yourui commited on
Commit
067ff11
1 Parent(s): f3742fe

first commit

Browse files
README.md ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ tags:
3
+ - generated_from_trainer
4
+ datasets:
5
+ - /data/lush-dev/zhangyoujin/models/Data/PromptCBLUE/bgiv3
6
+ model-index:
7
+ - name: baichuan_finetune_v3_071618
8
+ results: []
9
+ ---
10
+
11
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
12
+ should probably proofread and complete it, then remove this comment. -->
13
+
14
+ # baichuan_finetune_v3_071618
15
+
16
+ This model is a fine-tuned version of [/data/lush-dev/zhangyoujin/models/Transformers/baichuan-13B/](https://huggingface.co//data/lush-dev/zhangyoujin/models/Transformers/baichuan-13B/) on the /data/lush-dev/zhangyoujin/models/Data/PromptCBLUE/bgiv3 dataset.
17
+
18
+ ## Model description
19
+
20
+ More information needed
21
+
22
+ ## Intended uses & limitations
23
+
24
+ More information needed
25
+
26
+ ## Training and evaluation data
27
+
28
+ More information needed
29
+
30
+ ## Training procedure
31
+
32
+ ### Training hyperparameters
33
+
34
+ The following hyperparameters were used during training:
35
+ - learning_rate: 2e-05
36
+ - train_batch_size: 1
37
+ - eval_batch_size: 8
38
+ - seed: 42
39
+ - distributed_type: multi-GPU
40
+ - num_devices: 8
41
+ - total_train_batch_size: 8
42
+ - total_eval_batch_size: 64
43
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
44
+ - lr_scheduler_type: cosine
45
+ - num_epochs: 2.0
46
+
47
+ ### Training results
48
+
49
+
50
+
51
+ ### Framework versions
52
+
53
+ - Transformers 4.30.2
54
+ - Pytorch 2.0.1+cu118
55
+ - Datasets 2.12.0
56
+ - Tokenizers 0.13.3
config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "_name_or_path": "/data/lush-dev/zhangyoujin/models/Transformers/baichuan-13B/",
4
+ "architectures": [
5
+ "BaichuanForCausalLM"
6
+ ],
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_baichuan.BaichuanConfig",
9
+ "AutoModelForCausalLM": "modeling_baichuan.BaichuanForCausalLM"
10
+ },
11
+ "bos_token_id": 1,
12
+ "eos_token_id": 2,
13
+ "hidden_act": "silu",
14
+ "hidden_size": 5120,
15
+ "initializer_range": 0.02,
16
+ "intermediate_size": 13696,
17
+ "model_max_length": 4096,
18
+ "model_type": "baichuan",
19
+ "num_attention_heads": 40,
20
+ "num_hidden_layers": 40,
21
+ "pad_token_id": 0,
22
+ "rms_norm_eps": 1e-06,
23
+ "tie_word_embeddings": false,
24
+ "torch_dtype": "bfloat16",
25
+ "transformers_version": "4.30.2",
26
+ "use_cache": true,
27
+ "vocab_size": 64000
28
+ }
configuration_baichuan.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
2
+
3
+ from transformers.configuration_utils import PretrainedConfig
4
+
5
+ class BaichuanConfig(PretrainedConfig):
6
+ model_type = "baichuan"
7
+ keys_to_ignore_at_inference = ["past_key_values"]
8
+
9
+ def __init__(
10
+ self,
11
+ vocab_size=64000,
12
+ hidden_size=5120,
13
+ intermediate_size=13696,
14
+ num_hidden_layers=40,
15
+ num_attention_heads=40,
16
+ hidden_act="silu",
17
+ model_max_length=4096,
18
+ initializer_range=0.02,
19
+ rms_norm_eps=1e-6,
20
+ use_cache=True,
21
+ pad_token_id=0,
22
+ bos_token_id=1,
23
+ eos_token_id=2,
24
+ tie_word_embeddings=False,
25
+ gradient_checkpointing=False,
26
+ **kwargs,
27
+ ):
28
+ self.vocab_size = vocab_size
29
+ self.model_max_length = model_max_length
30
+ self.hidden_size = hidden_size
31
+ self.intermediate_size = intermediate_size
32
+ self.num_hidden_layers = num_hidden_layers
33
+ self.num_attention_heads = num_attention_heads
34
+ self.hidden_act = hidden_act
35
+ self.initializer_range = initializer_range
36
+ self.rms_norm_eps = rms_norm_eps
37
+ self.use_cache = use_cache
38
+ self.gradient_checkpointing = gradient_checkpointing,
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
+ )
46
+
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": "4.30.2"
7
+ }
modeling_baichuan.py ADDED
@@ -0,0 +1,536 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
2
+
3
+ import math
4
+ from typing import List, Optional, Tuple, Union
5
+
6
+ import torch
7
+ import torch.utils.checkpoint
8
+ from torch.nn import CrossEntropyLoss
9
+ from transformers import PreTrainedModel
10
+ from transformers.activations import ACT2FN
11
+ from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
12
+ from transformers.utils import logging
13
+ from transformers.generation.utils import GenerationConfig
14
+
15
+ from .configuration_baichuan import BaichuanConfig
16
+
17
+ logger = logging.get_logger(__name__)
18
+
19
+ def _get_interleave(n):
20
+ def _get_interleave_power_of_2(n):
21
+ start = (2 ** (-2 ** -(math.log2(n) - 3)))
22
+ ratio = start
23
+ return [start * ratio ** i for i in range(n)]
24
+
25
+ if math.log2(n).is_integer():
26
+ return _get_interleave_power_of_2(n)
27
+ else:
28
+ closest_power_of_2 = 2 ** math.floor(math.log2(n))
29
+ return _get_interleave_power_of_2(closest_power_of_2) + \
30
+ _get_interleave(2 * closest_power_of_2)[0::2][:n - closest_power_of_2]
31
+
32
+ def _fill_with_neg_inf(t):
33
+ """FP16-compatible function that fills a tensor with -inf."""
