YaTharThShaRma999 commited on
Commit
e56e3fa
1 Parent(s): 7600eb9

Upload CogVLMForCausalLM

Browse files
README.md ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: transformers
3
+ tags: []
4
+ ---
5
+
6
+ # Model Card for Model ID
7
+
8
+ <!-- Provide a quick summary of what the model is/does. -->
9
+
10
+
11
+
12
+ ## Model Details
13
+
14
+ ### Model Description
15
+
16
+ <!-- Provide a longer summary of what this model is. -->
17
+
18
+ This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
19
+
20
+ - **Developed by:** [More Information Needed]
21
+ - **Funded by [optional]:** [More Information Needed]
22
+ - **Shared by [optional]:** [More Information Needed]
23
+ - **Model type:** [More Information Needed]
24
+ - **Language(s) (NLP):** [More Information Needed]
25
+ - **License:** [More Information Needed]
26
+ - **Finetuned from model [optional]:** [More Information Needed]
27
+
28
+ ### Model Sources [optional]
29
+
30
+ <!-- Provide the basic links for the model. -->
31
+
32
+ - **Repository:** [More Information Needed]
33
+ - **Paper [optional]:** [More Information Needed]
34
+ - **Demo [optional]:** [More Information Needed]
35
+
36
+ ## Uses
37
+
38
+ <!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
39
+
40
+ ### Direct Use
41
+
42
+ <!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
43
+
44
+ [More Information Needed]
45
+
46
+ ### Downstream Use [optional]
47
+
48
+ <!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
49
+
50
+ [More Information Needed]
51
+
52
+ ### Out-of-Scope Use
53
+
54
+ <!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
55
+
56
+ [More Information Needed]
57
+
58
+ ## Bias, Risks, and Limitations
59
+
60
+ <!-- This section is meant to convey both technical and sociotechnical limitations. -->
61
+
62
+ [More Information Needed]
63
+
64
+ ### Recommendations
65
+
66
+ <!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
67
+
68
+ Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
69
+
70
+ ## How to Get Started with the Model
71
+
72
+ Use the code below to get started with the model.
73
+
74
+ [More Information Needed]
75
+
76
+ ## Training Details
77
+
78
+ ### Training Data
79
+
80
+ <!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
81
+
82
+ [More Information Needed]
83
+
84
+ ### Training Procedure
85
+
86
+ <!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
87
+
88
+ #### Preprocessing [optional]
89
+
90
+ [More Information Needed]
91
+
92
+
93
+ #### Training Hyperparameters
94
+
95
+ - **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
96
+
97
+ #### Speeds, Sizes, Times [optional]
98
+
99
+ <!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
100
+
101
+ [More Information Needed]
102
+
103
+ ## Evaluation
104
+
105
+ <!-- This section describes the evaluation protocols and provides the results. -->
106
+
107
+ ### Testing Data, Factors & Metrics
108
+
109
+ #### Testing Data
110
+
111
+ <!-- This should link to a Dataset Card if possible. -->
112
+
113
+ [More Information Needed]
114
+
115
+ #### Factors
116
+
117
+ <!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
118
+
119
+ [More Information Needed]
120
+
121
+ #### Metrics
122
+
123
+ <!-- These are the evaluation metrics being used, ideally with a description of why. -->
124
+
125
+ [More Information Needed]
126
+
127
+ ### Results
128
+
129
+ [More Information Needed]
130
+
131
+ #### Summary
132
+
133
+
134
+
135
+ ## Model Examination [optional]
136
+
137
+ <!-- Relevant interpretability work for the model goes here -->
138
+
139
+ [More Information Needed]
140
+
141
+ ## Environmental Impact
142
+
143
+ <!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
144
+
145
+ Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
146
+
147
+ - **Hardware Type:** [More Information Needed]
148
+ - **Hours used:** [More Information Needed]
149
+ - **Cloud Provider:** [More Information Needed]
150
+ - **Compute Region:** [More Information Needed]
151
+ - **Carbon Emitted:** [More Information Needed]
152
+
153
+ ## Technical Specifications [optional]
154
+
155
+ ### Model Architecture and Objective
156
+
157
+ [More Information Needed]
158
+
159
+ ### Compute Infrastructure
160
+
161
+ [More Information Needed]
162
+
163
+ #### Hardware
164
+
165
+ [More Information Needed]
166
+
167
+ #### Software
168
+
169
+ [More Information Needed]
170
+
171
+ ## Citation [optional]
172
+
173
+ <!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
174
+
175
+ **BibTeX:**
176
+
177
+ [More Information Needed]
178
+
179
+ **APA:**
180
+
181
+ [More Information Needed]
182
+
183
+ ## Glossary [optional]
184
+
185
+ <!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
186
+
187
+ [More Information Needed]
188
+
189
+ ## More Information [optional]
190
+
191
+ [More Information Needed]
192
+
193
+ ## Model Card Authors [optional]
194
+
195
+ [More Information Needed]
196
+
