Matt commited on
Commit
d4dbe8b
1 Parent(s): d3397bb

Remove custom code

Browse files
configuration_internlm.py DELETED
@@ -1,120 +0,0 @@
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
- """ InternLM model configuration"""
21
-
22
- from transformers.utils import logging
23
- from transformers.configuration_utils import PretrainedConfig
24
-
25
-
26
- logger = logging.get_logger(__name__)
27
-
28
- INTERNLM_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
29
-
30
-
31
- class InternLMConfig(PretrainedConfig):
32
- r"""
33
- This is the configuration class to store the configuration of a [`InternLMModel`]. It is used to instantiate an InternLM
34
- model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
35
- defaults will yield a similar configuration to that of the InternLM-7B.
36
-
37
- Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
38
- documentation from [`PretrainedConfig`] for more information.
39
-
40
-
41
- Args:
42
- vocab_size (`int`, *optional*, defaults to 32000):
43
- Vocabulary size of the InternLM model. Defines the number of different tokens that can be represented by the
44
- `inputs_ids` passed when calling [`InternLMModel`]
45
- hidden_size (`int`, *optional*, defaults to 4096):
46
- Dimension of the hidden representations.
47
- intermediate_size (`int`, *optional*, defaults to 11008):
48
- Dimension of the MLP representations.
49
- num_hidden_layers (`int`, *optional*, defaults to 32):
50
- Number of hidden layers in the Transformer encoder.
51
- num_attention_heads (`int`, *optional*, defaults to 32):
52
- Number of attention heads for each attention layer in the Transformer encoder.
53
- hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
54
- The non-linear activation function (function or string) in the decoder.
55
- max_position_embeddings (`int`, *optional*, defaults to 2048):
56
- The maximum sequence length that this model might ever be used with. Typically set this to something large
57
- just in case (e.g., 512 or 1024 or 2048).
58
- initializer_range (`float`, *optional*, defaults to 0.02):
59
- The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
60
- rms_norm_eps (`float`, *optional*, defaults to 1e-12):
61
- The epsilon used by the rms normalization layers.
62
- use_cache (`bool`, *optional*, defaults to `True`):
63
- Whether or not the model should return the last key/values attentions (not used by all models). Only
64
- relevant if `config.is_decoder=True`.
65
- tie_word_embeddings(`bool`, *optional*, defaults to `False`):
66
- Whether to tie weight embeddings
67
- Example:
68
-
69
- ```python
70
- >>> from transformers import InternLMModel, InternLMConfig
71
-
72
- >>> # Initializing a InternLM internlm-7b style configuration
73
- >>> configuration = InternLMConfig()
74
-
75
- >>> # Initializing a model from the internlm-7b style configuration
76
- >>> model = InternLMModel(configuration)
77
-
78
- >>> # Accessing the model configuration
79
- >>> configuration = model.config
80
- ```"""
81
- model_type = "internlm"
82
- _auto_class = "AutoConfig"
83
-
84
- def __init__(
85
- self,
86
- vocab_size=103168,
87
- hidden_size=4096,
88
- intermediate_size=11008,
89
- num_hidden_layers=32,
90
- num_attention_heads=32,
91
- hidden_act="silu",
92
- max_position_embeddings=2048,
93
- initializer_range=0.02,
94
- rms_norm_eps=1e-6,
95
- use_cache=True,
96
- pad_token_id=0,
97
- bos_token_id=1,
98
- eos_token_id=2,
99
- tie_word_embeddings=False,
100
- bias=True,
101
- **kwargs,
102
- ):
103
- self.vocab_size = vocab_size
104
- self.max_position_embeddings = max_position_embeddings
105
- self.hidden_size = hidden_size
106
- self.intermediate_size = intermediate_size
107
- self.num_hidden_layers = num_hidden_layers
108
- self.num_attention_heads = num_attention_heads
109
- self.hidden_act = hidden_act
110
- self.initializer_range = initializer_range
111
- self.rms_norm_eps = rms_norm_eps
112
- self.use_cache = use_cache
113
- self.bias = bias
114
- super().__init__(
115
- pad_token_id=pad_token_id,
116
- bos_token_id=bos_token_id,
117
- eos_token_id=eos_token_id,
118
- tie_word_embeddings=tie_word_embeddings,
119
- **kwargs,
120
- )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
modeling_internlm.py DELETED
@@ -1,996 +0,0 @@
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
- """ PyTorch InternLM model."""
21
- import math
22
- from typing import List, Optional, Tuple, Union
23
- import threading, queue
24
-
25
- import torch
26
- import torch.utils.checkpoint
27
- from torch import nn
28
- from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
29
-
30
- from transformers.activations import ACT2FN
31
- from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
32
- from transformers.modeling_utils import PreTrainedModel
33
- from transformers.generation.streamers import BaseStreamer
34
- from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
35
- from .configuration_internlm import InternLMConfig
36
-
37
-
38
- logger = logging.get_logger(__name__)
39
-
40
- _CONFIG_FOR_DOC = "InternLMConfig"
41
-
42
- # Copied from transformers.models.bart.modeling_bart._make_causal_mask
43
- def _make_causal_mask(
44
- input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
45
- ):
46
- """
47
- Make causal mask used for bi-directional self-attention.
