hiyouga commited on
Commit
e901b0a
1 Parent(s): 70f18bc

Upload 10 files

Browse files
config.json ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "/home/incoming/zhengyw/baichuan/",
3
+ "architectures": [
4
+ "BaiChuanForCausalLM"
5
+ ],
6
+ "auto_map": {
7
+ "AutoConfig": "configuration_baichuan.BaiChuanConfig",
8
+ "AutoModelForCausalLM": "modeling_baichuan.BaiChuanForCausalLM"
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": 4096,
17
+ "model_type": "baichuan",
18
+ "num_attention_heads": 32,
19
+ "num_hidden_layers": 32,
20
+ "pad_token_id": 0,
21
+ "rms_norm_eps": 1e-06,
22
+ "tie_word_embeddings": false,
23
+ "torch_dtype": "float16",
24
+ "transformers_version": "4.30.1",
25
+ "use_cache": true,
26
+ "vocab_size": 64000
27
+ }
configuration_baichuan.py ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
+ # and OPT implementations in this library. It has been modified from its
6
+ # original forms to accommodate minor architectural differences compared
7
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
8
+ #
9
+ # Licensed under the Apache License, Version 2.0 (the "License");
10
+ # you may not use this file except in compliance with the License.
11
+ # You may obtain a copy of the License at
12
+ #
13
+ # http://www.apache.org/licenses/LICENSE-2.0
14
+ #
15
+ # Unless required by applicable law or agreed to in writing, software
16
+ # distributed under the License is distributed on an "AS IS" BASIS,
17
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
+ # See the License for the specific language governing permissions and
19
+ # limitations under the License.
20
+
21
+ from transformers.configuration_utils import PretrainedConfig
22
+ from transformers.utils import logging
23
+
24
+
25
+ logger = logging.get_logger(__name__)
26
+
27
+
28
+ class BaiChuanConfig(PretrainedConfig):
29
+ model_type = "baichuan"
30
+ keys_to_ignore_at_inference = ["past_key_values"]
31
+
32
+ def __init__(
33
+ self,
34
+ vocab_size=64000,
35
+ hidden_size=4096,
36
+ intermediate_size=11008,
37
+ num_hidden_layers=32,
38
+ num_attention_heads=32,
39
+ hidden_act="silu",
40
+ max_position_embeddings=4096,
41
+ initializer_range=0.02,
42
+ rms_norm_eps=1e-6,
43
+ use_cache=True,
44
+ pad_token_id=0,
45
+ bos_token_id=1,
46
+ eos_token_id=2,
47
+ tie_word_embeddings=False,
48
+ **kwargs,
49
+ ):
50
+ self.vocab_size = vocab_size
51
+ self.max_position_embeddings = max_position_embeddings
52
+ self.hidden_size = hidden_size
53
+ self.intermediate_size = intermediate_size
54
+ self.num_hidden_layers = num_hidden_layers
55
+ self.num_attention_heads = num_attention_heads
56
+ self.hidden_act = hidden_act
57
+ self.initializer_range = initializer_range
58
+ self.rms_norm_eps = rms_norm_eps
59
+ self.use_cache = use_cache
60
+ super().__init__(
61
+ pad_token_id=pad_token_id,
62
+ bos_token_id=bos_token_id,
63
+ eos_token_id=eos_token_id,
64
+ tie_word_embeddings=tie_word_embeddings,
65
+ **kwargs,
66
+ )
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": "4.30.1"
7
+ }
handler.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from typing import Dict, List, Any
3
+ from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
4
+
5
+ # get dtype
6
+ dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] == 8 else torch.float16
7
+
8
+
9
+ class EndpointHandler:
10
+ def __init__(self, path=""):
11
+ # load the model
12
+ tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
13
+ model = AutoModelForCausalLM.from_pretrained(path, device_map="auto", torch_dtype=dtype, trust_remote_code=True)
14
+ # create inference pipeline
15
+ self.pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer)
16
+
17
+ def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
18
+ inputs = data.pop("inputs", data)
19
+ parameters = data.pop("parameters", None)
20
+
21
+ # pass inputs with all kwargs in data
22
+ if parameters is not None:
23
+ prediction = self.pipeline(inputs, **parameters)
24
+ else:
25
+ prediction = self.pipeline(inputs)
26
+ # postprocess the prediction
27
+ return prediction
modeling_baichuan.py ADDED
@@ -0,0 +1,678 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
+ # and OPT implementations in this library. It has been modified from its
6
+ # original forms to accommodate minor architectural differences compared
7
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
8
+ #
9
+ # Licensed under the Apache License, Version 2.0 (the "License");
10
+ # you may not use this file except in compliance with the License.
11
+ # You may obtain a copy of the License at
12
+ #
13
+ # http://www.apache.org/licenses/LICENSE-2.0
14
+ #
15
+ # Unless required by applicable law or agreed to in writing, software
16
+ # distributed under the License is distributed on an "AS IS" BASIS,
17
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
+ # See the License for the specific language governing permissions and
19
+ # limitations under the License.
20
+ from .configuration_baichuan import BaiChuanConfig
21
+ from transformers import PreTrainedModel, add_start_docstrings
22
+ from transformers.activations import ACT2FN
23
+ from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, \
24
+ SequenceClassifierOutputWithPast
25
+ from transformers.utils import logging, add_start_docstrings_to_model_forward, replace_return_docstrings
26
+
27
+ import math
28
+ from typing import List, Optional, Tuple, Union
29
+
30
+ import torch
31
+ import torch.utils.checkpoint
32
+ from torch import nn
33
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
34
+
35
+
36
+ logger = logging.get_logger(__name__)
37
+
38
+ # Copied from transformers.models.bart.modeling_bart._make_causal_mask
39
+ def _make_causal_mask(
40
+ input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
41
+ ):
42
+ """
43
+ Make causal mask used for bi-directional self-attention.