34
+ return t.float().fill_(float("-inf")).type_as(t)
35
+
36
+ def _gen_alibi_mask(n_head, max_pos):
37
+ slopes = torch.Tensor(_get_interleave(n_head))
38
+ alibi = slopes.unsqueeze(1).unsqueeze(1) * torch.arange(max_pos).unsqueeze(0).unsqueeze(0).expand(
39
+ n_head, -1, -1)
40
+ alibi = alibi.view(n_head, 1, max_pos)
41
+ alibi_mask = torch.triu(
42
+ _fill_with_neg_inf(torch.zeros([max_pos, max_pos])), 1
43
+ )
44
+ alibi_mask = alibi_mask.unsqueeze(0) + alibi
45
+ return alibi_mask
46
+
47
+
48
+ class RMSNorm(torch.nn.Module):
49
+ def __init__(self, hidden_size, epsilon=1e-6):
50
+ super().__init__()
51
+ self.weight = torch.nn.Parameter(torch.empty(hidden_size))
52
+ self.epsilon = epsilon
53
+
54
+ def forward(self, hidden_states):
55
+ variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
56
+ hidden_states = hidden_states * torch.rsqrt(variance + self.epsilon)
57
+
58
+ # convert into half-precision
59
+ if self.weight.dtype in [torch.float16, torch.bfloat16]:
60
+ hidden_states = hidden_states.to(self.weight.dtype)
61
+
62
+ return self.weight * hidden_states
63
+
64
+
65
+ class MLP(torch.nn.Module):
66
+ def __init__(
67
+ self,
68
+ hidden_size: int,
69
+ intermediate_size: int,
70
+ hidden_act: str,
71
+ ):
72
+ super().__init__()
73
+ self.gate_proj = torch.nn.Linear(hidden_size, intermediate_size, bias=False)
74
+ self.down_proj = torch.nn.Linear(intermediate_size, hidden_size, bias=False)
75
+ self.up_proj = torch.nn.Linear(hidden_size, intermediate_size, bias=False)
76
+ self.act_fn = ACT2FN[hidden_act]
77
+
78
+ def forward(self, x):
79
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
80
+
81
+
82
+ class BaichuanAttention(torch.nn.Module):
83
+
84
+ def __init__(self, config: BaichuanConfig):
85
+ super().__init__()
86
+ self.config = config
87
+ self.hidden_size = config.hidden_size
88
+ self.num_heads = config.num_attention_heads
89
+ self.head_dim = self.hidden_size // self.num_heads
90
+ self.max_position_embeddings = config.model_max_length
91
+
92
+ if (self.head_dim * self.num_heads) != self.hidden_size:
93
+ raise ValueError(
94
+ f"hidden_size {self.hidden_size} is not divisible by num_heads {self.num_heads}"
95
+ )
96
+ self.W_pack = torch.nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=False)
97
+ self.o_proj = torch.nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
98
+
99
+ def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
100
+ return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
101
+
102
+ def forward(
103
+ self,
104
+ hidden_states: torch.Tensor,
105
+ attention_mask: Optional[torch.Tensor] = None,
106
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
107
+ output_attentions: bool = False,
108
+ use_cache: bool = False,
109
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
110
+
111
+ bsz, q_len, _ = hidden_states.size()
112
+
113
+ proj = self.W_pack(hidden_states)
114
+ proj = proj.unflatten(-1, (3, self.hidden_size)).unsqueeze(0).transpose(0, -2).squeeze(-2)
115
+ query_states = proj[0].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
116
+ key_states = proj[1].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
117
+ value_states = proj[2].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
118
+
119
+ kv_seq_len = key_states.shape[-2]
120
+ if past_key_value is not None:
121
+ kv_seq_len += past_key_value[0].shape[-2]
122
+
123
+ if past_key_value is not None:
124
+ # reuse k, v, self_attention
125
+ key_states = torch.cat([past_key_value[0], key_states], dim=2)
126
+ value_states = torch.cat([past_key_value[1], value_states], dim=2)
127
+
128
+ past_key_value = (key_states, value_states) if use_cache else None
129
+
130
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
131
+
132
+ if attention_mask is not None:
133
+ if attn_weights.size(-2) == 1:
134
+ attention_mask = attention_mask[:, -1:, :]
135
+ attn_weights = attn_weights + attention_mask.unsqueeze(0)
136
+ attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
137
+
138
+ attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1)
139
+ attn_output = torch.matmul(attn_weights, value_states)
140
+
141
+ attn_output = attn_output.transpose(1, 2)
142
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
143
+ attn_output = self.o_proj(attn_output)
144
+
145
+ if not output_attentions:
146
+ attn_weights = None
147
+
148
+ return attn_output, attn_weights, past_key_value
149
+
150
+
151
+ class BaichuanLayer(torch.nn.Module):
152
+ def __init__(self, config: BaichuanConfig):
153
+ super().__init__()
154
+ self.hidden_size = config.hidden_size
155
+ self.self_attn = BaichuanAttention(config=config)
156
+ self.mlp = MLP(
157
+ hidden_size=self.hidden_size,
158
+ intermediate_size=config.intermediate_size,
159
+ hidden_act=config.hidden_act,
160
+ )
161
+ self.input_layernorm = RMSNorm(config.hidden_size, epsilon=config.rms_norm_eps)
162
+ self.post_attention_layernorm = RMSNorm(config.hidden_size, epsilon=config.rms_norm_eps)
