README.md DELETED
@@ -1,242 +0,0 @@
1
- ---
2
- language:
3
- - en
4
- inference: false
5
- tags:
6
- - pytorch
7
- - causal-lm
8
- - Cerebras
9
- - BTLM
10
- datasets:
11
- - cerebras/SlimPajama-627B
12
- pipeline_tag: text-generation
13
- license: apache-2.0
14
- ---
15
-
16
- # BTLM-3B-8k-base
17
-
18
- [Bittensor Language Model (BTLM-3B-8k-base)](https://www.cerebras.net/blog/btlm-3b-8k-7b-performance-in-a-3-billion-parameter-model/) is a 3 billion parameter language model with an 8k context length trained on 627B tokens of [SlimPajama](https://huggingface.co/datasets/cerebras/SlimPajama-627B). BTLM-3B-8k-base sets a new standard for 3B parameter models, outperforming models trained on hundreds of billions more tokens and achieving comparable performance to open 7B parameter models. BTLM-3B-8k-base can also be quantized to 4-bit to fit in devices with as little as 3GB of memory. The model is made available with an Apache 2.0 license for commercial use.
19
-
20
- BTLM was trained by [Cerebras](https://www.cerebras.net/) in partnership with [Opentensor](https://opentensor.ai/) on the newly unveiled [Condor Galaxy 1 (CG-1) supercomputer](https://www.cerebras.net/blog/introducing-condor-galaxy-1-a-4-exaflop-supercomputer-for-generative-ai/), the first public deliverable of the G42-Cerebras strategic partnership.
21
-
22
- BTLM-3B-8k was trained with a similar architecture to [CerebrasGPT](https://arxiv.org/abs/2304.03208) with the addition of [SwiGLU](https://arxiv.org/abs/2002.05202) nonlinearity, [ALiBi](https://arxiv.org/abs/2108.12409) position embeddings, and [maximal update parameterization (muP)](https://arxiv.org/abs/2203.03466). The model was trained for 1 epoch of SlimPajama-627B. 75% of training was performed with 2k sequence length. The final 25% of training was performed at 8k sequence length to enable long sequence applications
23
-
24
- Read [our paper](https://arxiv.org/abs/2309.11568) for more details!
25
-
26
- ## BTLM-3B-8k Highlights
27
-
28
- BTLM-3B-8k-base:
29
- - **Licensed for commercial use** (Apache 2.0).
30
- - **[State of the art 3B parameter model](#performance-vs-3b-models)**.
31
- - **Provides 7B model performance in a 3B model** via performance enhancements from [ALiBi](https://arxiv.org/abs/2108.12409), [SwiGLU](https://arxiv.org/abs/2002.05202), [maximal update parameterization (muP)](https://arxiv.org/abs/2203.03466) and the the extensively deduplicated and cleaned [SlimPajama-627B dataset](https://huggingface.co/datasets/cerebras/SlimPajama-627B).
32
- - **[Fits in devices with as little as 3GB of memory](#memory-requirements) when quantized to 4-bit**.
33
- - **One of few 3B models that supports 8k sequence length** thanks to ALiBi.
34
- - **Requires 71% fewer training FLOPs, has 58% smaller memory footprint** for inference than comparable 7B models.
35
-
36
- ## Usage
37
- *Note: Transformers does not support muP for all models, so BTLM-3B-8k-base requires a custom model class. This causes a situation where users must either (1) enable `trust_remote_code=True` when loading the model or (2) acknowledge the warning about code execution upon loading the model.*
38
-
39
- #### With generate():
40
- ```python
41
- from transformers import AutoTokenizer, AutoModelForCausalLM
42
-
43
- # Load the tokenizer and model
44
- tokenizer = AutoTokenizer.from_pretrained("cerebras/btlm-3b-8k-base")
45
- model = AutoModelForCausalLM.from_pretrained("cerebras/btlm-3b-8k-base", trust_remote_code=True, torch_dtype="auto")
46
-
47
- # Set the prompt for generating text
48
- prompt = "Albert Einstein was known for "
49
-
50
- # Tokenize the prompt and convert to PyTorch tensors
51
- inputs = tokenizer(prompt, return_tensors="pt")
52
-
53
- # Generate text using the model
54
- outputs = model.generate(
55
- **inputs,
56
- num_beams=5,
57
- max_new_tokens=50,
58
- early_stopping=True,
59
- no_repeat_ngram_size=2
60
- )
61
-
62
- # Convert the generated token IDs back to text
63
- generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)
64
-
65
- # Print the generated text
66
- print(generated_text[0])
67
- ```
68
-
69
- #### With pipeline:
70
- ```python
71
- from transformers import AutoTokenizer, AutoModelForCausalLM
72
- from transformers import pipeline
73
-
74
- # Load the tokenizer and model
75
- tokenizer = AutoTokenizer.from_pretrained("cerebras/btlm-3b-8k-base")
76
- model = AutoModelForCausalLM.from_pretrained("cerebras/btlm-3b-8k-base", trust_remote_code=True, torch_dtype="auto")
77
-
78
- # Set the prompt for text generation
79
- prompt = """Isaac Newton was a """
80
-
81
- # Create a text generation pipeline
82
- pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
83
-
84
- # Generate text using the pipeline
85
- generated_text = pipe(
86
- prompt,
87
- max_length=50,
88
- do_sample=False,
89
- no_repeat_ngram_size=2)[0]
90
-
91
- # Print the generated text
92
- print(generated_text['generated_text'])
93
- ```
94
-
95
- ## Evaluations and Comparisons to Other Models
96
-
97
- ### Memory Requirements
98
- ![figure_1_image](./figure_1_memory_footprint.png)
99
- Figure 1. Memory requirements of different model sizes and quantization schemes
100
-
101
- ### Quality, Training Cost, Memory Footprint, Inference Speed
102
- ![figure_2_image](./figure_2_half_the_size_twice_the_speed.png)
103
- Figure 2: Comparisons of quality, memory footprint & inference cost between BTLM-3B-8K and 7B model families.
104
-
105
- ### Performance vs 3B models
106
- ![table_1_image](./table_1_downstream_performance_3b.png)
107
- Table 1: Performance at 3B model size. Detailed down-stream tasks comparisons. MMLU task performance is reported using 5-shot, other tasks are 0-shot.
108
-
109
- ![figure_3_image](./figure_3_performance_vs_3b_models.png)
110
- Figure 3: Performance at 3B model size
111
-
112
- ### Performance vs 7B models
113
- ![table_2_image](./table_2_downstream_performance_7b.png)
114
- Table 2: Performance at 7B model size. Detailed down-stream tasks comparisons. MMLU task performance is reported using 5-shot, everything else is 0-shot.
115
-
116
- ![figure_4_image](./figure_4_performance_vs_7b_models.jpg)
117
- Figure 4: Performance at 7B model size
118
-
119
- ## Long Sequence Lengths
120
- To enable long sequence applications, we use ALiBi position embeddings and trained on 470B tokens at the context length of 2,048 followed by 157B of tokens trained at 8,192 context length. To assess BTLM’s long sequence capability, we evaluate it on SlimPajama test set with 32,768 context length and plot loss at each token position. Although ALiBi allows extrapolation in theory, 2,048 context length training alone does not extrapolate well in practice. Thankfully variable sequence length training allows for substantially improved extrapolation. BTLM-3B extrapolates well up to 10k context length but the performance degrades slightly beyond this.
121
-
122
- ![figure_5_image](./figure_5_xentropy_with_sequence_lengths.svg)
123
- Figure 5: BTLM-3B model's cross-entropy evaluation on the SlimPajama’s test set. Inference performed on the extrapolated sequence length of 32,768 tokens.
124
-
125
- ## Model Details
126
- - Developed by: [Cerebras Systems](https://www.cerebras.net/) and [Opentensor](https://opentensor.ai/) with generous support from [G42 Cloud](https://www.g42cloud.com/) and [IIAI](https://www.inceptioniai.org/en/)
127
- - License: Apache 2.0
128
- - Model type: Decoder-only Language Model
129
- - Architecture: GPT-2 style architecture with SwiGLU, ALiBi, and muP
130
- - Data set: SlimPajama-627B
131
- - Tokenizer: Byte Pair Encoding
132
- - Vocabulary Size: 50257
133
- - Sequence Length: 8192
134
- - Optimizer: AdamW
135
- - Positional Encoding: ALiBi
136
- - Language: English
137
- - Learn more: [BTLM-3B-8k blog](https://www.cerebras.net/blog/btlm-3b-8k-7b-performance-in-a-3-billion-parameter-model/)
138
- - Paper: [BTLM-3B-8K: 7B Parameter Performance in a 3B Parameter Model](https://arxiv.org/abs/2309.11568)
139
-
140
- ## To continue training with PyTorch and Maximal Update Parameterization
141
-
142
- ```python
143
- from transformers import AutoModelForCausalLM
144
- import torch
145
-
146
- model = AutoModelForCausalLM.from_pretrained("cerebras/btlm-3b-8k-base", trust_remote_code=True)
147
-
148
- # Get the parameter groups for the muP optimizer
149
- param_groups = model.get_mup_param_groups(lr=1e-3, weight_decay=0.1)
150
-
151
- # Set up the optimizer using AdamW with muP parameters
152
- optimizer = torch.optim.AdamW(
153
- param_groups,
154
- betas=(0.9, 0.95),
155
- eps=1e-8
156
- )
157
- ```
158
-
159
- Ensure the following muP parameters are passed in your config, otherwise your model will default to standard parameterization
160
- - `mup_width_scale: <float>`
161
- - `mup_embeddings_scale: <float>`
162
- - `mup_output_alpha: <float>`
163
- - `mup_scale_qk_dot_by_d: true`
164
-
165
- ## To extend the context length with Position Interpolation
166
-
167
- ### During inference (without fine-tuning):
168
- It's possible to extend the context length to 2x the training context length without degradation in performance using dynamic linear scaling. Dynamic linear scaling adjusts the slopes of ALiBi with a factor of `input_seq_len/train_seq_len` when `input_seq_len` is larger than `train_seq_len`. Check the details in our paper [Position Interpolation Improves ALiBi Extrapolation](https://arxiv.org/abs/2310.13017). To enable dynamic linear scaling, update `config.json` as follows:
169
- ```json
170
- # update `n_positions` with the maximum context length will be
171
- # encountered during inference (e.g. 16384 tokens)
172
- "n_positions": 16384,
173
-
174
- # specify `train_seq_len` in `alibi_scaling` parameter
175
- "alibi_scaling": {
176
- "type": "linear",
177
- "train_seq_len": 8192
178
- }
179
- ```
180
-
181
- ### Using fine-tuning + position interpolation:
182
- Performing fine-tuning with position interpolation can help achieve greater extrapolation lengths. The scaling factor should be fixed to `finetuning_seq_len/train_seq_len`. To enable fixed linear scaling, update `config.json` as follows:
183
- ```json
184
- # update `n_positions` with the fine-tuning context length (e.g. 32768 tokens)
185
- "n_positions": 32768,
186
-
187
- # specify the scaling `factor` in `alibi_scaling` parameter
188
- "alibi_scaling": {
189
- "type": "linear",
190
- "factor": 4.0
191
- }
192
- ```
193
-
194
- ## Uses and Limitations
195
-
196
- ### Intended Use
197
- The primary intended use is to further research into large language models. BTLM-3B-8k-base can be used as a foundation model for NLP, applications, ethics, and alignment research. We release these models with a fully permissive Apache license for the community to use freely.
198
-
199
- You may fine-tune and adapt BTLM-3B-8k-base model via either Cerebras [Model Studio](https://www.cerebras.net/product-cloud/) or third-party libraries. Further safety-related testing and mitigations should be applied before using the BTLM-3B-8k-base in production downstream applications.
200
-
201
- ### Out of Scope Use
202
- BTLM-3B-8k-base was trained on SlimPajama, with primarily English language, and is not recommended for machine translation tasks. BTLM-3B-8k-base has not been tuned for instruction-following or chat-based use cases.
203
-
204
- ### Generation Samples
205
- ```
206
- Prompt: What is a manometer?"
207
- Generation: "A manometric device is an instrument that measures the pressure of a fluid."
208
- ```
209
-
210
- ```
211
- Prompt: "What Robert Oppenheimer is famous for?"
212
- Generation: “Robert Opperman was a physicist who was the director of the Manhattan Project, which was responsible for the development of nuclear weapons.”
213
- ```
214
-
215
- ```
216
- Prompt: "What was Albert Einstein known for?"
217
- Generation: “Albert Einstein was a German-born theoretical physicist who developed the theory of relativity, one of the two pillars of modern physics (alongside quantum mechanics).”
218
- ```
219
-
220
- ```
221
- Prompt: “Who is Isaac Newton?”
222
- Generation: “Isaac Newton was an English physicist, mathematician, astronomer, alchemist, and theologian. He is widely recognized as one of the most influential scientists of all time and as a key figure in the scientific revolution.”
