Zoyd commited on
Commit
caa40ac
1 Parent(s): dd31418

Upload folder using huggingface_hub

Browse files
README.md ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ license_link: https://huggingface.co/microsoft/Phi-3-medium-4k-instruct/resolve/main/LICENSE
4
+
5
+ language:
6
+ - multilingual
7
+ pipeline_tag: text-generation
8
+ tags:
9
+ - nlp
10
+ - code
11
+ inference:
12
+ parameters:
13
+ temperature: 0.7
14
+ widget:
15
+ - messages:
16
+ - role: user
17
+ content: Can you provide ways to eat combinations of bananas and dragonfruits?
18
+ ---
19
+
20
+ # Phi-3-medium-4k-instruct-abliterated-v3
21
+
22
+
23
+ [My Jupyter "cookbook" to replicate the methodology can be found here, refined library coming soon](https://huggingface.co/failspy/llama-3-70B-Instruct-abliterated/blob/main/ortho_cookbook.ipynb)
24
+
25
+ #### Phi-3-abliterated statement
26
+
27
+ Took me a while to wizard this one up. It’s been a while since I’ve released a Phi-3 model. In the past I accidentally missed an item required in the model release process - hallucination testing.
28
+
29
+ This model has been tested and though it is more likely to hallucinate than the original model in my experience, it is generally as stable as the original.
30
+
31
+ Now that the new Phi-3 models are out, I'm working on completing this abliteration process quickly and then will release the other models as soon as possible. 🏇
32
+
33
+ ## Summary
34
+
35
+ This is [microsoft/Phi-3-medium-4k-instruct](https://huggingface.co/microsoft/Phi-3-medium-4k-instruct) with orthogonalized bfloat16 safetensor weights, generated with a refined methodology based on that which was described in the preview paper/blog post: '[Refusal in LLMs is mediated by a single direction](https://www.alignmentforum.org/posts/jGuXSZgv6qfdhMCuJ/refusal-in-llms-is-mediated-by-a-single-direction)' which I encourage you to read to understand more.
36
+
37
+ [GGUF Quants](https://huggingface.co/failspy/Phi-3-medium-4k-instruct-abliterated-v3-GGUF)
38
+
39
+ ## Hang on, "abliterated"? Orthogonalization? Ablation? What is this?
40
+
41
+ TL;DR: This model has had certain weights manipulated to "inhibit" the model's ability to express refusal. It is not in anyway _guaranteed_ that it won't refuse you, understand your request, it may still lecture you about ethics/safety, etc. It is tuned in all other respects the same as the original 70B instruct model was, just with the strongest refusal directions orthogonalized out.
42
+
43
+ **TL;TL;DR;DR: It's uncensored in the purest form I can manage -- no new or changed behaviour in any other respect from the original model.**
44
+
45
+ As far as "abliterated": it's just a fun play-on-words using the original "ablation" term used in the original paper to refer to removing features, which I made up particularly to differentiate the model from "uncensored" fine-tunes.
46
+ Ablate + obliterated = Abliterated
47
+
48
+ Anyways, orthogonalization/ablation are both aspects to refer to the same thing here, the technique in which the refusal feature was "ablated" from the model was via orthogonalization.
49
+
50
+ ## A little more on the methodology, and why this is interesting
51
+
52
+ To me, ablation (or applying the methodology for the inverse, "augmentation") seems to be good for inducing/removing very specific features that you'd have to spend way too many tokens on encouraging or discouraging in your system prompt.
53
+ Instead, you just apply your system prompt in the ablation script against a blank system prompt on the same dataset and orthogonalize for the desired behaviour in the final model weights.
54
+
55
+ > Why this over fine-tuning?
56
+
57
+ Ablation is much more surgical in nature whilst also being effectively executed with a _lot_ less data than fine-tuning, which I think is its main advantage.
58
+
59
+ As well, and its most valuable aspect is it keeps as much of the original model's knowledge and training intact, whilst removing its tendency to behave in one very specific undesireable manner. (In this case, refusing user requests.)
60
+
61
+ Fine tuning is still exceptionally useful and the go-to for broad behaviour changes; however, you may be able to get close to your desired behaviour with very few samples using the ablation/augmentation techniques.
