cekal commited on
Commit
300e678
1 Parent(s): 258e8d9

Upload 16 files

Browse files
README.md ADDED
@@ -0,0 +1,228 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ tags:
4
+ - Composer
5
+ - MosaicML
6
+ - llm-foundry
7
+ - StreamingDatasets
8
+ datasets:
9
+ - mc4
10
+ - c4
11
+ - togethercomputer/RedPajama-Data-1T
12
+ - bigcode/the-stack
13
+ - allenai/s2orc
14
+ inference: false
15
+ ---
16
+
17
+ # MPT-7B
18
+
19
+ MPT-7B is a decoder-style transformer pretrained from scratch on 1T tokens of English text and code.
20
+ This model was trained by [MosaicML](https://www.mosaicml.com).
21
+
22
+ MPT-7B is part of the family of MosaicPretrainedTransformer (MPT) models, which use a modified transformer architecture optimized for efficient training and inference.
23
+
24
+ These architectural changes include performance-optimized layer implementations and the elimination of context length limits by replacing
25
+ positional embeddings with Attention with Linear Biases ([ALiBi](https://arxiv.org/abs/2108.12409)).
26
+ Thanks to these modifications, MPT models can be trained with high throughput efficiency and stable convergence.
27
+ MPT models can also be served efficiently with both standard HuggingFace pipelines and NVIDIA's [FasterTransformer](https://github.com/NVIDIA/FasterTransformer).
28
+
29
+ This model uses the MosaicML LLM codebase, which can be found in the [llm-foundry repository](https://github.com/mosaicml/llm-foundry). It was trained by MosaicML’s NLP team on the [MosaicML platform](https://www.mosaicml.com/training) for LLM pretraining, finetuning, and inference.
30
+
31
+ ### How is this model different?
32
+
33
+ MPT-7B is
34
+
35
+ * **Licensed for the possibility of commercial use** (unlike [LLaMA](https://arxiv.org/abs/2302.13971)).
36
+ * **Trained on a large amount of data** (1T tokens like [LLaMA](https://arxiv.org/abs/2302.13971) vs. 300B for [Pythia](https://github.com/EleutherAI/pythia), 300B for [OpenLLaMA](https://github.com/openlm-research/open_llama), and 800B for [StableLM](https://github.com/Stability-AI/StableLM)).
37
+ * **Prepared to handle extremely long inputs** thanks to [ALiBi](https://arxiv.org/abs/2108.12409) (we finetuned [MPT-7B-StoryWriter-65k+](https://huggingface.co/mosaicml/mpt-7b-storywriter) on up to 65k inputs and can handle up to 84k vs. 2k-4k for other open source models).
38
+ * **Capable of fast training and inference** (via [FlashAttention](https://arxiv.org/pdf/2205.14135.pdf) and [FasterTransformer](https://github.com/NVIDIA/FasterTransformer))
39
+ * **Equipped with highly efficient open-source training code** via the [llm-foundry repository](https://github.com/mosaicml/llm-foundry)
40
+
41
+ ### Models finetuned off MPT-7B:
42
+
43
+ The following models are finetuned on MPT-7B:
44
+
45
+ * [MPT-7B-StoryWriter-65k+](https://huggingface.co/mosaicml/mpt-7b-storywriter): a model designed to read and write fictional stories with super long context lengths.
46
+ Built by finetuning MPT-7B with a context length of 65k tokens on a filtered fiction subset of the [books3 dataset](https://huggingface.co/datasets/the_pile_books3).
47
+ At inference time, thanks to [ALiBi](https://arxiv.org/abs/2108.12409), MPT-7B-StoryWriter-65k+ can extrapolate even beyond 65k tokens.
48
+ We demonstrate generations as long as 80k tokens on a single A100-80GB GPU in our [blogpost](www.mosaicml.com/blog/mpt-7b).
49
+ * License: Apache 2.0
50
+
51
+ * [MPT-7B-Instruct](https://huggingface.co/mosaicml/mpt-7b-instruct): a model for short-form instruction following.
52
+ Built by finetuning MPT-7B on a [dataset](https://huggingface.co/datasets/mosaicml/dolly_hhrlhf) we also release, derived from the [Databricks Dolly-15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k) and the [Anthropic Helpful and Harmless (HH-RLHF)](https://huggingface.co/datasets/Anthropic/hh-rlhf) datasets.
53
+ * License: _CC-By-SA-3.0_
54
+ * [Demo on Hugging Face Spaces](https://huggingface.co/spaces/mosaicml/mpt-7b-instruct)
55
+
56
+ * [MPT-7B-Chat](https://huggingface.co/mosaicml/mpt-7b-chat): a chatbot-like model for dialogue generation.
57
+ Built by finetuning MPT-7B on the [ShareGPT-Vicuna](https://huggingface.co/datasets/jeffwan/sharegpt_vicuna), [HC3](https://huggingface.co/datasets/Hello-SimpleAI/HC3),
58
+ [Alpaca](https://huggingface.co/datasets/tatsu-lab/alpaca), [HH-RLHF](https://huggingface.co/datasets/Anthropic/hh-rlhf), and [Evol-Instruct](https://huggingface.co/datasets/victor123/evol_instruct_70k) datasets.
59
+ * License: _CC-By-NC-SA-4.0_
60
+ * [Demo on Hugging Face Spaces](https://huggingface.co/spaces/mosaicml/mpt-7b-chat)
61
+
62
+ ## Model Date
63
+
64
+ May 5, 2023
65
+
66
+ ## Model License
67
+
68
+ Apache-2.0
69
+
70
+ ## Documentation
71
+
72
+ * [Blog post: Introducing MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs](https://www.mosaicml.com/blog/mpt-7b)
73
+ * [Codebase (mosaicml/llm-foundry repo)](https://github.com/mosaicml/llm-foundry/)
74
+ * Questions: Feel free to contact us via the [MosaicML Community Slack](https://join.slack.com/t/mosaicml-community/shared_invite/zt-1btms90mc-GipE2ufuPkKY0QBrmF3LSA)!
75
+
76
+
77
+ ## How to Use
78
+
79
+ This model is best used with the MosaicML [llm-foundry repository](https://github.com/mosaicml/llm-foundry) for training and finetuning.
80
+
81
+ ```python
82
+ import transformers
83
+ model = transformers.AutoModelForCausalLM.from_pretrained(
84
+ 'mosaicml/mpt-7b',
85
+ trust_remote_code=True
86
+ )
87
+ ```
88
+ Note: This model requires that `trust_remote_code=True` be passed to the `from_pretrained` method.
89
+ This is because we use a custom `MPT` model architecture that is not yet part of the Hugging Face `transformers` package.
90
+ `MPT` includes options for many training efficiency features such as [FlashAttention](https://arxiv.org/pdf/2205.14135.pdf), [ALiBi](https://arxiv.org/abs/2108.12409), [QK LayerNorm](https://arxiv.org/abs/2010.04245), and more.
91
+
92
+ To use the optimized [triton implementation](https://github.com/openai/triton) of FlashAttention, you can load the model with `attn_impl='triton'` and move the model to `bfloat16`:
93
+ ```python
94
+ config = transformers.AutoConfig.from_pretrained(
95
+ 'mosaicml/mpt-7b',
96
+ trust_remote_code=True
97
+ )
98
+ config.attn_config['attn_impl'] = 'triton'
99
+
100
+ model = transformers.AutoModelForCausalLM.from_pretrained(
101
+ 'mosaicml/mpt-7b',
102
+ config=config,
103
+ torch_dtype=torch.bfloat16,
104
+ trust_remote_code=True
105
+ )
106
+ model.to(device='cuda:0')
107
+ ```
108
+
109
+ Although the model was trained with a sequence length of 2048, ALiBi enables users to increase the maximum sequence length during finetuning and/or inference. For example:
110
+
111
+ ```python
112
+ config = transformers.AutoConfig.from_pretrained(
113
+ 'mosaicml/mpt-7b',
114
+ trust_remote_code=True
115
+ )
116
+ config.update({"max_seq_len": 4096})
117
+ model = transformers.AutoModelForCausalLM.from_pretrained(
118
+ 'mosaicml/mpt-7b',
119
+ config=config,
120
+ trust_remote_code=True
121
+ )
122
+ ```
123
+
124
+ This model was trained with the [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b) tokenizer.
125
+
126
+ ```python
127
+ from transformers import AutoTokenizer
128
+ tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b")
129
+ ```
130
+
131
+ ## Model Description
132
+
133
+ The architecture is a modification of a standard decoder-only transformer.
134
+
135
+ The model has been modified from a standard transformer in the following ways:
136
+ * It uses [FlashAttention](https://arxiv.org/pdf/2205.14135.pdf)
137
+ * It uses [ALiBi (Attention with Linear Biases)](https://arxiv.org/abs/2108.12409) and does not use positional embeddings
138
+ * It does not use biases
139
+
140
+
141
+ | Hyperparameter | Value |
142
+ |----------------|-------|
143
+ |n_parameters | 6.7B |
144
+ |n_layers | 32 |
145
+ | n_heads | 32 |
146
+ | d_model | 4096 |
147
+ | vocab size | 50432 |
148
+ | sequence length | 2048 |
149
+
150
+
151
+
152
+ ## Training Data
153
+
154
+ ### Streaming Datasets
155
+
156
+ Data was formatted using the MosaicML [StreamingDataset](https://github.com/mosaicml/streaming) library to host our data in object storage and efficiently stream it to our compute cluster during training.
157
+ StreamingDataset obviates the need to download the whole dataset before starting training, and allows instant resumption of training from any point in the dataset.
158
+
159
+
160
+ ### Data Mix
161
+
162
+ The model was trained for 1T tokens (with batch size 1760 and sequence length 2048). It was trained on the following data mix:
163
+
164
+
165
+ | Data Source | Number of Tokens in Source | Proportion | Effective Number of Tokens | Epochs |
166
+ |-------------|----------------------------|------------|----------------------------|--------|
167
+ | mC4 3.1.0 - English | 417.99 B | 0.33 | 330 B | 0.14 |
168
+ | C4 - English - SemDedup 80% | 100.42 B | 0.299 | 299 B | 2.98 |
169
+ | RedPajama - CommonCrawl | 878.45 B | 0.1 | 100 B | 0.11 |
170
+ | The Stack - Selected Languages | 463.78 B | 0.1 | 100 B | 0.22 |
171
+ | RedPajama - Wikipedia - En | 4.87 B | 0.04 | 40 B | 8.21 |
172
+ | The Stack - Markdown | 107.07 B | 0.035 | 35 B | 0.33 |
173
+ | S2ORC | 48.85 B | 0.033 | 33 B | 0.68 |
174
+ | RedPajama - Books | 26.02 B | 0.03 | 30B | 1.15 |
175
+ | RedPajama - arXiv | 28.10 B | 0.019 | 19 B | 0.68 |
176
+ | RedPajama - StackExchange | 20.54 B | 0.014 | 14 B |0.68 |
177
+
178
+ Samples for each batch were selected from one of the datasets with the probability specified above.
179
+ The examples were shuffled within each dataset, and each example was constructed from as many sequences from that dataset as were necessary to fill the 2048 sequence length.
180
+
181
+ The data was tokenized using the [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b) tokenizer. This BPE tokenizer has a number of desirable characteristics,
182
+ most of which are relevant for tokenizing code:
183
+ (1) It was trained on a diverse mix of data that includes code (The Pile)
184
+ (2) It applies consistent space delimitation, unlike the GPT2 tokenizer which tokenizes inconsistently depending on the presence of prefix spaces
185
+ (3) It contains tokens for repeated space characters, which allows superior compression of text with large amounts of repeated space characters.
186
+
187
+ The model vocabulary size of 50432 was set to be a multiple of 128 (as in [MEGATRON-LM](https://arxiv.org/abs/1909.08053)), model flop utilization (MFU) increased by up to four percentage points.
188
+
189
+ ### Training Configuration
190
+
191
+ This model was trained on 440 A100-40GBs for about 9.5 days using the [MosaicML Platform](https://www.mosaicml.com/platform).
192
+ The model was trained with sharded data parallelism using [FSDP](https://pytorch.org/docs/stable/fsdp.html) and used the [LION](https://arxiv.org/abs/2302.06675) optimizer.
193
+
194
+ ## Limitations and Biases
195
+
196
+ _The following language is modified from [EleutherAI's GPT-NeoX-20B](https://huggingface.co/EleutherAI/gpt-neox-20b)_
197
+
198
+ MPT-7B (Base) is **not** intended for deployment without finetuning.
199
+ It should not be used for human-facing interactions without further guardrails and user consent.
200
+
201
+ MPT-7B can produce factually incorrect output, and should not be relied on to produce factually accurate information.
202
+ MPT-7B was trained on various public datasets.
203
+ While great efforts have been taken to clean the pretraining data, it is possible that this model could generate lewd, biased or otherwise offensive outputs.
204
+
205
+
206
+ ## MosaicML Platform
207
+
208
+ If you're interested in [training](https://www.mosaicml.com/training) and [deploying](https://www.mosaicml.com/inference) your own MPT or LLMs on the MosaicML Platform, [sign up here](https://forms.mosaicml.com/demo?utm_source=huggingface&utm_medium=referral&utm_campaign=mpt-7b).
209
+
210
+ ## Disclaimer
211
+
212
+ The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please cosult an attorney before using this model for commercial purposes.
213
+
214
+ ## Citation
215
+
216
+ Please cite this model using the following format:
217
+
218
+ ```
219
+ @online{MosaicML2023Introducing,
220
+ author = {MosaicML NLP Team},
221
+ title = {Introducing MPT-7B: A New Standard for Open-Source,
222
+ ly Usable LLMs},
223
+ year = {2023},
224
+ url = {www.mosaicml.com/blog/mpt-7b},
225
+ note = {Accessed: 2023-03-28}, % change this date
226
+ urldate = {2023-03-28} % change this date
227
+ }
228
+ ```
adapt_tokenizer.py ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Union
2
+ from transformers import AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast
3
+ Tokenizer = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
4
+ NUM_SENTINEL_TOKENS: int = 100
5
+
6
+ def adapt_tokenizer_for_denoising(tokenizer: Tokenizer):
7
+ """Adds sentinel tokens and padding token (if missing).
8
+
9
+ Expands the tokenizer vocabulary to include sentinel tokens
10
+ used in mixture-of-denoiser tasks as well as a padding token.
11
+
12
+ All added tokens are added as special tokens. No tokens are
13
+ added if sentinel tokens and padding token already exist.
14
+ """
15
+ sentinels_to_add = [f'<extra_id_{i}>' for i in range(NUM_SENTINEL_TOKENS)]
16
+ tokenizer.add_tokens(sentinels_to_add, special_tokens=True)
17
+ if tokenizer.pad_token is None:
18
+ tokenizer.add_tokens('<pad>', special_tokens=True)
19
+ tokenizer.pad_token = '<pad>'
20
+ assert tokenizer.pad_token_id is not None
21
+ sentinels = ''.join([f'<extra_id_{i}>' for i in range(NUM_SENTINEL_TOKENS)])
22
+ _sentinel_token_ids = tokenizer(sentinels, add_special_tokens=False).input_ids
23
+ tokenizer.sentinel_token_ids = _sentinel_token_ids
24
+
25
+ class AutoTokenizerForMOD(AutoTokenizer):
26
+ """AutoTokenizer + Adaptation for MOD.
27
+
28
+ A simple wrapper around AutoTokenizer to make instantiating
29
+ an MOD-adapted tokenizer a bit easier.
30
+
31
+ MOD-adapted tokenizers have sentinel tokens (e.g., <extra_id_0>),
32
+ a padding token, and a property to get the token ids of the
33
+ sentinel tokens.
34
+ """
35
+
36
+ @classmethod
37
+ def from_pretrained(cls, *args, **kwargs):
38
+ """See `AutoTokenizer.from_pretrained` docstring."""