44
+ """
45
+ bsz, tgt_len = input_ids_shape
46
+ mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
47
+ mask_cond = torch.arange(mask.size(-1), device=device)
48
+ mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
49
+ mask = mask.to(dtype)
50
+
51
+ if past_key_values_length > 0:
52
+ mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
53
+ return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
54
+
55
+
56
+ # Copied from transformers.models.bart.modeling_bart._expand_mask
57
+ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
58
+ """
59
+ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
60
+ """
61
+ bsz, src_len = mask.size()
62
+ tgt_len = tgt_len if tgt_len is not None else src_len
63
+
64
+ expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
65
+
66
+ inverted_mask = 1.0 - expanded_mask
67
+
68
+ return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
69
+
70
+
71
+ class RMSNorm(nn.Module):
72
+ def __init__(self, hidden_size, eps=1e-6):
73
+ """
74
+ RMSNorm is equivalent to T5LayerNorm
75
+ """
76
+ super().__init__()
77
+ self.weight = nn.Parameter(torch.ones(hidden_size))
78
+ self.variance_epsilon = eps
79
+
80
+ def forward(self, hidden_states):
81
+ variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
82
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
83
+
84
+ # convert into half-precision if necessary
85
+ if self.weight.dtype in [torch.float16, torch.bfloat16]:
86
+ hidden_states = hidden_states.to(self.weight.dtype)
87
+
88
+ return self.weight * hidden_states
89
+
90
+
91
+ class RotaryEmbedding(torch.nn.Module):
92
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
93
+ super().__init__()
94
+ inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
95
+ self.register_buffer("inv_freq", inv_freq)
96
+
97
+ # Build here to make `torch.jit.trace` work.
98
+ self.max_seq_len_cached = max_position_embeddings
99
+ t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
100
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
101
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
102
+ emb = torch.cat((freqs, freqs), dim=-1)
103
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
104
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
105
+
106
+ def forward(self, x, seq_len=None):
107
+ # x: [bs, num_attention_heads, seq_len, head_size]
108
+ # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
109
+ if seq_len > self.max_seq_len_cached:
110
+ self.max_seq_len_cached = seq_len
111
+ t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
112
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
113
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
114
+ emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
115
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
116
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
117
+ return (
118
+ self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
119
+ self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
120
+ )
121
+
122
+
123
+ def rotate_half(x):
124
+ """Rotates half the hidden dims of the input."""
125
+ x1 = x[..., : x.shape[-1] // 2]
126
+ x2 = x[..., x.shape[-1] // 2:]
127
+ return torch.cat((-x2, x1), dim=-1)
128
+
129
+
130
+ def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
131
+ # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
132
+ cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
133
+ sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
134
+ cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
135
+ sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
136
+ q_embed = (q * cos) + (rotate_half(q) * sin)
137
+ k_embed = (k * cos) + (rotate_half(k) * sin)
138
+ return q_embed, k_embed
139
+
140
+
141
+ class MLP(nn.Module):
142
+ def __init__(
143
+ self,
144
+ hidden_size: int,
145
+ intermediate_size: int,
146
+ hidden_act: str,
147
+ ):
148
+ super().__init__()
149
+ self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
150
+ self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
151
+ self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
152
+ self.act_fn = ACT2FN[hidden_act]
153
+
154
+ def forward(self, x):
155
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
156
+
157
+
158
+ class Attention(nn.Module):
159
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
160
+
161
+ def __init__(self, config: BaiChuanConfig):
162
+ super().__init__()
163
+ self.config = config
164
+ self.hidden_size = config.hidden_size
165
+ self.num_heads = config.num_attention_heads
166
+ self.head_dim = self.hidden_size // self.num_heads
167
+ self.max_position_embeddings = config.max_position_embeddings
168
+
169
+ if (self.head_dim * self.num_heads) != self.hidden_size:
170
+ raise ValueError(
171
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
172
+ f" and `num_heads`: {self.num_heads})."
