Azure99 commited on
Commit
51ecf66
1 Parent(s): a2a5612

Upload model

Browse files
config.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "_name_or_path": "/mnt/output/baichuan2-7b_4-stage1",
4
+ "architectures": [
5
+ "BaichuanForCausalLM"
6
+ ],
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_baichuan.BaichuanConfig",
9
+ "AutoModel": "modeling_baichuan.BaichuanForCausalLM",
10
+ "AutoModelForCausalLM": "modeling_baichuan.BaichuanForCausalLM"
11
+ },
12
+ "bos_token_id": 1,
13
+ "eos_token_id": 2,
14
+ "hidden_act": "silu",
15
+ "hidden_size": 4096,
16
+ "initializer_range": 0.02,
17
+ "intermediate_size": 11008,
18
+ "max_position_embeddings": 4096,
19
+ "model_max_length": 4096,
20
+ "model_type": "baichuan",
21
+ "num_attention_heads": 32,
22
+ "num_hidden_layers": 32,
23
+ "pad_token_id": 0,
24
+ "rms_norm_eps": 1e-06,
25
+ "tie_word_embeddings": false,
26
+ "torch_dtype": "bfloat16",
27
+ "transformers_version": "4.33.2",
28
+ "use_cache": true,
29
+ "vocab_size": 125696
30
+ }
configuration_baichuan.py ADDED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 Baichuan Inc. All Rights Reserved.
2
+
3
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
4
+ #
5
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
6
+ # and OPT implementations in this library. It has been modified from its
7
+ # original forms to accommodate minor architectural differences compared
8
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
9
+ #
10
+ # Licensed under the Apache License, Version 2.0 (the "License");
11
+ # you may not use this file except in compliance with the License.
12
+ # You may obtain a copy of the License at
13
+ #
14
+ # http://www.apache.org/licenses/LICENSE-2.0
15
+ #
16
+ # Unless required by applicable law or agreed to in writing, software
17
+ # distributed under the License is distributed on an "AS IS" BASIS,
18
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
19
+ # See the License for the specific language governing permissions and
20
+ # limitations under the License.
21
+
22
+ from transformers.configuration_utils import PretrainedConfig
23
+ from transformers.utils import logging
24
+
25
+
26
+ logger = logging.get_logger(__name__)
27
+
28
+
29
+ class BaichuanConfig(PretrainedConfig):
30
+ model_type = "baichuan"
31
+ keys_to_ignore_at_inference = ["past_key_values"]
32
+
33
+ def __init__(
34
+ self,
35
+ vocab_size=125696,
36
+ hidden_size=4096,
37
+ intermediate_size=11008,
38
+ num_hidden_layers=32,
39
+ num_attention_heads=32,
40
+ hidden_act="silu",
41
+ max_position_embeddings=4096,
42
+ initializer_range=0.02,
43
+ rms_norm_eps=1e-6,
44
+ use_cache=True,
45
+ pad_token_id=0,
46
+ bos_token_id=1,
47
+ eos_token_id=2,
48
+ tie_word_embeddings=False,
49
+ **kwargs,
50
+ ):
51
+ self.vocab_size = vocab_size
52
+ self.max_position_embeddings = max_position_embeddings
53
+ self.hidden_size = hidden_size
54
+ self.intermediate_size = intermediate_size
55
+ self.num_hidden_layers = num_hidden_layers
56
+ self.num_attention_heads = num_attention_heads
57
+ self.hidden_act = hidden_act
58
+ self.initializer_range = initializer_range
59
+ self.rms_norm_eps = rms_norm_eps
60
+ self.use_cache = use_cache
61
+ super().__init__(
62
+ pad_token_id=pad_token_id,
63
+ bos_token_id=bos_token_id,
64
+ eos_token_id=eos_token_id,
65
+ tie_word_embeddings=tie_word_embeddings,
66
+ **kwargs,
67
+ )
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": "4.33.2"
7
+ }
generation_utils.py ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List
2
+ from queue import Queue
3
+
4
+ import torch
5
+
6
+
7
+ def build_chat_input(model, tokenizer, messages: List[dict], max_new_tokens: int=0):
8
+ def _parse_messages(messages, split_role="user"):
9
+ system, rounds = "", []
10
+ round = []
11
+ for i, message in enumerate(messages):
12
+ if message["role"] == "system":
13
+ assert i == 0
14
+ system = message["content"]
15
+ continue
16
+ if message["role"] == split_role and round:
17
+ rounds.append(round)
18
+ round = []
19
+ round.append(message)
20
+ if round:
21
+ rounds.append(round)
22
+ return system, rounds
23
+
24
+ max_new_tokens = max_new_tokens or model.generation_config.max_new_tokens
25
+ max_input_tokens = model.config.model_max_length - max_new_tokens
26
+ system, rounds = _parse_messages(messages, split_role="user")
27
+ system_tokens = tokenizer.encode(system)
28
+ max_history_tokens = max_input_tokens - len(system_tokens)
29
+
30
+ history_tokens = []
31
+ for round in rounds[::-1]:
32
+ round_tokens = []
33
+ for message in round:
34
+ if message["role"] == "user":
35
+ round_tokens.append(model.generation_config.user_token_id)
36
+ else:
37
+ round_tokens.append(model.generation_config.assistant_token_id)
38
+ round_tokens.extend(tokenizer.encode(message["content"]))
39
+ if len(history_tokens) == 0 or len(history_tokens) + len(round_tokens) <= max_history_tokens:
40
+ history_tokens = round_tokens + history_tokens # concat left
41
+ if len(history_tokens) < max_history_tokens:
42
+ continue
43
+ break
44
+
45
+ input_tokens = system_tokens + history_tokens
46
+ if messages[-1]["role"] != "assistant":
47
+ input_tokens.append(model.generation_config.assistant_token_id)
48
+ input_tokens = input_tokens[-max_input_tokens:] # truncate left
49
+ return torch.LongTensor([input_tokens]).to(model.device)
50
+
51
+
52
+ class TextIterStreamer:
53
+ def __init__(self, tokenizer, skip_prompt=False, skip_special_tokens=False):
54
+ self.tokenizer = tokenizer
55
+ self.skip_prompt = skip_prompt
56
+ self.skip_special_tokens = skip_special_tokens
57
+ self.tokens = []
58
+ self.text_queue = Queue()
59
+ self.next_tokens_are_prompt = True
60
+
61
+ def put(self, value):
62
+ if self.skip_prompt and self.next_tokens_are_prompt:
63
+ self.next_tokens_are_prompt = False
64
+ else:
65
+ if len(value.shape) > 1:
66
+ value = value[0]
67
+ self.tokens.extend(value.tolist())
68
+ self.text_queue.put(
69
+ self.tokenizer.decode(self.tokens, skip_special_tokens=self.skip_special_tokens))
70
+
71
+ def end(self):
72
+ self.text_queue.put(None)
73
+
74
+ def __iter__(self):
75
+ return self
76
+
77
+ def __next__(self):
78
+ value = self.text_queue.get()
79
+ if value is None:
80
+ raise StopIteration()
81
+ else:
82
+ return value
83
+
modeling_baichuan.py ADDED
@@ -0,0 +1,781 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 Baichuan Inc. All Rights Reserved.
2
+
3
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
4
+ #
5
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
6
+ # and OPT implementations in this library. It has been modified from its
7
+ # original forms to accommodate minor architectural differences compared
8
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
9
+ #
10
+ # Licensed under the Apache License, Version 2.0 (the "License");
11
+ # you may not use this file except in compliance with the License.
12
+ # You may obtain a copy of the License at
13
+ #
14
+ # http://www.apache.org/licenses/LICENSE-2.0
15
+ #
16
+ # Unless required by applicable law or agreed to in writing, software
17
+ # distributed under the License is distributed on an "AS IS" BASIS,
18
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
19
+ # See the License for the specific language governing permissions and
20
+ # limitations under the License.
