doberst commited on
Commit
aaa5013
1 Parent(s): ac50972

Upload 9 files

Browse files
config.txt ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "YiForCausalLM"
4
+ ],
5
+ "auto_map": {
6
+ "AutoConfig": "configuration_yi.YiConfig",
7
+ "AutoModel": "modeling_yi.YiModel",
8
+ "AutoModelForCausalLM":"modeling_yi.YiForCausalLM"
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": "Yi",
18
+ "num_attention_heads": 32,
19
+ "num_hidden_layers": 32,
20
+ "num_key_value_heads": 4,
21
+ "pad_token_id": 0,
22
+ "rms_norm_eps": 1e-05,
23
+ "rope_theta": 5000000.0,
24
+ "tie_word_embeddings": false,
25
+ "torch_dtype": "bfloat16",
26
+ "transformers_version": "4.34.0",
27
+ "use_cache": true,
28
+ "vocab_size": 64000
29
+ }
configuration_yi.txt ADDED
@@ -0,0 +1,121 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ Yi model configuration"""
2
+ from transformers.configuration_utils import PretrainedConfig
3
+ from transformers.utils import logging
4
+
5
+ logger = logging.get_logger(__name__)
6
+
7
+ Yi_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
8
+
9
+
10
+ class YiConfig(PretrainedConfig):
11
+ r"""
12
+ This is the configuration class to store the configuration of a [`YiModel`]. It is used to instantiate an Yi
13
+ model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
14
+ defaults will yield a similar configuration to that of the Yi model.
15
+
16
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
17
+ documentation from [`PretrainedConfig`] for more information.
18
+
19
+
20
+ Args:
21
+ vocab_size (`int`, *optional*, defaults to 64000):
22
+ Vocabulary size of the Yi model. Defines the number of different tokens that can be represented by the
23
+ `inputs_ids` passed when calling [`YiModel`]
24
+ hidden_size (`int`, *optional*, defaults to 4096):
25
+ Dimension of the hidden representations.
26
+ intermediate_size (`int`, *optional*, defaults to 11008):
27
+ Dimension of the MLP representations.
28
+ num_hidden_layers (`int`, *optional*, defaults to 32):
29
+ Number of hidden layers in the Transformer encoder.
30
+ num_attention_heads (`int`, *optional*, defaults to 32):
31
+ Number of attention heads for each attention layer in the Transformer encoder.
32
+ num_key_value_heads (`int`, *optional*):
33
+ This is the number of key_value heads that should be used to implement Grouped Query Attention. If
34
+ `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
35
+ `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
36
+ converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
37
+ by meanpooling all the original heads within that group. For more details checkout [this
38
+ paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
39
+ `num_attention_heads`.
40
+ hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
41
+ The non-linear activation function (function or string) in the decoder.
42
+ max_position_embeddings (`int`, *optional*, defaults to 4096):
43
+ The maximum sequence length that this model might ever be used with. Typically set this to something large
44
+ just in case (e.g., 512 or 1024 or 2048 or 4096).
45
+ initializer_range (`float`, *optional*, defaults to 0.02):
46
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
47
+ rms_norm_eps (`float`, *optional*, defaults to 1e-5):
48
+ The epsilon used by the rms normalization layers.
49
+ use_cache (`bool`, *optional*, defaults to `True`):
50
+ Whether or not the model should return the last key/values attentions (not used by all models). Only
51
+ relevant if `config.is_decoder=True`.
52
+ tie_word_embeddings(`bool`, *optional*, defaults to `False`):
53
+ Whether to tie weight embeddings
54
+ output_attentions (`bool`, *optional*, defaults to `False`):
55
+ Whether or not to output attentions.
56
+ rope_theta (`float`, *optional*, defaults to 5000000.0):
57
+ The base period of the RoPE embeddings.
58
+ Example:
59
+
60
+ ```python
61
+ >>> from transformers import YiModel, YiConfig
62
+
63
+ >>> # Initializing a Yi style configuration
64
+ >>> configuration = YiConfig()
65
+
66
+ >>> # Initializing a model from the Yi style configuration
67
+ >>> model = YiModel(configuration)
68
+
69
+ >>> # Accessing the model configuration
70
+ >>> configuration = model.config
71
+ ```"""
72
+ model_type = "Yi"
73
+ keys_to_ignore_at_inference = ["past_key_values"]
74
+
75
+ def __init__(
76
+ self,
77
+ vocab_size=64000,
78
+ hidden_size=4096,
79
+ intermediate_size=11008,
80
+ num_hidden_layers=32,
81
+ num_attention_heads=32,
82
+ num_key_value_heads=4,
83
+ hidden_act="silu",
84
+ max_position_embeddings=4096,
85
+ initializer_range=0.02,
86
+ rms_norm_eps=1e-5,
87
+ use_cache=True,
88
+ pad_token_id=0,
89
+ bos_token_id=1,
90
+ eos_token_id=2,
91
+ tie_word_embeddings=False,
92
+ output_attentions=False,
93
+ rope_theta=5000000.0,
94
+ **kwargs,
95
+ ):
96
+ self.vocab_size = vocab_size
97
+ self.max_position_embeddings = max_position_embeddings
98
+ self.hidden_size = hidden_size
99
+ self.intermediate_size = intermediate_size
100
+ self.num_hidden_layers = num_hidden_layers
101
+ self.num_attention_heads = num_attention_heads
102
+
103
+ # for backward compatibility
104
+ if num_key_value_heads is None:
105
+ num_key_value_heads = num_attention_heads
106
+
107
+ self.num_key_value_heads = num_key_value_heads
108
+ self.hidden_act = hidden_act
109
+ self.initializer_range = initializer_range
110
+ self.rms_norm_eps = rms_norm_eps
111
+ self.use_cache = use_cache
112
+ self.output_attentions = output_attentions
113
+ self.rope_theta = rope_theta
114
+
115
+ super().__init__(
116
+ pad_token_id=pad_token_id,
117
+ bos_token_id=bos_token_id,
118
+ eos_token_id=eos_token_id,
119
+ tie_word_embeddings=tie_word_embeddings,
120
+ **kwargs,
121
+ )
generation_config.txt 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": "4.34.0"
7
+ }
md5.txt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ e11decc690391b47f62e217a3faca830 pytorch_model-00001-of-00002.bin
2
+ ea6e2b5eddc1416a101361efb286d79c pytorch_model-00002-of-00002.bin
3
+ 291724ef50f729e45d68f474a7755bbc tokenizer.model
modeling_yi.txt ADDED
@@ -0,0 +1,1028 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ PyTorch Yi model."""
