afrizalha commited on
Commit
f52d7cb
1 Parent(s): f2100e3

Upload 2 files

Browse files
Files changed (2) hide show
  1. configuration_phi3.py +213 -0
  2. modeling_phi3.py +1606 -0
configuration_phi3.py ADDED
@@ -0,0 +1,213 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ """ Phi-3 model configuration"""
17
+
18
+
19
+ from transformers.configuration_utils import PretrainedConfig
20
+ from transformers.utils import logging
21
+
22
+
23
+ logger = logging.get_logger(__name__)
24
+
25
+ PHI3_PRETRAINED_CONFIG_ARCHIVE_MAP = {
26
+ "microsoft/Phi-3-mini-4k-instruct": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/resolve/main/config.json",
27
+ "microsoft/Phi-3-mini-128k-instruct": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/main/config.json",
28
+ }
29
+
30
+
31
+ class Phi3Config(PretrainedConfig):
32
+ r"""
33
+ This is the configuration class to store the configuration of a [`Phi3Model`]. It is used to instantiate a Phi-3
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
36
+ [microsoft/Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct).
37
+
38
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
39
+ documentation from [`PretrainedConfig`] for more information.
40
+
41
+ Args:
42
+ vocab_size (`int`, *optional*, defaults to 32064):
43
+ Vocabulary size of the Phi-3 model. Defines the number of different tokens that can be represented by the
44
+ `inputs_ids` passed when calling [`Phi3Model`].
45
+ hidden_size (`int`, *optional*, defaults to 3072):
46
+ Dimension of the hidden representations.
47
+ intermediate_size (`int`, *optional*, defaults to 8192):
48
+ Dimension of the MLP representations.
49
+ num_hidden_layers (`int`, *optional*, defaults to 32):
50
+ Number of hidden layers in the Transformer decoder.
51
+ num_attention_heads (`int`, *optional*, defaults to 32):
52
+ Number of attention heads for each attention layer in the Transformer decoder.
53
+ num_key_value_heads (`int`, *optional*):
54
+ This is the number of key_value heads that should be used to implement Grouped Query Attention. If
55
+ `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
56
+ `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
57
+ converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
58
+ by meanpooling all the original heads within that group. For more details checkout [this
59
+ paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
60
+ `num_attention_heads`.
61
+ resid_pdrop (`float`, *optional*, defaults to 0.0):
62
+ Dropout probability for mlp outputs.
63
+ embd_pdrop (`int`, *optional*, defaults to 0.0):
64
+ The dropout ratio for the embeddings.
65
+ attention_dropout (`float`, *optional*, defaults to 0.0):
66
+ The dropout ratio after computing the attention scores.
67
+ hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
68
+ The non-linear activation function (function or string) in the decoder.
69
+ max_position_embeddings (`int`, *optional*, defaults to 4096):
70
+ The maximum sequence length that this model might ever be used with.
71
+ original_max_position_embeddings (`int`, *optional*, defaults to 4096):
72
+ The maximum sequence length that this model was trained with. This is used to determine the size of the
73
+ original RoPE embeddings when using long scaling.
74
+ initializer_range (`float`, *optional*, defaults to 0.02):
75
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
76
+ rms_norm_eps (`float`, *optional*, defaults to 1e-05):
77
+ The epsilon value used for the RMSNorm.
78
+ use_cache (`bool`, *optional*, defaults to `True`):
79
+ Whether or not the model should return the last key/values attentions (not used by all models). Only
80
+ relevant if `config.is_decoder=True`. Whether to tie weight embeddings or not.
81
+ tie_word_embeddings (`bool`, *optional*, defaults to `False`):
82
+ Whether to tie weight embeddings
83
+ rope_theta (`float`, *optional*, defaults to 10000.0):
84
+ The base period of the RoPE embeddings.
85
+ rope_scaling (`dict`, *optional*):
86
+ The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must
87
+ contain the following keys: `type`, `short_factor` and `long_factor`. The `type` must be either `su` or `yarn` and
88
+ the `short_factor` and `long_factor` must be lists of numbers with the same length as the hidden size
89
+ divided by the number of attention heads divided by 2.
90
+ bos_token_id (`int`, *optional*, defaults to 1):
91
+ The id of the "beginning-of-sequence" token.
92
+ eos_token_id (`int`, *optional*, defaults to 32000):
93
+ The id of the "end-of-sequence" token.
94
+ pad_token_id (`int`, *optional*, defaults to 32000):
95
+ The id of the padding token.
96
+ sliding_window (`int`, *optional*):
97
+ Sliding window attention window size. If `None`, no sliding window is applied.
98
+
99
+ Example:
100
+
101
+ ```python
102
+ >>> from transformers import Phi3Model, Phi3Config
103
+
104
+ >>> # Initializing a Phi-3 style configuration
105
+ >>> configuration = Phi3Config.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
106
+
107
+ >>> # Initializing a model from the configuration
108
+ >>> model = Phi3Model(configuration)
109
+
110
+ >>> # Accessing the model configuration
111
+ >>> configuration = model.config
112
+ ```"""
113
+
114
+ model_type = "phi3"
115
+ keys_to_ignore_at_inference = ["past_key_values"]
116
+
117
+ def __init__(
118
+ self,
119
+ vocab_size=4096,
120
+ hidden_size=512,
121
+ intermediate_size=1024,
122
+ num_hidden_layers=8,
123
+ num_attention_heads=8,
124
+ num_key_value_heads=None,
125
+ resid_pdrop=0.0,
126
+ embd_pdrop=0.0,
127
+ attention_dropout=0.0,
128
+ hidden_act="silu",
129
+ max_position_embeddings=32,
130
+ original_max_position_embeddings=32,
131
+ initializer_range=0.02,
132
+ rms_norm_eps=1e-5,
133
+ use_cache=True,
134
+ tie_word_embeddings=False,
135
+ rope_theta=3200.0,
136
+ rope_scaling=None,
137
+ bos_token_id=0,
138
+ eos_token_id=1,
139
+ pad_token_id=4,
140
+ sliding_window=None,
141
+ **kwargs,
142
+ ):
143
+ self.vocab_size = vocab_size
144
+ self.hidden_size = hidden_size
145
+ self.intermediate_size = intermediate_size
146
+ self.num_hidden_layers = num_hidden_layers
147
+ self.num_attention_heads = num_attention_heads
148
+
149
+ if num_key_value_heads is None:
150
+ num_key_value_heads = num_attention_heads
151
+
152
+ self.num_key_value_heads = num_key_value_heads
153
+ self.resid_pdrop = resid_pdrop
154
+ self.embd_pdrop = embd_pdrop
155
+ self.attention_dropout = attention_dropout
156
+ self.hidden_act = hidden_act
157
+ self.max_position_embeddings = max_position_embeddings
158
+ self.original_max_position_embeddings = original_max_position_embeddings
159
+ self.initializer_range = initializer_range
160
+ self.rms_norm_eps = rms_norm_eps
161
+ self.use_cache = use_cache
162
+ self.rope_theta = rope_theta
163
+ self.rope_scaling = rope_scaling
164
+ self._rope_scaling_validation()
165
+ self.sliding_window = sliding_window
166
+
167
+ super().__init__(
168
+ bos_token_id=bos_token_id,
169
+ eos_token_id=eos_token_id,
170
+ pad_token_id=pad_token_id,
171
+ tie_word_embeddings=tie_word_embeddings,
172
+ **kwargs,
173
+ )
174
+
175
+ def _rope_scaling_validation(self):
176
+ """
177
+ Validate the `rope_scaling` configuration.