21
+
22
+
23
+ from .configuration_baichuan import BaichuanConfig
24
+ from .generation_utils import build_chat_input, TextIterStreamer
25
+
26
+ import math
27
+ from typing import List, Optional, Tuple, Union
28
+ from threading import Thread
29
+
30
+ import torch
31
+ import torch.utils.checkpoint
32
+ from torch import nn
33
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
34
+ from torch.nn import functional as F
35
+ from transformers import PreTrainedModel, PretrainedConfig
36
+ from transformers.activations import ACT2FN
37
+ from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
38
+ from transformers.generation.utils import GenerationConfig
39
+ from transformers.utils import logging, ContextManagers
40
+
41
+ import os
42
+ from contextlib import contextmanager
43
+ logger = logging.get_logger(__name__)
44
+
45
+ try:
46
+ from xformers import ops as xops
47
+ except ImportError:
48
+ xops = None
49
+ logger.warning(
50
+ "Xformers is not installed correctly. If you want to use memory_efficient_attention to accelerate training use the following command to install Xformers\npip install xformers."
51
+ )
52
+
53
+
54
+ # Copied from transformers.models.bart.modeling_bart._make_causal_mask
55
+ def _make_causal_mask(
56
+ input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
57
+ ):
58
+ """
59
+ Make causal mask used for bi-directional self-attention.
60
+ """
61
+ bsz, tgt_len = input_ids_shape
62
+ mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
63
+ mask_cond = torch.arange(mask.size(-1), device=device)
64
+ mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
65
+ mask = mask.to(dtype)
66
+
67
+ if past_key_values_length > 0:
68
+ mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
69
+ return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
70
+
71
+ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
72
+ """
73
+ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
74
+ """
75
+ if len(mask.size()) == 3:
76
+ bsz, src_len, _ = mask.size()
77
+ tgt_len = tgt_len if tgt_len is not None else src_len
78
+ expanded_mask = mask[:,None,:,:].expand(bsz, 1, tgt_len, src_len).to(dtype)
79
+ else:
80
+ bsz, src_len = mask.size()
81
+ tgt_len = tgt_len if tgt_len is not None else src_len
82
+ expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
83
+
84
+ inverted_mask = 1.0 - expanded_mask
85
+
86
+ return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
87
+
88
+
89
+ class RMSNorm(nn.Module):
90
+ def __init__(self, hidden_size, eps=1e-6):
91
+ """
92
+ RMSNorm is equivalent to T5LayerNorm
93
+ """
94
+ super().__init__()
95
+ self.weight = nn.Parameter(torch.ones(hidden_size))
96
+ self.variance_epsilon = eps
97
+
98
+ def forward(self, hidden_states):
99
+ variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
100
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
101
+
102
+ # convert into half-precision if necessary
103
+ if self.weight.dtype in [torch.float16, torch.bfloat16]:
104
+ hidden_states = hidden_states.to(self.weight.dtype)
105
+
106
+ return self.weight * hidden_states
107
+
108
+
109
+ class RotaryEmbedding(torch.nn.Module):
110
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
111
+ super().__init__()
112
+ self.inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
113
+ self.max_seq_len_cached = max_position_embeddings
114
+ t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=torch.float32)
115
+ freqs = torch.outer(t, self.inv_freq)
116
+ emb = torch.cat((freqs, freqs), dim=-1)
117
+ self.cos_cached = emb.cos()[None, None, :, :].to(torch.float32)
118
+ self.sin_cached = emb.sin()[None, None, :, :].to(torch.float32)
119
+ def forward(self, x, seq_len=None):
120
+ # x: [bs, num_attention_heads, seq_len, head_size]
121
+ # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
122
+ if seq_len > self.max_seq_len_cached:
123
+ self.max_seq_len_cached = seq_len
124
+ t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=torch.float32)
125
+ freqs = torch.outer(t, self.inv_freq)
126
+ emb = torch.cat((freqs, freqs), dim=-1)
127
+ self.cos_cached = emb.cos()[None, None, :, :].to(torch.float32).to(x.device)
128
+ self.sin_cached = emb.sin()[None, None, :, :].to(torch.float32).to(x.device)
129
+ elif self.cos_cached.device != x.device:
130
+ self.cos_cached = self.cos_cached.to(x.device)
131
+ self.sin_cached = self.sin_cached.to(x.device)
132
+ return (
133
+ self.cos_cached[:, :, :seq_len, ...],
134
+ self.sin_cached[:, :, :seq_len, ...],
135
+ )
136
+
137
+
138
+ def rotate_half(x):
139
+ """Rotates half the hidden dims of the input."""
140
+ x1 = x[..., : x.shape[-1] // 2]
141
+ x2 = x[..., x.shape[-1] // 2:]
142
+ return torch.cat((-x2, x1), dim=-1)
143
+
144
+
145
+ def apply_rotary_pos_emb(q, k, cos_, sin_, position_ids):
146
+ cos = cos_.squeeze(1).squeeze(0) # [seq_len, dim]
147
+ sin = sin_.squeeze(1).squeeze(0) # [seq_len, dim]
148
+ cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
149
+ sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
150
+ q_embed = (q.float() * cos) + (rotate_half(q.float()) * sin)
151
+ k_embed = (k.float() * cos) + (rotate_half(k.float()) * sin)
152
+ return q_embed.to(q.dtype), k_embed.to(k.dtype)
153
+
154
+
155
+ class MLP(nn.Module):
156
+ def __init__(
157
+ self,
158
+ hidden_size: int,
159
+ intermediate_size: int,
160
+ hidden_act: str,
161
+ ):
162
+ super().__init__()
163
+ self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
164
+ self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
165
+ self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
166
+ self.act_fn = ACT2FN[hidden_act]
167
+
168
+ def forward(self, x):
169
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
170
+
171
+
172
+ class Attention(nn.Module):
173
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
174
+ def __init__(self, config: BaichuanConfig):
175
+ super().__init__()
176
+ self.config = config
177
+ self.hidden_size = config.hidden_size
178
+ self.num_heads = config.num_attention_heads
179
+ self.head_dim = self.hidden_size // self.num_heads
180
+ self.max_position_embeddings = config.max_position_embeddings
181
+
182
+ if (self.head_dim * self.num_heads) != self.hidden_size:
183
+ raise ValueError(
184
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
185
+ f" and `num_heads`: {self.num_heads})."