62
+ It may also be a useful step to add to your model refinement: orthogonalize -> fine-tune or vice-versa.
63
+
64
+ I haven't really gotten around to exploring this model stacked with fine-tuning, I encourage others to give it a shot if they've got the capacity.
65
+
66
+ > Okay, fine, but why V3? There's no V2?
67
+
68
+ Well, I released a V2 of an abliterated model a while back for Meta-Llama-3-8B under Cognitive Computations.
69
+ It ended up being not worth it to try V2 with larger models, I wanted to refine the model before wasting compute cycles on what might not even be a better model.
70
+ I am however quite pleased about this latest methodology, it seems to have induced fewer hallucinations.
71
+ So to show that it's a new fancy methodology from even that of the 8B V2, I decided to do a Microsoft and double up on my version jump because it's *such* an advancement (or so the excuse went, when in actuality it was because too many legacy but actively used Microsoft libraries checked for 'Windows 9' in the OS name to detect Windows 95/98 as one.)
72
+
73
+ ## Quirkiness awareness notice
74
+
75
+ This model may come with interesting quirks, with the methodology being so new. I encourage you to play with the model, and post any quirks you notice in the community tab, as that'll help us further understand what this orthogonalization has in the way of side effects.
76
+
77
+ If you manage to develop further improvements, please share! This is really the most basic way to use ablation, but there are other possibilities that I believe are as-yet unexplored.
78
+
79
+ Additionally, feel free to reach out in any way about this. I'm on the Cognitive Computations Discord, I'm watching the Community tab, reach out! I'd love to see this methodology used in other ways, and so would gladly support whoever whenever I can.
config.json ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "./microsoft/Phi-3-medium-4k-instruct",
3
+ "architectures": [
4
+ "Phi3ForCausalLM"
5
+ ],
6
+ "attention_dropout": 0.0,
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_phi3.Phi3Config",
9
+ "AutoModelForCausalLM": "modeling_phi3.Phi3ForCausalLM"
10
+ },
11
+ "bos_token_id": 1,
12
+ "embd_pdrop": 0.0,
13
+ "eos_token_id": 32000,
14
+ "hidden_act": "silu",
15
+ "hidden_size": 5120,
16
+ "initializer_range": 0.02,
17
+ "intermediate_size": 17920,
18
+ "max_position_embeddings": 4096,
19
+ "model_type": "phi3",
20
+ "num_attention_heads": 40,
21
+ "num_hidden_layers": 40,
22
+ "num_key_value_heads": 10,
23
+ "original_max_position_embeddings": 4096,
24
+ "pad_token_id": 32000,
25
+ "resid_pdrop": 0.0,
26
+ "rms_norm_eps": 1e-05,
27
+ "rope_scaling": null,
28
+ "rope_theta": 10000.0,
29
+ "sliding_window": 2047,
30
+ "tie_word_embeddings": false,
31
+ "torch_dtype": "bfloat16",
32
+ "transformers_version": "4.42.0.dev0",
33
+ "use_cache": true,
34
+ "vocab_size": 32064,
35
+ "quantization_config": {
36
+ "quant_method": "exl2",
37
+ "version": "0.1.1",
38
+ "bits": 2.2,
39
+ "head_bits": 6,
40
+ "calibration": {
41
+ "rows": 100,
42
+ "length": 2048,
43
+ "dataset": "(default)"
44
+ }
45
+ }
46
+ }
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=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_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
+ )
generation_config.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 1,
4
+ "eos_token_id": [
5
+ 32000,
6
+ 32001,
7
+ 32007
8
+ ],
9
+ "pad_token_id": 32000,
10
+ "transformers_version": "4.42.0.dev0"
11
+ }
model.safetensors.index.json ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 27920476160
4
+ },
5
+ "weight_map": {
6
+ "lm_head.weight": "model-00006-of-00006.safetensors",
7
+ "model.embed_tokens.weight": "model-00001-of-00006.safetensors",
8
+ "model.layers.0.input_layernorm.weight": "model-00001-of-00006.safetensors",
9
+ "model.layers.0.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
10
+ "model.layers.0.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
11
+ "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
12
+ "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
13
+ "model.layers.0.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
14
+ "model.layers.1.input_layernorm.weight": "model-00001-of-00006.safetensors",
15
+ "model.layers.1.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
16
+ "model.layers.1.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
17
+ "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
18
+ "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
19
+ "model.layers.1.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
20
+ "model.layers.10.input_layernorm.weight": "model-00002-of-00006.safetensors",
21
+ "model.layers.10.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
22
+ "model.layers.10.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
23
+ "model.layers.10.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
24
+ "model.layers.10.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
25
+ "model.layers.10.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
26
+ "model.layers.11.input_layernorm.weight": "model-00002-of-00006.safetensors",
27
+ "model.layers.11.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
28