2
+ import math
3
+ from typing import List, Optional, Tuple, Union
4
+
5
+ import torch.utils.checkpoint
6
+ from einops import repeat
7
+ from torch import nn
8
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
9
+ from transformers.activations import ACT2FN
10
+ from transformers.modeling_outputs import (
11
+ BaseModelOutputWithPast,
12
+ CausalLMOutputWithPast,
13
+ SequenceClassifierOutputWithPast,
14
+ )
15
+ from transformers.modeling_utils import PreTrainedModel
16
+ from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS
17
+ from transformers.utils import (
18
+ add_start_docstrings,
19
+ add_start_docstrings_to_model_forward,
20
+ logging,
21
+ replace_return_docstrings,
22
+ )
23
+
24
+ from .configuration_yi import YiConfig
25
+
26
+ is_flash_attn_available = True
27
+ try:
28
+ from flash_attn import flash_attn_func
29
+ except Exception:
30
+ is_flash_attn_available = False
31
+
32
+ logger = logging.get_logger(__name__)
33
+
34
+ _CONFIG_FOR_DOC = "YiConfig"
35
+
36
+
37
+ # Copied from transformers.models.bart.modeling_bart._make_causal_mask
38
+ def _make_causal_mask(
39
+ input_ids_shape: torch.Size,
40
+ dtype: torch.dtype,
41
+ device: torch.device,
42
+ past_key_values_length: int = 0,
43
+ ):
44
+ """
45
+ Make causal mask used for bi-directional self-attention.
46
+ """
47
+ bsz, tgt_len = input_ids_shape
48
+ mask = torch.full(
49
+ (tgt_len, tgt_len),
50
+ torch.tensor(torch.finfo(dtype).min, device=device),
51
+ device=device,
52
+ )
53
+ mask_cond = torch.arange(mask.size(-1), device=device)
54
+ mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
55
+ mask = mask.to(dtype)
56
+
57
+ if past_key_values_length > 0:
58
+ mask = torch.cat(
59
+ [
60
+ torch.zeros(
61
+ tgt_len, past_key_values_length, dtype=dtype, device=device
62
+ ),
63
+ mask,
64
+ ],
65
+ dim=-1,
66
+ )
67
+ return mask[None, None, :, :].expand(
68
+ bsz, 1, tgt_len, tgt_len + past_key_values_length
69
+ )
70
+
71
+
72
+ # Copied from transformers.models.bart.modeling_bart._expand_mask
73
+ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
74
+ """
75
+ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
76
+ """
77
+ bsz, src_len = mask.size()
78
+ tgt_len = tgt_len if tgt_len is not None else src_len
79
+
80
+ expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
81
+
82
+ inverted_mask = 1.0 - expanded_mask
83
+
84
+ return inverted_mask.masked_fill(
85
+ inverted_mask.to(torch.bool), torch.finfo(dtype).min
86
+ )
87
+
88
+
89
+ class YiRMSNorm(nn.Module):
90
+ def __init__(self, hidden_size, eps=1e-5):
91
+ """
92
+ YiRMSNorm 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
+ input_dtype = hidden_states.dtype
100
+ hidden_states = hidden_states.to(torch.float32)
101
+ variance = hidden_states.pow(2).mean(-1, keepdim=True)
102
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
103
+
104
+ return self.weight * hidden_states.to(input_dtype)
105
+
106
+
107
+ ALL_LAYERNORM_LAYERS.append(YiRMSNorm)
108
+
109
+
110
+ class YiRotaryEmbedding(torch.nn.Module):
111
+ def __init__(self, dim, max_position_embeddings=4096, base=5000000, device=None):
112
+ super().__init__()
113
+
114
+ self.dim = dim
115
+ self.max_position_embeddings = max_position_embeddings
116
+ self.base = base
117
+
118
+ # Build here to make `torch.jit.trace` work.
119
+ self._set_cos_sin_cache(seq_len=max_position_embeddings, device=device)
120
+
121
+ def _set_cos_sin_cache(self, seq_len, device):
122
+ self.max_seq_len_cached = seq_len
123
+ inv_freq = 1.0 / (
124
+ self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim)
125
+ )
126
+ t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.float32)
127
+ freqs = torch.einsum("i,j->ij", t, inv_freq)
128
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
129
+ emb = torch.cat((freqs, freqs), dim=-1)
130
+ self.register_buffer(
131
+ "cos_cached", emb.cos()[None, None, :, :], persistent=False
132
+ )
133
+ self.register_buffer(
134
+ "sin_cached", emb.sin()[None, None, :, :], persistent=False
135
+ )
136
+
137
+ def forward(self, x, seq_len=None):
138
+ # x: [bs, num_attention_heads, seq_len, head_size]
139
+ if seq_len > self.max_seq_len_cached:
140
+ self._set_cos_sin_cache(seq_len=seq_len, device=x.device)
141
+
142
+ return (
143
+ self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
144
+ self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
145
+ )
146
+
147
+
148
+ def rotate_half(x):
149
+ """Rotates half the hidden dims of the input."""
150
+ x1 = x[..., : x.shape[-1] // 2]
151
+ x2 = x[..., x.shape[-1] // 2 :]
152
+ return torch.cat((-x2, x1), dim=-1)
153
+
154
+
155
+ def apply_rotary_pos_emb(q, k, cos, sin, position_ids, flash_attn_available):
156
+ # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
157
+ cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
158
+ sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
159
+ expand_dim = 2 if flash_attn_available else 1
160
+ cos = cos[position_ids].unsqueeze(expand_dim) # [bs, seq_len, dim]
161
+ sin = sin[position_ids].unsqueeze(expand_dim) # [bs, seq_len, dim]
162
+ q_embed = (q * cos) + (rotate_half(q) * sin)
163
+ k_embed = (k * cos) + (rotate_half(k) * sin)
164
+ return q_embed, k_embed
165
+
166
+
167
+ class YiMLP(nn.Module):
168
+ def __init__(self, hidden_size: int, intermediate_size: int, hidden_act: str):
169
+ super().__init__()
170
+ self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
171
+ self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
172
+ self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
173
+ self.act_fn = ACT2FN[hidden_act]
174
+
175
+ def forward(self, x):
176
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
177
+
178
+
179
+ class YiAttention(nn.Module):
180
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
181
+
182
+ def __init__(self, config: YiConfig):
183
+ super().__init__()
184
+ self.config = config
185
+ self.hidden_size = config.hidden_size
186
+ self.num_heads = config.num_attention_heads
187
+ self.head_dim = self.hidden_size // self.num_heads
188
+ self.num_key_value_heads = config.num_key_value_heads
189
+ self.num_key_value_groups = self.num_heads // self.num_key_value_heads
190
+ self.max_position_embeddings = config.max_position_embeddings
191
+
192
+ if (self.head_dim * self.num_heads) != self.hidden_size:
193
+ raise ValueError(
194
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
195
+ f" and `num_heads`: {self.num_heads})."