39
+ tokenizer = super().from_pretrained(*args, **kwargs)
40
+ adapt_tokenizer_for_denoising(tokenizer)
41
+ return tokenizer
attention.py ADDED
@@ -0,0 +1,276 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Attention layers."""
2
+ import math
3
+ import warnings
4
+ from typing import Optional
5
+ import torch
6
+ import torch.nn as nn
7
+ from einops import rearrange
8
+ from torch import nn
9
+ from .norm import LPLayerNorm
10
+
11
+ def _reset_is_causal(num_query_tokens: int, num_key_tokens: int, original_is_causal: bool):
12
+ if original_is_causal and num_query_tokens != num_key_tokens:
13
+ if num_query_tokens != 1:
14
+ raise NotImplementedError('MPT does not support query and key with different number of tokens, unless number of query tokens is 1.')
15
+ else:
16
+ return False
17
+ return original_is_causal
18
+
19
+ def scaled_multihead_dot_product_attention(query, key, value, n_heads, softmax_scale=None, attn_bias=None, key_padding_mask=None, is_causal=False, dropout_p=0.0, training=False, needs_weights=False, multiquery=False):
20
+ q = rearrange(query, 'b s (h d) -> b h s d', h=n_heads)
21
+ k = rearrange(key, 'b s (h d) -> b h d s', h=1 if multiquery else n_heads)
22
+ v = rearrange(value, 'b s (h d) -> b h s d', h=1 if multiquery else n_heads)
23
+ min_val = torch.finfo(q.dtype).min
24
+ (b, _, s_q, d) = q.shape
25
+ s_k = k.size(-1)
26
+ if softmax_scale is None:
27
+ softmax_scale = 1 / math.sqrt(d)
28
+ attn_weight = q.matmul(k) * softmax_scale
29
+ if attn_bias is not None:
30
+ if attn_bias.size(-1) != 1 and attn_bias.size(-1) != s_k or (attn_bias.size(-2) != 1 and attn_bias.size(-2) != s_q):
31
+ raise RuntimeError(f'attn_bias (shape: {attn_bias.shape}) is expected to broadcast to shape: {attn_weight.shape}.')
32
+ attn_weight = attn_weight + attn_bias
33
+ if key_padding_mask is not None:
34
+ if attn_bias is not None:
35
+ warnings.warn('Propogating key_padding_mask to the attention module ' + 'and applying it within the attention module can cause ' + 'unneccessary computation/memory usage. Consider integrating ' + 'into attn_bias once and passing that to each attention ' + 'module instead.')
36
+ attn_weight = attn_weight.masked_fill(~key_padding_mask.view((b, 1, 1, s_k)), min_val)
37
+ if is_causal:
38
+ s = max(s_q, s_k)
39
+ causal_mask = attn_weight.new_ones(s, s, dtype=torch.float16)
40
+ causal_mask = causal_mask.tril()
41
+ causal_mask = causal_mask.to(torch.bool)
42
+ causal_mask = ~causal_mask
43
+ causal_mask = causal_mask[-s_q:, -s_k:]
44
+ attn_weight = attn_weight.masked_fill(causal_mask.view(1, 1, s_q, s_k), min_val)
45
+ attn_weight = torch.softmax(attn_weight, dim=-1)
46
+ if dropout_p:
47
+ attn_weight = torch.nn.functional.dropout(attn_weight, p=dropout_p, training=training, inplace=True)
48
+ out = attn_weight.matmul(v)
49
+ out = rearrange(out, 'b h s d -> b s (h d)')
50
+ if needs_weights:
51
+ return (out, attn_weight)
52
+ return (out, None)
53
+
54
+ def check_valid_inputs(*tensors, valid_dtypes=[torch.float16, torch.bfloat16]):
55
+ for tensor in tensors:
56
+ if tensor.dtype not in valid_dtypes:
57
+ raise TypeError(f'tensor.dtype={tensor.dtype!r} must be in valid_dtypes={valid_dtypes!r}.')
58
+ if not tensor.is_cuda:
59
+ raise TypeError(f'Inputs must be cuda tensors (tensor.is_cuda={tensor.is_cuda!r}).')
60
+
61
+ def flash_attn_fn(query, key, value, n_heads, softmax_scale=None, attn_bias=None, key_padding_mask=None, is_causal=False, dropout_p=0.0, training=False, needs_weights=False, multiquery=False):
62
+ try:
63
+ from flash_attn import bert_padding, flash_attn_interface
64
+ except:
65
+ raise RuntimeError('Please install flash-attn==1.0.3.post0')
66
+ check_valid_inputs(query, key, value)
67
+ if attn_bias is not None:
68
+ raise NotImplementedError(f'attn_bias not implemented for flash attn.')
69
+ (batch_size, seqlen) = query.shape[:2]
70
+ if key_padding_mask is None:
71
+ key_padding_mask = torch.ones_like(key[:, :, 0], dtype=torch.bool)
72
+ query_padding_mask = key_padding_mask[:, -query.size(1):]
73
+ (query_unpad, indices_q, cu_seqlens_q, max_seqlen_q) = bert_padding.unpad_input(query, query_padding_mask)
74
+ query_unpad = rearrange(query_unpad, 'nnz (h d) -> nnz h d', h=n_heads)
75
+ (key_unpad, _, cu_seqlens_k, max_seqlen_k) = bert_padding.unpad_input(key, key_padding_mask)
76
+ key_unpad = rearrange(key_unpad, 'nnz (h d) -> nnz h d', h=1 if multiquery else n_heads)
77
+ (value_unpad, _, _, _) = bert_padding.unpad_input(value, key_padding_mask)
78
+ value_unpad = rearrange(value_unpad, 'nnz (h d) -> nnz h d', h=1 if multiquery else n_heads)
79
+ if multiquery:
80
+ key_unpad = key_unpad.expand(key_unpad.size(0), n_heads, key_unpad.size(-1))
81
+ value_unpad = value_unpad.expand(value_unpad.size(0), n_heads, value_unpad.size(-1))
82
+ dropout_p = dropout_p if training else 0.0
83
+ reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
84
+ output_unpad = flash_attn_interface.flash_attn_unpadded_func(query_unpad, key_unpad, value_unpad, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, dropout_p, softmax_scale=softmax_scale, causal=reset_is_causal, return_attn_probs=needs_weights)
85
+ output = bert_padding.pad_input(rearrange(output_unpad, 'nnz h d -> nnz (h d)'), indices_q, batch_size, seqlen)
86
+ return (output, None)
87
+
88
+ def triton_flash_attn_fn(query, key, value, n_heads, softmax_scale=None, attn_bias=None, key_padding_mask=None, is_causal=False, dropout_p=0.0, training=False, needs_weights=False, multiquery=False):
89
+ try:
90
+ from flash_attn import flash_attn_triton
91
+ except:
92
+ raise RuntimeError('Please install flash-attn==1.0.3.post0 and triton==2.0.0.dev20221202')
93
+ check_valid_inputs(query, key, value)
94
+ if dropout_p:
95
+ raise NotImplementedError(f'Dropout not implemented for attn_impl: triton.')
96
+ if needs_weights:
97
+ raise NotImplementedError(f'attn_impl: triton cannot return attn weights.')
98
+ if key_padding_mask is not None:
99
+ warnings.warn('Propagating key_padding_mask to the attention module ' + 'and applying it within the attention module can cause ' + 'unnecessary computation/memory usage. Consider integrating ' + 'into attn_bias once and passing that to each attention ' + 'module instead.')
100
+ (b_size, s_k) = key_padding_mask.shape[:2]
101
+ if attn_bias is None:
102
+ attn_bias = query.new_zeros(b_size, 1, 1, s_k)
103
+ attn_bias = attn_bias.masked_fill(~key_padding_mask.view((b_size, 1, 1, s_k)), torch.finfo(query.dtype).min)
104
+ query = rearrange(query, 'b s (h d) -> b s h d', h=n_heads)
105
+ key = rearrange(key, 'b s (h d) -> b s h d', h=1 if multiquery else n_heads)
106
+ value = rearrange(value, 'b s (h d) -> b s h d', h=1 if multiquery else n_heads)
107
+ if multiquery:
108
+ key = key.expand(*key.shape[:2], n_heads, key.size(-1))
109
+ value = value.expand(*value.shape[:2], n_heads, value.size(-1))
110
+ reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
111
+ attn_output = flash_attn_triton.flash_attn_func(query, key, value, attn_bias, reset_is_causal, softmax_scale)
112
+ output = attn_output.view(*attn_output.shape[:2], -1)
113
+ return (output, None)
114
+
115
+ class MultiheadAttention(nn.Module):
116
+ """Multi-head self attention.
117
+
118
+ Using torch or triton attention implemetation enables user to also use
119
+ additive bias.
120
+ """
121
+
122
+ def __init__(self, d_model: int, n_heads: int, attn_impl: str='triton', clip_qkv: Optional[float]=None, qk_ln: bool=False, softmax_scale: Optional[float]=None, attn_pdrop: float=0.0, low_precision_layernorm: bool=False, device: Optional[str]=None):
123
+ super().__init__()
124
+ self.attn_impl = attn_impl
125
+ self.clip_qkv = clip_qkv
126
+ self.qk_ln = qk_ln
127
+ self.d_model = d_model
128
+ self.n_heads = n_heads
129
+ self.softmax_scale = softmax_scale
130
+ if self.softmax_scale is None:
131
+ self.softmax_scale = 1 / math.sqrt(self.d_model / self.n_heads)
132
+ self.attn_dropout_p = attn_pdrop
133
+ self.Wqkv = nn.Linear(self.d_model, 3 * self.d_model, device=device)
134
+ fuse_splits = (d_model, 2 * d_model)
135
+ self.Wqkv._fused = (0, fuse_splits)
136
+ if self.qk_ln:
137
+ layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
138
+ self.q_ln = layernorm_class(self.d_model, device=device)
139
+ self.k_ln = layernorm_class(self.d_model, device=device)
140
+ if self.attn_impl == 'flash':
141
+ self.attn_fn = flash_attn_fn
142
+ elif self.attn_impl == 'triton':
143
+ self.attn_fn = triton_flash_attn_fn
144
+ warnings.warn('While `attn_impl: triton` can be faster than `attn_impl: flash` ' + 'it uses more memory. When training larger models this can trigger ' + 'alloc retries which hurts performance. If encountered, we recommend ' + 'using `attn_impl: flash` if your model does not use `alibi` or `prefix_lm`.')
145
+ elif self.attn_impl == 'torch':
146
+ self.attn_fn = scaled_multihead_dot_product_attention
147
+ if torch.cuda.is_available():
148
+ warnings.warn('Using `attn_impl: torch`. If your model does not use `alibi` or ' + '`prefix_lm` we recommend using `attn_impl: flash` otherwise ' + 'we recommend using `attn_impl: triton`.')
149
+ else:
150
+ raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
151
+ self.out_proj = nn.Linear(self.d_model, self.d_model, device=device)
152
+ self.out_proj._is_residual = True
153
+
154
+ def forward(self, x, past_key_value=None, attn_bias=None, attention_mask=None, is_causal=True, needs_weights=False):
155
+ qkv = self.Wqkv(x)
156
+ if self.clip_qkv:
157
+ qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
158
+ (query, key, value) = qkv.chunk(3, dim=2)
159
+ key_padding_mask = attention_mask
160
+ if self.qk_ln:
161
+ dtype = query.dtype
162
+ query = self.q_ln(query).to(dtype)
163
+ key = self.k_ln(key).to(dtype)
164
+ if past_key_value is not None:
165
+ if len(past_key_value) != 0:
166
+ key = torch.cat([past_key_value[0], key], dim=1)
167
+ value = torch.cat([past_key_value[1], value], dim=1)
168
+ past_key_value = (key, value)
169
+ if attn_bias is not None:
170
+ attn_bias = attn_bias[:, :, -query.size(1):, -key.size(1):]
171
+ (context, attn_weights) = self.attn_fn(query, key, value, self.n_heads, softmax_scale=self.softmax_scale, attn_bias=attn_bias, key_padding_mask=key_padding_mask, is_causal=is_causal, dropout_p=self.attn_dropout_p, training=self.training, needs_weights=needs_weights)
172
+ return (self.out_proj(context), attn_weights, past_key_value)
173
+
174
+ class MultiQueryAttention(nn.Module):
175
+ """Multi-Query self attention.
176
+
177
+ Using torch or triton attention implemetation enables user to also use
178
+ additive bias.
179
+ """
180
+
181
+ def __init__(self, d_model: int, n_heads: int, attn_impl: str='triton', clip_qkv: Optional[float]=None, qk_ln: bool=False, softmax_scale: Optional[float]=None, attn_pdrop: float=0.0, low_precision_layernorm: bool=False, device: Optional[str]=None):
182
+ super().__init__()
183
+ self.attn_impl = attn_impl
184
+ self.clip_qkv = clip_qkv
185
+ self.qk_ln = qk_ln
186
+ self.d_model = d_model
187
+ self.n_heads = n_heads
188
+ self.head_dim = d_model // n_heads
189
+ self.softmax_scale = softmax_scale
190
+ if self.softmax_scale is None:
191
+ self.softmax_scale = 1 / math.sqrt(self.head_dim)
192
+ self.attn_dropout_p = attn_pdrop
193
+ self.Wqkv = nn.Linear(d_model, d_model + 2 * self.head_dim, device=device)
194
+ fuse_splits = (d_model, d_model + self.head_dim)
195
+ self.Wqkv._fused = (0, fuse_splits)
196
+ if self.qk_ln:
197
+ layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
198
+ self.q_ln = layernorm_class(d_model, device=device)
199
+ self.k_ln = layernorm_class(self.head_dim, device=device)
200
+ if self.attn_impl == 'flash':
201
+ self.attn_fn = flash_attn_fn
202
+ elif self.attn_impl == 'triton':
203
+ self.attn_fn = triton_flash_attn_fn
204
+ warnings.warn('While `attn_impl: triton` can be faster than `attn_impl: flash` ' + 'it uses more memory. When training larger models this can trigger ' + 'alloc retries which hurts performance. If encountered, we recommend ' + 'using `attn_impl: flash` if your model does not use `alibi` or `prefix_lm`.')
205
+ elif self.attn_impl == 'torch':
206
+ self.attn_fn = scaled_multihead_dot_product_attention
207
+ if torch.cuda.is_available():
208
+ warnings.warn('Using `attn_impl: torch`. If your model does not use `alibi` or ' + '`prefix_lm` we recommend using `attn_impl: flash` otherwise ' + 'we recommend using `attn_impl: triton`.')