186
+ )
187
+ self.W_pack = nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=False)
188
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
189
+ self.rotary_emb = RotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
190
+
191
+ def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
192
+ return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
193
+
194
+ def forward(
195
+ self,
196
+ hidden_states: torch.Tensor,
197
+ attention_mask: Optional[torch.Tensor] = None,
198
+ position_ids: Optional[torch.LongTensor] = None,
199
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
200
+ output_attentions: bool = False,
201
+ use_cache: bool = False,
202
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
203
+ bsz, q_len, _ = hidden_states.size()
204
+
205
+ proj = self.W_pack(hidden_states)
206
+ proj = proj.unflatten(-1, (3, self.hidden_size)).unsqueeze(0).transpose(0, -2).squeeze(-2)
207
+ query_states = proj[0].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
208
+ key_states = proj[1].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
209
+ value_states = proj[2].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
210
+
211
+ kv_seq_len = key_states.shape[-2]
212
+ if past_key_value is not None:
213
+ kv_seq_len += past_key_value[0].shape[-2]
214
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
215
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
216
+ # [bsz, nh, t, hd]
217
+
218
+ if past_key_value is not None:
219
+ # reuse k, v, self_attention
220
+ key_states = torch.cat([past_key_value[0], key_states], dim=2)
221
+ value_states = torch.cat([past_key_value[1], value_states], dim=2)
222
+
223
+ past_key_value = (key_states, value_states) if use_cache else None
224
+ if xops is not None and self.training:
225
+ attn_weights = None
226
+ query_states = query_states.transpose(1, 2)
227
+ key_states = key_states.transpose(1, 2)
228
+ value_states = value_states.transpose(1, 2)
229
+ attn_output = xops.memory_efficient_attention(
230
+ query_states, key_states, value_states, attn_bias=xops.LowerTriangularMask()
231
+ )
232
+ else:
233
+ with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=True):
234
+ attn_output = F.scaled_dot_product_attention(query_states, key_states, value_states, attn_mask = attention_mask)
235
+ attn_output = attn_output.transpose(1, 2)
236
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
237
+ attn_output = self.o_proj(attn_output)
238
+
239
+ if not output_attentions:
240
+ attn_weights = None
241
+
242
+ return attn_output, attn_weights, past_key_value
243
+
244
+
245
+ class DecoderLayer(nn.Module):
246
+ def __init__(self, config: BaichuanConfig):
247
+ super().__init__()
248
+ self.hidden_size = config.hidden_size
249
+ self.self_attn = Attention(config=config)
250
+ self.mlp = MLP(
251
+ hidden_size=self.hidden_size,
252
+ intermediate_size=config.intermediate_size,
253
+ hidden_act=config.hidden_act,
254
+ )
255
+ self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
256
+ self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
257
+
258
+ def forward(
259
+ self,
260
+ hidden_states: torch.Tensor,
261
+ attention_mask: Optional[torch.Tensor] = None,
262
+ position_ids: Optional[torch.LongTensor] = None,
263
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
264
+ output_attentions: Optional[bool] = False,
265
+ use_cache: Optional[bool] = False,
266
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
267
+
268
+ residual = hidden_states
269
+
270
+ hidden_states = self.input_layernorm(hidden_states)
271
+
272
+ # Self Attention
273
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
274
+ hidden_states=hidden_states,
275
+ attention_mask=attention_mask,
276
+ position_ids=position_ids,
277
+ past_key_value=past_key_value,
278
+ output_attentions=output_attentions,
279
+ use_cache=use_cache,
280
+ )
281
+ hidden_states = residual + hidden_states
282
+
283
+ # Fully Connected
284
+ residual = hidden_states
285
+ hidden_states = self.post_attention_layernorm(hidden_states)
286
+ hidden_states = self.mlp(hidden_states)
287
+ hidden_states = residual + hidden_states
288
+
289
+ outputs = (hidden_states,)
290
+
291
+ if output_attentions:
292
+ outputs += (self_attn_weights,)
293
+
294
+ if use_cache:
295
+ outputs += (present_key_value,)
296
+
297
+ return outputs
298
+
299
+
300
+ class BaichuanPreTrainedModel(PreTrainedModel):
301
+ config_class = BaichuanConfig
302
+ base_model_prefix = "model"
303
+ supports_gradient_checkpointing = True
304
+ _no_split_modules = ["DecoderLayer"]
305
+ _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
306
+
307
+ def _init_weights(self, module):
308
+ std = self.config.initializer_range
309
+ if isinstance(module, nn.Linear):
310
+ module.weight.data.normal_(mean=0.0, std=std)
311
+ if module.bias is not None:
312
+ module.bias.data.zero_()
313
+ elif isinstance(module, nn.Embedding):
314
+ module.weight.data.normal_(mean=0.0, std=std)
315
+ if module.padding_idx is not None:
316
+ module.weight.data[module.padding_idx].zero_()
317
+
318
+ def _set_gradient_checkpointing(self, module, value=False):
319
+ if isinstance(module, BaichuanModel):
320
+ module.gradient_checkpointing = value
321
+
322
+
323
+ class BaichuanModel(BaichuanPreTrainedModel):
324
+ def __init__(self, config: BaichuanConfig):
325
+ super().__init__(config)
326
+ self.padding_idx = config.pad_token_id
327
+ self.vocab_size = config.vocab_size
328
+
329
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
330
+ self.layers = nn.ModuleList([DecoderLayer(config) for _ in range(config.num_hidden_layers)])
331
+ self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
332
+
333
+ self.gradient_checkpointing = False
334
+ # Initialize weights and apply final processing
335
+ self.post_init()
336
+
337
+ def get_input_embeddings(self):
338
+ return self.embed_tokens
339
+
340
+ def set_input_embeddings(self, value):
341
+ self.embed_tokens = value
342
+
343
+ # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
344
+ def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
345
+ # create causal mask
346
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
347
+ combined_attention_mask = None
348
+ if input_shape[-1] > 1:
349
+ combined_attention_mask = _make_causal_mask(
350
+ input_shape,
351
+ inputs_embeds.dtype,
352
+ device=inputs_embeds.device,
353
+ past_key_values_length=past_key_values_length,
354
+ )
355
+
356
+ if attention_mask is not None:
357
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
358
+ expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
359
+ inputs_embeds.device
360
+ )
361
+ combined_attention_mask = (
362
+ expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
363
+ )
364
+
365
+ return combined_attention_mask
366
+
367
+ def forward(
368
+ self,
369
+ input_ids: torch.LongTensor = None,
370
+ attention_mask: Optional[torch.Tensor] = None,
371
+ position_ids: Optional[torch.LongTensor] = None,
372
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
373
+ inputs_embeds: Optional[torch.FloatTensor] = None,
374
+ use_cache: Optional[bool] = None,
375
+ output_attentions: Optional[bool] = None,
376
+ output_hidden_states: Optional[bool] = None,
377
+ return_dict: Optional[bool] = None,
378
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
379
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
380
+ output_hidden_states = (
381
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
382
+ )
383
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
384
+
385
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
386
+
387
+ # retrieve input_ids and inputs_embeds
388
+ if input_ids is not None and inputs_embeds is not None:
389
+ raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
390
+ elif input_ids is not None:
391
+ batch_size, seq_length = input_ids.shape
392
+ elif inputs_embeds is not None:
393
+ batch_size, seq_length, _ = inputs_embeds.shape
394
+ else:
395
+ raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
396
+
397
+ seq_length_with_past = seq_length
398
+ past_key_values_length = 0
399
+
400
+ if past_key_values is not None:
401
+ past_key_values_length = past_key_values[0][0].shape[2]
402
+ seq_length_with_past = seq_length_with_past + past_key_values_length
403
+
404
+ if position_ids is None:
405
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
406
+ position_ids = torch.arange(
407
+ past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
408
+ )
409
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
410
+ else:
411
+ position_ids = position_ids.view(-1, seq_length).long()
412
+
413
+ if inputs_embeds is None:
414
+ inputs_embeds = self.embed_tokens(input_ids)
415
+ # embed positions
416
+ if attention_mask is None:
417
+ attention_mask = torch.ones(
418
+ (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
419
+ )
420
+ attention_mask = self._prepare_decoder_attention_mask(
421
+ attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
422
+ )
423
+
424
+ hidden_states = inputs_embeds
425
+
426
+ if self.gradient_checkpointing and self.training:
427
+ if use_cache:
428
+ logger.warning_once(
429
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
430
+ )
431
+ use_cache = False
432
+
433
+ # decoder layers
434
+ all_hidden_states = () if output_hidden_states else None
435
+ all_self_attns = () if output_attentions else None
436
+ next_decoder_cache = () if use_cache else None
437
+
438
+ for idx, decoder_layer in enumerate(self.layers):
439
+ if output_hidden_states:
440
+ all_hidden_states += (hidden_states,)
441
+
442
+ past_key_value = past_key_values[idx] if past_key_values is not None else None
443
+
444
+ if self.gradient_checkpointing and self.training:
445
+
446
+ def create_custom_forward(module):
447
+ def custom_forward(*inputs):
448
+ # None for past_key_value
449
+ return module(*inputs, output_attentions, None)
450
+
451
+ return custom_forward
452
+
453
+ layer_outputs = torch.utils.checkpoint.checkpoint(
454
+ create_custom_forward(decoder_layer),
455
+ hidden_states,
456
+ attention_mask,
457
+ position_ids,
458
+ None,
459
+ )
460
+ else:
461
+ layer_outputs = decoder_layer(
462
+ hidden_states,
463
+ attention_mask=attention_mask,
464
+ position_ids=position_ids,
465
+ past_key_value=past_key_value,
466
+ output_attentions=output_attentions,
467
+ use_cache=use_cache,
468
+ )
469
+
470
+ hidden_states = layer_outputs[0]
471
+
472
+ if use_cache:
473
+ next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
474
+
475
+ if output_attentions:
476
+ all_self_attns += (layer_outputs[1],)
477
+
478
+ hidden_states = self.norm(hidden_states)
479
+
480
+ # add hidden states from the last decoder layer
481
+ if output_hidden_states:
482
+ all_hidden_states += (hidden_states,)
483
+
484
+ next_cache = next_decoder_cache if use_cache else None
485
+ if not return_dict:
486
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
487
+ return BaseModelOutputWithPast(
488
+ last_hidden_state=hidden_states,
489
+ past_key_values=next_cache,
490
+ hidden_states=all_hidden_states,
491
+ attentions=all_self_attns,
492
+ )
493
+
494
+
495
+ class NormHead(nn.Module):
496
+ def __init__(self, hidden_size, vocab_size, bias=False):
497
+ super().__init__()
498
+ self.weight = nn.Parameter(torch.empty((vocab_size, hidden_size)))
499
+ nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))
500
+ self.first_flag = True
501
+
502
+ def forward(self, hidden_states):
503
+ if self.training:
504
+ norm_weight = nn.functional.normalize(self.weight)
505
+ elif self.first_flag:
506
+ self.first_flag = False
507
+ self.weight = nn.Parameter(nn.functional.normalize(self.weight))
508
+ norm_weight = self.weight
509
+ else:
510
+ norm_weight = self.weight
511
+ return nn.functional.linear(hidden_states, norm_weight)
512
+
513
+ _init_weights = True
514
+ @contextmanager
515
+ def no_init_weights(_enable=True):
516
+ global _init_weights
517
+ old_init_weights = _init_weights
518
+ if _enable:
519
+ _init_weights = False
520
+ try:
521
+ yield
522
+ finally:
523
+ _init_weights = old_init_weights
524
+
525
+ class BaichuanForCausalLM(BaichuanPreTrainedModel):
526
+ def __init__(self, config, *model_args, **model_kwargs):
527
+ super().__init__(config, *model_args, **model_kwargs)
528
+ self.model = BaichuanModel(config)
529
+
530
+ self.lm_head = NormHead(config.hidden_size, config.vocab_size, bias=False)
531
+ if hasattr(config, "quantization_config") and config.quantization_config['load_in_4bit']:
532
+ try:
533
+ from .quantizer import quantize_offline, init_model_weight_int4
534
+ except ImportError:
535
+ raise ImportError(f"Needs QLinear to run quantize.")
