token
|
1161 |
+
outputs = self(
|
1162 |
+
**model_inputs,
|
1163 |
+
return_dict=True,
|
1164 |
+
output_attentions=False,
|
1165 |
+
output_hidden_states=False,
|
1166 |
+
)
|
1167 |
+
|
1168 |
+
next_token_logits = outputs.logits[:, -1, :]
|
1169 |
+
|
1170 |
+
# pre-process distribution
|
1171 |
+
next_token_scores = logits_processor(input_ids, next_token_logits)
|
1172 |
+
next_token_scores = logits_warper(input_ids, next_token_scores)
|
1173 |
+
|
1174 |
+
# sample
|
1175 |
+
probs = nn.functional.softmax(next_token_scores, dim=-1)
|
1176 |
+
if generation_config.do_sample:
|
1177 |
+
next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
|
1178 |
+
else:
|
1179 |
+
next_tokens = torch.argmax(probs, dim=-1)
|
1180 |
+
# update generated ids, model inputs, and length for next step
|
1181 |
+
input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
|
1182 |
+
model_kwargs = self._update_model_kwargs_for_generation(
|
1183 |
+
outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
|
1184 |
+
)
|
1185 |
+
unfinished_sequences = unfinished_sequences.mul(
|
1186 |
+
next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0)
|
1187 |
+
)
|
1188 |
+
if return_past_key_values:
|
1189 |
+
yield input_ids, outputs.past_key_values
|
1190 |
+
else:
|
1191 |
+
yield input_ids
|
1192 |
+
# stop when each sentence is finished, or if we exceed the maximum length
|
1193 |
+
if unfinished_sequences.max() == 0 or stopping_criteria(input_ids, scores):
|
1194 |
+
break
|
1195 |
+
|
1196 |
+
def quantize(self, bits: int, empty_init=False, device=None, **kwargs):
|
1197 |
+
if bits == 0:
|
1198 |
+
return
|
1199 |
+
|
1200 |
+
from .quantization import quantize
|
1201 |
+
|
1202 |
+
if self.quantized:
|
1203 |
+
logger.info("Already quantized.")
|
1204 |
+
return self
|
1205 |
+
|
1206 |
+
self.quantized = True
|
1207 |
+
|
1208 |
+
self.config.quantization_bit = bits
|
1209 |
+
|
1210 |
+
self.transformer.encoder = quantize(self.transformer.encoder, bits, empty_init=empty_init, device=device,
|
1211 |
+
**kwargs)
|
1212 |
+
return self
|
1213 |
+
|
1214 |
+
|
1215 |
+
class ChatGLMForSequenceClassification(ChatGLMPreTrainedModel):
|
1216 |
+
def __init__(self, config: ChatGLMConfig, empty_init=True, device=None):
|
1217 |
+
super().__init__(config)
|
1218 |
+
|
1219 |
+
self.num_labels = config.num_labels
|
1220 |
+
self.transformer = ChatGLMModel(config, empty_init=empty_init, device=device)
|
1221 |
+
|
1222 |
+
self.classifier_head = nn.Linear(config.hidden_size, config.num_labels, bias=True, dtype=torch.half)
|
1223 |
+
if config.classifier_dropout is not None:
|
1224 |
+
self.dropout = nn.Dropout(config.classifier_dropout)
|
1225 |
+
else:
|
1226 |
+
self.dropout = None
|
1227 |
+
self.config = config
|
1228 |
+
|
1229 |
+
if self.config.quantization_bit:
|
1230 |
+
self.quantize(self.config.quantization_bit, empty_init=True)
|
1231 |
+
|
1232 |
+
def forward(
|
1233 |
+
self,
|
1234 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1235 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1236 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1237 |
+
full_attention_mask: Optional[torch.Tensor] = None,
|
1238 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
|
1239 |
+
inputs_embeds: Optional[torch.LongTensor] = None,
|
1240 |
+
labels: Optional[torch.LongTensor] = None,
|
1241 |
+
use_cache: Optional[bool] = None,
|
1242 |
+
output_hidden_states: Optional[bool] = None,
|
1243 |
+
return_dict: Optional[bool] = None,
|
1244 |
+
) -> Union[Tuple[torch.Tensor, ...], SequenceClassifierOutputWithPast]:
|
1245 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1246 |
+
|
1247 |
+
transformer_outputs = self.transformer(
|
1248 |
+
input_ids=input_ids,
|
1249 |
+
position_ids=position_ids,
|
1250 |
+
attention_mask=attention_mask,
|
1251 |
+
full_attention_mask=full_attention_mask,
|
1252 |
+
past_key_values=past_key_values,
|
1253 |
+
inputs_embeds=inputs_embeds,
|
1254 |
+
use_cache=use_cache,
|
1255 |
+
output_hidden_states=output_hidden_states,
|
1256 |
+
return_dict=return_dict,
|
1257 |
+
)
|
1258 |
+
|
1259 |
+
hidden_states = transformer_outputs[0]
|
1260 |
+
pooled_hidden_states = hidden_states[-1]
|
1261 |
+
if self.dropout is not None:
|
1262 |
+
pooled_hidden_states = self.dropout(pooled_hidden_states)
|
1263 |
+
logits = self.classifier_head(pooled_hidden_states)
|
1264 |
+
|
1265 |
+
loss = None
|
1266 |
+
if labels is not None:
|
1267 |
+
if self.config.problem_type is None:
|
1268 |
+
if self.num_labels == 1:
|
1269 |
+
self.config.problem_type = "regression"
|
1270 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
1271 |
+
self.config.problem_type = "single_label_classification"
|
1272 |
+
else:
|
1273 |
+
self.config.problem_type = "multi_label_classification"
|
1274 |
+
|
1275 |
+
if self.config.problem_type == "regression":
|
1276 |
+
loss_fct = MSELoss()
|
1277 |
+
if self.num_labels == 1:
|
1278 |
+
loss = loss_fct(logits.squeeze().float(), labels.squeeze())
|
1279 |
+
else:
|
1280 |
+
loss = loss_fct(logits.float(), labels)
|
1281 |
+
elif self.config.problem_type == "single_label_classification":
|
1282 |
+
loss_fct = CrossEntropyLoss()
|
1283 |
+
loss = loss_fct(logits.view(-1, self.num_labels).float(), labels.view(-1))
|
1284 |
+
elif self.config.problem_type == "multi_label_classification":
|
1285 |
+
loss_fct = BCEWithLogitsLoss()
|
1286 |
+
loss = loss_fct(logits.float(), labels.view(-1, self.num_labels))
|
1287 |
+
|
1288 |
+
if not return_dict:
|
1289 |
+
output = (logits,) + transformer_outputs[1:]
|
1290 |
+
return ((loss,) + output) if loss is not None else output
|
1291 |
+
|
1292 |
+
return SequenceClassifierOutputWithPast(
|
1293 |
+
loss=loss,
|
1294 |
+
logits=logits,
|
1295 |
+
past_key_values=transformer_outputs.past_key_values,
|
1296 |
+
hidden_states=transformer_outputs.hidden_states,
|
1297 |
+
attentions=transformer_outputs.attentions,
|
1298 |
+
)
|
models/tokenization_chatglm.py
ADDED
@@ -0,0 +1,300 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import json
|
2 |
+
import os
|
3 |
+
import re
|
4 |
+
from typing import List, Optional, Union, Dict
|
5 |
+
from sentencepiece import SentencePieceProcessor
|
6 |
+
from transformers import PreTrainedTokenizer
|
7 |
+
from transformers.utils import logging, PaddingStrategy
|
8 |
+
from transformers.tokenization_utils_base import EncodedInput, BatchEncoding
|
9 |
+
|
10 |
+
|
11 |
+
class SPTokenizer:
|
12 |
+
def __init__(self, model_path: str):
|
13 |
+
# reload tokenizer
|
14 |
+
assert os.path.isfile(model_path), model_path
|
15 |
+
self.sp_model = SentencePieceProcessor(model_file=model_path)
|
16 |
+
|
17 |
+
# BOS / EOS token IDs
|
18 |
+
self.n_words: int = self.sp_model.vocab_size()
|
19 |
+
self.bos_id: int = self.sp_model.bos_id()
|
20 |
+
self.eos_id: int = self.sp_model.eos_id()
|
21 |
+
self.pad_id: int = self.sp_model.unk_id()
|
22 |
+
assert self.sp_model.vocab_size() == self.sp_model.get_piece_size()
|
23 |
+
|
24 |
+
role_special_tokens = ["<|system|>", "<|user|>", "<|assistant|>", "<|observation|>"]
|
25 |
+
special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", "sop", "eop"] + role_special_tokens
|
26 |
+
self.special_tokens = {}
|
27 |
+
self.index_special_tokens = {}
|
28 |
+
for token in special_tokens:
|
29 |
+
self.special_tokens[token] = self.n_words
|
30 |
+
self.index_special_tokens[self.n_words] = token
|
31 |
+
self.n_words += 1
|
32 |
+
self.role_special_token_expression = "|".join([re.escape(token) for token in role_special_tokens])
|
33 |
+
|
34 |
+
def tokenize(self, s: str, encode_special_tokens=False):
|
35 |
+
if encode_special_tokens:
|
36 |
+
last_index = 0
|
37 |
+
t = []
|
38 |
+
for match in re.finditer(self.role_special_token_expression, s):
|
39 |
+
if last_index < match.start():
|
40 |
+
t.extend(self.sp_model.EncodeAsPieces(s[last_index:match.start()]))
|
41 |
+
t.append(s[match.start():match.end()])
|
42 |
+
last_index = match.end()
|
43 |
+
if last_index < len(s):
|
44 |
+
t.extend(self.sp_model.EncodeAsPieces(s[last_index:]))
|
45 |
+
return t
|
46 |
+
else:
|
47 |
+
return self.sp_model.EncodeAsPieces(s)
|
48 |
+
|
49 |
+
def encode(self, s: str, bos: bool = False, eos: bool = False) -> List[int]:
|
50 |
+
assert type(s) is str
|
51 |
+
t = self.sp_model.encode(s)
|
52 |
+
if bos:
|
53 |
+
t = [self.bos_id] + t
|
54 |
+
if eos:
|
55 |
+
t = t + [self.eos_id]
|
56 |
+
return t
|
57 |
+
|
58 |
+
def decode(self, t: List[int]) -> str:
|
59 |
+
text, buffer = "", []
|
60 |
+
for token in t:
|
61 |
+
if token in self.index_special_tokens:
|
62 |
+
if buffer:
|
63 |
+
text += self.sp_model.decode(buffer)
|
64 |
+
buffer = []
|
65 |
+
text += self.index_special_tokens[token]
|
66 |
+
else:
|
67 |
+
buffer.append(token)
|
68 |
+
if buffer:
|
69 |
+
text += self.sp_model.decode(buffer)
|
70 |
+
return text
|
71 |
+
|
72 |
+
def decode_tokens(self, tokens: List[str]) -> str:
|
73 |
+
text = self.sp_model.DecodePieces(tokens)
|
74 |
+
return text
|
75 |
+
|
76 |
+
def convert_token_to_id(self, token):
|
77 |
+
""" Converts a token (str) in an id using the vocab. """
|
78 |
+
if token in self.special_tokens:
|
79 |
+
return self.special_tokens[token]
|
80 |
+
return self.sp_model.PieceToId(token)
|
81 |
+
|
82 |
+
def convert_id_to_token(self, index):
|
83 |
+
"""Converts an index (integer) in a token (str) using the vocab."""