+ "model.layers.11.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
29
+ "model.layers.11.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
30
+ "model.layers.11.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
31
+ "model.layers.11.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
32
+ "model.layers.12.input_layernorm.weight": "model-00002-of-00006.safetensors",
33
+ "model.layers.12.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
34
+ "model.layers.12.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
35
+ "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
36
+ "model.layers.12.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
37
+ "model.layers.12.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
38
+ "model.layers.13.input_layernorm.weight": "model-00002-of-00006.safetensors",
39
+ "model.layers.13.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
40
+ "model.layers.13.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
41
+ "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
42
+ "model.layers.13.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
43
+ "model.layers.13.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
44
+ "model.layers.14.input_layernorm.weight": "model-00003-of-00006.safetensors",
45
+ "model.layers.14.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
46
+ "model.layers.14.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
47
+ "model.layers.14.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
48
+ "model.layers.14.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
49
+ "model.layers.14.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
50
+ "model.layers.15.input_layernorm.weight": "model-00003-of-00006.safetensors",
51
+ "model.layers.15.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
52
+ "model.layers.15.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
53
+ "model.layers.15.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
54
+ "model.layers.15.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
55
+ "model.layers.15.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
56
+ "model.layers.16.input_layernorm.weight": "model-00003-of-00006.safetensors",
57
+ "model.layers.16.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
58
+ "model.layers.16.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
59
+ "model.layers.16.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
60
+ "model.layers.16.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
61
+ "model.layers.16.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
62
+ "model.layers.17.input_layernorm.weight": "model-00003-of-00006.safetensors",
63
+ "model.layers.17.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
64
+ "model.layers.17.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
65
+ "model.layers.17.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
66
+ "model.layers.17.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
67
+ "model.layers.17.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
68
+ "model.layers.18.input_layernorm.weight": "model-00003-of-00006.safetensors",
69
+ "model.layers.18.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
70
+ "model.layers.18.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
71
+ "model.layers.18.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
72
+ "model.layers.18.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
73
+ "model.layers.18.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
74
+ "model.layers.19.input_layernorm.weight": "model-00003-of-00006.safetensors",
75
+ "model.layers.19.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
76
+ "model.layers.19.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
77
+ "model.layers.19.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
78
+ "model.layers.19.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
79
+ "model.layers.19.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
80
+ "model.layers.2.input_layernorm.weight": "model-00001-of-00006.safetensors",
81
+ "model.layers.2.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
82
+ "model.layers.2.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
83
+ "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
84
+ "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
85
+ "model.layers.2.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
86
+ "model.layers.20.input_layernorm.weight": "model-00003-of-00006.safetensors",
87
+ "model.layers.20.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
88
+ "model.layers.20.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
89
+ "model.layers.20.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
90
+ "model.layers.20.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
91
+ "model.layers.20.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
92
+ "model.layers.21.input_layernorm.weight": "model-00004-of-00006.safetensors",
93
+ "model.layers.21.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
94
+ "model.layers.21.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
95
+ "model.layers.21.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
96
+ "model.layers.21.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
97
+ "model.layers.21.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