536
+ quantize_offline(self, 4)
537
+ # Initialize weights and apply final processing
538
+ self.post_init()
539
+
540
+ def get_input_embeddings(self):
541
+ return self.model.embed_tokens
542
+
543
+ def set_input_embeddings(self, value):
544
+ self.model.embed_tokens = value
545
+
546
+ def get_output_embeddings(self):
547
+ return self.lm_head
548
+
549
+ def set_output_embeddings(self, new_embeddings):
550
+ self.lm_head = new_embeddings
551
+
552
+ def set_decoder(self, decoder):
553
+ self.model = decoder
554
+
555
+ def get_decoder(self):
556
+ return self.model
557
+
558
+ @classmethod
559
+ def from_pretrained(
560
+ cls,
561
+ pretrained_model_name_or_path: Optional[Union[str, os.PathLike]],
562
+ *model_args,
563
+ config: Optional[Union[PretrainedConfig, str, os.PathLike]] = None,
564
+ cache_dir: Optional[Union[str, os.PathLike]] = None,
565
+ ignore_mismatched_sizes: bool = False,
566
+ force_download: bool = False,
567
+ local_files_only: bool = False,
568
+ token: Optional[Union[str, bool]] = None,
569
+ revision: str = "main",
570
+ use_safetensors: bool = None,
571
+ **kwargs,
572
+ ):
573
+ # Load config if we don't provide a configuration
574
+ if not isinstance(config, PretrainedConfig):
575
+ config_path = config if config is not None else pretrained_model_name_or_path
576
+ config, model_kwargs = cls.config_class.from_pretrained(
577
+ config_path,
578
+ cache_dir=cache_dir,
579
+ return_unused_kwargs=True,
580
+ force_download=force_download,
581
+ resume_download=False,
582
+ proxies=None,
583
+ local_files_only=local_files_only,
584
+ token=token,
585
+ revision=revision,
586
+ subfolder="",
587
+ _from_auto=False,
588
+ _from_pipeline=None,
589
+ **kwargs,
590
+ )
591
+ else:
592
+ model_kwargs = kwargs
593
+
594
+ if hasattr(config, "quantization_config") and config.quantization_config['load_in_4bit']:
595
+ try:
596
+ from .quantizer import init_model_weight_int4
597
+ from accelerate import init_empty_weights, dispatch_model, infer_auto_device_map
598
+ from accelerate.utils import CustomDtype
599
+ from accelerate.utils import get_balanced_memory
600
+ except ImportError:
601
+ raise ImportError(f"Needs import model weight init func to run quantize.")
602
+ # Instantiate model.
603
+ init_contexts = [no_init_weights(_enable=True)]
604
+ init_contexts.append(init_empty_weights())
605
+ with ContextManagers(init_contexts):
606
+ model = cls(config)
607
+
608
+ model_file = os.path.join(pretrained_model_name_or_path, 'pytorch_model.bin')
609
+ state_dict = torch.load(model_file, map_location="cpu")
610
+ model.is_quantized = True
611
+
612
+ device_map = kwargs.pop("device_map", None)
613
+ torch_dtype = kwargs.pop("torch_dtype", None)
614
+
615
+ kwargs = {"no_split_module_classes": model._no_split_modules}
616
+ target_dtype = CustomDtype.INT4
617
+ max_memory = get_balanced_memory(
618
+ model,
619
+ dtype=target_dtype,
620
+ low_zero=(device_map == "balanced_low_0"),
621
+ max_memory=None,
622
+ **kwargs,
623
+ )
624
+ kwargs["max_memory"] = max_memory
625
+
626
+ device_map = infer_auto_device_map(model, dtype=target_dtype, **kwargs)
627
+ model = init_model_weight_int4(config, model, state_dict)
628
+
629
+ # Set model in evaluation mode to deactivate DropOut modules by default
630
+ model.eval()
631
+ # If it is a model with generation capabilities, attempt to load the generation config
632
+ if model.can_generate():
633
+ try:
634
+ model.generation_config = GenerationConfig.from_pretrained(
635
+ pretrained_model_name_or_path,
636
+ cache_dir=cache_dir,
637
+ force_download=force_download,
638
+ resume_download=False,
639
+ proxies=None,
640
+ local_files_only=local_files_only,
641
+ token=token,
642
+ revision=revision,
643
+ subfolder="",
644
+ _from_auto=False,
645
+ _from_pipeline=None,
646
+ **kwargs,
647
+ )
648
+ except (OSError, TypeError):
649
+ logger.info(
650
+ "Generation config file not found, using a generation config created from the model config."