209
+ else:
210
+ raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
211
+ self.out_proj = nn.Linear(self.d_model, self.d_model, device=device)
212
+ self.out_proj._is_residual = True
213
+
214
+ def forward(self, x, past_key_value=None, attn_bias=None, attention_mask=None, is_causal=True, needs_weights=False):
215
+ qkv = self.Wqkv(x)
216
+ if self.clip_qkv:
217
+ qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
218
+ (query, key, value) = qkv.split([self.d_model, self.head_dim, self.head_dim], dim=2)
219
+ key_padding_mask = attention_mask
220
+ if self.qk_ln:
221
+ dtype = query.dtype
222
+ query = self.q_ln(query).to(dtype)
223
+ key = self.k_ln(key).to(dtype)
224
+ if past_key_value is not None:
225
+ if len(past_key_value) != 0:
226
+ key = torch.cat([past_key_value[0], key], dim=1)
227
+ value = torch.cat([past_key_value[1], value], dim=1)
228
+ past_key_value = (key, value)
229
+ if attn_bias is not None:
230
+ attn_bias = attn_bias[:, :, -query.size(1):, -key.size(1):]
231
+ (context, attn_weights) = self.attn_fn(query, key, value, self.n_heads, softmax_scale=self.softmax_scale, attn_bias=attn_bias, key_padding_mask=key_padding_mask, is_causal=is_causal, dropout_p=self.attn_dropout_p, training=self.training, needs_weights=needs_weights, multiquery=True)
232
+ return (self.out_proj(context), attn_weights, past_key_value)
233
+
234
+ def attn_bias_shape(attn_impl, n_heads, seq_len, alibi, prefix_lm, causal, use_sequence_id):
235
+ if attn_impl == 'flash':
236
+ return None
237
+ elif attn_impl in ['torch', 'triton']:
238
+ if alibi:
239
+ if (prefix_lm or not causal) or use_sequence_id:
240
+ return (1, n_heads, seq_len, seq_len)
241
+ return (1, n_heads, 1, seq_len)
242
+ elif prefix_lm or use_sequence_id:
243
+ return (1, 1, seq_len, seq_len)
244
+ return None
245
+ else:
246
+ raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
247
+
248
+ def build_attn_bias(attn_impl, attn_bias, n_heads, seq_len, causal=False, alibi=False, alibi_bias_max=8):
249
+ if attn_impl == 'flash':
250
+ return None
251
+ elif attn_impl in ['torch', 'triton']:
252
+ if alibi:
253
+ (device, dtype) = (attn_bias.device, attn_bias.dtype)
254
+ attn_bias = attn_bias.add(build_alibi_bias(n_heads, seq_len, full=not causal, alibi_bias_max=alibi_bias_max, device=device, dtype=dtype))
255
+ return attn_bias
256
+ else:
257
+ raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
258
+
259
+ def gen_slopes(n_heads, alibi_bias_max=8, device=None):
260
+ _n_heads = 2 ** math.ceil(math.log2(n_heads))
261
+ m = torch.arange(1, _n_heads + 1, dtype=torch.float32, device=device)
262
+ m = m.mul(alibi_bias_max / _n_heads)
263
+ slopes = 1.0 / torch.pow(2, m)
264
+ if _n_heads != n_heads:
265
+ slopes = torch.concat([slopes[1::2], slopes[::2]])[:n_heads]
266
+ return slopes.view(1, n_heads, 1, 1)
267
+
268
+ def build_alibi_bias(n_heads, seq_len, full=False, alibi_bias_max=8, device=None, dtype=None):
269
+ alibi_bias = torch.arange(1 - seq_len, 1, dtype=torch.int32, device=device).view(1, 1, 1, seq_len)
270
+ if full:
271
+ alibi_bias = alibi_bias - torch.arange(1 - seq_len, 1, dtype=torch.int32, device=device).view(1, 1, seq_len, 1)
272
+ alibi_bias = alibi_bias.abs().mul(-1)
273
+ slopes = gen_slopes(n_heads, alibi_bias_max, device=device)
274
+ alibi_bias = alibi_bias * slopes
275
+ return alibi_bias.to(dtype=dtype)
276
+ ATTN_CLASS_REGISTRY = {'multihead_attention': MultiheadAttention, 'multiquery_attention': MultiQueryAttention}
blocks.py ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GPT Blocks used for the GPT Model."""
2
+ from typing import Dict, Optional, Tuple
3
+ import torch
4
+ import torch.nn as nn
5
+ from .attention import ATTN_CLASS_REGISTRY
6
+ from .norm import NORM_CLASS_REGISTRY
7
+
8
+ class MPTMLP(nn.Module):
9
+
10
+ def __init__(self, d_model: int, expansion_ratio: int, device: Optional[str]=None):
11
+ super().__init__()
12
+ self.up_proj = nn.Linear(d_model, expansion_ratio * d_model, device=device)
13
+ self.act = nn.GELU(approximate='none')
14
+ self.down_proj = nn.Linear(expansion_ratio * d_model, d_model, device=device)
15
+ self.down_proj._is_residual = True
16
+
17
+ def forward(self, x):
18
+ return self.down_proj(self.act(self.up_proj(x)))
19
+
20
+ class MPTBlock(nn.Module):
21
+
22
+ def __init__(self, d_model: int, n_heads: int, expansion_ratio: int, attn_config: Dict={'attn_type': 'multihead_attention', 'attn_pdrop': 0.0, 'attn_impl': 'triton', 'qk_ln': False, 'clip_qkv': None, 'softmax_scale': None, 'prefix_lm': False, 'attn_uses_sequence_id': False, 'alibi': False, 'alibi_bias_max': 8}, resid_pdrop: float=0.0, norm_type: str='low_precision_layernorm', device: Optional[str]=None, **kwargs):
23
+ del kwargs
24
+ super().__init__()
25
+ norm_class = NORM_CLASS_REGISTRY[norm_type.lower()]
26
+ attn_class = ATTN_CLASS_REGISTRY[attn_config['attn_type']]
27
+ self.norm_1 = norm_class(d_model, device=device)
28
+ self.attn = attn_class(attn_impl=attn_config['attn_impl'], clip_qkv=attn_config['clip_qkv'], qk_ln=attn_config['qk_ln'], softmax_scale=attn_config['softmax_scale'], attn_pdrop=attn_config['attn_pdrop'], d_model=d_model, n_heads=n_heads, device=device)
29
+ self.norm_2 = norm_class(d_model, device=device)
30
+ self.ffn = MPTMLP(d_model=d_model, expansion_ratio=expansion_ratio, device=device)
31
+ self.resid_attn_dropout = nn.Dropout(resid_pdrop)
32
+ self.resid_ffn_dropout = nn.Dropout(resid_pdrop)
33
+
34
+ def forward(self, x: torch.Tensor, past_key_value: Optional[Tuple[torch.Tensor]]=None, attn_bias: Optional[torch.Tensor]=None, attention_mask: Optional[torch.ByteTensor]=None, is_causal: bool=True) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor]]]:
35
+ a = self.norm_1(x)
36
+ (b, _, past_key_value) = self.attn(a, past_key_value=past_key_value, attn_bias=attn_bias, attention_mask=attention_mask, is_causal=is_causal)
37
+ x = x + self.resid_attn_dropout(b)
38
+ m = self.norm_2(x)
39
+ n = self.ffn(m)
40
+ x = x + self.resid_ffn_dropout(n)
41
+ return (x, past_key_value)
config (1).json ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "MPTForCausalLM"
4
+ ],
5
+ "attn_config": {
6
+ "alibi": true,
7
+ "alibi_bias_max": 8,
8
+ "attn_impl": "torch",
9
+ "attn_pdrop": 0,
10
+ "attn_type": "multihead_attention",
11
+ "attn_uses_sequence_id": false,
12
+ "clip_qkv": null,
13
+ "prefix_lm": false,
14
+ "qk_ln": false,
15
+ "softmax_scale": null
16
+ },
17
+ "auto_map": {
18
+ "AutoConfig": "configuration_mpt.MPTConfig",
19
+ "AutoModelForCausalLM": "modeling_mpt.MPTForCausalLM"
20
+ },
21
+ "d_model": 4096,
22
+ "emb_pdrop": 0,
23
+ "embedding_fraction": 1.0,
24
+ "expansion_ratio": 4,
25
+ "init_config": {
26
+ "emb_init_std": null,
27
+ "emb_init_uniform_lim": null,
28
+ "fan_mode": "fan_in",
29
+ "init_div_is_residual": true,
30
+ "init_gain": 0,
31
+ "init_nonlinearity": "relu",
32
+ "init_std": 0.02,
33
+ "name": "kaiming_normal_",
34
+ "verbose": 0
35
+ },
36
+ "init_device": "cpu",
37
+ "learned_pos_emb": true,
38
+ "logit_scale": null,
39
+ "max_seq_len": 2048,
40
+ "model_type": "mpt",
41
+ "n_heads": 32,
42
+ "n_layers": 32,
43
+ "no_bias": true,
44
+ "norm_type": "low_precision_layernorm",
45
+ "resid_pdrop": 0,
46
+ "tokenizer_name": "EleutherAI/gpt-neox-20b",
47
+ "torch_dtype": "bfloat16",
48
+ "transformers_version": "4.28.1",
49
+ "use_cache": false,
50
+ "verbose": 0,
51
+ "vocab_size": 50432
52
+ }
configuration_mpt.py ADDED
@@ -0,0 +1,118 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A HuggingFace-style model configuration."""
2
+ from typing import Dict, Optional, Union
3
+ from transformers import PretrainedConfig
4
+ attn_config_defaults: Dict = {'attn_type': 'multihead_attention', 'attn_pdrop': 0.0, 'attn_impl': 'triton', 'qk_ln': False, 'clip_qkv': None, 'softmax_scale': None, 'prefix_lm': False, 'attn_uses_sequence_id': False, 'alibi': False, 'alibi_bias_max': 8}
5
+ init_config_defaults: Dict = {'name': 'kaiming_normal_', 'fan_mode': 'fan_in', 'init_nonlinearity': 'relu'}
6
+
7
+ class MPTConfig(PretrainedConfig):
8
+ model_type = 'mpt'
9
+
10
+ def __init__(self, d_model: int=2048, n_heads: int=16, n_layers: int=24, expansion_ratio: int=4, max_seq_len: int=2048, vocab_size: int=50368, resid_pdrop: float=0.0, emb_pdrop: float=0.0, learned_pos_emb: bool=True, attn_config: Dict=attn_config_defaults, init_device: str='cpu', logit_scale: Optional[Union[float, str]]=None, no_bias: bool=False, verbose: int=0, embedding_fraction: float=1.0, norm_type: str='low_precision_layernorm', use_cache: bool=False, init_config: Dict=init_config_defaults, **kwargs):
11
+ """The MPT configuration class.
12
+
13
+ Args:
14
+ d_model (int): The size of the embedding dimension of the model.
15
+ n_heads (int): The number of attention heads.
16
+ n_layers (int): The number of layers in the model.
17
+ expansion_ratio (int): The ratio of the up/down scale in the MLP.
18
+ max_seq_len (int): The maximum sequence length of the model.
19
+ vocab_size (int): The size of the vocabulary.
20
+ resid_pdrop (float): The dropout probability applied to the attention output before combining with residual.
21
+ emb_pdrop (float): The dropout probability for the embedding layer.
22
+ learned_pos_emb (bool): Whether to use learned positional embeddings
23
+ attn_config (Dict): A dictionary used to configure the model's attention module:
24
+ attn_type (str): type of attention to use. Options: multihead_attention, multiquery_attention
25
+ attn_pdrop (float): The dropout probability for the attention layers.
26
+ attn_impl (str): The attention implementation to use. One of 'torch', 'flash', or 'triton'.
27
+ qk_ln (bool): Whether to apply layer normalization to the queries and keys in the attention layer.
28
+ clip_qkv (Optional[float]): If not None, clip the queries, keys, and values in the attention layer to
29
+ this value.
30
+ softmax_scale (Optional[float]): If not None, scale the softmax in the attention layer by this value. If None,
31
+ use the default scale of ``1/sqrt(d_keys)``.
32
+ prefix_lm (Optional[bool]): Whether the model should operate as a Prefix LM. This requires passing an
33
+ extra `prefix_mask` argument which indicates which tokens belong to the prefix. Tokens in the prefix
34
+ can attend to one another bi-directionally. Tokens outside the prefix use causal attention.
35
+ attn_uses_sequence_id (Optional[bool]): Whether to restrict attention to tokens that have the same sequence_id.
36
+ When the model is in `train` mode, this requires passing an extra `sequence_id` argument which indicates
37
+ which sub-sequence each token belongs to.
38
+ Defaults to ``False`` meaning any provided `sequence_id` will be ignored.
39
+ alibi (bool): Whether to use the alibi bias instead of position embeddings.
40
+ alibi_bias_max (int): The maximum value of the alibi bias.
41
+ init_device (str): The device to use for parameter initialization.
42
+ logit_scale (Optional[Union[float, str]]): If not None, scale the logits by this value.
43
+ no_bias (bool): Whether to use bias in all layers.
44
+ verbose (int): The verbosity level. 0 is silent.
45
+ embedding_fraction (float): The fraction to scale the gradients of the embedding layer by.
46
+ norm_type (str): choose type of norm to use
47
+ multiquery_attention (bool): Whether to use multiquery attention implementation.
48
+ use_cache (bool): Whether or not the model should return the last key/values attentions
49
+ init_config (Dict): A dictionary used to configure the model initialization:
50
+ init_config.name: The parameter initialization scheme to use. Options: 'default_', 'baseline_',
51
+ 'kaiming_uniform_', 'kaiming_normal_', 'neox_init_', 'small_init_', 'xavier_uniform_', or
52
+ 'xavier_normal_'. These mimic the parameter initialization methods in PyTorch.
53
+ init_div_is_residual (Union[int, float, str, bool]): Value to divide initial weights by if ``module._is_residual`` is True.
54
+ emb_init_std (Optional[float]): The standard deviation of the normal distribution used to initialize the embedding layer.
55
+ emb_init_uniform_lim (Optional[Union[Tuple[float, float], float]]): The lower and upper limits of the uniform distribution
56
+ used to initialize the embedding layer. Mutually exclusive with ``emb_init_std``.
57
+ init_std (float): The standard deviation of the normal distribution used to initialize the model,
58
+ if using the baseline_ parameter initialization scheme.
59
+ init_gain (float): The gain to use for parameter initialization with kaiming or xavier initialization schemes.
60
+ fan_mode (str): The fan mode to use for parameter initialization with kaiming initialization schemes.
61
+ init_nonlinearity (str): The nonlinearity to use for parameter initialization with kaiming initialization schemes.
62
+ ---
63
+ See llmfoundry.models.utils.param_init_fns.py for info on other param init config options
64
+ """
65
+ self.d_model = d_model
66
+ self.n_heads = n_heads
67
+ self.n_layers = n_layers
68
+ self.expansion_ratio = expansion_ratio
69
+ self.max_seq_len = max_seq_len
70
+ self.vocab_size = vocab_size
71
+ self.resid_pdrop = resid_pdrop
72
+ self.emb_pdrop = emb_pdrop
73
+ self.learned_pos_emb = learned_pos_emb
74
+ self.attn_config = attn_config
75
+ self.init_device = init_device
76
+ self.logit_scale = logit_scale
77
+ self.no_bias = no_bias
78
+ self.verbose = verbose
79
+ self.embedding_fraction = embedding_fraction
80
+ self.norm_type = norm_type
81
+ self.use_cache = use_cache
82
+ self.init_config = init_config
83
+ if 'name' in kwargs:
84
+ del kwargs['name']
85
+ if 'loss_fn' in kwargs:
86
+ del kwargs['loss_fn']
87
+ super().__init__(**kwargs)
88
+ self._validate_config()
89
+
90
+ def _set_config_defaults(self, config, config_defaults):
91
+ for (k, v) in config_defaults.items():
92
+ if k not in config:
93
+ config[k] = v
94
+ return config
95
+
96
+ def _validate_config(self):
97
+ self.attn_config = self._set_config_defaults(self.attn_config, attn_config_defaults)
98
+ self.init_config = self._set_config_defaults(self.init_config, init_config_defaults)
99
+ if self.d_model % self.n_heads != 0:
100
+ raise ValueError('d_model must be divisible by n_heads')
101
+ if any((prob < 0 or prob > 1 for prob in [self.attn_config['attn_pdrop'], self.resid_pdrop, self.emb_pdrop])):
102
+ raise ValueError("self.attn_config['attn_pdrop'], resid_pdrop, emb_pdrop are probabilities and must be between 0 and 1")
103
+ if self.attn_config['attn_impl'] not in ['torch', 'flash', 'triton']:
104
+ raise ValueError(f"Unknown attn_impl={self.attn_config['attn_impl']}")
105
+ if self.attn_config['prefix_lm'] and self.attn_config['attn_impl'] not in ['torch', 'triton']:
106
+ raise NotImplementedError('prefix_lm only implemented with torch and triton attention.')
107
+ if self.attn_config['alibi'] and self.attn_config['attn_impl'] not in ['torch', 'triton']:
108
+ raise NotImplementedError('alibi only implemented with torch and triton attention.')
109
+ if self.attn_config['attn_uses_sequence_id'] and self.attn_config['attn_impl'] not in ['torch', 'triton']:
110
+ raise NotImplementedError('attn_uses_sequence_id only implemented with torch and triton attention.')