163
+
164
+ def forward(
165
+ self,
166
+ hidden_states: torch.Tensor,
167
+ attention_mask: Optional[torch.Tensor] = None,
168
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
169
+ output_attentions: Optional[bool] = False,
170
+ use_cache: Optional[bool] = False,
171
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
172
+
173
+ residual = hidden_states
174
+
175
+ hidden_states = self.input_layernorm(hidden_states)
176
+
177
+ # Self Attention
178
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
179
+ hidden_states=hidden_states,
180
+ attention_mask=attention_mask,
181
+ past_key_value=past_key_value,
182
+ output_attentions=output_attentions,
183
+ use_cache=use_cache,
184
+ )
185
+ hidden_states = residual + hidden_states
186
+
187
+ # Fully Connected
188
+ residual = hidden_states
189
+ hidden_states = self.post_attention_layernorm(hidden_states)
190
+ hidden_states = self.mlp(hidden_states)
191
+ hidden_states = residual + hidden_states
192
+
193
+ outputs = (hidden_states,)
194
+
195
+ if use_cache:
196
+ outputs += (present_key_value,)
197
+
198
+ return outputs
199
+
200
+
201
+ class BaichuanPreTrainedModel(PreTrainedModel):
202
+ config_class = BaichuanConfig
203
+ base_model_prefix = "model"
204
+ supports_gradient_checkpointing = True
205
+ _no_split_modules = ["BaichuanLayer"]
206
+ _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
207
+
208
+ def _init_weights(self, module):
209
+ std = self.config.initializer_range
210
+ if isinstance(module, torch.nn.Linear):
211
+ module.weight.data.normal_(mean=0.0, std=std)
212
+ if module.bias is not None:
213
+ module.bias.data.zero_()
214
+ elif isinstance(module, torch.nn.Embedding):
215
+ module.weight.data.normal_(mean=0.0, std=std)
216
+ if module.padding_idx is not None:
217
+ module.weight.data[module.padding_idx].zero_()
218
+
219
+ def _set_gradient_checkpointing(self, module, value=False):
220
+ if isinstance(module, BaichuanModel):
221
+ module.gradient_checkpointing = value
222
+
223
+
224
+
225
+ class BaichuanModel(BaichuanPreTrainedModel):
226
+ def __init__(self, config: BaichuanConfig):
227
+ super().__init__(config)
228
+ self.padding_idx = config.pad_token_id
229
+ self.vocab_size = config.vocab_size
230
+ self.n_head = config.num_attention_heads
231
+ self.embed_tokens = torch.nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
232
+ self.layers = torch.nn.ModuleList([BaichuanLayer(config) for _ in range(config.num_hidden_layers)])
233
+ self.norm = RMSNorm(config.hidden_size, epsilon=config.rms_norm_eps)
234
+
235
+ self.gradient_checkpointing = config.gradient_checkpointing
236
+ self.post_init()
237
+ self.max_cache_pos = config.model_max_length
238
+ self.first_run = True
239
+
240
+ def get_input_embeddings(self):
241
+ return self.embed_tokens
242
+
243
+ def set_input_embeddings(self, value):
244
+ self.embed_tokens = value
245
+
246
+ def get_alibi_mask(self, tensor, seq_length_with_past):
247
+ if self.first_run:
248
+ self.first_run = False
249
+ self.register_buffer("future_mask", _gen_alibi_mask(self.n_head, self.max_cache_pos).to(tensor), persistent=False)
250
+ if seq_length_with_past > self.max_cache_pos:
251
+ self.max_cache_pos = seq_length_with_past
252
+ self.register_buffer("future_mask", _gen_alibi_mask(self.n_head, self.max_cache_pos).to(tensor), persistent=False)
253
+ mask = self.future_mask[:self.n_head, :seq_length_with_past, :seq_length_with_past]
254
+ return mask
255
+
256
+ def forward(
257
+ self,
258
+ input_ids: torch.LongTensor = None,
259
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
260
+ inputs_embeds: Optional[torch.FloatTensor] = None,
261
+ use_cache: Optional[bool] = False,
262
+ output_attentions: Optional[bool] = False,
263
+ output_hidden_states: Optional[bool] = False,
264
+ return_dict: Optional[bool] = True,
265
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
266
+
267
+
268
+ if input_ids is not None and inputs_embeds is not None:
269
+ raise ValueError("You cannot provide both input_ids and inputs_embeds simultaneously")
270
+ elif input_ids is not None:
271
+ batch_size, seq_length = input_ids.shape
272
+ elif inputs_embeds is not None:
273
+ batch_size, seq_length, _ = inputs_embeds.shape
274
+ else:
275
+ raise ValueError("You need to provide input_ids or inputs_embeds")
276
+
277
+ seq_length_with_past = seq_length
278
+
279
+ if past_key_values is not None:
280
+ past_key_values_length = past_key_values[0][0].shape[2]
281
+ seq_length_with_past = seq_length_with_past + past_key_values_length
282
+
283
+ if inputs_embeds is None:
284
+ inputs_embeds = self.embed_tokens(input_ids)
285
+
286
+ # embed positions
287
+ attention_mask = self.get_alibi_mask(inputs_embeds, seq_length_with_past)
288
+
289
+ hidden_states = inputs_embeds
290
+
291
+ if self.gradient_checkpointing and self.training:
292
+ if use_cache:
293
+ logger.warning_once(
294
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
295
+ )
296
+ use_cache = False