|
84 |
+
if index in self.index_special_tokens:
|
85 |
+
return self.index_special_tokens[index]
|
86 |
+
if index in [self.eos_id, self.bos_id, self.pad_id] or index < 0:
|
87 |
+
return ""
|
88 |
+
return self.sp_model.IdToPiece(index)
|
89 |
+
|
90 |
+
|
91 |
+
class ChatGLMTokenizer(PreTrainedTokenizer):
|
92 |
+
vocab_files_names = {"vocab_file": "tokenizer.model"}
|
93 |
+
|
94 |
+
model_input_names = ["input_ids", "attention_mask", "position_ids"]
|
95 |
+
|
96 |
+
def __init__(self, vocab_file, padding_side="left", clean_up_tokenization_spaces=False, encode_special_tokens=False,
|
97 |
+
**kwargs):
|
98 |
+
self.name = "GLMTokenizer"
|
99 |
+
|
100 |
+
self.vocab_file = vocab_file
|
101 |
+
self.tokenizer = SPTokenizer(vocab_file)
|
102 |
+
self.special_tokens = {
|
103 |
+
"<bos>": self.tokenizer.bos_id,
|
104 |
+
"<eos>": self.tokenizer.eos_id,
|
105 |
+
"<pad>": self.tokenizer.pad_id
|
106 |
+
}
|
107 |
+
self.encode_special_tokens = encode_special_tokens
|
108 |
+
super().__init__(padding_side=padding_side, clean_up_tokenization_spaces=clean_up_tokenization_spaces,
|
109 |
+
encode_special_tokens=encode_special_tokens,
|
110 |
+
**kwargs)
|
111 |
+
|
112 |
+
def get_command(self, token):
|
113 |
+
if token in self.special_tokens:
|
114 |
+
return self.special_tokens[token]
|
115 |
+
assert token in self.tokenizer.special_tokens, f"{token} is not a special token for {self.name}"
|
116 |
+
return self.tokenizer.special_tokens[token]
|
117 |
+
|
118 |
+
@property
|
119 |
+
def unk_token(self) -> str:
|
120 |
+
return "<unk>"
|
121 |
+
|
122 |
+
@property
|
123 |
+
def pad_token(self) -> str:
|
124 |
+
return "<unk>"
|
125 |
+
|
126 |
+
@property
|
127 |
+
def pad_token_id(self):
|
128 |
+
return self.get_command("<pad>")
|
129 |
+
|
130 |
+
@property
|
131 |
+
def eos_token(self) -> str:
|
132 |
+
return "</s>"
|
133 |
+
|
134 |
+
@property
|
135 |
+
def eos_token_id(self):
|
136 |
+
return self.get_command("<eos>")
|
137 |
+
|
138 |
+
@property
|
139 |
+
def vocab_size(self):
|
140 |
+
return self.tokenizer.n_words
|
141 |
+
|
142 |
+
def get_vocab(self):
|
143 |
+
""" Returns vocab as a dict """
|
144 |
+
vocab = {self._convert_id_to_token(i): i for i in range(self.vocab_size)}
|
145 |
+
vocab.update(self.added_tokens_encoder)
|
146 |
+
return vocab
|
147 |
+
|
148 |
+
def _tokenize(self, text, **kwargs):
|
149 |
+
return self.tokenizer.tokenize(text, encode_special_tokens=self.encode_special_tokens)
|
150 |
+
|
151 |
+
def _convert_token_to_id(self, token):
|
152 |
+
""" Converts a token (str) in an id using the vocab. """
|
153 |
+
return self.tokenizer.convert_token_to_id(token)
|
154 |
+
|
155 |
+
def _convert_id_to_token(self, index):
|
156 |
+
"""Converts an index (integer) in a token (str) using the vocab."""
|
157 |
+
return self.tokenizer.convert_id_to_token(index)
|
158 |
+
|
159 |
+
def convert_tokens_to_string(self, tokens: List[str]) -> str:
|
160 |
+
return self.tokenizer.decode_tokens(tokens)
|
161 |
+
|
162 |
+
def save_vocabulary(self, save_directory, filename_prefix=None):
|
163 |
+
"""
|
164 |
+
Save the vocabulary and special tokens file to a directory.
|
165 |
+
|
166 |
+
Args:
|
167 |
+
save_directory (`str`):
|
168 |
+
The directory in which to save the vocabulary.
|
169 |
+
filename_prefix (`str`, *optional*):
|
170 |
+
An optional prefix to add to the named of the saved files.
|
171 |
+
|
172 |
+
Returns:
|
173 |
+
`Tuple(str)`: Paths to the files saved.
|
174 |
+
"""
|
175 |
+
if os.path.isdir(save_directory):
|
176 |
+
vocab_file = os.path.join(
|
177 |
+
save_directory, self.vocab_files_names["vocab_file"]
|
178 |
+
)
|
179 |
+
else:
|
180 |
+
vocab_file = save_directory
|
181 |
+
|
182 |
+
with open(self.vocab_file, 'rb') as fin:
|
183 |
+
proto_str = fin.read()
|
184 |
+
|
185 |
+
with open(vocab_file, "wb") as writer:
|
186 |
+
writer.write(proto_str)
|
187 |
+
|
188 |
+
return (vocab_file,)
|
189 |
+
|
190 |
+
def get_prefix_tokens(self):
|
191 |
+
prefix_tokens = [self.get_command("[gMASK]"), self.get_command("sop")]
|
192 |
+
return prefix_tokens
|
193 |
+
|
194 |
+
def build_single_message(self, role, metadata, message):
|
195 |
+
assert role in ["system", "user", "assistant", "observation"], role
|
196 |
+
role_tokens = [self.get_command(f"<|{role}|>")] + self.tokenizer.encode(f"{metadata}\n")
|
197 |
+
message_tokens = self.tokenizer.encode(message)
|
198 |
+
tokens = role_tokens + message_tokens
|
199 |
+
return tokens
|
200 |
+
|
201 |
+
def build_chat_input(self, query, history=None, role="user"):
|
202 |
+
if history is None:
|
203 |
+
history = []
|
204 |
+
input_ids = []
|
205 |
+
for item in history:
|
206 |
+
content = item["content"]
|
207 |
+
if item["role"] == "system" and "tools" in item:
|
208 |
+
content = content + "\n" + json.dumps(item["tools"], indent=4, ensure_ascii=False)
|
209 |
+
input_ids.extend(self.build_single_message(item["role"], item.get("metadata", ""), content))
|
210 |
+
input_ids.extend(self.build_single_message(role, "", query))
|
211 |
+
input_ids.extend([self.get_command("<|assistant|>")])
|
212 |
+
return self.batch_encode_plus([input_ids], return_tensors="pt", is_split_into_words=True)
|
213 |
+
|
214 |
+
def build_inputs_with_special_tokens(
|
215 |
+
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
|
216 |
+
) -> List[int]:
|
217 |
+
"""
|
218 |
+
Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
|
219 |
+
adding special tokens. A BERT sequence has the following format:
|
220 |
+
|
221 |
+
- single sequence: `[CLS] X [SEP]`
|
222 |
+
- pair of sequences: `[CLS] A [SEP] B [SEP]`
|
223 |
+
|
224 |
+
Args:
|
225 |
+
token_ids_0 (`List[int]`):
|
226 |
+
List of IDs to which the special tokens will be added.
|
227 |
+
token_ids_1 (`List[int]`, *optional*):
|
228 |
+
Optional second list of IDs for sequence pairs.
|
229 |
+
|
230 |
+
Returns:
|
231 |
+
`List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
|
232 |
+
"""
|
233 |
+
prefix_tokens = self.get_prefix_tokens()
|
234 |
+
token_ids_0 = prefix_tokens + token_ids_0
|
235 |
+
if token_ids_1 is not None:
|
236 |
+
token_ids_0 = token_ids_0 + token_ids_1 + [self.get_command("<eos>")]
|
237 |
+
return token_ids_0
|
238 |
+
|
239 |
+
def _pad(
|
240 |
+
self,
|
241 |
+
encoded_inputs: Union[Dict[str, EncodedInput], BatchEncoding],
|
242 |
+
max_length: Optional[int] = None,
|
243 |
+
padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
|
244 |
+
pad_to_multiple_of: Optional[int] = None,
|
245 |
+
return_attention_mask: Optional[bool] = None,
|
246 |
+
) -> dict:
|
247 |
+
"""
|
248 |
+
Pad encoded inputs (on left/right and up to predefined length or max length in the batch)
|
249 |
+
|
250 |
+
Args:
|
251 |
+
encoded_inputs:
|
252 |
+
Dictionary of tokenized inputs (`List[int]`) or batch of tokenized inputs (`List[List[int]]`).
|
253 |
+
max_length: maximum length of the returned list and optionally padding length (see below).
|
254 |
+
Will truncate by taking into account the special tokens.
|
255 |
+
padding_strategy: PaddingStrategy to use for padding.
|
256 |
+
|
257 |
+
- PaddingStrategy.LONGEST Pad to the longest sequence in the batch
|
258 |
+
- PaddingStrategy.MAX_LENGTH: Pad to the max length (default)
|
259 |
+
- PaddingStrategy.DO_NOT_PAD: Do not pad
|
260 |
+
The tokenizer padding sides are defined in self.padding_side:
|
261 |
+
|
262 |
+
- 'left': pads on the left of the sequences
|
263 |
+
- 'right': pads on the right of the sequences
|
264 |
+
pad_to_multiple_of: (optional) Integer if set will pad the sequence to a multiple of the provided value.
|
265 |
+
This is especially useful to enable the use of Tensor Core on NVIDIA hardware with compute capability
|
266 |
+
`>= 7.5` (Volta).