651
+ )
652
+ pass
653
+
654
+ if device_map is not None:
655
+ dispatch_model(model, device_map=device_map)
656
+
657
+ return model
658
+ return super(BaichuanForCausalLM, cls).from_pretrained(pretrained_model_name_or_path, *model_args,
659
+ config=config, cache_dir=cache_dir, ignore_mismatched_sizes=ignore_mismatched_sizes,
660
+ force_download=force_download, local_files_only=local_files_only, token=token, revision=revision,
661
+ use_safetensors=use_safetensors, **kwargs)
662
+
663
+ def forward(
664
+ self,
665
+ input_ids: torch.LongTensor = None,
666
+ attention_mask: Optional[torch.Tensor] = None,
667
+ position_ids: Optional[torch.LongTensor] = None,
668
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
669
+ inputs_embeds: Optional[torch.FloatTensor] = None,
670
+ labels: Optional[torch.LongTensor] = None,
671
+ use_cache: Optional[bool] = None,
672
+ output_attentions: Optional[bool] = None,
673
+ output_hidden_states: Optional[bool] = None,
674
+ return_dict: Optional[bool] = None,
675
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
676
+
677
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
678
+ output_hidden_states = (
679
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
680
+ )
681
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
682
+
683
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
684
+ outputs = self.model(
685
+ input_ids=input_ids,
686
+ attention_mask=attention_mask,
687
+ position_ids=position_ids,
688
+ past_key_values=past_key_values,
689
+ inputs_embeds=inputs_embeds,
690
+ use_cache=use_cache,
691
+ output_attentions=output_attentions,
692
+ output_hidden_states=output_hidden_states,
693
+ return_dict=return_dict,
694
+ )
695
+
696
+ hidden_states = outputs[0]
697
+ logits = self.lm_head(hidden_states)
698
+ loss = None
699
+ if labels is not None:
700
+ # Shift so that tokens < n predict n
701
+ shift_logits = logits[..., :-1, :].contiguous()
702
+ shift_labels = labels[..., 1:].contiguous()
703
+ # Flatten the tokens
704
+ loss_fct = CrossEntropyLoss()
705
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
706
+ shift_labels = shift_labels.view(-1)
707
+ # Enable model parallelism
708
+ shift_labels = shift_labels.to(shift_logits.device)
709
+ loss = loss_fct(shift_logits, shift_labels)
710
+
711
+ if not return_dict:
712
+ output = (logits,) + outputs[1:]
713
+ return (loss,) + output if loss is not None else output
714
+
715
+ return CausalLMOutputWithPast(
716
+ loss=loss,
717
+ logits=logits,
718
+ past_key_values=outputs.past_key_values,
719
+ hidden_states=outputs.hidden_states,
720
+ attentions=outputs.attentions,
721
+ )
722
+
723
+ def prepare_inputs_for_generation(
724
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
725
+ ):
726
+ if past_key_values:
727
+ input_ids = input_ids[:, -1:]
728
+
729
+ position_ids = kwargs.get("position_ids", None)
730
+ if attention_mask is not None and position_ids is None:
731
+ # create position_ids on the fly for batch generation
732
+ position_ids = attention_mask.long().cumsum(-1) - 1
733
+ position_ids.masked_fill_(attention_mask == 0, 1)
734
+ if past_key_values:
735
+ position_ids = position_ids[:, -1].unsqueeze(-1)
736
+
737
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
738
+ if inputs_embeds is not None and past_key_values is None:
739
+ model_inputs = {"inputs_embeds": inputs_embeds}
740
+ else:
741
+ model_inputs = {"input_ids": input_ids}
742
+
743
+ model_inputs.update(
744
+ {
745
+ "position_ids": position_ids,
746
+ "past_key_values": past_key_values,
747
+ "use_cache": kwargs.get("use_cache"),
748
+ "attention_mask": attention_mask,
749
+ }
750
+ )
751
+ return model_inputs
752
+
753
+ @staticmethod
754
+ def _reorder_cache(past_key_values, beam_idx):
755
+ reordered_past = ()
756
+ for layer_past in past_key_values:
757
+ reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
758
+ return reordered_past
759
+
760
+ def quantize(self, bits: int):
761
+ try:
762
+ from .quantizer import quantize_online
763
+ except ImportError:
764
+ raise ImportError(f"Needs QLinear to run quantize.")
765
+ return quantize_online(self, bits)
766
+
767
+ def chat(self, tokenizer, messages: List[dict], stream=False,
768
+ generation_config: Optional[GenerationConfig]=None):
769
+ generation_config = generation_config or self.generation_config
770
+ input_ids = build_chat_input(self, tokenizer, messages, generation_config.max_new_tokens)
771
+ if stream:
772
+ streamer = TextIterStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
773
+ Thread(target=self.generate, kwargs=dict(
774
+ inputs=input_ids, streamer=streamer,
775
+ generation_config=generation_config,
776
+ )).start()
777
+ return streamer
778
+ else:
779
+ outputs = self.generate(input_ids, generation_config=generation_config)
780
+ response = tokenizer.decode(outputs[0][len(input_ids[0]):], skip_special_tokens=True)
781
+ return response
pytorch_model-00001-of-00002.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:734d25d350a1cf43100fb6538e0980822e7253d37ffd9964700828091f8a52a1
3
+ size 9934622796
pytorch_model-00002-of-00002.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cf4ba7632cfd333e85795d4235bd5f72bef84481340be81a987041888929b6bc
3
+ size 5077401163
pytorch_model.bin.index.json ADDED
@@ -0,0 +1,234 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 15011946496
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.1.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
16
+ "model.layers.1.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
17
+ "model.layers.1.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
18
+ "model.layers.1.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
19
+ "model.layers.1.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
20
+ "model.layers.1.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
21
+ "model.layers.1.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
22
+ "model.layers.10.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
23
+ "model.layers.10.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
24
+ "model.layers.10.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
25
+ "model.layers.10.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
26
+ "model.layers.10.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
27
+ "model.layers.10.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
28
+ "model.layers.10.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
29
+ "model.layers.11.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
30
+ "model.layers.11.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
31
+ "model.layers.11.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
32
+ "model.layers.11.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
33
+ "model.layers.11.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
34
+ "model.layers.11.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
35
+ "model.layers.11.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
36
+ "model.layers.12.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
37
+ "model.layers.12.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
38
+ "model.layers.12.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
39
+ "model.layers.12.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
40
+ "model.layers.12.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
41
+ "model.layers.12.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
42
+ "model.layers.12.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
43
+ "model.layers.13.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
44
+ "model.layers.13.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
45
+ "model.layers.13.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
46
+ "model.layers.13.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
47
+ "model.layers.13.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
48
+ "model.layers.13.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
49
+ "model.layers.13.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
50
+ "model.layers.14.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
51
+ "model.layers.14.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
52
+ "model.layers.14.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
53
+ "model.layers.14.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
54
+ "model.layers.14.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
55
+ "model.layers.14.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
56
+ "model.layers.14.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
57
+ "model.layers.15.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
58
+ "model.layers.15.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
59
+ "model.layers.15.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
60
+ "model.layers.15.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
61
+ "model.layers.15.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
62
+ "model.layers.15.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
63
+ "model.layers.15.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
64
+ "model.layers.16.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
65
+ "model.layers.16.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
66
+ "model.layers.16.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
67
+ "model.layers.16.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
68
+ "model.layers.16.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
69
+ "model.layers.16.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
70
+ "model.layers.16.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
71
+ "model.layers.17.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
72
+ "model.layers.17.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
73
+ "model.layers.17.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
74
+ "model.layers.17.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
75
+ "model.layers.17.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
76
+ "model.layers.17.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
77
+ "model.layers.17.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
78
+ "model.layers.18.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
79
+ "model.layers.18.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
80
+ "model.layers.18.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
81
+ "model.layers.18.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
82
+ "model.layers.18.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
83
+ "model.layers.18.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
84
+ "model.layers.18.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
85
+ "model.layers.19.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
86
+ "model.layers.19.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
87
+ "model.layers.19.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
88
+ "model.layers.19.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
89
+ "model.layers.19.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
90
+ "model.layers.19.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
91
+ "model.layers.19.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
92
+ "model.layers.2.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
93
+ "model.layers.2.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
94
+ "model.layers.2.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
95
+ "model.layers.2.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
96
+ "model.layers.2.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
97
+ "model.layers.2.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
98
+ "model.layers.2.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
