mgoin commited on
Commit
c2786da
1 Parent(s): 7725cbc

Upload 2 files

Browse files
Files changed (2) hide show
  1. configuration_phi3.py +227 -0
  2. modeling_phi3.py +1570 -0
configuration_phi3.py ADDED
@@ -0,0 +1,227 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 `longrope` 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=32064,
120
+ hidden_size=3072,
121
+ intermediate_size=8192,
122
+ num_hidden_layers=32,
123
+ num_attention_heads=32,
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=4096,
130
+ original_max_position_embeddings=4096,
131
+ initializer_range=0.02,
132
+ rms_norm_eps=1e-5,
133
+ use_cache=True,
134
+ tie_word_embeddings=False,
135
+ rope_theta=10000.0,
136
+ rope_scaling=None,
137
+ bos_token_id=1,
138
+ eos_token_id=32000,
139
+ pad_token_id=32000,
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_adjustment()
165
+ self._rope_scaling_validation()
166
+ self.sliding_window = sliding_window
167
+
168
+ super().__init__(
169
+ bos_token_id=bos_token_id,
170
+ eos_token_id=eos_token_id,
171
+ pad_token_id=pad_token_id,
172
+ tie_word_embeddings=tie_word_embeddings,
173
+ **kwargs,
174
+ )
175
+
176
+ def _rope_scaling_adjustment(self):
177
+ """
178
+ Adjust the `type` of the `rope_scaling` configuration for backward compatibility.
179
+ """
180
+ if self.rope_scaling is None:
181
+ return
182
+
183
+ rope_scaling_type = self.rope_scaling.get("type", None)
184
+
185
+ # For backward compatibility if previous version used "su" or "yarn"
186
+ if rope_scaling_type is not None and rope_scaling_type in ["su", "yarn"]:
187
+ self.rope_scaling["type"] = "longrope"
188
+
189
+ def _rope_scaling_validation(self):
190
+ """
191
+ Validate the `rope_scaling` configuration.
192
+ """
193
+ if self.rope_scaling is None:
194
+ return
195
+
196
+ if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 3:
197
+ raise ValueError(
198
+ "`rope_scaling` must be a dictionary with three fields, `type`, `short_factor` and `long_factor`, "
199
+ f"got {self.rope_scaling}"
200
+ )
201
+ rope_scaling_type = self.rope_scaling.get("type", None)
202
+ rope_scaling_short_factor = self.rope_scaling.get("short_factor", None)
203
+ rope_scaling_long_factor = self.rope_scaling.get("long_factor", None)
204
+ if rope_scaling_type is None or rope_scaling_type not in ["longrope"]:
205
+ raise ValueError(f"`rope_scaling`'s type field must be one of ['longrope'], got {rope_scaling_type}")
206
+ if not (
207
+ isinstance(rope_scaling_short_factor, list)
208
+ and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor)
209
+ ):
210
+ raise ValueError(
211
+ f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}"
212
+ )
213
+ if not len(rope_scaling_short_factor) == self.hidden_size // self.num_attention_heads // 2:
214
+ raise ValueError(
215
+ f"`rope_scaling`'s short_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_short_factor)}"
216
+ )
217
+ if not (
218
+ isinstance(rope_scaling_long_factor, list)
219
+ and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor)
220
+ ):
221
+ raise ValueError(
222
+ f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}"
223
+ )
224
+ if not len(rope_scaling_long_factor) == self.hidden_size // self.num_attention_heads // 2:
225
+ raise ValueError(
226
+ f"`rope_scaling`'s long_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_long_factor)}"
227
+ )
modeling_phi3.py ADDED
@@ -0,0 +1,1570 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 Phi3LongRoPEScaledRotaryEmbedding(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 = seq_len or 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
+ # Copied from transformers.models.llama.modeling_llama.rotate_half
184
+ def rotate_half(x):
185
+ """Rotates half the hidden dims of the input."""
186
+ x1 = x[..., : x.shape[-1] // 2]
187
+ x2 = x[..., x.shape[-1] // 2 :]
188
+ return torch.cat((-x2, x1), dim=-1)
189
+
190
+
191
+ # Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
192
+ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
193
+ """Applies Rotary Position Embedding to the query and key tensors.
194
+
195
+ Args:
196
+ q (`torch.Tensor`): The query tensor.
197
+ k (`torch.Tensor`): The key tensor.
198
+ cos (`torch.Tensor`): The cosine part of the rotary embedding.
199
+ sin (`torch.Tensor`): The sine part of the rotary embedding.
200
+ position_ids (`torch.Tensor`, *optional*):
201
+ Deprecated and unused.
202
+ unsqueeze_dim (`int`, *optional*, defaults to 1):
203
+ The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
204
+ sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
205
+ that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
206
+ k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
207
+ cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
208
+ the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
209
+ Returns:
210
+ `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
211
+ """
212
+ cos = cos.unsqueeze(unsqueeze_dim)
213
+ sin = sin.unsqueeze(unsqueeze_dim)
214
+ q_embed = (q * cos) + (rotate_half(q) * sin)
215
+ k_embed = (k * cos) + (rotate_half(k) * sin)
216
+ return q_embed, k_embed
217
+
218
+
219
+ class Phi3MLP(nn.Module):
220
+ def __init__(self, config):
221
+ super().__init__()
222
+
223
+ self.config = config
224
+ self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
225
+ self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
226
+
227
+ self.activation_fn = ACT2FN[config.hidden_act]
228
+
229
+ def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
230
+ up_states = self.gate_up_proj(hidden_states)
231
+
232
+ gate, up_states = up_states.chunk(2, dim=-1)
233
+ up_states = up_states * self.activation_fn(gate)
234
+
235
+ return self.down_proj(up_states)
236
+
237
+
238
+ # Copied from transformers.models.llama.modeling_llama.repeat_kv with llama->phi
239
+ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
240
+ """
241
+ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
242
+ num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
243
+ """
244
+ batch, num_key_value_heads, slen, head_dim = hidden_states.shape
245
+ if n_rep == 1:
246
+ return hidden_states
247
+ hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
248
+ return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
249
+
250
+
251
+ class Phi3Attention(nn.Module):
252
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
253
+
254
+ def __init__(self, config: Phi3Config, layer_idx: Optional[int] = None):
255
+ super().__init__()
256
+ self.config = config
257
+ self.layer_idx = layer_idx
258
+ if layer_idx is None:
259
+ logger.warning_once(
260
+ f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
261
+ "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
262
+ "when creating this class."