|
267 |
+
return_attention_mask:
|
268 |
+
(optional) Set to False to avoid returning attention mask (default: set to model specifics)
|
269 |
+
"""
|
270 |
+
# Load from model defaults
|
271 |
+
assert self.padding_side == "left"
|
272 |
+
|
273 |
+
required_input = encoded_inputs[self.model_input_names[0]]
|
274 |
+
seq_length = len(required_input)
|
275 |
+
|
276 |
+
if padding_strategy == PaddingStrategy.LONGEST:
|
277 |
+
max_length = len(required_input)
|
278 |
+
|
279 |
+
if max_length is not None and pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0):
|
280 |
+
max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of
|
281 |
+
|
282 |
+
needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len(required_input) != max_length
|
283 |
+
|
284 |
+
# Initialize attention mask if not present.
|
285 |
+
if "attention_mask" not in encoded_inputs:
|
286 |
+
encoded_inputs["attention_mask"] = [1] * seq_length
|
287 |
+
|
288 |
+
if "position_ids" not in encoded_inputs:
|
289 |
+
encoded_inputs["position_ids"] = list(range(seq_length))
|
290 |
+
|
291 |
+
if needs_to_be_padded:
|
292 |
+
difference = max_length - len(required_input)
|
293 |
+
|
294 |
+
if "attention_mask" in encoded_inputs:
|
295 |
+
encoded_inputs["attention_mask"] = [0] * difference + encoded_inputs["attention_mask"]
|
296 |
+
if "position_ids" in encoded_inputs:
|
297 |
+
encoded_inputs["position_ids"] = [0] * difference + encoded_inputs["position_ids"]
|
298 |
+
encoded_inputs[self.model_input_names[0]] = [self.pad_token_id] * difference + required_input
|
299 |
+
|
300 |
+
return encoded_inputs
|
models/unet_2d_condition.py
ADDED
@@ -0,0 +1,1318 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2024 The HuggingFace Team. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
from dataclasses import dataclass
|
15 |
+
from typing import Any, Dict, List, Optional, Tuple, Union
|
16 |
+
|
17 |
+
import torch
|
18 |
+
import torch.nn as nn
|
19 |
+
import torch.utils.checkpoint
|
20 |
+
|
21 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
22 |
+
from diffusers.loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin
|
23 |
+
from diffusers.loaders.single_file_model import FromOriginalModelMixin
|
24 |
+
from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers
|
25 |
+
from diffusers.models.activations import get_activation
|
26 |
+
from diffusers.models.attention_processor import (
|
27 |
+
ADDED_KV_ATTENTION_PROCESSORS,
|
28 |
+
CROSS_ATTENTION_PROCESSORS,
|
29 |
+
Attention,
|
30 |
+
AttentionProcessor,
|
31 |
+
AttnAddedKVProcessor,
|
32 |
+
AttnProcessor,
|
33 |
+
)
|
34 |
+
from diffusers.models.embeddings import (
|
35 |
+
GaussianFourierProjection,
|
36 |
+
GLIGENTextBoundingboxProjection,
|
37 |
+
ImageHintTimeEmbedding,
|
38 |
+
ImageProjection,
|
39 |
+
ImageTimeEmbedding,
|
40 |
+
TextImageProjection,
|
41 |
+
TextImageTimeEmbedding,
|
42 |
+
TextTimeEmbedding,
|
43 |
+
TimestepEmbedding,
|
44 |
+
Timesteps,
|
45 |
+
)
|
46 |
+
from diffusers.models.modeling_utils import ModelMixin
|
47 |
+
|
48 |
+
try:
|
49 |
+
from diffusers.models.unet_2d_blocks import (
|
50 |
+
get_down_block,
|
51 |
+
get_mid_block,
|
52 |
+
get_up_block,
|
53 |
+
)
|
54 |
+
except:
|
55 |
+
from diffusers.models.unets.unet_2d_blocks import (
|
56 |
+
get_down_block,
|
57 |
+
get_mid_block,
|
58 |
+
get_up_block,
|
59 |
+
)
|
60 |
+
|
61 |
+
|
62 |
+
|
63 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
64 |
+
|
65 |
+
|
66 |
+
@dataclass
|
67 |
+
class UNet2DConditionOutput(BaseOutput):
|
68 |
+
"""
|
69 |
+
The output of [`UNet2DConditionModel`].
|
70 |
+
|
71 |
+
Args:
|
72 |
+
sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)`):
|
73 |
+
The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
|
74 |
+
"""
|
75 |
+
|
76 |
+
sample: torch.Tensor = None
|
77 |
+
|
78 |
+
|
79 |
+
class UNet2DConditionModel(
|
80 |
+
ModelMixin, ConfigMixin, FromOriginalModelMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin
|
81 |
+
):
|
82 |
+
r"""
|
83 |
+
A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
|
84 |
+
shaped output.
|
85 |
+
|
86 |
+
This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
|
87 |
+
for all models (such as downloading or saving).
|
88 |
+
|
89 |
+
Parameters:
|
90 |
+
sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
|
91 |
+
Height and width of input/output sample.
|
92 |
+
in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
|
93 |
+
out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
|
94 |
+
center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
|
95 |
+
flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
|
96 |
+
Whether to flip the sin to cos in the time embedding.
|
97 |
+
freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
|
98 |
+
down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
|
99 |
+
The tuple of downsample blocks to use.
|
100 |
+
mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
|
101 |
+
Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
|
102 |
+
`UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
|
103 |
+
up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
|
104 |
+
The tuple of upsample blocks to use.
|
105 |
+
only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
|
106 |
+
Whether to include self-attention in the basic transformer blocks, see
|
107 |
+
[`~models.attention.BasicTransformerBlock`].
|
108 |
+
block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
|
109 |
+
The tuple of output channels for each block.
|
110 |
+
layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
|
111 |
+
downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
|
112 |
+
mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
|
113 |
+
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
114 |
+
act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
|
115 |
+
norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
|
116 |
+
If `None`, normalization and activation layers is skipped in post-processing.
|
117 |
+
norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
|
118 |
+
cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
|
119 |
+
The dimension of the cross attention features.
|
120 |
+
transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
|
121 |
+
The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
|
122 |
+
[`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`],
|
123 |
+
[`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
|
124 |
+
reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
|
125 |
+
The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
|
126 |
+
blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
|
127 |
+
[`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`],
|
128 |
+
[`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
|
129 |
+
encoder_hid_dim (`int`, *optional*, defaults to None):
|
130 |
+
If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
|
131 |
+
dimension to `cross_attention_dim`.
|
132 |
+
encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
|
133 |
+
If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
|
134 |
+
embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
|
135 |
+
attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
|
136 |
+
num_attention_heads (`int`, *optional*):
|
137 |
+
The number of attention heads. If not defined, defaults to `attention_head_dim`
|
138 |
+
resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
|
139 |
+
for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
|
140 |
+
class_embed_type (`str`, *optional*, defaults to `None`):
|
141 |
+
The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
|
142 |
+
`"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
|
143 |
+
addition_embed_type (`str`, *optional*, defaults to `None`):
|
144 |
+
Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
|
145 |
+
"text". "text" will use the `TextTimeEmbedding` layer.
|
146 |
+
addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
|
147 |
+
Dimension for the timestep embeddings.
|
148 |
+
num_class_embeds (`int`, *optional*, defaults to `None`):
|
149 |
+
Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
|
150 |
+
class conditioning with `class_embed_type` equal to `None`.
|
151 |
+
time_embedding_type (`str`, *optional*, defaults to `positional`):
|
152 |
+
The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
|
153 |
+
time_embedding_dim (`int`, *optional*, defaults to `None`):
|
154 |
+
An optional override for the dimension of the projected time embedding.
|
155 |
+
time_embedding_act_fn (`str`, *optional*, defaults to `None`):
|
156 |
+
Optional activation function to use only once on the time embeddings before they are passed to the rest of
|
157 |
+
the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
|
158 |
+
timestep_post_act (`str`, *optional*, defaults to `None`):
|
159 |
+
The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
|
160 |
+
time_cond_proj_dim (`int`, *optional*, defaults to `None`):
|
161 |
+
The dimension of `cond_proj` layer in the timestep embedding.
|
162 |
+
conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
|
163 |
+
conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
|
164 |
+
projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
|
165 |
+
`class_embed_type="projection"`. Required when `class_embed_type="projection"`.
|
166 |
+
class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
|
167 |
+
embeddings with the class embeddings.
|
168 |
+
mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
|
169 |
+
Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
|
170 |
+
`only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
|
171 |
+
`only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
|
172 |
+
otherwise.
|
173 |
+
"""
|
174 |
+
|
175 |
+
_supports_gradient_checkpointing = True
|
176 |
+
_no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D", "CrossAttnUpBlock2D"]
|
177 |
+
|
178 |
+
@register_to_config
|
179 |
+
def __init__(
|
180 |
+
self,
|
181 |
+
sample_size: Optional[int] = None,
|
182 |
+
in_channels: int = 4,
|
183 |
+
out_channels: int = 4,
|
184 |
+
center_input_sample: bool = False,
|
185 |
+
flip_sin_to_cos: bool = True,
|
186 |
+
freq_shift: int = 0,
|
187 |
+
down_block_types: Tuple[str] = (
|
188 |
+
"CrossAttnDownBlock2D",
|
189 |
+
"CrossAttnDownBlock2D",
|
190 |
+
"CrossAttnDownBlock2D",
|
191 |
+
"DownBlock2D",
|
192 |
+
),
|
193 |
+
mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
|
194 |
+
up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
|
195 |
+
only_cross_attention: Union[bool, Tuple[bool]] = False,
|
196 |
+
block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
|
197 |
+
layers_per_block: Union[int, Tuple[int]] = 2,
|
198 |
+
downsample_padding: int = 1,
|
199 |
+
mid_block_scale_factor: float = 1,
|
200 |
+
dropout: float = 0.0,
|
201 |
+
act_fn: str = "silu",
|
202 |
+
norm_num_groups: Optional[int] = 32,
|
203 |
+
norm_eps: float = 1e-5,
|
204 |
+
cross_attention_dim: Union[int, Tuple[int]] = 1280,
|
205 |
+
transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
|
206 |
+
reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
|
207 |
+
encoder_hid_dim: Optional[int] = None,
|
208 |
+
encoder_hid_dim_type: Optional[str] = None,
|
209 |
+
attention_head_dim: Union[int, Tuple[int]] = 8,
|
210 |
+
num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
|
211 |
+
dual_cross_attention: bool = False,
|
212 |
+
use_linear_projection: bool = False,
|
213 |
+
class_embed_type: Optional[str] = None,
|
214 |
+
addition_embed_type: Optional[str] = None,
|
215 |
+
addition_time_embed_dim: Optional[int] = None,
|
216 |
+
num_class_embeds: Optional[int] = None,
|
217 |
+
upcast_attention: bool = False,
|
218 |
+
resnet_time_scale_shift: str = "default",
|
219 |
+
resnet_skip_time_act: bool = False,
|
220 |
+
resnet_out_scale_factor: float = 1.0,
|
221 |
+
time_embedding_type: str = "positional",
|
222 |
+
time_embedding_dim: Optional[int] = None,
|
223 |
+
time_embedding_act_fn: Optional[str] = None,
|
224 |
+
timestep_post_act: Optional[str] = None,
|
225 |
+
time_cond_proj_dim: Optional[int] = None,
|
226 |
+
conv_in_kernel: int = 3,
|
227 |
+
conv_out_kernel: int = 3,
|
228 |
+
projection_class_embeddings_input_dim: Optional[int] = None,
|
229 |
+
attention_type: str = "default",
|
230 |
+
class_embeddings_concat: bool = False,
|
231 |
+
mid_block_only_cross_attention: Optional[bool] = None,
|
232 |
+
cross_attention_norm: Optional[str] = None,
|
233 |
+
addition_embed_type_num_heads: int = 64,
|
234 |
+
):
|
235 |
+
super().__init__()
|
236 |
+
|
237 |
+
self.sample_size = sample_size
|
238 |
+
|
239 |
+
if num_attention_heads is not None:
|
240 |
+
raise ValueError(
|
241 |
+
"At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
|
242 |
+
)