223
- ```
224
-
225
- ```
226
- Prompt: "Where does the blood get pumped to after it leaves the right ventricle?"
227
- Generation: “The right atrium receives the deoxygenated blood from the pulmonary artery. The right side of the heart is the only side that receives de-oxygened...”
228
- ```
229
-
230
- ```
231
- Prompt: "Peter lost his wallet after attending a local concert. To help him, Bob gave David 10 dollars. David then passed it on to Peter. Now what is the total amount of money Peter has?"
232
- Generation: “A. $10”
233
- ```
234
-
235
- ## Risk, Bias, Ethical Considerations
236
- - **Human life:** The outputs from this model may or may not align with human values. The risk needs to be thoroughly investigated before deploying this model in a production environment where it can directly impact human life.
237
- - **Risks and harms:** There may be distributional bias in the [RedPajama dataset](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T) that can manifest in various forms in the downstream model deployment. There are other risks associated with large language models such as amplifying stereotypes, memorizing training data, or revealing private or secure information.
238
-
239
- ## Acknowledgements
240
- We are thankful to all Cerebras engineers that made this work possible.
241
-
242
- We would like to acknowledge the generous support of G42 Cloud and the Inception Institute of Artificial Intelligence for providing compute time on Condor Galaxy 1.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
config.json DELETED
@@ -1,40 +0,0 @@
1
- {
2
- "_name_or_path": "cerebras/btlm-3b-8k-base",
3
- "activation_function": "swiglu",
4
- "architectures": [
5
- "BTLMLMHeadModel"
6
- ],
7
- "attn_pdrop": 0.0,
8
- "auto_map": {
9
- "AutoConfig": "configuration_btlm.BTLMConfig",
10
- "AutoModel": "modeling_btlm.BTLMModel",
11
- "AutoModelForCausalLM": "modeling_btlm.BTLMLMHeadModel",
12
- "AutoModelForQuestionAnswering": "modeling_btlm.BTLMForQuestionAnswering",
13
- "AutoModelForSequenceClassification": "modeling_btlm.BTLMForSequenceClassification",
14
- "AutoModelForTokenClassification": "modeling_btlm.BTLMForTokenClassification"
15
- },
16
- "bos_token_id": 50256,
17
- "embd_pdrop": 0.0,
18
- "mup_embeddings_scale": 14.6,
19
- "eos_token_id": 50256,
20
- "initializer_range": 0.073,
21
- "layer_norm_epsilon": 1e-05,
22
- "model_type": "btlm",
23
- "n_embd": 2560,
24
- "n_head": 32,
25
- "n_inner": 6826,
26
- "n_layer": 32,
27
- "n_positions": 8192,
28
- "mup_output_alpha": 2.2200000000000003,
29
- "position_embedding_type": "alibi",
30
- "reorder_and_upcast_attn": false,
31
- "resid_pdrop": 0.0,
32
- "scale_attn_by_inverse_layer_idx": false,
33
- "scale_attn_weights": true,
34
- "mup_scale_qk_dot_by_d": true,
35
- "torch_dtype": "bfloat16",
36
- "transformers_version": "4.30.0",
37
- "use_cache": true,
38
- "vocab_size": 50257,
39
- "mup_width_scale": 0.1
40
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
configuration_btlm.py DELETED
@@ -1,201 +0,0 @@
1
- # coding=utf-8
2
- # Copyright 2023 The OpenAI Team Authors and HuggingFace Inc. team.
3
- # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
4
- # Copyright 2023 Cerebras Systems.
5
- #
6
- # Licensed under the Apache License, Version 2.0 (the "License");
7
- # you may not use this file except in compliance with the License.
8
- # You may obtain a copy of the License at
9
- #
10
- # http://www.apache.org/licenses/LICENSE-2.0
11
- #
12
- # Unless required by applicable law or agreed to in writing, software
13
- # distributed under the License is distributed on an "AS IS" BASIS,
14
- # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15
- # See the License for the specific language governing permissions and
16
- # limitations under the License.
17
- """ BTLM configuration"""
18
-
19
- from transformers.configuration_utils import PretrainedConfig
20
- from transformers.utils import logging
21
-
22
-
23
- logger = logging.get_logger(__name__)
24
-
25
- BTLM_PRETRAINED_CONFIG_ARCHIVE_MAP = {
26
- "cerebras/btlm-3b-8k-base": "https://huggingface.co/cerebras/btlm-3b-8k-base/resolve/main/config.json",
27
- }
28
-
29
-
30
- class BTLMConfig(PretrainedConfig):
31
- """
32
- This is the configuration class to store the configuration of a [`BTLMModel`]. It is used to instantiate a BTLM
33
- model according to the specified arguments, defining the model architecture.
34
-
35
- Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
36
- documentation from [`PretrainedConfig`] for more information.
37
-
38
-
39
- Args:
40
- vocab_size (`int`, *optional*, defaults to 50257):
41
- Vocabulary size of the BTLM model. Defines the number of different tokens that can be represented by the
42
- `inputs_ids` passed when calling [`BTLMModel`].
43
- n_positions (`int`, *optional*, defaults to 1024):
44
- The maximum sequence length that this model might ever be used with. Typically set this to something large
45
- just in case (e.g., 512 or 1024 or 2048).
46
- n_embd (`int`, *optional*, defaults to 768):
47
- Dimensionality of the embeddings and hidden states.
48
- n_layer (`int`, *optional*, defaults to 12):
49
- Number of hidden layers in the Transformer encoder.
50
- n_head (`int`, *optional*, defaults to 12):
51
- Number of attention heads for each attention layer in the Transformer encoder.
52
- n_inner (`int`, *optional*, defaults to None):
53
- Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd
54
- activation_function (`str`, *optional*, defaults to `"gelu"`):
55
- Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new", "swiglu"]`.
56
- resid_pdrop (`float`, *optional*, defaults to 0.1):
57
- The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
58
- embd_pdrop (`float`, *optional*, defaults to 0.1):
59
- The dropout ratio for the embeddings.
60
- attn_pdrop (`float`, *optional*, defaults to 0.1):
61
- The dropout ratio for the attention.
62
- layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
63
- The epsilon to use in the layer normalization layers.
64
- initializer_range (`float`, *optional*, defaults to 0.02):
65
- The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
66
- scale_attn_weights (`bool`, *optional*, defaults to `True`):
67
- Scale attention weights by dividing by sqrt(hidden_size)..
68
- use_cache (`bool`, *optional*, defaults to `True`):
69
- Whether or not the model should return the last key/values attentions (not used by all models).
70
- scale_attn_by_inverse_layer_idx (`bool`, *optional*, defaults to `False`):
71
- Whether to additionally scale attention weights by `1 / layer_idx + 1`.
72
- reorder_and_upcast_attn (`bool`, *optional*, defaults to `False`):
73
- Whether to scale keys (K) prior to computing attention (dot-product) and upcast attention
74
- dot-product/softmax to float() when training with mixed precision.
75
- position_embedding_type (`str`, *optional*, defaults to `"learned"`):
76
- Positional embedding can be either `"alibi"` or `"learned"`.
77
- mup_width_scale (`float`, *optional*, defaults to 1.0):
78
- muP parameter to scale learning rate and initializers. Calculated as (`d_model,0 / d_model`), where
79
- `d_model` is the model's width and `d_model,0` is the proxy model's width.
80
- mup_embeddings_scale (`float`, *optional*, defaults to 1.0):
81
- muP parameter to scale token and position embeddings.
82
- mup_output_alpha (`float`, *optional*, defaults to 1.0):
83
- muP parameter to scale output logits (`output_logits_scale = mup_output_alpha * mup_width_scale`).
84
- mup_scale_qk_dot_by_d (`bool`, *optional*, defaults to `False`):
85
- Scale attention weights by dividing by hidden_size instead of sqrt(hidden_size). Need to set
86
- scale_attn_weights to `True` as well.
87
- alibi_scaling (`Dict`, *optional*):
88
- Dictionary containing the scaling configuration for ALiBi embeddings. Currently only supports linear
89
- scaling strategy. Can specify either the scaling `factor` (must be a float greater than 1) for fixed scaling
90
- or `train_seq_len` for dynamic scaling on input samples with sequence length > `train_seq_len`. The expected
91
- formats are `{"type": strategy name, "factor": scaling factor}` or
92
- `{"type": strategy name, "train_seq_len": training sequence length}`.
93
-
94
- Example:
95
-
96
- ```python
97
- >>> from transformers import BTLMConfig, BTLMModel
98
-
99
- >>> # Initializing a BTLM configuration
100
- >>> configuration = BTLMConfig()
101
-
102
- >>> # Initializing a model (with random weights) from the configuration
103
- >>> model = BTLMModel(configuration)
104
-
105
- >>> # Accessing the model configuration
106
- >>> configuration = model.config
107
- ```"""
108
-
109
- model_type = "btlm"
110
- keys_to_ignore_at_inference = ["past_key_values"]
111
- attribute_map = {
112
- "hidden_size": "n_embd",
113
- "max_position_embeddings": "n_positions",
114
- "num_attention_heads": "n_head",
115
- "num_hidden_layers": "n_layer",
116
- }
117
-
118
- def __init__(
119
- self,
120
- vocab_size=50257,
121
- n_positions=1024,
122
- n_embd=768,
123
- n_layer=12,
124
- n_head=12,
125
- n_inner=None,
126
- activation_function="gelu_new",
127
- resid_pdrop=0.1,
128
- embd_pdrop=0.1,
129
- attn_pdrop=0.1,
130
- layer_norm_epsilon=1e-5,
131
- initializer_range=0.02,
132
- scale_attn_weights=True,
133
- use_cache=True,
134
- bos_token_id=50256,
135
- eos_token_id=50256,
136
- scale_attn_by_inverse_layer_idx=False,
137
- reorder_and_upcast_attn=False,
138
- position_embedding_type="learned",
139
- mup_width_scale=1.0,
140
- mup_embeddings_scale=1.0,
141
- mup_output_alpha=1.0,
142
- mup_scale_qk_dot_by_d=False,
143
- alibi_scaling=None,
144
- **kwargs,
145
- ):
146
- self.vocab_size = vocab_size
147
- self.n_positions = n_positions
148
- self.n_embd = n_embd
149
- self.n_layer = n_layer
150
- self.n_head = n_head
151
- self.n_inner = n_inner
152
- self.activation_function = activation_function
153
- self.resid_pdrop = resid_pdrop
154
- self.embd_pdrop = embd_pdrop
155
- self.attn_pdrop = attn_pdrop
156
- self.layer_norm_epsilon = layer_norm_epsilon
157
- self.initializer_range = initializer_range
158
- self.scale_attn_weights = scale_attn_weights
159
- self.use_cache = use_cache
160
- self.scale_attn_by_inverse_layer_idx = scale_attn_by_inverse_layer_idx
161
- self.reorder_and_upcast_attn = reorder_and_upcast_attn
162
-
163
- self.bos_token_id = bos_token_id
164
- self.eos_token_id = eos_token_id
165
-
166
- self.position_embedding_type = position_embedding_type
167
- self.mup_width_scale = mup_width_scale
168
- self.mup_embeddings_scale = mup_embeddings_scale
169
- self.mup_output_alpha = mup_output_alpha
170
- self.mup_scale_qk_dot_by_d = mup_scale_qk_dot_by_d
171
-
172
- self.alibi_scaling = alibi_scaling
173
- self._alibi_scaling_validation()
174
-
175
- super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
176
-
177
- def _alibi_scaling_validation(self):
178
- """
179
- Validate the `alibi_scaling` configuration.
180
- """
181
- if self.alibi_scaling is None:
182
- return
183
-
184
- if not isinstance(self.alibi_scaling, dict) or len(self.alibi_scaling) != 2:
185
- raise ValueError(
186
- "`alibi_scaling` must be a dictionary with two fields, `type` and `factor` or `type` and `train_seq_len`, "
187
- f"got {self.alibi_scaling}"
188
- )
189
- alibi_scaling_type = self.alibi_scaling.get("type", None)
190
- alibi_scaling_factor = self.alibi_scaling.get("factor", None)
191
- alibi_dynamic_scaling = self.alibi_scaling.get("train_seq_len", None)
192
- if alibi_scaling_type is None or alibi_scaling_type != "linear":
193
- raise ValueError(
194
- f"`alibi_scaling`'s type field must be 'linear', got {alibi_scaling_type}"
195
- )
196
- if alibi_scaling_factor is not None:
197
- if not isinstance(alibi_scaling_factor, float) or alibi_scaling_factor <= 1.0:
198
- raise ValueError(f"`alibi_scaling`'s factor field must be a float > 1.0, got {alibi_scaling_factor}")
199
- if alibi_dynamic_scaling is not None:
200
- if not isinstance(alibi_dynamic_scaling, int) or alibi_dynamic_scaling <= 1:
201
- raise ValueError(f"`alibi_scaling`'s `train_seq_len` field must be an integer > 1, got {alibi_dynamic_scaling}")
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
figure_1_memory_footprint.png DELETED
Binary file (301 kB)
 
