shaohang commited on
Commit
1956c78
1 Parent(s): bca5743

Upload 3 files

Browse files
Files changed (3) hide show
  1. configuration_baichuan.py +66 -0
  2. handler.py +27 -0
  3. modeling_baichuan.py +671 -0
configuration_baichuan.py ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
+ # and OPT implementations in this library. It has been modified from its
6
+ # original forms to accommodate minor architectural differences compared
7
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
8
+ #
9
+ # Licensed under the Apache License, Version 2.0 (the "License");
10
+ # you may not use this file except in compliance with the License.
11
+ # You may obtain a copy of the License at
12
+ #
13
+ # http://www.apache.org/licenses/LICENSE-2.0
14
+ #
15
+ # Unless required by applicable law or agreed to in writing, software
16
+ # distributed under the License is distributed on an "AS IS" BASIS,
17
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
+ # See the License for the specific language governing permissions and
19
+ # limitations under the License.
20
+
21
+ from transformers.configuration_utils import PretrainedConfig
22
+ from transformers.utils import logging
23
+
24
+
25
+ logger = logging.get_logger(__name__)
26
+
27
+
28
+ class BaiChuanConfig(PretrainedConfig):
29
+ model_type = "baichuan"
30
+ keys_to_ignore_at_inference = ["past_key_values"]
31
+
32
+ def __init__(
33
+ self,
34
+ vocab_size=64000,
35
+ hidden_size=4096,
36
+ intermediate_size=11008,
37
+ num_hidden_layers=32,
38
+ num_attention_heads=32,
39
+ hidden_act="silu",
40
+ max_position_embeddings=4096,
41
+ initializer_range=0.02,
42
+ rms_norm_eps=1e-6,
43
+ use_cache=True,
44
+ pad_token_id=0,
45
+ bos_token_id=1,
46
+ eos_token_id=2,
47
+ tie_word_embeddings=False,
48
+ **kwargs,
49
+ ):
50
+ self.vocab_size = vocab_size
51
+ self.max_position_embeddings = max_position_embeddings
52
+ self.hidden_size = hidden_size
53
+ self.intermediate_size = intermediate_size
54
+ self.num_hidden_layers = num_hidden_layers
55
+ self.num_attention_heads = num_attention_heads
56
+ self.hidden_act = hidden_act
57
+ self.initializer_range = initializer_range
58
+ self.rms_norm_eps = rms_norm_eps
59
+ self.use_cache = use_cache
60
+ super().__init__(
61
+ pad_token_id=pad_token_id,
62
+ bos_token_id=bos_token_id,
63
+ eos_token_id=eos_token_id,
64
+ tie_word_embeddings=tie_word_embeddings,
65
+ **kwargs,
66
+ )
handler.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from typing import Dict, List, Any
3
+ from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
4
+
5
+ # get dtype
6
+ dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] == 8 else torch.float16
7
+
8
+
9
+ class EndpointHandler:
10
+ def __init__(self, path=""):
11
+ # load the model
12
+ tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
13
+ model = AutoModelForCausalLM.from_pretrained(path, device_map="auto", torch_dtype=dtype, trust_remote_code=True)
14
+ # create inference pipeline
15
+ self.pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer)
16
+
17
+ def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
18
+ inputs = data.pop("inputs", data)
19
+ parameters = data.pop("parameters", None)
20
+
21
+ # pass inputs with all kwargs in data
22
+ if parameters is not None:
23
+ prediction = self.pipeline(inputs, **parameters)
24
+ else:
25
+ prediction = self.pipeline(inputs)
26
+ # postprocess the prediction
27
+ return prediction
modeling_baichuan.py ADDED
@@ -0,0 +1,671 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
+ # and OPT implementations in this library. It has been modified from its
6
+ # original forms to accommodate minor architectural differences compared
7
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
8
+ #
9
+ # Licensed under the Apache License, Version 2.0 (the "License");
10
+ # you may not use this file except in compliance with the License.
11
+ # You may obtain a copy of the License at
12
+ #
13
+ # http://www.apache.org/licenses/LICENSE-2.0
14
+ #
15
+ # Unless required by applicable law or agreed to in writing, software
16
+ # distributed under the License is distributed on an "AS IS" BASIS,
17
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
+ # See the License for the specific language governing permissions and
19
+ # limitations under the License.