|
243 |
+
|
244 |
+
# If `num_attention_heads` is not defined (which is the case for most models)
|
245 |
+
# it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
|
246 |
+
# The reason for this behavior is to correct for incorrectly named variables that were introduced
|
247 |
+
# when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
|
248 |
+
# Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
|
249 |
+
# which is why we correct for the naming here.
|
250 |
+
num_attention_heads = num_attention_heads or attention_head_dim
|
251 |
+
|
252 |
+
# Check inputs
|
253 |
+
self._check_config(
|
254 |
+
down_block_types=down_block_types,
|
255 |
+
up_block_types=up_block_types,
|
256 |
+
only_cross_attention=only_cross_attention,
|
257 |
+
block_out_channels=block_out_channels,
|
258 |
+
layers_per_block=layers_per_block,
|
259 |
+
cross_attention_dim=cross_attention_dim,
|
260 |
+
transformer_layers_per_block=transformer_layers_per_block,
|
261 |
+
reverse_transformer_layers_per_block=reverse_transformer_layers_per_block,
|
262 |
+
attention_head_dim=attention_head_dim,
|
263 |
+
num_attention_heads=num_attention_heads,
|
264 |
+
)
|
265 |
+
|
266 |
+
# input
|
267 |
+
conv_in_padding = (conv_in_kernel - 1) // 2
|
268 |
+
self.conv_in = nn.Conv2d(
|
269 |
+
in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
|
270 |
+
)
|
271 |
+
|
272 |
+
# time
|
273 |
+
time_embed_dim, timestep_input_dim = self._set_time_proj(
|
274 |
+
time_embedding_type,
|
275 |
+
block_out_channels=block_out_channels,
|
276 |
+
flip_sin_to_cos=flip_sin_to_cos,
|
277 |
+
freq_shift=freq_shift,
|
278 |
+
time_embedding_dim=time_embedding_dim,
|
279 |
+
)
|
280 |
+
|
281 |
+
self.time_embedding = TimestepEmbedding(
|
282 |
+
timestep_input_dim,
|
283 |
+
time_embed_dim,
|
284 |
+
act_fn=act_fn,
|
285 |
+
post_act_fn=timestep_post_act,
|
286 |
+
cond_proj_dim=time_cond_proj_dim,
|
287 |
+
)
|
288 |
+
|
289 |
+
self._set_encoder_hid_proj(
|
290 |
+
encoder_hid_dim_type,
|
291 |
+
cross_attention_dim=cross_attention_dim,
|
292 |
+
encoder_hid_dim=encoder_hid_dim,
|
293 |
+
)
|
294 |
+
|
295 |
+
# class embedding
|
296 |
+
self._set_class_embedding(
|
297 |
+
class_embed_type,
|
298 |
+
act_fn=act_fn,
|
299 |
+
num_class_embeds=num_class_embeds,
|
300 |
+
projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
|
301 |
+
time_embed_dim=time_embed_dim,
|
302 |
+
timestep_input_dim=timestep_input_dim,
|
303 |
+
)
|
304 |
+
|
305 |
+
self._set_add_embedding(
|
306 |
+
addition_embed_type,
|
307 |
+
addition_embed_type_num_heads=addition_embed_type_num_heads,
|
308 |
+
addition_time_embed_dim=addition_time_embed_dim,
|
309 |
+
cross_attention_dim=cross_attention_dim,
|
310 |
+
encoder_hid_dim=encoder_hid_dim,
|
311 |
+
flip_sin_to_cos=flip_sin_to_cos,
|
312 |
+
freq_shift=freq_shift,
|
313 |
+
projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
|
314 |
+
time_embed_dim=time_embed_dim,
|
315 |
+
)
|
316 |
+
|
317 |
+
if time_embedding_act_fn is None:
|
318 |
+
self.time_embed_act = None
|
319 |
+
else:
|
320 |
+
self.time_embed_act = get_activation(time_embedding_act_fn)
|
321 |
+
|
322 |
+
self.down_blocks = nn.ModuleList([])
|
323 |
+
self.up_blocks = nn.ModuleList([])
|
324 |
+
|
325 |
+
if isinstance(only_cross_attention, bool):
|
326 |
+
if mid_block_only_cross_attention is None:
|
327 |
+
mid_block_only_cross_attention = only_cross_attention
|
328 |
+
|
329 |
+
only_cross_attention = [only_cross_attention] * len(down_block_types)
|
330 |
+
|
331 |
+
if mid_block_only_cross_attention is None:
|
332 |
+
mid_block_only_cross_attention = False
|
333 |
+
|
334 |
+
if isinstance(num_attention_heads, int):
|
335 |
+
num_attention_heads = (num_attention_heads,) * len(down_block_types)
|
336 |
+
|
337 |
+
if isinstance(attention_head_dim, int):
|
338 |
+
attention_head_dim = (attention_head_dim,) * len(down_block_types)
|
339 |
+
|
340 |
+
if isinstance(cross_attention_dim, int):
|
341 |
+
cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
|
342 |
+
|
343 |
+
if isinstance(layers_per_block, int):
|
344 |
+
layers_per_block = [layers_per_block] * len(down_block_types)
|
345 |
+
|
346 |
+
if isinstance(transformer_layers_per_block, int):
|
347 |
+
transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
|
348 |
+
|
349 |
+
if class_embeddings_concat:
|
350 |
+
# The time embeddings are concatenated with the class embeddings. The dimension of the
|
351 |
+
# time embeddings passed to the down, middle, and up blocks is twice the dimension of the
|
352 |
+
# regular time embeddings
|
353 |
+
blocks_time_embed_dim = time_embed_dim * 2
|
354 |
+
else:
|
355 |
+
blocks_time_embed_dim = time_embed_dim
|
356 |
+
|
357 |
+
# down
|
358 |
+
output_channel = block_out_channels[0]
|
359 |
+
for i, down_block_type in enumerate(down_block_types):
|
360 |
+
input_channel = output_channel
|
361 |
+
output_channel = block_out_channels[i]
|
362 |
+
is_final_block = i == len(block_out_channels) - 1
|
363 |
+
|
364 |
+
down_block = get_down_block(
|
365 |
+
down_block_type,
|
366 |
+
num_layers=layers_per_block[i],
|
367 |
+
transformer_layers_per_block=transformer_layers_per_block[i],
|
368 |
+
in_channels=input_channel,
|
369 |
+
out_channels=output_channel,
|
370 |
+
temb_channels=blocks_time_embed_dim,
|
371 |
+
add_downsample=not is_final_block,
|
372 |
+
resnet_eps=norm_eps,
|
373 |
+
resnet_act_fn=act_fn,
|
374 |
+
resnet_groups=norm_num_groups,
|
375 |
+
cross_attention_dim=cross_attention_dim[i],
|
376 |
+
num_attention_heads=num_attention_heads[i],
|
377 |
+
downsample_padding=downsample_padding,
|
378 |
+
dual_cross_attention=dual_cross_attention,
|
379 |
+
use_linear_projection=use_linear_projection,
|
380 |
+
only_cross_attention=only_cross_attention[i],
|
381 |
+
upcast_attention=upcast_attention,
|
382 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
383 |
+
attention_type=attention_type,
|
384 |
+
resnet_skip_time_act=resnet_skip_time_act,
|
385 |
+
resnet_out_scale_factor=resnet_out_scale_factor,
|
386 |
+
cross_attention_norm=cross_attention_norm,
|
387 |
+
attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
|
388 |
+
dropout=dropout,
|
389 |
+
)
|
390 |
+
self.down_blocks.append(down_block)
|
391 |
+
|
392 |
+
# mid
|
393 |
+
self.mid_block = get_mid_block(
|
394 |
+
mid_block_type,
|
395 |
+
temb_channels=blocks_time_embed_dim,
|
396 |
+
in_channels=block_out_channels[-1],
|
397 |
+
resnet_eps=norm_eps,
|
398 |
+
resnet_act_fn=act_fn,
|
399 |
+
resnet_groups=norm_num_groups,
|
400 |
+
output_scale_factor=mid_block_scale_factor,
|
401 |
+
transformer_layers_per_block=transformer_layers_per_block[-1],
|
402 |
+
num_attention_heads=num_attention_heads[-1],
|
403 |
+
cross_attention_dim=cross_attention_dim[-1],
|
404 |
+
dual_cross_attention=dual_cross_attention,
|
405 |
+
use_linear_projection=use_linear_projection,
|
406 |
+
mid_block_only_cross_attention=mid_block_only_cross_attention,
|
407 |
+
upcast_attention=upcast_attention,
|
408 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
409 |
+
attention_type=attention_type,
|
410 |
+
resnet_skip_time_act=resnet_skip_time_act,
|
411 |
+
cross_attention_norm=cross_attention_norm,
|
412 |
+
attention_head_dim=attention_head_dim[-1],
|
413 |
+
dropout=dropout,
|
414 |
+
)
|
415 |
+
|
416 |
+
# count how many layers upsample the images
|
417 |
+
self.num_upsamplers = 0
|
418 |
+
|
419 |
+
# up
|
420 |
+
reversed_block_out_channels = list(reversed(block_out_channels))
|
421 |
+
reversed_num_attention_heads = list(reversed(num_attention_heads))
|
422 |
+
reversed_layers_per_block = list(reversed(layers_per_block))
|
423 |
+
reversed_cross_attention_dim = list(reversed(cross_attention_dim))
|
424 |
+
reversed_transformer_layers_per_block = (
|
425 |
+
list(reversed(transformer_layers_per_block))
|
426 |
+
if reverse_transformer_layers_per_block is None
|
427 |
+
else reverse_transformer_layers_per_block
|
428 |
+
)
|
429 |
+
only_cross_attention = list(reversed(only_cross_attention))
|
430 |
+
|
431 |
+
output_channel = reversed_block_out_channels[0]
|
432 |
+
for i, up_block_type in enumerate(up_block_types):
|
433 |
+
is_final_block = i == len(block_out_channels) - 1
|
434 |
+
|
435 |