178
+ """
179
+ if self.rope_scaling is None:
180
+ return
181
+
182
+ if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 3:
183
+ raise ValueError(
184
+ "`rope_scaling` must be a dictionary with three fields, `type`, `short_factor` and `long_factor`, "
185
+ f"got {self.rope_scaling}"
186
+ )
187
+ rope_scaling_type = self.rope_scaling.get("type", None)
188
+ rope_scaling_short_factor = self.rope_scaling.get("short_factor", None)
189
+ rope_scaling_long_factor = self.rope_scaling.get("long_factor", None)
190
+ if rope_scaling_type is None or rope_scaling_type not in ["su", "yarn"]:
191
+ raise ValueError(f"`rope_scaling`'s type field must be one of ['su', 'yarn'], got {rope_scaling_type}")
192
+ if not (
193
+ isinstance(rope_scaling_short_factor, list)
194
+ and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor)
195
+ ):
196
+ raise ValueError(
197
+ f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}"
198
+ )
199
+ if not len(rope_scaling_short_factor) == self.hidden_size // self.num_attention_heads // 2:
200
+ raise ValueError(
201
+ f"`rope_scaling`'s short_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_short_factor)}"
202
+ )
203
+ if not (
204
+ isinstance(rope_scaling_long_factor, list)
205
+ and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor)
206
+ ):
207
+ raise ValueError(
208
+ f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}"
209
+ )
210
+ if not len(rope_scaling_long_factor) == self.hidden_size // self.num_attention_heads // 2:
211
+ raise ValueError(
212
+ f"`rope_scaling`'s long_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_long_factor)}"
213
+ )
modeling_phi3.py ADDED
@@ -0,0 +1,1606 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ """ PyTorch Phi-3 model."""
17
+
18
+ import inspect
19
+ import math
20
+ import warnings
21
+ from typing import List, Optional, Tuple, Union
22
+
23
+ import torch
24
+ import torch.nn.functional as F
25
+ import torch.utils.checkpoint
26
+ from torch import nn
27
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
28
+
29
+ from transformers.activations import ACT2FN
30
+ from transformers.cache_utils import Cache, DynamicCache
31
+ from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
32
+ from transformers.modeling_outputs import (
33
+ BaseModelOutputWithPast,
34
+ CausalLMOutputWithPast,
35
+ SequenceClassifierOutputWithPast,
36
+ TokenClassifierOutput,
37
+ )
38
+ from transformers.modeling_utils import PreTrainedModel
39
+ from transformers.utils import (
40
+ add_code_sample_docstrings,
41
+ add_start_docstrings,
42
+ add_start_docstrings_to_model_forward,
43
+ is_flash_attn_2_available,
44
+ is_flash_attn_greater_or_equal_2_10,
45
+ logging,
46
+ replace_return_docstrings,
47
+ )
48
+ from .configuration_phi3 import Phi3Config
49
+
50
+
51
+ logger = logging.get_logger(__name__)
52
+
53
+ # Transformers scans dependencies in the modeling file, causing issues on conditional loading. The regex only ignores try/catch blocks, but not if statements
54
+ # if is_flash_attn_2_available():
55
+ _flash_supports_window_size = False
56
+ try:
57
+ from flash_attn import flash_attn_func, flash_attn_varlen_func
58
+ from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
59
+
60
+ _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
61
+ except ImportError as error:
62
+ logger.warning(
63
+ f"`flash-attention` package not found, consider installing for better performance: {error}."
64
+ )
65
+ if not _flash_supports_window_size:
66
+ logger.warning(
67
+ "Current `flash-attention` does not support `window_size`. Either upgrade or use `attn_implementation='eager'`."
68
+ )
69
+
70
+ _CHECKPOINT_FOR_DOC = "microsoft/Phi-3-mini-4k-instruct"
71
+ _CONFIG_FOR_DOC = "Phi3Config"
72
+
73
+ PHI3_PRETRAINED_MODEL_ARCHIVE_LIST = [
74
+ "microsoft/Phi-3-mini-4k-instruct",
75
+ "microsoft/Phi-3-mini-128k-instruct",
76
+ # See all Phi-3 models at https://huggingface.co/models?filter=Phi-3
77
+ ]
78
+
79
+
80
+ # Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Phi3
81
+ class Phi3RMSNorm(nn.Module):
82
+ def __init__(self, hidden_size, eps=1e-6):
83
+ """
84
+ Phi3RMSNorm is equivalent to T5LayerNorm
85
+ """
86
+ super().__init__()
87
+ self.weight = nn.Parameter(torch.ones(hidden_size))
88
+ self.variance_epsilon = eps
89
+
90
+ def forward(self, hidden_states):
91
+ input_dtype = hidden_states.dtype
92
+ hidden_states = hidden_states.to(torch.float32)
93
+ variance = hidden_states.pow(2).mean(-1, keepdim=True)
94
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
95
+ return self.weight * hidden_states.to(input_dtype)
96
+
97
+
98
+ # Copied from transformers.models.llama.modeling_llama._get_unpad_data
99
+ def _get_unpad_data(attention_mask):
100
+ seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
101
+ indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
102
+ max_seqlen_in_batch = seqlens_in_batch.max().item()
103
+ cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
104
+ return (
105
+ indices,
106
+ cu_seqlens,
107
+ max_seqlen_in_batch,
108
+ )
109
+
110
+
111
+ # Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with gemma->phi3, Gemma->Phi3
112
+ class Phi3RotaryEmbedding(nn.Module):
113
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
114
+ super().__init__()
115
+
116
+ self.dim = dim
117
+ self.max_position_embeddings = max_position_embeddings
118
+ self.base = base
119
+ self.register_buffer("inv_freq", None, persistent=False)
120
+
121
+ @torch.no_grad()
122
+ def forward(self, x, position_ids, seq_len=None):
123
+ # x: [bs, num_attention_heads, seq_len, head_size]
124
+ if self.inv_freq is None:
125
+ self.inv_freq = 1.0 / (
126
+ self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
127
+ )
128
+ inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
129
+ position_ids_expanded = position_ids[:, None, :].float()
130
+ # Force float32 since bfloat16 loses precision on long contexts
131
+ # See https://github.com/huggingface/transformers/pull/29285
132
+ device_type = x.device.type
133
+ device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
134
+ with torch.autocast(device_type=device_type, enabled=False):
135
+ freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
136
+ emb = torch.cat((freqs, freqs), dim=-1)
137
+ cos = emb.cos()
138
+ sin = emb.sin()
139
+ return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
140
+
141
+
142
+ class Phi3SuScaledRotaryEmbedding(Phi3RotaryEmbedding):
143
+ def __init__(self, dim, config, device=None):
144
+ super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
145
+
146
+ self.short_factor = config.rope_scaling["short_factor"]
147
+ self.long_factor = config.rope_scaling["long_factor"]
148
+ self.original_max_position_embeddings = config.original_max_position_embeddings
149
+
150
+ @torch.no_grad()
151
+ def forward(self, x, position_ids, seq_len=None):
152
+ seq_len = torch.max(position_ids) + 1
153
+ if seq_len > self.original_max_position_embeddings:
154
+ ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
155
+ else:
156
+ ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
157
+
158
+ inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
159
+ self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
160
+
161
+ inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
162
+ position_ids_expanded = position_ids[:, None, :].float()
163
+
164
+ # Force float32 since bfloat16 loses precision on long contexts
165
+ # See https://github.com/huggingface/transformers/pull/29285
166
+ device_type = x.device.type
167
+ device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
168
+ with torch.autocast(device_type=device_type, enabled=False):
169
+ freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
170
+ emb = torch.cat((freqs, freqs), dim=-1)
171
+
172
+ scale = self.max_position_embeddings / self.original_max_position_embeddings
173
+ if scale <= 1.0:
174
+ scaling_factor = 1.0
175
+ else:
176
+ scaling_factor = math.sqrt(1 + math.log(scale) / math.log(self.original_max_position_embeddings))
177
+
178
+ cos = emb.cos() * scaling_factor
179
+ sin = emb.sin() * scaling_factor
180
+ return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
181
+
182
+
183
+ class Phi3YarnScaledRotaryEmbedding(Phi3RotaryEmbedding):
184
+ def __init__(self, dim, config, device=None):
185
+ super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
186
+
187
+ self.short_factor = config.rope_scaling["short_factor"]
188
+ self.long_factor = config.rope_scaling["long_factor"]
189
+ self.original_max_position_embeddings = config.original_max_position_embeddings
190
+
191
+ @torch.no_grad()
192
+ def forward(self, x, position_ids, seq_len=None):
193
+ seq_len = torch.max(position_ids) + 1
194
+ if seq_len > self.original_max_position_embeddings:
195
+ ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
196
+ else:
197
+ ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
198
+
199
+ inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
200
+ self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
201
+
202
+ inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
203
+ position_ids_expanded = position_ids[:, None, :].float()
204
+
205
+ # Force float32 since bfloat16 loses precision on long contexts
206
+ # See https://github.com/huggingface/transformers/pull/29285
207
+ device_type = x.device.type
208
+ device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
209
+ with torch.autocast(device_type=device_type, enabled=False):
210
+ freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
211
+ emb = torch.cat((freqs, freqs), dim=-1)
212
+
213
+ scale = self.max_position_embeddings / self.original_max_position_embeddings
214
+ if scale <= 1.0:
215
+ scaling_factor = 1.0
216
+ else:
217
+ scaling_factor = 0.1 * math.log(scale) + 1.0
218
+
219
+ cos = emb.cos() * scaling_factor
220
+ sin = emb.sin() * scaling_factor
221
+ return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
222
+
223
+
224
+ # Copied from transformers.models.llama.modeling_llama.rotate_half
225
+ def rotate_half(x):
226
+ """Rotates half the hidden dims of the input."""