20
+ from .configuration_baichuan import BaiChuanConfig
21
+ from transformers import PreTrainedModel, add_start_docstrings
22
+ from transformers.activations import ACT2FN
23
+ from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, \
24
+ SequenceClassifierOutputWithPast
25
+ from transformers.utils import logging, add_start_docstrings_to_model_forward, replace_return_docstrings
26
+
27
+ import math
28
+ from typing import List, Optional, Tuple, Union
29
+
30
+ import torch
31
+ import torch.utils.checkpoint
32
+ from torch import nn
33
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
34
+
35
+
36
+ logger = logging.get_logger(__name__)
37
+
38
+ # Copied from transformers.models.bart.modeling_bart._make_causal_mask
39
+ def _make_causal_mask(
40
+ input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
41
+ ):
42
+ """
43
+ Make causal mask used for bi-directional self-attention.
44
+ """
45
+ bsz, tgt_len = input_ids_shape
46
+ mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
47
+ mask_cond = torch.arange(mask.size(-1), device=device)
48
+ mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
49
+ mask = mask.to(dtype)
50
+
51
+ if past_key_values_length > 0:
52
+ mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
53
+ return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
54
+
55
+
56
+ # Copied from transformers.models.bart.modeling_bart._expand_mask
57
+ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
58
+ """
59
+ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
60
+ """
61
+ bsz, src_len = mask.size()
62
+ tgt_len = tgt_len if tgt_len is not None else src_len
63
+
64
+ expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
65
+
66
+ inverted_mask = 1.0 - expanded_mask
67
+
68
+ return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
69
+
70
+
71
+ class RMSNorm(nn.Module):
72
+ def __init__(self, hidden_size, eps=1e-6):
73
+ """
74
+ RMSNorm is equivalent to T5LayerNorm
75
+ """
76
+ super().__init__()
77
+ self.weight = nn.Parameter(torch.ones(hidden_size))
78
+ self.variance_epsilon = eps
79
+
80
+ def forward(self, hidden_states):
81
+ variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
82
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
83
+
84
+ # convert into half-precision if necessary
85
+ if self.weight.dtype in [torch.float16, torch.bfloat16]:
86
+ hidden_states = hidden_states.to(self.weight.dtype)
87
+
88
+ return self.weight * hidden_states
89
+
90
+
91
+ class RotaryEmbedding(torch.nn.Module):
92
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
93
+ super().__init__()
94
+ inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
95
+ self.register_buffer("inv_freq", inv_freq)
96
+
97
+ # Build here to make `torch.jit.trace` work.
98
+ self.max_seq_len_cached = max_position_embeddings
99
+ t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
100
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
101
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
102
+ emb = torch.cat((freqs, freqs), dim=-1)
103
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
104
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
105
+
106
+ def forward(self, x, seq_len=None):
107
+ # x: [bs, num_attention_heads, seq_len, head_size]
108
+ # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
109
+ if seq_len > self.max_seq_len_cached:
110
+ self.max_seq_len_cached = seq_len
111
+ t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
112
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
113
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
114
+ emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
115
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
116
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
117
+ return (
118
+ self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
119
+ self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
120
+ )
121
+
122
+
123
+ def rotate_half(x):
124
+ """Rotates half the hidden dims of the input."""
125
+ x1 = x[..., : x.shape[-1] // 2]
126
+ x2 = x[..., x.shape[-1] // 2:]
127
+ return torch.cat((-x2, x1), dim=-1)
128
+
129
+
130
+ def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
131
+ # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
132
+ cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
133
+ sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
134
+ cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
135
+ sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
136
+ q_embed = (q * cos) + (rotate_half(q) * sin)
137
+ k_embed = (k * cos) + (rotate_half(k) * sin)
138
+ return q_embed, k_embed
139
+
140
+
141
+ class MLP(nn.Module):
142
+ def __init__(
143
+ self,
144
+ hidden_size: int,
145
+ intermediate_size: int,
146
+ hidden_act: str,
147
+ ):
148
+ super().__init__()
149
+ self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
150
+ self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
151
+ self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
152
+ self.act_fn = ACT2FN[hidden_act]
153
+
154
+ def forward(self, x):
155
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
156
+
157
+
158
+ class Attention(nn.Module):
159
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
160
+
161
+ def __init__(self, config: BaiChuanConfig):
162
+ super().__init__()
163
+ self.config = config
164
+ self.hidden_size = config.hidden_size
165
+ self.num_heads = config.num_attention_heads
166
+ self.head_dim = self.hidden_size // self.num_heads
167
+ self.max_position_embeddings = config.max_position_embeddings
168
+
169
+ if (self.head_dim * self.num_heads) != self.hidden_size:
170
+ raise ValueError(
171
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
172
+ f" and `num_heads`: {self.num_heads})."