173
+ )
174
+ # self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
175
+ # self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
176
+ # self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
177
+ self.W_pack = nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=False)
178
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
179
+ self.rotary_emb = RotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
180
+
181
+ def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
182
+ return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
183
+
184
+ def forward(
185
+ self,
186
+ hidden_states: torch.Tensor,
187
+ attention_mask: Optional[torch.Tensor] = None,
188
+ position_ids: Optional[torch.LongTensor] = None,
189
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
190
+ output_attentions: bool = False,
191
+ use_cache: bool = False,
192
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
193
+ bsz, q_len, _ = hidden_states.size()
194
+
195
+ proj = self.W_pack(hidden_states)
196
+ proj = proj.unflatten(-1, (3, self.hidden_size)).unsqueeze(0).transpose(0, -2).squeeze(-2)
197
+ query_states = proj[0].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1,
198
+ 2) # batch_size x source_len x hidden_size
199
+ key_states = proj[1].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1,
200
+ 2) # batch_size x target_len x head_size
201
+ value_states = proj[2].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1,
202
+ 2) # batch_size x source_len x hidden_size
203
+
204
+ # query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
205
+ # key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
206
+ # value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
207
+
208
+ kv_seq_len = key_states.shape[-2]
209
+ if past_key_value is not None:
210
+ kv_seq_len += past_key_value[0].shape[-2]
211
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
212
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
213
+ # [bsz, nh, t, hd]
214
+
215
+ if past_key_value is not None:
216
+ # reuse k, v, self_attention
217
+ key_states = torch.cat([past_key_value[0], key_states], dim=2)
218
+ value_states = torch.cat([past_key_value[1], value_states], dim=2)
219
+
220
+ past_key_value = (key_states, value_states) if use_cache else None
221
+
222
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
223
+
224
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
225
+ raise ValueError(
226
+ f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
227
+ f" {attn_weights.size()}"
228
+ )
229
+
230
+ if attention_mask is not None:
231
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
232
+ raise ValueError(
233
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
234
+ )
235
+ attn_weights = attn_weights + attention_mask
236
+ attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
237
+
238
+ # upcast attention to fp32
239
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
240
+ attn_output = torch.matmul(attn_weights, value_states)
241
+
242
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
243
+ raise ValueError(
244
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
245
+ f" {attn_output.size()}"
246
+ )
247
+
248
+ attn_output = attn_output.transpose(1, 2)
249
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
250
+
251
+ attn_output = self.o_proj(attn_output)
252
+
253
+ if not output_attentions:
254
+ attn_weights = None
255
+
256
+ return attn_output, attn_weights, past_key_value
257
+
258
+
259
+ class DecoderLayer(nn.Module):
260
+ def __init__(self, config: BaiChuanConfig):
261
+ super().__init__()
262
+ self.hidden_size = config.hidden_size
263
+ self.self_attn = Attention(config=config)
264
+ self.mlp = MLP(
265
+ hidden_size=self.hidden_size,
266
+ intermediate_size=config.intermediate_size,
267
+ hidden_act=config.hidden_act,
268
+ )
269
+ self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
270
+ self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
271
+
272
+ def forward(
273
+ self,
274
+ hidden_states: torch.Tensor,
275
+ attention_mask: Optional[torch.Tensor] = None,
276
+ position_ids: Optional[torch.LongTensor] = None,
277
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
278
+ output_attentions: Optional[bool] = False,
279
+ use_cache: Optional[bool] = False,
280
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
281
+ """
282
+ Args:
283
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
284
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
285
+ `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
286
+ output_attentions (`bool`, *optional*):
287
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
288
+ returned tensors for more detail.
289
+ use_cache (`bool`, *optional*):
290
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
291
+ (see `past_key_values`).
292
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
293
+ """
294
+
295
+ residual = hidden_states
296
+
297
+ hidden_states = self.input_layernorm(hidden_states)
298
+
299
+ # Self Attention
300
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
301
+ hidden_states=hidden_states,
302
+ attention_mask=attention_mask,
303
+ position_ids=position_ids,
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)
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
325
+
326
+
327
+ class PreTrainedModel(PreTrainedModel):
328
+ config_class = BaiChuanConfig
329
+ base_model_prefix = "model"
330
+ supports_gradient_checkpointing = True
331
+ _no_split_modules = ["DecoderLayer"]
332
+ _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
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
+ def _set_gradient_checkpointing(self, module, value=False):
346
+ if isinstance(module, Model):
347
+ module.gradient_checkpointing = value
348
+
349
+
350
+ class Model(PreTrainedModel):
351
+ """
352
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`DecoderLayer`]
353
+
354
+ Args:
355
+ config: BaiChuanConfig
356
+ """
357
+
358
+ def __init__(self, config: BaiChuanConfig):
359
+ super().__init__(config)
360
+ self.padding_idx = config.pad_token_id
361
+ self.vocab_size = config.vocab_size
362
+
363
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
364
+ self.layers = nn.ModuleList([DecoderLayer(config) for _ in range(config.num_hidden_layers)])
365
+ self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
366
+
367
+ self.gradient_checkpointing = False
368
+ # Initialize weights and apply final processing
369
+ self.post_init()
370
+
371
+ def get_input_embeddings(self):
372
+ return self.embed_tokens
373
+
374
+ def set_input_embeddings(self, value):
375
+ self.embed_tokens = value
376
+
377
+ # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
378
+ def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
379
+ # create causal mask
380
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
381
+ combined_attention_mask = None