196
+ )
197
+ self.q_proj = nn.Linear(
198
+ self.hidden_size, self.num_heads * self.head_dim, bias=False
199
+ )
200
+ self.k_proj = nn.Linear(
201
+ self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
202
+ )
203
+ self.v_proj = nn.Linear(
204
+ self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
205
+ )
206
+ self.o_proj = nn.Linear(
207
+ self.num_heads * self.head_dim, self.hidden_size, bias=False
208
+ )
209
+
210
+ self.rotary_emb = YiRotaryEmbedding(
211
+ self.head_dim,
212
+ max_position_embeddings=self.max_position_embeddings,
213
+ base=self.config.rope_theta,
214
+ )
215
+
216
+ def forward(
217
+ self,
218
+ hidden_states: torch.Tensor,
219
+ attention_mask: Optional[torch.Tensor] = None,
220
+ position_ids: Optional[torch.LongTensor] = None,
221
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
222
+ output_attentions: bool = False,
223
+ use_cache: bool = False,
224
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
225
+ bsz, q_len, _ = hidden_states.size()
226
+
227
+ query_states = self.q_proj(hidden_states).view(
228
+ bsz, q_len, self.num_heads, self.head_dim
229
+ )
230
+
231
+ key_states = self.k_proj(hidden_states).view(
232
+ bsz, q_len, self.num_key_value_heads, self.head_dim
233
+ )
234
+ value_states = self.v_proj(hidden_states).view(
235
+ bsz, q_len, self.num_key_value_heads, self.head_dim
236
+ )
237
+
238
+ if not is_flash_attn_available:
239
+ if self.num_key_value_groups > 1:
240
+ key_states = repeat(
241
+ key_states, f"b n h d -> b n (h {self.num_key_value_groups}) d"
242
+ )
243
+ value_states = repeat(
244
+ value_states, f"b n h d -> b n (h {self.num_key_value_groups}) d"
245
+ )
246
+
247
+ # b n h d -> b h n d
248
+ query_states = query_states.transpose(1, 2)
249
+ key_states = key_states.transpose(1, 2)
250
+ value_states = value_states.transpose(1, 2)
251
+
252
+ seq_dim = 1 if is_flash_attn_available else 2
253
+ kv_seq_len = key_states.shape[seq_dim]
254
+ if past_key_value is not None:
255
+ kv_seq_len += past_key_value[0].shape[seq_dim]
256
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
257
+ query_states, key_states = apply_rotary_pos_emb(
258
+ query_states, key_states, cos, sin, position_ids, is_flash_attn_available
259
+ )
260
+
261
+ if past_key_value is not None:
262
+ # reuse k, v, self_attention
263
+ key_states = torch.cat([past_key_value[0], key_states], dim=seq_dim)
264
+ value_states = torch.cat([past_key_value[1], value_states], dim=seq_dim)
265
+
266
+ past_key_value = (key_states, value_states) if use_cache else None
267
+
268
+ if is_flash_attn_available:
269
+ attn_output = flash_attn_func(
270
+ query_states, key_states, value_states, dropout_p=0.0, causal=True
271
+ )
272
+ else:
273
+ attn_weights = torch.matmul(
274
+ query_states, key_states.transpose(2, 3)
275
+ ) / math.sqrt(self.head_dim)
276
+
277
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
278
+ raise ValueError(
279
+ f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
280
+ f" {attn_weights.size()}"
281
+ )
282
+
283
+ if attention_mask is not None:
284
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
285
+ raise ValueError(
286
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is"
287
+ f"{attention_mask.size()}"
288
+ )
289
+ attn_weights = attn_weights + attention_mask
290
+ dtype_min = torch.tensor(
291
+ torch.finfo(attn_weights.dtype).min,
292
+ device=attn_weights.device,
293
+ dtype=attn_weights.dtype,
294
+ )
295
+ attn_weights = torch.max(attn_weights, dtype_min)
296
+
297
+ # upcast attention to fp32
298
+ attn_weights = nn.functional.softmax(
299
+ attn_weights, dim=-1, dtype=torch.float32
300
+ ).to(query_states.dtype)
301
+ attn_output = torch.matmul(attn_weights, value_states)
302
+
303
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
304
+ raise ValueError(
305
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
306
+ f" {attn_output.size()}"
307
+ )
308
+
309
+ if not is_flash_attn_available:
310
+ attn_output = attn_output.transpose(1, 2)
311
+
312
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
313
+
314
+ attn_output = self.o_proj(attn_output)
315
+
316
+ if not output_attentions:
317
+ attn_weights = None
318
+
319
+ return attn_output, attn_weights, past_key_value
320
+
321
+
322
+ class YiDecoderLayer(nn.Module):
323
+ def __init__(self, config: YiConfig):
324
+ super().__init__()
325
+
326
+ self.hidden_size = config.hidden_size
327
+ self.self_attn = YiAttention(config=config)
328
+ self.mlp = YiMLP(
329
+ hidden_size=self.hidden_size,
330
+ intermediate_size=config.intermediate_size,
331
+ hidden_act=config.hidden_act,
332
+ )
333
+
334
+ self.ln1 = YiRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
335
+ self.ln2 = YiRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
336
+
337
+ def forward(
338
+ self,
339
+ hidden_states: torch.Tensor,
340
+ attention_mask: Optional[torch.Tensor] = None,
341
+ position_ids: Optional[torch.LongTensor] = None,
342
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
343
+ output_attentions: Optional[bool] = False,
344
+ use_cache: Optional[bool] = False,
345
+ ) -> Tuple[
346
+ torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]
347
+ ]:
348
+ """
349
+ Args:
350
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
351
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
352
+ `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
353
+ output_attentions (`bool`, *optional*):
354
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
355
+ returned tensors for more detail.