197
+ ## Model Card Contact
198
+
199
+ [More Information Needed]
200
+
201
+
config.json ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "THUDM/cogvlm-grounding-generalist-hf",
3
+ "architectures": [
4
+ "CogVLMForCausalLM"
5
+ ],
6
+ "auto_map": {
7
+ "AutoConfig": "configuration_cogvlm.CogVLMConfig",
8
+ "AutoModelForCausalLM": "modeling_cogvlm.CogVLMForCausalLM"
9
+ },
10
+ "bos_token_id": 1,
11
+ "eos_token_id": 2,
12
+ "hidden_act": "silu",
13
+ "hidden_size": 4096,
14
+ "initializer_range": 0.02,
15
+ "intermediate_size": 11008,
16
+ "max_position_embeddings": 2048,
17
+ "num_attention_heads": 32,
18
+ "num_hidden_layers": 32,
19
+ "pad_token_id": 0,
20
+ "quantization_config": {
21
+ "bnb_4bit_compute_dtype": "float32",
22
+ "bnb_4bit_quant_type": "fp4",
23
+ "bnb_4bit_use_double_quant": false,
24
+ "llm_int8_enable_fp32_cpu_offload": false,
25
+ "llm_int8_has_fp16_weight": false,
26
+ "llm_int8_skip_modules": null,
27
+ "llm_int8_threshold": 6.0,
28
+ "load_in_4bit": true,
29
+ "load_in_8bit": false,
30
+ "quant_method": "bitsandbytes"
31
+ },
32
+ "rms_norm_eps": 1e-05,
33
+ "template_version": "base",
34
+ "tie_word_embeddings": false,
35
+ "torch_dtype": "float16",
36
+ "transformers_version": "4.37.2",
37
+ "use_cache": true,
38
+ "vision_config": {
39
+ "dropout_prob": 0.0,
40
+ "hidden_act": "gelu",
41
+ "hidden_size": 1792,
42
+ "image_size": 490,
43
+ "in_channels": 3,
44
+ "intermediate_size": 15360,
45
+ "layer_norm_eps": 1e-06,
46
+ "num_heads": 16,
47
+ "num_hidden_layers": 63,
48
+ "num_positions": 1226,
49
+ "patch_size": 14
50
+ },
51
+ "vocab_size": 32000
52
+ }
configuration_cogvlm.py ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Literal
2
+ from transformers import PretrainedConfig
3
+
4
+
5
+ class CogVLMConfig(PretrainedConfig):
6
+ _auto_class = "AutoConfig"
7
+
8
+ def __init__(
9
+ self,
10
+ vocab_size=32000,
11
+ hidden_size=4096,
12
+ intermediate_size=11008,
13
+ num_hidden_layers=32,
14
+ num_attention_heads=32,
15
+ hidden_act='silu',
16
+ max_position_embeddings=2048,
17
+ initializer_range=0.02,
18
+ rms_norm_eps=1e-06,
19
+ template_version: Literal["base", "chat"] = "chat",
20
+
21
+ pad_token_id=0,
22
+ bos_token_id=1,
23
+ eos_token_id=2,
24
+ tie_word_embeddings=False,
25
+ use_cache=True,
26
+ **kwargs,
27
+ ):
28
+ self.hidden_size = hidden_size
29
+ self.intermediate_size = intermediate_size
30
+ self.num_attention_heads = num_attention_heads
31
+ self.max_position_embeddings = max_position_embeddings
32
+ self.rms_norm_eps = rms_norm_eps
33
+ self.initializer_range = initializer_range
34
+ self.vocab_size = vocab_size
35
+ self.num_hidden_layers = num_hidden_layers
36
+ self.hidden_act = hidden_act
37
+ self.template_version = template_version
38
+ self.use_cache = use_cache
39
+ super().__init__(
40
+ pad_token_id=pad_token_id,
41
+ bos_token_id=bos_token_id,
42
+ eos_token_id=eos_token_id,
43
+ tie_word_embeddings=tie_word_embeddings,
44
+ **kwargs,
45
+ )
generation_config.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 1,
4
+ "eos_token_id": 2,
5
+ "pad_token_id": 0,
6
+ "transformers_version": "4.37.2"
7
+ }
model-00001-of-00003.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:395cf2c82cd877909eb3c347d0393c475fa1c1def0ed4bf2dd534767fd9dfb46
3
+ size 4993012638
model-00002-of-00003.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4df2aa6ad66d3139c002e39ed77c7caade6de7a31646cb76f7f3bab670e8dd45
3
+ size 4998869581
model-00003-of-00003.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:05aa81ff280c89c663e43fea01f151eaa44c4a9e1ea1e726ae9a8974692df9af
3
+ size 315688854
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
modeling_cogvlm.py ADDED
@@ -0,0 +1,834 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """largely copy from llama and adapt for cogvlm"""
2
+ import warnings
3
+ from typing import TYPE_CHECKING, Optional, Tuple, List, Union, Literal, Dict, Any
4
+
5
+ import math
6
+ import torch
7
+ from torch import nn
8
+ from torch.nn import functional as F
9
+ from torch.nn import CrossEntropyLoss
10
+ from torchvision import transforms
11
+ from einops import rearrange
12
+
13
+ from transformers import PreTrainedModel, PreTrainedTokenizer
14
+ from transformers.utils.logging import get_logger
15
+ from transformers.activations import ACT2FN
16
+ from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
17
+
18
+ from .configuration_cogvlm import CogVLMConfig
19
+ from .visual import EVA2CLIPModel
20
+
21
+ if TYPE_CHECKING:
22
+ from transformers.utils import ModelOutput
23
+
24
+ logger = get_logger(__name__)
25
+
26
+ LANGUAGE_TOKEN_TYPE = 0
27
+ VISION_TOKEN_TYPE = 1
28
+
29
+
30
+ # Copied from transformers.models.bart.modeling_bart._make_causal_mask
31
+ def _make_causal_mask(
32
+ input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
33
+ ):
34
+ """
35
+ Make causal mask used for bi-directional self-attention.