227
+ x1 = x[..., : x.shape[-1] // 2]
228
+ x2 = x[..., x.shape[-1] // 2 :]
229
+ return torch.cat((-x2, x1), dim=-1)
230
+
231
+
232
+ # Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
233
+ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
234
+ """Applies Rotary Position Embedding to the query and key tensors.
235
+
236
+ Args:
237
+ q (`torch.Tensor`): The query tensor.
238
+ k (`torch.Tensor`): The key tensor.
239
+ cos (`torch.Tensor`): The cosine part of the rotary embedding.
240
+ sin (`torch.Tensor`): The sine part of the rotary embedding.
241
+ position_ids (`torch.Tensor`, *optional*):
242
+ Deprecated and unused.
243
+ unsqueeze_dim (`int`, *optional*, defaults to 1):
244
+ The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
245
+ sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
246
+ that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
247
+ k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
248
+ cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
249
+ the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
250
+ Returns:
251
+ `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
252
+ """
253
+ cos = cos.unsqueeze(unsqueeze_dim)
254
+ sin = sin.unsqueeze(unsqueeze_dim)
255
+ q_embed = (q * cos) + (rotate_half(q) * sin)
256
+ k_embed = (k * cos) + (rotate_half(k) * sin)
257
+ return q_embed, k_embed
258
+
259
+
260
+ class Phi3MLP(nn.Module):
261
+ def __init__(self, config):
262
+ super().__init__()
263
+
264
+ self.config = config
265
+ self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
266
+ self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
267
+
268
+ self.activation_fn = ACT2FN[config.hidden_act]
269
+
270
+ def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
271
+ up_states = self.gate_up_proj(hidden_states)
272
+
273
+ gate, up_states = up_states.chunk(2, dim=-1)
274
+ up_states = up_states * self.activation_fn(gate)
275
+
276
+ return self.down_proj(up_states)
277
+
278
+
279
+ # Copied from transformers.models.llama.modeling_llama.repeat_kv with llama->phi
280
+ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
281
+ """
282
+ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
283
+ num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
284
+ """
285
+ batch, num_key_value_heads, slen, head_dim = hidden_states.shape
286
+ if n_rep == 1:
287
+ return hidden_states
288
+ hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
289
+ return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
290
+
291
+
292
+ class Phi3Attention(nn.Module):
293
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
294
+
295
+ def __init__(self, config: Phi3Config, layer_idx: Optional[int] = None):
296
+ super().__init__()
297
+ self.config = config
298
+ self.layer_idx = layer_idx
299
+ if layer_idx is None:
300
+ logger.warning_once(
301
+ f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
302
+ "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
303
+ "when creating this class."
304
+ )
305
+
306
+ self.attention_dropout = config.attention_dropout
307
+ self.hidden_size = config.hidden_size
308
+ self.num_heads = config.num_attention_heads
309
+ self.head_dim = self.hidden_size // self.num_heads
310
+ self.num_key_value_heads = config.num_key_value_heads
311
+ self.num_key_value_groups = self.num_heads // self.num_key_value_heads
312
+ self.max_position_embeddings = config.max_position_embeddings
313
+ self.original_max_position_embeddings = config.original_max_position_embeddings
314
+ self.rope_theta = config.rope_theta
315
+ self.rope_scaling = config.rope_scaling
316
+ self.is_causal = True
317
+
318
+ if (self.head_dim * self.num_heads) != self.hidden_size:
319
+ raise ValueError(
320
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
321
+ f" and `num_heads`: {self.num_heads})."
322
+ )
323
+
324
+ op_size = self.num_heads * self.head_dim + 2 * (self.num_key_value_heads * self.head_dim)
325
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
326
+ self.qkv_proj = nn.Linear(self.hidden_size, op_size, bias=False)
327
+ self._init_rope()
328
+
329
+ def _init_rope(self):
330
+ if self.rope_scaling is None:
331
+ self.rotary_emb = Phi3RotaryEmbedding(
332
+ self.head_dim,
333
+ max_position_embeddings=self.max_position_embeddings,
334
+ base=self.rope_theta,
335
+ )
336
+ else:
337
+ scaling_type = self.config.rope_scaling["type"]
338
+ if scaling_type == "su":
339
+ self.rotary_emb = Phi3SuScaledRotaryEmbedding(self.head_dim, self.config)
340
+ elif scaling_type == "yarn":
341
+ self.rotary_emb = Phi3YarnScaledRotaryEmbedding(self.head_dim, self.config)
342
+ else:
343
+ raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
344
+
345
+ def forward(
346
+ self,
347
+ hidden_states: torch.Tensor,
348
+ attention_mask: Optional[torch.Tensor] = None,
349
+ position_ids: Optional[torch.LongTensor] = None,
350
+ past_key_value: Optional[Cache] = None,
351
+ output_attentions: bool = False,
352
+ use_cache: bool = False,
353
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
354
+ logger.warning_once("You are not running the flash-attention implementation, expect numerical differences.")
355
+
356
+ bsz, q_len, _ = hidden_states.size()
357
+
358
+ qkv = self.qkv_proj(hidden_states)
359
+ query_pos = self.num_heads * self.head_dim
360
+ query_states = qkv[..., :query_pos]
361
+ key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
362
+ value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
363
+
364
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
365
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
366
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
367
+
368
+ kv_seq_len = key_states.shape[-2]
369
+ if past_key_value is not None:
370
+ if self.layer_idx is None:
371
+ raise ValueError(
372
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
373
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
374
+ "with a layer index."
375
+ )
376
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
377
+ cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
378
+
379
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
380
+
381
+ if past_key_value is not None:
382
+ cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
383
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
384
+
385
+ # repeat k/v heads if n_kv_heads < n_heads
386
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
387
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
388
+
389
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
390
+
391
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
392
+ raise ValueError(
393
+ f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
394
+ f" {attn_weights.size()}"
395
+ )
396
+
397
+ if attention_mask is not None:
398
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
399
+ raise ValueError(
400
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
401
+ )
402
+ attn_weights = attn_weights + attention_mask
403
+
404
+ # upcast attention to fp32
405
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype)
406
+ attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
407
+
408
+ attn_output = torch.matmul(attn_weights, value_states)
409
+
410
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
411
+ raise ValueError(
412
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
413
+ f" {attn_output.size()}"
414
+ )
415
+
416
+ attn_output = attn_output.transpose(1, 2).contiguous()
417
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
418
+
419
+ attn_output = self.o_proj(attn_output)
420
+
421
+ if not output_attentions:
422
+ attn_weights = None
423
+
424
+ return attn_output, attn_weights, past_key_value
425
+
426
+
427
+ class Phi3FlashAttention2(Phi3Attention):
428
+ """
429
+ Phi-3 flash attention module. This module inherits from `Phi3Attention` as the weights of the module stays
430
+ untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
431
+ flash attention and deal with padding tokens in case the input contains any of them.