+
prev_output_channel = output_channel
|
436 |
+
output_channel = reversed_block_out_channels[i]
|
437 |
+
input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
|
438 |
+
|
439 |
+
# add upsample block for all BUT final layer
|
440 |
+
if not is_final_block:
|
441 |
+
add_upsample = True
|
442 |
+
self.num_upsamplers += 1
|
443 |
+
else:
|
444 |
+
add_upsample = False
|
445 |
+
|
446 |
+
up_block = get_up_block(
|
447 |
+
up_block_type,
|
448 |
+
num_layers=reversed_layers_per_block[i] + 1,
|
449 |
+
transformer_layers_per_block=reversed_transformer_layers_per_block[i],
|
450 |
+
in_channels=input_channel,
|
451 |
+
out_channels=output_channel,
|
452 |
+
prev_output_channel=prev_output_channel,
|
453 |
+
temb_channels=blocks_time_embed_dim,
|
454 |
+
add_upsample=add_upsample,
|
455 |
+
resnet_eps=norm_eps,
|
456 |
+
resnet_act_fn=act_fn,
|
457 |
+
resolution_idx=i,
|
458 |
+
resnet_groups=norm_num_groups,
|
459 |
+
cross_attention_dim=reversed_cross_attention_dim[i],
|
460 |
+
num_attention_heads=reversed_num_attention_heads[i],
|
461 |
+
dual_cross_attention=dual_cross_attention,
|
462 |
+
use_linear_projection=use_linear_projection,
|
463 |
+
only_cross_attention=only_cross_attention[i],
|
464 |
+
upcast_attention=upcast_attention,
|
465 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
466 |
+
attention_type=attention_type,
|
467 |
+
resnet_skip_time_act=resnet_skip_time_act,
|
468 |
+
resnet_out_scale_factor=resnet_out_scale_factor,
|
469 |
+
cross_attention_norm=cross_attention_norm,
|
470 |
+
attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
|
471 |
+
dropout=dropout,
|
472 |
+
)
|
473 |
+
self.up_blocks.append(up_block)
|
474 |
+
prev_output_channel = output_channel
|
475 |
+
|
476 |
+
# out
|
477 |
+
if norm_num_groups is not None:
|
478 |
+
self.conv_norm_out = nn.GroupNorm(
|
479 |
+
num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
|
480 |
+
)
|
481 |
+
|
482 |
+
self.conv_act = get_activation(act_fn)
|
483 |
+
|
484 |
+
else:
|
485 |
+
self.conv_norm_out = None
|
486 |
+
self.conv_act = None
|
487 |
+
|
488 |
+
conv_out_padding = (conv_out_kernel - 1) // 2
|
489 |
+
self.conv_out = nn.Conv2d(
|
490 |
+
block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
|
491 |
+
)
|
492 |
+
|
493 |
+
self._set_pos_net_if_use_gligen(attention_type=attention_type, cross_attention_dim=cross_attention_dim)
|
494 |
+
|
495 |
+
def _check_config(
|
496 |
+
self,
|
497 |
+
down_block_types: Tuple[str],
|
498 |
+
up_block_types: Tuple[str],
|
499 |
+
only_cross_attention: Union[bool, Tuple[bool]],
|
500 |
+
block_out_channels: Tuple[int],
|
501 |
+
layers_per_block: Union[int, Tuple[int]],
|
502 |
+
cross_attention_dim: Union[int, Tuple[int]],
|
503 |
+
transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple[int]]],
|
504 |
+
reverse_transformer_layers_per_block: bool,
|
505 |
+
attention_head_dim: int,
|
506 |
+
num_attention_heads: Optional[Union[int, Tuple[int]]],
|
507 |
+
):
|
508 |
+
if len(down_block_types) != len(up_block_types):
|
509 |
+
raise ValueError(
|
510 |
+
f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
|
511 |
+
)
|
512 |
+
|
513 |
+
if len(block_out_channels) != len(down_block_types):
|
514 |
+
raise ValueError(
|
515 |
+
f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
|
516 |
+
)
|
517 |
+
|
518 |
+
if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
|
519 |
+
raise ValueError(
|
520 |
+
f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
|
521 |
+
)
|
522 |
+
|
523 |
+
if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
|
524 |
+
raise ValueError(
|
525 |
+
f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
|
526 |
+
)
|
527 |
+
|
528 |
+
if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
|
529 |
+
raise ValueError(
|
530 |
+
f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
|
531 |
+
)
|
532 |
+
|
533 |
+
if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
|
534 |
+
raise ValueError(
|
535 |
+
f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
|
536 |
+
)
|
537 |
+
|
538 |
+
if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
|
539 |
+
raise ValueError(
|
540 |
+
f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
|
541 |
+
)
|
542 |
+
if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
|
543 |
+
for layer_number_per_block in transformer_layers_per_block:
|
544 |
+
if isinstance(layer_number_per_block, list):
|
545 |
+
raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
|
546 |
+
|
547 |
+
def _set_time_proj(
|
548 |
+
self,
|
549 |
+
time_embedding_type: str,
|
550 |
+
block_out_channels: int,
|
551 |
+
flip_sin_to_cos: bool,
|
552 |
+
freq_shift: float,
|
553 |
+
time_embedding_dim: int,
|
554 |
+
) -> Tuple[int, int]:
|
555 |
+
if time_embedding_type == "fourier":
|
556 |
+
time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
|
557 |
+
if time_embed_dim % 2 != 0:
|
558 |
+
raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
|
559 |
+
self.time_proj = GaussianFourierProjection(
|
560 |
+
time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
|
561 |
+
)
|
562 |
+
timestep_input_dim = time_embed_dim
|
563 |
+
elif time_embedding_type == "positional":
|
564 |
+
time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
|
565 |
+
|
566 |
+
self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
|
567 |
+
timestep_input_dim = block_out_channels[0]
|
568 |
+
else:
|
569 |
+
raise ValueError(
|
570 |
+
f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
|
571 |
+
)
|
572 |
+
|
573 |
+
return time_embed_dim, timestep_input_dim
|
574 |
+
|
575 |
+
def _set_encoder_hid_proj(
|
576 |
+
self,
|
577 |
+
encoder_hid_dim_type: Optional[str],
|
578 |
+
cross_attention_dim: Union[int, Tuple[int]],
|
579 |
+
encoder_hid_dim: Optional[int],
|
580 |
+
):
|
581 |
+
if encoder_hid_dim_type is None and encoder_hid_dim is not None:
|
582 |
+
encoder_hid_dim_type = "text_proj"
|
583 |
+
self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
|
584 |
+
logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
|
585 |
+
|
586 |
+
if encoder_hid_dim is None and encoder_hid_dim_type is not None:
|
587 |
+
raise ValueError(
|
588 |
+
f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
|
589 |
+
)
|
590 |
+
|
591 |
+
if encoder_hid_dim_type == "text_proj":
|
592 |
+
self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
|
593 |
+
elif encoder_hid_dim_type == "text_image_proj":
|
594 |
+
# image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
|
595 |
+
# they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
|
596 |
+
# case when `addition_embed_type == "text_image_proj"` (Kandinsky 2.1)`
|
597 |
+
self.encoder_hid_proj = TextImageProjection(
|
598 |
+
text_embed_dim=encoder_hid_dim,
|
599 |
+
image_embed_dim=cross_attention_dim,
|
600 |
+
cross_attention_dim=cross_attention_dim,
|
601 |
+
)
|
602 |
+
elif encoder_hid_dim_type == "image_proj":
|
603 |
+
# Kandinsky 2.2
|
604 |
+
self.encoder_hid_proj = ImageProjection(
|
605 |
+
image_embed_dim=encoder_hid_dim,
|
606 |
+
cross_attention_dim=cross_attention_dim,
|
607 |
+
)
|
608 |
+
elif encoder_hid_dim_type is not None:
|
609 |
+
raise ValueError(
|
610 |
+
f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
|
611 |
+
)
|
612 |
+
else:
|
613 |
+
self.encoder_hid_proj = None
|
614 |
+
|
615 |
+
def _set_class_embedding(
|
616 |
+
self,
|
617 |
+
class_embed_type: Optional[str],
|
618 |
+
act_fn: str,
|
619 |
+
num_class_embeds: Optional[int],
|
620 |
+
projection_class_embeddings_input_dim: Optional[int],
|
621 |
+
time_embed_dim: int,
|
622 |
+
timestep_input_dim: int,
|
623 |
+
):
|
624 |
+
if class_embed_type is None and num_class_embeds is not None:
|
625 |
+
self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
|
626 |
+
elif class_embed_type == "timestep":
|
627 |
+
self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
|
628 |
+
elif class_embed_type == "identity":
|
629 |
+
self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
|
630 |
+
elif class_embed_type == "projection":
|
631 |
+
if projection_class_embeddings_input_dim is None:
|
632 |
+
raise ValueError(
|
633 |
+
"`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
|
634 |
+
)