36
+ """
37
+ bsz, tgt_len = input_ids_shape
38
+ mask = torch.full((tgt_len, tgt_len), torch.finfo(dtype).min, device=device)
39
+ mask_cond = torch.arange(mask.size(-1), device=device)
40
+ mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
41
+ mask = mask.to(dtype)
42
+
43
+ if past_key_values_length > 0:
44
+ mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
45
+ return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
46
+
47
+
48
+ # Copied from transformers.models.bart.modeling_bart._expand_mask
49
+ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
50
+ """
51
+ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
52
+ """
53
+ bsz, src_len = mask.size()
54
+ tgt_len = tgt_len if tgt_len is not None else src_len
55
+
56
+ expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
57
+
58
+ inverted_mask = 1.0 - expanded_mask
59
+
60
+ return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
61
+
62
+
63
+ class RMSNorm(nn.Module):
64
+ def __init__(self, hidden_size, eps=1e-6):
65
+ super().__init__()
66
+ self.weight = nn.Parameter(torch.ones(hidden_size))
67
+ self.variance_epsilon = eps
68
+
69
+ def forward(self, hidden_states):
70
+ input_dtype = hidden_states.dtype
71
+ hidden_states = hidden_states.to(torch.float32)
72
+ variance = hidden_states.pow(2).mean(-1, keepdim=True)
73
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
74
+ return (self.weight * hidden_states).to(input_dtype)
75
+
76
+
77
+ class MLP(nn.Module):
78
+ def __init__(self, config):
79
+ super().__init__()
80
+ self.hidden_size = config.hidden_size
81
+ self.intermediate_size = config.intermediate_size
82
+ self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
83
+ self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
84
+ self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
85
+ self.act_fn = ACT2FN[config.hidden_act]
86
+
87
+ def forward(self, x):
88
+ down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
89
+ return down_proj
90
+
91
+
92
+ def get_expert_mask(token_type_ids: "torch.LongTensor(B, L)") -> "[torch.BoolTensor(B, L), torch.BoolTensor(B, L)]":
93
+ vision_token_mask = torch.zeros_like(token_type_ids, dtype=torch.bool)
94
+ vision_token_mask[:, :-1] = (token_type_ids[:, :-1] == VISION_TOKEN_TYPE) & (token_type_ids[:, 1:] == VISION_TOKEN_TYPE)
95
+ language_token_mask = ~vision_token_mask
96
+ return vision_token_mask, language_token_mask
97
+
98
+
99
+ class VisionExpertMLP(nn.Module):
100
+ def __init__(self, config):
101
+ super().__init__()
102
+ self.language_mlp = MLP(config)
103
+ self.vision_mlp = MLP(config)
104
+
105
+ def forward(self, hidden_states: "torch.Tensor(B, L, D)", token_type_ids: "torch.LongTensor(B, L)"):
106
+ output = torch.empty(hidden_states.shape, dtype=hidden_states.dtype, device=hidden_states.device)
107
+ vision_token_mask, language_token_mask = get_expert_mask(token_type_ids)
108
+ output[vision_token_mask] = self.vision_mlp(hidden_states[vision_token_mask])
109
+ output[language_token_mask] = self.language_mlp(hidden_states[language_token_mask])
110
+ return output
111
+
112
+
113
+ def attention_fn(
114
+ query_layer: "torch.tensor(B, H, L, HD)",
115
+ key_layer: "torch.tensor(B, H, L, HD)",
116
+ value_layer: "torch.tensor(B, H, L, HD)",
117
+ attention_mask: "torch.tensor(B, H, L, HD)",
118
+ *,
119
+ scaling_attention_score: bool = True,
120
+ attention_dropout: nn.Module = None
121
+ ):
122
+ attention_mask_bool = (attention_mask == 0)
123
+ is_low_triangle = (attention_mask_bool == torch.ones_like(attention_mask_bool, dtype=torch.float).tril()).all()
124
+ is_full = (attention_mask_bool > 0).all()
125
+ if not (int(torch.__version__.split('.')[0]) >= 2):
126
+ warnings.warn("It's recommended to use torch2.0 or higher.")
127
+ if int(torch.__version__.split('.')[0]) >= 2 and scaling_attention_score and (is_full or is_low_triangle):
128
+ dropout_p = 0. if attention_dropout is None or not attention_dropout.training else attention_dropout.p
129
+ return torch.nn.functional.scaled_dot_product_attention(
130
+ query_layer, key_layer, value_layer,
131
+ attn_mask=None,
132
+ dropout_p=dropout_p,
133
+ is_causal=not is_full
134
+ )
135
+ else:
136
+ if scaling_attention_score:
137
+ query_layer = query_layer / math.sqrt(query_layer.shape[-1])
138
+ attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
139
+ attention_scores = attention_scores + attention_mask
140
+ attention_scores = nn.functional.softmax(attention_scores, dim=-1, dtype=torch.float32).to(query_layer.dtype)
141
+ if attention_dropout is not None:
142
+ attention_scores = attention_dropout(attention_scores)
143
+ context_layer = torch.matmul(attention_scores, value_layer)
144
+ return context_layer
145
+
146
+
147
+ class RotaryEmbedding(torch.nn.Module):
148
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
149
+ super().__init__()
150
+
151
+ self.dim = dim
152
+ self.max_position_embeddings = max_position_embeddings
153
+ self.base = base
154
+ inv_freq = self._compute_inv_freq(device)
155
+ self.register_buffer("inv_freq", inv_freq)
156
+ self.max_seq_len_cached = 0
157
+
158
+ def _compute_inv_freq(self, device=None):
159
+ return 1.0 / (
160
+ self.base
161
+ ** (torch.arange(0, self.dim, 2, device=device) / self.dim)
162
+ )
163
+
164
+ def _set_cos_sin_cache(self, seq_len, device, dtype):
165
+ self.max_seq_len_cached = seq_len
166
+ t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)
167
+
168
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
169
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
170
+ emb = torch.cat((freqs, freqs), dim=-1)
171
+ self.register_buffer("cos_cached", emb.cos()[:, None, :].to(dtype), persistent=False)
172
+ self.register_buffer("sin_cached", emb.sin()[:, None, :].to(dtype), persistent=False)
173
+
174
+ def forward(self, x, seq_len):
175
+ # x: [bs, num_attention_heads, seq_len, head_size]
176
+ if seq_len > self.max_seq_len_cached:
177
+ self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
178
+
179
+ return (
180
+ self.cos_cached[:seq_len, ...].to(dtype=x.dtype),
181
+ self.sin_cached[:seq_len, ...].to(dtype=x.dtype),
182
+ )
183
+
184
+
185
+ def rotate_half(x):
186
+ x1, x2 = x[..., :x.shape[-1] // 2], x[..., x.shape[-1] // 2:]
187
+ return torch.cat((-x2, x1), dim=x1.ndim - 1)
188
+
189
+
190
+ def apply_rotary_pos_emb_index_bhs(q, k, cos, sin, position_id):
191
+ # batch_size, num_head, seq_len, hidden_size
192
+ cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(1), \
193
+ F.embedding(position_id, sin.squeeze(1)).unsqueeze(1)
194
+ q, k = (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)
195
+ return q, k
196
+
197
+
198
+ class VisionExpertAttention(nn.Module):
199
+ def __init__(self, config):
200
+ super().__init__()
201
+ self.config = config
202
+ self.hidden_size = config.hidden_size
203
+ self.num_heads = config.num_attention_heads
204
+ self.head_dim = self.hidden_size // self.num_heads
205
+ self.max_position_embeddings = config.max_position_embeddings
206
+
207
+ self.rotary_emb = RotaryEmbedding(self.head_dim)
208
+ self.vision_expert_query_key_value = nn.Linear(self.hidden_size, self.hidden_size * 3, bias=False)
209
+ self.vision_expert_dense = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
210
+ self.language_expert_query_key_value = nn.Linear(self.hidden_size, self.hidden_size * 3, bias=False)
211
+ self.language_expert_dense = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
212
+
213
+ def _transpose_for_scores(self, tensor):
214
+ """Transpose a 3D tensor [B, L, H*HD] into a 4D tensor with size [B H L HD]."""