263
+ )
264
+
265
+ self.attention_dropout = config.attention_dropout
266
+ self.hidden_size = config.hidden_size
267
+ self.num_heads = config.num_attention_heads
268
+ self.head_dim = self.hidden_size // self.num_heads
269
+ self.num_key_value_heads = config.num_key_value_heads
270
+ self.num_key_value_groups = self.num_heads // self.num_key_value_heads
271
+ self.max_position_embeddings = config.max_position_embeddings
272
+ self.original_max_position_embeddings = config.original_max_position_embeddings
273
+ self.rope_theta = config.rope_theta
274
+ self.rope_scaling = config.rope_scaling
275
+ self.is_causal = True
276
+
277
+ if (self.head_dim * self.num_heads) != self.hidden_size:
278
+ raise ValueError(
279
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
280
+ f" and `num_heads`: {self.num_heads})."
281
+ )
282
+
283
+ op_size = self.num_heads * self.head_dim + 2 * (self.num_key_value_heads * self.head_dim)
284
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
285
+ self.qkv_proj = nn.Linear(self.hidden_size, op_size, bias=False)
286
+ self._init_rope()
287
+
288
+ def _init_rope(self):
289
+ if self.rope_scaling is None:
290
+ self.rotary_emb = Phi3RotaryEmbedding(
291
+ self.head_dim,
292
+ max_position_embeddings=self.max_position_embeddings,
293
+ base=self.rope_theta,
294
+ )
295
+ else:
296
+ scaling_type = self.config.rope_scaling["type"]
297
+ if scaling_type == "longrope":
298
+ self.rotary_emb = Phi3LongRoPEScaledRotaryEmbedding(self.head_dim, self.config)
299
+ else:
300
+ raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
301
+
302
+ def forward(
303
+ self,
304
+ hidden_states: torch.Tensor,
305
+ attention_mask: Optional[torch.Tensor] = None,
306
+ position_ids: Optional[torch.LongTensor] = None,
307
+ past_key_value: Optional[Cache] = None,
308
+ output_attentions: bool = False,
309
+ use_cache: bool = False,
310
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
311
+ logger.warning_once("You are not running the flash-attention implementation, expect numerical differences.")
312
+
313
+ bsz, q_len, _ = hidden_states.size()
314
+
315
+ qkv = self.qkv_proj(hidden_states)
316
+ query_pos = self.num_heads * self.head_dim
317
+ query_states = qkv[..., :query_pos]
318
+ key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
319
+ value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
320
+
321
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
322
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
323
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
324
+
325
+ kv_seq_len = key_states.shape[-2]
326
+ if past_key_value is not None:
327
+ if self.layer_idx is None:
328
+ raise ValueError(
329
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
330
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
331
+ "with a layer index."
332
+ )
333
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
334
+ cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
335
+
336
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
337
+
338
+ if past_key_value is not None:
339
+ cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
340
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
341
+
342
+ # repeat k/v heads if n_kv_heads < n_heads
343
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
344
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
345
+
346
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
347
+
348
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
349
+ raise ValueError(
350
+ f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
351
+ f" {attn_weights.size()}"
352
+ )
353
+
354
+ if attention_mask is not None:
355
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
356
+ raise ValueError(
357
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
358
+ )
359
+ attn_weights = attn_weights + attention_mask
360
+
361
+ # upcast attention to fp32
362
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype)
363
+ attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
364
+
365
+ attn_output = torch.matmul(attn_weights, value_states)
366
+
367
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
368
+ raise ValueError(
369
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
370
+ f" {attn_output.size()}"
371
+ )
372
+
373
+ attn_output = attn_output.transpose(1, 2).contiguous()
374
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
375
+
376
+ attn_output = self.o_proj(attn_output)
377
+
378
+ if not output_attentions:
379
+ attn_weights = None
380
+
381
+ return attn_output, attn_weights, past_key_value
382
+
383
+
384
+ class Phi3FlashAttention2(Phi3Attention):
385
+ """
386
+ Phi-3 flash attention module. This module inherits from `Phi3Attention` as the weights of the module stays
387
+ untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
388
+ flash attention and deal with padding tokens in case the input contains any of them.
389
+ """
390
+
391
+ # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
392
+ def __init__(self, *args, **kwargs):
393
+ super().__init__(*args, **kwargs)
394
+
395
+ # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
396
+ # 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.
397
+ # 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).
398
+ self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
399
+
400
+ def forward(
401
+ self,
402
+ hidden_states: torch.Tensor,
403
+ attention_mask: Optional[torch.LongTensor] = None,
404
+ position_ids: Optional[torch.LongTensor] = None,
405
+ past_key_value: Optional[Cache] = None,
406
+ output_attentions: bool = False,
407
+ use_cache: bool = False,
408
+ **kwargs,
409
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
410
+ # Phi3FlashAttention2 attention does not support output_attentions
411
+
412
+ if not _flash_supports_window_size:
413
+ logger.warning_once(
414
+ "The current flash attention version does not support sliding window attention. Please use `attn_implementation='eager'` or upgrade flash-attn library."
415
+ )
416
+ raise ValueError("The current flash attention version does not support sliding window attention.")
417
+
418
+ output_attentions = False
419
+
420
+ if "padding_mask" in kwargs:
421
+ warnings.warn(
422
+ "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
423
+ )
424
+
425
+ # overwrite attention_mask with padding_mask
426
+ attention_mask = kwargs.pop("padding_mask")
427
+
428
+ bsz, q_len, _ = hidden_states.size()
429
+
430
+ qkv = self.qkv_proj(hidden_states)
431
+ query_pos = self.num_heads * self.head_dim
432
+ query_states = qkv[..., :query_pos]
433
+ key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
434
+ value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
435
+
436
+ # Flash attention requires the input to have the shape
437
+ # batch_size x seq_length x head_dim x hidden_dim
438
+ # therefore we just need to keep the original shape
439
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
440
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
441
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
442
+
443
+ kv_seq_len = key_states.shape[-2]
444
+ if past_key_value is not None:
445
+ if self.layer_idx is None:
446
+ raise ValueError(
447
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
448
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
449
+ "with a layer index."
450
+ )
451
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
452
+
453
+ # Because the input can be padded, the absolute sequence length depends on the max position id.