48
- """
49
- bsz, tgt_len = input_ids_shape
50
- mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
51
- mask_cond = torch.arange(mask.size(-1), device=device)
52
- mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
53
- mask = mask.to(dtype)
54
-
55
- if past_key_values_length > 0:
56
- mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
57
- return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
58
-
59
-
60
- # Copied from transformers.models.bart.modeling_bart._expand_mask
61
- def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
62
- """
63
- Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
64
- """
65
- bsz, src_len = mask.size()
66
- tgt_len = tgt_len if tgt_len is not None else src_len
67
-
68
- expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
69
-
70
- inverted_mask = 1.0 - expanded_mask
71
-
72
- return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
73
-
74
-
75
- class InternLMRMSNorm(nn.Module):
76
- def __init__(self, hidden_size, eps=1e-6):
77
- """
78
- InternLMRMSNorm is equivalent to T5LayerNorm
79
- """
80
- super().__init__()
81
- self.weight = nn.Parameter(torch.ones(hidden_size))
82
- self.variance_epsilon = eps
83
-
84
- def forward(self, hidden_states):
85
- variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
86
- hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
87
-
88
- # convert into half-precision if necessary
89
- if self.weight.dtype in [torch.float16, torch.bfloat16]:
90
- hidden_states = hidden_states.to(self.weight.dtype)
91
-
92
- return self.weight * hidden_states
93
-
94
-
95
- class InternLMRotaryEmbedding(torch.nn.Module):
96
- def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
97
- super().__init__()
98
- inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
99
- self.register_buffer("inv_freq", inv_freq, persistent=False)
100
-
101
- # Build here to make `torch.jit.trace` work.
102
- self.max_seq_len_cached = max_position_embeddings
103
- t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
104
- freqs = torch.einsum("i,j->ij", t, self.inv_freq)
105
- # Different from paper, but it uses a different permutation in order to obtain the same calculation
106
- emb = torch.cat((freqs, freqs), dim=-1)
107
- self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
108
- self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
109
-
110
- def forward(self, x, seq_len=None):
111
- # x: [bs, num_attention_heads, seq_len, head_size]
112
- # This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
113
- if seq_len > self.max_seq_len_cached:
114
- self.max_seq_len_cached = seq_len
115
- t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
116
- freqs = torch.einsum("i,j->ij", t, self.inv_freq)
117
- # Different from paper, but it uses a different permutation in order to obtain the same calculation
118
- emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
119
- self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
120
- self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
121
- return (
122
- self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
123
- self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
124
- )
125
-
126
-
127
- def rotate_half(x):
128
- """Rotates half the hidden dims of the input."""
129
- x1 = x[..., : x.shape[-1] // 2]
130
- x2 = x[..., x.shape[-1] // 2 :]
131
- return torch.cat((-x2, x1), dim=-1)
132
-
133
-
134
- def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
135
- # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
136
- cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
137
- sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
138
- cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
139
- sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
140
- q_embed = (q * cos) + (rotate_half(q) * sin)
141
- k_embed = (k * cos) + (rotate_half(k) * sin)
142
- return q_embed, k_embed
143
-
144
-
145
- class InternLMMLP(nn.Module):
146
- def __init__(
147
- self,
148
- hidden_size: int,
149
- intermediate_size: int,
150
- hidden_act: str,
151
- ):
152
- super().__init__()
153
- self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
154
- self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
155
- self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
156
- self.act_fn = ACT2FN[hidden_act]
157
-
158
- def forward(self, x):
159
- return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
160
-
161
-
162
- class InternLMAttention(nn.Module):
163
- """Multi-headed attention from 'Attention Is All You Need' paper"""
164
-
165
- def __init__(self, config: InternLMConfig):
166
- super().__init__()
167
- self.config = config
168
- self.hidden_size = config.hidden_size
169
- self.num_heads = config.num_attention_heads
170
- self.head_dim = self.hidden_size // self.num_heads
171
- self.max_position_embeddings = config.max_position_embeddings
172
-
173
- if (self.head_dim * self.num_heads) != self.hidden_size:
174
- raise ValueError(
175
- f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
176
- f" and `num_heads`: {self.num_heads})."