98
+ "model.layers.22.input_layernorm.weight": "model-00004-of-00006.safetensors",
99
+ "model.layers.22.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
100
+ "model.layers.22.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
101
+ "model.layers.22.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
102
+ "model.layers.22.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
103
+ "model.layers.22.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
104
+ "model.layers.23.input_layernorm.weight": "model-00004-of-00006.safetensors",
105
+ "model.layers.23.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
106
+ "model.layers.23.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
107
+ "model.layers.23.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
108
+ "model.layers.23.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
109
+ "model.layers.23.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
110
+ "model.layers.24.input_layernorm.weight": "model-00004-of-00006.safetensors",
111
+ "model.layers.24.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
112
+ "model.layers.24.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
113
+ "model.layers.24.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
114
+ "model.layers.24.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
115
+ "model.layers.24.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
116
+ "model.layers.25.input_layernorm.weight": "model-00004-of-00006.safetensors",
117
+ "model.layers.25.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
118
+ "model.layers.25.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
119
+ "model.layers.25.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
120
+ "model.layers.25.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
121
+ "model.layers.25.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
122
+ "model.layers.26.input_layernorm.weight": "model-00004-of-00006.safetensors",
123
+ "model.layers.26.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
124
+ "model.layers.26.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
125
+ "model.layers.26.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
126
+ "model.layers.26.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
127
+ "model.layers.26.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
128
+ "model.layers.27.input_layernorm.weight": "model-00004-of-00006.safetensors",
129
+ "model.layers.27.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
130
+ "model.layers.27.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
131
+ "model.layers.27.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
132
+ "model.layers.27.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
133
+ "model.layers.27.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
134
+ "model.layers.28.input_layernorm.weight": "model-00005-of-00006.safetensors",
135
+ "model.layers.28.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
136
+ "model.layers.28.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
137
+ "model.layers.28.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
138
+ "model.layers.28.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
139
+ "model.layers.28.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
140
+ "model.layers.29.input_layernorm.weight": "model-00005-of-00006.safetensors",
141
+ "model.layers.29.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
142
+ "model.layers.29.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
143
+ "model.layers.29.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
144
+ "model.layers.29.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
145
+ "model.layers.29.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
146
+ "model.layers.3.input_layernorm.weight": "model-00001-of-00006.safetensors",
147
+ "model.layers.3.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
148
+ "model.layers.3.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
149
+ "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
150
+ "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
151
+ "model.layers.3.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
152
+ "model.layers.30.input_layernorm.weight": "model-00005-of-00006.safetensors",
153
+ "model.layers.30.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
154
+ "model.layers.30.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
155
+ "model.layers.30.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
156
+ "model.layers.30.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
157
+ "model.layers.30.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
158
+ "model.layers.31.input_layernorm.weight": "model-00005-of-00006.safetensors",
159
+ "model.layers.31.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
160
+ "model.layers.31.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
161
+ "model.layers.31.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
162
+ "model.layers.31.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
163
+ "model.layers.31.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
164
+ "model.layers.32.input_layernorm.weight": "model-00005-of-00006.safetensors",
165
+ "model.layers.32.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