454
+ rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item() + 1)
455
+ cos, sin = self.rotary_emb(value_states, position_ids, seq_len=rotary_seq_len)
456
+
457
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
458
+
459
+ use_sliding_windows = (
460
+ _flash_supports_window_size
461
+ and getattr(self.config, "sliding_window", None) is not None
462
+ and kv_seq_len > self.config.sliding_window
463
+ )
464
+
465
+ if past_key_value is not None:
466
+ # Activate slicing cache only if the config has a value `sliding_windows` attribute
467
+ cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
468
+ if (
469
+ getattr(self.config, "sliding_window", None) is not None
470
+ and kv_seq_len > self.config.sliding_window
471
+ and cache_has_contents
472
+ ):
473
+ slicing_tokens = 1 - self.config.sliding_window
474
+
475
+ past_key = past_key_value[self.layer_idx][0]
476
+ past_value = past_key_value[self.layer_idx][1]
477
+
478
+ past_key = past_key[:, :, slicing_tokens:, :].contiguous()
479
+ past_value = past_value[:, :, slicing_tokens:, :].contiguous()
480
+
481
+ if past_key.shape[-2] != self.config.sliding_window - 1:
482
+ raise ValueError(
483
+ f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
484
+ f" {past_key.shape}"
485
+ )
486
+
487
+ if attention_mask is not None:
488
+ attention_mask = attention_mask[:, slicing_tokens:]
489
+ attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
490
+
491
+ cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
492
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
493
+
494
+ # repeat k/v heads if n_kv_heads < n_heads
495
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
496
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
497
+
498
+ attn_dropout = self.attention_dropout if self.training else 0.0
499
+
500
+ # In PEFT, usually we cast the layer norms in float32 for training stability reasons
501
+ # therefore the input hidden states gets silently casted in float32. Hence, we need
502
+ # cast them back in the correct dtype just to be sure everything works as expected.
503
+ # This might slowdown training & inference so it is recommended to not cast the LayerNorms
504
+ # in fp32.
505
+
506
+ if query_states.dtype == torch.float32:
507
+ if torch.is_autocast_enabled():
508
+ target_dtype = torch.get_autocast_gpu_dtype()
509
+ # Handle the case where the model is quantized
510
+ elif hasattr(self.config, "_pre_quantization_dtype"):
511
+ target_dtype = self.config._pre_quantization_dtype
512
+ else:
513
+ target_dtype = self.qkv_proj.weight.dtype
514
+
515
+ logger.warning_once(
516
+ f"The input hidden states seems to be silently casted in float32, this might be related to"
517
+ f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
518
+ f" {target_dtype}."
519
+ )
520
+
521
+ query_states = query_states.to(target_dtype)
522
+ key_states = key_states.to(target_dtype)
523
+ value_states = value_states.to(target_dtype)
524
+
525
+ # Reashape to the expected shape for Flash Attention
526
+ query_states = query_states.transpose(1, 2)
527
+ key_states = key_states.transpose(1, 2)
528
+ value_states = value_states.transpose(1, 2)
529
+
530
+ attn_output = self._flash_attention_forward(
531
+ query_states,
532
+ key_states,
533
+ value_states,
534
+ attention_mask,
535
+ q_len,
536
+ dropout=attn_dropout,
537
+ use_sliding_windows=use_sliding_windows,
538
+ )
539
+
540
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
541
+ attn_output = self.o_proj(attn_output)
542
+
543
+ if not output_attentions:
544
+ attn_weights = None
545
+
546
+ return attn_output, attn_weights, past_key_value
547
+
548
+ # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._flash_attention_forward
549
+ def _flash_attention_forward(
550
+ self,
551
+ query_states,
552
+ key_states,
553
+ value_states,
554
+ attention_mask,
555
+ query_length,
556
+ dropout=0.0,
557
+ softmax_scale=None,
558
+ use_sliding_windows=False,
559
+ ):
560
+ """
561
+ Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
562
+ first unpad the input, then computes the attention scores and pad the final attention scores.
563
+
564
+ Args:
565
+ query_states (`torch.Tensor`):
566
+ Input query states to be passed to Flash Attention API
567
+ key_states (`torch.Tensor`):
568
+ Input key states to be passed to Flash Attention API
569
+ value_states (`torch.Tensor`):
570
+ Input value states to be passed to Flash Attention API
571
+ attention_mask (`torch.Tensor`):
572
+ The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
573
+ position of padding tokens and 1 for the position of non-padding tokens.
574
+ dropout (`float`):
575
+ Attention dropout
576
+ softmax_scale (`float`, *optional*):
577
+ The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
578
+ use_sliding_windows (`bool`, *optional*):
579
+ Whether to activate sliding window attention.
580
+ """
581
+ if not self._flash_attn_uses_top_left_mask:
582
+ causal = self.is_causal
583
+ else:
584
+ # 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__.
585
+ causal = self.is_causal and query_length != 1
586
+
587
+ # Contains at least one padding token in the sequence
588
+ if attention_mask is not None:
589
+ batch_size = query_states.shape[0]
590
+ query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
591
+ query_states, key_states, value_states, attention_mask, query_length
592
+ )
593
+
594
+ cu_seqlens_q, cu_seqlens_k = cu_seq_lens
595
+ max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
596
+
597
+ if not use_sliding_windows:
598
+ attn_output_unpad = flash_attn_varlen_func(
599
+ query_states,
600
+ key_states,
601
+ value_states,
602
+ cu_seqlens_q=cu_seqlens_q,
603
+ cu_seqlens_k=cu_seqlens_k,
604
+ max_seqlen_q=max_seqlen_in_batch_q,
605
+ max_seqlen_k=max_seqlen_in_batch_k,
606
+ dropout_p=dropout,
607
+ softmax_scale=softmax_scale,
608
+ causal=causal,
609
+ )
610
+ else:
611
+ attn_output_unpad = flash_attn_varlen_func(
612
+ query_states,
613
+ key_states,
614
+ value_states,
615
+ cu_seqlens_q=cu_seqlens_q,
616
+ cu_seqlens_k=cu_seqlens_k,
617
+ max_seqlen_q=max_seqlen_in_batch_q,
618
+ max_seqlen_k=max_seqlen_in_batch_k,
619
+ dropout_p=dropout,
620
+ softmax_scale=softmax_scale,
621
+ causal=causal,
622
+ window_size=(self.config.sliding_window, self.config.sliding_window),
623
+ )
624
+
625
+ attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
626
+ else:
627
+ if not use_sliding_windows:
628
+ attn_output = flash_attn_func(
629
+ query_states,
630
+ key_states,
631
+ value_states,
632
+ dropout,
633
+ softmax_scale=softmax_scale,
634
+ causal=causal,
635
+ )
636
+ else:
637
+ attn_output = flash_attn_func(
638
+ query_states,
639
+ key_states,
640
+ value_states,
641
+ dropout,
642
+ softmax_scale=softmax_scale,
643
+ causal=causal,
644
+ window_size=(self.config.sliding_window, self.config.sliding_window),
645
+ )
646
+
647
+ return attn_output
648
+
649
+ # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input
650
+ def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
651
+ batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
652
+
653
+ # On the first iteration we need to properly re-create the padding mask
654
+ # by slicing it on the proper place
655
+ if kv_seq_len != attention_mask.shape[-1]:
656
+ attention_mask_num_tokens = attention_mask.shape[-1]
657
+ attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
658
+
659
+ indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
660
+
661
+ key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
662
+ value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
663
+
664
+ if query_length == kv_seq_len:
665
+ query_layer = index_first_axis(
666
+ query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
667
+ )
668
+ cu_seqlens_q = cu_seqlens_k
669
+ max_seqlen_in_batch_q = max_seqlen_in_batch_k
670
+ indices_q = indices_k
671
+ elif query_length == 1:
672
+ max_seqlen_in_batch_q = 1
673
+ cu_seqlens_q = torch.arange(
674
+ batch_size + 1, dtype=torch.int32, device=query_layer.device
675
+ ) # There is a memcpy here, that is very bad.