297
+
298
+ # decoder layers
299
+ all_hidden_states = () if output_hidden_states else None
300
+ all_self_attns = () if output_attentions else None
301
+ next_decoder_cache = () if use_cache else None
302
+
303
+ for idx, decoder_layer in enumerate(self.layers):
304
+ if output_hidden_states:
305
+ all_hidden_states += (hidden_states,)
306
+
307
+ past_key_value = past_key_values[idx] if past_key_values is not None else None
308
+
309
+ if self.gradient_checkpointing and self.training:
310
+
311
+ def create_custom_forward(module):
312
+ def custom_forward(*inputs):
313
+ # None for past_key_value
314
+ return module(*inputs, output_attentions, None)
315
+
316
+ return custom_forward
317
+
318
+ layer_outputs = torch.utils.checkpoint.checkpoint(
319
+ create_custom_forward(decoder_layer),
320
+ hidden_states,
321
+ attention_mask,
322
+ None,
323
+ )
324
+ else:
325
+ layer_outputs = decoder_layer(
326
+ hidden_states,
327
+ attention_mask=attention_mask,
328
+ past_key_value=past_key_value,
329
+ output_attentions=output_attentions,
330
+ use_cache=use_cache,
331
+ )
332
+
333
+ hidden_states = layer_outputs[0]
334
+
335
+ if use_cache:
336
+ next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
337
+
338
+ if output_attentions:
339
+ all_self_attns += (layer_outputs[1],)
340
+
341
+ hidden_states = self.norm(hidden_states)
342
+
343
+ # add hidden states from the last decoder layer
344
+ if output_hidden_states:
345
+ all_hidden_states += (hidden_states,)
346
+
347
+ next_cache = next_decoder_cache if use_cache else None
348
+ if not return_dict:
349
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
350
+ return BaseModelOutputWithPast(
351
+ last_hidden_state=hidden_states,
352
+ past_key_values=next_cache,
353
+ hidden_states=all_hidden_states,
354
+ attentions=all_self_attns,
355
+ )
356
+
357
+
358
+ class BaichuanForCausalLM(BaichuanPreTrainedModel):
359
+ def __init__(self, config):
360
+ super().__init__(config)
361
+ self.model = BaichuanModel(config)
362
+ self.lm_head = torch.nn.Linear(config.hidden_size, config.vocab_size, bias=False)
363
+
364
+ # Initialize weights and apply final processing
365
+ self.post_init()
366
+
367
+ def forward(
368
+ self,
369
+ input_ids: torch.LongTensor = None,
370
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
371
+ inputs_embeds: Optional[torch.FloatTensor] = None,
372
+ labels: Optional[torch.LongTensor] = None,
373
+ use_cache: Optional[bool] = None,
374
+ output_attentions: Optional[bool] = False,
375
+ output_hidden_states: Optional[bool] = False,
376
+ return_dict: Optional[bool] = True,
377
+ **kwargs
378
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
379
+
380
+
381
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
382
+ outputs = self.model(
383
+ input_ids=input_ids,
384
+ past_key_values=past_key_values,
385
+ inputs_embeds=inputs_embeds,
386
+ use_cache=use_cache,
387
+ output_attentions=output_attentions,
388
+ output_hidden_states=output_hidden_states,
389
+ return_dict=return_dict,
390
+ )
391
+
392
+ hidden_states = outputs[0]
393
+ logits = self.lm_head(hidden_states)
394
+
395
+ loss = None
396
+ if labels is not None:
397
+ # Shift so that tokens < n predict n
398
+ shift_logits = logits[..., :-1, :].contiguous()
399
+ shift_labels = labels[..., 1:].contiguous()
400
+ # Flatten the tokens
401
+ loss_fct = CrossEntropyLoss()
402
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
403
+ shift_labels = shift_labels.view(-1)
404
+ # Enable model parallelism
405
+ shift_labels = shift_labels.to(shift_logits.device)
406
+ loss = loss_fct(shift_logits, shift_labels)
407
+
408
+ if not return_dict:
409
+ output = (logits,) + outputs[1:]
410
+ return (loss,) + output if loss is not None else output
411
+
412
+ return CausalLMOutputWithPast(
413
+ loss=loss,
414
+ logits=logits,
415
+ past_key_values=outputs.past_key_values,
416
+ hidden_states=outputs.hidden_states,
417
+ attentions=outputs.attentions,
418
+ )
419
+
420
+ def prepare_inputs_for_generation(
421
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
422
+ ):
423
+ if past_key_values:
424
+ input_ids = input_ids[:, -1:]
425
+
426
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
427
+ if inputs_embeds is not None and past_key_values is None:
428
+ model_inputs = {"inputs_embeds": inputs_embeds}
429
+ else:
430
+ model_inputs = {"input_ids": input_ids}
431
+
432
+ model_inputs.update(
433
+ {
434
+ "past_key_values": past_key_values,
435
+ "use_cache": kwargs.get("use_cache"),
436
+ }
437
+ )
438
+ return model_inputs
439
+
440
+ @staticmethod
441
+ def _reorder_cache(past_key_values, beam_idx):
442
+ return tuple(
443
+ tuple(past_state.index_select(0, beam_idx) for past_state in layer_past)
444
+ for layer_past in past_key_values
445
+ )
446
+
447
+
448
+ def quantize(self, bits: int):
449
+ try:
450
+ from .quantizer import QLinear
451
+ except ImportError:
452
+ raise ImportError(
453
+ f"Needs QLinear to run quantize."