166
+ "model.layers.32.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
167
+ "model.layers.32.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
168
+ "model.layers.32.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
169
+ "model.layers.32.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
170
+ "model.layers.33.input_layernorm.weight": "model-00005-of-00006.safetensors",
171
+ "model.layers.33.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
172
+ "model.layers.33.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
173
+ "model.layers.33.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
174
+ "model.layers.33.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
175
+ "model.layers.33.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
176
+ "model.layers.34.input_layernorm.weight": "model-00005-of-00006.safetensors",
177
+ "model.layers.34.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
178
+ "model.layers.34.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
179
+ "model.layers.34.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
180
+ "model.layers.34.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
181
+ "model.layers.34.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
182
+ "model.layers.35.input_layernorm.weight": "model-00006-of-00006.safetensors",
183
+ "model.layers.35.mlp.down_proj.weight": "model-00006-of-00006.safetensors",
184
+ "model.layers.35.mlp.gate_up_proj.weight": "model-00006-of-00006.safetensors",
185
+ "model.layers.35.post_attention_layernorm.weight": "model-00006-of-00006.safetensors",
186
+ "model.layers.35.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
187
+ "model.layers.35.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
188
+ "model.layers.36.input_layernorm.weight": "model-00006-of-00006.safetensors",
189
+ "model.layers.36.mlp.down_proj.weight": "model-00006-of-00006.safetensors",
190
+ "model.layers.36.mlp.gate_up_proj.weight": "model-00006-of-00006.safetensors",
191
+ "model.layers.36.post_attention_layernorm.weight": "model-00006-of-00006.safetensors",
192
+ "model.layers.36.self_attn.o_proj.weight": "model-00006-of-00006.safetensors",
193
+ "model.layers.36.self_attn.qkv_proj.weight": "model-00006-of-00006.safetensors",
194
+ "model.layers.37.input_layernorm.weight": "model-00006-of-00006.safetensors",
195
+ "model.layers.37.mlp.down_proj.weight": "model-00006-of-00006.safetensors",
196
+ "model.layers.37.mlp.gate_up_proj.weight": "model-00006-of-00006.safetensors",
197
+ "model.layers.37.post_attention_layernorm.weight": "model-00006-of-00006.safetensors",
198
+ "model.layers.37.self_attn.o_proj.weight": "model-00006-of-00006.safetensors",
199
+ "model.layers.37.self_attn.qkv_proj.weight": "model-00006-of-00006.safetensors",
200
+ "model.layers.38.input_layernorm.weight": "model-00006-of-00006.safetensors",
201
+ "model.layers.38.mlp.down_proj.weight": "model-00006-of-00006.safetensors",
202
+ "model.layers.38.mlp.gate_up_proj.weight": "model-00006-of-00006.safetensors",
203
+ "model.layers.38.post_attention_layernorm.weight": "model-00006-of-00006.safetensors",
204
+ "model.layers.38.self_attn.o_proj.weight": "model-00006-of-00006.safetensors",
205
+ "model.layers.38.self_attn.qkv_proj.weight": "model-00006-of-00006.safetensors",
206
+ "model.layers.39.input_layernorm.weight": "model-00006-of-00006.safetensors",
207
+ "model.layers.39.mlp.down_proj.weight": "model-00006-of-00006.safetensors",
208
+ "model.layers.39.mlp.gate_up_proj.weight": "model-00006-of-00006.safetensors",
209
+ "model.layers.39.post_attention_layernorm.weight": "model-00006-of-00006.safetensors",
210
+ "model.layers.39.self_attn.o_proj.weight": "model-00006-of-00006.safetensors",
211
+ "model.layers.39.self_attn.qkv_proj.weight": "model-00006-of-00006.safetensors",
212
+ "model.layers.4.input_layernorm.weight": "model-00001-of-00006.safetensors",
213
+ "model.layers.4.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
214
+ "model.layers.4.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
215
+ "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
216
+ "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
217
+ "model.layers.4.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
218
+ "model.layers.5.input_layernorm.weight": "model-00001-of-00006.safetensors",
219
+ "model.layers.5.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
220
+ "model.layers.5.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
221
+ "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
222
+ "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
223
+ "model.layers.5.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
224
+ "model.layers.6.input_layernorm.weight": "model-00002-of-00006.safetensors",
225
+ "model.layers.6.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
226
+ "model.layers.6.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
227
+ "model.layers.6.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
228
+ "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
229
+ "model.layers.6.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
230
+ "model.layers.7.input_layernorm.weight": "model-00002-of-00006.safetensors",
231
+ "model.layers.7.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