111
+ if self.embedding_fraction > 1 or self.embedding_fraction <= 0:
112
+ raise ValueError('model.embedding_fraction must be between 0 (exclusive) and 1 (inclusive)!')
113
+ if isinstance(self.logit_scale, str) and self.logit_scale != 'inv_sqrt_d_model':
114
+ raise ValueError(f"self.logit_scale={self.logit_scale!r} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'.")
115
+ if self.init_config.get('name', None) is None:
116
+ raise ValueError(f"self.init_config={self.init_config!r} 'name' needs to be set.")
117
+ if not self.learned_pos_emb and (not self.attn_config['alibi']):
118
+ raise ValueError(f'Positional information must be provided to the model using either learned_pos_emb or alibi.')
generation_config.json ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "transformers_version": "4.28.1",
4
+ "use_cache": false
5
+ }
hf_prefixlm_converter.py ADDED
@@ -0,0 +1,415 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Converts Huggingface Causal LM to Prefix LM.
2
+
3
+ Conversion does lightweight surgery on a HuggingFace
4
+ Causal LM to convert it to a Prefix LM.
5
+
6
+ Prefix LMs accepts a `bidirectional_mask` input in `forward`
7
+ and treat the input prompt as the prefix in `generate`.
8
+ """
9
+ import math
10
+ import warnings
11
+ from types import MethodType
12
+ from typing import Any, Dict, List, Optional, Tuple, Union
13
+ import torch
14
+ from transformers.models.bloom.modeling_bloom import BaseModelOutputWithPastAndCrossAttentions, BloomForCausalLM, BloomModel, CausalLMOutputWithCrossAttentions, CrossEntropyLoss
15
+ from transformers.models.bloom.modeling_bloom import _expand_mask as _expand_mask_bloom
16
+ from transformers.models.bloom.modeling_bloom import _make_causal_mask as _make_causal_mask_bloom
17
+ from transformers.models.bloom.modeling_bloom import logging
18
+ from transformers.models.gpt2.modeling_gpt2 import GPT2LMHeadModel
19
+ from transformers.models.gpt_neo.modeling_gpt_neo import GPTNeoForCausalLM
20
+ from transformers.models.gpt_neox.modeling_gpt_neox import GPTNeoXForCausalLM
21
+ from transformers.models.gptj.modeling_gptj import GPTJForCausalLM
22
+ from transformers.models.opt.modeling_opt import OPTForCausalLM
23
+ from transformers.models.opt.modeling_opt import _expand_mask as _expand_mask_opt
24
+ from transformers.models.opt.modeling_opt import _make_causal_mask as _make_causal_mask_opt
25
+ logger = logging.get_logger(__name__)
26
+ _SUPPORTED_GPT_MODELS = (GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM)
27
+ CAUSAL_GPT_TYPES = Union[GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM]
28
+
29
+ def _convert_gpt_causal_lm_to_prefix_lm(model: CAUSAL_GPT_TYPES) -> CAUSAL_GPT_TYPES:
30
+ """Converts a GPT-style Causal LM to a Prefix LM.
31
+
32
+ Supported HuggingFace model classes:
33
+ - `GPT2LMHeadModel`
34
+ - `GPTNeoForCausalLM`
35
+ - `GPTNeoXForCausalLM`
36
+ - `GPTJForCausalLM`
37
+
38
+ See `convert_hf_causal_lm_to_prefix_lm` for more details.
39
+ """
40
+ if hasattr(model, '_prefix_lm_converted'):
41
+ return model
42
+ assert isinstance(model, _SUPPORTED_GPT_MODELS)
43
+ assert model.config.add_cross_attention == False, 'Only supports GPT-style decoder-only models'
44
+
45
+ def _get_attn_modules(model: CAUSAL_GPT_TYPES) -> List[torch.nn.Module]:
46
+ """Helper that gets a list of the model's attention modules.
47
+
48
+ Each module has a `bias` buffer used for causal masking. The Prefix LM
49
+ conversion adds logic to dynamically manipulate these biases to support
50
+ Prefix LM attention masking.
51
+ """
52
+ attn_modules = []
53
+ if isinstance(model, GPTNeoXForCausalLM):
54
+ blocks = model.gpt_neox.layers
55
+ else:
56
+ blocks = model.transformer.h
57
+ for block in blocks:
58
+ if isinstance(model, GPTNeoForCausalLM):
59
+ if block.attn.attention_type != 'global':
60
+ continue
61
+ attn_module = block.attn.attention
62
+ elif isinstance(model, GPTNeoXForCausalLM):
63
+ attn_module = block.attention
64
+ else:
65
+ attn_module = block.attn
66
+ attn_modules.append(attn_module)
67
+ return attn_modules
68
+ setattr(model, '_original_forward', getattr(model, 'forward'))
69
+ setattr(model, '_original_generate', getattr(model, 'generate'))
70
+
71
+ def forward(self: CAUSAL_GPT_TYPES, input_ids: Optional[torch.LongTensor]=None, past_key_values: Optional[Tuple[Tuple[torch.Tensor]]]=None, attention_mask: Optional[torch.FloatTensor]=None, bidirectional_mask: Optional[torch.Tensor]=None, token_type_ids: Optional[torch.LongTensor]=None, position_ids: Optional[torch.LongTensor]=None, head_mask: Optional[torch.FloatTensor]=None, inputs_embeds: Optional[torch.FloatTensor]=None, labels: Optional[torch.LongTensor]=None, use_cache: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, return_dict: Optional[bool]=None):
72
+ """Wraps original forward to enable PrefixLM attention."""
73
+
74
+ def call_og_forward():
75
+ if isinstance(self, GPTNeoXForCausalLM):
76
+ return self._original_forward(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, head_mask=head_mask, inputs_embeds=inputs_embeds, labels=labels, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict)
77
+ else:
78
+ return self._original_forward(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, token_type_ids=token_type_ids, position_ids=position_ids, head_mask=head_mask, inputs_embeds=inputs_embeds, labels=labels, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict)
79
+ if bidirectional_mask is None:
80
+ return call_og_forward()
81
+ assert isinstance(bidirectional_mask, torch.Tensor)
82
+ attn_modules = _get_attn_modules(model)
83
+ (b, s) = bidirectional_mask.shape
84
+ max_length = attn_modules[0].bias.shape[-1]
85
+ if s > max_length:
86
+ raise ValueError(f'bidirectional_mask sequence length (={s}) exceeds the ' + f'max length allowed by the model ({max_length}).')
87
+ assert s <= max_length
88
+ if s < max_length:
89
+ pad = torch.zeros((int(b), int(max_length - s)), dtype=bidirectional_mask.dtype, device=bidirectional_mask.device)
90
+ bidirectional_mask = torch.cat([bidirectional_mask, pad], dim=1)
91
+ bidirectional = bidirectional_mask.unsqueeze(1).unsqueeze(1)
92
+ for attn_module in attn_modules:
93
+ attn_module.bias.data = torch.logical_or(attn_module.bias.data, bidirectional)
94
+ output = call_og_forward()
95
+ for attn_module in attn_modules:
96
+ attn_module.bias.data = torch.tril(attn_module.bias.data[0, 0])[None, None]
97
+ return output
98
+
99
+ def generate(self: CAUSAL_GPT_TYPES, *args: tuple, **kwargs: Dict[str, Any]):
100
+ """Wraps original generate to enable PrefixLM attention."""
101
+ attn_modules = _get_attn_modules(model)
102
+ for attn_module in attn_modules:
103
+ attn_module.bias.data[:] = 1
104
+ output = self._original_generate(*args, **kwargs)
105
+ for attn_module in attn_modules:
106
+ attn_module.bias.data = torch.tril(attn_module.bias.data[0, 0])[None, None]
107
+ return output
108
+ setattr(model, 'forward', MethodType(forward, model))
109
+ setattr(model, 'generate', MethodType(generate, model))
110
+ setattr(model, '_prefix_lm_converted', True)
111
+ return model
112
+
113
+ def _convert_bloom_causal_lm_to_prefix_lm(model: BloomForCausalLM) -> BloomForCausalLM:
114
+ """Converts a BLOOM Causal LM to a Prefix LM.
115
+
116
+ Supported HuggingFace model classes:
117
+ - `BloomForCausalLM`
118
+
119
+ See `convert_hf_causal_lm_to_prefix_lm` for more details.
120
+ """
121
+ if hasattr(model, '_prefix_lm_converted'):
122
+ return model
123
+ assert isinstance(model, BloomForCausalLM)
124
+ assert model.config.add_cross_attention == False, 'Only supports BLOOM decoder-only models'
125
+
126
+ def _prepare_attn_mask(self: BloomModel, attention_mask: torch.Tensor, bidirectional_mask: Optional[torch.Tensor], input_shape: Tuple[int, int], past_key_values_length: int) -> torch.BoolTensor:
127
+ combined_attention_mask = None
128
+ device = attention_mask.device
129
+ (_, src_length) = input_shape
130
+ if src_length > 1:
131
+ combined_attention_mask = _make_causal_mask_bloom(input_shape, device=device, past_key_values_length=past_key_values_length)
132
+ if bidirectional_mask is not None:
133
+ assert attention_mask.shape == bidirectional_mask.shape
134
+ expanded_bidirectional_mask = _expand_mask_bloom(bidirectional_mask, tgt_length=src_length)
135
+ combined_attention_mask = torch.logical_and(combined_attention_mask, expanded_bidirectional_mask)
136
+ expanded_attn_mask = _expand_mask_bloom(attention_mask, tgt_length=src_length)
137
+ combined_attention_mask = expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask | combined_attention_mask
138
+ return combined_attention_mask
139
+
140
+ def _build_alibi_tensor(self: BloomModel, batch_size: int, query_length: int, key_length: int, dtype: torch.dtype, device: torch.device) -> torch.Tensor:
141
+ num_heads = self.config.n_head
142
+ closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
143
+ base = torch.tensor(2 ** (-2 ** (-(math.log2(closest_power_of_2) - 3))), device=device, dtype=torch.float32)
144
+ powers = torch.arange(1, 1 + closest_power_of_2, device=device, dtype=torch.int32)
145
+ slopes = torch.pow(base, powers)
146
+ if closest_power_of_2 != num_heads:
147
+ extra_base = torch.tensor(2 ** (-2 ** (-(math.log2(2 * closest_power_of_2) - 3))), device=device, dtype=torch.float32)
148
+ num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
149
+ extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, device=device, dtype=torch.int32)
150
+ slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)
151
+ qa = torch.arange(query_length, device=device, dtype=torch.int32).view(-1, 1)
152
+ ka = torch.arange(key_length, device=device, dtype=torch.int32).view(1, -1)
153
+ diffs = qa - ka + key_length - query_length
154
+ diffs = -diffs.abs()
155
+ alibi = slopes.view(1, num_heads, 1, 1) * diffs.view(1, 1, query_length, key_length)
156
+ alibi = alibi.expand(batch_size, -1, -1, -1).reshape(-1, query_length, key_length)
157
+ return alibi.to(dtype)
158
+ KeyValueT = Tuple[torch.Tensor, torch.Tensor]
159
+
160
+ def forward(self: BloomModel, input_ids: Optional[torch.LongTensor]=None, past_key_values: Optional[Tuple[KeyValueT, ...]]=None, attention_mask: Optional[torch.Tensor]=None, bidirectional_mask: Optional[torch.Tensor]=None, head_mask: Optional[torch.LongTensor]=None, inputs_embeds: Optional[torch.LongTensor]=None, use_cache: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, return_dict: Optional[bool]=None, **deprecated_arguments) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPastAndCrossAttentions]:
161
+ if deprecated_arguments.pop('position_ids', False) is not False:
162
+ warnings.warn('`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. ' + 'You can safely ignore passing `position_ids`.', FutureWarning)
163
+ if len(deprecated_arguments) > 0:
164
+ raise ValueError(f'Got unexpected arguments: {deprecated_arguments}')
165
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
166
+ output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
167
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
168
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
169
+ if input_ids is not None and inputs_embeds is not None:
170
+ raise ValueError('You cannot specify both input_ids and inputs_embeds at the same time')
171
+ elif input_ids is not None:
172
+ (batch_size, seq_length) = input_ids.shape
173
+ elif inputs_embeds is not None:
174
+ (batch_size, seq_length, _) = inputs_embeds.shape
175
+ else:
176
+ raise ValueError('You have to specify either input_ids or inputs_embeds')
177
+ if past_key_values is None:
178
+ past_key_values = tuple([None] * len(self.h))
179
+ head_mask = self.get_head_mask(head_mask, self.config.n_layer)
180
+ if inputs_embeds is None:
181
+ inputs_embeds = self.word_embeddings(input_ids)
182
+ hidden_states = self.word_embeddings_layernorm(inputs_embeds)
183
+ presents = () if use_cache else None
184
+ all_self_attentions = () if output_attentions else None
185
+ all_hidden_states = () if output_hidden_states else None
186
+ seq_length_with_past = seq_length
187
+ past_key_values_length = 0
188
+ if past_key_values[0] is not None:
189
+ tmp = past_key_values[0][0]
190
+ past_key_values_length = tmp.shape[2]
191
+ seq_length_with_past = seq_length_with_past + past_key_values_length
192
+ if attention_mask is None:
193
+ attention_mask = torch.ones((batch_size, seq_length_with_past), device=hidden_states.device)
194
+ else:
195
+ attention_mask = attention_mask.to(hidden_states.device)
196
+ alibi = self._build_alibi_tensor(batch_size=batch_size, query_length=seq_length, key_length=seq_length_with_past, dtype=hidden_states.dtype, device=hidden_states.device)
197
+ causal_mask = self._prepare_attn_mask(attention_mask, bidirectional_mask, input_shape=(batch_size, seq_length), past_key_values_length=past_key_values_length)
198
+ for (i, (block, layer_past)) in enumerate(zip(self.h, past_key_values)):
199
+ if output_hidden_states:
200
+ hst = (hidden_states,)
201
+ all_hidden_states = all_hidden_states + hst
202
+ if self.gradient_checkpointing and self.training:
203
+ if use_cache:
204
+ logger.warning('`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...')
205
+ use_cache = False
206
+
207
+ def create_custom_forward(module):
208
+
209
+ def custom_forward(*inputs):
210
+ return module(*inputs, use_cache=use_cache, output_attentions=output_attentions)
211
+ return custom_forward
212
+ outputs = torch.utils.checkpoint.checkpoint(create_custom_forward(block), hidden_states, alibi, causal_mask, head_mask[i])
213
+ else:
214
+ outputs = block(hidden_states, layer_past=layer_past, attention_mask=causal_mask, head_mask=head_mask[i], use_cache=use_cache, output_attentions=output_attentions, alibi=alibi)
215
+ hidden_states = outputs[0]
216
+ if use_cache is True:
217
+ presents = presents + (outputs[1],)
218
+ if output_attentions:
219
+ oa = (outputs[2 if use_cache else 1],)
220
+ all_self_attentions = all_self_attentions + oa
221
+ hidden_states = self.ln_f(hidden_states)
222
+ if output_hidden_states:
223
+ hst = (hidden_states,)
224
+ all_hidden_states = all_hidden_states + hst
225
+ if not return_dict:
226
+ return tuple((v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None))
227
+ return BaseModelOutputWithPastAndCrossAttentions(last_hidden_state=hidden_states, past_key_values=presents, hidden_states=all_hidden_states, attentions=all_self_attentions)
228
+ setattr(model.transformer, '_prepare_attn_mask', MethodType(_prepare_attn_mask, model.transformer))
229
+ setattr(model.transformer, '_build_alibi_tensor', MethodType(_build_alibi_tensor, model.transformer))
230
+ setattr(model.transformer, 'forward', MethodType(forward, model.transformer))
231
+ KeyValueT = Tuple[torch.Tensor, torch.Tensor]
232
+
233
+ def forward(self: BloomForCausalLM, input_ids: Optional[torch.LongTensor]=None, past_key_values: Optional[Tuple[KeyValueT, ...]]=None, attention_mask: Optional[torch.Tensor]=None, bidirectional_mask: Optional[torch.Tensor]=None, head_mask: Optional[torch.Tensor]=None, inputs_embeds: Optional[torch.Tensor]=None, labels: Optional[torch.Tensor]=None, use_cache: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, return_dict: Optional[bool]=None, **deprecated_arguments) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
234
+ """Replacement forward method for BloomCausalLM."""