382
+ if input_shape[-1] > 1:
383
+ combined_attention_mask = _make_causal_mask(
384
+ input_shape,
385
+ inputs_embeds.dtype,
386
+ device=inputs_embeds.device,
387
+ past_key_values_length=past_key_values_length,
388
+ )
389
+
390
+ if attention_mask is not None:
391
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
392
+ expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
393
+ inputs_embeds.device
394
+ )
395
+ combined_attention_mask = (
396
+ expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
397
+ )
398
+
399
+ return combined_attention_mask
400
+
401
+ def forward(
402
+ self,
403
+ input_ids: torch.LongTensor = None,
404
+ attention_mask: Optional[torch.Tensor] = None,
405
+ position_ids: Optional[torch.LongTensor] = None,
406
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
407
+ inputs_embeds: Optional[torch.FloatTensor] = None,
408
+ use_cache: Optional[bool] = None,
409
+ output_attentions: Optional[bool] = None,
410
+ output_hidden_states: Optional[bool] = None,
411
+ return_dict: Optional[bool] = None,
412
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
413
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
414
+ output_hidden_states = (
415
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
416
+ )
417
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
418
+
419
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
420
+
421
+ # retrieve input_ids and inputs_embeds
422
+ if input_ids is not None and inputs_embeds is not None:
423
+ raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
424
+ elif input_ids is not None:
425
+ batch_size, seq_length = input_ids.shape
426
+ elif inputs_embeds is not None:
427
+ batch_size, seq_length, _ = inputs_embeds.shape
428
+ else:
429
+ raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
430
+
431
+ seq_length_with_past = seq_length
432
+ past_key_values_length = 0
433
+
434
+ if past_key_values is not None:
435
+ past_key_values_length = past_key_values[0][0].shape[2]
436
+ seq_length_with_past = seq_length_with_past + past_key_values_length
437
+
438
+ if position_ids is None:
439
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
440
+ position_ids = torch.arange(
441
+ past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
442
+ )
443
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
444
+ else:
445
+ position_ids = position_ids.view(-1, seq_length).long()
446
+
447
+ if inputs_embeds is None:
448
+ inputs_embeds = self.embed_tokens(input_ids)
449
+ # embed positions
450
+ if attention_mask is None:
451
+ attention_mask = torch.ones(
452
+ (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
453
+ )
454
+ attention_mask = self._prepare_decoder_attention_mask(
455
+ attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
456
+ )
457
+
458
+ hidden_states = inputs_embeds
459
+
460
+ if self.gradient_checkpointing and self.training:
461
+ if use_cache:
462
+ logger.warning_once(
463
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
464
+ )
465
+ use_cache = False
466
+
467
+ # decoder layers
468
+ all_hidden_states = () if output_hidden_states else None
469
+ all_self_attns = () if output_attentions else None
470
+ next_decoder_cache = () if use_cache else None
471
+
472
+ for idx, decoder_layer in enumerate(self.layers):
473
+ if output_hidden_states:
474
+ all_hidden_states += (hidden_states,)
475
+
476
+ past_key_value = past_key_values[idx] if past_key_values is not None else None
477
+
478
+ if self.gradient_checkpointing and self.training:
479
+
480
+ def create_custom_forward(module):
481
+ def custom_forward(*inputs):
482
+ # None for past_key_value
483
+ return module(*inputs, output_attentions, None)
484
+
485
+ return custom_forward
486
+
487
+ layer_outputs = torch.utils.checkpoint.checkpoint(
488
+ create_custom_forward(decoder_layer),
489
+ hidden_states,
490
+ attention_mask,
491
+ position_ids,
492
+ None,
493
+ )
494
+ else:
495
+ layer_outputs = decoder_layer(
496
+ hidden_states,
497
+ attention_mask=attention_mask,
498
+ position_ids=position_ids,
499
+ past_key_value=past_key_value,
500
+ output_attentions=output_attentions,
501
+ use_cache=use_cache,
502
+ )
503
+
504
+ hidden_states = layer_outputs[0]
505
+
506
+ if use_cache:
507
+ next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
508
+
509
+ if output_attentions:
510
+ all_self_attns += (layer_outputs[1],)
511
+
512
+ hidden_states = self.norm(hidden_states)
513
+
514
+ # add hidden states from the last decoder layer
515
+ if output_hidden_states:
516
+ all_hidden_states += (hidden_states,)
517
+
518
+ next_cache = next_decoder_cache if use_cache else None
519
+ if not return_dict:
520
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
521
+ return BaseModelOutputWithPast(
522
+ last_hidden_state=hidden_states,
523
+ past_key_values=next_cache,
524
+ hidden_states=all_hidden_states,
525
+ attentions=all_self_attns,
526
+ )
527
+
528
+
529
+ class BaiChuanForCausalLM(PreTrainedModel):
530
+ def __init__(self, config):
531
+ super().__init__(config)
532
+ self.model = Model(config)
533
+
534
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
535
+
536
+ # Initialize weights and apply final processing
537
+ self.post_init()
538
+
539
+ def get_input_embeddings(self):
540
+ return self.model.embed_tokens
541
+
542
+ def set_input_embeddings(self, value):
543
+ self.model.embed_tokens = value
544
+
545
+ def get_output_embeddings(self):
546
+ return self.lm_head
547
+
548
+ def set_output_embeddings(self, new_embeddings):
549
+ self.lm_head = new_embeddings
550
+
551
+ def set_decoder(self, decoder):
552
+ self.model = decoder
553
+
554
+ def get_decoder(self):
555
+ return self.model
556
+
557
+ def forward(
558
+ self,
559
+ input_ids: torch.LongTensor = None,
560
+ attention_mask: Optional[torch.Tensor] = None,
561
+ position_ids: Optional[torch.LongTensor] = None,
562
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
563
+ inputs_embeds: Optional[torch.FloatTensor] = None,
564
+ labels: Optional[torch.LongTensor] = None,
565
+ use_cache: Optional[bool] = None,
566
+ output_attentions: Optional[bool] = None,
567
+ output_hidden_states: Optional[bool] = None,
568
+ return_dict: Optional[bool] = None,
569
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
570
+ r"""
571
+ Args:
572
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
573
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
574
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
575
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
576
+
577
+ Returns:
578
+
579
+ Example:
580
+
581
+ ```python
582
+ >>> from transformers import AutoTokenizer, ModelForCausalLM
583
+
584
+ >>> model = ModelForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
585
+ >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
586
+
587
+ >>> prompt = "Hey, are you consciours? Can you talk to me?"