|
635 |
+
# The projection `class_embed_type` is the same as the timestep `class_embed_type` except
|
636 |
+
# 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
|
637 |
+
# 2. it projects from an arbitrary input dimension.
|
638 |
+
#
|
639 |
+
# Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
|
640 |
+
# When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
|
641 |
+
# As a result, `TimestepEmbedding` can be passed arbitrary vectors.
|
642 |
+
self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
|
643 |
+
elif class_embed_type == "simple_projection":
|
644 |
+
if projection_class_embeddings_input_dim is None:
|
645 |
+
raise ValueError(
|
646 |
+
"`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
|
647 |
+
)
|
648 |
+
self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
|
649 |
+
else:
|
650 |
+
self.class_embedding = None
|
651 |
+
|
652 |
+
def _set_add_embedding(
|
653 |
+
self,
|
654 |
+
addition_embed_type: str,
|
655 |
+
addition_embed_type_num_heads: int,
|
656 |
+
addition_time_embed_dim: Optional[int],
|
657 |
+
flip_sin_to_cos: bool,
|
658 |
+
freq_shift: float,
|
659 |
+
cross_attention_dim: Optional[int],
|
660 |
+
encoder_hid_dim: Optional[int],
|
661 |
+
projection_class_embeddings_input_dim: Optional[int],
|
662 |
+
time_embed_dim: int,
|
663 |
+
):
|
664 |
+
if addition_embed_type == "text":
|
665 |
+
if encoder_hid_dim is not None:
|
666 |
+
text_time_embedding_from_dim = encoder_hid_dim
|
667 |
+
else:
|
668 |
+
text_time_embedding_from_dim = cross_attention_dim
|
669 |
+
|
670 |
+
self.add_embedding = TextTimeEmbedding(
|
671 |
+
text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
|
672 |
+
)
|
673 |
+
elif addition_embed_type == "text_image":
|
674 |
+
# text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
|
675 |
+
# they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
|
676 |
+
# case when `addition_embed_type == "text_image"` (Kandinsky 2.1)`
|
677 |
+
self.add_embedding = TextImageTimeEmbedding(
|
678 |
+
text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
|
679 |
+
)
|
680 |
+
elif addition_embed_type == "text_time":
|
681 |
+
self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
|
682 |
+
self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
|
683 |
+
elif addition_embed_type == "image":
|
684 |
+
# Kandinsky 2.2
|
685 |
+
self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
|
686 |
+
elif addition_embed_type == "image_hint":
|
687 |
+
# Kandinsky 2.2 ControlNet
|
688 |
+
self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
|
689 |
+
elif addition_embed_type is not None:
|
690 |
+
raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
|
691 |
+
|
692 |
+
def _set_pos_net_if_use_gligen(self, attention_type: str, cross_attention_dim: int):
|
693 |
+
if attention_type in ["gated", "gated-text-image"]:
|
694 |
+
positive_len = 768
|
695 |
+
if isinstance(cross_attention_dim, int):
|
696 |
+
positive_len = cross_attention_dim
|
697 |
+
elif isinstance(cross_attention_dim, (list, tuple)):
|
698 |
+
positive_len = cross_attention_dim[0]
|
699 |
+
|
700 |
+
feature_type = "text-only" if attention_type == "gated" else "text-image"
|
701 |
+
self.position_net = GLIGENTextBoundingboxProjection(
|
702 |
+
positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
|
703 |
+
)
|
704 |
+
|
705 |
+
@property
|
706 |
+
def attn_processors(self) -> Dict[str, AttentionProcessor]:
|
707 |
+
r"""
|
708 |
+
Returns:
|
709 |
+
`dict` of attention processors: A dictionary containing all attention processors used in the model with
|
710 |
+
indexed by its weight name.
|
711 |
+
"""
|
712 |
+
# set recursively
|
713 |
+
processors = {}
|
714 |
+
|
715 |
+
def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
|
716 |
+
if hasattr(module, "get_processor"):
|
717 |
+
processors[f"{name}.processor"] = module.get_processor()
|
718 |
+
|
719 |
+
for sub_name, child in module.named_children():
|
720 |
+
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
|
721 |
+
|
722 |
+
return processors
|
723 |
+
|
724 |
+
for name, module in self.named_children():
|
725 |
+
fn_recursive_add_processors(name, module, processors)
|
726 |
+
|
727 |
+
return processors
|
728 |
+
|
729 |
+
def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
|
730 |
+
r"""
|
731 |
+
Sets the attention processor to use to compute attention.
|
732 |
+
|
733 |
+
Parameters:
|
734 |
+
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
|
735 |
+
The instantiated processor class or a dictionary of processor classes that will be set as the processor
|
736 |
+
for **all** `Attention` layers.
|
737 |
+
|
738 |
+
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
|
739 |
+
processor. This is strongly recommended when setting trainable attention processors.
|
740 |
+
|
741 |
+
"""
|
742 |
+
count = len(self.attn_processors.keys())
|
743 |
+
|
744 |
+
if isinstance(processor, dict) and len(processor) != count:
|
745 |
+
raise ValueError(
|
746 |
+
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
|
747 |
+
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
|
748 |
+
)
|
749 |
+
|
750 |
+
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
|
751 |
+
if hasattr(module, "set_processor"):
|
752 |
+
if not isinstance(processor, dict):
|
753 |
+
module.set_processor(processor)
|
754 |
+
else:
|
755 |
+
module.set_processor(processor.pop(f"{name}.processor"))
|
756 |
+
|
757 |
+
for sub_name, child in module.named_children():
|
758 |
+
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
|
759 |
+
|
760 |
+
for name, module in self.named_children():
|
761 |
+
fn_recursive_attn_processor(name, module, processor)
|
762 |
+
|
763 |
+
def set_default_attn_processor(self):
|
764 |
+
"""
|
765 |
+
Disables custom attention processors and sets the default attention implementation.
|
766 |
+
"""
|
767 |
+
if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
|
768 |
+
processor = AttnAddedKVProcessor()
|
769 |
+
elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
|
770 |
+
processor = AttnProcessor()
|
771 |
+
else:
|
772 |
+
raise ValueError(
|
773 |
+
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
|
774 |
+
)
|
775 |
+
|
776 |
+
self.set_attn_processor(processor)
|
777 |
+
|
778 |
+
def set_attention_slice(self, slice_size: Union[str, int, List[int]] = "auto"):
|
779 |
+
r"""
|
780 |
+
Enable sliced attention computation.
|
781 |
+
|
782 |
+
When this option is enabled, the attention module splits the input tensor in slices to compute attention in
|
783 |
+
several steps. This is useful for saving some memory in exchange for a small decrease in speed.
|
784 |
+
|
785 |
+
Args:
|
786 |
+
slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
|
787 |
+
When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
|
788 |
+
`"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
|
789 |
+
provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
|
790 |
+
must be a multiple of `slice_size`.
|
791 |
+
"""
|
792 |
+
sliceable_head_dims = []
|
793 |
+
|
794 |
+
def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
|
795 |
+
if hasattr(module, "set_attention_slice"):
|
796 |
+
sliceable_head_dims.append(module.sliceable_head_dim)
|
797 |
+
|
798 |
+
for child in module.children():
|
799 |
+
fn_recursive_retrieve_sliceable_dims(child)
|
800 |
+
|
801 |
+
# retrieve number of attention layers
|
802 |
+
for module in self.children():
|
803 |
+
fn_recursive_retrieve_sliceable_dims(module)
|
804 |
+
|
805 |
+
num_sliceable_layers = len(sliceable_head_dims)
|
806 |
+
|
807 |
+
if slice_size == "auto":
|
808 |
+
# half the attention head size is usually a good trade-off between
|
809 |
+
# speed and memory
|
810 |
+
slice_size = [dim // 2 for dim in sliceable_head_dims]
|
811 |
+
elif slice_size == "max":
|
812 |
+
# make smallest slice possible
|
813 |
+
slice_size = num_sliceable_layers * [1]
|
814 |
+
|
815 |
+
slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
|
816 |
+
|
817 |
+
if len(slice_size) != len(sliceable_head_dims):
|
818 |
+
raise ValueError(
|
819 |
+
f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
|
820 |
+
f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
|
821 |
+
)
|
822 |
+
|
823 |
+
for i in range(len(slice_size)):
|
824 |
+
size = slice_size[i]
|
825 |
+
dim = sliceable_head_dims[i]
|
826 |
+
if size is not None and size > dim:
|
827 |
+
raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
|
828 |
+
|
829 |
+
# Recursively walk through all the children.
|
830 |
+
# Any children which exposes the set_attention_slice method
|
831 |
+
# gets the message
|
832 |
+
def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
|
833 |
+
if hasattr(module, "set_attention_slice"):
|
834 |
+
module.set_attention_slice(slice_size.pop())
|
835 |
+
|
836 |
+
for child in module.children():
|
837 |
+
fn_recursive_set_attention_slice(child, slice_size)
|
838 |
+
|
839 |
+
reversed_slice_size = list(reversed(slice_size))
|
840 |
+
for module in self.children():
|
841 |
+
fn_recursive_set_attention_slice(module, reversed_slice_size)
|
842 |
+
|
843 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
844 |
+
if hasattr(module, "gradient_checkpointing"):
|
845 |
+
module.gradient_checkpointing = value
|
846 |
+
|
847 |
+
def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
|
848 |
+
r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
|
849 |
+
|
850 |
+
The suffixes after the scaling factors represent the stage blocks where they are being applied.
|
851 |
+
|
852 |
+
Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
|
853 |
+
are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
|
854 |
+
|
855 |
+
Args:
|
856 |
+
s1 (`float`):
|
857 |
+
Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
|
858 |
+
mitigate the "oversmoothing effect" in the enhanced denoising process.
|
859 |
+
s2 (`float`):
|
860 |
+
Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
|
861 |
+
mitigate the "oversmoothing effect" in the enhanced denoising process.
|
862 |
+
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
|
863 |
+
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
|
864 |
+
"""
|
865 |
+
for i, upsample_block in enumerate(self.up_blocks):
|
866 |
+
setattr(upsample_block, "s1", s1)
|
867 |
+
setattr(upsample_block, "s2", s2)
|
868 |
+
setattr(upsample_block, "b1", b1)
|
869 |
+
setattr(upsample_block, "b2", b2)
|
870 |
+
|
871 |
+
def disable_freeu(self):
|
872 |
+
"""Disables the FreeU mechanism."""
|
873 |
+
freeu_keys = {"s1", "s2", "b1", "b2"}
|
874 |
+
for i, upsample_block in enumerate(self.up_blocks):
|
875 |
+
for k in freeu_keys:
|
876 |
+
if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
|
877 |
+
setattr(upsample_block, k, None)
|
878 |
+
|
879 |
+
def fuse_qkv_projections(self):
|
880 |
+
"""
|
881 |
+
Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
|
882 |
+
are fused. For cross-attention modules, key and value projection matrices are fused.
|
883 |
+
|
884 |
+
<Tip warning={true}>
|
885 |
+
|
886 |
+
This API is 🧪 experimental.