235
+ if deprecated_arguments.pop('position_ids', False) is not False:
236
+ warnings.warn('`position_ids` have no functionality in BLOOM and will be removed ' + 'in v5.0.0. You can safely ignore passing `position_ids`.', FutureWarning)
237
+ if len(deprecated_arguments) > 0:
238
+ raise ValueError(f'Got unexpected arguments: {deprecated_arguments}')
239
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
240
+ transformer_outputs = self.transformer(input_ids, past_key_values=past_key_values, attention_mask=attention_mask, bidirectional_mask=bidirectional_mask, head_mask=head_mask, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict)
241
+ hidden_states = transformer_outputs[0]
242
+ lm_logits = self.lm_head(hidden_states)
243
+ loss = None
244
+ if labels is not None:
245
+ shift_logits = lm_logits[..., :-1, :].contiguous()
246
+ shift_labels = labels[..., 1:].contiguous()
247
+ (batch_size, seq_length, vocab_size) = shift_logits.shape
248
+ loss_fct = CrossEntropyLoss()
249
+ loss = loss_fct(shift_logits.view(batch_size * seq_length, vocab_size), shift_labels.view(batch_size * seq_length))
250
+ if not return_dict:
251
+ output = (lm_logits,) + transformer_outputs[1:]
252
+ return (loss,) + output if loss is not None else output
253
+ return CausalLMOutputWithCrossAttentions(loss=loss, logits=lm_logits, past_key_values=transformer_outputs.past_key_values, hidden_states=transformer_outputs.hidden_states, attentions=transformer_outputs.attentions)
254
+
255
+ def prepare_inputs_for_generation(self: BloomForCausalLM, input_ids: torch.LongTensor, past: Optional[torch.Tensor]=None, attention_mask: Optional[torch.Tensor]=None, **kwargs) -> dict:
256
+ if past:
257
+ input_ids = input_ids[:, -1].unsqueeze(-1)
258
+ bidirectional_mask = None
259
+ if past[0][0].shape[0] == input_ids.shape[0]:
260
+ past = self._convert_to_bloom_cache(past)
261
+ else:
262
+ bidirectional_mask = torch.ones_like(input_ids)
263
+ return {'input_ids': input_ids, 'past_key_values': past, 'use_cache': True, 'attention_mask': attention_mask, 'bidirectional_mask': bidirectional_mask}
264
+ setattr(model, 'forward', MethodType(forward, model))
265
+ setattr(model, 'prepare_inputs_for_generation', MethodType(prepare_inputs_for_generation, model))
266
+ setattr(model, '_prefix_lm_converted', True)
267
+ return model
268
+
269
+ def _convert_opt_causal_lm_to_prefix_lm(model: OPTForCausalLM) -> OPTForCausalLM:
270
+ """Converts an OPT Causal LM to a Prefix LM.
271
+
272
+ Supported HuggingFace model classes:
273
+ - `OPTForCausalLM`
274
+
275
+ See `convert_hf_causal_lm_to_prefix_lm` for more details.
276
+ """
277
+ if hasattr(model, '_prefix_lm_converted'):
278
+ return model
279
+ assert isinstance(model, OPTForCausalLM)
280
+ assert model.config.add_cross_attention == False, 'Only supports OPT decoder-only models'
281
+ setattr(model, '_original_forward', getattr(model, 'forward'))
282
+ setattr(model, '_original_generate', getattr(model, 'generate'))
283
+ model.model.decoder.bidirectional_mask = None
284
+
285
+ def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
286
+ combined_attention_mask = None
287
+ if input_shape[-1] > 1:
288
+ if self.bidirectional_mask == 'g':
289
+ (bsz, src_length) = input_shape
290
+ combined_attention_mask = torch.zeros((bsz, 1, src_length, src_length + past_key_values_length), dtype=inputs_embeds.dtype, device=inputs_embeds.device)
291
+ else:
292
+ combined_attention_mask = _make_causal_mask_opt(input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length).to(inputs_embeds.device)
293
+ if self.bidirectional_mask is not None:
294
+ assert attention_mask.shape == self.bidirectional_mask.shape
295
+ expanded_bidirectional_mask = _expand_mask_opt(self.bidirectional_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(inputs_embeds.device)
296
+ combined_attention_mask = torch.maximum(expanded_bidirectional_mask, combined_attention_mask)
297
+ if attention_mask is not None:
298
+ expanded_attn_mask = _expand_mask_opt(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(inputs_embeds.device)
299
+ combined_attention_mask = expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
300
+ return combined_attention_mask
301
+ setattr(model.model.decoder, '_prepare_decoder_attention_mask', MethodType(_prepare_decoder_attention_mask, model.model.decoder))
302
+
303
+ def forward(self: OPTForCausalLM, input_ids: Optional[torch.LongTensor]=None, attention_mask: Optional[torch.Tensor]=None, bidirectional_mask: Optional[torch.ByteTensor]=None, head_mask: Optional[torch.Tensor]=None, past_key_values: Optional[List[torch.FloatTensor]]=None, inputs_embeds: Optional[torch.FloatTensor]=None, labels: Optional[torch.LongTensor]=None, use_cache: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, return_dict: Optional[bool]=None):
304
+
305
+ def call_og_forward():
306
+ return self._original_forward(input_ids=input_ids, attention_mask=attention_mask, head_mask=head_mask, past_key_values=past_key_values, inputs_embeds=inputs_embeds, labels=labels, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict)
307
+ if bidirectional_mask is None:
308
+ return call_og_forward()
309
+ self.model.decoder.bidirectional_mask = bidirectional_mask
310
+ try:
311
+ outputs = call_og_forward()
312
+ except:
313
+ self.model.decoder.bidirectional_mask = None
314
+ raise
315
+ self.model.decoder.bidirectional_mask = None
316
+ return outputs
317
+
318
+ def generate(self: OPTForCausalLM, *args: tuple, **kwargs: Dict[str, Any]):
319
+ """Wraps original generate to enable PrefixLM-style attention."""
320
+ self.model.decoder.bidirectional_mask = 'g'
321
+ try:
322
+ output = self._original_generate(*args, **kwargs)
323
+ except:
324
+ self.model.decoder.bidirectional_mask = None
325
+ raise
326
+ self.model.decoder.bidirectional_mask = None
327
+ return output
328
+ setattr(model, 'forward', MethodType(forward, model))
329
+ setattr(model, 'generate', MethodType(generate, model))
330
+ setattr(model, '_prefix_lm_converted', True)
331
+ return model
332
+ _SUPPORTED_HF_MODELS = _SUPPORTED_GPT_MODELS + (BloomForCausalLM, OPTForCausalLM)
333
+ CAUSAL_LM_TYPES = Union[GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM, BloomForCausalLM, OPTForCausalLM]
334
+
335
+ def convert_hf_causal_lm_to_prefix_lm(model: CAUSAL_LM_TYPES) -> CAUSAL_LM_TYPES:
336
+ """Converts a HuggingFace Causal LM to a Prefix LM.
337
+
338
+ Supported HuggingFace model classes:
339
+ - `GPT2LMHeadModel`
340
+ - `GPTNeoForCausalLM`
341
+ - `GPTNeoXForCausalLM`
342
+ - `GPTJForCausalLM`
343
+ - `BloomForCausalLM`
344
+ - `OPTForCausalLM`
345
+
346
+ Conversion to a Prefix LM is done by modifying the `forward` method, and possibly also the
347
+ `generate` method and/or select underlying methods depending on the model class.
348
+
349
+ These changes preserve the model API, but add a new input to `forward`: "bidirectional_mask".
350
+
351
+ Notes on training:
352
+ To actually train the converted model as a Prefix LM, training batches will need to indicate
353
+ the prefix/target structure by including `bidirectional_mask` as part of the batch inputs.
354
+
355
+ **This is not a standard input and requires custom layers either within or after your dataloader.**
356
+
357
+ In addition to adding `bidirectional_mask` to the batch, this custom code should modify `labels`
358
+ such that `batch['labels'][batch['bidirectional_mask'] == 1] == -100`.
359
+ That is, the prefix portion of the sequence should not generate any loss. Loss should only be
360
+ generated by the target portion of the sequence.
361
+
362
+ Notes on `GPTNeoForCausalLM`:
363
+ To simplify the implementation, "global" and "local" attention layers are handled differently.
364
+ For "global" layers, we handle conversion as described above. For "local" layers, which use a
365
+ causal attention mask within a restricted local window, we do not alter the masking.
366
+
367
+ Notes on `forward` method conversion:
368
+ After conversion, the `forward` method will handle a new input, `bidirectional_mask`,
369
+ which should be a [batch_size, seq_length] byte tensor, where 1 indicates token positions
370
+ belonging to the prefix (prefix tokens can attend to one another bidirectionally), and
371
+ 0 indicates token positions belonging to the target.
372
+
373
+ The new `forward` method will incorporate `bidirectional_mask` (if supplied) into the existing
374
+ causal mask, call the original `forward` method, and (if the causal mask is a buffer) reset
375
+ the causal masks before returning the result.
376
+
377
+ Notes on `generate` method conversion:
378
+ After conversion, the `generate` method will have the same signature but will internally
379
+ convert all causal masks to be purely bidirectional, call the original `generate` method, and
380
+ (where appropriate) reset the causal masks before returning the result.
381
+
382
+ This works thanks to the logic of the HuggingFace `generate` API, which first encodes the token
383
+ "prompt" passed to `generate` (which is treated as the prefix) and then sequentially generates
384
+ each new token. Encodings are cached as generation happens, so all prefix tokens can attend to one
385
+ another (as expected in a Prefix LM) and generated tokens can only attend to prefix tokens and
386
+ previously-generated tokens (also as expected in a Prefix LM).
387
+
388
+ To preserve the API, the original methods are renamed to `_original_forward` and
389
+ `_original_generate`, and replaced with new `forward` and `generate` methods that wrap
390
+ them, respectively. Although implementation details vary by model class.
391
+ """
392
+ if isinstance(model, _SUPPORTED_GPT_MODELS):
393
+ return _convert_gpt_causal_lm_to_prefix_lm(model)
394
+ elif isinstance(model, BloomForCausalLM):
395
+ return _convert_bloom_causal_lm_to_prefix_lm(model)
396
+ elif isinstance(model, OPTForCausalLM):
397
+ return _convert_opt_causal_lm_to_prefix_lm(model)
398
+ else:
399
+ raise TypeError(f'Cannot convert model to Prefix LM. ' + f'Model does not belong to set of supported HF models:' + f'\n{_SUPPORTED_HF_MODELS}')
400
+
401
+ def add_bidirectional_mask_if_missing(batch: Dict[str, Any]):
402
+ """Attempts to add bidirectional_mask to batch if missing.
403
+
404
+ Raises:
405
+ KeyError if bidirectional_mask is missing and can't be inferred
406
+ """
407
+ if 'bidirectional_mask' not in batch:
408
+ if batch.get('mode', None) == 'icl_task':
409
+ batch['bidirectional_mask'] = batch['attention_mask'].clone()
410
+ for (i, continuation_indices) in enumerate(batch['continuation_indices']):
411
+ batch['bidirectional_mask'][i, continuation_indices] = 0
412
+ elif 'labels' in batch and 'attention_mask' in batch:
413
+ batch['bidirectional_mask'] = torch.logical_and(torch.eq(batch['attention_mask'], 1), torch.eq(batch['labels'], -100)).type_as(batch['attention_mask'])
414
+ else:
415
+ raise KeyError('No bidirectional_mask in batch and not sure how to construct one.')
meta_init_context.py ADDED
@@ -0,0 +1,94 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from contextlib import contextmanager
2
+ import torch
3
+ import torch.nn as nn
4
+
5
+ @contextmanager
6
+ def init_empty_weights(include_buffers: bool=False):
7
+ """Meta initialization context manager.
8
+
9
+ A context manager under which models are initialized with all parameters
10
+ on the meta device, therefore creating an empty model. Useful when just
11
+ initializing the model would blow the available RAM.
12
+
13
+ Args:
14
+ include_buffers (`bool`, *optional*, defaults to `False`): Whether or
15
+ not to also put all buffers on the meta device while initializing.
16
+
17
+ Example:
18
+ ```python
19
+ import torch.nn as nn
20
+
21
+ # Initialize a model with 100 billions parameters in no time and without using any RAM.
22
+ with init_empty_weights():
23
+ tst = nn.Sequential(*[nn.Linear(10000, 10000) for _ in range(1000)])
24
+ ```
25
+
26
+ <Tip warning={true}>
27
+
28
+ Any model created under this context manager has no weights. As such you can't do something like
29
+ `model.to(some_device)` with it. To load weights inside your empty model, see [`load_checkpoint_and_dispatch`].
30
+
31
+ </Tip>
32
+ """
33
+ with init_on_device(torch.device('meta'), include_buffers=include_buffers) as f:
34
+ yield f
35
+
36
+ @contextmanager
37
+ def init_on_device(device: torch.device, include_buffers: bool=False):
38
+ """Device initialization context manager.
39
+
40
+ A context manager under which models are initialized with all parameters
41
+ on the specified device.
42
+
43
+ Args:
44
+ device (`torch.device`): Device to initialize all parameters on.
45
+ include_buffers (`bool`, *optional*, defaults to `False`): Whether or
46
+ not to also put all buffers on the meta device while initializing.
47
+
48
+ Example:
49
+ ```python
50
+ import torch.nn as nn
51
+
52
+ with init_on_device(device=torch.device("cuda")):
53
+ tst = nn.Liner(100, 100) # on `cuda` device
54
+ ```
55
+ """
56
+ old_register_parameter = nn.Module.register_parameter
57
+ if include_buffers:
58
+ old_register_buffer = nn.Module.register_buffer
59
+
60
+ def register_empty_parameter(module, name, param):
61
+ old_register_parameter(module, name, param)
62
+ if param is not None:
63
+ param_cls = type(module._parameters[name])
64
+ kwargs = module._parameters[name].__dict__
65
+ module._parameters[name] = param_cls(module._parameters[name].to(device), **kwargs)
66
+
67
+ def register_empty_buffer(module, name, buffer):
68
+ old_register_buffer(module, name, buffer)
69
+ if buffer is not None:
70
+ module._buffers[name] = module._buffers[name].to(device)
71
+ if include_buffers:
72
+ tensor_constructors_to_patch = {torch_function_name: getattr(torch, torch_function_name) for torch_function_name in ['empty', 'zeros', 'ones', 'full']}
73
+ else:
74
+ tensor_constructors_to_patch = {}
75
+
76
+ def patch_tensor_constructor(fn):
77
+
78
+ def wrapper(*args, **kwargs):
79
+ kwargs['device'] = device
80
+ return fn(*args, **kwargs)
81
+ return wrapper
82
+ try:
83
+ nn.Module.register_parameter = register_empty_parameter
84
+ if include_buffers:
85
+ nn.Module.register_buffer = register_empty_buffer
86
+ for torch_function_name in tensor_constructors_to_patch.keys():
87
+ setattr(torch, torch_function_name, patch_tensor_constructor(getattr(torch, torch_function_name)))
88
+ yield
89
+ finally:
90
+ nn.Module.register_parameter = old_register_parameter
91
+ if include_buffers:
92
+ nn.Module.register_buffer = old_register_buffer
93
+ for (torch_function_name, old_torch_function) in tensor_constructors_to_patch.items():
94
+ setattr(torch, torch_function_name, old_torch_function)
modeling_mpt.py ADDED
@@ -0,0 +1,290 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A simple, flexible implementation of a GPT model.