588
+ >>> inputs = tokenizer(prompt, return_tensors="pt")
589
+
590
+ >>> # Generate
591
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
592
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
593
+ "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
594
+ ```"""
595
+
596
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
597
+ output_hidden_states = (
598
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
599
+ )
600
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
601
+
602
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
603
+ outputs = self.model(
604
+ input_ids=input_ids,
605
+ attention_mask=attention_mask,
606
+ position_ids=position_ids,
607
+ past_key_values=past_key_values,
608
+ inputs_embeds=inputs_embeds,
609
+ use_cache=use_cache,
610
+ output_attentions=output_attentions,
611
+ output_hidden_states=output_hidden_states,
612
+ return_dict=return_dict,
613
+ )
614
+
615
+ hidden_states = outputs[0]
616
+ logits = self.lm_head(hidden_states)
617
+
618
+ loss = None
619
+ if labels is not None:
620
+ # Shift so that tokens < n predict n
621
+ shift_logits = logits[..., :-1, :].contiguous()
622
+ shift_labels = labels[..., 1:].contiguous()
623
+ # Flatten the tokens
624
+ loss_fct = CrossEntropyLoss()
625
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
626
+ shift_labels = shift_labels.view(-1)
627
+ # Enable model parallelism
628
+ shift_labels = shift_labels.to(shift_logits.device)
629
+ loss = loss_fct(shift_logits, shift_labels)
630
+
631
+ if not return_dict:
632
+ output = (logits,) + outputs[1:]
633
+ return (loss,) + output if loss is not None else output
634
+
635
+ return CausalLMOutputWithPast(
636
+ loss=loss,
637
+ logits=logits,
638
+ past_key_values=outputs.past_key_values,
639
+ hidden_states=outputs.hidden_states,
640
+ attentions=outputs.attentions,
641
+ )
642
+
643
+ def prepare_inputs_for_generation(
644
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
645
+ ):
646
+ if past_key_values:
647
+ input_ids = input_ids[:, -1:]
648
+
649
+ position_ids = kwargs.get("position_ids", None)
650
+ if attention_mask is not None and position_ids is None:
651
+ # create position_ids on the fly for batch generation
652
+ position_ids = attention_mask.long().cumsum(-1) - 1
653
+ position_ids.masked_fill_(attention_mask == 0, 1)
654
+ if past_key_values:
655
+ position_ids = position_ids[:, -1].unsqueeze(-1)
656
+
657
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
658
+ if inputs_embeds is not None and past_key_values is None:
659
+ model_inputs = {"inputs_embeds": inputs_embeds}
660
+ else:
661
+ model_inputs = {"input_ids": input_ids}
662
+
663
+ model_inputs.update(
664
+ {
665
+ "position_ids": position_ids,
666
+ "past_key_values": past_key_values,
667
+ "use_cache": kwargs.get("use_cache"),
668
+ "attention_mask": attention_mask,
669
+ }
670
+ )
671
+ return model_inputs
672
+
673
+ @staticmethod
674
+ def _reorder_cache(past_key_values, beam_idx):
675
+ reordered_past = ()
676
+ for layer_past in past_key_values:
677
+ reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
678
+ return reordered_past
pytorch_model.bin.index.json ADDED
@@ -0,0 +1,266 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 14001123328
4
+ },
5
+ "weight_map": {
6
+ "lm_head.weight": "pytorch_model-00002-of-00002.bin",
7
+ "model.embed_tokens.weight": "pytorch_model-00001-of-00002.bin",
8
+ "model.layers.0.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
9
+ "model.layers.0.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
10
+ "model.layers.0.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
11
+ "model.layers.0.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
12
+ "model.layers.0.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
13
+ "model.layers.0.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
14
+ "model.layers.0.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
15
+ "model.layers.0.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
16
+ "model.layers.1.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
17
+ "model.layers.1.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
18
+ "model.layers.1.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
19
+ "model.layers.1.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
20
+ "model.layers.1.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
21
+ "model.layers.1.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
22
+ "model.layers.1.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
23
+ "model.layers.1.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
24
+ "model.layers.10.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
25
+ "model.layers.10.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
26
+ "model.layers.10.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
27
+ "model.layers.10.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
28
+ "model.layers.10.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
29
+ "model.layers.10.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
30
+ "model.layers.10.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
31
+ "model.layers.10.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
32
+ "model.layers.11.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
33
+ "model.layers.11.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
34
+ "model.layers.11.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
35
+ "model.layers.11.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
36
+ "model.layers.11.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
37
+ "model.layers.11.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
38
+ "model.layers.11.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
39
+ "model.layers.11.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
40
+ "model.layers.12.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
41
+ "model.layers.12.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
42
+ "model.layers.12.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
43
+ "model.layers.12.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
44
+ "model.layers.12.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
45
+ "model.layers.12.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
46
+ "model.layers.12.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
47
+ "model.layers.12.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
48
+ "model.layers.13.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
49
+ "model.layers.13.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
50
+ "model.layers.13.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
51
+ "model.layers.13.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
52
+ "model.layers.13.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
53
+ "model.layers.13.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
54
+ "model.layers.13.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
55
+ "model.layers.13.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
56
+ "model.layers.14.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
57
+ "model.layers.14.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
58
+ "model.layers.14.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
59
+ "model.layers.14.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
60
+ "model.layers.14.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
61
+ "model.layers.14.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
62
+ "model.layers.14.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
63
+ "model.layers.14.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
64
+ "model.layers.15.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
65
+ "model.layers.15.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
66
+ "model.layers.15.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
67
+ "model.layers.15.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
68
+ "model.layers.15.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
69
+ "model.layers.15.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
70
+ "model.layers.15.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
71
+ "model.layers.15.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
72
+ "model.layers.16.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
73
+ "model.layers.16.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
74
+ "model.layers.16.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
75
+ "model.layers.16.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
76
+ "model.layers.16.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
77
+ "model.layers.16.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
78
+ "model.layers.16.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
79
+ "model.layers.16.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
80
+ "model.layers.17.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
81
+ "model.layers.17.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
82
+ "model.layers.17.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
83