432
+ """
433
+
434
+ # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
435
+ def __init__(self, *args, **kwargs):
436
+ super().__init__(*args, **kwargs)
437
+
438
+ # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
439
+ # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
440
+ # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
441
+ self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
442
+
443
+ def forward(
444
+ self,
445
+ hidden_states: torch.Tensor,
446
+ attention_mask: Optional[torch.LongTensor] = None,
447
+ position_ids: Optional[torch.LongTensor] = None,
448
+ past_key_value: Optional[Cache] = None,
449
+ output_attentions: bool = False,
450
+ use_cache: bool = False,
451
+ **kwargs,
452
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
453
+ # Phi3FlashAttention2 attention does not support output_attentions
454
+
455
+ if not _flash_supports_window_size:
456
+ logger.warning_once(
457
+ "The current flash attention version does not support sliding window attention. Please use `attn_implementation='eager'` or upgrade flash-attn library."
458
+ )
459
+ raise ValueError("The current flash attention version does not support sliding window attention.")
460
+
461
+ output_attentions = False
462
+
463
+ if "padding_mask" in kwargs:
464
+ warnings.warn(
465
+ "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
466
+ )
467
+
468
+ # overwrite attention_mask with padding_mask
469
+ attention_mask = kwargs.pop("padding_mask")
470
+
471
+ bsz, q_len, _ = hidden_states.size()
472
+
473
+ qkv = self.qkv_proj(hidden_states)
474
+ query_pos = self.num_heads * self.head_dim
475
+ query_states = qkv[..., :query_pos]
476
+ key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
477
+ value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
478
+
479
+ # Flash attention requires the input to have the shape
480
+ # batch_size x seq_length x head_dim x hidden_dim
481
+ # therefore we just need to keep the original shape
482
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
483
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
484
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
485
+
486
+ kv_seq_len = key_states.shape[-2]
487
+ if past_key_value is not None:
488
+ if self.layer_idx is None:
489
+ raise ValueError(
490
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
491
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
492
+ "with a layer index."
493
+ )
494
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
495
+
496
+ # Because the input can be padded, the absolute sequence length depends on the max position id.
497
+ rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
498
+ cos, sin = self.rotary_emb(value_states, position_ids, seq_len=rotary_seq_len)
499
+
500
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
501
+
502
+ use_sliding_windows = (
503
+ _flash_supports_window_size
504
+ and getattr(self.config, "sliding_window", None) is not None
505
+ and kv_seq_len > self.config.sliding_window
506
+ )
507
+
508
+ if past_key_value is not None:
509
+ # Activate slicing cache only if the config has a value `sliding_windows` attribute
510
+ cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
511
+ if (
512
+ getattr(self.config, "sliding_window", None) is not None
513
+ and kv_seq_len > self.config.sliding_window
514
+ and cache_has_contents
515
+ ):
516
+ slicing_tokens = 1 - self.config.sliding_window
517
+
518
+ past_key = past_key_value[self.layer_idx][0]
519
+ past_value = past_key_value[self.layer_idx][1]
520
+
521
+ past_key = past_key[:, :, slicing_tokens:, :].contiguous()
522
+ past_value = past_value[:, :, slicing_tokens:, :].contiguous()
523
+
524
+ if past_key.shape[-2] != self.config.sliding_window - 1:
525
+ raise ValueError(
526
+ f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
527
+ f" {past_key.shape}"
528
+ )
529
+
530
+ if attention_mask is not None:
531
+ attention_mask = attention_mask[:, slicing_tokens:]
532
+ attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
533
+
534
+ cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
535
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
536
+
537
+ # repeat k/v heads if n_kv_heads < n_heads
538
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
539
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
540
+
541
+ attn_dropout = self.attention_dropout if self.training else 0.0
542
+
543
+ # In PEFT, usually we cast the layer norms in float32 for training stability reasons
544
+ # therefore the input hidden states gets silently casted in float32. Hence, we need
545
+ # cast them back in the correct dtype just to be sure everything works as expected.
546
+ # This might slowdown training & inference so it is recommended to not cast the LayerNorms
547
+ # in fp32.
548
+
549
+ if query_states.dtype == torch.float32:
550
+ if torch.is_autocast_enabled():
551
+ target_dtype = torch.get_autocast_gpu_dtype()
552
+ # Handle the case where the model is quantized
553
+ elif hasattr(self.config, "_pre_quantization_dtype"):
554
+ target_dtype = self.config._pre_quantization_dtype
555
+ else:
556
+ target_dtype = self.qkv_proj.weight.dtype
557
+
558
+ logger.warning_once(
559
+ f"The input hidden states seems to be silently casted in float32, this might be related to"
560
+ f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
561
+ f" {target_dtype}."
562
+ )
563
+
564
+ query_states = query_states.to(target_dtype)
565
+ key_states = key_states.to(target_dtype)
566
+ value_states = value_states.to(target_dtype)
567
+
568
+ # Reashape to the expected shape for Flash Attention
569
+ query_states = query_states.transpose(1, 2)
570
+ key_states = key_states.transpose(1, 2)
571
+ value_states = value_states.transpose(1, 2)
572
+
573
+ attn_output = self._flash_attention_forward(
574
+ query_states,
575
+ key_states,
576
+ value_states,
577
+ attention_mask,
578
+ q_len,
579
+ dropout=attn_dropout,
580
+ use_sliding_windows=use_sliding_windows,
581
+ )
582
+
583
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
584
+ attn_output = self.o_proj(attn_output)
585
+
586
+ if not output_attentions:
587
+ attn_weights = None
588
+
589
+ return attn_output, attn_weights, past_key_value
590
+
591
+ # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._flash_attention_forward
592
+ def _flash_attention_forward(
593
+ self,
594
+ query_states,
595
+ key_states,
596
+ value_states,
597
+ attention_mask,
598
+ query_length,
599
+ dropout=0.0,
600
+ softmax_scale=None,
601
+ use_sliding_windows=False,
602
+ ):
603
+ """
604
+ Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
605
+ first unpad the input, then computes the attention scores and pad the final attention scores.
606
+
607
+ Args:
608
+ query_states (`torch.Tensor`):
609
+ Input query states to be passed to Flash Attention API
610
+ key_states (`torch.Tensor`):
611
+ Input key states to be passed to Flash Attention API
612
+ value_states (`torch.Tensor`):
613
+ Input value states to be passed to Flash Attention API
614
+ attention_mask (`torch.Tensor`):
615
+ The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
616
+ position of padding tokens and 1 for the position of non-padding tokens.
617
+ dropout (`float`):
618
+ Attention dropout
619
+ softmax_scale (`float`, *optional*):
620
+ The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
621
+ use_sliding_windows (`bool`, *optional*):
622
+ Whether to activate sliding window attention.