454
+ )
455
+
456
+ for layer in self.model.layers:
457
+ layer.self_attn.W_pack = QLinear(
458
+ bits=bits,
459
+ weight=layer.self_attn.W_pack.weight,
460
+ bias = None,
461
+ )
462
+ layer.self_attn.o_proj = QLinear(
463
+ bits=bits,
464
+ weight=layer.self_attn.o_proj.weight,
465
+ bias = None,
466
+ )
467
+ layer.mlp.gate_proj = QLinear(
468
+ bits=bits,
469
+ weight=layer.mlp.gate_proj.weight,
470
+ bias = None,
471
+ )
472
+ layer.mlp.down_proj = QLinear(
473
+ bits=bits,
474
+ weight=layer.mlp.down_proj.weight,
475
+ bias = None,
476
+ )
477
+ layer.mlp.up_proj = QLinear(
478
+ bits=bits,
479
+ weight=layer.mlp.up_proj.weight,
480
+ bias = None,
481
+ )
482
+ return self
483
+
484
+ def _build_chat_input(self, tokenizer, messages: List[dict], max_new_tokens: int=0):
485
+ max_new_tokens = max_new_tokens or self.generation_config.max_new_tokens
486
+ max_input_tokens = self.config.model_max_length - max_new_tokens
487
+ max_input_tokens = max(self.config.model_max_length // 2, max_input_tokens)
488
+ total_input, round_input = [], []
489
+ for i, message in enumerate(messages[::-1]):
490
+ content_tokens = tokenizer.encode(message['content'])
491
+ if message['role'] == 'user':
492
+ round_input = [self.generation_config.user_token_id] + content_tokens + round_input
493
+ if total_input and len(total_input) + len(round_input) > max_input_tokens:
494
+ break
495
+ else:
496
+ total_input = round_input + total_input
497
+ if len(total_input) >= max_input_tokens:
498
+ break
499
+ else:
500
+ round_input = []
501
+ elif message['role'] == 'assistant':
502
+ round_input = [
503
+ self.generation_config.assistant_token_id
504
+ ] + content_tokens + [
505
+ self.generation_config.eos_token_id
506
+ ] + round_input
507
+ else:
508
+ raise ValueError(f"message role not supported yet: {message['role']}")
509
+ total_input = total_input[-max_input_tokens:] # truncate left
510
+ total_input.append(self.generation_config.assistant_token_id)
511
+ total_input = torch.LongTensor([total_input]).to(self.device)
512
+ return total_input
513
+
514
+ @torch.no_grad()
515
+ def chat(self, tokenizer, messages: List[dict], stream=False,
516
+ generation_config: Optional[GenerationConfig]=None):
517
+ generation_config = generation_config or self.generation_config
518
+ input_ids = self._build_chat_input(tokenizer, messages, generation_config.max_new_tokens)
519
+ if stream:
520
+ from transformers_stream_generator.main import NewGenerationMixin, StreamGenerationConfig
521
+ self.__class__.generate = NewGenerationMixin.generate
522
+ self.__class__.sample_stream = NewGenerationMixin.sample_stream
523
+ stream_config = StreamGenerationConfig(**generation_config.to_dict(), do_stream=True)
524
+
525
+ def stream_generator():
526
+ outputs = []
527
+ for token in self.generate(input_ids, generation_config=stream_config):
528
+ outputs.append(token.item())
529
+ yield tokenizer.decode(outputs, skip_special_tokens=True)
530
+
531
+ return stream_generator()
532
+ else:
533
+ self.__class__.generate = PreTrainedModel.generate # disable stream
534
+ outputs = self.generate(input_ids, generation_config=generation_config)
535
+ response = tokenizer.decode(outputs[0][len(input_ids[0]):], skip_special_tokens=True)
536
+ return response
quantizer.py ADDED
@@ -0,0 +1,123 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
2
+
3
+ import torch
4
+ from typing import List
5
+ import bz2
6
+ import base64
7
+ import ctypes
8
+ from transformers.utils import logging
9
+ logger = logging.get_logger(__name__)
10
+
11
+ try:
12
+ from cpm_kernels.kernels.base import LazyKernelCModule, KernelFunction, round_up
13
+
14
+ class Kernel:
15
+ def __init__(self, code: bytes, function_names: List[str]):
16
+ self.code = code
17
+ self._function_names = function_names
18
+ self._cmodule = LazyKernelCModule(self.code)
19
+
20
+ for name in self._function_names:
21
+ setattr(self, name, KernelFunction(self._cmodule, name))
22
+ quantization_code = "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"
23
+ kernels = Kernel(
24
+ bz2.decompress(base64.b64decode(quantization_code)),
25
+ [
26
+ "int4_to_fp16",
27
+ "fp16_to_int4",
28
+ "int8_to_fp16",
29