177
- )
178
- self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.bias)
179
- self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.bias)
180
- self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.bias)
181
- self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.bias)
182
- self.rotary_emb = InternLMRotaryEmbedding(self.head_dim, max_position_embeddings=self.max_position_embeddings)
183
-
184
- def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
185
- return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
186
-
187
- def forward(
188
- self,
189
- hidden_states: torch.Tensor,
190
- attention_mask: Optional[torch.Tensor] = None,
191
- position_ids: Optional[torch.LongTensor] = None,
192
- past_key_value: Optional[Tuple[torch.Tensor]] = None,
193
- output_attentions: bool = False,
194
- use_cache: bool = False,
195
- ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
196
- bsz, q_len, _ = hidden_states.size()
197
-
198
- query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
199
- key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
200
- value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
201
-
202
- kv_seq_len = key_states.shape[-2]
203
- if past_key_value is not None:
204
- kv_seq_len += past_key_value[0].shape[-2]
205
- cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
206
- query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
207
- # [bsz, nh, t, hd]
208
-
209
- if past_key_value is not None:
210
- # reuse k, v, self_attention
211
- key_states = torch.cat([past_key_value[0], key_states], dim=2)
212
- value_states = torch.cat([past_key_value[1], value_states], dim=2)
213
-
214
- past_key_value = (key_states, value_states) if use_cache else None
215
-
216
- attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
217
-
218
- if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
219
- raise ValueError(
220
- f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
221
- f" {attn_weights.size()}"
222
- )
223
-
224
- if attention_mask is not None:
225
- if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
226
- raise ValueError(
227
- f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
228
- )
229
- attn_weights = attn_weights + attention_mask
230
- attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
231
-
232
- # upcast attention to fp32
233
- attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
234
- attn_output = torch.matmul(attn_weights, value_states)
235
-
236
- if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
237
- raise ValueError(
238
- f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
239
- f" {attn_output.size()}"
240
- )
241
-
242
- attn_output = attn_output.transpose(1, 2)
243
- attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
244
-
245
- attn_output = self.o_proj(attn_output)
246
-
247
- if not output_attentions:
248
- attn_weights = None
249
-
250
- return attn_output, attn_weights, past_key_value
251
-
252
-
253
- class InternLMDecoderLayer(nn.Module):
254
- def __init__(self, config: InternLMConfig):
255
- super().__init__()
256
- self.hidden_size = config.hidden_size
257
- self.self_attn = InternLMAttention(config=config)
258
- self.mlp = InternLMMLP(
259
- hidden_size=self.hidden_size,
260
- intermediate_size=config.intermediate_size,
261
- hidden_act=config.hidden_act,
262
- )
263
- self.input_layernorm = InternLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
264
- self.post_attention_layernorm = InternLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
265
-
266
- def forward(
267
- self,
268
- hidden_states: torch.Tensor,
269
- attention_mask: Optional[torch.Tensor] = None,
270
- position_ids: Optional[torch.LongTensor] = None,
271
- past_key_value: Optional[Tuple[torch.Tensor]] = None,
272
- output_attentions: Optional[bool] = False,
273
- use_cache: Optional[bool] = False,
274
- ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
275
- """
276
- Args:
277
- hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
278
- attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
279
- `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
280
- output_attentions (`bool`, *optional*):
281
- Whether or not to return the attentions tensors of all attention layers. See `attentions` under
282
- returned tensors for more detail.
283
- use_cache (`bool`, *optional*):
284
- If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
285
- (see `past_key_values`).
286
- past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
287
- """
288
-
289
- residual = hidden_states
290
-
291
- hidden_states = self.input_layernorm(hidden_states)
292
-
293
- # Self Attention
294
- hidden_states, self_attn_weights, present_key_value = self.self_attn(
295
- hidden_states=hidden_states,
296
- attention_mask=attention_mask,
297
- position_ids=position_ids,
298
- past_key_value=past_key_value,
299
- output_attentions=output_attentions,
300
- use_cache=use_cache,
301
- )
302
- hidden_states = residual + hidden_states
303
-
304
- # Fully Connected
305
- residual = hidden_states
306
- hidden_states = self.post_attention_layernorm(hidden_states)
307
- hidden_states = self.mlp(hidden_states)
308
- hidden_states = residual + hidden_states
309
-
310
- outputs = (hidden_states,)
311
-
312
- if output_attentions:
313
- outputs += (self_attn_weights,)
314
-
315
- if use_cache:
316
- outputs += (present_key_value,)
317
-
318
- return outputs
319
-
320
-
321
- INTERNLM_START_DOCSTRING = r"""
322
- This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
323
- library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
324
- etc.)
325
-
326
- This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
327
- Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
328
- and behavior.
329
-
330
- Parameters:
331
- config ([`InternLMConfig`]):
332
- Model configuration class with all the parameters of the model. Initializing with a config file does not
333
- load the weights associated with the model, only the configuration. Check out the
334
- [`~PreTrainedModel.from_pretrained`] method to load the model weights.
335
- """
336
-
337
-
338
- @add_start_docstrings(
339
- "The bare InternLM Model outputting raw hidden-states without any specific head on top.",
340
- INTERNLM_START_DOCSTRING,
341
- )
342
- class InternLMPreTrainedModel(PreTrainedModel):
343
- config_class = InternLMConfig
344
- base_model_prefix = "model"
345
- supports_gradient_checkpointing = True
346
- _no_split_modules = ["InternLMDecoderLayer"]
347
- _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
348
-
349
- def _init_weights(self, module):
350
- std = self.config.initializer_range
351
- if isinstance(module, nn.Linear):
352
- module.weight.data.normal_(mean=0.0, std=std)
353
- if module.bias is not None:
354
- module.bias.data.zero_()
355
- elif isinstance(module, nn.Embedding):
356
- module.weight.data.normal_(mean=0.0, std=std)
357
- if module.padding_idx is not None:
358
- module.weight.data[module.padding_idx].zero_()
359
-
360
- def _set_gradient_checkpointing(self, module, value=False):
361
- if isinstance(module, InternLMModel):
362
- module.gradient_checkpointing = value
363
-
364
-
365
- INTERNLM_INPUTS_DOCSTRING = r"""
366
- Args:
367
- input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
368
- Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
369
- it.
370
-
371
- Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
372
- [`PreTrainedTokenizer.__call__`] for details.
373
-
374
- [What are input IDs?](../glossary#input-ids)
375
- attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
376
- Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
377
-
378
- - 1 for tokens that are **not masked**,
379
- - 0 for tokens that are **masked**.
380
-
381
- [What are attention masks?](../glossary#attention-mask)
382
-
383
- Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
384
- [`PreTrainedTokenizer.__call__`] for details.
385
-
386
- If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
387
- `past_key_values`).
388
-
389
- If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
390
- and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
391
- information on the default strategy.
392
-
393
- - 1 indicates the head is **not masked**,
394
- - 0 indicates the head is **masked**.
395
- position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
396
- Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
397
- config.n_positions - 1]`.
398
-
399
- [What are position IDs?](../glossary#position-ids)
400
- past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
401
- Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
402
- `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
403
- `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
404
-
405
- Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
406
- blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
407
-
408
- If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
409
- don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
410
- `decoder_input_ids` of shape `(batch_size, sequence_length)`.