676
+ indices_q = cu_seqlens_q[:-1]
677
+ query_layer = query_layer.squeeze(1)
678
+ else:
679
+ # The -q_len: slice assumes left padding.
680
+ attention_mask = attention_mask[:, -query_length:]
681
+ query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
682
+
683
+ return (
684
+ query_layer,
685
+ key_layer,
686
+ value_layer,
687
+ indices_q,
688
+ (cu_seqlens_q, cu_seqlens_k),
689
+ (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
690
+ )
691
+
692
+
693
+ # copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Phi3
694
+ # TODO @Arthur no longer copied from LLama after static cache
695
+ class Phi3SdpaAttention(Phi3Attention):
696
+ """
697
+ Phi3 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
698
+ `Phi3Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
699
+ SDPA API.
700
+ """
701
+
702
+ # Adapted from Phi3Attention.forward
703
+ def forward(
704
+ self,
705
+ hidden_states: torch.Tensor,
706
+ attention_mask: Optional[torch.Tensor] = None,
707
+ position_ids: Optional[torch.LongTensor] = None,
708
+ past_key_value: Optional[Cache] = None,
709
+ output_attentions: bool = False,
710
+ use_cache: bool = False,
711
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
712
+ if output_attentions:
713
+ # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
714
+ logger.warning_once(
715
+ "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, "
716
+ '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.'
717
+ )
718
+ return super().forward(
719
+ hidden_states=hidden_states,
720
+ attention_mask=attention_mask,
721
+ position_ids=position_ids,
722
+ past_key_value=past_key_value,
723
+ output_attentions=output_attentions,
724
+ use_cache=use_cache,
725
+ )
726
+
727
+ bsz, q_len, _ = hidden_states.size()
728
+
729
+ qkv = self.qkv_proj(hidden_states)
730
+ query_pos = self.num_heads * self.head_dim
731
+ query_states = qkv[..., :query_pos]
732
+ key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
733
+ value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
734
+
735
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
736
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
737
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
738
+
739
+ kv_seq_len = key_states.shape[-2]
740
+ if past_key_value is not None:
741
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
742
+ cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
743
+
744
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
745
+
746
+ if past_key_value is not None:
747
+ cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
748
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
749
+
750
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
751
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
752
+
753
+ if attention_mask is not None:
754
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
755
+ raise ValueError(
756
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
757
+ )
758
+
759
+ # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
760
+ # Reference: https://github.com/pytorch/pytorch/issues/112577.
761
+ if query_states.device.type == "cuda" and attention_mask is not None:
762
+ query_states = query_states.contiguous()
763
+ key_states = key_states.contiguous()
764
+ value_states = value_states.contiguous()
765
+
766
+ attn_output = torch.nn.functional.scaled_dot_product_attention(
767
+ query_states,
768
+ key_states,
769
+ value_states,
770
+ attn_mask=attention_mask,
771
+ dropout_p=self.attention_dropout if self.training else 0.0,
772
+ # 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.
773
+ is_causal=self.is_causal and attention_mask is None and q_len > 1,
774
+ )
775
+
776
+ attn_output = attn_output.transpose(1, 2).contiguous()
777
+ attn_output = attn_output.view(bsz, q_len, self.hidden_size)
778
+
779
+ attn_output = self.o_proj(attn_output)
780
+
781
+ return attn_output, None, past_key_value
782
+
783
+
784
+ PHI3_ATTENTION_CLASSES = {
785
+ "eager": Phi3Attention,
786
+ "flash_attention_2": Phi3FlashAttention2,
787
+ "sdpa": Phi3SdpaAttention,
788
+ }
789
+
790
+
791
+ class Phi3DecoderLayer(nn.Module):
792
+ def __init__(self, config: Phi3Config, layer_idx: int):
793
+ super().__init__()
794
+
795
+ self.config = config
796
+ self.self_attn = PHI3_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
797
+
798
+ self.mlp = Phi3MLP(config)
799
+ self.input_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
800
+
801
+ self.resid_attn_dropout = nn.Dropout(config.resid_pdrop)
802
+ self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop)
803
+ self.post_attention_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
804
+
805
+ def forward(
806
+ self,
807
+ hidden_states: torch.Tensor,
808
+ attention_mask: Optional[torch.Tensor] = None,
809
+ position_ids: Optional[torch.LongTensor] = None,
810
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
811
+ output_attentions: Optional[bool] = False,
812
+ use_cache: Optional[bool] = False,
813
+ **kwargs,
814
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
815
+ if "padding_mask" in kwargs:
816
+ warnings.warn(
817
+ "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
818
+ )
819
+ """
820
+ Args:
821
+ hidden_states (`torch.FloatTensor`):
822
+ input to the layer of shape `(batch, seq_len, embed_dim)`
823
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
824
+ `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
825
+ position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
826
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range
827
+ `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
828
+ output_attentions (`bool`, *optional*):
829
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
830
+ returned tensors for more detail.
831
+ use_cache (`bool`, *optional*):
832
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
833
+ (see `past_key_values`).