215
+ new_tensor_shape = tensor.size()[:-1] + (self.num_heads, self.head_dim)
216
+ tensor = tensor.view(*new_tensor_shape)
217
+ return tensor.permute(0, 2, 1, 3)
218
+
219
+ def forward(
220
+ self,
221
+ hidden_states: torch.Tensor,
222
+ token_type_ids: torch.LongTensor,
223
+ position_ids: torch.LongTensor,
224
+ attention_mask: Optional[torch.Tensor] = None,
225
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
226
+ output_attentions: bool = False,
227
+ use_cache: bool = False,
228
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
229
+ bsz, q_len, _ = hidden_states.size()
230
+ vision_token_mask, language_token_mask = get_expert_mask(token_type_ids)
231
+
232
+ shape = list(hidden_states.shape)
233
+ shape[-1] = shape[-1] * 3
234
+ mixed_raw_layer = torch.empty(shape, dtype=hidden_states.dtype, device=hidden_states.device)
235
+ mixed_raw_layer[vision_token_mask] = self.vision_expert_query_key_value(hidden_states[vision_token_mask])
236
+ mixed_raw_layer[language_token_mask] = self.language_expert_query_key_value(hidden_states[language_token_mask])
237
+
238
+ query_states, key_states, value_states = torch.split(mixed_raw_layer, self.hidden_size, dim=-1)
239
+ query_states = self._transpose_for_scores(query_states) # B, H, L, HD
240
+ key_states = self._transpose_for_scores(key_states) # B, H, L, HD
241
+ value_states = self._transpose_for_scores(value_states) # B, H, L, HD
242
+
243
+ kv_seq_len = key_states.shape[-2]
244
+ if past_key_value is not None:
245
+ kv_seq_len += past_key_value[0].shape[-2]
246
+ cos, sin = self.rotary_emb(value_states, seq_len=position_ids.max() + 1)
247
+ query_states, key_states = apply_rotary_pos_emb_index_bhs(query_states, key_states, cos, sin, position_ids)
248
+
249
+ if past_key_value is not None:
250
+ key_states = torch.cat([past_key_value[0], key_states], dim=2)
251
+ value_states = torch.cat([past_key_value[1], value_states], dim=2)
252
+
253
+ past_key_value = (key_states, value_states) if use_cache else None
254
+
255
+ context_layer = attention_fn(
256
+ query_layer=query_states, key_layer=key_states, value_layer=value_states, attention_mask=attention_mask,
257
+ scaling_attention_score=True, attention_dropout=None)
258
+ if context_layer.size() != (bsz, self.num_heads, q_len, self.head_dim):
259
+ raise ValueError(
260
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
261
+ f" {context_layer.size()}"
262
+ )
263
+ context_layer = context_layer.transpose(1, 2).contiguous().reshape(bsz, q_len, self.hidden_size)
264
+
265
+ attn_output = torch.empty(context_layer.shape, dtype=hidden_states.dtype, device=hidden_states.device)
266
+ attn_output[vision_token_mask] = self.vision_expert_dense(context_layer[vision_token_mask])
267
+ attn_output[language_token_mask] = self.language_expert_dense(context_layer[language_token_mask])
268
+
269
+ if output_attentions:
270
+ warnings.warn("output_attentions is not implemented.")
271
+
272
+ return attn_output, None, past_key_value
273
+
274
+
275
+ class CogVLMDecoderLayer(nn.Module):
276
+ def __init__(self, config):
277
+ super().__init__()
278
+ self.hidden_size = config.hidden_size
279
+ self.self_attn = VisionExpertAttention(config=config)
280
+ self.mlp = VisionExpertMLP(config)
281
+ self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
282
+ self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
283
+
284
+ def forward(
285
+ self,
286
+ hidden_states: torch.Tensor,
287
+ token_type_ids: torch.LongTensor,
288
+ position_ids: torch.LongTensor,
289
+ attention_mask: Optional[torch.Tensor] = None,
290
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
291
+ output_attentions: Optional[bool] = False,
292
+ use_cache: Optional[bool] = False,
293
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
294
+ residual = hidden_states
295
+
296
+ hidden_states = self.input_layernorm(hidden_states)
297
+
298
+ # Self Attention
299
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
300
+ hidden_states=hidden_states,
301
+ token_type_ids=token_type_ids,
302
+ position_ids=position_ids,
303
+ attention_mask=attention_mask,
304
+ past_key_value=past_key_value,
305
+ output_attentions=output_attentions,
306
+ use_cache=use_cache,
307
+ )
308
+ hidden_states = residual + hidden_states
309
+
310
+ # Fully Connected
311
+ residual = hidden_states
312
+ hidden_states = self.post_attention_layernorm(hidden_states)
313
+ hidden_states = self.mlp(hidden_states, token_type_ids=token_type_ids)
314
+ hidden_states = residual + hidden_states
315
+
316
+ outputs = (hidden_states,)
317
+
318
+ if output_attentions:
319
+ outputs += (self_attn_weights,)
320
+
321
+ if use_cache:
322
+ outputs += (present_key_value,)
323
+
324
+ return outputs # type: ignore
325
+
326
+
327
+ class CogVLMPreTrainedModel(PreTrainedModel):
328
+ config_class = CogVLMConfig
329
+ base_model_prefix = "model"
330
+ supports_gradient_checkpointing = False
331
+ _no_split_modules = ["CogVLMDecoderLayer", "TransformerLayer"]
332
+ _skip_keys_device_placement = "past_key_values"
333
+
334
+ def _init_weights(self, module):
335
+ std = self.config.initializer_range
336
+ if isinstance(module, nn.Linear):
337
+ module.weight.data.normal_(mean=0.0, std=std)
338
+ if module.bias is not None:
339
+ module.bias.data.zero_()
340
+ elif isinstance(module, nn.Embedding):
341
+ module.weight.data.normal_(mean=0.0, std=std)
342
+ if module.padding_idx is not None:
343
+ module.weight.data[module.padding_idx].zero_()
344
+
345
+
346
+ def is_empty(images_list: Optional[List[List[torch.Tensor]]]):
347
+ if images_list is None or len(images_list) == 0:
348
+ return True
349
+ for image_list in images_list:
350
+ if len(image_list):