+ "fp16_to_int8",
30
+ "int4_to_bf16",
31
+ "bf16_to_int4",
32
+ "int8_to_bf16",
33
+ "bf16_to_int8",
34
+ ],
35
+ )
36
+ except Exception as exception:
37
+ kernels = None
38
+ logger.warning("Failed to load kernels:" + str(exception))
39
+
40
+ def quant4(weight: torch.Tensor, scale: torch.Tensor):
41
+ stream = torch.cuda.current_stream()
42
+ num_row = weight.size(0)
43
+ num_chan_fp16 = weight.size(1)
44
+ # 4bit
45
+ num_chan_int = num_chan_fp16 // 8
46
+ qweight = torch.zeros((num_row, num_chan_int), dtype=torch.int32, device=weight.device)
47
+ intweight = torch.empty(num_row, num_chan_fp16, dtype = torch.int32)
48
+ intweight = torch.clip(torch.round(weight.to(scale.dtype) / scale[:, None]),-16, 15).to(dtype=torch.int32)
49
+
50
+ for j in range(num_chan_int):
51
+ qweight[:, j] = ((intweight[:, j*8+7] & 0x0f) << 28) \
52
+ | ((intweight[:, j*8+6] & 0x0f) << 24) \
53
+ | ((intweight[:, j*8+5] & 0x0f) << 20) \
54
+ | ((intweight[:, j*8+4] & 0x0f) << 16) \
55
+ | ((intweight[:, j*8+3] & 0x0f) << 12) \
56
+ | ((intweight[:, j*8+2] & 0x0f) << 8) \
57
+ | ((intweight[:, j*8+1] & 0x0f) << 4) \
58
+ | ((intweight[:, j*8] & 0x0f))
59
+ return qweight
60
+
61
+ def dequant4(qweight: torch.Tensor, scale: torch.Tensor, input: torch.Tensor):
62
+ stream = torch.cuda.current_stream()
63
+ num_row = qweight.size(0)
64
+ num_chan_int = qweight.size(1)
65
+ # 4bit
66
+ num_chan_fp16 = num_chan_int * 8
67
+
68
+ out = torch.empty((num_row, num_chan_fp16), dtype=input.dtype, device=qweight.device)
69
+
70
+ blockDim = (128, 1, 1)
71
+ gridDim = ((num_chan_int + blockDim[0] - 1) // blockDim[0], num_row, 1)
72
+ if input.dtype == torch.bfloat16:
73
+ kernels.int4_to_bf16(
74
+ gridDim,
75
+ blockDim,
76
+ 0,
77
+ stream,
78
+ [ctypes.c_void_p(out.data_ptr()), ctypes.c_void_p(qweight.data_ptr()),
79
+ ctypes.c_void_p(scale.data_ptr()), ctypes.c_int32(num_row), ctypes.c_int32(num_chan_int), ctypes.c_int32(num_chan_fp16)],
80
+ )
81
+ elif input.dtype == torch.float16:
82
+ kernels.int4_to_fp16(
83
+ gridDim,
84
+ blockDim,
85
+ 0,
86
+ stream,
87
+ [ctypes.c_void_p(out.data_ptr()), ctypes.c_void_p(qweight.data_ptr()),
88
+ ctypes.c_void_p(scale.data_ptr()), ctypes.c_int32(num_row), ctypes.c_int32(num_chan_int), ctypes.c_int32(num_chan_fp16)],
89
+ )
90
+ return out
91
+
92
+ class QLinear(torch.nn.Module):
93
+ def __init__(self, bits: int, weight: torch.Tensor, bias=None):
94
+ super().__init__()
95
+ self.quant_bits = bits
96
+ self.scale = weight.abs().max(dim=-1).values / ((2 ** (bits - 1)) - 1)
97
+ self.scale = self.scale.to(torch.float32)
98
+ if self.quant_bits == 4:
99
+ self.weight = quant4(weight, self.scale)
100
+ elif self.quant_bits == 8:
101
+ self.weight = torch.round(weight.to(self.scale.dtype) / self.scale[:, None]).to(torch.int8)
102
+ if self.quant_bits == 8:
103
+ self.weight = self.weight.T
104
+ self.bias = None
105
+
106
+ def forward(self, input):
107
+ if self.quant_bits == 4:
108
+ assert(input.dtype == torch.bfloat16 or input.dtype == torch.float16)
109
+
110
+ if self.weight.device != input.device:
111
+ self.weight = self.weight.to(input.device)
112
+ self.scale = self.scale.to(input.device)
113
+
114
+ if self.quant_bits == 4:
115
+ self.scale = self.scale.to(input.dtype)
116
+ rweight = dequant4(self.weight, self.scale, input).T
117
+ output = torch.matmul(input, rweight)
118
+ elif self.quant_bits == 8:
119
+ rweight = self.weight.to(input.dtype) * self.scale.to(input.dtype)
120
+ output = torch.matmul(input, rweight)
121
+ if self.bias is not None:
122
+ output = output + self.bias
123
+ return output
special_tokens_map.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<s>",
4
+ "lstrip": false,
5
+ "normalized": true,
6
+ "rstrip": false,
7
+ "single_word": true
8
+ },
9
+ "eos_token": {
10
+ "content": "</s>",
11
+ "lstrip": false,
12
+ "normalized": true,
13
+ "rstrip": false,
14
+ "single_word": true
15
+ },
16
+ "pad_token": {
17
+ "content": "<unk>",
18
+ "lstrip": false,
19
+ "normalized": true,
20
+ "rstrip": false,
21
+ "single_word": true
22
+ },
23
+ "unk_token": {
24
+ "content": "<unk>",
25
+ "lstrip": false,
26
+ "normalized": true,
27
+ "rstrip": false,
28
+ "single_word": true