99
+ "model.layers.20.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
100
+ "model.layers.20.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
101
+ "model.layers.20.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
102
+ "model.layers.20.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
103
+ "model.layers.20.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
104
+ "model.layers.20.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
105
+ "model.layers.20.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
106
+ "model.layers.21.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
107
+ "model.layers.21.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
108
+ "model.layers.21.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
109
+ "model.layers.21.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
110
+ "model.layers.21.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
111
+ "model.layers.21.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
112
+ "model.layers.21.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
113
+ "model.layers.22.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
114
+ "model.layers.22.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
115
+ "model.layers.22.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
116
+ "model.layers.22.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
117
+ "model.layers.22.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
118
+ "model.layers.22.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
119
+ "model.layers.22.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
120
+ "model.layers.23.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
121
+ "model.layers.23.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
122
+ "model.layers.23.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
123
+ "model.layers.23.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
124
+ "model.layers.23.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
125
+ "model.layers.23.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
126
+ "model.layers.23.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
127
+ "model.layers.24.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
128
+ "model.layers.24.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
129
+ "model.layers.24.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
130
+ "model.layers.24.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
131
+ "model.layers.24.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
132
+ "model.layers.24.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
133
+ "model.layers.24.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
134
+ "model.layers.25.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
135
+ "model.layers.25.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
136
+ "model.layers.25.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
137
+ "model.layers.25.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
138
+ "model.layers.25.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
139
+ "model.layers.25.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
140
+ "model.layers.25.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
141
+ "model.layers.26.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
142
+ "model.layers.26.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
143
+ "model.layers.26.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
144
+ "model.layers.26.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
145
+ "model.layers.26.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
146
+ "model.layers.26.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
147
+ "model.layers.26.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
148
+ "model.layers.27.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
149
+ "model.layers.27.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
150
+ "model.layers.27.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
151
+ "model.layers.27.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
152
+ "model.layers.27.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
153
+ "model.layers.27.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
154
+ "model.layers.27.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
155
+ "model.layers.28.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
156
+ "model.layers.28.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
157
+ "model.layers.28.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
158
+ "model.layers.28.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
159
+ "model.layers.28.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
160
+ "model.layers.28.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
161
+ "model.layers.28.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
162
+ "model.layers.29.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
163
+ "model.layers.29.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
164
+ "model.layers.29.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
165
+ "model.layers.29.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
166
+ "model.layers.29.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
167
+ "model.layers.29.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
168
+ "model.layers.29.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
169
+ "model.layers.3.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
170
+ "model.layers.3.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
171
+ "model.layers.3.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
172
+ "model.layers.3.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
173
+ "model.layers.3.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
174
+ "model.layers.3.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
175
+ "model.layers.3.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
176
+ "model.layers.30.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
177
+ "model.layers.30.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
178
+ "model.layers.30.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
179
+ "model.layers.30.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
180
+ "model.layers.30.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
181
+ "model.layers.30.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
182
+ "model.layers.30.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
183
+ "model.layers.31.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
184
+ "model.layers.31.mlp.down_proj.weight": "pytorch_model-00002-of-00002.bin",
185
+ "model.layers.31.mlp.gate_proj.weight": "pytorch_model-00002-of-00002.bin",
186
+ "model.layers.31.mlp.up_proj.weight": "pytorch_model-00002-of-00002.bin",
187
+ "model.layers.31.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
188
+ "model.layers.31.self_attn.W_pack.weight": "pytorch_model-00002-of-00002.bin",
189
+ "model.layers.31.self_attn.o_proj.weight": "pytorch_model-00002-of-00002.bin",
190
+ "model.layers.4.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
191
+ "model.layers.4.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
192
+ "model.layers.4.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
193
+ "model.layers.4.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
194
+ "model.layers.4.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
195
+ "model.layers.4.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
196
+ "model.layers.4.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
197
+ "model.layers.5.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
198
+ "model.layers.5.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
199
+ "model.layers.5.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
200
+ "model.layers.5.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
201
+ "model.layers.5.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
202
+ "model.layers.5.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
203
+ "model.layers.5.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
204
+ "model.layers.6.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
205
+ "model.layers.6.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
206
+ "model.layers.6.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
207
+ "model.layers.6.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
208
+ "model.layers.6.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
209
+ "model.layers.6.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
210
+ "model.layers.6.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
211
+ "model.layers.7.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
212
+ "model.layers.7.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
213
+ "model.layers.7.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
214
+ "model.layers.7.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
215
+ "model.layers.7.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
216
+ "model.layers.7.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
217
+ "model.layers.7.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
218
+ "model.layers.8.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
219
+ "model.layers.8.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
220
+ "model.layers.8.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
221
+ "model.layers.8.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
222
+ "model.layers.8.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
223
+ "model.layers.8.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
224
+ "model.layers.8.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
225
+ "model.layers.9.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
226
+ "model.layers.9.mlp.down_proj.weight": "pytorch_model-00001-of-00002.bin",
227
+ "model.layers.9.mlp.gate_proj.weight": "pytorch_model-00001-of-00002.bin",
228
+ "model.layers.9.mlp.up_proj.weight": "pytorch_model-00001-of-00002.bin",
229
+ "model.layers.9.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
230
+ "model.layers.9.self_attn.W_pack.weight": "pytorch_model-00001-of-00002.bin",
231
+ "model.layers.9.self_attn.o_proj.weight": "pytorch_model-00001-of-00002.bin",
232
+ "model.norm.weight": "pytorch_model-00002-of-00002.bin"
233
+ }
234
+ }
quantizer.py ADDED
@@ -0,0 +1,210 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import bitsandbytes as bnb
2
+ from bitsandbytes.nn.modules import Params4bit, Int8Params
3
+ import torch
4
+
5
+ def Params4bitCuda(self, device):
6
+ self.data = self.data.cuda(device)
7
+ self.quant_state[0] = self.quant_state[0].cuda(device)
8
+ self.quant_state[4][0] = self.quant_state[4][0].cuda(device)
9
+ self.quant_state[4][1][0] = self.quant_state[4][1][0].cuda(device)
10
+ self.quant_state[4][1][1] = self.quant_state[4][1][1].cuda(device)
11
+
12
+ self.quant_state[6] = self.quant_state[6].cuda(device)
13
+ return self
14
+
15
+ class Linear4bitOnline(torch.nn.Module):
16
+ def __init__(self, weight, bias, quant_type):
17
+ super().__init__()
18
+ self.weight = Params4bit(
19
+ weight.data, requires_grad=False, compress_statistics=True, quant_type=quant_type
20
+ )
21
+ self.compute_dtype = None
22
+ #self.weight.cuda(weight.device)
23
+ self.bias = bias
24
+
25
+ def forward(self, x: torch.Tensor):
26
+ # weights are cast automatically as Int8Params, but the bias has to be cast manually
27
+ if self.bias is not None and self.bias.dtype != x.dtype:
28
+ self.bias.data = self.bias.data.to(x.dtype)
29
+
30
+ if getattr(self.weight, "quant_state", None) is None:
31
+ print(
32
+ "FP4 quantization state not initialized. Please call .cuda() or .to(device) on the LinearFP4 layer first."