232
+ "model.layers.7.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
233
+ "model.layers.7.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
234
+ "model.layers.7.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
235
+ "model.layers.7.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
236
+ "model.layers.8.input_layernorm.weight": "model-00002-of-00006.safetensors",
237
+ "model.layers.8.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
238
+ "model.layers.8.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
239
+ "model.layers.8.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
240
+ "model.layers.8.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
241
+ "model.layers.8.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
242
+ "model.layers.9.input_layernorm.weight": "model-00002-of-00006.safetensors",
243
+ "model.layers.9.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
244
+ "model.layers.9.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
245
+ "model.layers.9.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
246
+ "model.layers.9.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
247
+ "model.layers.9.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
248
+ "model.norm.weight": "model-00006-of-00006.safetensors"
249
+ }
250
+ }
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-attenton` 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
+ )
output.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dd201977409e552381012aa35d4eb608fe9ae04863cf95161a07a1c31e195f9d
3
+ size 4228807128
special_tokens_map.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<s>",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "<|endoftext|>",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": {
17
+ "content": "<|endoftext|>",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ },
23
+ "unk_token": {
24
+ "content": "<unk>",
25
+ "lstrip": false,
26
+ "normalized": false,
27
+ "rstrip": false,
28
+ "single_word": false
29
+ }
30
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
3
+ size 499723
tokenizer_config.json ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": false,
3
+ "add_eos_token": false,
4
+ "added_tokens_decoder": {
5
+ "0": {
6
+ "content": "<unk>",
7
+ "lstrip": false,
8
+ "normalized": false,
9
+ "rstrip": false,
10
+ "single_word": false,
11
+ "special": true
12
+ },
13
+ "1": {
14
+ "content": "<s>",
15
+ "lstrip": false,
16
+ "normalized": false,
17
+ "rstrip": false,
18
+ "single_word": false,
19
+ "special": true
20
+ },
21
+ "2": {
22
+ "content": "</s>",
23
+ "lstrip": false,
24
+ "normalized": false,
25
+ "rstrip": true,
26
+ "single_word": false,
27
+ "special": false
28
+ },
29
+ "32000": {
30
+ "content": "<|endoftext|>",
31
+ "lstrip": false,
32
+ "normalized": false,
33
+ "rstrip": false,
34
+ "single_word": false,
35
+ "special": true
36
+ },
37
+ "32001": {
38
+ "content": "<|assistant|>",
39
+ "lstrip": false,
40
+ "normalized": false,
41
+ "rstrip": true,
42
+ "single_word": false,
43
+ "special": true
44
+ },
45
+ "32002": {
46
+ "content": "<|placeholder1|>",
47
+ "lstrip": false,
48
+ "normalized": false,
49
+ "rstrip": true,
50
+ "single_word": false,
51
+ "special": true
52
+ },
53
+ "32003": {
54
+ "content": "<|placeholder2|>",
55
+ "lstrip": false,
56
+ "normalized": false,
57
+ "rstrip": true,
58
+ "single_word": false,
59
+ "special": true
60
+ },
61
+ "32004": {
62
+ "content": "<|placeholder3|>",
63
+ "lstrip": false,
64
+ "normalized": false,
65
+ "rstrip": true,
66
+ "single_word": false,
67
+ "special": true
68
+ },
69
+ "32005": {
70
+ "content": "<|placeholder4|>",
71
+ "lstrip": false,
72
+ "normalized": false,
73
+ "rstrip": true,
74
+ "single_word": false,
75
+ "special": true
76
+ },
77
+ "32006": {
78
+ "content": "<|system|>",
79
+ "lstrip": false,
80
+ "normalized": false,
81
+ "rstrip": true,
82
+ "single_word": false,
83
+ "special": true
84
+ },
85
+ "32007": {
86
+ "content": "<|end|>",
87
+ "lstrip": false,
88
+ "normalized": false,
89
+ "rstrip": true,
90
+ "single_word": false,
91
+ "special": true
92
+ },
93
+ "32008": {
94
+ "content": "<|placeholder5|>",
95
+ "lstrip": false,
96
+ "normalized": false,
97
+ "rstrip": true,
98
+ "single_word": false,
99
+ "special": true
100
+ },
101
+ "32009": {
102
+ "content": "<|placeholder6|>",
103
+ "lstrip": false,
104
+ "normalized": false,
105
+ "rstrip": true,
106
+ "single_word": false,
107
+ "special": true
108
+ },
109
+ "32010": {
110
+ "content": "<|user|>",
111
+ "lstrip": false,
112
+ "normalized": false,
113
+ "rstrip": true,
114
+ "single_word": false,
115
+ "special": true
116
+ }
117
+ },
118
+ "bos_token": "<s>",
119
+ "chat_template": "{% for message in messages %}{% if (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif (message['role'] == 'assistant') %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}",
120
+ "clean_up_tokenization_spaces": false,
121
+ "eos_token": "<|endoftext|>",
122
+ "legacy": false,
123
+ "model_max_length": 4096,
124
+ "pad_token": "<|endoftext|>",
125
+ "padding_side": "left",
126
+ "sp_model_kwargs": {},
127
+ "tokenizer_class": "LlamaTokenizer",
128
+ "unk_token": "<unk>",
129
+ "use_default_system_prompt": false
130
+ }