figure_2_half_the_size_twice_the_speed.png DELETED
Binary file (342 kB)
 
figure_3_performance_vs_3b_models.png DELETED
Binary file (238 kB)
 
figure_4_performance_vs_7b_models.jpg DELETED
Binary file (102 kB)
 
figure_5_xentropy_with_sequence_lengths.png DELETED
Binary file (52.1 kB)
 
figure_5_xentropy_with_sequence_lengths.svg DELETED
generation_config.json DELETED
@@ -1,6 +0,0 @@
1
- {
2
- "_from_model_config": true,
3
- "bos_token_id": 50256,
4
- "eos_token_id": 50256,
5
- "transformers_version": "4.30.0"
6
- }
 
 
 
 
 
 
 
merges.txt DELETED
The diff for this file is too large to render. See raw diff
 
modeling_btlm.py DELETED
@@ -1,1605 +0,0 @@
1
- # coding=utf-8
2
- # Copyright 2023 The OpenAI Team Authors and HuggingFace Inc. team.
3
- # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
4
- # Copyright 2023 Cerebras Systems.
5
- #
6
- # Licensed under the Apache License, Version 2.0 (the "License");
7
- # you may not use this file except in compliance with the License.
8
- # You may obtain a copy of the License at
9
- #
10
- # http://www.apache.org/licenses/LICENSE-2.0
11
- #
12
- # Unless required by applicable law or agreed to in writing, software
13
- # distributed under the License is distributed on an "AS IS" BASIS,
14
- # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15
- # See the License for the specific language governing permissions and
16
- # limitations under the License.
17
- """ PyTorch BTLM model."""
18
-
19
- import math
20
- import os
21
- import warnings
22
- from typing import Optional, Tuple, Union
23
-
24
- import torch
25
- from torch import Tensor, nn
26
- from torch.cuda.amp import autocast
27
- from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
28
-
29
- from transformers.activations import ACT2FN
30
- from transformers.modeling_outputs import (
31
- BaseModelOutputWithPastAndCrossAttentions,
32
- CausalLMOutputWithCrossAttentions,
33
- QuestionAnsweringModelOutput,
34
- SequenceClassifierOutputWithPast,
35
- TokenClassifierOutput,
36
- )
37
- from transformers.modeling_utils import PreTrainedModel
38
- from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
39
- from transformers.utils import (
40
- add_code_sample_docstrings,
41
- add_start_docstrings,
42
- add_start_docstrings_to_model_forward,
43
- logging,
44
- )
45
- from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
46
- from .configuration_btlm import BTLMConfig
47
-
48
-
49
- logger = logging.get_logger(__name__)
50
-
51
- _CHECKPOINT_FOR_DOC = "cerebras/btlm-3b-8k-base"
52
- _CONFIG_FOR_DOC = "BTLMConfig"
53
-
54
- BTLM_PRETRAINED_MODEL_ARCHIVE_LIST = [
55
- "cerebras/btlm-3b-8k-base",
56
- # See all BTLM models at https://huggingface.co/models?filter=btlm
57
- ]
58
-
59
-
60
- class SwiGLUActivation(nn.Module):
61
- def forward(self, x1: Tensor, x2: Tensor) -> Tensor:
62
- return x1 * nn.functional.silu(x2)
63
-
64
-
65
- class AlibiPositionEmbeddingLayer(nn.Module):
66
- def __init__(self, num_heads, alibi_scaling=None):
67
- super(AlibiPositionEmbeddingLayer, self).__init__()
68
-
69
- self.num_heads = num_heads
70
- self.alibi_scaling = alibi_scaling
71
- slopes = torch.tensor(AlibiPositionEmbeddingLayer._get_alibi_slopes(num_heads)).unsqueeze(-1)
72
- self.slopes = nn.parameter.Parameter(slopes, requires_grad=False)
73
-
74
- def forward(
75
- self,
76
- seq_length,
77
- key_length,
78
- cached_qk_len,
79
- ):
80
- context_position = torch.arange(
81
- cached_qk_len, cached_qk_len + seq_length, device=self.slopes.device
82
- )[:, None]
83
- memory_position = torch.arange(
84
- key_length + cached_qk_len, device=self.slopes.device
85
- )[None, :]
86
- relative_position = memory_position - context_position
87
- relative_position = torch.abs(relative_position).unsqueeze(0).expand(self.num_heads, -1, -1)
88
-
89
- if self.alibi_scaling is None:
90
- scale = 1.0
91
- elif self.alibi_scaling.get("factor") is not None:
92
- scale = self.alibi_scaling["factor"]
93
- elif relative_position.shape[-1] > self.alibi_scaling["train_seq_len"]:
94
- scale = relative_position.shape[-1] / self.alibi_scaling["train_seq_len"]
95
- else:
96
- scale = 1.0
97
-
98
- alibi = (self.slopes / -scale).unsqueeze(1) * relative_position
99
- return alibi
100
-
101
- @staticmethod
102
- def _get_alibi_slopes(n):
103
- def get_slopes_power_of_2(n):
104
- start = 2 ** (-(2 ** -(math.log2(n) - 3)))
105
- ratio = start
106
- return [start * ratio**i for i in range(n)]
107
-
108
- if math.log2(n).is_integer():
109
- return get_slopes_power_of_2(
110
- n
111
- ) # In the paper, we only train models that have 2^a heads for some a. This function has
112
- else: # some good properties that only occur when the input is a power of 2. To maintain that even
113
- closest_power_of_2 = 2 ** math.floor(
114
- math.log2(n)
115
- ) # when the number of heads is not a power of 2, we use this workaround.
116
- return (
117
- get_slopes_power_of_2(closest_power_of_2)
118
- + AlibiPositionEmbeddingLayer._get_alibi_slopes(2 * closest_power_of_2)[0::2][: n - closest_power_of_2]
119
- )
120
-
121
-
122
- def load_tf_weights_in_btlm(model, config, btlm_checkpoint_path):
123
- """Load tf checkpoints in a pytorch model"""
124
- try:
125
- import re
126
-
127
- import tensorflow as tf
128
- except ImportError:
129
- logger.error(
130
- "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
131
- "https://www.tensorflow.org/install/ for installation instructions."
132
- )
133
- raise
134
- tf_path = os.path.abspath(btlm_checkpoint_path)
135
- logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
136
- # Load weights from TF model
137
- init_vars = tf.train.list_variables(tf_path)
138
- names = []
139
- arrays = []
140
- for name, shape in init_vars:
141
- logger.info(f"Loading TF weight {name} with shape {shape}")
142
- array = tf.train.load_variable(tf_path, name)
143
- names.append(name)
144
- arrays.append(array.squeeze())
145
-
146
- for name, array in zip(names, arrays):
147
- name = name[6:] # skip "model/"
148
- name = name.split("/")
149
- pointer = model
150
- for m_name in name:
151
- if re.fullmatch(r"[A-Za-z]+\d+", m_name):
152
- scope_names = re.split(r"(\d+)", m_name)
153
- else:
154
- scope_names = [m_name]
155
- if scope_names[0] == "w" or scope_names[0] == "g":
156
- pointer = getattr(pointer, "weight")
157
- elif scope_names[0] == "b":
158
- pointer = getattr(pointer, "bias")
159
- elif scope_names[0] == "wpe" or scope_names[0] == "wte":
160
- pointer = getattr(pointer, scope_names[0])
161
- pointer = getattr(pointer, "weight")
162
- else:
163
- pointer = getattr(pointer, scope_names[0])
164
- if len(scope_names) >= 2:
165
- num = int(scope_names[1])
166
- pointer = pointer[num]
167
- try:
168
- assert (
169
- pointer.shape == array.shape
170
- ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
171
- except AssertionError as e:
172
- e.args += (pointer.shape, array.shape)
173
- raise
174
- logger.info(f"Initialize PyTorch weight {name}")
175
- pointer.data = torch.from_numpy(array)
176
- return model
177
-
178
-
179
- class BTLMAttention(nn.Module):
180
- def __init__(self, config, is_cross_attention=False, layer_idx=None):
181
- super().__init__()
182
-
183
- max_positions = config.max_position_embeddings
184
- self.register_buffer(
185
- "bias",
186
- torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view(
187
- 1, 1, max_positions, max_positions
188
- ),
189
- persistent=False,
190
- )
191
- self.register_buffer("masked_bias", torch.tensor(-1e4), persistent=False)
192
-
193
- self.embed_dim = config.hidden_size
194
- self.num_heads = config.num_attention_heads
195
- self.head_dim = self.embed_dim // self.num_heads
196
- self.split_size = self.embed_dim
197
- if self.head_dim * self.num_heads != self.embed_dim:
198
- raise ValueError(
199
- f"`embed_dim` must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
200
- f" {self.num_heads})."
201
- )
202
-
203
- self.scale_attn_weights = config.scale_attn_weights
204
- self.is_cross_attention = is_cross_attention
205
-
206
- # Layer-wise attention scaling, reordering, and upcasting
207
- self.scale_attn_by_inverse_layer_idx = config.scale_attn_by_inverse_layer_idx
208
- self.layer_idx = layer_idx
209
- self.reorder_and_upcast_attn = config.reorder_and_upcast_attn
210
-
211
- if self.is_cross_attention:
212
- self.c_attn = Conv1D(2 * self.embed_dim, self.embed_dim)
213
- self.q_attn = Conv1D(self.embed_dim, self.embed_dim)
214
- else:
215
- self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim)
216
- self.c_proj = Conv1D(self.embed_dim, self.embed_dim)
217
-
218
- self.attn_dropout = nn.Dropout(config.attn_pdrop)
219
- self.resid_dropout = nn.Dropout(config.resid_pdrop)
220
-
221
- self.pruned_heads = set()
222
-
223
- self.attn_scale_power = 1.0 if config.mup_scale_qk_dot_by_d else 0.5
224
-
225
- def prune_heads(self, heads):
226
- if len(heads) == 0:
227
- return
228
- heads, index = find_pruneable_heads_and_indices(heads, self.num_heads, self.head_dim, self.pruned_heads)
229
- index_attn = torch.cat([index, index + self.split_size, index + (2 * self.split_size)])
230
-
231
- # Prune conv1d layers
232
- self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
233
- self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
234
-
235
- # Update hyper params
236
- self.split_size = (self.split_size // self.num_heads) * (self.num_heads - len(heads))
237
- self.num_heads = self.num_heads - len(heads)
238
- self.pruned_heads = self.pruned_heads.union(heads)
239
-
240
- def _attn(self, query, key, value, attention_mask=None, head_mask=None, position_bias=None):
241
- attn_weights = torch.matmul(query, key.transpose(-1, -2))
242
-
243
- if self.scale_attn_weights:
244
- attn_weights = attn_weights / torch.full(
245
- [], value.size(-1) ** self.attn_scale_power, dtype=attn_weights.dtype, device=attn_weights.device
246
- )
247
-
248
- # Layer-wise attention scaling
249
- if self.scale_attn_by_inverse_layer_idx:
250
- attn_weights = attn_weights / float(self.layer_idx + 1)
251
-
252
- if not self.is_cross_attention:
253
- # if only "normal" attention layer implements causal mask
254
- query_length, key_length = query.size(-2), key.size(-2)
255
- causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]
256
- mask_value = torch.finfo(attn_weights.dtype).min
257
- # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
258
- # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
259
- mask_value = torch.full([], mask_value, dtype=attn_weights.dtype).to(attn_weights.device)
260
- attn_weights = torch.where(causal_mask, attn_weights.to(attn_weights.dtype), mask_value)
261
-
262
- if attention_mask is not None:
263
- # Apply the attention mask
264
- attn_weights = attn_weights + attention_mask
265
-
266
- if position_bias is not None:
267
- attn_weights += position_bias.type_as(attn_weights).unsqueeze(0)
268
- attn_weights = nn.functional.softmax(attn_weights, dim=-1)
269
-
270
- # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op otherwise
271