|
887 |
+
|
888 |
+
</Tip>
|
889 |
+
"""
|
890 |
+
self.original_attn_processors = None
|
891 |
+
|
892 |
+
for _, attn_processor in self.attn_processors.items():
|
893 |
+
if "Added" in str(attn_processor.__class__.__name__):
|
894 |
+
raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
|
895 |
+
|
896 |
+
self.original_attn_processors = self.attn_processors
|
897 |
+
|
898 |
+
for module in self.modules():
|
899 |
+
if isinstance(module, Attention):
|
900 |
+
module.fuse_projections(fuse=True)
|
901 |
+
|
902 |
+
def unfuse_qkv_projections(self):
|
903 |
+
"""Disables the fused QKV projection if enabled.
|
904 |
+
|
905 |
+
<Tip warning={true}>
|
906 |
+
|
907 |
+
This API is 🧪 experimental.
|
908 |
+
|
909 |
+
</Tip>
|
910 |
+
|
911 |
+
"""
|
912 |
+
if self.original_attn_processors is not None:
|
913 |
+
self.set_attn_processor(self.original_attn_processors)
|
914 |
+
|
915 |
+
def get_time_embed(
|
916 |
+
self, sample: torch.Tensor, timestep: Union[torch.Tensor, float, int]
|
917 |
+
) -> Optional[torch.Tensor]:
|
918 |
+
timesteps = timestep
|
919 |
+
if not torch.is_tensor(timesteps):
|
920 |
+
# TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
|
921 |
+
# This would be a good case for the `match` statement (Python 3.10+)
|
922 |
+
is_mps = sample.device.type == "mps"
|
923 |
+
if isinstance(timestep, float):
|
924 |
+
dtype = torch.float32 if is_mps else torch.float64
|
925 |
+
else:
|
926 |
+
dtype = torch.int32 if is_mps else torch.int64
|
927 |
+
timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
|
928 |
+
elif len(timesteps.shape) == 0:
|
929 |
+
timesteps = timesteps[None].to(sample.device)
|
930 |
+
|
931 |
+
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
932 |
+
timesteps = timesteps.expand(sample.shape[0])
|
933 |
+
|
934 |
+
t_emb = self.time_proj(timesteps)
|
935 |
+
# `Timesteps` does not contain any weights and will always return f32 tensors
|
936 |
+
# but time_embedding might actually be running in fp16. so we need to cast here.
|
937 |
+
# there might be better ways to encapsulate this.
|
938 |
+
t_emb = t_emb.to(dtype=sample.dtype)
|
939 |
+
return t_emb
|
940 |
+
|
941 |
+
def get_class_embed(self, sample: torch.Tensor, class_labels: Optional[torch.Tensor]) -> Optional[torch.Tensor]:
|
942 |
+
class_emb = None
|
943 |
+
if self.class_embedding is not None:
|
944 |
+
if class_labels is None:
|
945 |
+
raise ValueError("class_labels should be provided when num_class_embeds > 0")
|
946 |
+
|
947 |
+
if self.config.class_embed_type == "timestep":
|
948 |
+
class_labels = self.time_proj(class_labels)
|
949 |
+
|
950 |
+
# `Timesteps` does not contain any weights and will always return f32 tensors
|
951 |
+
# there might be better ways to encapsulate this.
|
952 |
+
class_labels = class_labels.to(dtype=sample.dtype)
|
953 |
+
|
954 |
+
class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
|
955 |
+
return class_emb
|
956 |
+
|
957 |
+
def get_aug_embed(
|
958 |
+
self, emb: torch.Tensor, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
|
959 |
+
) -> Optional[torch.Tensor]:
|
960 |
+
aug_emb = None
|
961 |
+
if self.config.addition_embed_type == "text":
|
962 |
+
aug_emb = self.add_embedding(encoder_hidden_states)
|
963 |
+
elif self.config.addition_embed_type == "text_image":
|
964 |
+
# Kandinsky 2.1 - style
|
965 |
+
if "image_embeds" not in added_cond_kwargs:
|
966 |
+
raise ValueError(
|
967 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
|
968 |
+
)
|
969 |
+
|
970 |
+
image_embs = added_cond_kwargs.get("image_embeds")
|
971 |
+
text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
|
972 |
+
aug_emb = self.add_embedding(text_embs, image_embs)
|
973 |
+
elif self.config.addition_embed_type == "text_time":
|
974 |
+
# SDXL - style
|
975 |
+
if "text_embeds" not in added_cond_kwargs:
|
976 |
+
raise ValueError(
|
977 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
|
978 |
+
)
|
979 |
+
text_embeds = added_cond_kwargs.get("text_embeds")
|
980 |
+
if "time_ids" not in added_cond_kwargs:
|
981 |
+
raise ValueError(
|
982 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
|
983 |
+
)
|
984 |
+
time_ids = added_cond_kwargs.get("time_ids")
|
985 |
+
time_embeds = self.add_time_proj(time_ids.flatten())
|
986 |
+
time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
|
987 |
+
add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
|
988 |
+
add_embeds = add_embeds.to(emb.dtype)
|
989 |
+
aug_emb = self.add_embedding(add_embeds)
|
990 |
+
elif self.config.addition_embed_type == "image":
|
991 |
+
# Kandinsky 2.2 - style
|
992 |
+
if "image_embeds" not in added_cond_kwargs:
|
993 |
+
raise ValueError(
|
994 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
|
995 |
+
)
|
996 |
+
image_embs = added_cond_kwargs.get("image_embeds")
|
997 |
+
aug_emb = self.add_embedding(image_embs)
|
998 |
+
elif self.config.addition_embed_type == "image_hint":
|
999 |
+
# Kandinsky 2.2 - style
|
1000 |
+
if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
|
1001 |
+
raise ValueError(
|
1002 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
|
1003 |
+
)
|
1004 |
+
image_embs = added_cond_kwargs.get("image_embeds")
|
1005 |
+
hint = added_cond_kwargs.get("hint")
|
1006 |
+
aug_emb = self.add_embedding(image_embs, hint)
|
1007 |
+
return aug_emb
|
1008 |
+
|
1009 |
+
def process_encoder_hidden_states(
|
1010 |
+
self, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
|
1011 |
+
) -> torch.Tensor:
|
1012 |
+
if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
|
1013 |
+
encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
|
1014 |
+
elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
|
1015 |
+
# Kandinsky 2.1 - style
|
1016 |
+
if "image_embeds" not in added_cond_kwargs:
|
1017 |
+
raise ValueError(
|
1018 |
+
f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
|
1019 |
+
)
|
1020 |
+
|
1021 |
+
image_embeds = added_cond_kwargs.get("image_embeds")
|
1022 |
+
encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
|
1023 |
+
elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
|
1024 |
+
# Kandinsky 2.2 - style
|
1025 |
+
if "image_embeds" not in added_cond_kwargs:
|
1026 |
+
raise ValueError(
|
1027 |
+
f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
|
1028 |
+
)
|
1029 |
+
image_embeds = added_cond_kwargs.get("image_embeds")
|
1030 |
+
encoder_hidden_states = self.encoder_hid_proj(image_embeds)
|
1031 |
+
elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
|
1032 |
+
if "image_embeds" not in added_cond_kwargs:
|
1033 |
+
raise ValueError(
|
1034 |
+
f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
|
1035 |
+
)
|
1036 |
+
|
1037 |
+
if hasattr(self, 'text_encoder_hid_proj') and not self.text_encoder_hid_proj is None:
|
1038 |
+
encoder_hidden_states = self.text_encoder_hid_proj( encoder_hidden_states )
|
1039 |
+
|
1040 |
+
image_embeds = added_cond_kwargs.get("image_embeds")
|
1041 |
+
image_embeds = self.encoder_hid_proj(image_embeds)
|
1042 |
+
encoder_hidden_states = (encoder_hidden_states, image_embeds)
|
1043 |
+
return encoder_hidden_states
|
1044 |
+
|
1045 |
+
def forward(
|
1046 |
+
self,
|
1047 |
+
sample: torch.Tensor,
|
1048 |
+
timestep: Union[torch.Tensor, float, int],
|
1049 |
+
encoder_hidden_states: torch.Tensor,
|
1050 |
+
class_labels: Optional[torch.Tensor] = None,
|
1051 |
+
timestep_cond: Optional[torch.Tensor] = None,
|
1052 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1053 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
1054 |
+
added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
|
1055 |
+
down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
|
1056 |
+
mid_block_additional_residual: Optional[torch.Tensor] = None,
|
1057 |
+
down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
|
1058 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
1059 |
+
return_dict: bool = True,
|
1060 |
+
) -> Union[UNet2DConditionOutput, Tuple]:
|
1061 |
+
r"""
|
1062 |
+
The [`UNet2DConditionModel`] forward method.
|
1063 |
+
|
1064 |
+
Args:
|
1065 |
+
sample (`torch.Tensor`):
|
1066 |
+
The noisy input tensor with the following shape `(batch, channel, height, width)`.
|
1067 |
+
timestep (`torch.Tensor` or `float` or `int`): The number of timesteps to denoise an input.
|
1068 |
+
encoder_hidden_states (`torch.Tensor`):
|
1069 |
+
The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
|
1070 |
+
class_labels (`torch.Tensor`, *optional*, defaults to `None`):
|
1071 |
+
Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
|
1072 |
+
timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
|
1073 |
+
Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
|
1074 |
+
through the `self.time_embedding` layer to obtain the timestep embeddings.
|
1075 |
+
attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
|
1076 |
+
An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
|
1077 |
+
is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
|
1078 |
+
negative values to the attention scores corresponding to "discard" tokens.
|
1079 |
+
cross_attention_kwargs (`dict`, *optional*):
|
1080 |
+
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
1081 |
+
`self.processor` in
|
1082 |
+
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
1083 |
+
added_cond_kwargs: (`dict`, *optional*):
|
1084 |
+
A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
|
1085 |
+
are passed along to the UNet blocks.
|
1086 |
+
down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
|
1087 |
+
A tuple of tensors that if specified are added to the residuals of down unet blocks.
|
1088 |
+
mid_block_additional_residual: (`torch.Tensor`, *optional*):
|
1089 |
+
A tensor that if specified is added to the residual of the middle unet block.
|
1090 |
+
down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
|
1091 |
+
additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
|
1092 |
+
encoder_attention_mask (`torch.Tensor`):
|
1093 |
+
A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
|
1094 |
+
`True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
|
1095 |
+
which adds large negative values to the attention scores corresponding to "discard" tokens.
|
1096 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
1097 |
+
Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
|
1098 |
+
tuple.
|
1099 |
+
|
1100 |
+
Returns:
|
1101 |
+
[`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
|
1102 |
+
If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned,
|
1103 |
+
otherwise a `tuple` is returned where the first element is the sample tensor.