411
- inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
412
- Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
413
- is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
414
- model's internal embedding lookup matrix.
415
- use_cache (`bool`, *optional*):
416
- If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
417
- `past_key_values`).
418
- output_attentions (`bool`, *optional*):
419
- Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
420
- tensors for more detail.
421
- output_hidden_states (`bool`, *optional*):
422
- Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
423
- more detail.
424
- return_dict (`bool`, *optional*):
425
- Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
426
- """
427
-
428
-
429
- @add_start_docstrings(
430
- "The bare InternLM Model outputting raw hidden-states without any specific head on top.",
431
- INTERNLM_START_DOCSTRING,
432
- )
433
- class InternLMModel(InternLMPreTrainedModel):
434
- """
435
- Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`InternLMDecoderLayer`]
436
-
437
- Args:
438
- config: InternLMConfig
439
- """
440
- _auto_class = "AutoModel"
441
-
442
- def __init__(self, config: InternLMConfig):
443
- super().__init__(config)
444
- self.padding_idx = config.pad_token_id
445
- self.vocab_size = config.vocab_size
446
-
447
- self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
448
- self.layers = nn.ModuleList([InternLMDecoderLayer(config) for _ in range(config.num_hidden_layers)])
449
- self.norm = InternLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
450
-
451
- self.gradient_checkpointing = False
452
- # Initialize weights and apply final processing
453
- self.post_init()
454
-
455
- def get_input_embeddings(self):
456
- return self.embed_tokens
457
-
458
- def set_input_embeddings(self, value):
459
- self.embed_tokens = value
460
-
461
- # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
462
- def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
463
- # create causal mask
464
- # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
465
- combined_attention_mask = None
466
- if input_shape[-1] > 1:
467
- combined_attention_mask = _make_causal_mask(
468
- input_shape,
469
- inputs_embeds.dtype,
470
- device=inputs_embeds.device,
471
- past_key_values_length=past_key_values_length,
472
- )
473
-
474
- if attention_mask is not None:
475
- # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
476
- expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
477
- inputs_embeds.device
478
- )
479
- combined_attention_mask = (
480
- expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
481
- )
482
-
483
- return combined_attention_mask
484
-
485
- @add_start_docstrings_to_model_forward(INTERNLM_INPUTS_DOCSTRING)
486
- def forward(
487
- self,
488
- input_ids: torch.LongTensor = None,
489
- attention_mask: Optional[torch.Tensor] = None,
490
- position_ids: Optional[torch.LongTensor] = None,
491
- past_key_values: Optional[List[torch.FloatTensor]] = None,
492
- inputs_embeds: Optional[torch.FloatTensor] = None,
493
- use_cache: Optional[bool] = None,
494
- output_attentions: Optional[bool] = None,
495
- output_hidden_states: Optional[bool] = None,
496
- return_dict: Optional[bool] = None,
497
- ) -> Union[Tuple, BaseModelOutputWithPast]:
498
- output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
499
- output_hidden_states = (
500
- output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
501
- )
502
- use_cache = use_cache if use_cache is not None else self.config.use_cache
503
-
504
- return_dict = return_dict if return_dict is not None else self.config.use_return_dict
505
-
506
- # retrieve input_ids and inputs_embeds
507
- if input_ids is not None and inputs_embeds is not None:
508
- raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
509
- elif input_ids is not None:
510
- batch_size, seq_length = input_ids.shape
511
- elif inputs_embeds is not None:
512
- batch_size, seq_length, _ = inputs_embeds.shape
513
- else:
514
- raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
515
-
516
- seq_length_with_past = seq_length
517
- past_key_values_length = 0
518
-
519
- if past_key_values is not None:
520
- past_key_values_length = past_key_values[0][0].shape[2]
521
- seq_length_with_past = seq_length_with_past + past_key_values_length
522
-
523
- if position_ids is None:
524
- device = input_ids.device if input_ids is not None else inputs_embeds.device
525
- position_ids = torch.arange(
526
- past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
527
- )
528
- position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
529
- else:
530
- position_ids = position_ids.view(-1, seq_length).long()
531
-
532
- if inputs_embeds is None:
533
- inputs_embeds = self.embed_tokens(input_ids)
534
- # embed positions
535
- if attention_mask is None:
536
- attention_mask = torch.ones(
537
- (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
538
- )
539
- attention_mask = self._prepare_decoder_attention_mask(
540
- attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
541
- )
542
-
543
- hidden_states = inputs_embeds
544
-
545
- if self.gradient_checkpointing and self.training:
546
- if use_cache:
547
- logger.warning_once(
548
- "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
549
- )
550
- use_cache = False
551
-
552
- # decoder layers
553
- all_hidden_states = () if output_hidden_states else None
554
- all_self_attns = () if output_attentions else None
555
- next_decoder_cache = () if use_cache else None
556
-
557