- attn_weights = attn_weights.type(value.dtype)
272
- attn_weights = self.attn_dropout(attn_weights)
273
-
274
- # Mask heads if we want to
275
- if head_mask is not None:
276
- attn_weights = attn_weights * head_mask
277
-
278
- attn_output = torch.matmul(attn_weights, value)
279
-
280
- return attn_output, attn_weights
281
-
282
- def _upcast_and_reordered_attn(self, query, key, value, attention_mask=None, head_mask=None, position_bias=None):
283
- # Use `torch.baddbmm` (a bit more efficient w/ alpha param for scaling -- from Megatron-LM)
284
- bsz, num_heads, q_seq_len, dk = query.size()
285
- _, _, k_seq_len, _ = key.size()
286
-
287
- # Preallocate attn_weights for `baddbmm`
288
- attn_weights = torch.empty(bsz * num_heads, q_seq_len, k_seq_len, dtype=torch.float32, device=query.device)
289
-
290
- # Compute Scale Factor
291
- scale_factor = 1.0
292
- if self.scale_attn_weights:
293
- scale_factor /= float(value.size(-1)) ** self.attn_scale_power
294
-
295
- if self.scale_attn_by_inverse_layer_idx:
296
- scale_factor /= float(self.layer_idx + 1)
297
-
298
- # Upcast (turn off autocast) and reorder (Scale K by 1 / root(dk))
299
- with autocast(enabled=False):
300
- q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(-1, dk, k_seq_len)
301
- attn_weights = torch.baddbmm(attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor)
302
- attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)
303
-
304
- if not self.is_cross_attention:
305
- # if only "normal" attention layer implements causal mask
306
- query_length, key_length = query.size(-2), key.size(-2)
307
- causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]
308
- mask_value = torch.finfo(attn_weights.dtype).min
309
- # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
310
- # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
311
- mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(attn_weights.device)
312
- attn_weights = torch.where(causal_mask, attn_weights, mask_value)
313
-
314
- if attention_mask is not None:
315
- # Apply the attention mask
316
- attn_weights = attn_weights + attention_mask
317
-
318
- if position_bias is not None:
319
- attn_weights += position_bias.type_as(attn_weights).unsqueeze(0)
320
- attn_weights = nn.functional.softmax(attn_weights, dim=-1)
321
-
322
- # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op if otherwise
323
- if attn_weights.dtype != torch.float32:
324
- raise RuntimeError("Error with upcasting, attn_weights does not have dtype torch.float32")
325
- attn_weights = attn_weights.type(value.dtype)
326
- attn_weights = self.attn_dropout(attn_weights)
327
-
328
- # Mask heads if we want to
329
- if head_mask is not None:
330
- attn_weights = attn_weights * head_mask
331
-
332
- attn_output = torch.matmul(attn_weights, value)
333
-
334
- return attn_output, attn_weights
335
-
336
- def _split_heads(self, tensor, num_heads, attn_head_size):
337
- """
338
- Splits hidden_size dim into attn_head_size and num_heads
339
- """
340
- new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
341
- tensor = tensor.view(new_shape)
342
- return tensor.permute(0, 2, 1, 3) # (batch, head, seq_length, head_features)
343
-
344
- def _merge_heads(self, tensor, num_heads, attn_head_size):
345
- """
346
- Merges attn_head_size dim and num_attn_heads dim into hidden_size
347
- """
348
- tensor = tensor.permute(0, 2, 1, 3).contiguous()
349
- new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
350
- return tensor.view(new_shape)
351
-
352
- def forward(
353
- self,
354
- hidden_states: Optional[Tuple[torch.FloatTensor]],
355
- layer_past: Optional[Tuple[torch.Tensor]] = None,
356
- attention_mask: Optional[torch.FloatTensor] = None,
357
- head_mask: Optional[torch.FloatTensor] = None,
358
- encoder_hidden_states: Optional[torch.Tensor] = None,
359
- encoder_attention_mask: Optional[torch.FloatTensor] = None,
360
- use_cache: Optional[bool] = False,
361
- output_attentions: Optional[bool] = False,
362
- position_bias: Optional[torch.FloatTensor] = None,
363
- ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
364
- if encoder_hidden_states is not None:
365
- if not hasattr(self, "q_attn"):
366
- raise ValueError(
367
- "If class is used as cross attention, the weights `q_attn` have to be defined. "
368
- "Please make sure to instantiate class with `BTLMAttention(..., is_cross_attention=True)`."
369
- )
370
-
371
- query = self.q_attn(hidden_states)
372
- key, value = self.c_attn(encoder_hidden_states).split(self.split_size, dim=2)
373
- attention_mask = encoder_attention_mask
374
- else:
375
- query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
376
-
377
- query = self._split_heads(query, self.num_heads, self.head_dim)
378
- key = self._split_heads(key, self.num_heads, self.head_dim)
379
- value = self._split_heads(value, self.num_heads, self.head_dim)
380
-
381
- if layer_past is not None:
382
- past_key, past_value = layer_past
383
- key = torch.cat((past_key, key), dim=-2)
384
- value = torch.cat((past_value, value), dim=-2)
385
-
386
- if use_cache is True:
387
- present = (key, value)
388
- else:
389
- present = None
390
-
391
- if self.reorder_and_upcast_attn:
392
- attn_output, attn_weights = self._upcast_and_reordered_attn(
393
- query, key, value, attention_mask, head_mask, position_bias
394
- )
395
- else:
396
- attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask, position_bias)
397
-
398
- attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)
399
- attn_output = self.c_proj(attn_output)
400
- attn_output = self.resid_dropout(attn_output)
401
-
402
- outputs = (attn_output, present)
403
- if output_attentions:
404
- outputs += (attn_weights,)
405
-
406
- return outputs # a, present, (attentions)
407
-
408
-
409
- class BTLMMLP(nn.Module):
410
- def __init__(self, intermediate_size, config):
411
- super().__init__()
412
- embed_dim = config.hidden_size
413
- self.swiglu = config.activation_function == "swiglu"
414
- self.c_fc = Conv1D(intermediate_size, embed_dim)
415
- self.c_fc2 = Conv1D(intermediate_size, embed_dim) if self.swiglu else None
416
- self.c_proj = Conv1D(embed_dim, intermediate_size)
417
- self.act = SwiGLUActivation() if self.swiglu else ACT2FN[config.activation_function]
418
- self.dropout = nn.Dropout(config.resid_pdrop)
419
-
420
- def forward(self, hidden_states: Optional[Tuple[torch.FloatTensor]]) -> torch.FloatTensor:
421
- if self.swiglu:
422
- hidden_states2 = self.c_fc2(hidden_states)
423
- hidden_states = self.c_fc(hidden_states)
424
- hidden_states = self.act(hidden_states, hidden_states2) if self.swiglu else self.act(hidden_states)
425
- hidden_states = self.c_proj(hidden_states)
426
- hidden_states = self.dropout(hidden_states)
427
- return hidden_states
428
-
429
-
430
- class BTLMBlock(nn.Module):
431
- def __init__(self, config, layer_idx=None):
432
- super().__init__()
433
- hidden_size = config.hidden_size
434
- inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
435
-
436
- self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
437
- self.attn = BTLMAttention(config, layer_idx=layer_idx)
438
- self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
439
-
440
- if config.add_cross_attention:
441
- self.crossattention = BTLMAttention(config, is_cross_attention=True, layer_idx=layer_idx)
442
- self.ln_cross_attn = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
443
-
444
- self.mlp = BTLMMLP(inner_dim, config)
445
-
446
- def forward(
447
- self,
448
- hidden_states: Optional[Tuple[torch.FloatTensor]],
449
- layer_past: Optional[Tuple[torch.Tensor]] = None,
450
- attention_mask: Optional[torch.FloatTensor] = None,
451
- head_mask: Optional[torch.FloatTensor] = None,
452
- encoder_hidden_states: Optional[torch.Tensor] = None,
453
- encoder_attention_mask: Optional[torch.FloatTensor] = None,
454
- use_cache: Optional[bool] = False,
455
- output_attentions: Optional[bool] = False,
456
- position_bias: Optional[torch.FloatTensor] = None,
457
- ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
458
- residual = hidden_states
459
- hidden_states = self.ln_1(hidden_states)
460
- attn_outputs = self.attn(
461
- hidden_states,
462
- layer_past=layer_past,
463
- attention_mask=attention_mask,
464
- head_mask=head_mask,
465
- use_cache=use_cache,
466
- output_attentions=output_attentions,
467
- position_bias=position_bias,
468
- )
469
- attn_output = attn_outputs[0] # output_attn: a, present, (attentions)
470
- outputs = attn_outputs[1:]
471
- # residual connection
472
- hidden_states = attn_output + residual
473
-
474
- if encoder_hidden_states is not None:
475
- # add one self-attention block for cross-attention
476
- if not hasattr(self, "crossattention"):
477
- raise ValueError(
478
- f"If `encoder_hidden_states` are passed, {self} has to be instantiated with "
479
- "cross-attention layers by setting `config.add_cross_attention=True`"
480
- )
481
- residual = hidden_states
482
- hidden_states = self.ln_cross_attn(hidden_states)
483
- cross_attn_outputs = self.crossattention(
484
- hidden_states,
485
- attention_mask=attention_mask,
486
- head_mask=head_mask,
487
- encoder_hidden_states=encoder_hidden_states,
488
- encoder_attention_mask=encoder_attention_mask,
489
- output_attentions=output_attentions,
490
- position_bias=position_bias,
491
- )
492
- attn_output = cross_attn_outputs[0]
493
- # residual connection
494
- hidden_states = residual + attn_output
495
- outputs = outputs + cross_attn_outputs[2:] # add cross attentions if we output attention weights
496
-
497
- residual = hidden_states
498
- hidden_states = self.ln_2(hidden_states)
499
- feed_forward_hidden_states = self.mlp(hidden_states)
500
- # residual connection
501
- hidden_states = residual + feed_forward_hidden_states
502
-
503
- if use_cache:
504
- outputs = (hidden_states,) + outputs
505
- else:
506
- outputs = (hidden_states,) + outputs[1:]
507
-
508
- return outputs # hidden_states, present, (attentions, cross_attentions)
509
-
510
-
511
- class BTLMPreTrainedModel(PreTrainedModel):
512
- """
513
- An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
514
- models.
515
- """
516
-
517
- config_class = BTLMConfig
518
- load_tf_weights = load_tf_weights_in_btlm
519
- base_model_prefix = "transformer"
520
- is_parallelizable = True
521
- supports_gradient_checkpointing = True
522
- _no_split_modules = ["BTLMBlock"]
523
- _skip_keys_device_placement = "past_key_values"
524
-
525
- def __init__(self, *inputs, **kwargs):
526
- super().__init__(*inputs, **kwargs)
527
-
528
- def _init_weights(self, module):
529
- """Initialize the weights."""
530
- mup_init_scale = math.sqrt(self.config.mup_width_scale)
531
- if isinstance(module, (nn.Linear, Conv1D)):
532
- # Slightly different from the TF version which uses truncated_normal for initialization
533
- # cf https://github.com/pytorch/pytorch/pull/5617
534
- module.weight.data.normal_(mean=0.0, std=(self.config.initializer_range * mup_init_scale))
535
- if module.bias is not None:
536
- module.bias.data.zero_()
537