623
+ """
624
+ if not self._flash_attn_uses_top_left_mask:
625
+ causal = self.is_causal
626
+ else:
627
+ # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
628
+ causal = self.is_causal and query_length != 1
629
+
630
+ # Contains at least one padding token in the sequence
631
+ if attention_mask is not None:
632
+ batch_size = query_states.shape[0]
633
+ query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
634
+ query_states, key_states, value_states, attention_mask, query_length
635
+ )
636
+
637
+ cu_seqlens_q, cu_seqlens_k = cu_seq_lens
638
+ max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
639
+
640
+ if not use_sliding_windows:
641
+ attn_output_unpad = flash_attn_varlen_func(
642
+ query_states,
643
+ key_states,
644
+ value_states,
645
+ cu_seqlens_q=cu_seqlens_q,
646
+ cu_seqlens_k=cu_seqlens_k,
647
+ max_seqlen_q=max_seqlen_in_batch_q,
648
+ max_seqlen_k=max_seqlen_in_batch_k,
649
+ dropout_p=dropout,
650
+ softmax_scale=softmax_scale,
651
+ causal=causal,
652
+ )
653
+ else:
654
+ attn_output_unpad = flash_attn_varlen_func(
655
+ query_states,
656
+ key_states,
657
+ value_states,
658
+ cu_seqlens_q=cu_seqlens_q,
659
+ cu_seqlens_k=cu_seqlens_k,
660
+ max_seqlen_q=max_seqlen_in_batch_q,
661
+ max_seqlen_k=max_seqlen_in_batch_k,
662
+ dropout_p=dropout,
663
+ softmax_scale=softmax_scale,
664
+ causal=causal,
665
+ window_size=(self.config.sliding_window, self.config.sliding_window),
666
+ )
667
+
668
+ attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
669
+ else:
670
+ if not use_sliding_windows:
671
+ attn_output = flash_attn_func(
672
+ query_states,
673
+ key_states,
674
+ value_states,
675
+ dropout,
676
+ softmax_scale=softmax_scale,
677
+ causal=causal,
678
+ )
679
+ else:
680
+ attn_output = flash_attn_func(
681
+ query_states,
682
+ key_states,
683
+ value_states,
684
+ dropout,
685
+ softmax_scale=softmax_scale,
686
+ causal=causal,
687
+ window_size=(self.config.sliding_window, self.config.sliding_window),
688
+ )
689
+
690
+ return attn_output
691
+
692
+ # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input
693
+ def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
694
+ batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
695
+
696
+ # On the first iteration we need to properly re-create the padding mask
697
+ # by slicing it on the proper place
698
+ if kv_seq_len != attention_mask.shape[-1]:
699
+ attention_mask_num_tokens = attention_mask.shape[-1]
700
+ attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
701
+
702
+ indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
703
+
704
+ key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
705
+ value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
706
+
707
+ if query_length == kv_seq_len:
708
+ query_layer = index_first_axis(
709
+ query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
710
+ )
711
+ cu_seqlens_q = cu_seqlens_k
712
+ max_seqlen_in_batch_q = max_seqlen_in_batch_k
713
+ indices_q = indices_k
714
+ elif query_length == 1:
715
+ max_seqlen_in_batch_q = 1
716
+ cu_seqlens_q = torch.arange(
717
+ batch_size + 1, dtype=torch.int32, device=query_layer.device
718
+ ) # There is a memcpy here, that is very bad.
719
+ indices_q = cu_seqlens_q[:-1]
720
+ query_layer = query_layer.squeeze(1)
721
+ else:
722
+ # The -q_len: slice assumes left padding.
723
+ attention_mask = attention_mask[:, -query_length:]
724
+ query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
725
+
726
+ return (
727
+ query_layer,
728
+ key_layer,
729
+ value_layer,
730
+ indices_q,
731
+ (cu_seqlens_q, cu_seqlens_k),
732
+ (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
733
+ )
734
+
735
+
736
+ # copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Phi3
737
+ # TODO @Arthur no longer copied from LLama after static cache
738
+ class Phi3SdpaAttention(Phi3Attention):
739
+ """
740
+ Phi3 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
741
+ `Phi3Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
742
+ SDPA API.
743
+ """
744
+
745
+ # Adapted from Phi3Attention.forward
746
+ def forward(
747
+ self,
748
+ hidden_states: torch.Tensor,
749
+ attention_mask: Optional[torch.Tensor] = None,
750
+ position_ids: Optional[torch.LongTensor] = None,
751
+ past_key_value: Optional[Cache] = None,
752
+ output_attentions: bool = False,
753
+ use_cache: bool = False,
754
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
755
+ if output_attentions:
756
+ # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
757
+ logger.warning_once(
758
+ "Phi3Model is using Phi3SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
759
+ 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
760
+ )
761
+ return super().forward(
762
+ hidden_states=hidden_states,
763
+ attention_mask=attention_mask,
764
+ position_ids=position_ids,
765
+ past_key_value=past_key_value,
766
+ output_attentions=output_attentions,
767
+ use_cache=use_cache,
768
+ )
769
+
770
+ bsz, q_len, _ = hidden_states.size()
771
+
772
+ qkv = self.qkv_proj(hidden_states)
773
+ query_pos = self.num_heads * self.head_dim
774
+ query_states = qkv[..., :query_pos]
775
+ key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
776
+ value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
777
+
778
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
779
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
780
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
781
+
782
+ kv_seq_len = key_states.shape[-2]
783
+ if past_key_value is not None:
784
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
785
+ cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
786
+
787
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
788
+
789
+ if past_key_value is not None:
790
+ cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
791
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
792
+
793
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
794
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
795
+
796
+ if attention_mask is not None:
797
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
798
+ raise ValueError(
799
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
800
+ )
801
+
802
+ # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
803
+ # Reference: https://github.com/pytorch/pytorch/issues/112577.
804
+ if query_states.device.type == "cuda" and attention_mask is not None:
805
+ query_states = query_states.contiguous()
806
+ key_states = key_states.contiguous()
807
+ value_states = value_states.contiguous()
808
+
809
+ attn_output = torch.nn.functional.scaled_dot_product_attention(
810
+ query_states,
811
+ key_states,
812
+ value_states,
813
+ attn_mask=attention_mask,
814
+ dropout_p=self.attention_dropout if self.training else 0.0,
815
+ # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
816
+ is_causal=self.is_causal and attention_mask is None and q_len > 1,
817
+ )
818
+
819
+ attn_output = attn_output.transpose(1, 2).contiguous()
820
+ attn_output = attn_output.view(bsz, q_len, self.hidden_size)
821
+
822
+ attn_output = self.o_proj(attn_output)
823
+
824
+ return attn_output, None, past_key_value
825
+
826
+
827
+ PHI3_ATTENTION_CLASSES = {
828
+ "eager": Phi3Attention,
829
+ "flash_attention_2": Phi3FlashAttention2,
830
+ "sdpa": Phi3SdpaAttention,
831
+ }
832
+
833
+
834
+ class Phi3DecoderLayer(nn.Module):
835
+ def __init__(self, config: Phi3Config, layer_idx: int):
836
+ super().__init__()
837
+
838
+ self.config = config
839
+ self.self_attn = PHI3_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
840
+
841
+ self.mlp = Phi3MLP(config)
842
+ self.input_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
843
+
844
+ self.resid_attn_dropout = nn.Dropout(config.resid_pdrop)
845
+ self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop)
846
+ self.post_attention_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
847
+
848
+ def forward(
849
+ self,
850
+ hidden_states: torch.Tensor,
851
+ attention_mask: Optional[torch.Tensor] = None,
852
+ position_ids: Optional[torch.LongTensor] = None,
853
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
854
+ output_attentions: Optional[bool] = False,
855
+ use_cache: Optional[bool] = False,
856
+ **kwargs,
857
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
858
+ if "padding_mask" in kwargs:
859
+ warnings.warn(
860
+ "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
861
+ )
862
+ """
863
+ Args:
864
+ hidden_states (`torch.FloatTensor`):
865
+ input to the layer of shape `(batch, seq_len, embed_dim)`
866
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
867
+ `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
868
+ position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
869
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range
870
+ `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
871
+ output_attentions (`bool`, *optional*):
872
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
873
+ returned tensors for more detail.
874
+ use_cache (`bool`, *optional*):
875
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
876
+ (see `past_key_values`).
877
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
878
+ """
879
+
880
+ residual = hidden_states
881
+
882
+ hidden_states = self.input_layernorm(hidden_states)
883
+
884
+ # Self Attention
885
+ attn_outputs, self_attn_weights, present_key_value = self.self_attn(
886
+ hidden_states=hidden_states,
887
+ attention_mask=attention_mask,
888
+ position_ids=position_ids,
889
+ past_key_value=past_key_value,
890
+ output_attentions=output_attentions,
891
+ use_cache=use_cache,
892
+ )
893
+
894
+ hidden_states = residual + self.resid_attn_dropout(attn_outputs)
895
+
896
+ residual = hidden_states
897
+ hidden_states = self.post_attention_layernorm(hidden_states)
898
+ hidden_states = self.mlp(hidden_states)
899
+ hidden_states = residual + self.resid_mlp_dropout(hidden_states)
900
+
901
+ outputs = (hidden_states,)
902
+
903
+ if output_attentions:
904
+ outputs += (self_attn_weights,)
905
+
906
+ if use_cache:
907
+ outputs += (present_key_value,)
908
+
909
+ return outputs
910
+
911
+
912
+ PHI3_START_DOCSTRING = r"""
913
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
914
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
915
+ etc.)
916
+
917
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
918
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
919
+ and behavior.