33
+ )
34
+ inp_dtype = x.dtype
35
+ if self.compute_dtype is not None:
36
+ x = x.to(self.compute_dtype)
37
+
38
+ bias = None if self.bias is None else self.bias.to(self.compute_dtype)
39
+ out = bnb.matmul_4bit(
40
+ x, self.weight.t(), bias=bias, quant_state=self.weight.quant_state
41
+ )
42
+
43
+ out = out.to(inp_dtype)
44
+
45
+ return out
46
+
47
+ class Linear8bitLtOnline(torch.nn.Module):
48
+ def __init__(
49
+ self,
50
+ weight,
51
+ bias,
52
+ has_fp16_weights=True,
53
+ memory_efficient_backward=False,
54
+ threshold=0.0,
55
+ index=None,
56
+ ):
57
+ super().__init__()
58
+ assert (
59
+ not memory_efficient_backward
60
+ ), "memory_efficient_backward is no longer required and the argument is deprecated in 0.37.0 and will be removed in 0.39.0"
61
+ self.state = bnb.MatmulLtState()
62
+ self.index = index
63
+
64
+ # Necessary for stacked layers
65
+ self.state.threshold = threshold
66
+ self.state.has_fp16_weights = has_fp16_weights
67
+ self.state.memory_efficient_backward = memory_efficient_backward
68
+ if threshold > 0.0 and not has_fp16_weights:
69
+ self.state.use_pool = True
70
+
71
+ self.weight = Int8Params(
72
+ weight.data,
73
+ has_fp16_weights=has_fp16_weights,
74
+ requires_grad=has_fp16_weights,
75
+ )
76
+ self.bias = bias
77
+
78
+ def init_8bit_state(self):
79
+ self.state.CB = self.weight.CB
80
+ self.state.SCB = self.weight.SCB
81
+ self.weight.CB = None
82
+ self.weight.SCB = None
83
+
84
+ def forward(self, x: torch.Tensor):
85
+ self.state.is_training = self.training
86
+ if self.weight.CB is not None:
87
+ self.init_8bit_state()
88
+
89
+ # weights are cast automatically as Int8Params, but the bias has to be cast manually
90
+ if self.bias is not None and self.bias.dtype != x.dtype:
91
+ self.bias.data = self.bias.data.to(x.dtype)
92
+
93
+ out = bnb.matmul(x, self.weight, bias=self.bias, state=self.state)
94
+
95
+ if not self.state.has_fp16_weights:
96
+ if self.state.CB is not None and self.state.CxB is not None:
97
+ # we converted 8-bit row major to turing/ampere format in the first inference pass
98
+ # we no longer need the row-major weight
99
+ del self.state.CB
100
+ self.weight.data = self.state.CxB
101
+ return out
102
+
103
+ def quantize_offline(model, bits: int):
104
+ assert (bits == 4), f'bits: {bits} is not supported'
105
+
106
+ for i, layer in enumerate(model.model.layers):
107
+ layer.self_attn.W_pack = bnb.nn.Linear4bit(
108
+ layer.self_attn.W_pack.weight.shape[1],
109
+ layer.self_attn.W_pack.weight.shape[0],
110
+ False,
111
+ torch.float16,
112
+ compress_statistics=True,
113
+ quant_type="nf4",
114
+ )
115
+ layer.self_attn.o_proj = bnb.nn.Linear4bit(
116
+ layer.self_attn.o_proj.weight.shape[1],
117
+ layer.self_attn.o_proj.weight.shape[0],
118
+ False,
119
+ torch.float16,
120
+ compress_statistics=True,
121
+ quant_type="nf4",
122
+ )
123
+
124
+ layer.mlp.gate_proj = bnb.nn.Linear4bit(
125
+ layer.mlp.gate_proj.weight.shape[1],
126
+ layer.mlp.gate_proj.weight.shape[0],
127
+ False,
128
+ torch.float16,
129
+ compress_statistics=True,
130
+ quant_type="nf4",
131
+ )
132
+ layer.mlp.down_proj = bnb.nn.Linear4bit(
133
+ layer.mlp.down_proj.weight.shape[1],
134
+ layer.mlp.down_proj.weight.shape[0],
135
+ False,
136
+ torch.float16,
137
+ compress_statistics=True,
138
+ quant_type="nf4",
139
+ )
140
+ layer.mlp.up_proj = bnb.nn.Linear4bit(
141
+ layer.mlp.up_proj.weight.shape[1],
142
+ layer.mlp.up_proj.weight.shape[0],
143
+ False,
144
+ torch.float16,
145
+ compress_statistics=True,
146
+ quant_type="nf4",
147
+ )
148
+ return model
149
+
150
+ def quantize_online(model, bits: int):
151
+ def quant(weight, bias=None):
152
+ if bits == 8:
153
+ linear = Linear8bitLtOnline(
154
+ weight,
155
+ bias,
156
+ has_fp16_weights=False,
157
+ threshold=6.0,
158
+ )
159
+ if bias is not None:
160
+ linear.bias = torch.nn.Parameter(bias)
161
+ elif bits == 4:
162
+ linear = Linear4bitOnline(
163
+ weight,
164
+ bias,
165
+ quant_type="nf4", #fp4/nf4
166
+ )
167
+ else:
168
+ raise ValueError("quantize only support 4/8 bit")
169
+ return linear
170
+
171
+ for i, layer in enumerate(model.model.layers):
172
+ layer.self_attn.W_pack = quant(layer.self_attn.W_pack.weight)
173
+ layer.self_attn.o_proj = quant(layer.self_attn.o_proj.weight)
174
+ layer.mlp.gate_proj = quant(layer.mlp.gate_proj.weight)
175
+ layer.mlp.down_proj = quant(layer.mlp.down_proj.weight)
176
+ layer.mlp.up_proj = quant(layer.mlp.up_proj.weight)
177
+ return model
178
+
179
+ def init_model_weight_int4(config, model, state_dict):
180
+ #replace Params4bit.cuda with Params4bitCuda
181
+ Params4bit.cuda = Params4bitCuda
182
+
183
+ for i in range(config.num_hidden_layers):
184
+ weight_data = state_dict[f'model.layers.{i}.self_attn.W_pack.weight.data']
185
+ weight_quant_state = state_dict[f'model.layers.{i}.self_attn.W_pack.weight.quant_state']
186
+ model.model.layers[i].self_attn.W_pack.weight = Params4bit(weight_data, requires_grad=False, quant_state=weight_quant_state)
187
+
188
+ weight_data = state_dict[f'model.layers.{i}.self_attn.o_proj.weight.data']
189
+ weight_quant_state = state_dict[f'model.layers.{i}.self_attn.o_proj.weight.quant_state']
190
+ model.model.layers[i].self_attn.o_proj.weight = Params4bit(weight_data, requires_grad=False, quant_state=weight_quant_state)
191
+
192
+ weight_data = state_dict[f'model.layers.{i}.mlp.gate_proj.weight.data']
193
+ weight_quant_state = state_dict[f'model.layers.{i}.mlp.gate_proj.weight.quant_state']
194
+ model.model.layers[i].mlp.gate_proj.weight = Params4bit(weight_data, requires_grad=False, quant_state=weight_quant_state)
195
+
196
+ weight_data = state_dict[f'model.layers.{i}.mlp.up_proj.weight.data']
197
+ weight_quant_state = state_dict[f'model.layers.{i}.mlp.up_proj.weight.quant_state']
198
+ model.model.layers[i].mlp.up_proj.weight = Params4bit(weight_data, requires_grad=False, quant_state=weight_quant_state)
199
+
200
+ weight_data = state_dict[f'model.layers.{i}.mlp.down_proj.weight.data']
201
+ weight_quant_state = state_dict[f'model.layers.{i}.mlp.down_proj.weight.quant_state']
202
+ model.model.layers[i].mlp.down_proj.weight = Params4bit(weight_data, requires_grad=False, quant_state=weight_quant_state)
203
+
204
+ model.model.layers[i].input_layernorm.weight = state_dict[f'model.layers.{i}.input_layernorm.weight']
205
+ model.model.layers[i].post_attention_layernorm.weight = state_dict[f'model.layers.{i}.post_attention_layernorm.weight']
206
+
207
+ model.model.embed_tokens.weight = state_dict['model.embed_tokens.weight']
208
+ model.model.norm.weight = state_dict['model.norm.weight']
209
+ model.lm_head.weight = state_dict['lm_head.weight']
210
+ return model
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": 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": {
17
+ "content": "<unk>",
18
+ "lstrip": false,
19
+ "normalized": true,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ },
23
+ "unk_token": {
24
+ "content": "<unk>",
25
+ "lstrip": false,
26
+ "normalized": true,
27
+ "rstrip": false,
28
+ "single_word": false
29
+ }
30
+ }
tokenization_baichuan.py ADDED
@@ -0,0 +1,251 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 Baichuan Inc. All Rights Reserved.
2
+
3
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
4
+ #
5
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
6
+ # and OPT implementations in this library. It has been modified from its
7
+ # original forms to accommodate minor architectural differences compared
8
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
9
+ #
10
+ # Licensed under the Apache License, Version 2.0 (the "License");
11
+ # you may not use this file except in compliance with the License.
12
+ # You may obtain a copy of the License at
13
+ #
14
+ # http://www.apache.org/licenses/LICENSE-2.0
15
+ #
16
+ # Unless required by applicable law or agreed to in writing, software
17
+ # distributed under the License is distributed on an "AS IS" BASIS,
18
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
19
+ # See the License for the specific language governing permissions and
20
+ # limitations under the License.