- for idx, decoder_layer in enumerate(self.layers):
558
- if output_hidden_states:
559
- all_hidden_states += (hidden_states,)
560
-
561
- past_key_value = past_key_values[idx] if past_key_values is not None else None
562
-
563
- if self.gradient_checkpointing and self.training:
564
-
565
- def create_custom_forward(module):
566
- def custom_forward(*inputs):
567
- # None for past_key_value
568
- return module(*inputs, output_attentions, None)
569
-
570
- return custom_forward
571
-
572
- layer_outputs = torch.utils.checkpoint.checkpoint(
573
- create_custom_forward(decoder_layer),
574
- hidden_states,
575
- attention_mask,
576
- position_ids,
577
- None,
578
- )
579
- else:
580
- layer_outputs = decoder_layer(
581
- hidden_states,
582
- attention_mask=attention_mask,
583
- position_ids=position_ids,
584
- past_key_value=past_key_value,
585
- output_attentions=output_attentions,
586
- use_cache=use_cache,
587
- )
588
-
589
- hidden_states = layer_outputs[0]
590
-
591
- if use_cache:
592
- next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
593
-
594
- if output_attentions:
595
- all_self_attns += (layer_outputs[1],)
596
-
597
- hidden_states = self.norm(hidden_states)
598
-
599
- # add hidden states from the last decoder layer
600
- if output_hidden_states:
601
- all_hidden_states += (hidden_states,)
602
-
603
- next_cache = next_decoder_cache if use_cache else None
604
- if not return_dict:
605
- return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
606
- return BaseModelOutputWithPast(
607
- last_hidden_state=hidden_states,
608
- past_key_values=next_cache,
609
- hidden_states=all_hidden_states,
610
- attentions=all_self_attns,
611
- )
612
-
613
-
614
- class InternLMForCausalLM(InternLMPreTrainedModel):
615
- _auto_class = "AutoModelForCausalLM"
616
-
617
- def __init__(self, config):
618
- super().__init__(config)
619
- self.model = InternLMModel(config)
620
-
621
- self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
622
-
623
- # Initialize weights and apply final processing
624
- self.post_init()
625
-
626
- def get_input_embeddings(self):
627
- return self.model.embed_tokens
628
-
629
- def set_input_embeddings(self, value):
630
- self.model.embed_tokens = value
631
-
632
- def get_output_embeddings(self):
633
- return self.lm_head
634
-
635
- def set_output_embeddings(self, new_embeddings):
636
- self.lm_head = new_embeddings
637
-
638
- def set_decoder(self, decoder):
639
- self.model = decoder
640
-
641
- def get_decoder(self):
642
- return self.model
643
-
644
- @add_start_docstrings_to_model_forward(INTERNLM_INPUTS_DOCSTRING)
645
- @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
646
- def forward(
647
- self,
648
- input_ids: torch.LongTensor = None,
649
- attention_mask: Optional[torch.Tensor] = None,
650
- position_ids: Optional[torch.LongTensor] = None,
651
- past_key_values: Optional[List[torch.FloatTensor]] = None,
652
- inputs_embeds: Optional[torch.FloatTensor] = None,
653
- labels: Optional[torch.LongTensor] = None,
654
- use_cache: Optional[bool] = None,
655
- output_attentions: Optional[bool] = None,
656
- output_hidden_states: Optional[bool] = None,
657
- return_dict: Optional[bool] = None,
658
- ) -> Union[Tuple, CausalLMOutputWithPast]:
659
- r"""
660
- Args:
661
- labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
662
- Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
663
- config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
664
- (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
665
-
666
- Returns:
667
-
668
- Example:
669
-
670
- ```python
671
- >>> from transformers import AutoTokenizer, InternLMForCausalLM
672
-
673
- >>> model = InternLMForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
674
- >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
675
-
676
- >>> prompt = "Hey, are you consciours? Can you talk to me?"
677
- >>> inputs = tokenizer(prompt, return_tensors="pt")
678
-
679
- >>> # Generate
680
- >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
681
- >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
682
- "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you."
683
- ```"""
684
-
685
- output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
686
- output_hidden_states = (
687
- output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
688
- )
689
- return_dict = return_dict if return_dict is not None else self.config.use_return_dict
690
-
691
- # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
692
- outputs = self.model(
693
- input_ids=input_ids,
694
- attention_mask=attention_mask,
695
- position_ids=position_ids,
696
- past_key_values=past_key_values,
697
- inputs_embeds=inputs_embeds,
698
- use_cache=use_cache,
699
- output_attentions=output_attentions,
700
- output_hidden_states=output_hidden_states,
701
- return_dict=return_dict,
702
- )
703
-
704
- hidden_states = outputs[0]
705
- logits = self.lm_head(hidden_states)
706
-
707
- loss = None
708
- if labels is not None:
709
- # Shift so that tokens < n predict n
710
- shift_logits = logits[..., :-1, :].contiguous()
711
- shift_labels = labels[..., 1:].contiguous()
712
- # Flatten the tokens
713
- loss_fct = CrossEntropyLoss()
714
- shift_logits = shift_logits.view(-1, self.config.vocab_size)
715
- shift_labels = shift_labels.view(-1)
716
- # Enable model parallelism
717
- shift_labels = shift_labels.to(shift_logits.device)
718
- loss = loss_fct(shift_logits, shift_labels)
719
-
720
- if not return_dict:
721
- output = (logits,) + outputs[1:]
722