|
1104 |
+
"""
|
1105 |
+
# By default samples have to be AT least a multiple of the overall upsampling factor.
|
1106 |
+
# The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
|
1107 |
+
# However, the upsampling interpolation output size can be forced to fit any upsampling size
|
1108 |
+
# on the fly if necessary.
|
1109 |
+
default_overall_up_factor = 2**self.num_upsamplers
|
1110 |
+
|
1111 |
+
# upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
|
1112 |
+
forward_upsample_size = False
|
1113 |
+
upsample_size = None
|
1114 |
+
|
1115 |
+
for dim in sample.shape[-2:]:
|
1116 |
+
if dim % default_overall_up_factor != 0:
|
1117 |
+
# Forward upsample size to force interpolation output size.
|
1118 |
+
forward_upsample_size = True
|
1119 |
+
break
|
1120 |
+
|
1121 |
+
# ensure attention_mask is a bias, and give it a singleton query_tokens dimension
|
1122 |
+
# expects mask of shape:
|
1123 |
+
# [batch, key_tokens]
|
1124 |
+
# adds singleton query_tokens dimension:
|
1125 |
+
# [batch, 1, key_tokens]
|
1126 |
+
# this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
|
1127 |
+
# [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
|
1128 |
+
# [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
|
1129 |
+
if attention_mask is not None:
|
1130 |
+
# assume that mask is expressed as:
|
1131 |
+
# (1 = keep, 0 = discard)
|
1132 |
+
# convert mask into a bias that can be added to attention scores:
|
1133 |
+
# (keep = +0, discard = -10000.0)
|
1134 |
+
attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
|
1135 |
+
attention_mask = attention_mask.unsqueeze(1)
|
1136 |
+
|
1137 |
+
# convert encoder_attention_mask to a bias the same way we do for attention_mask
|
1138 |
+
if encoder_attention_mask is not None:
|
1139 |
+
encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
|
1140 |
+
encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
|
1141 |
+
|
1142 |
+
# 0. center input if necessary
|
1143 |
+
if self.config.center_input_sample:
|
1144 |
+
sample = 2 * sample - 1.0
|
1145 |
+
|
1146 |
+
# 1. time
|
1147 |
+
t_emb = self.get_time_embed(sample=sample, timestep=timestep)
|
1148 |
+
emb = self.time_embedding(t_emb, timestep_cond)
|
1149 |
+
aug_emb = None
|
1150 |
+
|
1151 |
+
class_emb = self.get_class_embed(sample=sample, class_labels=class_labels)
|
1152 |
+
if class_emb is not None:
|
1153 |
+
if self.config.class_embeddings_concat:
|
1154 |
+
emb = torch.cat([emb, class_emb], dim=-1)
|
1155 |
+
else:
|
1156 |
+
emb = emb + class_emb
|
1157 |
+
|
1158 |
+
aug_emb = self.get_aug_embed(
|
1159 |
+
emb=emb, encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
|
1160 |
+
)
|
1161 |
+
if self.config.addition_embed_type == "image_hint":
|
1162 |
+
aug_emb, hint = aug_emb
|
1163 |
+
sample = torch.cat([sample, hint], dim=1)
|
1164 |
+
|
1165 |
+
emb = emb + aug_emb if aug_emb is not None else emb
|
1166 |
+
|
1167 |
+
if self.time_embed_act is not None:
|
1168 |
+
emb = self.time_embed_act(emb)
|
1169 |
+
|
1170 |
+
encoder_hidden_states = self.process_encoder_hidden_states(
|
1171 |
+
encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
|
1172 |
+
)
|
1173 |
+
|
1174 |
+
# 2. pre-process
|
1175 |
+
sample = self.conv_in(sample)
|
1176 |
+
|
1177 |
+
# 2.5 GLIGEN position net
|
1178 |
+
if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
|
1179 |
+
cross_attention_kwargs = cross_attention_kwargs.copy()
|
1180 |
+
gligen_args = cross_attention_kwargs.pop("gligen")
|
1181 |
+
cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
|
1182 |
+
|
1183 |
+
# 3. down
|
1184 |
+
# we're popping the `scale` instead of getting it because otherwise `scale` will be propagated
|
1185 |
+
# to the internal blocks and will raise deprecation warnings. this will be confusing for our users.
|
1186 |
+
if cross_attention_kwargs is not None:
|
1187 |
+
cross_attention_kwargs = cross_attention_kwargs.copy()
|
1188 |
+
lora_scale = cross_attention_kwargs.pop("scale", 1.0)
|
1189 |
+
else:
|
1190 |
+
lora_scale = 1.0
|
1191 |
+
|
1192 |
+
if USE_PEFT_BACKEND:
|
1193 |
+
# weight the lora layers by setting `lora_scale` for each PEFT layer
|
1194 |
+
scale_lora_layers(self, lora_scale)
|
1195 |
+
|
1196 |
+
is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
|
1197 |
+
# using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
|
1198 |
+
is_adapter = down_intrablock_additional_residuals is not None
|
1199 |
+
# maintain backward compatibility for legacy usage, where
|
1200 |
+
# T2I-Adapter and ControlNet both use down_block_additional_residuals arg
|
1201 |
+
# but can only use one or the other
|
1202 |
+
if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
|
1203 |
+
deprecate(
|
1204 |
+
"T2I should not use down_block_additional_residuals",
|
1205 |
+
"1.3.0",
|
1206 |
+
"Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
|
1207 |
+
and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \
|
1208 |
+
for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
|
1209 |
+
standard_warn=False,
|
1210 |
+
)
|
1211 |
+
down_intrablock_additional_residuals = down_block_additional_residuals
|
1212 |
+
is_adapter = True
|
1213 |
+
|
1214 |
+
down_block_res_samples = (sample,)
|
1215 |
+
for downsample_block in self.down_blocks:
|
1216 |
+
if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
|
1217 |
+
# For t2i-adapter CrossAttnDownBlock2D
|
1218 |
+
additional_residuals = {}
|
1219 |
+
if is_adapter and len(down_intrablock_additional_residuals) > 0:
|
1220 |
+
additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
|
1221 |
+
|
1222 |
+
sample, res_samples = downsample_block(
|
1223 |
+
hidden_states=sample,
|
1224 |
+
temb=emb,
|
1225 |
+
encoder_hidden_states=encoder_hidden_states,
|
1226 |
+
attention_mask=attention_mask,
|
1227 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
1228 |
+
encoder_attention_mask=encoder_attention_mask,
|
1229 |
+
**additional_residuals,
|
1230 |
+
)
|
1231 |
+
else:
|
1232 |
+
sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
|
1233 |
+
if is_adapter and len(down_intrablock_additional_residuals) > 0:
|
1234 |
+
sample += down_intrablock_additional_residuals.pop(0)
|
1235 |
+
|
1236 |
+
down_block_res_samples += res_samples
|
1237 |
+
|
1238 |
+
if is_controlnet:
|
1239 |
+
new_down_block_res_samples = ()
|
1240 |
+
|
1241 |
+
for down_block_res_sample, down_block_additional_residual in zip(
|
1242 |
+
down_block_res_samples, down_block_additional_residuals
|
1243 |
+
):
|
1244 |
+
down_block_res_sample = down_block_res_sample + down_block_additional_residual
|
1245 |
+
new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
|
1246 |
+
|
1247 |
+
down_block_res_samples = new_down_block_res_samples
|
1248 |
+
|
1249 |
+
# 4. mid
|
1250 |
+
if self.mid_block is not None:
|
1251 |
+
if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
|
1252 |
+
sample = self.mid_block(
|
1253 |
+
sample,
|
1254 |
+
emb,
|
1255 |
+
encoder_hidden_states=encoder_hidden_states,
|
1256 |
+
attention_mask=attention_mask,
|
1257 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
1258 |
+
encoder_attention_mask=encoder_attention_mask,
|
1259 |
+
)
|
1260 |
+
else:
|
1261 |
+
sample = self.mid_block(sample, emb)
|
1262 |
+
|
1263 |
+
# To support T2I-Adapter-XL
|
1264 |
+
if (
|
1265 |
+
is_adapter
|
1266 |
+
and len(down_intrablock_additional_residuals) > 0
|
1267 |
+
and sample.shape == down_intrablock_additional_residuals[0].shape
|
1268 |
+
):
|
1269 |
+
sample += down_intrablock_additional_residuals.pop(0)
|
1270 |
+
|
1271 |
+
if is_controlnet:
|
1272 |
+
sample = sample + mid_block_additional_residual
|
1273 |
+
|
1274 |
+
# 5. up
|
1275 |
+
for i, upsample_block in enumerate(self.up_blocks):
|
1276 |
+
is_final_block = i == len(self.up_blocks) - 1
|
1277 |
+
|
1278 |
+
res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
|
1279 |
+
down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
|
1280 |
+
|
1281 |
+
# if we have not reached the final block and need to forward the
|
1282 |
+
# upsample size, we do it here
|
1283 |
+
if not is_final_block and forward_upsample_size:
|
1284 |
+
upsample_size = down_block_res_samples[-1].shape[2:]
|
1285 |
+
|
1286 |
+
if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
|
1287 |
+
sample = upsample_block(
|
1288 |
+
hidden_states=sample,
|
1289 |
+
temb=emb,
|
1290 |
+
res_hidden_states_tuple=res_samples,
|
1291 |
+
encoder_hidden_states=encoder_hidden_states,
|
1292 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
1293 |
+
upsample_size=upsample_size,
|
1294 |
+
attention_mask=attention_mask,
|
1295 |
+
encoder_attention_mask=encoder_attention_mask,
|
1296 |
+
)
|
1297 |
+
else:
|
1298 |
+
sample = upsample_block(
|
1299 |
+
hidden_states=sample,
|
1300 |
+
temb=emb,
|
1301 |
+
res_hidden_states_tuple=res_samples,
|
1302 |
+
upsample_size=upsample_size,
|
1303 |
+
)
|
1304 |
+
|
1305 |
+
# 6. post-process
|
1306 |
+
if self.conv_norm_out:
|
1307 |
+
sample = self.conv_norm_out(sample)
|
1308 |
+
sample = self.conv_act(sample)
|
1309 |
+
sample = self.conv_out(sample)
|
1310 |
+
|
1311 |
+
if USE_PEFT_BACKEND:
|
1312 |
+
# remove `lora_scale` from each PEFT layer
|
1313 |
+
unscale_lora_layers(self, lora_scale)
|
1314 |
+
|
1315 |
+
if not return_dict:
|
1316 |
+
return (sample,)
|
1317 |
+
|
1318 |
+
return UNet2DConditionOutput(sample=sample)
|