21
+
22
+ import os
23
+ from shutil import copyfile
24
+ from typing import Any, Dict, List, Optional, Tuple
25
+
26
+ import sentencepiece as spm
27
+
28
+ from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
29
+ from transformers.utils import logging
30
+
31
+
32
+ logger = logging.get_logger(__name__)
33
+
34
+ VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
35
+
36
+ PRETRAINED_VOCAB_FILES_MAP = {
37
+ "vocab_file": {},
38
+ "tokenizer_file": {},
39
+ }
40
+ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {}
41
+
42
+
43
+ class BaichuanTokenizer(PreTrainedTokenizer):
44
+ """
45
+ Construct a Baichuan tokenizer. Based on byte-level Byte-Pair-Encoding.
46
+
47
+ Args:
48
+ vocab_file (`str`):
49
+ Path to the vocabulary file.
50
+ """
51
+
52
+ vocab_files_names = VOCAB_FILES_NAMES
53
+ pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
54
+ max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
55
+ model_input_names = ["input_ids", "attention_mask"]
56
+
57
+ def __init__(
58
+ self,
59
+ vocab_file,
60
+ unk_token="<unk>",
61
+ bos_token="<s>",
62
+ eos_token="</s>",
63
+ pad_token=None,
64
+ sp_model_kwargs: Optional[Dict[str, Any]] = None,
65
+ add_bos_token=True,
66
+ add_eos_token=False,
67
+ clean_up_tokenization_spaces=False,
68
+ **kwargs,
69
+ ):
70
+ self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
71
+ bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
72
+ eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
73
+ unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
74
+ pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
75
+ super().__init__(
76
+ bos_token=bos_token,
77
+ eos_token=eos_token,
78
+ unk_token=unk_token,
79
+ pad_token=pad_token,
80
+ add_bos_token=add_bos_token,
81
+ add_eos_token=add_eos_token,
82
+ sp_model_kwargs=self.sp_model_kwargs,
83
+ clean_up_tokenization_spaces=clean_up_tokenization_spaces,
84
+ **kwargs,
85
+ )
86
+ self.vocab_file = vocab_file
87
+ self.add_bos_token = add_bos_token
88
+ self.add_eos_token = add_eos_token
89
+ self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
90
+ self.sp_model.Load(vocab_file)
91
+
92
+ def __getstate__(self):
93
+ state = self.__dict__.copy()
94
+ state["sp_model"] = None
95
+ return state
96
+
97
+ def __setstate__(self, d):
98
+ self.__dict__ = d
99
+ self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
100
+ self.sp_model.Load(self.vocab_file)
101
+
102
+ @property
103
+ def vocab_size(self):
104
+ """Returns vocab size"""
105
+ return self.sp_model.get_piece_size()
106
+
107
+ def get_vocab(self):
108
+ """Returns vocab as a dict"""
109
+ vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
110
+ vocab.update(self.added_tokens_encoder)
111
+ return vocab
112
+
113
+ def _tokenize(self, text):
114
+ """Returns a tokenized string."""
115
+ return self.sp_model.encode(text, out_type=str)
116
+
117
+ def _convert_token_to_id(self, token):
118
+ """Converts a token (str) in an id using the vocab."""
119
+ return self.sp_model.piece_to_id(token)
120
+
121
+ def _convert_id_to_token(self, index):
122
+ """Converts an index (integer) in a token (str) using the vocab."""
123
+ token = self.sp_model.IdToPiece(index)
124
+ return token
125
+
126
+ def convert_tokens_to_string(self, tokens):
127
+ """Converts a sequence of tokens (string) in a single string."""
128
+ current_sub_tokens = []
129
+ out_string = ""
130
+ prev_is_special = False
131
+ for i, token in enumerate(tokens):
132
+ # make sure that special tokens are not decoded using sentencepiece model
133
+ if token in self.all_special_tokens:
134
+ if not prev_is_special and i != 0:
135
+ out_string += " "
136
+ out_string += self.sp_model.decode(current_sub_tokens) + token
137
+ prev_is_special = True
138
+ current_sub_tokens = []
139
+ else:
140
+ current_sub_tokens.append(token)
141
+ prev_is_special = False
142
+ out_string += self.sp_model.decode(current_sub_tokens)
143
+ return out_string
144
+
145
+ def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
146
+ """
147
+ Save the vocabulary and special tokens file to a directory.
148
+
149
+ Args:
150
+ save_directory (`str`):
151
+ The directory in which to save the vocabulary.
152
+
153
+ Returns:
154
+ `Tuple(str)`: Paths to the files saved.
155
+ """
156
+ if not os.path.isdir(save_directory):
157
+ logger.error(f"Vocabulary path ({save_directory}) should be a directory")
158
+ return
159
+ out_vocab_file = os.path.join(
160
+ save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
161
+ )
162
+
163
+ if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
164
+ copyfile(self.vocab_file, out_vocab_file)
165
+ elif not os.path.isfile(self.vocab_file):
166
+ with open(out_vocab_file, "wb") as fi:
167
+ content_spiece_model = self.sp_model.serialized_model_proto()
168
+ fi.write(content_spiece_model)
169
+
170
+ return (out_vocab_file,)
171
+
172
+ def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
173
+ bos_token_id = [self.bos_token_id] if self.add_bos_token else []
174
+ eos_token_id = [self.eos_token_id] if self.add_eos_token else []
175
+
176
+ output = bos_token_id + token_ids_0 + eos_token_id
177
+
178
+ if token_ids_1 is not None:
179
+ output = output + bos_token_id + token_ids_1 + eos_token_id
180
+
181
+ return output
182
+
183
+ def get_special_tokens_mask(
184
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
185
+ ) -> List[int]:
186
+ """
187
+ Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
188
+ special tokens using the tokenizer `prepare_for_model` method.
189
+
190
+ Args:
191
+ token_ids_0 (`List[int]`):
192
+ List of IDs.
193
+ token_ids_1 (`List[int]`, *optional*):
194
+ Optional second list of IDs for sequence pairs.
195
+ already_has_special_tokens (`bool`, *optional*, defaults to `False`):
196
+ Whether or not the token list is already formatted with special tokens for the model.
197
+
198
+ Returns:
199
+ `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
200
+ """
201
+ if already_has_special_tokens:
202
+ return super().get_special_tokens_mask(
203
+ token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
204
+ )
205
+
206
+ bos_token_id = [1] if self.add_bos_token else []
207
+ eos_token_id = [1] if self.add_eos_token else []
208
+
209
+ if token_ids_1 is None:
210
+ return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
211
+ return (
212
+ bos_token_id
213
+ + ([0] * len(token_ids_0))
214
+ + eos_token_id
215
+ + bos_token_id
216
+ + ([0] * len(token_ids_1))
217
+ + eos_token_id
218
+ )
219
+
220
+ def create_token_type_ids_from_sequences(
221
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
222
+ ) -> List[int]:
223
+ """
224
+ Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
225
+ sequence pair mask has the following format:
226
+
227
+ ```
228
+ 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
229
+ | first sequence | second sequence |
230
+ ```
231
+
232
+ if token_ids_1 is None, only returns the first portion of the mask (0s).
233
+
234
+ Args:
235
+ token_ids_0 (`List[int]`):
236
+ List of ids.
237
+ token_ids_1 (`List[int]`, *optional*):
238
+ Optional second list of IDs for sequence pairs.
239
+
240
+ Returns:
241
+ `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
242
+ """
243
+ bos_token_id = [self.bos_token_id] if self.add_bos_token else []
244
+ eos_token_id = [self.eos_token_id] if self.add_eos_token else []
245
+
246
+ output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
247
+
248
+ if token_ids_1 is not None:
249
+ output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
250
+
251
+ return output
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:79452955be6b419a65984273a9f08af86042e1c2a75ee3ba989cbf620a133cc2
3
+ size 2001107
tokenizer_config.json ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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": 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
+ "padding_side": "left",
37
+ "sp_model_kwargs": {},
38
+ "split_special_tokens": false,
39
+ "tokenizer_class": "BaichuanTokenizer",
40
+ "unk_token": {
41
+ "__type": "AddedToken",
42
+ "content": "<unk>",
43
+ "lstrip": false,
44
+ "normalized": true,
45
+ "rstrip": false,
46
+ "single_word": true
47
+ },
48
+ "use_fast": false
49
+ }