351
+ return False
352
+ return True
353
+
354
+
355
+ def build_position_ids(x: "torch.BoolTensor(B, L)", attention_mask: Optional["torch.BoolTensor(B, L)"] = None) -> "torch.LongTensor(B, L)":
356
+ if attention_mask is not None:
357
+ tmp = x.clone()
358
+ tmp[~(attention_mask.bool())] = -1
359
+ else:
360
+ tmp = x.clone()
361
+ # image boi eoi token as LANGUAGE_TOKEN_TYPE
362
+ is_boi_eoi = torch.zeros_like(x, dtype=torch.bool)
363
+ is_boi_eoi[:, 1:] |= (tmp[:, 1:] == VISION_TOKEN_TYPE) & (tmp[:, :-1] == LANGUAGE_TOKEN_TYPE)
364
+ is_boi_eoi[:, 0] |= (tmp[:, 0] == VISION_TOKEN_TYPE)
365
+ is_boi_eoi[:, :-1] |= (tmp[:, :-1] == VISION_TOKEN_TYPE) & (tmp[:, 1:] == LANGUAGE_TOKEN_TYPE)
366
+ is_boi_eoi[:, -1] |= (tmp[:, -1] == VISION_TOKEN_TYPE)
367
+ tmp[is_boi_eoi] = LANGUAGE_TOKEN_TYPE
368
+ # final position ids
369
+ y = torch.zeros_like(x, dtype=torch.long)
370
+ y[:, 1:] = (tmp[:, 1:] == LANGUAGE_TOKEN_TYPE) | ((tmp[:, 1:] == VISION_TOKEN_TYPE) & (tmp[:, :-1] == LANGUAGE_TOKEN_TYPE))
371
+ y = y.cumsum(dim=-1)
372
+ return y
373
+
374
+
375
+ class CogVLMModel(CogVLMPreTrainedModel):
376
+ def __init__(self, config):
377
+ super().__init__(config)
378
+ self.padding_idx = config.pad_token_id
379
+ self.vocab_size = config.vocab_size
380
+
381
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
382
+ self.layers = nn.ModuleList([CogVLMDecoderLayer(config) for _ in range(config.num_hidden_layers)])
383
+ self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
384
+
385
+ self.vision = EVA2CLIPModel(config)
386
+
387
+ self.gradient_checkpointing = False
388
+ # Initialize weights and apply final processing
389
+ self.post_init()
390
+
391
+ def encode_images(self, images: List[List[torch.Tensor]]) -> torch.Tensor:
392
+ images_list, images = images, []
393
+
394
+ images = []
395
+ for image_list in images_list:
396
+ for image in image_list:
397
+ images.append(image)
398
+
399
+ images = torch.stack(images)
400
+ images_features = self.vision(images)
401
+ return images_features
402
+
403
+ def forward(
404
+ self,
405
+ input_ids: torch.LongTensor = None,
406
+ images: List[List[torch.Tensor]] = None,
407
+ token_type_ids: Optional[torch.LongTensor] = None,
408
+ attention_mask: Optional[torch.Tensor] = None,
409
+ position_ids: Optional[torch.LongTensor] = None,
410
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
411
+ inputs_embeds: Optional[torch.FloatTensor] = None,
412
+ use_cache: Optional[bool] = None,
413
+ output_attentions: Optional[bool] = None,
414
+ output_hidden_states: Optional[bool] = None,
415
+ return_dict: Optional[bool] = None,
416
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
417
+ """take care of image_encode, token_type_ids, position_ids and (attention_mask = None is fine)"""
418
+
419
+ if past_key_values is not None:
420
+ pass # generate mode with past_key_values. the image features are already mapped
421
+ else:
422
+ # not allow for inputs_embeds, because we want to process image feature
423
+ assert input_ids is not None and inputs_embeds is None, f"{input_ids} {inputs_embeds}"
424
+ if not is_empty(images): # multi-modality
425
+ assert token_type_ids is not None, f"multi-modality requires `token_type_ids`!"
426
+ assert len(input_ids) == len(images), f"{len(input_ids)} {len(images)}"
427
+ inputs_embeds = self.embed_tokens(input_ids)
428
+ images_features = self.encode_images(images)
429
+ images_features = rearrange(images_features, 'b n d -> (b n) d')
430
+ images_features = images_features.to(dtype=inputs_embeds.dtype, device=inputs_embeds.device)
431
+ inputs_embeds = inputs_embeds.index_put([token_type_ids == VISION_TOKEN_TYPE], images_features)
432
+ else: # single-modality
433
+ if token_type_ids is None:
434
+ token_type_ids = torch.ones_like(input_ids, dtype=torch.long, device=input_ids.device) * LANGUAGE_TOKEN_TYPE
435
+ assert not (token_type_ids == VISION_TOKEN_TYPE).any(), f"{(token_type_ids == VISION_TOKEN_TYPE).sum()}"
436
+ inputs_embeds = self.embed_tokens(input_ids)
437
+
438
+ if position_ids is None:
439
+ position_ids = build_position_ids(token_type_ids, attention_mask)
440
+ input_ids = None
441
+
442
+ return self.llm_forward(
443
+ input_ids=input_ids,
444
+ token_type_ids=token_type_ids,
445
+ attention_mask=attention_mask,
446
+ position_ids=position_ids,
447
+ past_key_values=past_key_values,
448
+ inputs_embeds=inputs_embeds,
449
+ use_cache=use_cache,
450
+ output_attentions=output_attentions,
451
+ output_hidden_states=output_hidden_states,
452
+ return_dict=return_dict,
453
+ )
454
+
455
+ def llm_forward(
456
+ self,
457
+ input_ids: torch.LongTensor = None,
458
+ token_type_ids: torch.LongTensor = None,
459
+ attention_mask: Optional[torch.Tensor] = None,
460
+ position_ids: Optional[torch.LongTensor] = None,
461
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
462
+ inputs_embeds: Optional[torch.FloatTensor] = None,
463
+ use_cache: Optional[bool] = None,
464
+ output_attentions: Optional[bool] = None,
465
+ output_hidden_states: Optional[bool] = None,
466
+ return_dict: Optional[bool] = None,
467
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
468
+ """largely copy from llama forward and adapt for cogvlm with `token_type_ids`"""
469
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
470
+ output_hidden_states = (
471
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
472
+ )
473
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
474
+
475
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
476
+
477
+ # retrieve input_ids and inputs_embeds
478
+ if input_ids is not None and inputs_embeds is not None:
479
+ raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