356
+ use_cache (`bool`, *optional*):
357
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
358
+ (see `past_key_values`).
359
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
360
+ """
361
+
362
+ residual = hidden_states
363
+
364
+ hidden_states = self.ln1(hidden_states)
365
+
366
+ # Self Attention
367
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
368
+ hidden_states=hidden_states,
369
+ attention_mask=attention_mask,
370
+ position_ids=position_ids,
371
+ past_key_value=past_key_value,
372
+ output_attentions=output_attentions,
373
+ use_cache=use_cache,
374
+ )
375
+ hidden_states = residual + hidden_states
376
+
377
+ # Fully Connected
378
+ residual = hidden_states
379
+ hidden_states = self.ln2(hidden_states)
380
+ hidden_states = self.mlp(hidden_states)
381
+ hidden_states = residual + hidden_states
382
+
383
+ outputs = (hidden_states,)
384
+
385
+ if output_attentions:
386
+ outputs += (self_attn_weights,)
387
+
388
+ if use_cache:
389
+ outputs += (present_key_value,)
390
+
391
+ return outputs
392
+
393
+
394
+ Yi_START_DOCSTRING = r"""
395
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
396
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
397
+ etc.)
398
+
399
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
400
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
401
+ and behavior.
402
+
403
+ Parameters:
404
+ config ([`YiConfig`]):
405
+ Model configuration class with all the parameters of the model. Initializing with a config file does not
406
+ load the weights associated with the model, only the configuration. Check out the
407
+ [`~PreTrainedModel.from_pretrained`] method to load the model weights.
408
+ """
409
+
410
+
411
+ @add_start_docstrings(
412
+ "The bare Yi Model outputting raw hidden-states without any specific head on top.",
413
+ Yi_START_DOCSTRING,
414
+ )
415
+ class YiPreTrainedModel(PreTrainedModel):
416
+ config_class = YiConfig
417
+ base_model_prefix = "model"
418
+ supports_gradient_checkpointing = True
419
+ _no_split_modules = ["YiDecoderLayer"]
420
+ _skip_keys_device_placement = "past_key_values"
421
+
422
+ def _init_weights(self, module):
423
+ std = self.config.initializer_range
424
+ if isinstance(module, nn.Linear):
425
+ module.weight.data.normal_(mean=0.0, std=std)
426
+ if module.bias is not None:
427
+ module.bias.data.zero_()
428
+ elif isinstance(module, nn.Embedding):
429
+ module.weight.data.normal_(mean=0.0, std=std)
430
+ if module.padding_idx is not None:
431
+ module.weight.data[module.padding_idx].zero_()
432
+
433
+ def _set_gradient_checkpointing(self, module, value=False):
434
+ if isinstance(module, YiModel):
435
+ module.gradient_checkpointing = value
436
+
437
+
438
+ Yi_INPUTS_DOCSTRING = r"""
439
+ Args:
440
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
441
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
442
+ it.
443
+
444
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
445
+ [`PreTrainedTokenizer.__call__`] for details.
446
+
447
+ [What are input IDs?](../glossary#input-ids)
448
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
449
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
450
+
451
+ - 1 for tokens that are **not masked**,
452
+ - 0 for tokens that are **masked**.
453
+
454
+ [What are attention masks?](../glossary#attention-mask)
455
+
456
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
457
+ [`PreTrainedTokenizer.__call__`] for details.
458
+
459
+ If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
460
+ `past_key_values`).
461
+
462
+ If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
463
+ and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
464
+ information on the default strategy.
465
+
466
+ - 1 indicates the head is **not masked**,
467
+ - 0 indicates the head is **masked**.
468
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
469
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
470
+ config.n_positions - 1]`.
471
+
472
+ [What are position IDs?](../glossary#position-ids)
473
+ past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
474
+ Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
475
+ `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
476
+ `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
477
+
478
+ Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
479
+ blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
480
+
481
+ If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
482
+ don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
483
+ `decoder_input_ids` of shape `(batch_size, sequence_length)`.
484
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
485
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
486
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
487
+ model's internal embedding lookup matrix.
488
+ use_cache (`bool`, *optional*):
489
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
490
+ `past_key_values`).
491
+ output_attentions (`bool`, *optional*):
492
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
493
+ tensors for more detail.
494
+ output_hidden_states (`bool`, *optional*):
495
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
496
+ more detail.