920
+
921
+ Parameters:
922
+ config ([`Phi3Config`]):
923
+ Model configuration class with all the parameters of the model. Initializing with a config file does not
924
+ load the weights associated with the model, only the configuration. Check out the
925
+ [`~PreTrainedModel.from_pretrained`] method to load the model weights.
926
+ """
927
+
928
+
929
+ @add_start_docstrings(
930
+ "The bare Phi-3 model outputting raw hidden-states without any specific head on top.",
931
+ PHI3_START_DOCSTRING,
932
+ )
933
+ class Phi3PreTrainedModel(PreTrainedModel):
934
+ config_class = Phi3Config
935
+ base_model_prefix = "model"
936
+ supports_gradient_checkpointing = True
937
+ _no_split_modules = ["Phi3DecoderLayer"]
938
+ _skip_keys_device_placement = "past_key_values"
939
+ _supports_flash_attn_2 = True
940
+ _supports_sdpa = False
941
+ _supports_cache_class = True
942
+
943
+ _version = "0.0.5"
944
+
945
+ def _init_weights(self, module):
946
+ std = self.config.initializer_range
947
+ if isinstance(module, nn.Linear):
948
+ module.weight.data.normal_(mean=0.0, std=std)
949
+ if module.bias is not None:
950
+ module.bias.data.zero_()
951
+ elif isinstance(module, nn.Embedding):
952
+ module.weight.data.normal_(mean=0.0, std=std)
953
+ if module.padding_idx is not None:
954
+ module.weight.data[module.padding_idx].zero_()
955
+
956
+
957
+ PHI3_INPUTS_DOCSTRING = r"""
958
+ Args:
959
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
960
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
961
+ it.
962
+
963
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
964
+ [`PreTrainedTokenizer.__call__`] for details.
965
+
966
+ [What are input IDs?](../glossary#input-ids)
967
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
968
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
969
+
970
+ - 1 for tokens that are **not masked**,
971
+ - 0 for tokens that are **masked**.
972
+
973
+ [What are attention masks?](../glossary#attention-mask)
974
+
975
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
976
+ [`PreTrainedTokenizer.__call__`] for details.
977
+
978
+ If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
979
+ `past_key_values`).
980
+
981
+ If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
982
+ and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
983
+ information on the default strategy.
984
+
985
+ - 1 indicates the head is **not masked**,
986
+ - 0 indicates the head is **masked**.
987
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
988
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
989
+ config.n_positions - 1]`.
990
+
991
+ [What are position IDs?](../glossary#position-ids)
992
+ past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
993
+ Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
994
+ blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
995
+ returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
996
+
997
+ Two formats are allowed:
998
+ - a [`~cache_utils.Cache`] instance;
999
+ - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
1000
+ shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
1001
+ cache format.
1002
+
1003
+ The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
1004
+ legacy cache format will be returned.
1005
+
1006
+ If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
1007
+ have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
1008
+ of shape `(batch_size, sequence_length)`.
1009
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
1010
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
1011
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
1012
+ model's internal embedding lookup matrix.
1013
+ use_cache (`bool`, *optional*):
1014
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
1015
+ `past_key_values`).
1016
+ output_attentions (`bool`, *optional*):
1017
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
1018
+ tensors for more detail.
1019
+ output_hidden_states (`bool`, *optional*):
1020
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
1021
+ more detail.
1022
+ return_dict (`bool`, *optional*):
1023
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
1024
+ """
1025
+
1026
+
1027
+ @add_start_docstrings(
1028
+ "The bare Phi-3 model outputting raw hidden-states without any specific head on top.",
1029
+ PHI3_START_DOCSTRING,
1030
+ )
1031
+ class Phi3Model(Phi3PreTrainedModel):
1032
+ """
1033
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Phi3DecoderLayer`]
1034
+
1035
+ Args:
1036
+ config: Phi3Config
1037
+ """
1038
+
1039
+ def __init__(self, config: Phi3Config):
1040
+ super().__init__(config)
1041
+ self.padding_idx = config.pad_token_id
1042
+ self.vocab_size = config.vocab_size
1043
+
1044
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
1045
+ self.embed_dropout = nn.Dropout(config.embd_pdrop)
1046
+ self.layers = nn.ModuleList(
1047
+ [Phi3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
1048
+ )
1049
+ self._attn_implementation = config._attn_implementation
1050
+ self.norm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
1051
+
1052
+ self.gradient_checkpointing = False
1053
+ # Initialize weights and apply final processing
1054
+ self.post_init()
1055
+
1056
+ def get_input_embeddings(self):
1057
+ return self.embed_tokens
1058
+
1059
+ def set_input_embeddings(self, value):
1060
+ self.embed_tokens = value
1061
+
1062
+ @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
1063
+ def forward(
1064
+ self,
1065
+ input_ids: torch.LongTensor = None,
1066
+ attention_mask: Optional[torch.Tensor] = None,
1067
+ position_ids: Optional[torch.LongTensor] = None,
1068
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1069
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1070
+ use_cache: Optional[bool] = None,
1071
+ output_attentions: Optional[bool] = None,
1072
+ output_hidden_states: Optional[bool] = None,
1073
+ return_dict: Optional[bool] = None,
1074
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
1075
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
1076
+ output_hidden_states = (
1077
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
1078
+ )
1079
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
1080
+
1081
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1082
+
1083
+ # retrieve input_ids and inputs_embeds
1084
+ if input_ids is not None and inputs_embeds is not None:
1085
+ raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
1086
+ elif input_ids is not None:
1087
+ batch_size, seq_length = input_ids.shape[:2]
1088
+ elif inputs_embeds is not None:
1089
+ batch_size, seq_length = inputs_embeds.shape[:2]
1090
+ else:
1091
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
1092
+
1093
+ past_key_values_length = 0
1094
+
1095
+ if self.gradient_checkpointing and self.training:
1096
+ if use_cache:
1097
+ logger.warning_once(
1098
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
1099
+ )
1100
+ use_cache = False
1101
+
1102
+ if use_cache:
1103
+ use_legacy_cache = not isinstance(past_key_values, Cache)
1104
+ if use_legacy_cache:
1105
+ past_key_values = DynamicCache.from_legacy_cache(past_key_values)
1106
+ past_key_values_length = past_key_values.get_usable_length(seq_length)
1107
+
1108
+ if position_ids is None:
1109
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
1110
+ position_ids = torch.arange(
1111
+ past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
1112
+ )
1113
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
1114
+ else:
1115
+ position_ids = position_ids.view(-1, seq_length).long()
1116
+
1117
+ if inputs_embeds is None:
1118
+ inputs_embeds = self.embed_tokens(input_ids)
1119
+
1120
+ if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
1121
+ is_padding_right = attention_mask[:, -1].sum().item() != batch_size
1122
+ if is_padding_right:
1123
+ raise ValueError(
1124
+ "You are attempting to perform batched generation with padding_side='right'"
1125
+ " this may lead to unexpected behaviour for Flash Attention version of Phi3. Make sure to "
1126
+ " call `tokenizer.padding_side = 'left'` before tokenizing the input. "
1127
+ )
1128
+
1129
+ if self._attn_implementation == "flash_attention_2":
1130
+ # 2d mask is passed through the layers
1131
+ attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
1132
+ else:
1133
+ # 4d mask is passed through the layers
1134
+ attention_mask = _prepare_4d_causal_attention_mask(
1135
+ attention_mask,
1136
+ (batch_size, seq_length),
1137
+ inputs_embeds,
1138
+ past_key_values_length,
1139
+ sliding_window=self.config.sliding_window,
1140
+ )
1141
+
1142
+ hidden_states = inputs_embeds
1143
+
1144
+ # decoder layers
1145
+ all_hidden_states = () if output_hidden_states else None
1146
+ all_self_attns = () if output_attentions else None
1147
+ next_decoder_cache = None
1148
+
1149
+ for decoder_layer in self.layers:
1150
+ if output_hidden_states:
1151
+ all_hidden_states += (hidden_states,)