- elif isinstance(module, nn.Embedding):
538
- module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
539
- if module.padding_idx is not None:
540
- module.weight.data[module.padding_idx].zero_()
541
- elif isinstance(module, nn.LayerNorm):
542
- module.bias.data.zero_()
543
- module.weight.data.fill_(1.0)
544
-
545
- # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
546
- # > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
547
- # > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
548
- # > -- GPT-2 :: https://openai.com/blog/better-language-models/
549
- #
550
- # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
551
- for name, p in module.named_parameters():
552
- if name == "c_proj.weight":
553
- # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
554
- stddev = self.config.initializer_range * mup_init_scale / math.sqrt(2 * self.config.n_layer)
555
- p.data.normal_(mean=0.0, std=stddev)
556
-
557
- def _set_gradient_checkpointing(self, module, value=False):
558
- if isinstance(module, BTLMModel):
559
- module.gradient_checkpointing = value
560
-
561
- def get_mup_param_groups(self, lr, weight_decay=0.0, decoupled_wd=True):
562
- """
563
- Returns list of dicts defining parameter groups for muP:
564
- group 0: most model params get scaled learning rate and weight decay.
565
- group 1: embedding layer gets non-scaled learning rate and weight decay.
566
- group 2: normalization layers and biases get non-scaled learning rate only.
567
-
568
- The output can be passed to Adam-base optimizers
569
- e.g.
570
- param_groups = model.get_mup_param_groups(lr=1e-3, weight_decay=0.1)
571
- torch.optim.AdamW(param_groups, betas=(0.9, 0.95), eps=1e-8)
572
- """
573
- norm_modules = (
574
- torch.nn.LayerNorm,
575
- torch.nn.BatchNorm1d,
576
- torch.nn.BatchNorm2d,
577
- torch.nn.BatchNorm3d,
578
- torch.nn.InstanceNorm1d,
579
- torch.nn.InstanceNorm2d,
580
- torch.nn.InstanceNorm3d,
581
- torch.nn.GroupNorm,
582
- torch.nn.SyncBatchNorm,
583
- torch.nn.LocalResponseNorm,
584
- )
585
-
586
- def get_group_index(param_name):
587
- for name, module in self.named_modules():
588
- if name in param_name:
589
- if isinstance(module, norm_modules):
590
- return 2
591
- elif isinstance(module, torch.nn.Embedding):
592
- return 1
593
- return 0
594
-
595
- width_scale = self.config.mup_width_scale
596
- new_param_groups = []
597
- new_param_groups.append({"params": [], "lr": lr * width_scale, "weight_decay": weight_decay})
598
- if not decoupled_wd:
599
- new_param_groups[0]["weight_decay"] /= width_scale
600
- new_param_groups.append({"params": [], "lr": lr, "weight_decay": weight_decay})
601
- new_param_groups.append({"params": [], "lr": lr, "weight_decay": 0.0})
602
-
603
- for name, param in self.named_parameters():
604
- if not param.requires_grad:
605
- continue
606
-
607
- if name.endswith("bias"):
608
- new_param_groups[2]["params"].append(param)
609
- else:
610
- new_param_groups[get_group_index(name)]["params"].append(param)
611
-
612
- for idx, param_group in enumerate(new_param_groups):
613
- if len(param_group["params"]) == 0:
614
- del new_param_groups[idx]
615
-
616
- return new_param_groups
617
-
618
-
619
- BTLM_START_DOCSTRING = r"""
620
-
621
- This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
622
- library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
623
- etc.)
624
-
625
- This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
626
- Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
627
- and behavior.
628
-
629
- Parameters:
630
- config ([`BTLMConfig`]): Model configuration class with all the parameters of the model.
631
- Initializing with a config file does not load the weights associated with the model, only the
632
- configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
633
- """
634
-
635
- BTLM_INPUTS_DOCSTRING = r"""
636
- Args:
637
- input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
638
- `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
639
- `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
640
- sequence tokens in the vocabulary.
641
-
642
- If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
643
- `input_ids`.
644
-
645
- Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
646
- [`PreTrainedTokenizer.__call__`] for details.
647
-
648
- [What are input IDs?](../glossary#input-ids)
649
- past_key_values (`Tuple[Tuple[torch.Tensor]]` of length `config.n_layers`):
650
- Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
651
- `past_key_values` output below). Can be used to speed up sequential decoding. The `input_ids` which have
652
- their past given to this model should not be passed as `input_ids` as they have already been computed.
653
- attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
654
- Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
655
-
656
- - 1 for tokens that are **not masked**,
657
- - 0 for tokens that are **masked**.
658
-
659
- If `past_key_values` is used, `attention_mask` needs to contain the masking strategy that was used for
660
- `past_key_values`. In other words, the `attention_mask` always has to have the length:
661
- `len(past_key_values) + len(input_ids)`
662
-
663
- [What are attention masks?](../glossary#attention-mask)
664
- token_type_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
665
- Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
666
- 1]`:
667
-
668
- - 0 corresponds to a *sentence A* token,
669
- - 1 corresponds to a *sentence B* token.
670
-
671
- [What are token type IDs?](../glossary#token-type-ids)
672
- position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
673
- Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
674
- config.max_position_embeddings - 1]`.
675
-
676
- [What are position IDs?](../glossary#position-ids)
677
- head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
678
- Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
679
-
680
- - 1 indicates the head is **not masked**,
681
- - 0 indicates the head is **masked**.
682
-
683
- inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
684
- Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
685
- is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
686
- model's internal embedding lookup matrix.
687
-
688
- If `past_key_values` is used, optionally only the last `inputs_embeds` have to be input (see
689
- `past_key_values`).
690
- use_cache (`bool`, *optional*):
691
- If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
692
- `past_key_values`).
693
- output_attentions (`bool`, *optional*):
694
- Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
695
- tensors for more detail.
696
- output_hidden_states (`bool`, *optional*):
697
- Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
698
- more detail.
699
- return_dict (`bool`, *optional*):
700
- Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
701
- """
702
- PARALLELIZE_DOCSTRING = r"""
703
- This is an experimental feature and is a subject to change at a moment's notice.
704
-
705
- Uses a device map to distribute attention modules of the model across several devices. If no device map is given,
706
- it will evenly distribute blocks across all devices.
707
-
708
- Args:
709
- device_map (`Dict[int, list]`, optional, defaults to None):
710
- A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
711
- automatically mapped to the first device (for esoteric reasons). That means that the first device should
712
- have fewer attention modules mapped to it than other devices. For reference, the gpt2 models have the
713
- following number of attention modules:
714
-
715
- - gpt2: 12
716
- - gpt2-medium: 24
717
- - gpt2-large: 36
718
- - gpt2-xl: 48
719
-
720
- Example:
721
-
722
- ```python
723
- # Here is an example of a device map on a machine with 4 GPUs using gpt2-xl, which has a total of 48 attention modules:
724
- model = GPT2LMHeadModel.from_pretrained("gpt2-xl")
725
- device_map = {
726
- 0: [0, 1, 2, 3, 4, 5, 6, 7, 8],
727
- 1: [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21],
728
- 2: [22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34],
729
- 3: [35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47],
730
- }
731
- model.parallelize(device_map)
732
- ```
733
- """
734
- DEPARALLELIZE_DOCSTRING = r"""
735
- Moves the model to cpu from a model parallel state.
736
-
737
- Example:
738
-
739
- ```python
740
- # On a 4 GPU machine with gpt2-large:
741
- model = GPT2LMHeadModel.from_pretrained("gpt2-large")
742
- device_map = {
743
- 0: [0, 1, 2, 3, 4, 5, 6, 7],
744
- 1: [8, 9, 10, 11, 12, 13, 14, 15],
745
- 2: [16, 17, 18, 19, 20, 21, 22, 23],
746
- 3: [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35],
747
- }
748
- model.parallelize(device_map) # Splits the model across several devices
749
- model.deparallelize() # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()
750
- ```
751
- """
752
-
753
-
754
- @add_start_docstrings(
755
- "The bare BTLM Model transformer outputting raw hidden-states without any specific head on top.",
756
- BTLM_START_DOCSTRING,
757
- )
758
- class BTLMModel(BTLMPreTrainedModel):
759
- _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"]
760
- _keys_to_ignore_on_load_missing = [r"attn.masked_bias", r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias"]
761
-
762
- def __init__(self, config):
763
- super().__init__(config)
764
-
765
- self.embed_dim = config.hidden_size
766
-
767
- self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
768
- self.wpe = (
769
- nn.Embedding(config.max_position_embeddings, self.embed_dim)
770
- if config.position_embedding_type != "alibi"
771
- else None
772
- )
773
- self.embeddings_scale = config.mup_embeddings_scale
774
-
775
- self.drop = nn.Dropout(config.embd_pdrop)
776
- self.h = nn.ModuleList([BTLMBlock(config, layer_idx=i) for i in range(config.num_hidden_layers)])
777
- self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
778
-
779
- self.relative_pe = (
780
- AlibiPositionEmbeddingLayer(config.num_attention_heads, config.alibi_scaling)
781
- if config.position_embedding_type == "alibi"
782
- else None
783
- )
784
-
785
- # Model parallel
786
- self.model_parallel = False
787
- self.device_map = None
788
- self.gradient_checkpointing = False
789
-
790
- # Initialize weights and apply final processing
791
- self.post_init()
792
-
793
- @add_start_docstrings(PARALLELIZE_DOCSTRING)
794
- def parallelize(self, device_map=None):
795
- # Check validity of device_map
796
- warnings.warn(
797
- "`BTLMModel.parallelize` is deprecated and will be removed in v5 of Transformers, you should load your"
798
- " model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
799
- " `device_map` but it needs to be a dictionary module_name to device, so for instance {'h.0': 0, 'h.1': 1,"
800
- " ...}",
801
- FutureWarning,
802
- )
803
- self.device_map = (
804
- get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map
805
- )
806
- assert_device_map(self.device_map, len(self.h))
807
- self.model_parallel = True