+ "model.layers.17.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
84
+ "model.layers.17.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
85
+ "model.layers.17.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
86
+ "model.layers.17.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
87
+ "model.layers.17.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
88
+ "model.layers.18.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
89
+ "model.layers.18.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
90
+ "model.layers.18.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
91
+ "model.layers.18.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
92
+ "model.layers.18.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
93
+ "model.layers.18.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
94
+ "model.layers.18.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
95
+ "model.layers.18.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
96
+ "model.layers.19.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
97
+ "model.layers.19.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
98
+ "model.layers.19.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
99
+ "model.layers.19.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
100
+ "model.layers.19.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
101
+ "model.layers.19.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
102
+ "model.layers.19.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
103
+ "model.layers.19.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
104
+ "model.layers.2.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
105
+ "model.layers.2.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
106
+ "model.layers.2.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
107
+ "model.layers.2.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
108
+ "model.layers.2.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
109
+ "model.layers.2.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
110
+ "model.layers.2.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
111
+ "model.layers.2.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
112
+ "model.layers.20.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
113
+ "model.layers.20.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
114
+ "model.layers.20.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
115
+ "model.layers.20.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
116
+ "model.layers.20.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
117
+ "model.layers.20.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
118
+ "model.layers.20.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
119
+ "model.layers.20.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
120
+ "model.layers.21.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
121
+ "model.layers.21.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
122
+ "model.layers.21.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
123
+ "model.layers.21.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
124
+ "model.layers.21.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
125
+ "model.layers.21.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
126
+ "model.layers.21.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
127
+ "model.layers.21.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
128
+ "model.layers.22.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
129
+ "model.layers.22.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
130
+ "model.layers.22.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
131
+ "model.layers.22.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
132
+ "model.layers.22.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
133
+ "model.layers.22.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
134
+ "model.layers.22.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
135
+ "model.layers.22.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
136
+ "model.layers.23.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
137
+ "model.layers.23.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
138
+ "model.layers.23.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
139
+ "model.layers.23.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
140
+ "model.layers.23.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
141
+ "model.layers.23.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
142
+ "model.layers.23.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
143
+ "model.layers.23.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
144
+ "model.layers.24.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
145
+ "model.layers.24.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
146
+ "model.layers.24.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
147
+ "model.layers.24.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
148
+ "model.layers.24.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
149
+ "model.layers.24.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
150
+ "model.layers.24.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
151
+ "model.layers.24.self_attn.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
152
+ "model.layers.25.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
153
+ "model.layers.25.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
154
+ "model.layers.25.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
155
+ "model.layers.25.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
156
+ "model.layers.25.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
157
+ "model.layers.25.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
158
+ "model.layers.25.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
159
+ "model.layers.25.self_attn.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
160
+ "model.layers.26.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
161
+ "model.layers.26.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
162
+ "model.layers.26.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
163
+ "model.layers.26.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
164
+ "model.layers.26.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
165
+ "model.layers.26.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
166
+ "model.layers.26.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
167
+ "model.layers.26.self_attn.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
168
+ "model.layers.27.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
169
+ "model.layers.27.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
170
+ "model.layers.27.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
171
+ "model.layers.27.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
172
+ "model.layers.27.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
173
+ "model.layers.27.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
174
+ "model.layers.27.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
175
+ "model.layers.27.self_attn.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
176
+ "model.layers.28.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
177
+ "model.layers.28.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
178
+ "model.layers.28.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
179
+ "model.layers.28.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
180
+ "model.layers.28.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
181
+ "model.layers.28.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
182
+ "model.layers.28.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
183
+ "model.layers.28.self_attn.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
184
+ "model.layers.29.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
185
+ "model.layers.29.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
186
+ "model.layers.29.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
187
+ "model.layers.29.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
188
+ "model.layers.29.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
189
+ "model.layers.29.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
190
+ "model.layers.29.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
191
+ "model.layers.29.self_attn.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
192
+ "model.layers.3.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
193
+ "model.layers.3.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
194
+ "model.layers.3.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
195
+ "model.layers.3.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
196
+ "model.layers.3.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
197
+ "model.layers.3.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
198
+ "model.layers.3.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
199
+ "model.layers.3.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
200
+ "model.layers.30.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