834
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
835
+ """
836
+
837
+ residual = hidden_states
838
+
839
+ hidden_states = self.input_layernorm(hidden_states)
840
+
841
+ # Self Attention
842
+ attn_outputs, self_attn_weights, present_key_value = self.self_attn(
843
+ hidden_states=hidden_states,
844
+ attention_mask=attention_mask,
845
+ position_ids=position_ids,
846
+ past_key_value=past_key_value,
847
+ output_attentions=output_attentions,
848
+ use_cache=use_cache,
849
+ )
850
+
851
+ hidden_states = residual + self.resid_attn_dropout(attn_outputs)
852
+
853
+ residual = hidden_states
854
+ hidden_states = self.post_attention_layernorm(hidden_states)
855
+ hidden_states = self.mlp(hidden_states)
856
+ hidden_states = residual + self.resid_mlp_dropout(hidden_states)
857
+
858
+ outputs = (hidden_states,)
859
+
860
+ if output_attentions:
861
+ outputs += (self_attn_weights,)
862
+
863
+ if use_cache:
864
+ outputs += (present_key_value,)
865
+
866
+ return outputs
867
+
868
+
869
+ PHI3_START_DOCSTRING = r"""
870
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
871
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
872
+ etc.)
873
+
874
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
875
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
876
+ and behavior.
877
+
878
+ Parameters:
879
+ config ([`Phi3Config`]):
880
+ Model configuration class with all the parameters of the model. Initializing with a config file does not
881
+ load the weights associated with the model, only the configuration. Check out the
882
+ [`~PreTrainedModel.from_pretrained`] method to load the model weights.
883
+ """
884
+
885
+
886
+ @add_start_docstrings(
887
+ "The bare Phi-3 model outputting raw hidden-states without any specific head on top.",
888
+ PHI3_START_DOCSTRING,
889
+ )
890
+ class Phi3PreTrainedModel(PreTrainedModel):
891
+ config_class = Phi3Config
892
+ base_model_prefix = "model"
893
+ supports_gradient_checkpointing = True
894
+ _no_split_modules = ["Phi3DecoderLayer"]
895
+ _skip_keys_device_placement = "past_key_values"
896
+ _supports_flash_attn_2 = True
897
+ _supports_sdpa = False
898
+ _supports_cache_class = True
899
+
900
+ _version = "0.0.5"
901
+
902
+ def _init_weights(self, module):
903
+ std = self.config.initializer_range
904
+ if isinstance(module, nn.Linear):
905
+ module.weight.data.normal_(mean=0.0, std=std)
906
+ if module.bias is not None:
907
+ module.bias.data.zero_()
908
+ elif isinstance(module, nn.Embedding):
909
+ module.weight.data.normal_(mean=0.0, std=std)
910
+ if module.padding_idx is not None:
911
+ module.weight.data[module.padding_idx].zero_()
912
+
913
+
914
+ PHI3_INPUTS_DOCSTRING = r"""
915
+ Args:
916
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
917
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
918
+ it.
919
+
920
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
921
+ [`PreTrainedTokenizer.__call__`] for details.
922
+
923
+ [What are input IDs?](../glossary#input-ids)
924
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
925
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
926
+
927
+ - 1 for tokens that are **not masked**,
928
+ - 0 for tokens that are **masked**.
929
+
930
+ [What are attention masks?](../glossary#attention-mask)
931
+
932
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
933
+ [`PreTrainedTokenizer.__call__`] for details.
934
+
935
+ If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
936
+ `past_key_values`).
937
+
938
+ If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
939
+ and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
940
+ information on the default strategy.
941
+
942
+ - 1 indicates the head is **not masked**,
943
+ - 0 indicates the head is **masked**.
944
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
945
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
946
+ config.n_positions - 1]`.
947
+
948
+ [What are position IDs?](../glossary#position-ids)
949
+ past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
950
+ Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
951
+ blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
952
+ returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
953
+
954
+ Two formats are allowed:
955
+ - a [`~cache_utils.Cache`] instance;
956
+ - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
957
+ shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
958
+ cache format.
959
+
960
+ The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
961
+ legacy cache format will be returned.
962
+
963
+ If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
964
+ have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
965
+ of shape `(batch_size, sequence_length)`.
966
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
967
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
968
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
969
+ model's internal embedding lookup matrix.
970
+ use_cache (`bool`, *optional*):
971
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
972
+ `past_key_values`).
973
+ output_attentions (`bool`, *optional*):
974
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
975
+ tensors for more detail.
976
+ output_hidden_states (`bool`, *optional*):
977
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
978
+ more detail.
979
+ return_dict (`bool`, *optional*):
980
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
981
+ """
982
+
983
+
984
+ @add_start_docstrings(
985
+ "The bare Phi-3 model outputting raw hidden-states without any specific head on top.",
986
+ PHI3_START_DOCSTRING,
987
+ )
988
+ class Phi3Model(Phi3PreTrainedModel):
989
+ """
990
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Phi3DecoderLayer`]
991
+
992
+ Args:
993
+ config: Phi3Config
994
+ """
995
+
996
+ def __init__(self, config: Phi3Config):
997
+ super().__init__(config)
998
+ self.padding_idx = config.pad_token_id
999
+ self.vocab_size = config.vocab_size
1000
+
1001
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
1002
+ self.embed_dropout = nn.Dropout(config.embd_pdrop)
1003
+ self.layers = nn.ModuleList(
1004
+ [Phi3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
1005
+ )
1006
+ self._attn_implementation = config._attn_implementation
1007
+ self.norm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
1008
+
1009
+ self.gradient_checkpointing = False
1010
+ # Initialize weights and apply final processing
1011
+ self.post_init()
1012
+
1013
+ def get_input_embeddings(self):
1014
+ return self.embed_tokens
1015
+
1016
+ def set_input_embeddings(self, value):
1017
+ self.embed_tokens = value
1018
+
1019
+ @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
1020
+ def forward(
1021
+ self,
1022
+ input_ids: torch.LongTensor = None,
1023
+ attention_mask: Optional[torch.Tensor] = None,
1024
+ position_ids: Optional[torch.LongTensor] = None,
1025
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1026
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1027
+ use_cache: Optional[bool] = None,
1028
+ output_attentions: Optional[bool] = None,
1029
+ output_hidden_states: Optional[bool] = None,
1030