497
+ return_dict (`bool`, *optional*):
498
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
499
+ """
500
+
501
+
502
+ @add_start_docstrings(
503
+ "The bare Yi Model outputting raw hidden-states without any specific head on top.",
504
+ Yi_START_DOCSTRING,
505
+ )
506
+ class YiModel(YiPreTrainedModel):
507
+ """
508
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`YiDecoderLayer`]
509
+
510
+ Args:
511
+ config: YiConfig
512
+ """
513
+
514
+ def __init__(self, config: YiConfig):
515
+ super().__init__(config)
516
+ self.padding_idx = config.pad_token_id
517
+ self.vocab_size = config.vocab_size
518
+
519
+ self.embed_tokens = nn.Embedding(
520
+ config.vocab_size, config.hidden_size, self.padding_idx
521
+ )
522
+ self.layers = nn.ModuleList(
523
+ [YiDecoderLayer(config) for _ in range(config.num_hidden_layers)]
524
+ )
525
+
526
+ self.norm = YiRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
527
+
528
+ self.gradient_checkpointing = False
529
+ # Initialize weights and apply final processing
530
+ self.post_init()
531
+
532
+ def get_input_embeddings(self):
533
+ return self.embed_tokens
534
+
535
+ def set_input_embeddings(self, value):
536
+ self.embed_tokens = value
537
+
538
+ # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
539
+ def _prepare_decoder_attention_mask(
540
+ self, attention_mask, input_ids, inputs_embeds, past_key_values_length
541
+ ):
542
+ input_shape = input_ids.shape if input_ids is not None else inputs_embeds.shape[:-1]
543
+ # create causal mask
544
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
545
+ combined_attention_mask = None
546
+ if input_shape[-1] > 1:
547
+ combined_attention_mask = _make_causal_mask(
548
+ input_shape,
549
+ inputs_embeds.dtype,
550
+ device=inputs_embeds.device,
551
+ past_key_values_length=past_key_values_length,
552
+ )
553
+
554
+ if attention_mask is not None:
555
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
556
+ expanded_attn_mask = _expand_mask(
557
+ attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
558
+ ).to(inputs_embeds.device)
559
+ combined_attention_mask = (
560
+ expanded_attn_mask
561
+ if combined_attention_mask is None
562
+ else expanded_attn_mask + combined_attention_mask
563
+ )
564
+
565
+ return combined_attention_mask
566
+
567
+ @add_start_docstrings_to_model_forward(Yi_INPUTS_DOCSTRING)
568
+ def forward(
569
+ self,
570
+ input_ids: torch.LongTensor = None,
571
+ attention_mask: Optional[torch.Tensor] = None,
572
+ position_ids: Optional[torch.LongTensor] = None,
573
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
574
+ inputs_embeds: Optional[torch.FloatTensor] = None,
575
+ use_cache: Optional[bool] = None,
576
+ output_attentions: Optional[bool] = None,
577
+ output_hidden_states: Optional[bool] = None,
578
+ return_dict: Optional[bool] = None,
579
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
580
+ output_attentions = (
581
+ output_attentions
582
+ if output_attentions is not None
583
+ else self.config.output_attentions
584
+ )
585
+ output_hidden_states = (
586
+ output_hidden_states
587
+ if output_hidden_states is not None
588
+ else self.config.output_hidden_states
589
+ )
590
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
591
+
592
+ return_dict = (
593
+ return_dict if return_dict is not None else self.config.use_return_dict
594
+ )
595
+
596
+ # retrieve input_ids and inputs_embeds
597
+ if input_ids is not None and inputs_embeds is not None:
598
+ raise ValueError(
599
+ "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
600
+ )
601
+ elif input_ids is not None:
602
+ batch_size, seq_length = input_ids.shape
603
+ elif inputs_embeds is not None:
604
+ batch_size, seq_length, _ = inputs_embeds.shape
605
+ else:
606
+ raise ValueError(
607
+ "You have to specify either decoder_input_ids or decoder_inputs_embeds"
608
+ )
609
+
610
+ seq_length_with_past = seq_length
611
+ past_key_values_length = 0
612
+
613
+ if past_key_values is not None:
614
+ past_key_values_length = past_key_values[0][0].shape[2]
615
+ seq_length_with_past = seq_length_with_past + past_key_values_length
616
+
617
+ if position_ids is None:
618
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
619
+ position_ids = torch.arange(
620
+ past_key_values_length,
621
+ seq_length + past_key_values_length,
622
+ dtype=torch.long,
623
+ device=device,
624
+ )
625
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
626
+ else:
627
+ position_ids = position_ids.view(-1, seq_length).long()
628
+
629
+ if inputs_embeds is None:
630
+ inputs_embeds = self.embed_tokens(input_ids)
631
+
632
+ if not is_flash_attn_available:
633
+ # embed positions
634
+ if attention_mask is None:
635
+ attention_mask = torch.ones(
636
+ (batch_size, seq_length_with_past),
637
+ dtype=torch.bool,
638
+ device=inputs_embeds.device,
639
+ )
640
+ attention_mask = self._prepare_decoder_attention_mask(
641
+ attention_mask,
642
+ input_ids,
643
+ inputs_embeds,
644
+ past_key_values_length,
645
+ )
646
+ else:
647
+ attention_mask = None
648
+
649
+ hidden_states = inputs_embeds
650
+ if self.gradient_checkpointing and self.training:
651
+ if use_cache:
652
+ logger.warning_once(
653
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
654
+ )
655
+ use_cache = False
656
+
657
+ # decoder layers
658
+ all_hidden_states = () if output_hidden_states else None
659
+ all_self_attns = () if output_attentions else None
660
+ next_decoder_cache = () if use_cache else None
661
+
662
+ for idx, decoder_layer in enumerate(self.layers):
663
+ if output_hidden_states:
664
+ all_hidden_states += (hidden_states,)
665
+
666
+ past_key_value = (
667
+ past_key_values[idx] if past_key_values is not None else None
668
+ )
669
+
670
+ if self.gradient_checkpointing and self.training:
671
+
672
+ def create_custom_forward(module):
673
+ def custom_forward(*inputs):
674
+ # None for past_key_value
675
+ return module(*inputs, past_key_value, output_attentions)
676
+
677
+ return custom_forward
678
+
679
+ layer_outputs = torch.utils.checkpoint.checkpoint(
680
+ create_custom_forward(decoder_layer),
681
+ hidden_states,
682
+ attention_mask,
683
+ position_ids,
684
+ )
685
+ else:
686
+ layer_outputs = decoder_layer(
687
+ hidden_states,
688
+ attention_mask=attention_mask,
689
+ position_ids=position_ids,
690
+ past_key_value=past_key_value,
691
+ output_attentions=output_attentions,
692
+ use_cache=use_cache,
693
+ )
694
+
695
+ hidden_states = layer_outputs[0]
696
+
697
+ if use_cache:
698
+ next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
699
+
700
+ if output_attentions:
701
+ all_self_attns += (layer_outputs[1],)
702
+
703
+ hidden_states = self.norm(hidden_states)
704
+ # add hidden states from the last decoder layer
705
+ if output_hidden_states:
706
+ all_hidden_states += (hidden_states,)
707
+
708
+ next_cache = next_decoder_cache if use_cache else None
709
+ if not return_dict:
710
+ return tuple(
711
+ v
712
+ for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]
713
+ if v is not None
714
+ )
715
+ return BaseModelOutputWithPast(
716
+ last_hidden_state=hidden_states,
717
+ past_key_values=next_cache,
718
+ hidden_states=all_hidden_states,
719
+ attentions=all_self_attns,
720
+ )
721
+
722
+
723
+ class YiForCausalLM(YiPreTrainedModel):
724
+ _tied_weights_keys = ["lm_head.weight"]
725
+
726
+ def __init__(self, config):
727
+ super().__init__(config)
728
+ self.model = YiModel(config)
729
+
730
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
731
+
732
+ # Initialize weights and apply final processing
733
+ self.post_init()
734
+
735
+ def get_input_embeddings(self):
736
+ return self.model.embed_tokens
737
+
738
+ def set_input_embeddings(self, value):
739
+ self.model.embed_tokens = value
740
+
741
+ def get_output_embeddings(self):
742
+ return self.lm_head
743
+
744
+ def set_output_embeddings(self, new_embeddings):
745
+ self.lm_head = new_embeddings
746
+
747
+ def set_decoder(self, decoder):
748
+ self.model = decoder
749
+
750
+ def get_decoder(self):
751
+ return self.model
752
+
753
+ @add_start_docstrings_to_model_forward(Yi_INPUTS_DOCSTRING)
754
+ @replace_return_docstrings(
755
+ output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
756
+ )
757
+ def forward(
758
+ self,
759
+ input_ids: torch.LongTensor = None,
760
+ attention_mask: Optional[torch.Tensor] = None,
761
+ position_ids: Optional[torch.LongTensor] = None,
762
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
763
+ inputs_embeds: Optional[torch.FloatTensor] = None,
764
+ labels: Optional[torch.LongTensor] = None,
765
+ use_cache: Optional[bool] = None,
766
+ output_attentions: Optional[bool] = None,
767
+ output_hidden_states: Optional[bool] = None,
768
+ return_dict: Optional[bool] = None,
769
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
770
+ r"""
771
+ Args:
772
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
773
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
774
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
775
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
776
+
777
+ Returns:
778
+
779
+ Example:
780
+
781
+ ```python
782
+ >>> from transformers import AutoTokenizer, YiForCausalLM
783
+
784
+ >>> model = YiForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
785
+ >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
786
+
787
+ >>> prompt = "Hey, are you conscious? Can you talk to me?"