173
+ )
174
+ self.W_pack = nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=False)
175
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
176
+ self.rotary_emb = RotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
177
+
178
+ def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
179
+ return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
180
+
181
+ def forward(
182
+ self,
183
+ hidden_states: torch.Tensor,
184
+ attention_mask: Optional[torch.Tensor] = None,
185
+ position_ids: Optional[torch.LongTensor] = None,
186
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
187
+ output_attentions: bool = False,
188
+ use_cache: bool = False,
189
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
190
+ bsz, q_len, _ = hidden_states.size()
191
+
192
+ proj = self.W_pack(hidden_states)
193
+ proj = proj.unflatten(-1, (3, self.hidden_size)).unsqueeze(0).transpose(0, -2).squeeze(-2)
194
+ query_states = proj[0].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1,
195
+ 2) # batch_size x source_len x hidden_size
196
+ key_states = proj[1].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1,
197
+ 2) # batch_size x target_len x head_size
198
+ value_states = proj[2].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1,
199
+ 2) # batch_size x source_len x hidden_size
200
+
201
+ kv_seq_len = key_states.shape[-2]
202
+ if past_key_value is not None:
203
+ kv_seq_len += past_key_value[0].shape[-2]
204
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
205
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
206
+ # [bsz, nh, t, hd]
207
+
208
+ if past_key_value is not None:
209
+ # reuse k, v, self_attention
210
+ key_states = torch.cat([past_key_value[0], key_states], dim=2)
211
+ value_states = torch.cat([past_key_value[1], value_states], dim=2)
212
+
213
+ past_key_value = (key_states, value_states) if use_cache else None
214
+
215
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
216
+
217
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
218
+ raise ValueError(
219
+ f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
220
+ f" {attn_weights.size()}"
221
+ )
222
+
223
+ if attention_mask is not None:
224
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
225
+ raise ValueError(
226
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
227
+ )
228
+ attn_weights = attn_weights + attention_mask
229
+ attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
230
+
231
+ # upcast attention to fp32
232
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
233
+ attn_output = torch.matmul(attn_weights, value_states)
234
+
235
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
236
+ raise ValueError(
237
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
238
+ f" {attn_output.size()}"
239
+ )
240
+
241
+ attn_output = attn_output.transpose(1, 2)
242
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
243
+
244
+ attn_output = self.o_proj(attn_output)
245
+
246
+ if not output_attentions:
247
+ attn_weights = None
248
+
249
+ return attn_output, attn_weights, past_key_value
250
+
251
+
252
+ class DecoderLayer(nn.Module):
253
+ def __init__(self, config: BaiChuanConfig):
254
+ super().__init__()
255
+ self.hidden_size = config.hidden_size
256
+ self.self_attn = Attention(config=config)
257
+ self.mlp = MLP(
258
+ hidden_size=self.hidden_size,
259
+ intermediate_size=config.intermediate_size,
260
+ hidden_act=config.hidden_act,
261
+ )
262
+ self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
263
+ self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
264
+
265
+ def forward(
266
+ self,
267
+ hidden_states: torch.Tensor,
268
+ attention_mask: Optional[torch.Tensor] = None,
269
+ position_ids: Optional[torch.LongTensor] = None,
270
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
271
+ output_attentions: Optional[bool] = False,
272
+ use_cache: Optional[bool] = False,
273
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
274
+ """
275
+ Args:
276
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
277
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
278
+ `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
279
+ output_attentions (`bool`, *optional*):
280
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
281
+ returned tensors for more detail.
282
+ use_cache (`bool`, *optional*):
283
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
284
+ (see `past_key_values`).