+ return_dict: Optional[bool] = None,
1031
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
1032
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
1033
+ output_hidden_states = (
1034
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
1035
+ )
1036
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
1037
+
1038
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1039
+
1040
+ # retrieve input_ids and inputs_embeds
1041
+ if input_ids is not None and inputs_embeds is not None:
1042
+ raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
1043
+ elif input_ids is not None:
1044
+ batch_size, seq_length = input_ids.shape[:2]
1045
+ elif inputs_embeds is not None:
1046
+ batch_size, seq_length = inputs_embeds.shape[:2]
1047
+ else:
1048
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
1049
+
1050
+ past_key_values_length = 0
1051
+
1052
+ if self.gradient_checkpointing and self.training:
1053
+ if use_cache:
1054
+ logger.warning_once(
1055
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
1056
+ )
1057
+ use_cache = False
1058
+
1059
+ if use_cache:
1060
+ use_legacy_cache = not isinstance(past_key_values, Cache)
1061
+ if use_legacy_cache:
1062
+ past_key_values = DynamicCache.from_legacy_cache(past_key_values)
1063
+ past_key_values_length = past_key_values.get_usable_length(seq_length)
1064
+
1065
+ if position_ids is None:
1066
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
1067
+ position_ids = torch.arange(
1068
+ past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
1069
+ )
1070
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
1071
+ else:
1072
+ position_ids = position_ids.view(-1, seq_length).long()
1073
+
1074
+ if inputs_embeds is None:
1075
+ inputs_embeds = self.embed_tokens(input_ids)
1076
+
1077
+ if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
1078
+ is_padding_right = attention_mask[:, -1].sum().item() != batch_size
1079
+ if is_padding_right:
1080
+ raise ValueError(
1081
+ "You are attempting to perform batched generation with padding_side='right'"
1082
+ " this may lead to unexpected behaviour for Flash Attention version of Phi3. Make sure to "
1083
+ " call `tokenizer.padding_side = 'left'` before tokenizing the input. "
1084
+ )
1085
+
1086
+ if self._attn_implementation == "flash_attention_2":
1087
+ # 2d mask is passed through the layers
1088
+ attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
1089
+ else:
1090
+ # 4d mask is passed through the layers
1091
+ attention_mask = _prepare_4d_causal_attention_mask(
1092
+ attention_mask,
1093
+ (batch_size, seq_length),
1094
+ inputs_embeds,
1095
+ past_key_values_length,
1096
+ sliding_window=self.config.sliding_window,
1097
+ )
1098
+
1099
+ hidden_states = inputs_embeds
1100
+
1101
+ # decoder layers
1102
+ all_hidden_states = () if output_hidden_states else None
1103
+ all_self_attns = () if output_attentions else None
1104
+ next_decoder_cache = None
1105
+
1106
+ for decoder_layer in self.layers:
1107
+ if output_hidden_states:
1108
+ all_hidden_states += (hidden_states,)
1109
+
1110
+ if self.gradient_checkpointing and self.training:
1111
+ layer_outputs = self._gradient_checkpointing_func(
1112
+ decoder_layer.__call__,
1113
+ hidden_states,
1114
+ attention_mask,
1115
+ position_ids,
1116
+ past_key_values,
1117
+ output_attentions,
1118
+ use_cache,
1119
+ )
1120
+ else:
1121
+ layer_outputs = decoder_layer(
1122
+ hidden_states,
1123
+ attention_mask=attention_mask,
1124
+ position_ids=position_ids,
1125
+ past_key_value=past_key_values,
1126
+ output_attentions=output_attentions,
1127
+ use_cache=use_cache,
1128
+ )
1129
+
1130
+ hidden_states = layer_outputs[0]
1131
+
1132
+ if use_cache:
1133
+ next_decoder_cache = layer_outputs[2 if output_attentions else 1]
1134
+
1135
+ if output_attentions:
1136
+ all_self_attns += (layer_outputs[1],)
1137
+
1138
+ hidden_states = self.norm(hidden_states)
1139
+
1140
+ # add hidden states from the last decoder layer
1141
+ if output_hidden_states:
1142
+ all_hidden_states += (hidden_states,)
1143
+
1144
+ next_cache = None
1145
+ if use_cache:
1146
+ next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
1147
+ if not return_dict:
1148
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
1149
+ return BaseModelOutputWithPast(
1150
+ last_hidden_state=hidden_states,
1151
+ past_key_values=next_cache,
1152
+ hidden_states=all_hidden_states,
1153
+ attentions=all_self_attns,
1154
+ )
1155
+
1156
+
1157
+ class Phi3ForCausalLM(Phi3PreTrainedModel):
1158
+ _tied_weights_keys = ["lm_head.weight"]
1159
+
1160
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with Llama->Phi3
1161
+ def __init__(self, config):
1162
+ super().__init__(config)
1163
+ self.model = Phi3Model(config)
1164
+ self.vocab_size = config.vocab_size
1165
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
1166
+
1167
+ # Initialize weights and apply final processing
1168
+ self.post_init()
1169
+
1170
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_input_embeddings
1171
+ def get_input_embeddings(self):
1172
+ return self.model.embed_tokens
1173
+
1174
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_input_embeddings
1175
+ def set_input_embeddings(self, value):
1176
+ self.model.embed_tokens = value
1177
+
1178
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_output_embeddings
1179
+ def get_output_embeddings(self):
1180
+ return self.lm_head
1181
+
1182
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_output_embeddings
1183
+ def set_output_embeddings(self, new_embeddings):
1184
+ self.lm_head = new_embeddings
1185
+
1186
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_decoder
1187
+ def set_decoder(self, decoder):
1188
+ self.model = decoder
1189
+
1190
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_decoder
1191
+ def get_decoder(self):
1192
+ return self.model
1193
+
1194
+ # Ignore copy
1195
+ @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
1196
+ @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
1197
+ def forward(
1198
+ self,
1199
+ input_ids: torch.LongTensor = None,
1200
+ attention_mask: Optional[torch.Tensor] = None,
1201
+ position_ids: Optional[torch.LongTensor] = None,
1202
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1203
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1204
+ labels: Optional[torch.LongTensor] = None,
1205
+ use_cache: Optional[bool] = None,
1206
+ output_attentions: Optional[bool] = None,
1207
+ output_hidden_states: Optional[bool] = None,
1208
+ return_dict: Optional[bool] = None,
1209
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
1210
+ r"""
1211
+ Args:
1212
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1213
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
1214
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
1215
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
1216
+
1217
+ Returns:
1218
+
1219
+ Example:
1220
+
1221
+ ```python
1222
+ >>> from transformers import AutoTokenizer, Phi3ForCausalLM
1223
+
1224
+ >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
1225
+ >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
1226
+
1227
+ >>> prompt = "This is an example script ."