788
+ >>> inputs = tokenizer(prompt, return_tensors="pt")
789
+
790
+ >>> # Generate
791
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
792
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
793
+ "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
794
+ ```"""
795
+
796
+ output_attentions = (
797
+ output_attentions
798
+ if output_attentions is not None
799
+ else self.config.output_attentions
800
+ )
801
+ output_hidden_states = (
802
+ output_hidden_states
803
+ if output_hidden_states is not None
804
+ else self.config.output_hidden_states
805
+ )
806
+ return_dict = (
807
+ return_dict if return_dict is not None else self.config.use_return_dict
808
+ )
809
+
810
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
811
+ outputs = self.model(
812
+ input_ids=input_ids,
813
+ attention_mask=attention_mask,
814
+ position_ids=position_ids,
815
+ past_key_values=past_key_values,
816
+ inputs_embeds=inputs_embeds,
817
+ use_cache=use_cache,
818
+ output_attentions=output_attentions,
819
+ output_hidden_states=output_hidden_states,
820
+ return_dict=return_dict,
821
+ )
822
+
823
+ hidden_states = outputs[0]
824
+ logits = self.lm_head(hidden_states)
825
+
826
+ loss = None
827
+ if labels is not None:
828
+ # Shift so that tokens < n predict n
829
+ shift_logits = logits[..., :-1, :].contiguous()
830
+ shift_labels = labels[..., 1:].contiguous()
831
+ # Flatten the tokens
832
+ loss_fct = CrossEntropyLoss()
833
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
834
+ shift_labels = shift_labels.view(-1)
835
+ # Enable model parallelism
836
+ shift_labels = shift_labels.to(shift_logits.device)
837
+ loss = loss_fct(shift_logits, shift_labels)
838
+
839
+ if not return_dict:
840
+ output = (logits,) + outputs[1:]
841
+ return (loss,) + output if loss is not None else output
842
+
843
+ return CausalLMOutputWithPast(
844
+ loss=loss,
845
+ logits=logits,
846
+ past_key_values=outputs.past_key_values,
847
+ hidden_states=outputs.hidden_states,
848
+ attentions=outputs.attentions,
849
+ )
850
+
851
+ def prepare_inputs_for_generation(
852
+ self,
853
+ input_ids,
854
+ past_key_values=None,
855
+ attention_mask=None,
856
+ inputs_embeds=None,
857
+ **kwargs,
858
+ ):
859
+ if past_key_values:
860
+ input_ids = input_ids[:, -1:]
861
+
862
+ position_ids = kwargs.get("position_ids", None)
863
+ if attention_mask is not None and position_ids is None:
864
+ # create position_ids on the fly for batch generation
865
+ position_ids = attention_mask.long().cumsum(-1) - 1
866
+ position_ids.masked_fill_(attention_mask == 0, 1)
867
+ if past_key_values:
868
+ position_ids = position_ids[:, -1].unsqueeze(-1)
869
+
870
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
871
+ if inputs_embeds is not None and past_key_values is None:
872
+ model_inputs = {"inputs_embeds": inputs_embeds}
873
+ else:
874
+ model_inputs = {"input_ids": input_ids}
875
+
876
+ model_inputs.update(
877
+ {
878
+ "position_ids": position_ids,
879
+ "past_key_values": past_key_values,
880
+ "use_cache": kwargs.get("use_cache"),
881
+ "attention_mask": attention_mask,
882
+ }
883
+ )
884
+ return model_inputs
885
+
886
+ @staticmethod
887
+ def _reorder_cache(past_key_values, beam_idx):
888
+ reordered_past = ()
889
+ for layer_past in past_key_values:
890
+ reordered_past += (
891
+ tuple(
892
+ past_state.index_select(0, beam_idx.to(past_state.device))
893
+ for past_state in layer_past
894
+ ),
895
+ )
896
+ return reordered_past
897
+
898
+
899
+ @add_start_docstrings(
900
+ """
901
+ The Yi Model transformer with a sequence classification head on top (linear layer).
902
+
903
+ [`YiForSequenceClassification`] uses the last token in order to do the classification, as other causal models
904
+ (e.g. GPT-2) do.
905
+
906
+ Since it does classification on the last token, it requires to know the position of the last token. If a
907
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
908
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
909
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
910
+ each row of the batch).