480
+ elif input_ids is not None:
481
+ batch_size, seq_length = input_ids.shape
482
+ elif inputs_embeds is not None:
483
+ batch_size, seq_length, _ = inputs_embeds.shape
484
+ else:
485
+ raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
486
+
487
+ seq_length_with_past = seq_length
488
+ past_key_values_length = 0
489
+
490
+ if past_key_values is not None:
491
+ past_key_values_length = past_key_values[0][0].shape[2]
492
+ seq_length_with_past = seq_length_with_past + past_key_values_length
493
+
494
+ if position_ids is None:
495
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
496
+ position_ids = torch.arange(
497
+ past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
498
+ )
499
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
500
+ else:
501
+ position_ids = position_ids.view(-1, seq_length).long()
502
+
503
+ if inputs_embeds is None:
504
+ inputs_embeds = self.embed_tokens(input_ids)
505
+ # embed positions
506
+ if attention_mask is None:
507
+ attention_mask = torch.ones(
508
+ (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
509
+ )
510
+ attention_mask = self._prepare_decoder_attention_mask(
511
+ attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
512
+ )
513
+
514
+ hidden_states = inputs_embeds
515
+
516
+ # decoder layers
517
+ all_hidden_states = () if output_hidden_states else None
518
+ all_self_attns = () if output_attentions else None
519
+ next_decoder_cache = () if use_cache else None
520
+
521
+ for idx, decoder_layer in enumerate(self.layers):
522
+ if output_hidden_states:
523
+ all_hidden_states += (hidden_states,)
524
+
525
+ past_key_value = past_key_values[idx] if past_key_values is not None else None
526
+ layer_outputs = decoder_layer(
527
+ hidden_states,
528
+ token_type_ids=token_type_ids,
529
+ attention_mask=attention_mask,
530
+ position_ids=position_ids,
531
+ past_key_value=past_key_value,
532
+ output_attentions=output_attentions,
533
+ use_cache=use_cache,
534
+ )
535
+ hidden_states = layer_outputs[0]
536
+
537
+ if use_cache:
538
+ next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
539
+
540
+ if output_attentions:
541
+ all_self_attns += (layer_outputs[1],)
542
+
543
+ hidden_states = self.norm(hidden_states)
544
+
545
+ # add hidden states from the last decoder layer
546
+ if output_hidden_states:
547
+ all_hidden_states += (hidden_states,)
548
+
549
+ next_cache = next_decoder_cache if use_cache else None
550
+ if not return_dict:
551
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
552
+ return BaseModelOutputWithPast(
553
+ last_hidden_state=hidden_states,
554
+ past_key_values=next_cache,
555
+ hidden_states=all_hidden_states,
556
+ attentions=all_self_attns,
557
+ )
558
+
559
+ def get_input_embeddings(self):
560
+ return self.embed_tokens
561
+
562
+ def set_input_embeddings(self, value):
563
+ self.embed_tokens = value
564
+
565
+ # noinspection PyMethodMayBeStatic
566
+ # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
567
+ def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
568
+ # create causal mask
569
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
570
+ combined_attention_mask = None
571
+ if input_shape[-1] > 1:
572
+ combined_attention_mask = _make_causal_mask(
573
+ input_shape,
574
+ inputs_embeds.dtype,
575
+ device=inputs_embeds.device,
576
+ past_key_values_length=past_key_values_length,
577
+ )
578
+
579
+ if attention_mask is not None:
580
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
581
+ expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
582
+ inputs_embeds.device
583
+ )
584
+ combined_attention_mask = (
585
+ expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
586
+ )
587
+
588
+ return combined_attention_mask
589
+
590
+
591
+ def _history_to_prompt(signal_type, history, query):
592
+ if signal_type == 'base':
593
+ return query
594
+ elif signal_type == 'vqa':
595
+ answer_format = 'Short answer:'
596
+ elif signal_type == 'chat':
597
+ answer_format = 'Answer:'
598
+ else:
599
+ assert False, f"Unknown signal type {signal_type}"
600
+
601
+ prompt = ''
602
+ for i, (old_query, response) in enumerate(history):
603
+ prompt += 'Question: ' + old_query + " {} ".format(answer_format) + response + "\n"
604
+ prompt += 'Question: {} {}'.format(query, answer_format)
605
+ return prompt
606
+
607
+
608
+ class CogVLMForCausalLM(CogVLMPreTrainedModel):
609
+ _auto_class = "AutoModelForCausalLM"
610
+
611
+ def __init__(self, config):
612
+ super().__init__(config)
613
+ self.model = CogVLMModel(config)
614
+ self.vocab_size = config.vocab_size
615
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
616
+
617
+ # Initialize weights and apply final processing
618
+ self.post_init()
619
+
620
+ def get_input_embeddings(self):
621
+ return self.model.embed_tokens
622
+
623
+ def set_input_embeddings(self, value):
624
+ self.model.embed_tokens = value
625
+
626
+ def get_output_embeddings(self):
627
+ return self.lm_head
628
+
629
+ def set_output_embeddings(self, new_embeddings):
630
+ self.lm_head = new_embeddings
631
+
632
+ def set_decoder(self, decoder):
633
+ self.model = decoder
634
+
635
+ def get_decoder(self):
636
+ return self.model
637
+
638
+ def forward(
639
+ self,
640
+ input_ids: torch.LongTensor = None,
641
+ images: List[List[torch.Tensor]] = None,
642
+ token_type_ids: Optional[torch.LongTensor] = None,
643
+ attention_mask: Optional[torch.Tensor] = None,
644
+ position_ids: Optional[torch.LongTensor] = None,
645
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
646
+ inputs_embeds: Optional[torch.FloatTensor] = None,
647
+ use_cache: Optional[bool] = None,
648
+ output_attentions: Optional[bool] = None,