1228
+ >>> inputs = tokenizer(prompt, return_tensors="pt")
1229
+
1230
+ >>> # Generate
1231
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
1232
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
1233
+ 'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
1234
+ ```"""
1235
+
1236
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
1237
+ output_hidden_states = (
1238
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
1239
+ )
1240
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1241
+
1242
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
1243
+ outputs = self.model(
1244
+ input_ids=input_ids,
1245
+ attention_mask=attention_mask,
1246
+ position_ids=position_ids,
1247
+ past_key_values=past_key_values,
1248
+ inputs_embeds=inputs_embeds,
1249
+ use_cache=use_cache,
1250
+ output_attentions=output_attentions,
1251
+ output_hidden_states=output_hidden_states,
1252
+ return_dict=return_dict,
1253
+ )
1254
+
1255
+ hidden_states = outputs[0]
1256
+ logits = self.lm_head(hidden_states)
1257
+ logits = logits.float()
1258
+
1259
+ loss = None
1260
+ if labels is not None:
1261
+ # Shift so that tokens < n predict n
1262
+ shift_logits = logits[..., :-1, :].contiguous()
1263
+ shift_labels = labels[..., 1:].contiguous()
1264
+ # Flatten the tokens
1265
+ loss_fct = CrossEntropyLoss()
1266
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
1267
+ shift_labels = shift_labels.view(-1)
1268
+ # Enable model parallelism
1269
+ shift_labels = shift_labels.to(shift_logits.device)
1270
+ loss = loss_fct(shift_logits, shift_labels)
1271
+
1272
+ if not return_dict:
1273
+ output = (logits,) + outputs[1:]
1274
+ return (loss,) + output if loss is not None else output
1275
+
1276
+ return CausalLMOutputWithPast(
1277
+ loss=loss,
1278
+ logits=logits,
1279
+ past_key_values=outputs.past_key_values,
1280
+ hidden_states=outputs.hidden_states,
1281
+ attentions=outputs.attentions,
1282
+ )
1283
+
1284
+ # Copied from transformers.models.persimmon.modeling_persimmon.PersimmonForCausalLM.prepare_inputs_for_generation
1285
+ def prepare_inputs_for_generation(
1286
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
1287
+ ):
1288
+ # When the first time input length reached long and short factor switching point, enforce re-compute cache
1289
+ # It will cause downside of slower at this single token position, however, better than current failure.
1290
+ if past_key_values and self.config.rope_scaling and input_ids.shape[1] >= self.config.original_max_position_embeddings + 1:
1291
+ past_length = past_key_values.seen_tokens if isinstance(past_key_values, Cache) else past_key_values[0][0].shape[2]
1292
+ if past_length <= self.config.original_max_position_embeddings:
1293
+ past_key_values = None
1294
+
1295
+ if past_key_values is not None:
1296
+ if isinstance(past_key_values, Cache):
1297
+ cache_length = past_key_values.get_seq_length()
1298
+ past_length = past_key_values.seen_tokens
1299
+ max_cache_length = past_key_values.get_max_length()
1300
+ else:
1301
+ cache_length = past_length = past_key_values[0][0].shape[2]
1302
+ max_cache_length = None
1303
+
1304
+ # Keep only the unprocessed tokens:
1305
+ # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
1306
+ # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
1307
+ # input)
1308
+ if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
1309
+ input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
1310
+ # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
1311
+ # input_ids based on the past_length.
1312
+ elif past_length < input_ids.shape[1]:
1313
+ input_ids = input_ids[:, past_length:]
1314
+ # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
1315
+
1316
+ # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
1317
+ if (
1318
+ max_cache_length is not None
1319
+ and attention_mask is not None
1320
+ and cache_length + input_ids.shape[1] > max_cache_length
1321
+ ):
1322
+ attention_mask = attention_mask[:, -max_cache_length:]
1323
+
1324
+ position_ids = kwargs.get("position_ids", None)
1325
+ if attention_mask is not None and position_ids is None:
1326
+ # create position_ids on the fly for batch generation
1327
+ position_ids = attention_mask.long().cumsum(-1) - 1
1328
+ position_ids.masked_fill_(attention_mask == 0, 1)
1329
+ if past_key_values:
1330
+ position_ids = position_ids[:, -input_ids.shape[1] :]
1331
+
1332
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
1333
+ if inputs_embeds is not None and past_key_values is None:
1334
+ model_inputs = {"inputs_embeds": inputs_embeds}
1335
+ else:
1336
+ model_inputs = {"input_ids": input_ids}
1337
+
1338
+ model_inputs.update(
1339
+ {
1340
+ "position_ids": position_ids,
1341
+ "past_key_values": past_key_values,
1342
+ "use_cache": kwargs.get("use_cache"),
1343
+ "attention_mask": attention_mask,
1344
+ }
1345
+ )
1346
+ return model_inputs
1347
+
1348
+ @staticmethod
1349
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM._reorder_cache
1350
+ def _reorder_cache(past_key_values, beam_idx):
1351
+ reordered_past = ()
1352
+ for layer_past in past_key_values:
1353
+ reordered_past += (
1354
+ tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
1355
+ )
1356
+ return reordered_past
1357
+
1358
+
1359
+ @add_start_docstrings(
1360
+ """
1361
+ The [`Phi3Model`] with a sequence classification head on top (linear layer).
1362
+
1363
+ [`Phi3ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
1364
+ (e.g. GPT-2) do.
1365
+
1366
+ Since it does classification on the last token, it requires to know the position of the last token. If a
1367
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
1368
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
1369
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
1370
+ each row of the batch).