808
- self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))
809
- self.last_device = "cuda:" + str(max(self.device_map.keys()))
810
- self.wte = self.wte.to(self.first_device)
811
- if self.wpe is not None:
812
- self.wpe = self.wpe.to(self.first_device)
813
- # Load onto devices
814
- for k, v in self.device_map.items():
815
- for block in v:
816
- cuda_device = "cuda:" + str(k)
817
- self.h[block] = self.h[block].to(cuda_device)
818
- # ln_f to last
819
- self.ln_f = self.ln_f.to(self.last_device)
820
-
821
- @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
822
- def deparallelize(self):
823
- warnings.warn(
824
- "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
825
- FutureWarning,
826
- )
827
- self.model_parallel = False
828
- self.device_map = None
829
- self.first_device = "cpu"
830
- self.last_device = "cpu"
831
- self.wte = self.wte.to("cpu")
832
- if self.wpe is not None:
833
- self.wpe = self.wpe.to("cpu")
834
- for index in range(len(self.h)):
835
- self.h[index] = self.h[index].to("cpu")
836
- self.ln_f = self.ln_f.to("cpu")
837
- torch.cuda.empty_cache()
838
-
839
- def get_input_embeddings(self):
840
- return self.wte
841
-
842
- def set_input_embeddings(self, new_embeddings):
843
- self.wte = new_embeddings
844
-
845
- def _prune_heads(self, heads_to_prune):
846
- """
847
- Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
848
- """
849
- for layer, heads in heads_to_prune.items():
850
- self.h[layer].attn.prune_heads(heads)
851
-
852
- @add_start_docstrings_to_model_forward(BTLM_INPUTS_DOCSTRING)
853
- @add_code_sample_docstrings(
854
- checkpoint=_CHECKPOINT_FOR_DOC,
855
- output_type=BaseModelOutputWithPastAndCrossAttentions,
856
- config_class=_CONFIG_FOR_DOC,
857
- )
858
- def forward(
859
- self,
860
- input_ids: Optional[torch.LongTensor] = None,
861
- past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
862
- attention_mask: Optional[torch.FloatTensor] = None,
863
- token_type_ids: Optional[torch.LongTensor] = None,
864
- position_ids: Optional[torch.LongTensor] = None,
865
- head_mask: Optional[torch.FloatTensor] = None,
866
- inputs_embeds: Optional[torch.FloatTensor] = None,
867
- encoder_hidden_states: Optional[torch.Tensor] = None,
868
- encoder_attention_mask: Optional[torch.FloatTensor] = None,
869
- use_cache: Optional[bool] = None,
870
- output_attentions: Optional[bool] = None,
871
- output_hidden_states: Optional[bool] = None,
872
- return_dict: Optional[bool] = None,
873
- ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
874
- output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
875
- output_hidden_states = (
876
- output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
877
- )
878
- use_cache = use_cache if use_cache is not None else self.config.use_cache
879
- return_dict = return_dict if return_dict is not None else self.config.use_return_dict
880
-
881
- if input_ids is not None and inputs_embeds is not None:
882
- raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
883
- elif input_ids is not None:
884
- input_shape = input_ids.size()
885
- input_ids = input_ids.view(-1, input_shape[-1])
886
- batch_size = input_ids.shape[0]
887
- elif inputs_embeds is not None:
888
- input_shape = inputs_embeds.size()[:-1]
889
- batch_size = inputs_embeds.shape[0]
890
- else:
891
- raise ValueError("You have to specify either input_ids or inputs_embeds")
892
-
893
- device = input_ids.device if input_ids is not None else inputs_embeds.device
894
-
895
- if token_type_ids is not None:
896
- token_type_ids = token_type_ids.view(-1, input_shape[-1])
897
- if position_ids is not None:
898
- position_ids = position_ids.view(-1, input_shape[-1])
899
-
900
- if past_key_values is None:
901
- past_length = 0
902
- past_key_values = tuple([None] * len(self.h))
903
- else:
904
- past_length = past_key_values[0][0].size(-2)
905
- if position_ids is None:
906
- position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
907
- position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
908
-
909
- # BTLMAttention mask.
910
- if attention_mask is not None:
911
- if batch_size <= 0:
912
- raise ValueError("batch_size has to be defined and > 0")
913
- attention_mask = attention_mask.view(batch_size, -1)
914
- # We create a 3D attention mask from a 2D tensor mask.
915
- # Sizes are [batch_size, 1, 1, to_seq_length]
916
- # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
917
- # this attention mask is more simple than the triangular masking of causal attention
918
- # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
919
- attention_mask = attention_mask[:, None, None, :]
920
-
921
- # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
922
- # masked positions, this operation will create a tensor which is 0.0 for
923
- # positions we want to attend and the dtype's smallest value for masked positions.
924
- # Since we are adding it to the raw scores before the softmax, this is
925
- # effectively the same as removing these entirely.
926
- attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility
927
- attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
928
-
929
- # If a 2D or 3D attention mask is provided for the cross-attention
930
- # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
931
- if self.config.add_cross_attention and encoder_hidden_states is not None:
932
- encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
933
- encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
934
- if encoder_attention_mask is None:
935
- encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
936
- encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)
937
- else:
938
- encoder_attention_mask = None
939
-
940
- # Prepare head mask if needed
941
- # 1.0 in head_mask indicate we keep the head
942
- # attention_probs has shape bsz x n_heads x N x N
943
- # head_mask has shape n_layer x batch x n_heads x N x N
944
- head_mask = self.get_head_mask(head_mask, self.config.n_layer)
945
-
946
- if inputs_embeds is None:
947
- inputs_embeds = self.wte(input_ids)
948
- if self.wpe is not None:
949
- position_embeds = self.wpe(position_ids)
950
- hidden_states = inputs_embeds + position_embeds
951
- else:
952
- hidden_states = inputs_embeds
953
- hidden_states *= torch.tensor(
954
- float(self.embeddings_scale), dtype=hidden_states.dtype, device=hidden_states.device
955
- )
956
-
957
- if token_type_ids is not None:
958
- token_type_embeds = self.wte(token_type_ids)
959
- hidden_states = hidden_states + token_type_embeds
960
-
961
- hidden_states = self.drop(hidden_states)
962
-
963
- if self.relative_pe is not None:
964
- length = input_ids.shape[1]
965
- cached_kv_length = 0
966
- cached_kv = past_key_values[0]
967
- if cached_kv is not None:
968
- cached_kv_length = cached_kv[0].shape[-2]
969
- position_bias = self.relative_pe(length, length, cached_kv_length)
970
- else:
971
- position_bias = None
972
-
973
- output_shape = input_shape + (hidden_states.size(-1),)
974
-
975
- if self.gradient_checkpointing and self.training:
976
- if use_cache:
977
- logger.warning_once(
978
- "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
979
- )
980
- use_cache = False
981
-
982
- presents = () if use_cache else None
983
- all_self_attentions = () if output_attentions else None
984
- all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
985
- all_hidden_states = () if output_hidden_states else None
986
- for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
987
- # Model parallel
988
- if self.model_parallel:
989
- torch.cuda.set_device(hidden_states.device)
990
- # Ensure layer_past is on same device as hidden_states (might not be correct)
991
- if layer_past is not None:
992
- layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
993
- # Ensure that attention_mask is always on the same device as hidden_states
994
- if attention_mask is not None:
995
- attention_mask = attention_mask.to(hidden_states.device)
996
- if isinstance(head_mask, torch.Tensor):
997
- head_mask = head_mask.to(hidden_states.device)
998
- if output_hidden_states:
999
- all_hidden_states = all_hidden_states + (hidden_states,)
1000
-
1001
- if self.gradient_checkpointing and self.training:
1002
-
1003
- def create_custom_forward(module):
1004
- def custom_forward(*inputs):
1005
- # None for past_key_value
1006
- return module(*inputs, use_cache, output_attentions)
1007
-
1008
- return custom_forward
1009
-
1010
- outputs = torch.utils.checkpoint.checkpoint(
1011
- create_custom_forward(block),
1012
- hidden_states,
1013
- None,
1014
- attention_mask,
1015
- head_mask[i],
1016
- encoder_hidden_states,
1017
- encoder_attention_mask,
1018
- )
1019
- else:
1020
- outputs = block(
1021
- hidden_states,
1022
- layer_past=layer_past,
1023
- attention_mask=attention_mask,
1024
- head_mask=head_mask[i],
1025
- encoder_hidden_states=encoder_hidden_states,
1026
- encoder_attention_mask=encoder_attention_mask,
1027
- use_cache=use_cache,
1028
- output_attentions=output_attentions,
1029
- position_bias=position_bias,
1030
- )
1031
-
1032
- hidden_states = outputs[0]
1033
- if use_cache is True:
1034
- presents = presents + (outputs[1],)
1035
-
1036
- if output_attentions:
1037
- all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
1038
- if self.config.add_cross_attention:
1039
- all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],)
1040
-
1041
- # Model Parallel: If it's the last layer for that device, put things on the next device
1042
- if self.model_parallel:
1043
- for k, v in self.device_map.items():
1044
- if i == v[-1] and "cuda:" + str(k) != self.last_device:
1045
- hidden_states = hidden_states.to("cuda:" + str(k + 1))
1046
-
1047
- hidden_states = self.ln_f(hidden_states)
1048
-
1049
- hidden_states = hidden_states.view(output_shape)
1050
- # Add last hidden state
1051
- if output_hidden_states:
1052
- all_hidden_states = all_hidden_states + (hidden_states,)
1053
-
1054
- if not return_dict:
1055
- return tuple(
1056
- v
1057
- for v in [hidden_states, presents, all_hidden_states, all_self_attentions, all_cross_attentions]
1058
- if v is not None
1059
- )
1060
-
1061
- return BaseModelOutputWithPastAndCrossAttentions(
1062
- last_hidden_state=hidden_states,
1063
- past_key_values=presents,
1064
- hidden_states=all_hidden_states,
1065
- attentions=all_self_attentions,
1066
- cross_attentions=all_cross_attentions,
1067
- )
1068
-
1069
-
1070
- @add_start_docstrings(
1071
- """
1072
- The BTLM Model transformer with a language modeling head on top (linear layer with weights tied to the input
1073
- embeddings).
1074
- """,
1075
- BTLM_START_DOCSTRING,
1076
- )
1077
- class BTLMLMHeadModel(BTLMPreTrainedModel):
1078
- _keys_to_ignore_on_load_missing = [r"lm_head.weight"]
1079
- _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias"]
1080
-
1081
- def __init__(self, config):
1082
- super().__init__(config)
1083
- self.transformer = BTLMModel(config)
1084
- self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
1085