285
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
286
+ """
287
+
288
+ residual = hidden_states
289
+
290
+ hidden_states = self.input_layernorm(hidden_states)
291
+
292
+ # Self Attention
293
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
294
+ hidden_states=hidden_states,
295
+ attention_mask=attention_mask,
296
+ position_ids=position_ids,
297
+ past_key_value=past_key_value,
298
+ output_attentions=output_attentions,
299
+ use_cache=use_cache,
300
+ )
301
+ hidden_states = residual + hidden_states
302
+
303
+ # Fully Connected
304
+ residual = hidden_states
305
+ hidden_states = self.post_attention_layernorm(hidden_states)
306
+ hidden_states = self.mlp(hidden_states)
307
+ hidden_states = residual + hidden_states
308
+
309
+ outputs = (hidden_states,)
310
+
311
+ if output_attentions:
312
+ outputs += (self_attn_weights,)
313
+
314
+ if use_cache:
315
+ outputs += (present_key_value,)
316
+
317
+ return outputs
318
+
319
+
320
+ class PreTrainedModel(PreTrainedModel):
321
+ config_class = BaiChuanConfig
322
+ base_model_prefix = "model"
323
+ supports_gradient_checkpointing = True
324
+ _no_split_modules = ["DecoderLayer"]
325
+ _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
326
+
327
+ def _init_weights(self, module):
328
+ std = self.config.initializer_range
329
+ if isinstance(module, nn.Linear):
330
+ module.weight.data.normal_(mean=0.0, std=std)
331
+ if module.bias is not None:
332
+ module.bias.data.zero_()
333
+ elif isinstance(module, nn.Embedding):
334
+ module.weight.data.normal_(mean=0.0, std=std)
335
+ if module.padding_idx is not None:
336
+ module.weight.data[module.padding_idx].zero_()
337
+
338
+ def _set_gradient_checkpointing(self, module, value=False):
339
+ if isinstance(module, Model):
340
+ module.gradient_checkpointing = value
341
+
342
+
343
+ class Model(PreTrainedModel):
344
+ """
345
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`DecoderLayer`]
346
+
347
+ Args:
348
+ config: BaiChuanConfig
349
+ """
350
+
351
+ def __init__(self, config: BaiChuanConfig):
352
+ super().__init__(config)
353
+ self.padding_idx = config.pad_token_id
354
+ self.vocab_size = config.vocab_size
355
+
356
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
357
+ self.layers = nn.ModuleList([DecoderLayer(config) for _ in range(config.num_hidden_layers)])
358
+ self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
359
+
360
+ self.gradient_checkpointing = False
361
+ # Initialize weights and apply final processing
362
+ self.post_init()
363
+
364
+ def get_input_embeddings(self):
365
+ return self.embed_tokens
366
+
367
+ def set_input_embeddings(self, value):
368
+ self.embed_tokens = value
369
+
370
+ # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
371
+ def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
372
+ # create causal mask
373
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
374
+ combined_attention_mask = None
375
+ if input_shape[-1] > 1:
376
+ combined_attention_mask = _make_causal_mask(
377
+ input_shape,
378
+ inputs_embeds.dtype,
379
+ device=inputs_embeds.device,
380
+ past_key_values_length=past_key_values_length,
381
+ )
382
+
383
+ if attention_mask is not None:
384
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
385
+ expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
386
+ inputs_embeds.device
387
+ )
388
+ combined_attention_mask = (
389
+ expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
390
+ )
391
+
392
+ return combined_attention_mask
393
+
394
+ def forward(
395
+ self,
396
+ input_ids: torch.LongTensor = None,
397
+ attention_mask: Optional[torch.Tensor] = None,
398
+ position_ids: Optional[torch.LongTensor] = None,
399
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
400
+ inputs_embeds: Optional[torch.FloatTensor] = None,
401
+ use_cache: Optional[bool] = None,
402
+ output_attentions: Optional[bool] = None,
403
+ output_hidden_states: Optional[bool] = None,
404
+ return_dict: Optional[bool] = None,
405
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
406
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
407
+ output_hidden_states = (
408
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
409
+ )
410
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
411
+
412
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
413
+
414
+ # retrieve input_ids and inputs_embeds
415
+ if input_ids is not None and inputs_embeds is not None:
416
+ raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
417
+ elif input_ids is not None:
418
+ batch_size, seq_length = input_ids.shape
419
+ elif inputs_embeds is not None:
420
+ batch_size, seq_length, _ = inputs_embeds.shape
421
+ else:
422
+ raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
423
+
424
+ seq_length_with_past = seq_length
425
+ past_key_values_length = 0
426
+
427
+ if past_key_values is not None:
428
+ past_key_values_length = past_key_values[0][0].shape[2]