29
+ }
30
+ }
tokenization_baichuan.py ADDED
@@ -0,0 +1,232 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
2
+
3
+ import os
4
+ from shutil import copyfile
5
+ from typing import Any, Dict, List, Optional, Tuple
6
+
7
+ import sentencepiece as spm
8
+ from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
9
+ from transformers.utils import logging
10
+
11
+
12
+ logger = logging.get_logger(__name__)
13
+
14
+ VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
15
+
16
+ PRETRAINED_VOCAB_FILES_MAP = {
17
+ "vocab_file": {},
18
+ "tokenizer_file": {},
19
+ }
20
+ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {}
21
+
22
+
23
+ class BaichuanTokenizer(PreTrainedTokenizer):
24
+ """
25
+ Construct a Baichuan tokenizer. Based on byte-level Byte-Pair-Encoding.
26
+
27
+ Args:
28
+ vocab_file (`str`):
29
+ Path to the vocabulary file.
30
+ """
31
+
32
+ vocab_files_names = VOCAB_FILES_NAMES
33
+ pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
34
+ max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
35
+ model_input_names = ["input_ids", "attention_mask"]
36
+
37
+ def __init__(
38
+ self,
39
+ vocab_file,
40
+ unk_token="<unk>",
41
+ bos_token="<s>",
42
+ eos_token="</s>",
43
+ pad_token=None,
44
+ sp_model_kwargs: Optional[Dict[str, Any]] = None,
45
+ add_bos_token=True,
46
+ add_eos_token=False,
47
+ clean_up_tokenization_spaces=False,
48
+ **kwargs,
49
+ ):
50
+ self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
51
+ bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
52
+ eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
53
+ unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
54
+ pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
55
+ super().__init__(
56
+ bos_token=bos_token,
57
+ eos_token=eos_token,
58
+ unk_token=unk_token,
59
+ pad_token=pad_token,
60
+ add_bos_token=add_bos_token,
61
+ add_eos_token=add_eos_token,
62
+ sp_model_kwargs=self.sp_model_kwargs,
63
+ clean_up_tokenization_spaces=clean_up_tokenization_spaces,
64
+ **kwargs,
65
+ )
66
+ self.vocab_file = vocab_file
67
+ self.add_bos_token = add_bos_token
68
+ self.add_eos_token = add_eos_token
69
+ self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
70
+ self.sp_model.Load(vocab_file)
71
+
72
+ def __getstate__(self):
73
+ state = self.__dict__.copy()
74
+ state["sp_model"] = None
75
+ return state
76
+
77
+ def __setstate__(self, d):
78
+ self.__dict__ = d
79
+ self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
80
+ self.sp_model.Load(self.vocab_file)
81
+
82
+ @property
83
+ def vocab_size(self):
84
+ """Returns vocab size"""
85
+ return self.sp_model.get_piece_size()
86
+
87
+ def get_vocab(self):
88
+ """Returns vocab as a dict"""
89
+ vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
90
+ vocab.update(self.added_tokens_encoder)
91
+ return vocab
92
+
93
+ def _tokenize(self, text):
94
+ """Returns a tokenized string."""
95
+ return self.sp_model.encode(text, out_type=str)
96
+
97
+ def _convert_token_to_id(self, token):
98
+ """Converts a token (str) in an id using the vocab."""
99
+ return self.sp_model.piece_to_id(token)
100
+
101
+ def _convert_id_to_token(self, index):
102
+ """Converts an index (integer) in a token (str) using the vocab."""
103
+ token = self.sp_model.IdToPiece(index)
104
+ return token
105
+
106
+ def convert_tokens_to_string(self, tokens):
107
+ """Converts a sequence of tokens (string) in a single string."""
108
+ current_sub_tokens = []
109
+ out_string = ""
110
+ prev_is_special = False
111
+ for i, token in enumerate(tokens):
112
+ # make sure that special tokens are not decoded using sentencepiece model
113
+ if token in self.all_special_tokens:
114
+ if not prev_is_special and i != 0:
115
+ out_string += " "
116
+ out_string += self.sp_model.decode(current_sub_tokens) + token
117
+ prev_is_special = True
118
+ current_sub_tokens = []
119
+ else:
120
+ current_sub_tokens.append(token)
121
+ prev_is_special = False
122
+ out_string += self.sp_model.decode(current_sub_tokens)
123
+ return out_string
124
+
125
+ def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
126
+ """
127
+ Save the vocabulary and special tokens file to a directory.