- return (loss,) + output if loss is not None else output
723
-
724
- return CausalLMOutputWithPast(
725
- loss=loss,
726
- logits=logits,
727
- past_key_values=outputs.past_key_values,
728
- hidden_states=outputs.hidden_states,
729
- attentions=outputs.attentions,
730
- )
731
-
732
- def prepare_inputs_for_generation(
733
- self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
734
- ):
735
- if past_key_values:
736
- input_ids = input_ids[:, -1:]
737
-
738
- position_ids = kwargs.get("position_ids", None)
739
- if attention_mask is not None and position_ids is None:
740
- # create position_ids on the fly for batch generation
741
- position_ids = attention_mask.long().cumsum(-1) - 1
742
- position_ids.masked_fill_(attention_mask == 0, 1)
743
- if past_key_values:
744
- position_ids = position_ids[:, -1].unsqueeze(-1)
745
-
746
- # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
747
- if inputs_embeds is not None and past_key_values is None:
748
- model_inputs = {"inputs_embeds": inputs_embeds}
749
- else:
750
- model_inputs = {"input_ids": input_ids}
751
-
752
- model_inputs.update(
753
- {
754
- "position_ids": position_ids,
755
- "past_key_values": past_key_values,
756
- "use_cache": kwargs.get("use_cache"),
757
- "attention_mask": attention_mask,
758
- }
759
- )
760
- return model_inputs
761
-
762
- @staticmethod
763
- def _reorder_cache(past_key_values, beam_idx):
764
- reordered_past = ()
765
- for layer_past in past_key_values:
766
- reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
767
- return reordered_past
768
-
769
- def build_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = []):
770
- prompt = ""
771
- for record in history:
772
- prompt += f"""<|User|>:{record[0]}<eoh>\n<|Bot|>:{record[1]}<eoa>\n"""
773
- prompt += f"""<|User|>:{query}<eoh>\n<|Bot|>:"""
774
- return tokenizer([prompt], return_tensors="pt")
775
-
776
- @torch.no_grad()
777
- def chat(self,
778
- tokenizer,
779
- query: str,
780
- history: List[Tuple[str, str]] = [],
781
- streamer: Optional[BaseStreamer] = None,
782
- max_new_tokens: int = 1024,
783
- do_sample: bool = True,
784
- temperature: float = 0.8,
785
- top_p: float = 0.8,
786
- **kwargs):
787
- inputs = self.build_inputs(tokenizer, query, history)
788
- inputs = {k: v.to(self.device) for k, v in inputs.items() if torch.is_tensor(v)}
789
- outputs = self.generate(**inputs,
790
- streamer=streamer,
791
- max_new_tokens=max_new_tokens,
792
- do_sample=do_sample,
793
- temperature=temperature,
794
- top_p=top_p,
795
- **kwargs)
796
- outputs = outputs[0].cpu().tolist()[len(inputs["input_ids"][0]):]
797
- response = tokenizer.decode(outputs, skip_special_tokens=True)
798
- response = response.split("<eoa>")[0]
799
- history = history + [(query, response)]
800
- return response, history
801
-
802
- @torch.no_grad()
803
- def stream_chat(self,
804
- tokenizer,
805
- query: str,
806
- history: List[Tuple[str, str]] = [],
807
- max_new_tokens: int = 1024,
808
- do_sample: bool = True,
809
- temperature: float = 0.8,
810
- top_p: float = 0.8,
811
- **kwargs):
812
- """
813
- Return a generator in format: (response, history)
814
- Eg.
815
- ('你好,有什么可以帮助您的吗', [('你好', '你好,有什么可以帮助您的吗')])
816
- ('你好,有什么可以帮助您的吗?', [('你好', '你好,有什么可以帮助您的吗?')])
817
- """
818
-
819
- response_queue = queue.Queue(maxsize=20)
820
-
821
- class ChatStreamer(BaseStreamer):
822
- def __init__(self, tokenizer) -> None:
823
- super().__init__()
824
- self.tokenizer = tokenizer
825
- self.queue = response_queue
826
- self.query = query
827
- self.history = history
828
- self.response = ""
829
- self.received_inputs = False
830
- self.queue.put((self.response, history + [(self.query, self.response)]))
831
-
832
- def put(self, value):
833
- if len(value.shape) > 1 and value.shape[0] > 1:
834
- raise ValueError("ChatStreamer only supports batch size 1")
835
- elif len(value.shape) > 1:
836
- value = value[0]
837
-
838
- if not self.received_inputs:
839
- # The first received value is input_ids, ignore here
840
- self.received_inputs = True
841
- return
842
-
843
- token = self.tokenizer.decode([value[-1]], skip_special_tokens=True)
844
- if token.strip() != "<eoa>":
845
- self.response = self.response + token
846
- history = self.history + [(self.query, self.response)]
847
- self.queue.put((self.response, history))
848
-
849
- def end(self):
850
- self.queue.put(None)
851
-
852
- def stream_producer():
853
- return self.chat(
854
- tokenizer=tokenizer,
855
- query=query,
856
- streamer=ChatStreamer(tokenizer=tokenizer),
857
- history=history,
858
- max_new_tokens=max_new_tokens,
859
- do_sample=do_sample,
860
- temperature=temperature,
861
- top_p=top_p,
862
- **kwargs
863
- )
864
-
865
- def consumer():
866
- producer = threading.Thread(target=stream_producer)
867
- producer.start()
868
- while True:
869
- res = response_queue.get()
870
- if res is None:
871
- return
872
- yield res
873
-
874
- return consumer()
875
-
876
-
877
- @add_start_docstrings(
878
- """
879
- The InternLM Model transformer with a sequence classification head on top (linear layer).
880
-
881
- [`InternLMForSequenceClassification`] uses the last token in order to do the classification, as other causal models
882
- (e.g. GPT-2) do.
883
-
884
- Since it does classification on the last token, it requires to know the position of the last token. If a
885
- `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
886
- no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
887
- padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
888
- each row of the batch).