1152
+
1153
+ if self.gradient_checkpointing and self.training:
1154
+ layer_outputs = self._gradient_checkpointing_func(
1155
+ decoder_layer.__call__,
1156
+ hidden_states,
1157
+ attention_mask,
1158
+ position_ids,
1159
+ past_key_values,
1160
+ output_attentions,
1161
+ use_cache,
1162
+ )
1163
+ else:
1164
+ layer_outputs = decoder_layer(
1165
+ hidden_states,
1166
+ attention_mask=attention_mask,
1167
+ position_ids=position_ids,
1168
+ past_key_value=past_key_values,
1169
+ output_attentions=output_attentions,
1170
+ use_cache=use_cache,
1171
+ )
1172
+
1173
+ hidden_states = layer_outputs[0]
1174
+
1175
+ if use_cache:
1176
+ next_decoder_cache = layer_outputs[2 if output_attentions else 1]
1177
+
1178
+ if output_attentions:
1179
+ all_self_attns += (layer_outputs[1],)
1180
+
1181
+ hidden_states = self.norm(hidden_states)
1182
+
1183
+ # add hidden states from the last decoder layer
1184
+ if output_hidden_states:
1185
+ all_hidden_states += (hidden_states,)
1186
+
1187
+ next_cache = None
1188
+ if use_cache:
1189
+ next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
1190
+ if not return_dict:
1191
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
1192
+ return BaseModelOutputWithPast(
1193
+ last_hidden_state=hidden_states,
1194
+ past_key_values=next_cache,
1195
+ hidden_states=all_hidden_states,
1196
+ attentions=all_self_attns,
1197
+ )
1198
+
1199
+
1200
+ class Phi3ForCausalLM(Phi3PreTrainedModel):
1201
+ _tied_weights_keys = ["lm_head.weight"]
1202
+
1203
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with Llama->Phi3
1204
+ def __init__(self, config):
1205
+ super().__init__(config)
1206
+ self.model = Phi3Model(config)
1207
+ self.vocab_size = config.vocab_size
1208
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
1209
+
1210
+ # Initialize weights and apply final processing
1211
+ self.post_init()
1212
+
1213
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_input_embeddings
1214
+ def get_input_embeddings(self):
1215
+ return self.model.embed_tokens
1216
+
1217
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_input_embeddings
1218
+ def set_input_embeddings(self, value):
1219
+ self.model.embed_tokens = value
1220
+
1221
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_output_embeddings
1222
+ def get_output_embeddings(self):
1223
+ return self.lm_head
1224
+
1225
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_output_embeddings
1226
+ def set_output_embeddings(self, new_embeddings):
1227
+ self.lm_head = new_embeddings
1228
+
1229
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_decoder
1230
+ def set_decoder(self, decoder):
1231
+ self.model = decoder
1232
+
1233
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_decoder
1234
+ def get_decoder(self):
1235
+ return self.model
1236
+
1237
+ # Ignore copy
1238
+ @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
1239
+ @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
1240
+ def forward(
1241
+ self,
1242
+ input_ids: torch.LongTensor = None,
1243
+ attention_mask: Optional[torch.Tensor] = None,
1244
+ position_ids: Optional[torch.LongTensor] = None,
1245
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1246
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1247
+ labels: Optional[torch.LongTensor] = None,
1248
+ use_cache: Optional[bool] = None,
1249
+ output_attentions: Optional[bool] = None,
1250
+ output_hidden_states: Optional[bool] = None,
1251
+ return_dict: Optional[bool] = None,
1252
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
1253
+ r"""
1254
+ Args:
1255
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1256
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
1257
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
1258
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
1259
+
1260
+ Returns:
1261
+
1262
+ Example:
1263
+
1264
+ ```python
1265
+ >>> from transformers import AutoTokenizer, Phi3ForCausalLM
1266
+
1267
+ >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
1268
+ >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
1269
+
1270
+ >>> prompt = "This is an example script ."
1271
+ >>> inputs = tokenizer(prompt, return_tensors="pt")
1272
+
1273
+ >>> # Generate
1274
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
1275
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
1276
+ 'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
1277
+ ```"""
1278
+
1279
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
1280
+ output_hidden_states = (
1281
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
1282
+ )
1283
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1284
+
1285
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
1286
+ outputs = self.model(
1287
+ input_ids=input_ids,
1288
+ attention_mask=attention_mask,
1289
+ position_ids=position_ids,
1290
+ past_key_values=past_key_values,
1291
+ inputs_embeds=inputs_embeds,
1292
+ use_cache=use_cache,
1293
+ output_attentions=output_attentions,
1294
+ output_hidden_states=output_hidden_states,
1295
+ return_dict=return_dict,
1296
+ )
1297
+
1298
+ hidden_states = outputs[0]
1299
+ logits = self.lm_head(hidden_states)
1300
+ logits = logits.float()
1301
+
1302
+ loss = None
1303
+ if labels is not None:
1304
+ # Shift so that tokens < n predict n
1305
+ shift_logits = logits[..., :-1, :].contiguous()
1306
+ shift_labels = labels[..., 1:].contiguous()
1307
+ # Flatten the tokens
1308
+ loss_fct = CrossEntropyLoss()
1309
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
1310
+ shift_labels = shift_labels.view(-1)
1311
+ # Enable model parallelism
1312
+ shift_labels = shift_labels.to(shift_logits.device)
1313
+ loss = loss_fct(shift_logits, shift_labels)
1314
+
1315
+ if not return_dict:
1316
+ output = (logits,) + outputs[1:]
1317
+ return (loss,) + output if loss is not None else output
1318
+
1319
+ return CausalLMOutputWithPast(
1320
+ loss=loss,
1321
+ logits=logits,
1322
+ past_key_values=outputs.past_key_values,
1323
+ hidden_states=outputs.hidden_states,
1324
+ attentions=outputs.attentions,
1325
+ )
1326
+
1327
+ # Copied from transformers.models.persimmon.modeling_persimmon.PersimmonForCausalLM.prepare_inputs_for_generation
1328
+ def prepare_inputs_for_generation(
1329
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
1330
+ ):
1331
+ if past_key_values is not None:
1332
+ if isinstance(past_key_values, Cache):
1333
+ cache_length = past_key_values.get_seq_length()
1334
+ past_length = past_key_values.seen_tokens
1335
+ max_cache_length = past_key_values.get_max_length()
1336
+ else:
1337
+ cache_length = past_length = past_key_values[0][0].shape[2]
1338
+ max_cache_length = None
1339
+
1340
+ # Keep only the unprocessed tokens:
1341
+ # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
1342
+ # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
1343
+ # input)
1344
+ if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
1345
+ input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
1346
+ # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
1347
+ # input_ids based on the past_length.
1348
+ elif past_length < input_ids.shape[1]:
1349
+ input_ids = input_ids[:, past_length:]
1350
+ # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
1351
+
1352
+ # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
1353
+ if (
1354
+ max_cache_length is not None
1355
+ and attention_mask is not None
1356
+ and cache_length + input_ids.shape[1] > max_cache_length
1357
+ ):
1358
+ attention_mask = attention_mask[:, -max_cache_length:]
1359
+
1360
+ position_ids = kwargs.get("position_ids", None)
1361
+ if attention_mask is not None and position_ids is None:
1362
+ # create position_ids on the fly for batch generation
1363
+ position_ids = attention_mask.long().cumsum(-1) - 1
1364
+ position_ids.masked_fill_(attention_mask == 0, 1)
1365
+ if past_key_values:
1366
+ position_ids = position_ids[:, -input_ids.shape[1] :]
1367
+
1368
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
1369
+ if inputs_embeds is not None and past_key_values is None:
1370
+ model_inputs = {"inputs_embeds": inputs_embeds}
1371
+ else:
1372
+ model_inputs = {"input_ids": input_ids}
1373
+
1374
+ model_inputs.update(
1375
+ {
1376
+ "position_ids": position_ids,
1377
+ "past_key_values": past_key_values,
1378
+ "use_cache": kwargs.get("use_cache"),
1379
+ "attention_mask": attention_mask,
1380
+ }
1381
+ )
1382
+ return model_inputs
1383
+
1384
+ @staticmethod
1385
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM._reorder_cache
1386
+ def _reorder_cache(past_key_values, beam_idx):
1387
+ reordered_past = ()
1388
+ for layer_past in past_key_values:
1389
+ reordered_past += (
1390
+ tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
1391
+ )
1392
+ return reordered_past
1393
+
1394
+
1395
+ @add_start_docstrings(
1396
+ """
1397
+ The [`Phi3Model`] with a sequence classification head on top (linear layer).