2
+
3
+ Inspired by https://github.com/karpathy/minGPT/blob/master/mingpt/model.py
4
+ """
5
+ import math
6
+ import warnings
7
+ from typing import List, Optional, Tuple, Union
8
+ import torch
9
+ import torch.nn as nn
10
+ import torch.nn.functional as F
11
+ from transformers import PreTrainedModel, PreTrainedTokenizer, PreTrainedTokenizerFast
12
+ from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
13
+ from .attention import attn_bias_shape, build_attn_bias
14
+ from .blocks import MPTBlock
15
+ from .norm import NORM_CLASS_REGISTRY
16
+ from .configuration_mpt import MPTConfig
17
+ from .adapt_tokenizer import AutoTokenizerForMOD, adapt_tokenizer_for_denoising
18
+ from .hf_prefixlm_converter import add_bidirectional_mask_if_missing, convert_hf_causal_lm_to_prefix_lm
19
+ from .meta_init_context import init_empty_weights
20
+ from .param_init_fns import MODEL_INIT_REGISTRY, generic_param_init_fn_
21
+ Tokenizer = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
22
+
23
+ class MPTPreTrainedModel(PreTrainedModel):
24
+ config_class = MPTConfig
25
+ base_model_prefix = 'model'
26
+
27
+ class MPTModel(MPTPreTrainedModel):
28
+
29
+ def __init__(self, config: MPTConfig):
30
+ config._validate_config()
31
+ super().__init__(config)
32
+ self.attn_impl = config.attn_config['attn_impl']
33
+ self.prefix_lm = config.attn_config['prefix_lm']
34
+ self.attn_uses_sequence_id = config.attn_config['attn_uses_sequence_id']
35
+ self.alibi = config.attn_config['alibi']
36
+ self.alibi_bias_max = config.attn_config['alibi_bias_max']
37
+ if config.norm_type.lower() not in NORM_CLASS_REGISTRY.keys():
38
+ norm_options = ' | '.join(NORM_CLASS_REGISTRY.keys())
39
+ raise NotImplementedError(f'Requested norm type ({config.norm_type}) is not implemented within this repo (Options: {norm_options}).')
40
+ norm_class = NORM_CLASS_REGISTRY[config.norm_type.lower()]
41
+ self.embedding_fraction = config.embedding_fraction
42
+ self.wte = nn.Embedding(config.vocab_size, config.d_model, device=config.init_device)
43
+ if not self.alibi:
44
+ self.wpe = nn.Embedding(config.max_seq_len, config.d_model, device=config.init_device)
45
+ self.emb_drop = nn.Dropout(config.emb_pdrop)
46
+ self.blocks = nn.ModuleList([MPTBlock(device=config.init_device, **config.to_dict()) for _ in range(config.n_layers)])
47
+ self.norm_f = norm_class(config.d_model, device=config.init_device)
48
+ if config.init_device != 'meta':
49
+ self.apply(self.param_init_fn)
50
+ self.is_causal = not self.prefix_lm
51
+ self._attn_bias_initialized = False
52
+ self.attn_bias = None
53
+ self.attn_bias_shape = attn_bias_shape(self.attn_impl, config.n_heads, config.max_seq_len, self.alibi, prefix_lm=self.prefix_lm, causal=self.is_causal, use_sequence_id=self.attn_uses_sequence_id)
54
+ if config.no_bias:
55
+ for module in self.modules():
56
+ if hasattr(module, 'bias') and isinstance(module.bias, nn.Parameter):
57
+ if config.verbose:
58
+ warnings.warn(f'Removing bias ({module.bias}) from {module}.')
59
+ module.register_parameter('bias', None)
60
+ if config.verbose and config.verbose > 2:
61
+ print(self)
62
+ if 'verbose' not in self.config.init_config:
63
+ self.config.init_config['verbose'] = self.config.verbose
64
+ if self.config.init_config['verbose'] > 1:
65
+ init_fn_name = self.config.init_config['name']
66
+ warnings.warn(f'Using {init_fn_name} initialization.')
67
+
68
+ def get_input_embeddings(self):
69
+ return self.wte
70
+
71
+ def set_input_embeddings(self, value):
72
+ self.wte = value
73
+
74
+ @torch.no_grad()
75
+ def _attn_bias(self, device, dtype, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None):
76
+ if not self._attn_bias_initialized:
77
+ if self.attn_bias_shape:
78
+ self.attn_bias = torch.zeros(self.attn_bias_shape, device=device, dtype=dtype)
79
+ self.attn_bias = build_attn_bias(self.attn_impl, self.attn_bias, self.config.n_heads, self.config.max_seq_len, causal=self.is_causal, alibi=self.alibi, alibi_bias_max=self.alibi_bias_max)
80
+ self._attn_bias_initialized = True
81
+ if self.attn_impl == 'flash':
82
+ return (self.attn_bias, attention_mask)
83
+ if self.attn_bias is not None:
84
+ self.attn_bias = self.attn_bias.to(dtype=dtype, device=device)
85
+ attn_bias = self.attn_bias
86
+ if self.prefix_lm:
87
+ assert isinstance(attn_bias, torch.Tensor)
88
+ assert isinstance(prefix_mask, torch.Tensor)
89
+ attn_bias = self._apply_prefix_mask(attn_bias, prefix_mask)
90
+ if self.attn_uses_sequence_id and sequence_id is not None:
91
+ assert isinstance(attn_bias, torch.Tensor)
92
+ attn_bias = self._apply_sequence_id(attn_bias, sequence_id)
93
+ if attention_mask is not None:
94
+ s_k = attention_mask.shape[-1]
95
+ if attn_bias is None:
96
+ attn_bias = torch.zeros((1, 1, 1, s_k), device=device, dtype=dtype)
97
+ else:
98
+ attn_bias = attn_bias[:, :, :, -s_k:]
99
+ if prefix_mask is not None and attention_mask.shape != prefix_mask.shape:
100
+ raise ValueError(f'attention_mask shape={attention_mask.shape} ' + f'and prefix_mask shape={prefix_mask.shape} are not equal.')
101
+ min_val = torch.finfo(attn_bias.dtype).min
102
+ attn_bias = attn_bias.masked_fill(~attention_mask.view(-1, 1, 1, s_k), min_val)
103
+ return (attn_bias, None)
104
+
105
+ def _apply_prefix_mask(self, attn_bias: torch.Tensor, prefix_mask: torch.Tensor):
106
+ (s_k, s_q) = attn_bias.shape[-2:]
107
+ if s_k != self.config.max_seq_len or s_q != self.config.max_seq_len:
108
+ raise ValueError('attn_bias does not match the expected shape. ' + f'The last two dimensions should both be {self.config.max_length} ' + f'but are {s_k} and {s_q}.')
109
+ seq_len = prefix_mask.shape[-1]
110
+ if seq_len > self.config.max_seq_len:
111
+ raise ValueError(f'prefix_mask sequence length cannot exceed max_seq_len={self.config.max_seq_len}')
112
+ attn_bias = attn_bias[..., :seq_len, :seq_len]
113
+ causal = torch.tril(torch.ones((seq_len, seq_len), dtype=torch.bool, device=prefix_mask.device)).view(1, 1, seq_len, seq_len)
114
+ prefix = prefix_mask.view(-1, 1, 1, seq_len)
115
+ cannot_attend = ~torch.logical_or(causal, prefix.bool())
116
+ min_val = torch.finfo(attn_bias.dtype).min
117
+ attn_bias = attn_bias.masked_fill(cannot_attend, min_val)
118
+ return attn_bias
119
+
120
+ def _apply_sequence_id(self, attn_bias: torch.Tensor, sequence_id: torch.LongTensor):
121
+ seq_len = sequence_id.shape[-1]
122
+ if seq_len > self.config.max_seq_len:
123
+ raise ValueError(f'sequence_id sequence length cannot exceed max_seq_len={self.config.max_seq_len}')
124
+ attn_bias = attn_bias[..., :seq_len, :seq_len]
125
+ cannot_attend = torch.logical_not(torch.eq(sequence_id.view(-1, seq_len, 1), sequence_id.view(-1, 1, seq_len))).unsqueeze(1)
126
+ min_val = torch.finfo(attn_bias.dtype).min
127
+ attn_bias = attn_bias.masked_fill(cannot_attend, min_val)
128
+ return attn_bias
129
+
130
+ def forward(self, input_ids: torch.LongTensor, past_key_values: Optional[List[Tuple[torch.FloatTensor]]]=None, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None, return_dict: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, use_cache: Optional[bool]=None):
131
+ return_dict = return_dict if return_dict is not None else self.config.return_dict
132
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
133
+ if attention_mask is not None:
134
+ attention_mask = attention_mask.bool()
135
+ if prefix_mask is not None:
136
+ prefix_mask = prefix_mask.bool()
137
+ if not return_dict:
138
+ raise NotImplementedError('return_dict False is not implemented yet for MPT')
139
+ if output_attentions:
140
+ raise NotImplementedError('output_attentions is not implemented yet for MPT')
141
+ if attention_mask is not None and attention_mask[:, 0].sum() != attention_mask.shape[0] and self.training:
142
+ raise NotImplementedError('MPT does not support training with left padding.')
143
+ if self.prefix_lm and prefix_mask is None:
144
+ raise ValueError('prefix_mask is a required argument when MPT is configured with prefix_lm=True.')
145
+ if self.training:
146
+ if self.attn_uses_sequence_id and sequence_id is None:
147
+ raise ValueError('sequence_id is a required argument when MPT is configured with attn_uses_sequence_id=True ' + 'and the model is in train mode.')
148
+ elif self.attn_uses_sequence_id is False and sequence_id is not None:
149
+ warnings.warn('MPT received non-None input for `sequence_id` but is configured with attn_uses_sequence_id=False. ' + 'This input will be ignored. If you want the model to use `sequence_id`, set attn_uses_sequence_id to True.')
150
+ S = input_ids.size(1)
151
+ assert S <= self.config.max_seq_len, f'Cannot forward input with seq_len={S}, this model only supports seq_len<={self.config.max_seq_len}'
152
+ tok_emb = self.wte(input_ids)
153
+ if self.alibi:
154
+ x = tok_emb
155
+ else:
156
+ past_position = 0
157
+ if past_key_values is not None:
158
+ if len(past_key_values) != self.config.n_layers:
159
+ raise ValueError(f'past_key_values must provide a past_key_value for each attention ' + f'layer in the network (len(past_key_values)={len(past_key_values)!r}; self.config.n_layers={self.config.n_layers!r}).')
160
+ past_position = past_key_values[0][0].size(1)
161
+ if S + past_position > self.config.max_seq_len:
162
+ raise ValueError(f'Cannot forward input with past sequence length {past_position} and current sequence length {S + 1}, this model only supports total sequence length <= {self.config.max_seq_len}.')
163
+ pos = torch.arange(past_position, S + past_position, dtype=torch.long, device=input_ids.device).unsqueeze(0)
164
+ if attention_mask is not None:
165
+ pos = torch.clamp(pos - torch.cumsum((~attention_mask).to(torch.int32), dim=1)[:, past_position:], min=0)
166
+ pos_emb = self.wpe(pos)
167
+ x = tok_emb + pos_emb
168
+ if self.embedding_fraction == 1:
169
+ x = self.emb_drop(x)
170
+ else:
171
+ x_shrunk = x * self.embedding_fraction + x.detach() * (1 - self.embedding_fraction)
172
+ assert isinstance(self.emb_drop, nn.Module)
173
+ x = self.emb_drop(x_shrunk)
174
+ (attn_bias, attention_mask) = self._attn_bias(device=x.device, dtype=x.dtype, attention_mask=attention_mask, prefix_mask=prefix_mask, sequence_id=sequence_id)
175
+ if use_cache and past_key_values is None:
176
+ past_key_values = [() for _ in range(self.config.n_layers)]
177
+ all_hidden_states = () if output_hidden_states else None
178
+ for (b_idx, block) in enumerate(self.blocks):
179
+ if output_hidden_states:
180
+ assert all_hidden_states is not None
181
+ all_hidden_states = all_hidden_states + (x,)
182
+ past_key_value = past_key_values[b_idx] if past_key_values is not None else None
183
+ (x, past_key_value) = block(x, past_key_value=past_key_value, attn_bias=attn_bias, attention_mask=attention_mask, is_causal=self.is_causal)
184
+ if past_key_values is not None:
185
+ past_key_values[b_idx] = past_key_value
186
+ x = self.norm_f(x)
187
+ return BaseModelOutputWithPast(last_hidden_state=x, past_key_values=past_key_values, hidden_states=all_hidden_states)
188
+
189
+ def param_init_fn(self, module):
190
+ init_fn_name = self.config.init_config['name']
191
+ MODEL_INIT_REGISTRY[init_fn_name](module=module, n_layers=self.config.n_layers, d_model=self.config.d_model, **self.config.init_config)
192
+
193
+ def fsdp_wrap_fn(self, module):
194
+ return isinstance(module, MPTBlock)
195
+
196
+ def activation_checkpointing_fn(self, module):
197
+ return isinstance(module, MPTBlock)
198
+
199
+ class MPTForCausalLM(MPTPreTrainedModel):
200
+
201
+ def __init__(self, config: MPTConfig):
202
+ super().__init__(config)
203
+ if not config.tie_word_embeddings:
204
+ raise ValueError('MPTForCausalLM only supports tied word embeddings')
205
+ self.transformer = MPTModel(config)
206
+ self.logit_scale = None
207
+ if config.logit_scale is not None:
208
+ logit_scale = config.logit_scale
209
+ if isinstance(logit_scale, str):
210
+ if logit_scale == 'inv_sqrt_d_model':
211
+ logit_scale = 1 / math.sqrt(config.d_model)
212
+ else:
213
+ raise ValueError(f"logit_scale={logit_scale!r} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'.")
214
+ self.logit_scale = logit_scale
215
+
216
+ def get_input_embeddings(self):
217
+ return self.transformer.wte
218
+
219
+ def set_input_embeddings(self, value):
220
+ self.transformer.wte = value
221
+
222
+ def get_output_embeddings(self):
223
+ return self.transformer.wte
224
+
225
+ def set_output_embeddings(self, new_embeddings):
226
+ self.transformer.wte = new_embeddings
227
+
228
+ def set_decoder(self, decoder):
229
+ self.transformer = decoder
230
+
231
+ def get_decoder(self):
232
+ return self.transformer
233
+
234
+ def forward(self, input_ids: torch.LongTensor, past_key_values: Optional[List[Tuple[torch.FloatTensor]]]=None, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None, labels: Optional[torch.LongTensor]=None, return_dict: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, use_cache: Optional[bool]=None):
235
+ return_dict = return_dict if return_dict is not None else self.config.return_dict
236
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
237
+ outputs = self.transformer(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, prefix_mask=prefix_mask, sequence_id=sequence_id, return_dict=return_dict, output_attentions=output_attentions, output_hidden_states=output_hidden_states, use_cache=use_cache)
238
+ logits = F.linear(outputs.last_hidden_state, self.transformer.wte.weight)
239
+ if self.logit_scale is not None:
240
+ if self.logit_scale == 0:
241
+ warnings.warn(f'Multiplying logits by self.logit_scale={self.logit_scale!r}. This will produce uniform (uninformative) outputs.')
242
+ logits *= self.logit_scale
243
+ loss = None
244
+ if labels is not None:
245
+ labels = torch.roll(labels, shifts=-1)
246
+ labels[:, -1] = -100
247
+ loss = F.cross_entropy(logits.view(-1, logits.size(-1)), labels.to(logits.device).view(-1))
248
+ return CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states)
249
+
250
+ def param_init_fn(self, module):
251
+ init_fn_name = self.config.init_config['name']
252
+ MODEL_INIT_REGISTRY[init_fn_name](module=module, n_layers=self.config.n_layers, d_model=self.config.d_model, **self.config.init_config)
253
+
254
+ def fsdp_wrap_fn(self, module):
255
+ return isinstance(module, MPTBlock)
256
+
257
+ def activation_checkpointing_fn(self, module):
258
+ return isinstance(module, MPTBlock)
259
+
260
+ def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
261
+ if inputs_embeds is not None:
262
+ raise NotImplementedError('inputs_embeds is not implemented for MPT yet')
263
+ attention_mask = kwargs['attention_mask'].bool()
264
+ if attention_mask[:, -1].sum() != attention_mask.shape[0]:
265
+ raise NotImplementedError('MPT does not support generation with right padding.')