911
+ """,
912
+ Yi_START_DOCSTRING,
913
+ )
914
+ class YiForSequenceClassification(YiPreTrainedModel):
915
+ def __init__(self, config):
916
+ super().__init__(config)
917
+ self.num_labels = config.num_labels
918
+ self.model = YiModel(config)
919
+ self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
920
+
921
+ # Initialize weights and apply final processing
922
+ self.post_init()
923
+
924
+ def get_input_embeddings(self):
925
+ return self.model.embed_tokens
926
+
927
+ def set_input_embeddings(self, value):
928
+ self.model.embed_tokens = value
929
+
930
+ @add_start_docstrings_to_model_forward(Yi_INPUTS_DOCSTRING)
931
+ def forward(
932
+ self,
933
+ input_ids: torch.LongTensor = None,
934
+ attention_mask: Optional[torch.Tensor] = None,
935
+ position_ids: Optional[torch.LongTensor] = None,
936
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
937
+ inputs_embeds: Optional[torch.FloatTensor] = None,
938
+ labels: Optional[torch.LongTensor] = None,
939
+ use_cache: Optional[bool] = None,
940
+ output_attentions: Optional[bool] = None,
941
+ output_hidden_states: Optional[bool] = None,
942
+ return_dict: Optional[bool] = None,
943
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
944
+ r"""
945
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
946
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
947
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
948
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
949
+ """
950
+ return_dict = (
951
+ return_dict if return_dict is not None else self.config.use_return_dict
952
+ )
953
+
954
+ transformer_outputs = self.model(
955
+ input_ids,
956
+ attention_mask=attention_mask,
957
+ position_ids=position_ids,
958
+ past_key_values=past_key_values,
959
+ inputs_embeds=inputs_embeds,
960
+ use_cache=use_cache,
961
+ output_attentions=output_attentions,
962
+ output_hidden_states=output_hidden_states,
963
+ return_dict=return_dict,
964
+ )
965
+ hidden_states = transformer_outputs[0]
966
+ logits = self.score(hidden_states)
967
+
968
+ if input_ids is not None:
969
+ batch_size = input_ids.shape[0]
970
+ else:
971
+ batch_size = inputs_embeds.shape[0]
972
+
973
+ if self.config.pad_token_id is None and batch_size != 1:
974
+ raise ValueError(
975
+ "Cannot handle batch sizes > 1 if no padding token is defined."
976
+ )
977
+ if self.config.pad_token_id is None:
978
+ sequence_lengths = -1
979
+ else:
980
+ if input_ids is not None:
981
+ sequence_lengths = (
982
+ torch.eq(input_ids, self.config.pad_token_id).long().argmax(-1) - 1
983
+ ).to(logits.device)
984
+ else:
985
+ sequence_lengths = -1
986
+
987
+ pooled_logits = logits[
988
+ torch.arange(batch_size, device=logits.device), sequence_lengths
989
+ ]
990
+
991
+ loss = None
992
+ if labels is not None:
993
+ labels = labels.to(logits.device)
994
+ if self.config.problem_type is None:
995
+ if self.num_labels == 1:
996
+ self.config.problem_type = "regression"
997
+ elif self.num_labels > 1 and (
998
+ labels.dtype == torch.long or labels.dtype == torch.int
999
+ ):
1000
+ self.config.problem_type = "single_label_classification"
1001
+ else:
1002
+ self.config.problem_type = "multi_label_classification"
1003
+
1004
+ if self.config.problem_type == "regression":
1005
+ loss_fct = MSELoss()
1006
+ if self.num_labels == 1:
1007
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1008
+ else:
1009
+ loss = loss_fct(pooled_logits, labels)
1010
+ elif self.config.problem_type == "single_label_classification":
1011
+ loss_fct = CrossEntropyLoss()
1012
+ loss = loss_fct(
1013
+ pooled_logits.view(-1, self.num_labels), labels.view(-1)
1014
+ )
1015
+ elif self.config.problem_type == "multi_label_classification":
1016
+ loss_fct = BCEWithLogitsLoss()
1017
+ loss = loss_fct(pooled_logits, labels)
1018
+ if not return_dict:
1019
+ output = (pooled_logits,) + transformer_outputs[1:]
1020
+ return ((loss,) + output) if loss is not None else output
1021
+
1022
+ return SequenceClassifierOutputWithPast(
1023
+ loss=loss,
1024
+ logits=pooled_logits,
1025
+ past_key_values=transformer_outputs.past_key_values,
1026
+ hidden_states=transformer_outputs.hidden_states,
1027
+ attentions=transformer_outputs.attentions,
1028
+ )
tokenization_yi.txt ADDED
@@ -0,0 +1,255 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ from shutil import copyfile
3
+ from typing import Any, Dict, List, Optional, Tuple
4
+
5
+ import sentencepiece as spm
6
+ from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
7
+ from transformers.utils import logging
8
+
9
+ logger = logging.get_logger(__name__)
10
+
11
+ VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
12
+
13
+ PRETRAINED_VOCAB_FILES_MAP = {
14
+ "vocab_file": {},
15
+ "tokenizer_file": {},
16
+ }
17
+ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {}
18
+
19
+
20
+ class YiTokenizer(PreTrainedTokenizer):
21
+ """
22
+ Construct a Yi tokenizer. Based on byte-level Byte-Pair-Encoding.
23
+
24
+ Args:
25
+ vocab_file (`str`):
26
+ Path to the vocabulary file.
27
+ """
28
+
29
+ vocab_files_names = VOCAB_FILES_NAMES
30
+ pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
31
+ max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
32
+ model_input_names = ["input_ids", "attention_mask"]
33
+
34
+ def __init__(
35
+ self,
36
+ vocab_file,
37
+ unk_token="<unk>",
38
+ bos_token="<|startoftext|>",
39
+ eos_token="<|endoftext|>",
40
+ pad_token="<unk>",
41
+ sp_model_kwargs: Optional[Dict[str, Any]] = None,
42
+ add_bos_token=True,
43
+ add_eos_token=False,
44
+ clean_up_tokenization_spaces=False,
45
+ **kwargs,
46
+ ):
47
+ self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
48
+ bos_token = (
49
+ AddedToken(bos_token, lstrip=False, rstrip=False)
50
+ if isinstance(bos_token, str)
51
+ else bos_token
52
+ )
53
+ eos_token = (
54
+ AddedToken(eos_token, lstrip=False, rstrip=False)
55
+ if isinstance(eos_token, str)
56
+ else eos_token
57
+ )
58
+ unk_token = (
59
+ AddedToken(unk_token, lstrip=False, rstrip=False)
60
+ if isinstance(unk_token, str)
61
+ else unk_token
62
+ )
63
+ pad_token = (
64
+ AddedToken(pad_token, lstrip=False, rstrip=False)
65
+ if isinstance(pad_token, str)
66
+ else pad_token
67
+ )
68
+ self.vocab_file = vocab_file
69
+ self.add_bos_token = add_bos_token
70
+ self.add_eos_token = add_eos_token
71
+ self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
72
+ self.sp_model.Load(vocab_file)
73
+ super().__init__(
74
+ bos_token=bos_token,
75
+ eos_token=eos_token,
76
+ unk_token=unk_token,
77
+ pad_token=pad_token,
78
+ add_bos_token=add_bos_token,
79
+ add_eos_token=add_eos_token,
80
+ sp_model_kwargs=self.sp_model_kwargs,
81
+ clean_up_tokenization_spaces=clean_up_tokenization_spaces,
82
+ **kwargs,
83
+ )
84
+
85
+ def __getstate__(self):
86
+ state = self.__dict__.copy()
87
+ state["sp_model"] = None
88
+ return state
89
+
90
+ def __setstate__(self, d):
91
+ self.__dict__ = d
92
+ self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
93
+ self.sp_model.Load(self.vocab_file)
94
+
95
+ @property
96
+ def vocab_size(self):
97
+ """Returns vocab size"""
98
+ return self.sp_model.get_piece_size()
99
+
100
+ def get_vocab(self):
101
+ """Returns vocab as a dict"""
102
+ vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
103
+ vocab.update(self.added_tokens_encoder)
104
+ return vocab
105
+
106
+ def _tokenize(self, text):
107
+ """Returns a tokenized string."""