889
- """,
890
- INTERNLM_START_DOCSTRING,
891
- )
892
- class InternLMForSequenceClassification(InternLMPreTrainedModel):
893
- _keys_to_ignore_on_load_missing = [r"lm_head.weight"]
894
-
895
- def __init__(self, config):
896
- super().__init__(config)
897
- self.num_labels = config.num_labels
898
- self.model = InternLMModel(config)
899
- self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
900
-
901
- # Initialize weights and apply final processing
902
- self.post_init()
903
-
904
- def get_input_embeddings(self):
905
- return self.model.embed_tokens
906
-
907
- def set_input_embeddings(self, value):
908
- self.model.embed_tokens = value
909
-
910
- @add_start_docstrings_to_model_forward(INTERNLM_INPUTS_DOCSTRING)
911
- def forward(
912
- self,
913
- input_ids: torch.LongTensor = None,
914
- attention_mask: Optional[torch.Tensor] = None,
915
- position_ids: Optional[torch.LongTensor] = None,
916
- past_key_values: Optional[List[torch.FloatTensor]] = None,
917
- inputs_embeds: Optional[torch.FloatTensor] = None,
918
- labels: Optional[torch.LongTensor] = None,
919
- use_cache: Optional[bool] = None,
920
- output_attentions: Optional[bool] = None,
921
- output_hidden_states: Optional[bool] = None,
922
- return_dict: Optional[bool] = None,
923
- ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
924
- r"""
925
- labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
926
- Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
927
- config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
928
- `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
929
- """
930
- return_dict = return_dict if return_dict is not None else self.config.use_return_dict
931
-
932
- transformer_outputs = self.model(
933
- input_ids,
934
- attention_mask=attention_mask,
935
- position_ids=position_ids,
936
- past_key_values=past_key_values,
937
- inputs_embeds=inputs_embeds,
938
- use_cache=use_cache,
939
- output_attentions=output_attentions,
940
- output_hidden_states=output_hidden_states,
941
- return_dict=return_dict,
942
- )
943
- hidden_states = transformer_outputs[0]
944
- logits = self.score(hidden_states)
945
-
946
- if input_ids is not None:
947
- batch_size = input_ids.shape[0]
948
- else:
949
- batch_size = inputs_embeds.shape[0]
950
-
951
- if self.config.pad_token_id is None and batch_size != 1:
952
- raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
953
- if self.config.pad_token_id is None:
954
- sequence_lengths = -1
955
- else:
956
- if input_ids is not None:
957
- sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
958
- else:
959
- sequence_lengths = -1
960
-
961
- pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
962
-
963
- loss = None
964
- if labels is not None:
965
- labels = labels.to(logits.device)
966
- if self.config.problem_type is None:
967
- if self.num_labels == 1:
968
- self.config.problem_type = "regression"
969
- elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
970
- self.config.problem_type = "single_label_classification"
971
- else:
972
- self.config.problem_type = "multi_label_classification"
973
-
974
- if self.config.problem_type == "regression":
975
- loss_fct = MSELoss()
976
- if self.num_labels == 1:
977
- loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
978
- else:
979
- loss = loss_fct(pooled_logits, labels)
980
- elif self.config.problem_type == "single_label_classification":
981
- loss_fct = CrossEntropyLoss()
982
- loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
983
- elif self.config.problem_type == "multi_label_classification":
984
- loss_fct = BCEWithLogitsLoss()
985
- loss = loss_fct(pooled_logits, labels)
986
- if not return_dict:
987
- output = (pooled_logits,) + transformer_outputs[1:]
988
- return ((loss,) + output) if loss is not None else output
989
-
990
- return SequenceClassifierOutputWithPast(
991
- loss=loss,
992
- logits=pooled_logits,
993
- past_key_values=transformer_outputs.past_key_values,
994
- hidden_states=transformer_outputs.hidden_states,
995
- attentions=transformer_outputs.attentions,
996
- )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
tokenization_internlm.py DELETED
@@ -1,242 +0,0 @@
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
- """Tokenization classes for IntermLM."""
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 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
-
38
-
39
- class InternLMTokenizer(PreTrainedTokenizer):
40
- """
41
- Construct a InternLM tokenizer. Based on byte-level Byte-Pair-Encoding.
42
-
43
- Args:
44
- vocab_file (`str`):
45
- Path to the vocabulary file.
46
- """
47
-
48
- vocab_files_names = VOCAB_FILES_NAMES
49
- pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
50
- model_input_names = ["input_ids", "attention_mask"]
51
- _auto_class = "AutoTokenizer"
52
-
53
- def __init__(
54
- self,
55
- vocab_file,
56
- unk_token="<unk>",
57
- bos_token="<s>",
58
- eos_token="</s>",
59
- pad_token="</s>",
60
- sp_model_kwargs: Optional[Dict[str, Any]] = None,
61
- add_bos_token=True,
62
- add_eos_token=False,
63
- decode_with_prefix_space=False,
64
- clean_up_tokenization_spaces=False,
65
- **kwargs,
66
- ):
67
- self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
68
- super().__init__(
69
- bos_token=bos_token,
70
- eos_token=eos_token,
71
- unk_token=unk_token,
72
- pad_token=pad_token,
73
- clean_up_tokenization_spaces=clean_up_tokenization_spaces,
74
- **kwargs,
75
- )
76
- self.vocab_file = vocab_file
77
- self.add_bos_token = add_bos_token
78
- self.add_eos_token = add_eos_token
79
- self.decode_with_prefix_space = decode_with_prefix_space
80
- self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
81
- self.sp_model.Load(vocab_file)
82
- self._no_prefix_space_tokens = None
83
-
84
- """ Initialisation"""
85
-
86
- @property
87
- def no_prefix_space_tokens(self):
88
- if self._no_prefix_space_tokens is None:
89
- vocab = self.convert_ids_to_tokens(list(range(self.vocab_size)))
90
- self._no_prefix_space_tokens = {i for i, tok in enumerate(vocab) if not tok.startswith("▁")}
91
- return self._no_prefix_space_tokens
92
-
93
- @property
94
- def vocab_size(self):
95
- """Returns vocab size"""
96
- return self.sp_model.get_piece_size()
97
-
98
- @property
99
- def bos_token_id(self) -> Optional[int]:
100
- return self.sp_model.bos_id()
101
-
102
- @property
103
- def eos_token_id(self) -> Optional[int]:
104
- return self.sp_model.eos_id()
105
-
106
- def get_vocab(self):
107
- """Returns vocab as a dict"""
108
- vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
109
- vocab.update(self.added_tokens_encoder)
110
- return vocab
111
-
112
- def _tokenize(self, text):
113
- """Returns a tokenized string."""