649
+ output_hidden_states: Optional[bool] = None,
650
+ return_dict: Optional[bool] = None,
651
+ labels: Optional[torch.LongTensor] = None,
652
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
653
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
654
+ output_hidden_states = (
655
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
656
+ )
657
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
658
+
659
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
660
+ outputs = self.model(
661
+ input_ids=input_ids,
662
+ images=images,
663
+ token_type_ids=token_type_ids,
664
+ attention_mask=attention_mask,
665
+ position_ids=position_ids,
666
+ past_key_values=past_key_values,
667
+ inputs_embeds=inputs_embeds,
668
+ use_cache=use_cache,
669
+ output_attentions=output_attentions,
670
+ output_hidden_states=output_hidden_states,
671
+ return_dict=return_dict,
672
+ )
673
+
674
+ hidden_states = outputs[0]
675
+ logits = self.lm_head(hidden_states)
676
+ logits = logits.float()
677
+
678
+ loss = None
679
+ if labels is not None:
680
+ # Shift so that tokens < n predict n
681
+ shift_logits = logits[..., :-1, :].contiguous()
682
+ shift_labels = labels[..., 1:].contiguous()
683
+ # Flatten the tokens
684
+ loss_fct = CrossEntropyLoss()
685
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
686
+ shift_labels = shift_labels.view(-1)
687
+ # Enable model parallelism
688
+ shift_labels = shift_labels.to(shift_logits.device)
689
+ loss = loss_fct(shift_logits, shift_labels)
690
+
691
+ if not return_dict:
692
+ output = (logits,) + outputs[1:]
693
+ return (loss,) + output if loss is not None else output
694
+
695
+ return CausalLMOutputWithPast(
696
+ loss=loss,
697
+ logits=logits,
698
+ past_key_values=outputs.past_key_values,
699
+ hidden_states=outputs.hidden_states,
700
+ attentions=outputs.attentions,
701
+ )
702
+
703
+ def _prepare_attention_mask_for_generation(
704
+ self,
705
+ inputs: torch.Tensor,
706
+ pad_token_id: Optional[int],
707
+ eos_token_id: Optional[Union[int, List[int]]],
708
+ ) -> torch.LongTensor:
709
+ return torch.ones(inputs.shape[:2], dtype=torch.long, device=inputs.device) # type: ignore
710
+
711
+ def prepare_inputs_for_generation(
712
+ self, input_ids, token_type_ids, images=None, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
713
+ ):
714
+ # build position_ids if needed
715
+ position_ids = kwargs.get("position_ids", None)
716
+ if position_ids is None:
717
+ position_ids = build_position_ids(token_type_ids, attention_mask)
718
+
719
+ if past_key_values:
720
+ input_ids = input_ids[:, -1:]
721
+ token_type_ids = token_type_ids[:, -1:]
722
+ position_ids = position_ids[:, -1:]
723
+
724
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
725
+ if inputs_embeds is not None and past_key_values is None:
726
+ model_inputs = {"inputs_embeds": inputs_embeds}
727
+ else:
728
+ model_inputs = {"input_ids": input_ids}
729
+
730
+ model_inputs.update(
731
+ {
732
+ "token_type_ids": token_type_ids,
733
+ "images": images,
734
+ "position_ids": position_ids,
735
+ "past_key_values": past_key_values,
736
+ "use_cache": kwargs.get("use_cache"),
737
+ "attention_mask": attention_mask,
738
+ }
739
+ )
740
+ return model_inputs
741
+
742
+ def _update_model_kwargs_for_generation(
743
+ self,
744
+ outputs: "ModelOutput",
745
+ model_kwargs: Dict[str, Any],
746
+ is_encoder_decoder: bool = False,
747
+ standardize_cache_format: bool = False,
748
+ ) -> Dict[str, Any]:
749
+ # update past_key_values
750
+ model_kwargs["past_key_values"] = self._extract_past_from_model_output(
751
+ outputs, standardize_cache_format=standardize_cache_format
752
+ )
753
+ if getattr(outputs, "state", None) is not None:
754
+ model_kwargs["state"] = outputs.state
755
+
756
+ # update token_type_ids with last value
757
+ if "token_type_ids" in model_kwargs:
758
+ token_type_ids = model_kwargs["token_type_ids"]
759
+ new_token_type_ids = torch.ones(size=(token_type_ids.shape[0], 1), dtype=token_type_ids.dtype,
760
+ device=token_type_ids.device) * LANGUAGE_TOKEN_TYPE
761
+ model_kwargs["token_type_ids"] = torch.cat([token_type_ids, new_token_type_ids], dim=-1)
762
+
763
+ if not is_encoder_decoder:
764
+ # update attention mask
765
+ if "attention_mask" in model_kwargs:
766
+ attention_mask = model_kwargs["attention_mask"]
767
+ model_kwargs["attention_mask"] = torch.cat(
768
+ [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
769
+ )
770
+ else:
771
+ # update decoder attention mask
772
+ if "decoder_attention_mask" in model_kwargs:
773
+ decoder_attention_mask = model_kwargs["decoder_attention_mask"]
774
+ model_kwargs["decoder_attention_mask"] = torch.cat(
775
+ [decoder_attention_mask, decoder_attention_mask.new_ones((decoder_attention_mask.shape[0], 1))],
776
+ dim=-1,
777
+ )
778
+
779
+ return model_kwargs
780
+
781
+ def _reorder_cache(self, past_key_values, beam_idx):
782
+ reordered_past = ()
783
+ for layer_past in past_key_values:
784
+ reordered_past += (
785
+ tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
786
+ )
787
+ return reordered_past
788
+
789
+ def build_conversation_input_ids(
790
+ self,
791
+ tokenizer: "PreTrainedTokenizer",
792
+ *,
793
+ query: str,
794
+ history: Optional[List[Tuple[str, str]]] = None,
795
+ images: Optional[List["PIL.Image"]] = None,
796
+ template_version: Optional[Literal["base", "chat", "vqa"]] = None,
797
+ ):
798
+ image_size: int = self.config.vision_config['image_size']
799
+ patch_size: int = self.config.vision_config['patch_size']
800
+ template_version = template_version or self.config.template_version
801
+ assert images is None or len(images) <= 1, f"not support multi images by now."