266
+ if self.transformer.attn_uses_sequence_id and self.training:
267
+ sequence_id = torch.zeros_like(input_ids[:1])
268
+ else:
269
+ sequence_id = None
270
+ if past_key_values is not None:
271
+ input_ids = input_ids[:, -1].unsqueeze(-1)
272
+ if self.transformer.prefix_lm:
273
+ prefix_mask = torch.ones_like(attention_mask)
274
+ if kwargs.get('use_cache') == False:
275
+ raise NotImplementedError('MPT with prefix_lm=True does not support use_cache=False.')
276
+ else:
277
+ prefix_mask = None
278
+ return {'input_ids': input_ids, 'attention_mask': attention_mask, 'prefix_mask': prefix_mask, 'sequence_id': sequence_id, 'past_key_values': past_key_values, 'use_cache': kwargs.get('use_cache', True)}
279
+
280
+ @staticmethod
281
+ def _reorder_cache(past_key_values, beam_idx):
282
+ """Used by HuggingFace generate when using beam search with kv-caching.
283
+
284
+ See https://github.com/huggingface/transformers/blob/3ec7a47664ebe40c40f4b722f6bb1cd30c3821ec/src/transformers/models/gpt2/modeling_gpt2.py#L1122-L1133
285
+ for an example in transformers.
286
+ """
287
+ reordered_past = []
288
+ for layer_past in past_key_values:
289
+ reordered_past += [tuple((past_state.index_select(0, beam_idx) for past_state in layer_past))]
290
+ return reordered_past
norm.py ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+
3
+ def _cast_if_autocast_enabled(tensor):
4
+ if torch.is_autocast_enabled():
5
+ if tensor.device.type == 'cuda':
6
+ dtype = torch.get_autocast_gpu_dtype()
7
+ elif tensor.device.type == 'cpu':
8
+ dtype = torch.get_autocast_cpu_dtype()
9
+ else:
10
+ raise NotImplementedError()
11
+ return tensor.to(dtype=dtype)
12
+ return tensor
13
+
14
+ class LPLayerNorm(torch.nn.LayerNorm):
15
+
16
+ def __init__(self, normalized_shape, eps=1e-05, elementwise_affine=True, device=None, dtype=None):
17
+ super().__init__(normalized_shape=normalized_shape, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
18
+
19
+ def forward(self, x):
20
+ module_device = x.device
21
+ downcast_x = _cast_if_autocast_enabled(x)
22
+ downcast_weight = _cast_if_autocast_enabled(self.weight) if self.weight is not None else self.weight
23
+ downcast_bias = _cast_if_autocast_enabled(self.bias) if self.bias is not None else self.bias
24
+ with torch.autocast(enabled=False, device_type=module_device.type):
25
+ return torch.nn.functional.layer_norm(downcast_x, self.normalized_shape, downcast_weight, downcast_bias, self.eps)
26
+
27
+ def rms_norm(x, weight=None, eps=1e-05):
28
+ output = x / torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + eps)
29
+ if weight is not None:
30
+ return output * weight
31
+ return output
32
+
33
+ class RMSNorm(torch.nn.Module):
34
+
35
+ def __init__(self, normalized_shape, eps=1e-05, weight=True, dtype=None, device=None):
36
+ super().__init__()
37
+ self.eps = eps
38
+ if weight:
39
+ self.weight = torch.nn.Parameter(torch.ones(normalized_shape, dtype=dtype, device=device))
40
+ else:
41
+ self.register_parameter('weight', None)
42
+
43
+ def forward(self, x):
44
+ return rms_norm(x.float(), self.weight, self.eps).to(dtype=x.dtype)
45
+
46
+ class LPRMSNorm(RMSNorm):
47
+
48
+ def __init__(self, normalized_shape, eps=1e-05, weight=True, dtype=None, device=None):
49
+ super().__init__(normalized_shape=normalized_shape, eps=eps, weight=weight, dtype=dtype, device=device)
50
+
51
+ def forward(self, x):
52
+ downcast_x = _cast_if_autocast_enabled(x)
53
+ downcast_weight = _cast_if_autocast_enabled(self.weight) if self.weight is not None else self.weight
54
+ with torch.autocast(enabled=False, device_type=x.device.type):
55
+ return rms_norm(downcast_x, downcast_weight, self.eps).to(dtype=x.dtype)
56
+ NORM_CLASS_REGISTRY = {'layernorm': torch.nn.LayerNorm, 'low_precision_layernorm': LPLayerNorm, 'rmsnorm': RMSNorm, 'low_precision_rmsnorm': LPRMSNorm}
param_init_fns.py ADDED
@@ -0,0 +1,181 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import warnings
3
+ from collections.abc import Sequence
4
+ from functools import partial
5
+ from typing import Optional, Tuple, Union
6
+ import torch
7
+ from torch import nn
8
+ from .norm import NORM_CLASS_REGISTRY
9
+
10
+ def torch_default_param_init_fn_(module: nn.Module, verbose: int=0, **kwargs):
11
+ del kwargs
12
+ if verbose > 1:
13
+ warnings.warn(f"Initializing network using module's reset_parameters attribute")
14
+ if hasattr(module, 'reset_parameters'):
15
+ module.reset_parameters()
16
+
17
+ def fused_init_helper_(module: nn.Module, init_fn_):
18
+ _fused = getattr(module, '_fused', None)
19
+ if _fused is None:
20
+ raise RuntimeError(f'Internal logic error')
21
+ (dim, splits) = _fused
22
+ splits = (0, *splits, module.weight.size(dim))
23
+ for (s, e) in zip(splits[:-1], splits[1:]):
24
+ slice_indices = [slice(None)] * module.weight.ndim
25
+ slice_indices[dim] = slice(s, e)
26
+ init_fn_(module.weight[slice_indices])
27
+
28
+ def generic_param_init_fn_(module: nn.Module, init_fn_, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, verbose: int=0, **kwargs):
29
+ del kwargs
30
+ if verbose > 1:
31
+ warnings.warn(f'If model has bias parameters they are initialized to 0.')
32
+ init_div_is_residual = init_div_is_residual
33
+ if init_div_is_residual is False:
34
+ div_is_residual = 1.0
35
+ elif init_div_is_residual is True:
36
+ div_is_residual = math.sqrt(2 * n_layers)
37
+ elif isinstance(init_div_is_residual, float) or isinstance(init_div_is_residual, int):
38
+ div_is_residual = init_div_is_residual
39
+ elif isinstance(init_div_is_residual, str) and init_div_is_residual.isnumeric():
40
+ div_is_residual = float(init_div_is_residual)
41
+ else:
42
+ div_is_residual = 1.0
43
+ raise ValueError(f'Expected init_div_is_residual to be boolean or numeric, got {init_div_is_residual}')
44
+ if init_div_is_residual is not False:
45
+ if verbose > 1:
46
+ warnings.warn(f'Initializing _is_residual layers then dividing them by {div_is_residual:.3f}. ' + f'Set `init_div_is_residual: false` in init config to disable this.')
47
+ if isinstance(module, nn.Linear):
48
+ if hasattr(module, '_fused'):
49
+ fused_init_helper_(module, init_fn_)
50
+ else:
51
+ init_fn_(module.weight)
52
+ if module.bias is not None:
53
+ torch.nn.init.zeros_(module.bias)
54
+ if init_div_is_residual is not False and getattr(module, '_is_residual', False):
55
+ with torch.no_grad():
56
+ module.weight.div_(div_is_residual)
57
+ elif isinstance(module, nn.Embedding):
58
+ if emb_init_std is not None:
59
+ std = emb_init_std
60
+ if std == 0:
61
+ warnings.warn(f'Embedding layer initialized to 0.')
62
+ emb_init_fn_ = partial(torch.nn.init.normal_, mean=0.0, std=std)
63
+ if verbose > 1:
64
+ warnings.warn(f'Embedding layer initialized using normal distribution with mean=0 and std={std!r}.')
65
+ elif emb_init_uniform_lim is not None:
66
+ lim = emb_init_uniform_lim
67
+ if isinstance(lim, Sequence):
68
+ if len(lim) > 2:
69
+ raise ValueError(f'Uniform init requires a min and a max limit. User input: {lim}.')
70
+ if lim[0] == lim[1]:
71
+ warnings.warn(f'Embedding layer initialized to {lim[0]}.')
72
+ else:
73
+ if lim == 0:
74
+ warnings.warn(f'Embedding layer initialized to 0.')
75
+ lim = [-lim, lim]
76
+ (a, b) = lim
77
+ emb_init_fn_ = partial(torch.nn.init.uniform_, a=a, b=b)
78
+ if verbose > 1:
79
+ warnings.warn(f'Embedding layer initialized using uniform distribution in range {lim}.')
80
+ else:
81
+ emb_init_fn_ = init_fn_
82
+ emb_init_fn_(module.weight)
83
+ elif isinstance(module, tuple(set(NORM_CLASS_REGISTRY.values()))):
84
+ if verbose > 1:
85
+ warnings.warn(f'Norm weights are set to 1. If norm layer has a bias it is initialized to 0.')
86
+ if hasattr(module, 'weight') and module.weight is not None:
87
+ torch.nn.init.ones_(module.weight)
88
+ if hasattr(module, 'bias') and module.bias is not None:
89
+ torch.nn.init.zeros_(module.bias)
90
+ elif isinstance(module, nn.MultiheadAttention):
91
+ if module._qkv_same_embed_dim:
92
+ assert module.in_proj_weight is not None
93
+ assert module.q_proj_weight is None and module.k_proj_weight is None and (module.v_proj_weight is None)
94
+ assert d_model is not None
95
+ _d = d_model
96
+ splits = (0, _d, 2 * _d, 3 * _d)
97
+ for (s, e) in zip(splits[:-1], splits[1:]):
98
+ init_fn_(module.in_proj_weight[s:e])
99
+ else:
100
+ assert module.q_proj_weight is not None and module.k_proj_weight is not None and (module.v_proj_weight is not None)
101
+ assert module.in_proj_weight is None
102
+ init_fn_(module.q_proj_weight)
103
+ init_fn_(module.k_proj_weight)
104
+ init_fn_(module.v_proj_weight)
105
+ if module.in_proj_bias is not None:
106
+ torch.nn.init.zeros_(module.in_proj_bias)
107
+ if module.bias_k is not None:
108
+ torch.nn.init.zeros_(module.bias_k)
109
+ if module.bias_v is not None:
110
+ torch.nn.init.zeros_(module.bias_v)
111
+ init_fn_(module.out_proj.weight)
112
+ if init_div_is_residual is not False and getattr(module.out_proj, '_is_residual', False):
113
+ with torch.no_grad():
114
+ module.out_proj.weight.div_(div_is_residual)
115
+ if module.out_proj.bias is not None:
116
+ torch.nn.init.zeros_(module.out_proj.bias)
117
+ else:
118
+ for _ in module.parameters(recurse=False):
119
+ raise NotImplementedError(f'{module.__class__.__name__} parameters are not initialized by param_init_fn.')
120
+
121
+ def _normal_init_(std, mean=0.0):
122
+ return partial(torch.nn.init.normal_, mean=mean, std=std)
123
+
124
+ def _normal_param_init_fn_(module: nn.Module, std: float, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, verbose: int=0, **kwargs):
125
+ del kwargs
126
+ init_fn_ = _normal_init_(std=std)
127
+ if verbose > 1:
128
+ warnings.warn(f'Using torch.nn.init.normal_ init fn mean=0.0, std={std}')
129
+ generic_param_init_fn_(module=module, init_fn_=init_fn_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
130
+
131
+ def baseline_param_init_fn_(module: nn.Module, init_std: float, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, verbose: int=0, **kwargs):
132
+ del kwargs
133
+ if init_std is None:
134
+ raise ValueError("You must set model.init_config['init_std'] to a float value to use the default initialization scheme.")