- self.output_logits_scale = config.mup_output_alpha * config.mup_width_scale
1086
-
1087
- # Model parallel
1088
- self.model_parallel = False
1089
- self.device_map = None
1090
-
1091
- # Initialize weights and apply final processing
1092
- self.post_init()
1093
-
1094
- @add_start_docstrings(PARALLELIZE_DOCSTRING)
1095
- def parallelize(self, device_map=None):
1096
- warnings.warn(
1097
- "`BTLMLMHeadModel.parallelize` is deprecated and will be removed in v5 of Transformers, you should load"
1098
- " your model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
1099
- " `device_map` but it needs to be a dictionary module_name to device, so for instance {'transformer.h.0':"
1100
- " 0, 'transformer.h.1': 1, ...}",
1101
- FutureWarning,
1102
- )
1103
- self.device_map = (
1104
- get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
1105
- if device_map is None
1106
- else device_map
1107
- )
1108
- assert_device_map(self.device_map, len(self.transformer.h))
1109
- self.transformer.parallelize(self.device_map)
1110
- self.lm_head = self.lm_head.to(self.transformer.first_device)
1111
- self.model_parallel = True
1112
-
1113
- @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
1114
- def deparallelize(self):
1115
- warnings.warn(
1116
- "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
1117
- FutureWarning,
1118
- )
1119
- self.transformer.deparallelize()
1120
- self.transformer = self.transformer.to("cpu")
1121
- self.lm_head = self.lm_head.to("cpu")
1122
- self.model_parallel = False
1123
- torch.cuda.empty_cache()
1124
-
1125
- def get_output_embeddings(self):
1126
- return self.lm_head
1127
-
1128
- def set_output_embeddings(self, new_embeddings):
1129
- self.lm_head = new_embeddings
1130
-
1131
- def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
1132
- token_type_ids = kwargs.get("token_type_ids", None)
1133
- # only last token for inputs_ids if past is defined in kwargs
1134
- if past_key_values:
1135
- input_ids = input_ids[:, -1].unsqueeze(-1)
1136
- if token_type_ids is not None:
1137
- token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
1138
-
1139
- attention_mask = kwargs.get("attention_mask", None)
1140
- position_ids = kwargs.get("position_ids", None)
1141
-
1142
- if attention_mask is not None and position_ids is None:
1143
- # create position_ids on the fly for batch generation
1144
- position_ids = attention_mask.long().cumsum(-1) - 1
1145
- position_ids.masked_fill_(attention_mask == 0, 1)
1146
- if past_key_values:
1147
- position_ids = position_ids[:, -1].unsqueeze(-1)
1148
- else:
1149
- position_ids = None
1150
-
1151
- # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
1152
- if inputs_embeds is not None and past_key_values is None:
1153
- model_inputs = {"inputs_embeds": inputs_embeds}
1154
- else:
1155
- model_inputs = {"input_ids": input_ids}
1156
-
1157
- model_inputs.update(
1158
- {
1159
- "past_key_values": past_key_values,
1160
- "use_cache": kwargs.get("use_cache"),
1161
- "position_ids": position_ids,
1162
- "attention_mask": attention_mask,
1163
- "token_type_ids": token_type_ids,
1164
- }
1165
- )
1166
- return model_inputs
1167
-
1168
- @add_start_docstrings_to_model_forward(BTLM_INPUTS_DOCSTRING)
1169
- @add_code_sample_docstrings(
1170
- checkpoint=_CHECKPOINT_FOR_DOC,
1171
- output_type=CausalLMOutputWithCrossAttentions,
1172
- config_class=_CONFIG_FOR_DOC,
1173
- )
1174
- def forward(
1175
- self,
1176
- input_ids: Optional[torch.LongTensor] = None,
1177
- past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
1178
- attention_mask: Optional[torch.FloatTensor] = None,
1179
- token_type_ids: Optional[torch.LongTensor] = None,
1180
- position_ids: Optional[torch.LongTensor] = None,
1181
- head_mask: Optional[torch.FloatTensor] = None,
1182
- inputs_embeds: Optional[torch.FloatTensor] = None,
1183
- encoder_hidden_states: Optional[torch.Tensor] = None,
1184
- encoder_attention_mask: Optional[torch.FloatTensor] = None,
1185
- labels: Optional[torch.LongTensor] = None,
1186
- use_cache: Optional[bool] = None,
1187
- output_attentions: Optional[bool] = None,
1188
- output_hidden_states: Optional[bool] = None,
1189
- return_dict: Optional[bool] = None,
1190
- ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
1191
- r"""
1192
- labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1193
- Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
1194
- `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
1195
- are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
1196
- """
1197
- return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1198
-
1199
- transformer_outputs = self.transformer(
1200
- input_ids,
1201
- past_key_values=past_key_values,
1202
- attention_mask=attention_mask,
1203
- token_type_ids=token_type_ids,
1204
- position_ids=position_ids,
1205
- head_mask=head_mask,
1206
- inputs_embeds=inputs_embeds,
1207
- encoder_hidden_states=encoder_hidden_states,
1208
- encoder_attention_mask=encoder_attention_mask,
1209
- use_cache=use_cache,
1210
- output_attentions=output_attentions,
1211
- output_hidden_states=output_hidden_states,
1212
- return_dict=return_dict,
1213
- )
1214
- hidden_states = transformer_outputs[0]
1215
-
1216
- # Set device for model parallelism
1217
- if self.model_parallel:
1218
- torch.cuda.set_device(self.transformer.first_device)
1219
- hidden_states = hidden_states.to(self.lm_head.weight.device)
1220
-
1221
- lm_logits = self.lm_head(hidden_states)
1222
- lm_logits *= torch.tensor(float(self.output_logits_scale), dtype=lm_logits.dtype, device=lm_logits.device)
1223
-
1224
- loss = None
1225
- if labels is not None:
1226
- # move labels to correct device to enable model parallelism
1227
- labels = labels.to(lm_logits.device)
1228
- # Shift so that tokens < n predict n
1229
- shift_logits = lm_logits[..., :-1, :].contiguous()
1230
- shift_labels = labels[..., 1:].contiguous()
1231
- # Flatten the tokens
1232
- loss_fct = CrossEntropyLoss()
1233
- loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
1234
-
1235
- if not return_dict:
1236
- output = (lm_logits,) + transformer_outputs[1:]
1237
- return ((loss,) + output) if loss is not None else output
1238
-
1239
- return CausalLMOutputWithCrossAttentions(
1240
- loss=loss,
1241
- logits=lm_logits,
1242
- past_key_values=transformer_outputs.past_key_values,
1243
- hidden_states=transformer_outputs.hidden_states,
1244
- attentions=transformer_outputs.attentions,
1245
- cross_attentions=transformer_outputs.cross_attentions,
1246
- )
1247
-
1248
- @staticmethod
1249
- def _reorder_cache(
1250
- past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
1251
- ) -> Tuple[Tuple[torch.Tensor]]:
1252
- """
1253
- This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
1254
- [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
1255
- beam_idx at every generation step.
1256
- """
1257
- return tuple(
1258
- tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
1259
- for layer_past in past_key_values
1260
- )
1261
-
1262
-
1263
- @add_start_docstrings(
1264
- """
1265
- The BTLM Model transformer with a sequence classification head on top (linear layer).
1266
-
1267
- [`BTLMForSequenceClassification`] uses the last token in order to do the classification, as other causal models
1268
- (e.g. GPT-1) do.
1269
-
1270
- Since it does classification on the last token, it requires to know the position of the last token. If a
1271
- `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
1272
- no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
1273
- padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
1274
- each row of the batch).
1275
- """,
1276
- BTLM_START_DOCSTRING,
1277
- )
1278
- class BTLMForSequenceClassification(BTLMPreTrainedModel):
1279
- _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"]
1280
- _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"lm_head.weight"]
1281
-
1282
- def __init__(self, config):
1283
- super().__init__(config)
1284
- self.num_labels = config.num_labels
1285
- self.transformer = BTLMModel(config)
1286
- self.score = nn.Linear(config.n_embd, self.num_labels, bias=False)
1287
- self.output_logits_scale = config.mup_output_alpha * config.mup_width_scale
1288
-
1289
- # Model parallel
1290
- self.model_parallel = False
1291
- self.device_map = None
1292
-
1293
- # Initialize weights and apply final processing
1294
- self.post_init()
1295
-
1296
- @add_start_docstrings_to_model_forward(BTLM_INPUTS_DOCSTRING)
1297
- @add_code_sample_docstrings(
1298
- checkpoint="microsoft/DialogRPT-updown",
1299
- output_type=SequenceClassifierOutputWithPast,
1300
- config_class=_CONFIG_FOR_DOC,
1301
- )
1302
- def forward(
1303
- self,
1304
- input_ids: Optional[torch.LongTensor] = None,
1305
- past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
1306
- attention_mask: Optional[torch.FloatTensor] = None,
1307
- token_type_ids: Optional[torch.LongTensor] = None,
1308
- position_ids: Optional[torch.LongTensor] = None,
1309
- head_mask: Optional[torch.FloatTensor] = None,
1310
- inputs_embeds: Optional[torch.FloatTensor] = None,
1311
- labels: Optional[torch.LongTensor] = None,
1312
- use_cache: Optional[bool] = None,
1313
- output_attentions: Optional[bool] = None,
1314
- output_hidden_states: Optional[bool] = None,
1315
- return_dict: Optional[bool] = None,
1316
- ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
1317
- r"""
1318
- labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1319
- Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1320
- config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1321
- `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1322
- """
1323
- return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1324
-
1325
- transformer_outputs = self.transformer(
1326
- input_ids,
1327
- past_key_values=past_key_values,
1328
- attention_mask=attention_mask,
1329
- token_type_ids=token_type_ids,
1330
- position_ids=position_ids,
1331
- head_mask=head_mask,
1332
- inputs_embeds=inputs_embeds,
1333
- use_cache=use_cache,
1334
- output_attentions=output_attentions,
1335
- output_hidden_states=output_hidden_states,
1336
- return_dict=return_dict,
1337
- )
1338
- hidden_states = transformer_outputs[0]
1339
- logits = self.score(hidden_states)
1340
- logits *= torch.tensor(float(self.output_logits_scale), dtype=logits.dtype, device=logits.device)
1341
-
1342
- if input_ids is not None:
1343
- batch_size, sequence_length = input_ids.shape[:2]
1344
- else:
1345
- batch_size, sequence_length = inputs_embeds.shape[:2]
1346
-
1347
- assert (
1348
- self.config.pad_token_id is not None or batch_size == 1
1349
- ), "Cannot handle batch sizes > 1 if no padding token is defined."
1350
- if self.config.pad_token_id is None:
1351
- sequence_lengths = -1
1352
- else:
1353
- if input_ids is not None:
1354
- sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
1355
- else:
1356
- sequence_lengths = -1
1357
- logger.warning(
1358
- f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
1359
- "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