201
+ "model.layers.30.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
202
+ "model.layers.30.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
203
+ "model.layers.30.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
204
+ "model.layers.30.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
205
+ "model.layers.30.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
206
+ "model.layers.30.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
207
+ "model.layers.30.self_attn.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
208
+ "model.layers.31.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
209
+ "model.layers.31.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
210
+ "model.layers.31.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
211
+ "model.layers.31.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
212
+ "model.layers.31.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
213
+ "model.layers.31.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
214
+ "model.layers.31.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
215
+ "model.layers.31.self_attn.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
216
+ "model.layers.4.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
217
+ "model.layers.4.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
218
+ "model.layers.4.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
219
+ "model.layers.4.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
220
+ "model.layers.4.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
221
+ "model.layers.4.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
222
+ "model.layers.4.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
223
+ "model.layers.4.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
224
+ "model.layers.5.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
225
+ "model.layers.5.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
226
+ "model.layers.5.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
227
+ "model.layers.5.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
228
+ "model.layers.5.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
229
+ "model.layers.5.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
230
+ "model.layers.5.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
231
+ "model.layers.5.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
232
+ "model.layers.6.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
233
+ "model.layers.6.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
234
+ "model.layers.6.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
235
+ "model.layers.6.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
236
+ "model.layers.6.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
237
+ "model.layers.6.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
238
+ "model.layers.6.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
239
+ "model.layers.6.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
240
+ "model.layers.7.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
241
+ "model.layers.7.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
242
+ "model.layers.7.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
243
+ "model.layers.7.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
244
+ "model.layers.7.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
245
+ "model.layers.7.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
246
+ "model.layers.7.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
247
+ "model.layers.7.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
248
+ "model.layers.8.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
249
+ "model.layers.8.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
250
+ "model.layers.8.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
251
+ "model.layers.8.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
252
+ "model.layers.8.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
253
+ "model.layers.8.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
254
+ "model.layers.8.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
255
+ "model.layers.8.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
256
+ "model.layers.9.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
257
+ "model.layers.9.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
258
+ "model.layers.9.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
259
+ "model.layers.9.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
260
+ "model.layers.9.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
261
+ "model.layers.9.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
262
+ "model.layers.9.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
263
+ "model.layers.9.self_attn.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
264
+ "model.norm.weight": "pytorch_model-00002-of-00002.bin"
265
+ }
266
+ }
special_tokens_map.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<s>",
4
+ "lstrip": false,
5
+ "normalized": true,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "</s>",
11
+ "lstrip": false,
12
+ "normalized": true,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": "<unk>",
17
+ "unk_token": {
18
+ "content": "<unk>",
19
+ "lstrip": false,
20
+ "normalized": true,
21
+ "rstrip": false,
22
+ "single_word": false
23
+ }
24
+ }
tokenization_baichuan.py ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
+ # and OPT implementations in this library. It has been modified from its
6
+ # original forms to accommodate minor architectural differences compared
7
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
8
+ #
9
+ # Licensed under the Apache License, Version 2.0 (the "License");
10
+ # you may not use this file except in compliance with the License.
11
+ # You may obtain a copy of the License at
12
+ #
13
+ # http://www.apache.org/licenses/LICENSE-2.0
14
+ #
15
+ # Unless required by applicable law or agreed to in writing, software
16
+ # distributed under the License is distributed on an "AS IS" BASIS,
17
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
+ # See the License for the specific language governing permissions and
19
+ # limitations under the License.
20
+
21
+ import os
22
+ from shutil import copyfile
23
+ from typing import Any, Dict, List, Optional, Tuple
24
+
25
+ import sentencepiece as spm
26
+
27
+ from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
28
+ from transformers.utils import logging
29
+
30
+
31
+ logger = logging.get_logger(__name__)
32
+
33
+ VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
34
+
35
+ PRETRAINED_VOCAB_FILES_MAP = {
36
+ "vocab_file": {},
37
+ "tokenizer_file": {},
38
+ }
39
+ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {}
40
+
41
+
42
+ class BaiChuanTokenizer(PreTrainedTokenizer):
43
+ """
44
+ Construct a BaiChuan tokenizer. Based on byte-level Byte-Pair-Encoding.
45
+
46
+ Args:
47
+ vocab_file (`str`):
48
+ Path to the vocabulary file.
49
+ """
50
+
51
+ vocab_files_names = VOCAB_FILES_NAMES
52
+ pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
53
+ max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
54
+ model_input_names = ["input_ids", "attention_mask"]
55
+
56
+ def __init__(
57
+ self,
58
+ vocab_file,
59
+ unk_token="<unk>",
60
+ bos_token="<s>",
61
+ eos_token="</s>",
62
+ pad_token=None,
63
+ sp_model_kwargs: Optional[Dict[str, Any]] = None,
64
+ add_bos_token=True,
65
+ add_eos_token=False,
66
+ clean_up_tokenization_spaces=False,
67
+ **kwargs,
68
+ ):
69
+ self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
70
+ bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
71
+ eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
72
+ unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
73
+ pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
74
+ super().__init__(
75
+ bos_token=bos_token,
76
+ eos_token=eos_token,
77
+ unk_token=unk_token,
78
+ pad_token=pad_token,
79
+ add_bos_token=add_bos_token,
80
+ add_eos_token=add_eos_token,
81
+ sp_model_kwargs=self.sp_model_kwargs,
82
+ clean_up_tokenization_spaces=clean_up_tokenization_spaces,
83
+ **kwargs,
84
+ )
85
+ self.vocab_file = vocab_file
86
+ self.add_bos_token = add_bos_token
87
+ self.add_eos_token = add_eos_token
88
+ self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
89
+ self.sp_model.Load(vocab_file)
90
+
91
+ def __getstate__(self):
92
+ state = self.__dict__.copy()
93
+ state["sp_model"] = None
94
+ return state
95
+
96
+ def __setstate__(self, d):
97
+ self.__dict__ = d
98
+ self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
99
+ self.sp_model.Load(self.vocab_file)
100
+
101
+ @property
102
+ def vocab_size(self):
103
+ """Returns vocab size"""
104
+ return self.sp_model.get_piece_size()
105
+
106
+ def get_vocab(self):
107
+ """Returns vocab as a dict"""
108
+ vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
109
+ vocab.update(self.added_tokens_encoder)
110
+ return vocab
111
+
112
+ def _tokenize(self, text):
113
+ """Returns a tokenized string."""