1371
+ """,
1372
+ PHI3_START_DOCSTRING,
1373
+ )
1374
+ # Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Phi3, LLAMA->PHI3, self.transformer->self.model, transformer_outputs->model_outputs
1375
+ class Phi3ForSequenceClassification(Phi3PreTrainedModel):
1376
+ def __init__(self, config):
1377
+ super().__init__(config)
1378
+ self.num_labels = config.num_labels
1379
+ self.model = Phi3Model(config)
1380
+ self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
1381
+
1382
+ # Initialize weights and apply final processing
1383
+ self.post_init()
1384
+
1385
+ def get_input_embeddings(self):
1386
+ return self.model.embed_tokens
1387
+
1388
+ def set_input_embeddings(self, value):
1389
+ self.model.embed_tokens = value
1390
+
1391
+ @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
1392
+ def forward(
1393
+ self,
1394
+ input_ids: torch.LongTensor = None,
1395
+ attention_mask: Optional[torch.Tensor] = None,
1396
+ position_ids: Optional[torch.LongTensor] = None,
1397
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1398
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1399
+ labels: Optional[torch.LongTensor] = None,
1400
+ use_cache: Optional[bool] = None,
1401
+ output_attentions: Optional[bool] = None,
1402
+ output_hidden_states: Optional[bool] = None,
1403
+ return_dict: Optional[bool] = None,
1404
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
1405
+ r"""
1406
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1407
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1408
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1409
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1410
+ """
1411
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1412
+
1413
+ model_outputs = self.model(
1414
+ input_ids,
1415
+ attention_mask=attention_mask,
1416
+ position_ids=position_ids,
1417
+ past_key_values=past_key_values,
1418
+ inputs_embeds=inputs_embeds,
1419
+ use_cache=use_cache,
1420
+ output_attentions=output_attentions,
1421
+ output_hidden_states=output_hidden_states,
1422
+ return_dict=return_dict,
1423
+ )
1424
+ hidden_states = model_outputs[0]
1425
+ logits = self.score(hidden_states)
1426
+
1427
+ if input_ids is not None:
1428
+ batch_size = input_ids.shape[0]
1429
+ else:
1430
+ batch_size = inputs_embeds.shape[0]
1431
+
1432
+ if self.config.pad_token_id is None and batch_size != 1:
1433
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
1434
+ if self.config.pad_token_id is None:
1435
+ sequence_lengths = -1
1436
+ else:
1437
+ if input_ids is not None:
1438
+ # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
1439
+ sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
1440
+ sequence_lengths = sequence_lengths % input_ids.shape[-1]
1441
+ sequence_lengths = sequence_lengths.to(logits.device)
1442
+ else:
1443
+ sequence_lengths = -1
1444
+
1445
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
1446
+
1447
+ loss = None
1448
+ if labels is not None:
1449
+ labels = labels.to(logits.device)
1450
+ if self.config.problem_type is None:
1451
+ if self.num_labels == 1:
1452
+ self.config.problem_type = "regression"
1453
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1454
+ self.config.problem_type = "single_label_classification"
1455
+ else:
1456
+ self.config.problem_type = "multi_label_classification"
1457
+
1458
+ if self.config.problem_type == "regression":
1459
+ loss_fct = MSELoss()
1460
+ if self.num_labels == 1:
1461
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1462
+ else:
1463
+ loss = loss_fct(pooled_logits, labels)
1464
+ elif self.config.problem_type == "single_label_classification":
1465
+ loss_fct = CrossEntropyLoss()
1466
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
1467
+ elif self.config.problem_type == "multi_label_classification":
1468
+ loss_fct = BCEWithLogitsLoss()
1469
+ loss = loss_fct(pooled_logits, labels)
1470
+ if not return_dict:
1471
+ output = (pooled_logits,) + model_outputs[1:]
1472
+ return ((loss,) + output) if loss is not None else output
1473
+
1474
+ return SequenceClassifierOutputWithPast(
1475
+ loss=loss,
1476
+ logits=pooled_logits,
1477
+ past_key_values=model_outputs.past_key_values,
1478
+ hidden_states=model_outputs.hidden_states,
1479
+ attentions=model_outputs.attentions,
1480
+ )
1481
+
1482
+
1483
+ @add_start_docstrings(
1484
+ """
1485
+ [`Phi3Model`] with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
1486
+ Named-Entity-Recognition (NER) tasks.
1487
+ """,
1488
+ PHI3_START_DOCSTRING,
1489
+ )
1490
+ # Copied from transformers.models.mpt.modeling_mpt.MptForTokenClassification with Mpt->Phi3,MPT->PHI3,self.transformer->self.model,transformer_outputs->model_outputs
1491
+ class Phi3ForTokenClassification(Phi3PreTrainedModel):
1492
+ def __init__(self, config: Phi3Config):
1493
+ super().__init__(config)
1494
+ self.num_labels = config.num_labels
1495
+
1496
+ self.model = Phi3Model(config)
1497
+ if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
1498
+ classifier_dropout = config.classifier_dropout
1499
+ elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
1500
+ classifier_dropout = config.hidden_dropout
1501
+ else:
1502
+ classifier_dropout = 0.1
1503
+ self.dropout = nn.Dropout(classifier_dropout)
1504
+ self.classifier = nn.Linear(config.hidden_size, config.num_labels)
1505
+
1506
+ # Initialize weights and apply final processing
1507
+ self.post_init()
1508
+
1509
+ @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
1510
+ @add_code_sample_docstrings(
1511
+ checkpoint=_CHECKPOINT_FOR_DOC,
1512
+ output_type=TokenClassifierOutput,
1513
+ config_class=_CONFIG_FOR_DOC,
1514
+ )
1515
+ def forward(
1516
+ self,
1517
+ input_ids: Optional[torch.LongTensor] = None,
1518
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
1519
+ attention_mask: Optional[torch.Tensor] = None,
1520
+ inputs_embeds: Optional[torch.Tensor] = None,
1521
+ labels: Optional[torch.Tensor] = None,
1522
+ use_cache: Optional[bool] = None,
1523
+ output_attentions: Optional[bool] = None,
1524
+ output_hidden_states: Optional[bool] = None,
1525
+ return_dict: Optional[bool] = None,
1526
+ **deprecated_arguments,
1527
+ ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
1528
+ r"""
1529
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1530
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1531
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1532
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1533
+ """
1534
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1535
+
1536
+ model_outputs = self.model(
1537
+ input_ids,
1538
+ past_key_values=past_key_values,
1539
+ attention_mask=attention_mask,
1540
+ inputs_embeds=inputs_embeds,
1541
+ use_cache=use_cache,
1542
+ output_attentions=output_attentions,
1543
+ output_hidden_states=output_hidden_states,
1544
+ return_dict=return_dict,
1545
+ )
1546
+
1547
+ hidden_states = model_outputs[0]
1548
+ hidden_states = self.dropout(hidden_states)
1549
+ logits = self.classifier(hidden_states)
1550
+
1551
+ loss = None
1552
+ if labels is not None:
1553
+ # move labels to correct device to enable model parallelism
1554
+ labels = labels.to(logits.device)
1555
+ batch_size, seq_length = labels.shape
1556
+ loss_fct = CrossEntropyLoss()
1557
+ loss = loss_fct(
1558
+ logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length)
1559
+ )
1560
+
1561
+ if not return_dict:
1562
+ output = (logits,) + model_outputs[2:]
1563
+ return ((loss,) + output) if loss is not None else output
1564
+
1565
+ return TokenClassifierOutput(
1566
+ loss=loss,
1567
+ logits=logits,
1568
+ hidden_states=model_outputs.hidden_states,
1569
+ attentions=model_outputs.attentions,
1570
+ )