1360
- )
1361
-
1362
- pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
1363
-
1364
- loss = None
1365
- if labels is not None:
1366
- if self.config.problem_type is None:
1367
- if self.num_labels == 1:
1368
- self.config.problem_type = "regression"
1369
- elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1370
- self.config.problem_type = "single_label_classification"
1371
- else:
1372
- self.config.problem_type = "multi_label_classification"
1373
-
1374
- if self.config.problem_type == "regression":
1375
- loss_fct = MSELoss()
1376
- if self.num_labels == 1:
1377
- loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1378
- else:
1379
- loss = loss_fct(pooled_logits, labels)
1380
- elif self.config.problem_type == "single_label_classification":
1381
- loss_fct = CrossEntropyLoss()
1382
- loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
1383
- elif self.config.problem_type == "multi_label_classification":
1384
- loss_fct = BCEWithLogitsLoss()
1385
- loss = loss_fct(pooled_logits, labels)
1386
- if not return_dict:
1387
- output = (pooled_logits,) + transformer_outputs[1:]
1388
- return ((loss,) + output) if loss is not None else output
1389
-
1390
- return SequenceClassifierOutputWithPast(
1391
- loss=loss,
1392
- logits=pooled_logits,
1393
- past_key_values=transformer_outputs.past_key_values,
1394
- hidden_states=transformer_outputs.hidden_states,
1395
- attentions=transformer_outputs.attentions,
1396
- )
1397
-
1398
-
1399
- @add_start_docstrings(
1400
- """
1401
- BTLM Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
1402
- Named-Entity-Recognition (NER) tasks.
1403
- """,
1404
- BTLM_START_DOCSTRING,
1405
- )
1406
- class BTLMForTokenClassification(BTLMPreTrainedModel):
1407
- def __init__(self, config):
1408
- super().__init__(config)
1409
- self.num_labels = config.num_labels
1410
-
1411
- self.transformer = BTLMModel(config)
1412
- if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
1413
- classifier_dropout = config.classifier_dropout
1414
- elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
1415
- classifier_dropout = config.hidden_dropout
1416
- else:
1417
- classifier_dropout = 0.1
1418
- self.dropout = nn.Dropout(classifier_dropout)
1419
- self.classifier = nn.Linear(config.hidden_size, config.num_labels)
1420
- self.output_logits_scale = config.mup_output_alpha * config.mup_width_scale
1421
-
1422
- # Model parallel
1423
- self.model_parallel = False
1424
- self.device_map = None
1425
-
1426
- # Initialize weights and apply final processing
1427
- self.post_init()
1428
-
1429
- @add_start_docstrings_to_model_forward(BTLM_INPUTS_DOCSTRING)
1430
- # fmt: off
1431
- @add_code_sample_docstrings(
1432
- checkpoint="brad1141/gpt2-finetuned-comp2",
1433
- output_type=TokenClassifierOutput,
1434
- config_class=_CONFIG_FOR_DOC,
1435
- expected_loss=0.25,
1436
- expected_output=["Lead", "Lead", "Lead", "Position", "Lead", "Lead", "Lead", "Lead", "Lead", "Lead", "Lead", "Lead"],
1437
- )
1438
- # fmt: on
1439
- def forward(
1440
- self,
1441
- input_ids: Optional[torch.LongTensor] = None,
1442
- past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
1443
- attention_mask: Optional[torch.FloatTensor] = None,
1444
- token_type_ids: Optional[torch.LongTensor] = None,
1445
- position_ids: Optional[torch.LongTensor] = None,
1446
- head_mask: Optional[torch.FloatTensor] = None,
1447
- inputs_embeds: Optional[torch.FloatTensor] = None,
1448
- labels: Optional[torch.LongTensor] = None,
1449
- use_cache: Optional[bool] = None,
1450
- output_attentions: Optional[bool] = None,
1451
- output_hidden_states: Optional[bool] = None,
1452
- return_dict: Optional[bool] = None,
1453
- ) -> Union[Tuple, TokenClassifierOutput]:
1454
- r"""
1455
- labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1456
- Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1457
- config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1458
- `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1459
- """
1460
- return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1461
-
1462
- transformer_outputs = self.transformer(
1463
- input_ids,
1464
- past_key_values=past_key_values,
1465
- attention_mask=attention_mask,
1466
- token_type_ids=token_type_ids,
1467
- position_ids=position_ids,
1468
- head_mask=head_mask,
1469
- inputs_embeds=inputs_embeds,
1470
- use_cache=use_cache,
1471
- output_attentions=output_attentions,
1472
- output_hidden_states=output_hidden_states,
1473
- return_dict=return_dict,
1474
- )
1475
-
1476
- hidden_states = transformer_outputs[0]
1477
- hidden_states = self.dropout(hidden_states)
1478
- logits = self.classifier(hidden_states)
1479
- logits *= torch.tensor(float(self.output_logits_scale), dtype=logits.dtype, device=logits.device)
1480
-
1481
- loss = None
1482
- if labels is not None:
1483
- labels = labels.to(logits.device)
1484
- loss_fct = CrossEntropyLoss()
1485
- loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
1486
-
1487
- if not return_dict:
1488
- output = (logits,) + transformer_outputs[2:]
1489
- return ((loss,) + output) if loss is not None else output
1490
-
1491
- return TokenClassifierOutput(
1492
- loss=loss,
1493
- logits=logits,
1494
- hidden_states=transformer_outputs.hidden_states,
1495
- attentions=transformer_outputs.attentions,
1496
- )
1497
-
1498
-
1499
- @add_start_docstrings(
1500
- """
1501
- The BTLM Model transformer with a span classification head on top for extractive question-answering tasks like
1502
- SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
1503
- """,
1504
- BTLM_START_DOCSTRING,
1505
- )
1506
- class BTLMForQuestionAnswering(BTLMPreTrainedModel):
1507
- _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"]
1508
- _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias", r"lm_head.weight"]
1509
-
1510
- def __init__(self, config):
1511
- super().__init__(config)
1512
- self.num_labels = config.num_labels
1513
- self.transformer = BTLMModel(config)
1514
- self.qa_outputs = nn.Linear(config.hidden_size, 2)
1515
- self.output_logits_scale = config.mup_output_alpha * config.mup_width_scale
1516
-
1517
- # Model parallel
1518
- self.model_parallel = False
1519
- self.device_map = None
1520
- self.gradient_checkpointing = False
1521
-
1522
- # Initialize weights and apply final processing
1523
- self.post_init()
1524
-
1525
- @add_start_docstrings_to_model_forward(BTLM_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
1526
- @add_code_sample_docstrings(
1527
- checkpoint=_CHECKPOINT_FOR_DOC,
1528
- output_type=QuestionAnsweringModelOutput,
1529
- config_class=_CONFIG_FOR_DOC,
1530
- real_checkpoint=_CHECKPOINT_FOR_DOC,
1531
- )
1532
- def forward(
1533
- self,
1534
- input_ids: Optional[torch.LongTensor] = None,
1535
- attention_mask: Optional[torch.FloatTensor] = None,
1536
- token_type_ids: Optional[torch.LongTensor] = None,
1537
- position_ids: Optional[torch.LongTensor] = None,
1538
- head_mask: Optional[torch.FloatTensor] = None,
1539
- inputs_embeds: Optional[torch.FloatTensor] = None,
1540
- start_positions: Optional[torch.LongTensor] = None,
1541
- end_positions: Optional[torch.LongTensor] = None,
1542
- output_attentions: Optional[bool] = None,
1543
- output_hidden_states: Optional[bool] = None,
1544
- return_dict: Optional[bool] = None,
1545
- ) -> Union[Tuple, QuestionAnsweringModelOutput]:
1546
- r"""
1547
- start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1548
- Labels for position (index) of the start of the labelled span for computing the token classification loss.
1549
- Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
1550
- are not taken into account for computing the loss.
1551
- end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1552
- Labels for position (index) of the end of the labelled span for computing the token classification loss.
1553
- Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
1554
- are not taken into account for computing the loss.
1555
- """
1556
- return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1557
-
1558
- outputs = self.transformer(
1559
- input_ids,
1560
- attention_mask=attention_mask,
1561
- token_type_ids=token_type_ids,
1562
- position_ids=position_ids,
1563
- head_mask=head_mask,
1564
- inputs_embeds=inputs_embeds,
1565
- output_attentions=output_attentions,
1566
- output_hidden_states=output_hidden_states,
1567
- return_dict=return_dict,
1568
- )
1569
-
1570
- sequence_output = outputs[0]
1571
-
1572
- logits = self.qa_outputs(sequence_output)
1573
- logits *= torch.tensor(float(self.output_logits_scale), dtype=logits.dtype, device=logits.device)
1574
- start_logits, end_logits = logits.split(1, dim=-1)
1575
- start_logits = start_logits.squeeze(-1).contiguous()
1576
- end_logits = end_logits.squeeze(-1).contiguous()
1577
-
1578
- total_loss = None
1579
- if start_positions is not None and end_positions is not None:
1580
- # If we are on multi-GPU, split add a dimension
1581
- if len(start_positions.size()) > 1:
1582
- start_positions = start_positions.squeeze(-1).to(start_logits.device)
1583
- if len(end_positions.size()) > 1:
1584
- end_positions = end_positions.squeeze(-1).to(end_logits.device)
1585
- # sometimes the start/end positions are outside our model inputs, we ignore these terms
1586
- ignored_index = start_logits.size(1)
1587
- start_positions = start_positions.clamp(0, ignored_index)
1588
- end_positions = end_positions.clamp(0, ignored_index)
1589
-
1590
- loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
1591
- start_loss = loss_fct(start_logits, start_positions)
1592
- end_loss = loss_fct(end_logits, end_positions)
1593
- total_loss = (start_loss + end_loss) / 2
1594
-
1595
- if not return_dict:
1596
- output = (start_logits, end_logits) + outputs[2:]
1597
- return ((total_loss,) + output) if total_loss is not None else output
1598
-
1599
- return QuestionAnsweringModelOutput(
1600
- loss=total_loss,
1601
- start_logits=start_logits,
1602
- end_logits=end_logits,
1603
- hidden_states=outputs.hidden_states,
1604
- attentions=outputs.attentions,
1605
- )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
pytorch_model.bin DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:c99feaec6acd74a23d674bbc52ca50b7f1109a3537bcb883a85b104905945cde
3
- size 5292652837
 
 
 
 
special_tokens_map.json DELETED
@@ -1,5 +0,0 @@
1
- {
2
- "bos_token": "<|endoftext|>",
3
- "eos_token": "<|endoftext|>",
4
- "unk_token": "<|endoftext|>"
5
- }
 
 
 
 
 
 
table_1_downstream_performance_3b.png DELETED
Binary file (590 kB)
 
table_2_downstream_performance_7b.png DELETED
Binary file (243 kB)
 
tokenizer_config.json DELETED
@@ -1,9 +0,0 @@
1
- {
2
- "add_prefix_space": false,
3
- "bos_token": "<|endoftext|>",
4
- "clean_up_tokenization_spaces": true,
5
- "eos_token": "<|endoftext|>",
6
- "model_max_length": 8192,
7
- "tokenizer_class": "GPT2Tokenizer",
8
- "unk_token": "<|endoftext|>"
9
- }
 
 
 
 
 
 
 
 
 
 
vocab.json DELETED
The diff for this file is too large to render. See raw diff