xiaoyewuz-Ruster commited on
Commit
a1b1f8a
·
verified ·
1 Parent(s): 09a13c5

Upload zzjrabbit22.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. zzjrabbit22.py +250 -0
zzjrabbit22.py ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import os.path
3
+ from typing import Optional, Union
4
+
5
+ import torch
6
+ import torch.nn as nn
7
+ from tokenizers import Tokenizer
8
+ from transformers import (
9
+ GenerationMixin,
10
+ PretrainedConfig,
11
+ PreTrainedModel,
12
+ PreTrainedTokenizer,
13
+ )
14
+ from transformers.modeling_outputs import BaseModelOutput, CausalLMOutput
15
+
16
+
17
+ class ZZJRabbit22Config(PretrainedConfig):
18
+ model_type = "zzjrabbit22"
19
+
20
+ def __init__(
21
+ self,
22
+ num_layers: int = 12,
23
+ num_attention_heads: int = 8,
24
+ vocab_size: int = 10000,
25
+ hidden_size: int = 1024,
26
+ **kwargs,
27
+ ):
28
+ self.num_layers = num_layers
29
+ self.num_attention_heads = num_attention_heads
30
+ self.vocab_size = vocab_size
31
+ self.hidden_size = hidden_size
32
+ assert hidden_size % num_attention_heads == 0
33
+ super().__init__(**kwargs)
34
+
35
+
36
+ class ZZJRabbit22PE(nn.Module):
37
+ def __init__(self, hidden_size: int, max_len: int = 32768):
38
+ super().__init__()
39
+ pe = torch.zeros(max_len, hidden_size)
40
+ position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
41
+ div_term = torch.exp(
42
+ torch.arange(0, hidden_size, 2).float() * (-math.log(10000.0) / hidden_size)
43
+ )
44
+ pe[:, 0::2] = torch.sin(position * div_term)
45
+ pe[:, 1::2] = torch.cos(position * div_term)
46
+ pe = pe.unsqueeze(0).transpose(0, 1)
47
+ self.register_buffer("pe", pe)
48
+
49
+ def forward(self, x: torch.Tensor):
50
+ return x + self.pe[: x.size(0), :]
51
+
52
+
53
+ class ZZJRabbit22Attention(nn.Module):
54
+ def __init__(self, config: ZZJRabbit22Config):
55
+ super().__init__()
56
+ self.config = config
57
+ self.head_dim = config.hidden_size // config.num_attention_heads
58
+ self.q_proj = nn.Linear(config.hidden_size, config.hidden_size)
59
+ self.k_proj = nn.Linear(config.hidden_size, config.hidden_size)
60
+ self.v_proj = nn.Linear(config.hidden_size, config.hidden_size)
61
+ self.out_proj = nn.Linear(config.hidden_size, config.hidden_size)
62
+ self.dropout = nn.Dropout(0.1)
63
+
64
+ def forward(
65
+ self,
66
+ x: torch.Tensor,
67
+ key_padding_mask: Optional[torch.BoolTensor] = None,
68
+ attn_mask: Optional[torch.BoolTensor] = None,
69
+ ):
70
+ batch_size = x.size(0)
71
+ Q = (
72
+ self.q_proj(x)
73
+ .view(batch_size, -1, self.config.num_attention_heads, self.head_dim)
74
+ .transpose(1, 2)
75
+ )
76
+ K = (
77
+ self.k_proj(x)
78
+ .view(batch_size, -1, self.config.num_attention_heads, self.head_dim)
79
+ .transpose(1, 2)
80
+ )
81
+ V = (
82
+ self.v_proj(x)
83
+ .view(batch_size, -1, self.config.num_attention_heads, self.head_dim)
84
+ .transpose(1, 2)
85
+ )
86
+ scores = torch.matmul(Q, K.transpose(-2, -1)) / torch.sqrt(
87
+ torch.tensor(self.head_dim, dtype=torch.float32)
88
+ )
89
+ if key_padding_mask is not None:
90
+ scores = scores.masked_fill(
91
+ key_padding_mask.view(batch_size, 1, 1, -1), float("-inf")
92
+ )
93
+ if attn_mask is not None:
94
+ scores = scores.masked_fill(attn_mask, float("-inf"))
95
+ attn_weights = nn.functional.softmax(scores, dim=-1)
96
+ attn_weights = self.dropout(attn_weights)
97
+ context = torch.matmul(attn_weights, V)
98
+ context = context.transpose(1, 2).contiguous()
99
+ context = context.view(batch_size, -1, self.config.hidden_size)
100
+ return self.out_proj(context)
101
+
102
+
103
+ class ZZJRabbit22Layer(nn.Module):
104
+ def __init__(self, config: ZZJRabbit22Config):
105
+ super().__init__()
106
+ self.attn = ZZJRabbit22Attention(config)
107
+ self.l1 = nn.Linear(config.hidden_size, config.hidden_size)
108
+ self.l2 = nn.Linear(config.hidden_size, config.hidden_size)
109
+ self.activate = nn.ReLU()
110
+ self.norm = nn.RMSNorm(config.hidden_size)
111
+
112
+ def forward(
113
+ self, x: torch.Tensor, attention_mask: Optional[torch.Tensor] = None
114
+ ) -> torch.Tensor:
115
+ key_padding_mask = None
116
+ attn_mask = None