429
+ seq_length_with_past = seq_length_with_past + past_key_values_length
430
+
431
+ if position_ids is None:
432
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
433
+ position_ids = torch.arange(
434
+ past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
435
+ )
436
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
437
+ else:
438
+ position_ids = position_ids.view(-1, seq_length).long()
439
+
440
+ if inputs_embeds is None:
441
+ inputs_embeds = self.embed_tokens(input_ids)
442
+ # embed positions
443
+ if attention_mask is None:
444
+ attention_mask = torch.ones(
445
+ (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
446
+ )
447
+ attention_mask = self._prepare_decoder_attention_mask(
448
+ attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
449
+ )
450
+
451
+ hidden_states = inputs_embeds
452
+
453
+ if self.gradient_checkpointing and self.training:
454
+ if use_cache:
455
+ logger.warning_once(
456
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
457
+ )
458
+ use_cache = False
459
+
460
+ # decoder layers
461
+ all_hidden_states = () if output_hidden_states else None
462
+ all_self_attns = () if output_attentions else None
463
+ next_decoder_cache = () if use_cache else None
464
+
465
+ for idx, decoder_layer in enumerate(self.layers):
466
+ if output_hidden_states:
467
+ all_hidden_states += (hidden_states,)
468
+
469
+ past_key_value = past_key_values[idx] if past_key_values is not None else None
470
+
471
+ if self.gradient_checkpointing and self.training:
472
+
473
+ def create_custom_forward(module):
474
+ def custom_forward(*inputs):
475
+ # None for past_key_value
476
+ return module(*inputs, output_attentions, None)
477
+
478
+ return custom_forward
479
+
480
+ layer_outputs = torch.utils.checkpoint.checkpoint(
481
+ create_custom_forward(decoder_layer),
482
+ hidden_states,
483
+ attention_mask,
484
+ position_ids,
485
+ None,
486
+ )
487
+ else:
488
+ layer_outputs = decoder_layer(
489
+ hidden_states,
490
+ attention_mask=attention_mask,
491
+ position_ids=position_ids,
492
+ past_key_value=past_key_value,
493
+ output_attentions=output_attentions,
494
+ use_cache=use_cache,
495
+ )
496
+
497
+ hidden_states = layer_outputs[0]
498
+
499
+ if use_cache:
500
+ next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
501
+
502
+ if output_attentions:
503
+ all_self_attns += (layer_outputs[1],)
504
+
505
+ hidden_states = self.norm(hidden_states)
506
+
507
+ # add hidden states from the last decoder layer
508
+ if output_hidden_states:
509
+ all_hidden_states += (hidden_states,)
510
+
511
+ next_cache = next_decoder_cache if use_cache else None
512
+ if not return_dict:
513
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
514
+ return BaseModelOutputWithPast(
515
+ last_hidden_state=hidden_states,
516
+ past_key_values=next_cache,
517
+ hidden_states=all_hidden_states,
518
+ attentions=all_self_attns,
519
+ )
520
+
521
+
522
+ class BaiChuanForCausalLM(PreTrainedModel):
523
+ def __init__(self, config):
524
+ super().__init__(config)
525
+ self.model = Model(config)
526
+
527
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
528
+
529
+ # Initialize weights and apply final processing
530
+ self.post_init()
531
+
532
+ def get_input_embeddings(self):
533
+ return self.model.embed_tokens
534
+
535
+ def set_input_embeddings(self, value):
536
+ self.model.embed_tokens = value
537
+
538
+ def get_output_embeddings(self):
539
+ return self.lm_head
540
+
541
+ def set_output_embeddings(self, new_embeddings):
542
+ self.lm_head = new_embeddings
543
+
544
+ def set_decoder(self, decoder):
545
+ self.model = decoder
546
+
547
+ def get_decoder(self):
548
+ return self.model
549
+
550
+ def forward(
551
+ self,
552
+ input_ids: torch.LongTensor = None,
553
+ attention_mask: Optional[torch.Tensor] = None,
554
+ position_ids: Optional[torch.LongTensor] = None,
555
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
556
+ inputs_embeds: Optional[torch.FloatTensor] = None,
557
+ labels: Optional[torch.LongTensor] = None,
558
+ use_cache: Optional[bool] = None,
559
+ output_attentions: Optional[bool] = None,
560
+ output_hidden_states: Optional[bool] = None,
561
+ return_dict: Optional[bool] = None,
562
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
563
+ r"""
564
+ Args:
565
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
566
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
567
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
568
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
569
+
570
+ Returns:
571
+
572
+ Example:
573
+
574
+ ```python
575
+ >>> from transformers import AutoTokenizer, ModelForCausalLM
576
+
577
+ >>> model = ModelForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
578
+ >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
579
+
580
+ >>> prompt = "Hey, are you consciours? Can you talk to me?"