114
+ return self.sp_model.encode(text, out_type=str)
115
+
116
+ def _convert_token_to_id(self, token):
117
+ """Converts a token (str) in an id using the vocab."""
118
+ return self.sp_model.piece_to_id(token)
119
+
120
+ def _convert_id_to_token(self, index):
121
+ """Converts an index (integer) in a token (str) using the vocab."""
122
+ token = self.sp_model.IdToPiece(index)
123
+ return token
124
+
125
+ def convert_tokens_to_string(self, tokens):
126
+ """Converts a sequence of tokens (string) in a single string."""
127
+ current_sub_tokens = []
128
+ out_string = ""
129
+ prev_is_special = False
130
+ for i, token in enumerate(tokens):
131
+ # make sure that special tokens are not decoded using sentencepiece model
132
+ if token in self.all_special_tokens:
133
+ if not prev_is_special and i != 0:
134
+ out_string += " "
135
+ out_string += self.sp_model.decode(current_sub_tokens) + token
136
+ prev_is_special = True
137
+ current_sub_tokens = []
138
+ else:
139
+ current_sub_tokens.append(token)
140
+ prev_is_special = False
141
+ out_string += self.sp_model.decode(current_sub_tokens)
142
+ return out_string
143
+
144
+ def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
145
+ """
146
+ Save the vocabulary and special tokens file to a directory.
147
+
148
+ Args:
149
+ save_directory (`str`):
150
+ The directory in which to save the vocabulary.
151
+
152
+ Returns:
153
+ `Tuple(str)`: Paths to the files saved.
154
+ """
155
+ if not os.path.isdir(save_directory):
156
+ logger.error(f"Vocabulary path ({save_directory}) should be a directory")
157
+ return
158
+ out_vocab_file = os.path.join(
159
+ save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
160
+ )
161
+
162
+ if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
163
+ copyfile(self.vocab_file, out_vocab_file)
164
+ elif not os.path.isfile(self.vocab_file):
165
+ with open(out_vocab_file, "wb") as fi:
166
+ content_spiece_model = self.sp_model.serialized_model_proto()
167
+ fi.write(content_spiece_model)
168
+
169
+ return (out_vocab_file,)
170
+
171
+ def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
172
+ bos_token_id = [self.bos_token_id] if self.add_bos_token else []
173
+ eos_token_id = [self.eos_token_id] if self.add_eos_token else []
174
+
175
+ output = bos_token_id + token_ids_0 + eos_token_id
176
+
177
+ if token_ids_1 is not None:
178
+ output = output + bos_token_id + token_ids_1 + eos_token_id
179
+
180
+ return output
181
+
182
+ def get_special_tokens_mask(
183
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
184
+ ) -> List[int]:
185
+ """
186
+ Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
187
+ special tokens using the tokenizer `prepare_for_model` method.
188
+
189
+ Args:
190
+ token_ids_0 (`List[int]`):
191
+ List of IDs.
192
+ token_ids_1 (`List[int]`, *optional*):
193
+ Optional second list of IDs for sequence pairs.
194
+ already_has_special_tokens (`bool`, *optional*, defaults to `False`):
195
+ Whether or not the token list is already formatted with special tokens for the model.
196
+
197
+ Returns:
198
+ `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
199
+ """
200
+ if already_has_special_tokens:
201
+ return super().get_special_tokens_mask(
202
+ token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
203
+ )
204
+
205
+ bos_token_id = [1] if self.add_bos_token else []
206
+ eos_token_id = [1] if self.add_eos_token else []
207
+
208
+ if token_ids_1 is None:
209
+ return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
210
+ return (
211
+ bos_token_id
212
+ + ([0] * len(token_ids_0))
213
+ + eos_token_id
214
+ + bos_token_id
215
+ + ([0] * len(token_ids_1))
216
+ + eos_token_id
217
+ )
218
+
219
+ def create_token_type_ids_from_sequences(
220
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
221
+ ) -> List[int]:
222
+ """
223
+ Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
224
+ sequence pair mask has the following format:
225
+
226
+ ```
227
+ 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
228
+ | first sequence | second sequence |
229
+ ```
230
+
231
+ if token_ids_1 is None, only returns the first portion of the mask (0s).
232
+
233
+ Args:
234
+ token_ids_0 (`List[int]`):
235
+ List of ids.
236
+ token_ids_1 (`List[int]`, *optional*):
237
+ Optional second list of IDs for sequence pairs.
238
+
239
+ Returns:
240
+ `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
241
+ """
242
+ bos_token_id = [self.bos_token_id] if self.add_bos_token else []
243
+ eos_token_id = [self.eos_token_id] if self.add_eos_token else []
244
+
245
+ output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
246
+
247
+ if token_ids_1 is not None:
248
+ output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
249
+
250
+ return output
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:25ea06a6726b6d518808fbc0d0d6bc3f0e899ea9e1656b2ab9716fa674e024f4
3
+ size 2095253
tokenizer_config.json ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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": false
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": false
26
+ },
27
+ "model_max_length": 1000000000000000019884624838656,
28
+ "pad_token": null,
29
+ "padding_side": "left",
30
+ "sp_model_kwargs": {},
31
+ "tokenizer_class": "BaiChuanTokenizer",
32
+ "unk_token": {
33
+ "__type": "AddedToken",
34
+ "content": "<unk>",
35
+ "lstrip": false,
36
+ "normalized": true,
37
+ "rstrip": false,
38
+ "single_word": false
39
+ }
40
+ }