108
+ return self.sp_model.encode(text, out_type=str)
109
+
110
+ def _convert_token_to_id(self, token):
111
+ """Converts a token (str) in an id using the vocab."""
112
+ return self.sp_model.piece_to_id(token)
113
+
114
+ def _convert_id_to_token(self, index):
115
+ """Converts an index (integer) in a token (str) using the vocab."""
116
+ token = self.sp_model.IdToPiece(index)
117
+ return token
118
+
119
+ def convert_tokens_to_string(self, tokens):
120
+ """Converts a sequence of tokens (string) in a single string."""
121
+ current_sub_tokens = []
122
+ out_string = ""
123
+ prev_is_special = False
124
+ for i, token in enumerate(tokens):
125
+ # make sure that special tokens are not decoded using sentencepiece model
126
+ if token in self.all_special_tokens:
127
+ if not prev_is_special and i != 0:
128
+ out_string += " "
129
+ out_string += self.sp_model.decode(current_sub_tokens) + token
130
+ prev_is_special = True
131
+ current_sub_tokens = []
132
+ else:
133
+ current_sub_tokens.append(token)
134
+ prev_is_special = False
135
+ out_string += self.sp_model.decode(current_sub_tokens)
136
+ return out_string
137
+
138
+ def save_vocabulary(
139
+ self, save_directory, filename_prefix: Optional[str] = None
140
+ ) -> Tuple[str]:
141
+ """
142
+ Save the vocabulary and special tokens file to a directory.
143
+
144
+ Args:
145
+ save_directory (`str`):
146
+ The directory in which to save the vocabulary.
147
+
148
+ Returns:
149
+ `Tuple(str)`: Paths to the files saved.
150
+ """
151
+ if not os.path.isdir(save_directory):
152
+ logger.error(f"Vocabulary path ({save_directory}) should be a directory")
153
+ return
154
+ out_vocab_file = os.path.join(
155
+ save_directory,
156
+ (filename_prefix + "-" if filename_prefix else "")
157
+ + VOCAB_FILES_NAMES["vocab_file"],
158
+ )
159
+
160
+ if os.path.abspath(self.vocab_file) != os.path.abspath(
161
+ out_vocab_file
162
+ ) 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,
184
+ token_ids_0: List[int],
185
+ token_ids_1: Optional[List[int]] = None,
186
+ already_has_special_tokens: bool = False,
187
+ ) -> List[int]:
188
+ """
189
+ Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
190
+ special tokens using the tokenizer `prepare_for_model` method.
191
+
192
+ Args:
193
+ token_ids_0 (`List[int]`):
194
+ List of IDs.
195
+ token_ids_1 (`List[int]`, *optional*):
196
+ Optional second list of IDs for sequence pairs.
197
+ already_has_special_tokens (`bool`, *optional*, defaults to `False`):
198
+ Whether or not the token list is already formatted with special tokens for the model.
199
+
200
+ Returns:
201
+ `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
202
+ """
203
+ if already_has_special_tokens:
204
+ return super().get_special_tokens_mask(
205
+ token_ids_0=token_ids_0,
206
+ token_ids_1=token_ids_1,
207
+ already_has_special_tokens=True,
208
+ )
209
+
210
+ bos_token_id = [1] if self.add_bos_token else []
211
+ eos_token_id = [1] if self.add_eos_token else []
212
+
213
+ if token_ids_1 is None:
214
+ return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
215
+ return (
216
+ bos_token_id
217
+ + ([0] * len(token_ids_0))
218
+ + eos_token_id
219
+ + bos_token_id
220
+ + ([0] * len(token_ids_1))
221
+ + eos_token_id
222
+ )
223
+
224
+ def create_token_type_ids_from_sequences(
225
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
226
+ ) -> List[int]:
227
+ """
228
+ Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
229
+ sequence pair mask has the following format:
230
+
231
+ ```
232
+ 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
233
+ | first sequence | second sequence |
234
+ ```
235
+
236
+ if token_ids_1 is None, only returns the first portion of the mask (0s).
237
+
238
+ Args:
239
+ token_ids_0 (`List[int]`):
240
+ List of ids.
241
+ token_ids_1 (`List[int]`, *optional*):
242
+ Optional second list of IDs for sequence pairs.
243
+
244
+ Returns:
245
+ `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
246
+ """
247
+ bos_token_id = [self.bos_token_id] if self.add_bos_token else []
248
+ eos_token_id = [self.eos_token_id] if self.add_eos_token else []
249
+
250
+ output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
251
+
252
+ if token_ids_1 is not None:
253
+ output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
254
+
255
+ return output
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:386c49cf943d71aa110361135338c50e38beeff0a66593480421f37b319e1a39
3
+ size 1033105
tokenizer.txt ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.txt ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "auto_map": {
3
+ "AutoTokenizer": ["tokenization_yi.YiTokenizer", null]
4
+ },
5
+ "add_bos_token": false,
6
+ "add_eos_token": false,
7
+ "model_max_length": 4096,
8
+ "tokenizer_class": "YiTokenizer"
9
+ }