117
+ if self.training:
118
+ attn_mask = torch.gt(
119
+ torch.triu(torch.ones(x.size(-2), x.size(-2), device=x.device), 1), 0
120
+ )
121
+ if attention_mask is not None:
122
+ key_padding_mask = torch.lt(attention_mask, 1)
123
+ attn = self.attn(
124
+ x,
125
+ key_padding_mask=key_padding_mask,
126
+ attn_mask=attn_mask,
127
+ )[0]
128
+ x = self.norm(x + attn)
129
+ o = self.l1(x)
130
+ o = self.activate(o)
131
+ o = self.l2(o)
132
+ return self.norm(x + o)
133
+
134
+
135
+ class ZZJRabbit22Model(PreTrainedModel):
136
+ config_class = ZZJRabbit22Config
137
+
138
+ def __init__(self, config: ZZJRabbit22Config, **kwargs):
139
+ super().__init__(config, **kwargs)
140
+ self.config = config
141
+ self.emb = nn.Embedding(config.vocab_size, config.hidden_size)
142
+ self.pe = ZZJRabbit22PE(config.hidden_size)
143
+ self.layers = nn.ModuleList(
144
+ [ZZJRabbit22Layer(config) for _ in range(config.num_layers)]
145
+ )
146
+
147
+ def forward(
148
+ self,
149
+ input_ids: torch.Tensor,
150
+ return_dict: Optional[bool] = None,
151
+ attention_mask: Optional[torch.Tensor] = None,
152
+ **kwargs,
153
+ ):
154
+ res = self.emb(input_ids)
155
+ res = self.pe(res)
156
+ res = res
157
+ for l in self.layers:
158
+ res = l(res, attention_mask)
159
+ if not return_dict:
160
+ return (res,)
161
+ else:
162
+ return BaseModelOutput(res)
163
+
164
+
165
+ class ZZJRabbit22ForCausalLM(PreTrainedModel, GenerationMixin):
166
+ config_class = ZZJRabbit22Config
167
+
168
+ def __init__(self, config, **kwargs):
169
+ super().__init__(config, **kwargs)
170
+ self.model = ZZJRabbit22Model(config, **kwargs)
171
+ self.l = nn.Linear(config.hidden_size, config.vocab_size)
172
+
173
+ def forward(
174
+ self,
175
+ input_ids: torch.Tensor,
176
+ return_dict: Optional[bool] = None,
177
+ labels: Optional[torch.Tensor] = None,
178
+ attention_mask: Optional[torch.Tensor] = None,
179
+ logits_to_keep: Union[int, torch.Tensor] = 0,
180
+ **kwargs,
181
+ ):
182
+ # print(input_ids, return_dict, labels, attention_mask, logits_to_keep, kwargs)
183
+ hidden = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
184
+ logits = self.l(
185
+ hidden[
186
+ :,
187
+ slice(-logits_to_keep, None)
188
+ if isinstance(logits_to_keep, int)
189
+ else logits_to_keep,
190
+ :,
191
+ ]
192
+ )
193
+ if labels is not None:
194
+ loss = self.loss_function(
195
+ logits=logits,
196
+ labels=labels,
197
+ vocab_size=self.config.vocab_size,
198
+ **kwargs,
199
+ )
200
+ print(loss)
201
+ if not return_dict:
202
+ return (loss, logits) if labels is not None else (logits,)
203
+ else:
204
+ return (
205
+ CausalLMOutput(logits=logits, loss=loss)
206
+ if labels is not None
207
+ else CausalLMOutput(logits=logits)
208
+ )
209
+
210
+ @classmethod
211
+ def can_generate(cls):
212
+ return True
213
+
214
+ def prepare_inputs_for_generation(self, input_ids, **kwargs):
215
+ return {"input_ids": input_ids}
216
+
217
+
218
+ class ZZJRabbit22Tokenizer(PreTrainedTokenizer):
219
+ vocab_files_names = {"tokenizers_file": "tokenizer.json"}
220
+
221
+ def __init__(self, tokenizers_file, **kwargs):
222
+ self.internal = Tokenizer.from_file(tokenizers_file)
223
+ super().__init__(**kwargs)
224
+
225
+ def get_vocab(self):
226
+ return {self.internal.id_to_token(i): i for i in range(self.vocab_size)}
227
+
228
+ def tokenize(self, text, **kwargs):
229
+ return self.internal.encode(text).tokens
230
+
231
+ def convert_tokens_to_ids(self, tokens):
232
+ return (
233
+ self.internal.token_to_id(tokens)
234
+ if isinstance(tokens, str)
235
+ else [self.internal.token_to_id(t) for t in tokens]
236
+ )
237
+
238
+ def decode(self, tokens, skip_special_tokens=True, **kwargs):
239
+ if isinstance(tokens, torch.Tensor):
240
+ tokens = tokens.tolist()
241
+ return self.internal.decode(tokens, skip_special_tokens=skip_special_tokens)
242
+
243
+ @property
244
+ def vocab_size(self):
245
+ return self.internal.get_vocab_size()
246
+
247
+ def save_vocabulary(self, path, *args, **kwargs) -> tuple[str]:
248
+ p = os.path.join(path, "tokenizer.json")
249
+ self.internal.save(p)
250
+ return (p,)