1398
+
1399
+ [`Phi3ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
1400
+ (e.g. GPT-2) do.
1401
+
1402
+ Since it does classification on the last token, it requires to know the position of the last token. If a
1403
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
1404
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
1405
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
1406
+ each row of the batch).
1407
+ """,
1408
+ PHI3_START_DOCSTRING,
1409
+ )
1410
+ # Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Phi3, LLAMA->PHI3, self.transformer->self.model, transformer_outputs->model_outputs
1411
+ class Phi3ForSequenceClassification(Phi3PreTrainedModel):
1412
+ def __init__(self, config):
1413
+ super().__init__(config)
1414
+ self.num_labels = config.num_labels
1415
+ self.model = Phi3Model(config)
1416
+ self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
1417
+
1418
+ # Initialize weights and apply final processing
1419
+ self.post_init()
1420
+
1421
+ def get_input_embeddings(self):
1422
+ return self.model.embed_tokens
1423
+
1424
+ def set_input_embeddings(self, value):
1425
+ self.model.embed_tokens = value
1426
+
1427
+ @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
1428
+ def forward(
1429
+ self,
1430
+ input_ids: torch.LongTensor = None,
1431
+ attention_mask: Optional[torch.Tensor] = None,
1432
+ position_ids: Optional[torch.LongTensor] = None,
1433
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1434
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1435
+ labels: Optional[torch.LongTensor] = None,
1436
+ use_cache: Optional[bool] = None,
1437
+ output_attentions: Optional[bool] = None,
1438
+ output_hidden_states: Optional[bool] = None,
1439
+ return_dict: Optional[bool] = None,
1440
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
1441
+ r"""
1442
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1443
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1444
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1445
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1446
+ """
1447
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1448
+
1449
+ model_outputs = self.model(
1450
+ input_ids,
1451
+ attention_mask=attention_mask,
1452
+ position_ids=position_ids,
1453
+ past_key_values=past_key_values,
1454
+ inputs_embeds=inputs_embeds,
1455
+ use_cache=use_cache,
1456
+ output_attentions=output_attentions,
1457
+ output_hidden_states=output_hidden_states,
1458
+ return_dict=return_dict,
1459
+ )
1460
+ hidden_states = model_outputs[0]
1461
+ logits = self.score(hidden_states)
1462
+
1463
+ if input_ids is not None:
1464
+ batch_size = input_ids.shape[0]
1465
+ else:
1466
+ batch_size = inputs_embeds.shape[0]
1467
+
1468
+ if self.config.pad_token_id is None and batch_size != 1:
1469
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
1470
+ if self.config.pad_token_id is None:
1471
+ sequence_lengths = -1
1472
+ else:
1473
+ if input_ids is not None:
1474
+ # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
1475
+ sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
1476
+ sequence_lengths = sequence_lengths % input_ids.shape[-1]
1477
+ sequence_lengths = sequence_lengths.to(logits.device)
1478
+ else:
1479
+ sequence_lengths = -1
1480
+
1481
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
1482
+
1483
+ loss = None
1484
+ if labels is not None:
1485
+ labels = labels.to(logits.device)
1486
+ if self.config.problem_type is None:
1487
+ if self.num_labels == 1:
1488
+ self.config.problem_type = "regression"
1489
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1490
+ self.config.problem_type = "single_label_classification"
1491
+ else:
1492
+ self.config.problem_type = "multi_label_classification"
1493
+
1494
+ if self.config.problem_type == "regression":
1495
+ loss_fct = MSELoss()
1496
+ if self.num_labels == 1:
1497
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1498
+ else:
1499
+ loss = loss_fct(pooled_logits, labels)
1500
+ elif self.config.problem_type == "single_label_classification":
1501
+ loss_fct = CrossEntropyLoss()
1502
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
1503
+ elif self.config.problem_type == "multi_label_classification":
1504
+ loss_fct = BCEWithLogitsLoss()
1505
+ loss = loss_fct(pooled_logits, labels)
1506
+ if not return_dict:
1507
+ output = (pooled_logits,) + model_outputs[1:]
1508
+ return ((loss,) + output) if loss is not None else output
1509
+
1510
+ return SequenceClassifierOutputWithPast(
1511
+ loss=loss,
1512
+ logits=pooled_logits,
1513
+ past_key_values=model_outputs.past_key_values,
1514
+ hidden_states=model_outputs.hidden_states,
1515
+ attentions=model_outputs.attentions,
1516
+ )
1517
+
1518
+
1519
+ @add_start_docstrings(
1520
+ """
1521
+ [`Phi3Model`] with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
1522
+ Named-Entity-Recognition (NER) tasks.
1523
+ """,
1524
+ PHI3_START_DOCSTRING,
1525
+ )
1526
+ # Copied from transformers.models.mpt.modeling_mpt.MptForTokenClassification with Mpt->Phi3,MPT->PHI3,self.transformer->self.model,transformer_outputs->model_outputs
1527
+ class Phi3ForTokenClassification(Phi3PreTrainedModel):
1528
+ def __init__(self, config: Phi3Config):
1529
+ super().__init__(config)
1530
+ self.num_labels = config.num_labels
1531
+
1532
+ self.model = Phi3Model(config)
1533
+ if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
1534
+ classifier_dropout = config.classifier_dropout
1535
+ elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
1536
+ classifier_dropout = config.hidden_dropout
1537
+ else:
1538
+ classifier_dropout = 0.1
1539
+ self.dropout = nn.Dropout(classifier_dropout)
1540
+ self.classifier = nn.Linear(config.hidden_size, config.num_labels)
1541
+
1542
+ # Initialize weights and apply final processing
1543
+ self.post_init()
1544
+
1545
+ @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
1546
+ @add_code_sample_docstrings(
1547
+ checkpoint=_CHECKPOINT_FOR_DOC,
1548
+ output_type=TokenClassifierOutput,
1549
+ config_class=_CONFIG_FOR_DOC,
1550
+ )
1551
+ def forward(
1552
+ self,
1553
+ input_ids: Optional[torch.LongTensor] = None,
1554
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
1555
+ attention_mask: Optional[torch.Tensor] = None,
1556
+ inputs_embeds: Optional[torch.Tensor] = None,
1557
+ labels: Optional[torch.Tensor] = None,
1558
+ use_cache: Optional[bool] = None,
1559
+ output_attentions: Optional[bool] = None,
1560
+ output_hidden_states: Optional[bool] = None,
1561
+ return_dict: Optional[bool] = None,
1562
+ **deprecated_arguments,
1563
+ ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
1564
+ r"""
1565
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1566
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1567
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1568
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1569
+ """
1570
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1571
+
1572
+ model_outputs = self.model(
1573
+ input_ids,
1574
+ past_key_values=past_key_values,
1575
+ attention_mask=attention_mask,
1576
+ inputs_embeds=inputs_embeds,
1577
+ use_cache=use_cache,
1578
+ output_attentions=output_attentions,
1579
+ output_hidden_states=output_hidden_states,
1580
+ return_dict=return_dict,
1581
+ )
1582
+
1583
+ hidden_states = model_outputs[0]
1584
+ hidden_states = self.dropout(hidden_states)
1585
+ logits = self.classifier(hidden_states)
1586
+
1587
+ loss = None
1588
+ if labels is not None:
1589
+ # move labels to correct device to enable model parallelism
1590
+ labels = labels.to(logits.device)
1591
+ batch_size, seq_length = labels.shape
1592
+ loss_fct = CrossEntropyLoss()
1593
+ loss = loss_fct(
1594
+ logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length)
1595
+ )
1596
+
1597
+ if not return_dict:
1598
+ output = (logits,) + model_outputs[2:]
1599
+ return ((loss,) + output) if loss is not None else output
1600
+
1601
+ return TokenClassifierOutput(
1602
+ loss=loss,
1603
+ logits=logits,
1604
+ hidden_states=model_outputs.hidden_states,
1605
+ attentions=model_outputs.attentions,
1606
+ )