581
+ >>> inputs = tokenizer(prompt, return_tensors="pt")
582
+
583
+ >>> # Generate
584
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
585
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
586
+ "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
587
+ ```"""
588
+
589
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
590
+ output_hidden_states = (
591
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
592
+ )
593
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
594
+
595
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
596
+ outputs = self.model(
597
+ input_ids=input_ids,
598
+ attention_mask=attention_mask,
599
+ position_ids=position_ids,
600
+ past_key_values=past_key_values,
601
+ inputs_embeds=inputs_embeds,
602
+ use_cache=use_cache,
603
+ output_attentions=output_attentions,
604
+ output_hidden_states=output_hidden_states,
605
+ return_dict=return_dict,
606
+ )
607
+
608
+ hidden_states = outputs[0]
609
+ logits = self.lm_head(hidden_states)
610
+
611
+ loss = None
612
+ if labels is not None:
613
+ # Shift so that tokens < n predict n
614
+ shift_logits = logits[..., :-1, :].contiguous()
615
+ shift_labels = labels[..., 1:].contiguous()
616
+ # Flatten the tokens
617
+ loss_fct = CrossEntropyLoss()
618
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
619
+ shift_labels = shift_labels.view(-1)
620
+ # Enable model parallelism
621
+ shift_labels = shift_labels.to(shift_logits.device)
622
+ loss = loss_fct(shift_logits, shift_labels)
623
+
624
+ if not return_dict:
625
+ output = (logits,) + outputs[1:]
626
+ return (loss,) + output if loss is not None else output
627
+
628
+ return CausalLMOutputWithPast(
629
+ loss=loss,
630
+ logits=logits,
631
+ past_key_values=outputs.past_key_values,
632
+ hidden_states=outputs.hidden_states,
633
+ attentions=outputs.attentions,
634
+ )
635
+
636
+ def prepare_inputs_for_generation(
637
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
638
+ ):
639
+ if past_key_values:
640
+ input_ids = input_ids[:, -1:]
641
+
642
+ position_ids = kwargs.get("position_ids", None)
643
+ if attention_mask is not None and position_ids is None:
644
+ # create position_ids on the fly for batch generation
645
+ position_ids = attention_mask.long().cumsum(-1) - 1
646
+ position_ids.masked_fill_(attention_mask == 0, 1)
647
+ if past_key_values:
648
+ position_ids = position_ids[:, -1].unsqueeze(-1)
649
+
650
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
651
+ if inputs_embeds is not None and past_key_values is None:
652
+ model_inputs = {"inputs_embeds": inputs_embeds}
653
+ else:
654
+ model_inputs = {"input_ids": input_ids}
655
+
656
+ model_inputs.update(
657
+ {
658
+ "position_ids": position_ids,
659
+ "past_key_values": past_key_values,
660
+ "use_cache": kwargs.get("use_cache"),
661
+ "attention_mask": attention_mask,
662
+ }
663
+ )
664
+ return model_inputs
665
+
666
+ @staticmethod
667
+ def _reorder_cache(past_key_values, beam_idx):
668
+ reordered_past = ()
669
+ for layer_past in past_key_values:
670
+ reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
671
+ return reordered_past