802
+ history = history or []
803
+ text = _history_to_prompt(template_version, history, query)
804
+
805
+ input_ids = [tokenizer.bos_token_id]
806
+ token_type_ids = [LANGUAGE_TOKEN_TYPE]
807
+ if images is not None and len(images) == 1:
808
+ # vision
809
+ transform = transforms.Compose(
810
+ [
811
+ transforms.Resize(
812
+ (image_size, image_size), interpolation=transforms.InterpolationMode.BICUBIC
813
+ ),
814
+ transforms.ToTensor(),
815
+ transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
816
+ ]
817
+ )
818
+ images = [transform(images[0])]
819
+ # language
820
+ vision_token_num = (image_size // patch_size) * (image_size // patch_size) + 2
821
+ input_ids += [tokenizer.pad_token_id] * vision_token_num
822
+ token_type_ids += [VISION_TOKEN_TYPE] * vision_token_num
823
+ text_ids = tokenizer.encode(text, add_special_tokens=False)
824
+
825
+ input_ids += text_ids
826
+ token_type_ids += [LANGUAGE_TOKEN_TYPE] * len(text_ids)
827
+ attention_mask = [1] * len(input_ids)
828
+
829
+ return {
830
+ 'input_ids': torch.tensor(input_ids, dtype=torch.long),
831
+ 'token_type_ids': torch.tensor(token_type_ids, dtype=torch.long),
832
+ 'attention_mask': torch.tensor(attention_mask, dtype=torch.long),
833
+ 'images': images,
834
+ }
visual.py ADDED
@@ -0,0 +1,135 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch import nn
3
+ from argparse import Namespace
4
+ import xformers.ops as xops
5
+ from transformers.activations import ACT2FN
6
+
7
+
8
+ class PatchEmbedding(nn.Module):
9
+ def __init__(self, config):
10
+ super().__init__()
11
+ self.proj = nn.Conv2d(config.in_channels, config.hidden_size, kernel_size=config.patch_size, stride=config.patch_size)
12
+ self.cls_embedding = nn.Parameter(torch.zeros(1, config.hidden_size))
13
+ self.position_embedding = nn.Embedding(config.num_positions, config.hidden_size)
14
+
15
+ def forward(self, images: "tensor(B, C, H, W)") -> "tensor(B, L, D)":
16
+ x = self.proj(images)
17
+ x = x.flatten(2).transpose(1, 2)
18
+ cls_token = self.cls_embedding.expand(x.shape[0], -1, -1)
19
+ x = torch.cat((cls_token, x), dim=1)
20
+ x += self.position_embedding.weight.unsqueeze(0)
21
+ return x
22
+
23
+
24
+ class Attention(nn.Module):
25
+ def __init__(self, config):
26
+ super().__init__()
27
+ self.num_heads = config.num_heads
28
+ head_dim = config.hidden_size // config.num_heads
29
+ self.scale = head_dim ** -0.5
30
+ self.query_key_value = nn.Linear(config.hidden_size, config.hidden_size * 3)
31
+ self.dense = nn.Linear(config.hidden_size, config.hidden_size)
32
+ self.output_dropout = torch.nn.Dropout(config.dropout_prob)
33
+
34
+ def forward(self, x: "tensor(B, L, D)") -> "tensor(B, L, D)":
35
+ B, L, _ = x.shape
36
+ qkv = self.query_key_value(x)
37
+ qkv = qkv.reshape(B, L, 3, self.num_heads, -1).permute(2, 0, 1, 3, 4) # 3, B, L, H, D
38
+ q, k, v = qkv[0], qkv[1], qkv[2]
39
+
40
+ out = xops.memory_efficient_attention(
41
+ q, k, v, scale=self.scale,
42
+ )
43
+ output = self.dense(out.view(B, L, -1))
44
+ output = self.output_dropout(output)
45
+ return output
46
+
47
+ def attention(self, q, k, v):
48
+ attn_weights = torch.matmul(q * self.scale, k.transpose(-2, -1))
49
+ attn_weights = attn_weights.softmax(dim=-1)
50
+ output = torch.matmul(attn_weights, v)
51
+ return output
52
+
53
+
54
+ class MLP(nn.Module):
55
+ def __init__(self, config):
56
+ super().__init__()
57
+ self.config = config
58
+ self.activation_fn = ACT2FN[config.hidden_act]
59
+ self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
60
+ self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
61
+
62
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
63
+ x = self.fc1(x)
64
+ x = self.activation_fn(x)
65
+ x = self.fc2(x)
66
+ return x
67
+
68
+
69
+ class TransformerLayer(nn.Module):
70
+ def __init__(self, config):
71
+ super().__init__()
72
+ self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
73
+ self.attention = Attention(config)
74
+ self.mlp = MLP(config)
75
+ self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
76
+
77
+ def forward(self, hidden_states):
78
+ attention_input = hidden_states
79
+ attention_output = self.input_layernorm(self.attention(attention_input))
80
+ hidden_states = attention_input + attention_output
81
+ mlp_input = hidden_states
82
+ mlp_output = self.post_attention_layernorm(self.mlp(mlp_input))
83
+ output = mlp_input + mlp_output
84
+ return output
85
+
86
+
87
+ class Transformer(nn.Module):
88
+ def __init__(self, config):
89
+ super().__init__()
90
+ self.layers = nn.ModuleList([TransformerLayer(config) for _ in range(config.num_hidden_layers)])
91
+
92
+ def forward(self, hidden_states):
93
+ for layer_module in self.layers:
94
+ hidden_states = layer_module(hidden_states)
95
+ return hidden_states
96
+
97
+
98
+ class GLU(nn.Module):
99
+ def __init__(self, config, in_features):
100
+ super().__init__()
101
+ self.linear_proj = nn.Linear(in_features, config.hidden_size, bias=False)
102
+ self.norm1 = nn.LayerNorm(config.hidden_size)
103
+ self.act1 = nn.GELU()
104
+ self.act2 = nn.functional.silu
105
+ self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
106
+ self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
107
+ self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
108
+
109
+ def forward(self, x):
110
+ x = self.linear_proj(x)
111
+ x = self.act1(self.norm1(x))
112
+ x = self.act2(self.gate_proj(x)) * self.dense_h_to_4h(x)
113
+ x = self.dense_4h_to_h(x)
114
+ return x
115
+
116
+
117
+ class EVA2CLIPModel(nn.Module):
118
+ def __init__(self, config):
119
+ super().__init__()
120
+ vision_config = Namespace(**config.vision_config)
121
+ self.patch_embedding = PatchEmbedding(vision_config)
122
+ self.transformer = Transformer(vision_config)
123
+ self.linear_proj = GLU(config, in_features=vision_config.hidden_size)
124
+ self.boi = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
125
+ self.eoi = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
126
+
127
+ def forward(self, images: "tensor(B, C, H, W)") -> "tensor(B, L, D)":
128
+ x = self.patch_embedding(images)
129
+ x = self.transformer(x)
130
+ x = x[:, 1:]
131
+ x = self.linear_proj(x)
132
+ boi = self.boi.expand(x.shape[0], -1, -1)
133
+ eoi = self.eoi.expand(x.shape[0], -1, -1)
134
+ x = torch.cat((boi, x, eoi), dim=1)
135
+ return x