forked from pytorch/torchtune
-
Notifications
You must be signed in to change notification settings - Fork 0
/
_model_builders.py
189 lines (160 loc) · 6.05 KB
/
_model_builders.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
from typing import List
from torchtune.models.gemma._component_builders import gemma, lora_gemma
from torchtune.models.gemma.transformer import GemmaTransformerDecoder
from torchtune.modules.tokenizers import SentencePieceTokenizer
from torchtune.modules.peft import LORA_ATTN_MODULES
from functools import partial
"""
Model builders build specific instantiations using component builders. For example
the ``gemma_2b`` model builder uses the ``gemma`` component builder.
"""
def gemma_2b() -> GemmaTransformerDecoder:
"""
Builder for creating a Gemma 2B model initialized w/ the default 2b parameter values
from: https://blog.google/technology/developers/gemma-open-models/
Returns:
GemmaTransformerDecoder: Instantiation of Gemma 2B model
"""
return gemma(
vocab_size=256_000,
num_layers=18,
num_heads=8,
head_dim=256,
num_kv_heads=1,
embed_dim=2048,
intermediate_dim=16384,
max_seq_len=8192,
attn_dropout=0.0,
norm_eps=1e-6,
)
def gemma_tokenizer(path: str) -> SentencePieceTokenizer:
"""
Tokenizer for Gemma.
Args:
path (str): path to the tokenizer
Returns:
SentencePieceTokenizer: Instantiation of the Gemma tokenizer
"""
tokenizer = SentencePieceTokenizer(path)
tokenizer.pad_id = 0
return tokenizer
def lora_gemma_2b(
lora_attn_modules: List[LORA_ATTN_MODULES],
apply_lora_to_mlp: bool = False,
lora_rank: int = 8,
lora_alpha: float = 16,
quantize_base: bool = False,
) -> GemmaTransformerDecoder:
"""
Builder for creating a Gemma 2B model with LoRA enabled.
The Gemma defaults are the same as in :func:`~torchtune.models.gemma.gemma_2b`,
while LoRA default params are based on
https://github.com/tloen/alpaca-lora/blob/8bb8579e403dc78e37fe81ffbb253c413007323f/finetune.py#L41-L43.
Args:
lora_attn_modules (List[LORA_ATTN_MODULES]): list of which linear layers
LoRA should be applied to in each self-attention block. Options are
``{"q_proj", "k_proj", "v_proj", "output_proj"}``.
apply_lora_to_mlp (bool): whether to apply LoRA to the MLP in each transformer layer.
Default: False
lora_rank (int): rank of each low-rank approximation
lora_alpha (float): scaling factor for the low-rank approximation
quantize_base (bool): Whether to quantize base model weights
Returns:
GemmaTransformerDecoder: Instantiation of Gemma 2B model with LoRA applied
"""
return lora_gemma(
lora_attn_modules=lora_attn_modules,
apply_lora_to_mlp=apply_lora_to_mlp,
vocab_size=256_000,
num_layers=18,
num_heads=8,
head_dim=256,
num_kv_heads=1,
embed_dim=2048,
intermediate_dim=16384,
max_seq_len=8192,
attn_dropout=0.0,
norm_eps=1e-6,
lora_rank=lora_rank,
lora_alpha=lora_alpha,
lora_dropout=0.05,
quantize_base=quantize_base,
)
qlora_gemma_2b = partial(lora_gemma_2b, quantize_base=True)
qlora_gemma_2b.__doc__ = """
Builder for creating a Gemma model with QLoRA enabled. Base model weights in linear layers
that LoRA is applied to are quantized per the QLoRA paper: https://arxiv.org/abs/2305.14314.
Please see `lora_gemma_2b` for full API arguments.
"""
def gemma_7b() -> GemmaTransformerDecoder:
"""
Builder for creating a Gemma 7B model initialized w/ the default 7b parameter values
from: https://blog.google/technology/developers/gemma-open-models/
Returns:
GemmaTransformerDecoder: Instantiation of Gemma 7B model
"""
return gemma(
vocab_size=256_000,
num_layers=28,
num_heads=16,
head_dim=256,
num_kv_heads=16,
embed_dim=3072,
intermediate_dim=24576,
max_seq_len=8192,
attn_dropout=0.0,
norm_eps=1e-6,
)
def lora_gemma_7b(
lora_attn_modules: List[LORA_ATTN_MODULES],
apply_lora_to_mlp: bool = False,
lora_rank: int = 8,
lora_alpha: float = 16,
quantize_base: bool = False,
) -> GemmaTransformerDecoder:
"""
Builder for creating a Gemma 7B model with LoRA enabled.
The Gemma defaults are the same as in :func:`~torchtune.models.gemma.gemma_7b`,
while LoRA default params are based on
https://github.com/tloen/alpaca-lora/blob/8bb8579e403dc78e37fe81ffbb253c413007323f/finetune.py#L41-L43.
Args:
lora_attn_modules (List[LORA_ATTN_MODULES]): list of which linear layers
LoRA should be applied to in each self-attention block. Options are
``{"q_proj", "k_proj", "v_proj", "output_proj"}``.
apply_lora_to_mlp (bool): whether to apply LoRA to the MLP in each transformer layer.
Default: False
lora_rank (int): rank of each low-rank approximation
lora_alpha (float): scaling factor for the low-rank approximation
quantize_base (bool): Whether to quantize base model weights
Returns:
GemmaTransformerDecoder: Instantiation of Gemma 7B model with LoRA applied
"""
return lora_gemma(
lora_attn_modules=lora_attn_modules,
apply_lora_to_mlp=apply_lora_to_mlp,
vocab_size=256_000,
num_layers=28,
num_heads=16,
head_dim=256,
num_kv_heads=16,
embed_dim=3072,
intermediate_dim=24576,
max_seq_len=8192,
attn_dropout=0.0,
norm_eps=1e-6,
lora_rank=lora_rank,
lora_alpha=lora_alpha,
lora_dropout=0.05,
quantize_base=quantize_base,
)
qlora_gemma_7b = partial(lora_gemma_7b, quantize_base=True)
qlora_gemma_7b.__doc__ = """
Builder for creating a Gemma model with QLoRA enabled. Base model weights in linear layers
that LoRA is applied to are quantized per the QLoRA paper: https://arxiv.org/abs/2305.14314.
Please see `lora_gemma_7b` for full API arguments.
"""