128
+
129
+ Args:
130
+ save_directory (`str`):
131
+ The directory in which to save the vocabulary.
132
+
133
+ Returns:
134
+ `Tuple(str)`: Paths to the files saved.
135
+ """
136
+ if not os.path.isdir(save_directory):
137
+ logger.error(f"Vocabulary path ({save_directory}) should be a directory")
138
+ return
139
+ out_vocab_file = os.path.join(
140
+ save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
141
+ )
142
+
143
+ if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
144
+ copyfile(self.vocab_file, out_vocab_file)
145
+ elif not os.path.isfile(self.vocab_file):
146
+ with open(out_vocab_file, "wb") as fi:
147
+ content_spiece_model = self.sp_model.serialized_model_proto()
148
+ fi.write(content_spiece_model)
149
+
150
+ return (out_vocab_file,)
151
+
152
+ def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
153
+ bos_token_id = [self.bos_token_id] if self.add_bos_token else []
154
+ eos_token_id = [self.eos_token_id] if self.add_eos_token else []
155
+
156
+ output = bos_token_id + token_ids_0 + eos_token_id
157
+
158
+ if token_ids_1 is not None:
159
+ output = output + bos_token_id + token_ids_1 + eos_token_id
160
+
161
+ return output
162
+
163
+ def get_special_tokens_mask(
164
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
165
+ ) -> List[int]:
166
+ """
167
+ Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
168
+ special tokens using the tokenizer `prepare_for_model` method.
169
+
170
+ Args:
171
+ token_ids_0 (`List[int]`):
172
+ List of IDs.
173
+ token_ids_1 (`List[int]`, *optional*):
174
+ Optional second list of IDs for sequence pairs.
175
+ already_has_special_tokens (`bool`, *optional*, defaults to `False`):
176
+ Whether or not the token list is already formatted with special tokens for the model.
177
+
178
+ Returns:
179
+ `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
180
+ """
181
+ if already_has_special_tokens:
182
+ return super().get_special_tokens_mask(
183
+ token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
184
+ )
185
+
186
+ bos_token_id = [1] if self.add_bos_token else []
187
+ eos_token_id = [1] if self.add_eos_token else []
188
+
189
+ if token_ids_1 is None:
190
+ return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
191
+ return (
192
+ bos_token_id
193
+ + ([0] * len(token_ids_0))
194
+ + eos_token_id
195
+ + bos_token_id
196
+ + ([0] * len(token_ids_1))
197
+ + eos_token_id
198
+ )
199
+
200
+ def create_token_type_ids_from_sequences(
201
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
202
+ ) -> List[int]:
203
+ """
204
+ Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
205
+ sequence pair mask has the following format:
206
+
207
+ ```
208
+ 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
209
+ | first sequence | second sequence |
210
+ ```
211
+
212
+ if token_ids_1 is None, only returns the first portion of the mask (0s).
213
+
214
+ Args:
215
+ token_ids_0 (`List[int]`):
216
+ List of ids.
217
+ token_ids_1 (`List[int]`, *optional*):
218
+ Optional second list of IDs for sequence pairs.
219
+
220
+ Returns:
221
+ `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
222
+ """
223
+ bos_token_id = [self.bos_token_id] if self.add_bos_token else []
224
+ eos_token_id = [self.eos_token_id] if self.add_eos_token else []
225
+
226
+ output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
227
+
228
+ if token_ids_1 is not None:
229
+ output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
230
+
231
+ return output
232
+
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f7d1ab69d25c74644af5c5e4dcd1cc6e96d33783dbd257b6bdea55b643c72813
3
+ size 1136765
tokenizer_config.json ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": false,
3
+ "add_eos_token": false,
4
+ "auto_map": {
5
+ "AutoTokenizer": [
6
+ "tokenization_baichuan.BaichuanTokenizer",
7
+ null
8
+ ]
9
+ },
10
+ "bos_token": {
11
+ "__type": "AddedToken",
12
+ "content": "<s>",
13
+ "lstrip": false,
14
+ "normalized": true,
15
+ "rstrip": false,
16
+ "single_word": true
17
+ },
18
+ "clean_up_tokenization_spaces": false,
19
+ "eos_token": {
20
+ "__type": "AddedToken",
21
+ "content": "</s>",
22
+ "lstrip": false,
23
+ "normalized": true,
24
+ "rstrip": false,
25
+ "single_word": true
26
+ },
27
+ "model_max_length": 4096,
28
+ "pad_token": {
29
+ "__type": "AddedToken",
30
+ "content": "<unk>",
31
+ "lstrip": false,
32
+ "normalized": true,
33
+ "rstrip": false,
34
+ "single_word": true
35
+ },
36
+ "sp_model_kwargs": {},
37
+ "tokenizer_class": "BaichuanTokenizer",
38
+ "unk_token": {
39
+ "__type": "AddedToken",
40
+ "content": "<unk>",
41
+ "lstrip": false,
42
+ "normalized": true,
43
+ "rstrip": false,
44
+ "single_word": true
45
+ }
46
+ }
trainer_state.json ADDED
The diff for this file is too large to render. See raw diff
 
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:23b60247e9813e273844c899f6f926980be4248401b52eee6f57dcf7bb738eed
3
+ size 5755