135
+ _normal_param_init_fn_(module=module, std=init_std, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
136
+
137
+ def small_param_init_fn_(module: nn.Module, n_layers: int, d_model: int, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, verbose: int=0, **kwargs):
138
+ del kwargs
139
+ std = math.sqrt(2 / (5 * d_model))
140
+ _normal_param_init_fn_(module=module, std=std, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
141
+
142
+ def neox_param_init_fn_(module: nn.Module, n_layers: int, d_model: int, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, verbose: int=0, **kwargs):
143
+ """From section 2.3.1 of GPT-NeoX-20B:
144
+
145
+ An Open-Source AutoregressiveLanguage Model — Black et. al. (2022)
146
+ see https://github.com/EleutherAI/gpt-neox/blob/9610391ab319403cef079b438edd016a2443af54/megatron/model/init_functions.py#L151
147
+ and https://github.com/EleutherAI/gpt-neox/blob/main/megatron/model/transformer.py
148
+ """
149
+ del kwargs
150
+ residual_div = n_layers / math.sqrt(10)
151
+ if verbose > 1:
152
+ warnings.warn(f'setting init_div_is_residual to {residual_div}')
153
+ small_param_init_fn_(module=module, d_model=d_model, n_layers=n_layers, init_div_is_residual=residual_div, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
154
+
155
+ def kaiming_uniform_param_init_fn_(module: nn.Module, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, init_gain: float=0, fan_mode: str='fan_in', init_nonlinearity: str='leaky_relu', verbose: int=0, **kwargs):
156
+ del kwargs
157
+ if verbose > 1:
158
+ warnings.warn(f'Using nn.init.kaiming_uniform_ init fn with parameters: ' + f'a={init_gain}, mode={fan_mode}, nonlinearity={init_nonlinearity}')
159
+ kaiming_uniform_ = partial(nn.init.kaiming_uniform_, a=init_gain, mode=fan_mode, nonlinearity=init_nonlinearity)
160
+ generic_param_init_fn_(module=module, init_fn_=kaiming_uniform_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
161
+
162
+ def kaiming_normal_param_init_fn_(module: nn.Module, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, init_gain: float=0, fan_mode: str='fan_in', init_nonlinearity: str='leaky_relu', verbose: int=0, **kwargs):
163
+ del kwargs
164
+ if verbose > 1:
165
+ warnings.warn(f'Using nn.init.kaiming_normal_ init fn with parameters: ' + f'a={init_gain}, mode={fan_mode}, nonlinearity={init_nonlinearity}')
166
+ kaiming_normal_ = partial(torch.nn.init.kaiming_normal_, a=init_gain, mode=fan_mode, nonlinearity=init_nonlinearity)
167
+ generic_param_init_fn_(module=module, init_fn_=kaiming_normal_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
168
+
169
+ def xavier_uniform_param_init_fn_(module: nn.Module, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, init_gain: float=0, verbose: int=0, **kwargs):
170
+ del kwargs
171
+ xavier_uniform_ = partial(torch.nn.init.xavier_uniform_, gain=init_gain)
172
+ if verbose > 1:
173
+ warnings.warn(f'Using torch.nn.init.xavier_uniform_ init fn with parameters: ' + f'gain={init_gain}')
174
+ generic_param_init_fn_(module=module, init_fn_=xavier_uniform_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
175
+
176
+ def xavier_normal_param_init_fn_(module: nn.Module, n_layers: int, d_model: Optional[int]=None, init_div_is_residual: Union[int, float, str, bool]=True, emb_init_std: Optional[float]=None, emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]]=None, init_gain: float=0, verbose: int=0, **kwargs):
177
+ xavier_normal_ = partial(torch.nn.init.xavier_normal_, gain=init_gain)
178
+ if verbose > 1:
179
+ warnings.warn(f'Using torch.nn.init.xavier_normal_ init fn with parameters: ' + f'gain={init_gain}')
180
+ generic_param_init_fn_(module=module, init_fn_=xavier_normal_, d_model=d_model, n_layers=n_layers, init_div_is_residual=init_div_is_residual, emb_init_std=emb_init_std, emb_init_uniform_lim=emb_init_uniform_lim, verbose=verbose)
181
+ MODEL_INIT_REGISTRY = {'default_': torch_default_param_init_fn_, 'baseline_': baseline_param_init_fn_, 'kaiming_uniform_': kaiming_uniform_param_init_fn_, 'kaiming_normal_': kaiming_normal_param_init_fn_, 'neox_init_': neox_param_init_fn_, 'small_init_': small_param_init_fn_, 'xavier_uniform_': xavier_uniform_param_init_fn_, 'xavier_normal_': xavier_normal_param_init_fn_}
pytorch_model.bin.index (2).json ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 13298573312
4
+ },
5
+ "weight_map": {
6
+ "transformer.blocks.0.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
7
+ "transformer.blocks.0.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
8
+ "transformer.blocks.0.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
9
+ "transformer.blocks.0.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
10
+ "transformer.blocks.0.norm_1.weight": "pytorch_model-00001-of-00002.bin",
11
+ "transformer.blocks.0.norm_2.weight": "pytorch_model-00001-of-00002.bin",
12
+ "transformer.blocks.1.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
13
+ "transformer.blocks.1.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
14
+ "transformer.blocks.1.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
15
+ "transformer.blocks.1.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
16
+ "transformer.blocks.1.norm_1.weight": "pytorch_model-00001-of-00002.bin",
17
+ "transformer.blocks.1.norm_2.weight": "pytorch_model-00001-of-00002.bin",
18
+ "transformer.blocks.10.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
19
+ "transformer.blocks.10.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
20
+ "transformer.blocks.10.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
21
+ "transformer.blocks.10.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
22
+ "transformer.blocks.10.norm_1.weight": "pytorch_model-00001-of-00002.bin",
23
+ "transformer.blocks.10.norm_2.weight": "pytorch_model-00001-of-00002.bin",
24
+ "transformer.blocks.11.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
25
+ "transformer.blocks.11.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
26
+ "transformer.blocks.11.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
27
+ "transformer.blocks.11.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
28
+ "transformer.blocks.11.norm_1.weight": "pytorch_model-00001-of-00002.bin",
29
+ "transformer.blocks.11.norm_2.weight": "pytorch_model-00001-of-00002.bin",
30
+ "transformer.blocks.12.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
31
+ "transformer.blocks.12.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
32
+ "transformer.blocks.12.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
33
+ "transformer.blocks.12.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
34
+ "transformer.blocks.12.norm_1.weight": "pytorch_model-00001-of-00002.bin",
35
+ "transformer.blocks.12.norm_2.weight": "pytorch_model-00001-of-00002.bin",
36
+ "transformer.blocks.13.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
37
+ "transformer.blocks.13.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
38
+ "transformer.blocks.13.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
39
+ "transformer.blocks.13.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
40
+ "transformer.blocks.13.norm_1.weight": "pytorch_model-00001-of-00002.bin",
41
+ "transformer.blocks.13.norm_2.weight": "pytorch_model-00001-of-00002.bin",
42
+ "transformer.blocks.14.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
43
+ "transformer.blocks.14.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
44
+ "transformer.blocks.14.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
45
+ "transformer.blocks.14.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
46
+ "transformer.blocks.14.norm_1.weight": "pytorch_model-00001-of-00002.bin",
47
+ "transformer.blocks.14.norm_2.weight": "pytorch_model-00001-of-00002.bin",
48
+ "transformer.blocks.15.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
49
+ "transformer.blocks.15.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
50
+ "transformer.blocks.15.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
51
+ "transformer.blocks.15.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
52
+ "transformer.blocks.15.norm_1.weight": "pytorch_model-00001-of-00002.bin",
53
+ "transformer.blocks.15.norm_2.weight": "pytorch_model-00001-of-00002.bin",
54
+ "transformer.blocks.16.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
55
+ "transformer.blocks.16.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
56
+ "transformer.blocks.16.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
57
+ "transformer.blocks.16.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
58
+ "transformer.blocks.16.norm_1.weight": "pytorch_model-00001-of-00002.bin",
59
+ "transformer.blocks.16.norm_2.weight": "pytorch_model-00001-of-00002.bin",
60
+ "transformer.blocks.17.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
61
+ "transformer.blocks.17.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
62
+ "transformer.blocks.17.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
63
+ "transformer.blocks.17.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
64
+ "transformer.blocks.17.norm_1.weight": "pytorch_model-00001-of-00002.bin",
65
+ "transformer.blocks.17.norm_2.weight": "pytorch_model-00001-of-00002.bin",
66
+ "transformer.blocks.18.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
67
+ "transformer.blocks.18.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
68
+ "transformer.blocks.18.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
69
+ "transformer.blocks.18.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
70
+ "transformer.blocks.18.norm_1.weight": "pytorch_model-00001-of-00002.bin",
71
+ "transformer.blocks.18.norm_2.weight": "pytorch_model-00001-of-00002.bin",
72
+ "transformer.blocks.19.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
73
+ "transformer.blocks.19.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
74
+ "transformer.blocks.19.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
75
+ "transformer.blocks.19.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
76
+ "transformer.blocks.19.norm_1.weight": "pytorch_model-00001-of-00002.bin",
77
+ "transformer.blocks.19.norm_2.weight": "pytorch_model-00001-of-00002.bin",
78
+ "transformer.blocks.2.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
79
+ "transformer.blocks.2.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
80
+ "transformer.blocks.2.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
81
+ "transformer.blocks.2.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
82
+ "transformer.blocks.2.norm_1.weight": "pytorch_model-00001-of-00002.bin",
83
+ "transformer.blocks.2.norm_2.weight": "pytorch_model-00001-of-00002.bin",
84
+ "transformer.blocks.20.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
85
+ "transformer.blocks.20.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
86
+ "transformer.blocks.20.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
87
+ "transformer.blocks.20.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
88
+ "transformer.blocks.20.norm_1.weight": "pytorch_model-00001-of-00002.bin",
89
+ "transformer.blocks.20.norm_2.weight": "pytorch_model-00001-of-00002.bin",
90
+ "transformer.blocks.21.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
91
+ "transformer.blocks.21.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
92
+ "transformer.blocks.21.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
93
+ "transformer.blocks.21.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
94
+ "transformer.blocks.21.norm_1.weight": "pytorch_model-00001-of-00002.bin",
95
+ "transformer.blocks.21.norm_2.weight": "pytorch_model-00001-of-00002.bin",
96
+ "transformer.blocks.22.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
97
+ "transformer.blocks.22.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
98
+ "transformer.blocks.22.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
99
+ "transformer.blocks.22.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
100
+ "transformer.blocks.22.norm_1.weight": "pytorch_model-00001-of-00002.bin",
101
+ "transformer.blocks.22.norm_2.weight": "pytorch_model-00001-of-00002.bin",
102
+ "transformer.blocks.23.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
103
+ "transformer.blocks.23.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
104
+ "transformer.blocks.23.ffn.down_proj.weight": "pytorch_model-00002-of-00002.bin",
105
+ "transformer.blocks.23.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
106
+ "transformer.blocks.23.norm_1.weight": "pytorch_model-00001-of-00002.bin",
107
+ "transformer.blocks.23.norm_2.weight": "pytorch_model-00001-of-00002.bin",
108
+ "transformer.blocks.24.attn.Wqkv.weight": "pytorch_model-00002-of-00002.bin",
109
+ "transformer.blocks.24.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
110
+ "transformer.blocks.24.ffn.down_proj.weight": "pytorch_model-00002-of-00002.bin",
111
+ "transformer.blocks.24.ffn.up_proj.weight": "pytorch_model-00002-of-00002.bin",
112
+ "transformer.blocks.24.norm_1.weight": "pytorch_model-00002-of-00002.bin",
113
+ "transformer.blocks.24.norm_2.weight": "pytorch_model-00002-of-00002.bin",
114
+ "transformer.blocks.25.attn.Wqkv.weight": "pytorch_model-00002-of-00002.bin",
115
+ "transformer.blocks.25.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
116
+ "transformer.blocks.25.ffn.down_proj.weight": "pytorch_model-00002-of-00002.bin",
117
+ "transformer.blocks.25.ffn.up_proj.weight": "pytorch_model-00002-of-00002.bin",
118
+ "transformer.blocks.25.norm_1.weight": "pytorch_model-00002-of-00002.bin",
119
+ "transformer.blocks.25.norm_2.weight": "pytorch_model-00002-of-00002.bin",
120
+ "transformer.blocks.26.attn.Wqkv.weight": "pytorch_model-00002-of-00002.bin",
121
+ "transformer.blocks.26.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
122
+ "transformer.blocks.26.ffn.down_proj.weight": "pytorch_model-00002-of-00002.bin",
123
+ "transformer.blocks.26.ffn.up_proj.weight": "pytorch_model-00002-of-00002.bin",
124
+ "transformer.blocks.26.norm_1.weight": "pytorch_model-00002-of-00002.bin",
125
+ "transformer.blocks.26.norm_2.weight": "pytorch_model-00002-of-00002.bin",
126
+ "transformer.blocks.27.attn.Wqkv.weight": "pytorch_model-00002-of-00002.bin",
127
+ "transformer.blocks.27.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
128
+ "transformer.blocks.27.ffn.down_proj.weight": "pytorch_model-00002-of-00002.bin",
129
+ "transformer.blocks.27.ffn.up_proj.weight": "pytorch_model-00002-of-00002.bin",
130
+ "transformer.blocks.27.norm_1.weight": "pytorch_model-00002-of-00002.bin",
131
+ "transformer.blocks.27.norm_2.weight": "pytorch_model-00002-of-00002.bin",
132
+ "transformer.blocks.28.attn.Wqkv.weight": "pytorch_model-00002-of-00002.bin",
133
+ "transformer.blocks.28.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
134
+ "transformer.blocks.28.ffn.down_proj.weight": "pytorch_model-00002-of-00002.bin",
135
+ "transformer.blocks.28.ffn.up_proj.weight": "pytorch_model-00002-of-00002.bin",
136
+ "transformer.blocks.28.norm_1.weight": "pytorch_model-00002-of-00002.bin",
137
+ "transformer.blocks.28.norm_2.weight": "pytorch_model-00002-of-00002.bin",
138
+ "transformer.blocks.29.attn.Wqkv.weight": "pytorch_model-00002-of-00002.bin",
139
+ "transformer.blocks.29.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
140
+ "transformer.blocks.29.ffn.down_proj.weight": "pytorch_model-00002-of-00002.bin",
141
+ "transformer.blocks.29.ffn.up_proj.weight": "pytorch_model-00002-of-00002.bin",
142
+ "transformer.blocks.29.norm_1.weight": "pytorch_model-00002-of-00002.bin",
143
+ "transformer.blocks.29.norm_2.weight": "pytorch_model-00002-of-00002.bin",
144
+ "transformer.blocks.3.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
145
+ "transformer.blocks.3.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
146
+ "transformer.blocks.3.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
147
+ "transformer.blocks.3.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
148
+ "transformer.blocks.3.norm_1.weight": "pytorch_model-00001-of-00002.bin",
149
+ "transformer.blocks.3.norm_2.weight": "pytorch_model-00001-of-00002.bin",
150
+ "transformer.blocks.30.attn.Wqkv.weight": "pytorch_model-00002-of-00002.bin",
151
+ "transformer.blocks.30.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
152
+ "transformer.blocks.30.ffn.down_proj.weight": "pytorch_model-00002-of-00002.bin",
153
+ "transformer.blocks.30.ffn.up_proj.weight": "pytorch_model-00002-of-00002.bin",
154
+ "transformer.blocks.30.norm_1.weight": "pytorch_model-00002-of-00002.bin",
155
+ "transformer.blocks.30.norm_2.weight": "pytorch_model-00002-of-00002.bin",
156
+ "transformer.blocks.31.attn.Wqkv.weight": "pytorch_model-00002-of-00002.bin",
157
+ "transformer.blocks.31.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
158
+ "transformer.blocks.31.ffn.down_proj.weight": "pytorch_model-00002-of-00002.bin",
159
+ "transformer.blocks.31.ffn.up_proj.weight": "pytorch_model-00002-of-00002.bin",
160
+ "transformer.blocks.31.norm_1.weight": "pytorch_model-00002-of-00002.bin",
161
+ "transformer.blocks.31.norm_2.weight": "pytorch_model-00002-of-00002.bin",
162
+ "transformer.blocks.4.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
163
+ "transformer.blocks.4.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
164
+ "transformer.blocks.4.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
165
+ "transformer.blocks.4.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
166
+ "transformer.blocks.4.norm_1.weight": "pytorch_model-00001-of-00002.bin",
167
+ "transformer.blocks.4.norm_2.weight": "pytorch_model-00001-of-00002.bin",
168
+ "transformer.blocks.5.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
169
+ "transformer.blocks.5.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
170
+ "transformer.blocks.5.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
171
+ "transformer.blocks.5.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
172
+ "transformer.blocks.5.norm_1.weight": "pytorch_model-00001-of-00002.bin",
173
+ "transformer.blocks.5.norm_2.weight": "pytorch_model-00001-of-00002.bin",
174
+ "transformer.blocks.6.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
175
+ "transformer.blocks.6.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
176
+ "transformer.blocks.6.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
177
+ "transformer.blocks.6.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
178
+ "transformer.blocks.6.norm_1.weight": "pytorch_model-00001-of-00002.bin",
179
+ "transformer.blocks.6.norm_2.weight": "pytorch_model-00001-of-00002.bin",
180
+ "transformer.blocks.7.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
181
+ "transformer.blocks.7.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
182
+ "transformer.blocks.7.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
183
+ "transformer.blocks.7.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
184
+ "transformer.blocks.7.norm_1.weight": "pytorch_model-00001-of-00002.bin",
185
+ "transformer.blocks.7.norm_2.weight": "pytorch_model-00001-of-00002.bin",
186
+ "transformer.blocks.8.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
187
+ "transformer.blocks.8.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
188
+ "transformer.blocks.8.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
189
+ "transformer.blocks.8.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
190
+ "transformer.blocks.8.norm_1.weight": "pytorch_model-00001-of-00002.bin",
191
+ "transformer.blocks.8.norm_2.weight": "pytorch_model-00001-of-00002.bin",
192
+ "transformer.blocks.9.attn.Wqkv.weight": "pytorch_model-00001-of-00002.bin",
193
+ "transformer.blocks.9.attn.out_proj.weight": "pytorch_model-00001-of-00002.bin",
194
+ "transformer.blocks.9.ffn.down_proj.weight": "pytorch_model-00001-of-00002.bin",
195
+ "transformer.blocks.9.ffn.up_proj.weight": "pytorch_model-00001-of-00002.bin",
196
+ "transformer.blocks.9.norm_1.weight": "pytorch_model-00001-of-00002.bin",
197
+ "transformer.blocks.9.norm_2.weight": "pytorch_model-00001-of-00002.bin",
198
+ "transformer.norm_f.weight": "pytorch_model-00002-of-00002.bin",
199
+ "transformer.wte.weight": "pytorch_model-00001-of-00002.bin"
200
+ }
201
+ }
special_tokens_map.json ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ {
2
+ "bos_token": "<|endoftext|>",
3
+ "eos_token": "<|endoftext|>",
4
+ "unk_token": "<|endoftext|>"
5
+ }
tokenizer (1).json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config (2).json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "bos_token": "<|endoftext|>",
4
+ "clean_up_tokenization_spaces": true,
5
+ "eos_token": "<|endoftext|>",
6
+ "model_max_length": 2048,
7
+ "tokenizer_class": "GPTNeoXTokenizer",
8
+ "unk_token": "<|endoftext|>"
9
+ }