114
- return self.sp_model.encode(text, out_type=str)
115
-
116
- def _convert_token_to_id(self, token):
117
- """Converts a token (str) in an id using the vocab."""
118
- return self.sp_model.piece_to_id(token)
119
-
120
- def _convert_id_to_token(self, index):
121
- """Converts an index (integer) in a token (str) using the vocab."""
122
- token = self.sp_model.IdToPiece(index)
123
- return token
124
-
125
- def _maybe_add_prefix_space(self, tokens, decoded):
126
- if tokens and tokens[0] not in self.no_prefix_space_tokens:
127
- return " " + decoded
128
- else:
129
- return decoded
130
-
131
- def convert_tokens_to_string(self, tokens):
132
- """Converts a sequence of tokens (string) in a single string."""
133
- current_sub_tokens = []
134
- out_string = ""
135
- prev_is_special = False
136
- for token in tokens:
137
- # make sure that special tokens are not decoded using sentencepiece model
138
- if token in self.all_special_tokens:
139
- if not prev_is_special:
140
- out_string += " "
141
- out_string += self.sp_model.decode(current_sub_tokens) + token
142
- prev_is_special = True
143
- current_sub_tokens = []
144
- else:
145
- current_sub_tokens.append(token)
146
- prev_is_special = False
147
- out_string += self.sp_model.decode(current_sub_tokens)
148
- out_string = self.clean_up_tokenization(out_string)
149
- out_string = self._maybe_add_prefix_space(tokens=tokens, decoded=out_string)
150
- return out_string[1:]
151
-
152
- def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
153
- """
154
- Save the vocabulary and special tokens file to a directory.
155
-
156
- Args:
157
- save_directory (`str`):
158
- The directory in which to save the vocabulary.
159
-
160
- Returns:
161
- `Tuple(str)`: Paths to the files saved.
162
- """
163
- if not os.path.isdir(save_directory):
164
- logger.error(f"Vocabulary path ({save_directory}) should be a directory")
165
- return
166
- out_vocab_file = os.path.join(
167
- save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
168
- )
169
-
170
- if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
171
- copyfile(self.vocab_file, out_vocab_file)
172
- elif not os.path.isfile(self.vocab_file):
173
- with open(out_vocab_file, "wb") as fi:
174
- content_spiece_model = self.sp_model.serialized_model_proto()
175
- fi.write(content_spiece_model)
176
-
177
- return (out_vocab_file,)
178
-
179
- def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
180
- if self.add_bos_token:
181
- bos_token_ids = [self.bos_token_id]
182
- else:
183
- bos_token_ids = []
184
-
185
- output = bos_token_ids + token_ids_0
186
-
187
- if token_ids_1 is not None:
188
- output = output + token_ids_1
189
-
190
- if self.add_eos_token:
191
- output = output + [self.eos_token_id]
192
-
193
- return output
194
-
195
- def get_special_tokens_mask(
196
- self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
197
- ) -> List[int]:
198
- """
199
- Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
200
- special tokens using the tokenizer `prepare_for_model` method.
201
-
202
- Args:
203
- token_ids_0 (`List[int]`):
204
- List of IDs.
205
- token_ids_1 (`List[int]`, *optional*):
206
- Optional second list of IDs for sequence pairs.
207
- already_has_special_tokens (`bool`, *optional*, defaults to `False`):
208
- Whether or not the token list is already formatted with special tokens for the model.
209
-
210
- Returns:
211
- `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
212
- """
213
- if already_has_special_tokens:
214
- return super().get_special_tokens_mask(
215
- token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
216
- )
217
-
218
- if token_ids_1 is None:
219
- return [1] + ([0] * len(token_ids_0)) + [1]
220
- return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
221
-
222
- def create_token_type_ids_from_sequences(
223
- self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
224
- ) -> List[int]:
225
- """
226
- Create a mask from the two sequences passed to be used in a sequence-pair classification task. T5 does not make
227
- use of token type ids, therefore a list of zeros is returned.
228
-
229
- Args:
230
- token_ids_0 (`List[int]`):
231
- List of IDs.
232
- token_ids_1 (`List[int]`, *optional*):
233
- Optional second list of IDs for sequence pairs.
234
-
235
- Returns:
236
- `List[int]`: List of zeros.
237
- """
238
- eos = [self.eos_token_id]
239
-
240
- if token_ids_1 is None:
241
- return len(token_ids_0 + eos) * [0]
242
- return len(token_ids_0 + eos + token_ids_1 + eos) * [0]